# -*- coding: utf-8 -*- """unet_eff_final.ipynb Automatically generated by Colaboratory. Original file is located at https://colab.research.google.com/drive/1qdwBOU4hJn0LA2neC9Rzo1wA5Tec5h5N """ !pip uninstall tensorflow !pip install tensorflow==2.9 # Commented out IPython magic to ensure Python compatibility. import os import cv2 import glob import PIL import shutil import numpy as np import pandas as pd import seaborn as sns import matplotlib.pyplot as plt from skimage import data from skimage.util import montage import skimage.transform as skTrans from skimage.transform import rotate from skimage.transform import resize from PIL import Image, ImageOps # neural imaging import nibabel as nib import albumentations as A # ml libs import keras import keras.backend as K from keras.callbacks import CSVLogger import tensorflow as tf from tensorflow.keras.utils import plot_model from sklearn.preprocessing import MinMaxScaler from sklearn.model_selection import train_test_split from sklearn.metrics import classification_report from tensorflow.keras.models import * from tensorflow.keras.layers import * from tensorflow.keras.optimizers import * from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, EarlyStopping, TensorBoard from tensorflow.keras.layers.experimental import preprocessing # %matplotlib inline VOLUME_SLICES = 128 #裁減多於黑邊 VOLUME_START_AT = 13 #裁減多於黑邊 IMG_SIZE=128 #圖片大小 TRAIN_DATASET_PATH = '/content/MyDrive/BraTS/'#資料集路徑 """**移除瑕疵圖**""" def youn(path):#path為資料集的資料夾名稱 type_img=["flair","t1ce","t1","t2"] type_Area=list() for i in range(4): image = nib.load(TRAIN_DATASET_PATH+path+"/"+path+"_"+type_img[i]+".nii").get_fdata() image=image[:,:,106] # 正規化圖像 image = image / np.max(image) # 正規化圖像至 0-1 範圍 # 二值化圖像 _, threshold = cv2.threshold(image, 0, 1, cv2.THRESH_BINARY) # 查找輪廓 contours, _ = cv2.findContours(threshold.astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) for contour in contours: # 獲取矩形範圍 if len(contours)>1: ellipse= cv2.contourArea(contours[-1]) else: ellipse= cv2.contourArea(contour) type_Area.append(ellipse) return type_Area da=os.listdir(TRAIN_DATASET_PATH)#資料集路徑 image_len=list() for i in range(len(da)): image_len.append(youn(da[i]))#path為資料集的資料夾名稱 image_len=np.array(image_len) '''Dixon考驗法''' asc=list() for i in range(len(da)): asd=image_len[i,:] asd.sort() asd[3]=asd[3]+1 if (asd[1]-asd[0])/(asd[3]-asd[0])>0.964: asc.append("True") else: asc.append("False") train_ids=['BraTS19_CBICA_BAX_1', 'BraTS19_TCIA02_300_1', 'BraTS19_CBICA_BNR_1', 'BraTS19_CBICA_AYI_1', 'BraTS19_CBICA_ALX_1', 'BraTS19_CBICA_AQO_1', 'BraTS19_TCIA06_247_1', 'BraTS19_CBICA_ANI_1', 'BraTS19_CBICA_ABO_1', 'BraTS19_TCIA04_149_1', 'BraTS19_CBICA_BAP_1', 'BraTS19_2013_2_1', 'BraTS19_CBICA_ATD_1', 'BraTS19_CBICA_BHK_1', 'BraTS19_2013_3_1', 'BraTS19_CBICA_ANG_1', 'BraTS19_CBICA_ABY_1', 'BraTS19_CBICA_BHZ_1', 'BraTS19_TCIA02_322_1', 'BraTS19_CBICA_AQR_1', 'BraTS19_CBICA_BHM_1', 'BraTS19_CBICA_AQD_1', 'BraTS19_TCIA01_429_1', 'BraTS19_TCIA02_118_1', 'BraTS19_TCIA08_105_1', 'BraTS19_CBICA_BGO_1', 'BraTS19_CBICA_AYA_1', 'BraTS19_CBICA_BGX_1', 'BraTS19_CBICA_APZ_1', 'BraTS19_CBICA_ABM_1', 'BraTS19_CBICA_AOS_1', 'BraTS19_TCIA06_372_1', 'BraTS19_CBICA_ASN_1', 'BraTS19_CBICA_AYG_1', 'BraTS19_CBICA_ASO_1', 'BraTS19_TCIA02_368_1', 'BraTS19_CBICA_AZD_1', 'BraTS19_CBICA_AUQ_1', 'BraTS19_2013_5_1', 'BraTS19_TCIA02_290_1', 'BraTS19_CBICA_AVF_1', 'BraTS19_TCIA01_412_1', 'BraTS19_TCIA06_332_1', 'BraTS19_CBICA_AVJ_1', 'BraTS19_2013_11_1', 'BraTS19_CBICA_BGN_1', 'BraTS19_CBICA_AUW_1', 'BraTS19_CBICA_AOP_1', 'BraTS19_CBICA_ANV_1', 'BraTS19_TCIA03_121_1', 'BraTS19_TCIA02_151_1', 'BraTS19_TCIA08_234_1', 'BraTS19_CBICA_BKV_1', 'BraTS19_TCIA08_280_1', 'BraTS19_TCIA05_396_1', 'BraTS19_CBICA_ABE_1', 'BraTS19_CBICA_ATF_1', 'BraTS19_CBICA_BJY_1', 'BraTS19_TCIA02_321_1', 'BraTS19_TCIA02_471_1', 'BraTS19_CBICA_AQY_1', 'BraTS19_2013_14_1', 'BraTS19_CBICA_ANZ_1', 'BraTS19_TCIA02_208_1', 'BraTS19_CBICA_AVB_1', 'BraTS19_TCIA02_374_1', 'BraTS19_TMC_06290_1', 'BraTS19_CBICA_AQU_1', 'BraTS19_TCIA02_608_1', 