def val(model, dataloader, num_train_val, device, args):
all_dice = []
all_iou = []
all_acc = []
all_sen = []
all_spe = []
model.eval()
with torch.no_grad():
with tqdm(total=num_train_val, desc=f'VAL', unit='img') as pbar:
for batch in dataloader:
image = batch["image"]
label = batch["label"]
assert len(image.size()) == 4
assert len(label.size()) == 3
image = image.to(device, dtype=torch.float32)
label = label.to(device, dtype=torch.long)
outputs = model(image)
main_out = outputs["main_out"]
main_out = torch.exp(main_out).max(dim=1)[1] # 阈值处理
# 逐图片计算指标
for b in range(image.size()[0]):
hist = utils.fast_hist(label[b, :, :], main_out[b, :, :],
args.n_class)
dice, iou, acc, Sensitivity, Specificity = utils.cal_scores(
hist.cpu().numpy())
all_dice.append(list(dice))
all_iou.append(list(iou))
all_acc.append([acc])
all_sen.append(list(Sensitivity))
all_spe.append(list(Specificity))
pbar.update(image.size()[0])
# 验证集指标
mDice = np.array(all_dice).mean()
mIoU = np.array(all_iou).mean()
mAcc = np.array(all_acc).mean()
mSensitivity = np.array(all_sen).mean()
mSpecificity = np.array(all_spe).mean()
dice = list(np.array(all_dice).mean(axis=0))
print(f"dice{dice}")
print(
f'\r [VAL] mDice:{mDice:0.2f}, mIoU:{mIoU:0.2f}, mAcc:{mAcc:0.2f}, mSen:{mSensitivity:0.2f}, mSpec:{mSpecificity:0.2f}'
)
return mDice, mIoU, mAcc, mSensitivity, mSpecificity
def val_3D(model, device, args, num_fold):
data_root = args.val_data_root if hasattr(
args, "val_data_root") else args.data_root
target_root = args.val_target_root if hasattr(
args, "val_target_root") else args.data_root
val_image_root = os.path.join(data_root, f"f{num_fold}")
val_mask_root = os.path.join(target_root, f"f{num_fold}")
volumes = os.listdir(val_image_root)
all_dice = []
all_iou = []
all_acc = []
all_sen = []
all_spe = []
model.eval()
with torch.no_grad():
with tqdm(total=len(volumes), desc=f'VAL', unit='volume') as pbar:
for volume in volumes:
volume_img_path = os.path.join(val_image_root, volume)
volume_mask_path = os.path.join(val_mask_root, volume)
# print(volume_img_path, volume_mask_path)
v_dataset = volume_Dataset(volume_img_path, volume_mask_path,
args.crop_size)
v_dataloader = DataLoader(v_dataset,
batch_size=1,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False)
volume_pred = []
volume_label = []
for batch in v_dataloader:
image = batch["image"]
label = batch["label"]
image = image.to(device, dtype=torch.float32)
label = label.to(device, dtype=torch.long)
outputs = model(image)
main_out = outputs["main_out"]
main_out = torch.exp(main_out).max(dim=1)[1] # 阈值处理
volume_pred.append(main_out)
volume_label.append(label)
volume_pred = torch.cat(volume_pred, dim=0)
volume_label = torch.cat(volume_label, dim=0)
hist = utils.fast_hist(volume_label, volume_pred, args.n_class)
dice, iou, acc, Sensitivity, Specificity = utils.cal_scores(
hist.cpu().numpy(), drop_non=args.drop_non)
all_dice.append(list(dice))
all_iou.append(list(iou))
all_acc.append([acc])
all_sen.append(list(Sensitivity))
all_spe.append(list(Specificity))
pbar.update(1)
volume_pred = sitk.GetImageFromArray(
np.array(volume_pred.cpu().numpy(), np.int8))
sitk.WriteImage(volume_pred,
os.path.join(args.plot_save_dir, volume + '.nii.gz'))
# 验证集指标
if args.drop_non:
mDice = np.mean(
np.array(all_dice).sum(axis=0) /
np.sum(np.array(all_dice) > 0, axis=0))
mIoU = np.mean(
np.array(all_iou).sum(axis=0) /
np.sum(np.array(all_iou) > 0, axis=0))
mAcc = np.array(all_acc).mean()
mSensitivity = np.mean(
np.array(all_sen).sum(axis=0) /
np.sum(np.array(all_sen) > 0, axis=0))
mSpecificity = np.mean(
np.array(all_spe).sum(axis=0) /
np.sum(np.array(all_spe) > 0, axis=0))
else:
mDice = np.array(all_dice).mean()
mIoU = np.array(all_iou).mean()
mAcc = np.array(all_acc).mean()
mSensitivity = np.array(all_sen).mean()