'BraTS19_CBICA_APK_1', 'BraTS19_CBICA_ASF_1', 'BraTS19_CBICA_BBG_1', 'BraTS19_TCIA08_406_1', 'BraTS19_TCIA06_184_1', 'BraTS19_CBICA_AQJ_1', 'BraTS19_CBICA_AUR_1', 'BraTS19_CBICA_AYC_1', 'BraTS19_2013_19_1', 'BraTS19_CBICA_AWH_1', 'BraTS19_CBICA_AAP_1', 'BraTS19_CBICA_BGW_1', 'BraTS19_TCIA06_211_1', 'BraTS19_TCIA01_378_1', 'BraTS19_CBICA_AOO_1', 'BraTS19_TCIA08_242_1', 'BraTS19_TCIA03_257_1', 'BraTS19_TCIA02_283_1', 'BraTS19_CBICA_AXO_1', 'BraTS19_CBICA_BLJ_1', 'BraTS19_TCIA02_377_1', 'BraTS19_TCIA03_338_1', 'BraTS19_TCIA01_235_1', 'BraTS19_TCIA08_218_1', 'BraTS19_CBICA_AWV_1', 'BraTS19_CBICA_ARW_1', 'BraTS19_CBICA_AWG_1', 'BraTS19_CBICA_AOC_1', 'BraTS19_2013_23_1', 'BraTS19_TCIA01_460_1', 'BraTS19_TCIA01_190_1', 'BraTS19_2013_18_1', 'BraTS19_CBICA_AQZ_1', 'BraTS19_CBICA_BEM_1', 'BraTS19_CBICA_AQN_1', 'BraTS19_CBICA_BAN_1', 'BraTS19_TCIA01_390_1', 'BraTS19_TCIA01_147_1', 'BraTS19_TCIA02_222_1', 'BraTS19_CBICA_ASG_1', 'BraTS19_CBICA_AAB_1', 'BraTS19_CBICA_AUX_1', 'BraTS19_CBICA_AXN_1', 'BraTS19_CBICA_AVV_1', 'BraTS19_CBICA_AZH_1', 'BraTS19_TCIA08_469_1', 'BraTS19_TCIA04_437_1', 'BraTS19_2013_22_1', 'BraTS19_TCIA01_411_1', 'BraTS19_CBICA_AQT_1', 'BraTS19_CBICA_AXW_1', 'BraTS19_TCIA04_479_1', 'BraTS19_2013_4_1', 'BraTS19_CBICA_AXM_1', 'BraTS19_TMC_15477_1', 'BraTS19_CBICA_ARZ_1', 'BraTS19_TMC_11964_1', 'BraTS19_TCIA02_430_1', 'BraTS19_TCIA02_135_1', 'BraTS19_TCIA08_167_1', 'BraTS19_CBICA_AWI_1', 'BraTS19_TMC_21360_1', 'BraTS19_CBICA_AXJ_1', 'BraTS19_TCIA03_375_1', 'BraTS19_CBICA_ASK_1', 'BraTS19_TCIA01_150_1', 'BraTS19_CBICA_BGR_1', 'BraTS19_2013_17_1', 'BraTS19_CBICA_ANP_1', 'BraTS19_TCIA01_231_1', 'BraTS19_TCIA02_331_1', 'BraTS19_TMC_06643_1', 'BraTS19_TCIA02_606_1', 'BraTS19_CBICA_AQP_1', 'BraTS19_TCIA03_474_1', 'BraTS19_TMC_27374_1', 'BraTS19_2013_27_1', 'BraTS19_2013_10_1', 'BraTS19_CBICA_BFB_1', 'BraTS19_CBICA_APR_1', 'BraTS19_CBICA_AQV_1', 'BraTS19_TCIA01_180_1', 'BraTS19_CBICA_AUA_1', 'BraTS19_TCIA08_113_1', 'BraTS19_CBICA_ATB_1', 'BraTS19_CBICA_AUN_1', 'BraTS19_TCIA08_205_1', 'BraTS19_CBICA_ASE_1', 'BraTS19_TCIA01_425_1', 'BraTS19_CBICA_AAL_1', 'BraTS19_CBICA_ARF_1', 'BraTS19_TCIA03_265_1', 'BraTS19_CBICA_AQQ_1', 'BraTS19_TCIA02_370_1', 'BraTS19_TCIA08_319_1', 'BraTS19_CBICA_AMH_1', 'BraTS19_TCIA02_179_1', 'BraTS19_TCIA06_165_1', 'BraTS19_CBICA_BCF_1', 'BraTS19_CBICA_AQA_1', 'BraTS19_2013_12_1', 'BraTS19_TCIA04_328_1', 'BraTS19_2013_26_1', 'BraTS19_TCIA02_607_1', 'BraTS19_CBICA_BHQ_1', 'BraTS19_TCIA03_498_1', 'BraTS19_TCIA08_162_1', 'BraTS19_CBICA_AYW_1', 'BraTS19_CBICA_BGG_1', 'BraTS19_TCIA01_221_1', 'BraTS19_TCIA02_226_1', 'BraTS19_CBICA_AOD_1', 'BraTS19_TCIA10_640_1', 'BraTS19_TCIA10_639_1', 'BraTS19_TCIA10_261_1', 'BraTS19_TCIA10_387_1', 'BraTS19_TCIA12_480_1', 'BraTS19_TCIA13_653_1', 'BraTS19_TCIA13_633_1', 'BraTS19_TCIA12_249_1', 'BraTS19_TCIA12_470_1', 'BraTS19_TCIA10_420_1', 'BraTS19_2013_29_1', 'BraTS19_TCIA10_625_1', 'BraTS19_TCIA10_346_1', 'BraTS19_TCIA09_402_1', 'BraTS19_TCIA09_177_1', 'BraTS19_TCIA09_141_1', 'BraTS19_TCIA10_276_1', 'BraTS19_TCIA10_103_1', 'BraTS19_TCIA13_623_1', 'BraTS19_TCIA10_442_1', 'BraTS19_TCIA10_410_1', 'BraTS19_2013_16_1', 'BraTS19_TCIA10_628_1', 'BraTS19_TCIA12_298_1', 'BraTS19_2013_15_1', 'BraTS19_2013_28_1', 'BraTS19_TCIA10_241_1', 'BraTS19_TCIA10_130_1', 'BraTS19_TCIA13_618_1', 'BraTS19_TCIA13_650_1', 'BraTS19_TCIA10_490_1', 'BraTS19_TCIA13_654_1', 'BraTS19_TCIA10_449_1', 'BraTS19_TCIA13_615_1', 'BraTS19_TCIA10_413_1', 'BraTS19_TCIA10_282_1', 'BraTS19_2013_24_1', 'BraTS19_TCIA09_620_1', 'BraTS19_TCIA10_152_1', 'BraTS19_TCIA10_299_1', 'BraTS19_TCIA13_630_1', 'BraTS19_TCIA10_393_1', 'BraTS19_TCIA09_254_1', 'BraTS19_TCIA13_634_1', 'BraTS19_TCIA10_310_1', 'BraTS19_TCIA09_312_1', 'BraTS19_TMC_09043_1', 'BraTS19_TCIA10_109_1', 'BraTS19_TCIA13_621_1', 'BraTS19_TCIA10_632_1', 'BraTS19_TCIA10_266_1', 'BraTS19_TCIA12_101_1'] test_ids=['BraTS19_TCIA02_605_1', 'BraTS19_TCIA03_419_1', 'BraTS19_TCIA13_642_1', 'BraTS19_TCIA10_637_1', 'BraTS19_TCIA09_428_1', 'BraTS19_2013_1_1', 'BraTS19_CBICA_APY_1', 'BraTS19_TCIA03_296_1', 'BraTS19_CBICA_BDK_1', 'BraTS19_CBICA_BGE_1', 'BraTS19_2013_21_1', 'BraTS19_TCIA02_198_1', 'BraTS19_TCIA13_645_1', 'BraTS19_TCIA02_274_1', 'BraTS19_TCIA01_448_1', 'BraTS19_TCIA01_335_1', 'BraTS19_CBICA_AOH_1', 'BraTS19_CBICA_ALU_1', 'BraTS19_CBICA_AOZ_1', 'BraTS19_TCIA03_133_1', 'BraTS19_CBICA_ASW_1', 'BraTS19_TCIA04_361_1', 'BraTS19_TCIA10_351_1', 'BraTS19_TCIA10_408_1', 'BraTS19_TCIA09_255_1', 'BraTS19_2013_9_1', 'BraTS19_TCIA02_171_1', 'BraTS19_CBICA_AXQ_1', 'BraTS19_TCIA05_444_1', 'BraTS19_2013_13_1', 'BraTS19_CBICA_ASY_1', 'BraTS19_CBICA_BCL_1', 'BraTS19_CBICA_ASH_1', 'BraTS19_CBICA_ASU_1', 'BraTS19_CBICA_ASA_1', 'BraTS19_TCIA02_491_1', 'BraTS19_CBICA_AVT_1', 'BraTS19_CBICA_ABN_1', 'BraTS19_CBICA_AWX_1', 'BraTS19_CBICA_AXL_1', 'BraTS19_TCIA13_624_1', 'BraTS19_TCIA02_309_1', 'BraTS19_CBICA_AME_1', 'BraTS19_TCIA02_314_1', 'BraTS19_TCIA02_394_1', 'BraTS19_CBICA_ATN_1', 'BraTS19_TCIA10_175_1', 'BraTS19_TCIA10_629_1', 'BraTS19_CBICA_ATV_1'] val_ids=['BraTS19_2013_7_1', 'BraTS19_CBICA_AYU_1', 'BraTS19_TCIA05_277_1', 'BraTS19_CBICA_AVG_1', 'BraTS19_TCIA01_499_1', 'BraTS19_TCIA10_330_1', 'BraTS19_CBICA_AQG_1', 'BraTS19_TCIA02_168_1', 'BraTS19_CBICA_ALN_1', 'BraTS19_CBICA_ATX_1', 'BraTS19_TCIA10_202_1', 'BraTS19_CBICA_BIC_1', 'BraTS19_CBICA_BGT_1', 'BraTS19_TCIA09_462_1', 'BraTS19_2013_25_1', 'BraTS19_TCIA02_117_1', 'BraTS19_CBICA_BFP_1', 'BraTS19_TCIA06_603_1', 'BraTS19_CBICA_AAG_1', 'BraTS19_CBICA_BHB_1', 'BraTS19_CBICA_BHV_1', 'BraTS19_TCIA06_409_1', 'BraTS19_TCIA04_343_1', 'BraTS19_TCIA05_478_1', 'BraTS19_TCIA01_201_1', 'BraTS19_TCIA01_203_1', 'BraTS19_TCIA01_186_1', 'BraTS19_TMC_12866_1', 'BraTS19_TCIA02_455_1', 'BraTS19_TCIA02_473_1', 'BraTS19_TCIA09_451_1', 'BraTS19_TCIA01_401_1', 'BraTS19_TCIA04_111_1', 'BraTS19_TCIA09_493_1', 'BraTS19_TCIA03_138_1', 'BraTS19_TCIA04_192_1', 'BraTS19_TCIA03_199_1', 'BraTS19_TCIA01_131_1', 'BraTS19_CBICA_ASV_1', 'BraTS19_TCIA10_325_1', 'BraTS19_TMC_30014_1', 'BraTS19_CBICA_ATP_1', 'BraTS19_TCIA12_466_1', 'BraTS19_TCIA08_436_1', 'BraTS19_CBICA_ASR_1', 'BraTS19_TCIA10_644_1', 'BraTS19_TCIA08_278_1', 'BraTS19_CBICA_ABB_1', 'BraTS19_TCIA10_307_1'] """**資料集切片**""" def test(Batch_ids): cl_ass=list() for c, i in enumerate(Batch_ids): print(i) case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_flair.nii') flair = nib.load(data_path).get_fdata() data_path = os.path.join(case_path, f'{i}_t1ce.nii') ce = nib.load(data_path).get_fdata() data_path = os.path.join(case_path, f'{i}_t1.nii') t1 = nib.load(data_path).get_fdata() data_path = os.path.join(case_path, f'{i}_t2.nii') t2 = nib.load(data_path).get_fdata() data_path = os.path.join(case_path, f'{i}_seg.nii') seg = nib.load(data_path).get_fdata() img_non=np.zeros((240, 240, 4)) img_non_128=np.zeros((VOLUME_SLICES,128, 128, 4)) img_non_f=np.zeros((VOLUME_SLICES,128, 128)) img_non_1=np.zeros((VOLUME_SLICES,128, 128)) img_non_2=np.zeros((VOLUME_SLICES,128, 128)) img_non_c=np.zeros((VOLUME_SLICES,128, 128)) seg_128_128=np.zeros((VOLUME_SLICES,128, 128)) for j in range(VOLUME_SLICES): asd=cv2.resize(flair[:,:,j+VOLUME_START_AT], (IMG_SIZE, IMG_SIZE)) if np.std(asd)==0: img_non_128[j,:,:,0]=asd else: img_non_128[j,:,:,0]=(asd-np.mean(asd))/(np.std(asd)) asd=cv2.resize(ce[:,:,j+VOLUME_START_AT], (IMG_SIZE, IMG_SIZE)) if np.std(asd)==0: img_non_128[j,:,:,1]=asd else: img_non_128[j,:,:,1]=(asd-np.mean(asd))/(np.std(asd)) asd=cv2.resize(t1[:,:,j+VOLUME_START_AT], (IMG_SIZE, IMG_SIZE)) if np.std(asd)==0: img_non_128[j,:,:,2]=asd else: img_non_128[j,:,:,2]=(asd-np.mean(asd))/(np.std(asd)) asd=cv2.resize(t2[:,:,j+VOLUME_START_AT], (IMG_SIZE, IMG_SIZE)) if np.std(asd)==0: img_non_128[j,:,:,3]=asd else: img_non_128[j,:,:,3]=(asd-np.mean(asd))/(np.std(asd)) mask_e=np.round(cv2.resize(seg[:,:,j+VOLUME_START_AT], (IMG_SIZE, IMG_SIZE))) mask_e[mask_e==4]=1 mask_e[mask_e==3]=1 mask_e[mask_e==2]=1 seg_128_128[j,:,:]=mask_e #img_non_128=np.swapaxes(img_non_128,0,2) #img_non_128=np.swapaxes(img_non_128,1,2)冠狀面(1,2)矢狀面(0,1) #seg_128_128=np.swapaxes(seg_128_128,0,2) #seg_128_128=np.swapaxes(seg_128_128,1,2)冠狀面(1,2)矢狀面(0,1) for k in range(128): floder_name='/content/BraTS2021_Training_Data_v/'+i+'_'+str(k) os.mkdir(floder_name) cl_ass.append(i+'_'+str(k) img_name=floder_name+'/'+i+'_'+str(k)+'_img' seg_name=floder_name+'/'+i+'_'+str(k)+'_seg' np.save(img_name, img_non_128[k,:,:,:]) np.save(seg_name, seg_128_128[k,:,:]) oringin="rm -r /content/BraTS/"+i os.system(oringin) return cl_ass os.mkdir("/content/BraTS2021_Training_Data_v") train_id=test(train_ids) train_id=np.array(train_id) np.save("train_id", train_id) val_id=test(val_ids) val_id=np.array(val_id) np.save("val_id", val_id) test_id=test(test_ids) test_id=np.array(test_id) np.save("test_id", test_id) import random train_id=np.load("/content/train_id.npy") val_id=np.load("/content/val_id.npy") test_id=np.load("/content/test_id.npy") train_id=list(train_id) random.shuffle(train_id) val_id=list(val_id) random.shuffle(val_id) test_id=list(test_id) random.shuffle(test_id) TRAIN_DATASET_PATH = '/content/BraTS2021_Training_Data_v/'#訓練資料集路徑 """**骰子相似係數**""" def dice_test(y_true, y_pred, epsilon=1e-6): intersection = K.sum(y_true * y_pred) return (2. * intersection) / (K.sum(K.square(y_true)) + K.sum(K.square(y_pred)) + epsilon) """**全域骰子損失函數**""" def generalized_dice(y_true, y_pred): """ Generalized Dice Score https://arxiv.org/pdf/1707.03237 """ y_true = K.reshape(y_true,shape=(-1,1)) y_pred = K.reshape(y_pred,shape=(-1,1)) sum_p = K.sum(y_pred, -2) sum_r = K.sum(y_true, -2) sum_pr = K.sum(y_true * y_pred, -2) weights = K.pow(sum_r+K.epsilon() , -1)#+ K.epsilon() generalized_dice = (2 * K.sum(weights * sum_pr)) / (K.sum(weights * (sum_r + sum_p))) return generalized_dice def generalized_dice_loss(y_true, y_pred): return 1-generalized_dice(y_true, y_pred) """# **EffcientUnet_fc2** EffcientUnet搭配(flair,t1ce,t2) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_fc2_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc2_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_fc1** EffcientUnet搭配(flair,t1ce,t1) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_2c1** EffcientUnet搭配(t2,t1ce,t1) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,3]#t2 X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_2c1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_2c1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,3] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_f12** EffcientUnet搭配(flair,t1,t2) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,2]#t1 X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_2c1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_2c1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **UNET**""" K.clear_session() from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import ReLU def build_unet(inputs, ker_init, dropout): conv1 = Conv2D(32, 3, padding = 'same', kernel_initializer = ker_init)(inputs) conv1=BatchNormalization()(conv1) conv1=ReLU()(conv1) conv1 = Conv2D(32, 3, padding = 'same', kernel_initializer = ker_init)(conv1) conv1=BatchNormalization()(conv1) conv1=ReLU()(conv1) pool = MaxPooling2D(pool_size=(2, 2))(conv1) conv = Conv2D(64, 3, padding = 'same', kernel_initializer = ker_init)(pool) conv=BatchNormalization()(conv) conv=ReLU()(conv) conv = Conv2D(64, 3, padding = 'same', kernel_initializer = ker_init)(conv) conv=BatchNormalization()(conv) conv=ReLU()(conv) pool1 = MaxPooling2D(pool_size=(2, 2))(conv) conv2 = Conv2D(128, 3, padding = 'same', kernel_initializer = ker_init)(pool1) conv2=BatchNormalization()(conv2) conv2=ReLU()(conv2) conv2 = Conv2D(128, 3, padding = 'same', kernel_initializer = ker_init)(conv2) conv2=BatchNormalization()(conv2) conv2=ReLU()(conv2) pool2 = MaxPooling2D(pool_size=(2, 2))(conv2) conv3 = Conv2D(256, 3, padding = 'same', kernel_initializer = ker_init)(pool2) conv3=BatchNormalization()(conv3) conv3=ReLU()(conv3) conv3 = Conv2D(256, 3, padding = 'same', kernel_initializer = ker_init)(conv3) conv3=BatchNormalization()(conv3) conv3=ReLU()(conv3) pool4 = MaxPooling2D(pool_size=(2, 2))(conv3) conv5 = Conv2D(512, 3, padding = 'same', kernel_initializer = ker_init)(pool4) conv5=BatchNormalization()(conv5) conv5=ReLU()(conv5) conv5 = Conv2D(512, 3,padding = 'same', kernel_initializer = ker_init)(conv5) conv5=BatchNormalization()(conv5) conv5=ReLU()(conv5) drop5 = Dropout(dropout)(conv5) up7 = Conv2D(256, 2,activation=tf.keras.layers.ReLU(), padding = 'same', kernel_initializer = ker_init)(UpSampling2D(size = (2,2))(drop5)) merge7 = concatenate([conv3,up7], axis = 3) conv7 = Conv2D(256, 3, padding = 'same', kernel_initializer = ker_init)(merge7) conv7=BatchNormalization()(conv7) conv7=ReLU()(conv7) conv7 = Conv2D(256, 3, padding = 'same', kernel_initializer = ker_init)(conv7) conv7=BatchNormalization()(conv7) conv7=ReLU()(conv7) up8 = Conv2D(128, 2,activation=tf.keras.layers.ReLU(), padding = 'same', kernel_initializer = ker_init)(UpSampling2D(size = (2,2))(conv7)) merge8 = concatenate([conv2,up8], axis = 3) conv8 = Conv2D(128, 3, padding = 'same', kernel_initializer = ker_init)(merge8) conv8=BatchNormalization()(conv8) conv8=ReLU()(conv8) conv8 = Conv2D(128, 3, padding = 'same', kernel_initializer = ker_init)(conv8) conv8=BatchNormalization()(conv8) conv8=ReLU()(conv8) up9 = Conv2D(64, 2,activation=tf.keras.layers.ReLU(), padding = 'same', kernel_initializer = ker_init)(UpSampling2D(size = (2,2))(conv8)) merge9 = concatenate([conv,up9], axis = 3) conv9 = Conv2D(64, 3, padding = 'same', kernel_initializer = ker_init)(merge9) conv9=BatchNormalization()(conv9) conv9=ReLU()(conv9) conv9 = Conv2D(64, 3, padding = 'same', kernel_initializer = ker_init)(conv9) conv9=BatchNormalization()(conv9) conv9=ReLU()(conv9) up = Conv2D(32, 2,activation=tf.keras.layers.ReLU(), padding = 'same', kernel_initializer = ker_init)(UpSampling2D(size = (2,2))(conv9)) merge = concatenate([conv1,up], axis = 3) conv = Conv2D(32, 3, padding = 'same', kernel_initializer = ker_init)(merge) conv=BatchNormalization()(conv) conv=ReLU()(conv) conv = Conv2D(32, 3, padding = 'same', kernel_initializer = ker_init)(conv) conv=BatchNormalization()(conv) conv=ReLU()(conv) conv10 = Conv2D(1, (1,1), activation = 'sigmoid')(conv) return Model(inputs = inputs, outputs = conv10) input_layer = Input((IMG_SIZE, IMG_SIZE, 4)) model = build_unet(input_layer, 'he_normal', 0.2) model.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 4, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 X[c,:,:,3] = img[...,3]#t2 return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = model.fit(training_generator, epochs=30, validation_data = valid_generator ) model.save('unet_xyz.h5') model = tf.keras.models.load_model('unet_xyz.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,4)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] X[j,:,:,3] = img[...,3] return model.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") model.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = model.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_fc2** EffcientUnetB4搭配(flair,t1ce,t2) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) skip = Conv2D(num_filters, 1, padding="same")(skip) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) x = Add()([x, skip]) return x def decoder_block_f(inputs, skip, num_filters): x = Concatenate()([inputs, skip]) x = conv_block(x, num_filters) return x def decoder_block_l(inputs, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_fc1** EffcientUnetB4搭配(flair,t1ce,t1) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) skip = Conv2D(num_filters, 1, padding="same")(skip) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) x = Add()([x, skip]) return x def decoder_block_f(inputs, skip, num_filters): x = Concatenate()([inputs, skip]) x = conv_block(x, num_filters) return x def decoder_block_l(inputs, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_2c1** EffcientUnetB4搭配(2,t1ce,t1) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) skip = Conv2D(num_filters, 1, padding="same")(skip) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) x = Add()([x, skip]) return x def decoder_block_f(inputs, skip, num_filters): x = Concatenate()([inputs, skip]) x = conv_block(x, num_filters) return x def decoder_block_l(inputs, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,3]#t2 X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,3] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_f12** EffcientUnetB4搭配(flair,1,t2) """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout import tensorflow as tf print("TF Version: ", tf.__version__) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = Activation("relu")(x) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) skip = Conv2D(num_filters, 1, padding="same")(skip) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) x = Add()([x, skip]) return x def decoder_block_f(inputs, skip, num_filters): x = Concatenate()([inputs, skip]) x = conv_block(x, num_filters) return x def decoder_block_l(inputs, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block(b1, s4, 512) ## 32 d2 = decoder_block(d1, s3, 256) ## 64 d3 = decoder_block(d2, s2, 128) ## 128 d4 = decoder_block(d3, s1, 64) d5 = decoder_block(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,2]#t1 X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_fc2_attention_gate** EffcientUnet搭配(flair,t1ce,t2)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_fc2_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc2_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_fc1_attention_gate** EffcientUnet搭配(flair,t1ce,t1)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_2c1_attention_gate** EffcientUnet搭配(t2,t1ce,t1)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,3]#t2 X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_2c1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_2c1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,3] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **EffcientUnet_f12_attention_gate** EffcientUnet搭配(flair,t1,t2)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetV2S from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetV2S(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5i_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetV2S_UNET") return model input_shape = (128, 128, 3) effienet_1f2 = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_1f2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,2]#t1 X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_1f2.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_1f2.save('effienet_2c1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_2c1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_fc2_attention_gate** EffcientUnetB4搭配(flair,t1ce,t2)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output ## 32 """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_fc1_attention_gate** EffcientUnetB4搭配(flair,t1ce,t1)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output ## 32 """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_2c1_attention_gate** EffcientUnetB4搭配(2,t1ce,t1)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output ## 32 """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,3]#t2 X[c,:,:,1] = img[...,1]#t1ce X[c,:,:,2] = img[...,2]#t1 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,3] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# **Efficientunet b4_f12_attention_gate** EffcientUnetB4搭配(flair,1,t2)並加入attention_gate """ K.clear_session() from tensorflow.keras.layers import Conv2D, BatchNormalization, Activation, MaxPool2D, Conv2DTranspose, Concatenate, Input from tensorflow.keras.models import Model from tensorflow.keras.applications import EfficientNetB4 from tensorflow.keras.layers import Dropout from tensorflow.keras.layers import Add import tensorflow as tf print("TF Version: ", tf.__version__) def attention_gate(input_x, g): theta_x = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(input_x) phi_g = Conv2D(filters=input_x.shape[-1], kernel_size=(1, 1), strides=(1, 1), padding="same")(g) f = Activation("relu")(Add()([theta_x, phi_g])) psi = Conv2D(filters=1, kernel_size=(1, 1), strides=(1, 1), padding="same")(f) attention = Activation("sigmoid")(psi) return Multiply()([input_x, attention]) def conv_block(inputs, num_filters): x = Conv2D(num_filters, 3, padding="same")(inputs) x = BatchNormalization()(x) x = Activation("relu")(x) x = Conv2D(num_filters, 3, padding="same")(x) x = BatchNormalization()(x) x = Activation("relu")(x) return x def decoder_block_with_attention_and_residual(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) # 添加Attention Gate attention_output = attention_gate(x, skip) # 添加Residual Connection x = Add()([x, attention_output]) return x def decoder_block(inputs, skip, num_filters): x = Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(inputs) #skip= Conv2DTranspose(num_filters, (2, 2), strides=2, padding="same")(skip) x = Concatenate()([x, skip]) x = conv_block(x, num_filters) return x def build_effienet_unet(input_shape): """ Input """ inputs = Input(input_shape) """ Pre-trained Encoder """ encoder = EfficientNetB4(include_top=False, weights="imagenet", input_tensor=inputs) s0 = encoder.get_layer("input_1").output s1 = encoder.get_layer("block1b_add").output ## 256 s2 = encoder.get_layer("block2d_add").output ## 128 s3 = encoder.get_layer("block3d_add").output ## 64 s4 = encoder.get_layer("block5f_add").output ## 32 """ Bottleneck """ #b1 = encoder.get_layer("top_activation").output ## 16 b1 = encoder.layers[-1].output ## 32 """ Decoder """ d1 = decoder_block_with_attention_and_residual(b1, s4, 512) ## 32 d2 = decoder_block_with_attention_and_residual(d1, s3, 256) ## 64 d3 = decoder_block_with_attention_and_residual(d2, s2, 128) ## 128 d4 = decoder_block_with_attention_and_residual(d3, s1, 64) d5 = decoder_block_with_attention_and_residual(d4, s0, 32) """ Decoder """ """ Output """ outputs = Conv2D(1, 1, padding="same", activation="sigmoid")(d5) model = Model(inputs, outputs, name="EfficientNetB4_UNET") return model input_shape = (128, 128, 3) effienet_unet = build_effienet_unet(input_shape) #effienet_unet.summary() effienet_unet.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), metrics = ['accuracy', dice_test] ) class DataGenerator(tf.keras.utils.Sequence): 'Generates data for Keras' def __init__(self, list_IDs, dim=(IMG_SIZE,IMG_SIZE), batch_size = 64, n_channels = 3, shuffle=True): 'Initialization' self.dim = dim self.batch_size = batch_size self.list_IDs = list_IDs self.n_channels = n_channels self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.floor(len(self.list_IDs) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size] # Find list of IDs Batch_ids = [self.list_IDs[k] for k in indexes] # Generate data X, y = self.__data_generation(Batch_ids) return X, y def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.list_IDs)) if self.shuffle == True: np.random.shuffle(self.indexes) def __data_generation(self, Batch_ids): 'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels) # Initialization X = np.zeros((self.batch_size, *self.dim, self.n_channels)) y = np.zeros((self.batch_size, 128, 128)) # Generate data for c, i in enumerate(Batch_ids): case_path = os.path.join(TRAIN_DATASET_PATH, i) data_path = os.path.join(case_path, f'{i}_img.npy') img = np.load(data_path) data_path = os.path.join(case_path, f'{i}_seg.npy') seg = np.load(data_path) X[c,:,:,0] = img[...,0]#flair X[c,:,:,1] = img[...,2]#t1 X[c,:,:,2] = img[...,3]#t2 y[c,:,:]=seg return X, y training_generator = DataGenerator(train_id) valid_generator = DataGenerator(val_id) test_generator = DataGenerator(test_id) history = effienet_unet.fit(training_generator, epochs=30, validation_data = valid_generator ) effienet_unet.save('effienet_fc1_zxy.h5') effienet_1c2 = tf.keras.models.load_model('effienet_fc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE,3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] return effienet_1c2.predict(X, verbose=1) def predictByPath_X(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((VOLUME_SLICES, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y import numpy as np def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' p = predictByPath(path,case) origImage,gt=predictByPath_X(path,case) plt.figure(figsize=(18, 50)) f, axarr = plt.subplots(1,3, figsize = (18, 50)) axarr[0].imshow(origImage[0,:,:], cmap="gray") axarr[0].title.set_text('Original image flair') axarr[1].imshow(gt[0,:,:], cmap="Reds", interpolation='none', alpha=0.3) axarr[1].title.set_text('Ground truth') axarr[2].imshow(np.round(p[0,:,]), cmap="OrRd", interpolation='none', alpha=0.3) axarr[2].title.set_text('predicted') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93") effienet_1c2.compile(loss=generalized_dice_loss, optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), metrics = ['accuracy', dice_test] ) # Evaluate the model on the test data using `evaluate` print("Evaluate on test data") results = effienet_1c2.evaluate(test_generator, batch_size=100) print("test loss, test acc:", results) """# ***模型比較***""" unet = tf.keras.models.load_model('/content/drive/MyDrive/unet_3_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) effienet_fc1 = tf.keras.models.load_model('/content/drive/MyDrive/effienet_tc1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) effienet_fc2 = tf.keras.models.load_model('/content/drive/MyDrive/effienet_tc2_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) effienet_2c1 = tf.keras.models.load_model('/content/drive/MyDrive/effienet_2c1_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) effienet_f12 = tf.keras.models.load_model('/content/drive/MyDrive/effienet_f12_zxy.h5',custom_objects={ 'accuracy': dice_test,"dice_test": dice_test}, compile=False) def predictByPath_unet(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 4)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] X[j,:,:,3] = img[...,4] return unet.predict(X, verbose=1) def predictByPath_effienet_fc1(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_fc1.predict(X, verbose=1) def predictByPath_effienet_fc2(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,3] return effienet_fc2.predict(X, verbose=1) def predictByPath_effienet_2c1(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,3] X[j,:,:,1] = img[...,1] X[j,:,:,2] = img[...,2] return effienet_2c1.predict(X, verbose=1) def predictByPath_effienet_f12(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 3)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] X[j,:,:,1] = img[...,2] X[j,:,:,2] = img[...,3] return effienet_f12.predict(X, verbose=1) import numpy as np def predictByPath_seg(case_path,case): files = next(os.walk(case_path))[2] X = np.empty((1, IMG_SIZE, IMG_SIZE, 1)) y = np.empty((1, IMG_SIZE, IMG_SIZE)) vol_path = case_path + case + '_img.npy' img=np.load(vol_path) vol_path = case_path + case + '_seg.npy' seg=nib.load(vol_path) for j in range(1): X[j,:,:,0] = img[...,0] y[j,...] = seg return X,y def showPredictsById(case, start_slice = 60): path = TRAIN_DATASET_PATH + case + '/' gt = nib.load(path + case +'_seg.nii').get_fdata() origImage = nib.load(path + case +'_flair.nii').get_fdata() p = predictByPath_unet(path,case) p1 = predictByPath_effienet_fc1(path,case) p2= predictByPath_effienet_fc2(path,case) p3= predictByPath_effienet_2c1(path,case) p4= predictByPath_effienet_f12(path,case) o,gt_check=predictByPath_seg(path,case) f, axarr = plt.subplots(1,5, figsize = (25, 15)) asd=p asc=gt_check axarr[0].imshow(np.round(asd)+asc, cmap="hot") axarr[0].title.set_text('U-Net predicted') asd=p1 axarr[1].imshow(np.round(asd)+asc, cmap="hot") axarr[1].title.set_text('EfficienUnet_fc1 predicted') asd=p2 axarr[2].imshow(np.round(asd)+asc, cmap="hot") axarr[2].title.set_text('EfficienUnet_fc2 predicted') asd=p3 axarr[3].imshow(np.round(asd)+asc, cmap="hot") axarr[3].title.set_text('EfficienUnet_2c1 predicted') asd=p4 axarr[4].imshow(np.round(asd)+asc, cmap="hot") axarr[4].title.set_text('EfficienUnet_f12 predicted') axarr[0].axis('off') axarr[1].axis('off') axarr[2].axis('off') axarr[3].axis('off') axarr[4].axis('off') plt.show() showPredictsById(case="BraTS19_CBICA_ATN_1_93")