mSpecificity = np.array(all_spe).mean()
dice = list(
np.array(all_dice).sum(axis=0) /
np.sum(np.array(all_dice) > 0, axis=0))
print(f"dice{dice}")
print(
f'\r [VAL] mDice:{mDice:0.2f}, mIoU:{mIoU:0.2f}, mAcc:{mAcc:0.2f}, mSen:{mSensitivity:0.2f}, mSpec:{mSpecificity:0.2f}'
)
return mDice, mIoU, mAcc, mSensitivity, mSpecificity
def test_3D(model, device, args, num_fold=0):
# 导入模型, 选取每一折的最优模型
if os.path.exists(args.val_result_file):
with open(args.val_result_file, "r") as f:
reader = csv.reader(f)
val_result = list(reader)
best_epoch = utils.best_model_in_fold(val_result, num_fold)
else:
best_epoch = args.num_epochs
# 导入模型
model_list = os.listdir(args.checkpoint_dir[num_fold])
model_dir = [x for x in model_list if str(best_epoch) in x][0]
model_dir = os.path.join(args.checkpoint_dir[num_fold], model_dir)
if not os.path.exists(model_dir):
model_dir = os.path.join(args.checkpoint_dir[num_fold],
f'CP_epoch{best_epoch}.pth')
model.load_state_dict(torch.load(model_dir, map_location=device))
print(f'\rtest model loaded: [fold:{num_fold}] [best_epoch:{best_epoch}]')
data_root = args.val_data_root if hasattr(
args, "val_data_root") else args.data_root
target_root = args.val_target_root if hasattr(
args, "val_target_root") else args.data_root
val_image_root = os.path.join(data_root, f"f{num_fold}")
val_mask_root = os.path.join(target_root, f"f{num_fold}")
volumes = sorted(os.listdir(val_image_root))
all_dice = []
all_iou = []
all_acc = []
all_sen = []
all_spe = []
model.eval()
with torch.no_grad():
with tqdm(total=len(volumes), desc=f'VAL', unit='img') as pbar:
for volume in volumes:
volume_img_path = os.path.join(val_image_root, volume)
volume_mask_path = os.path.join(val_mask_root, volume)
v_dataset = volume_Dataset(volume_img_path, volume_mask_path,
args.crop_size)
v_dataloader = DataLoader(v_dataset,
batch_size=1,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True,
drop_last=False)
volume_pred = []
volume_label = []
for batch in v_dataloader:
image = batch["image"]
label = batch["label"]
image = image.to(device, dtype=torch.float32)
label = label.to(device, dtype=torch.long)
outputs = model(image)
main_out = outputs["main_out"]
main_out = torch.exp(main_out).max(dim=1)[1] # 阈值处理
volume_pred.append(main_out)
volume_label.append(label)
volume_pred = torch.cat(volume_pred, dim=0)
volume_label = torch.cat(volume_label, dim=0)
hist = utils.fast_hist(volume_label, volume_pred, args.n_class)
dice, iou, acc, Sensitivity, Specificity = utils.cal_scores(
hist.cpu().numpy(), drop_non=args.drop_non)
all_dice.append(list(dice))
all_iou.append(list(iou))
all_acc.append([acc])
all_sen.append(list(Sensitivity))
all_spe.append(list(Specificity))
# 写入每个测试数据的指标
test_result = [volume, dice.mean()]+list(dice)+[iou.mean()]+list(iou)+[acc] + \
[Sensitivity.mean()]+list(Sensitivity)+[Specificity.mean()]+list(Specificity)
with open(args.test_result_file, "a") as f:
w = csv.writer(f)
w.writerow(test_result)
if args.plot:
volume_pred = sitk.GetImageFromArray(
np.array(volume_pred.cpu().numpy(), np.int8))
sitk.WriteImage(
volume_pred,
os.path.join(args.plot_save_dir,
volume + f'_{dice.mean():.2f}.nii.gz'))
pbar.update(image.size()[0])
print(f"\r---------Fold {num_fold} Test Result---------")
if args.drop_non:
print(
f'mDice: {np.mean(np.array(all_dice).sum(axis=0)/np.sum(np.array(all_dice) > 0, axis=0))}'
)
print(
f'mIoU: {np.mean(np.array(all_iou).sum(axis=0)/np.sum(np.array(all_iou) > 0, axis=0))}'
)
print(f'mAcc: {np.array(all_acc).mean()}')
print(
f'mSens: {np.mean(np.array(all_sen).sum(axis=0)/np.sum(np.array(all_sen) > 0, axis=0))}'
)
print(
f'mSpec: {np.mean(np.array(all_spe).sum(axis=0)/np.sum(np.array(all_spe) > 0, axis=0))}'
)
else:
print(f'mDice: {np.array(all_dice).mean()}')
print(f'mIoU: {np.array(all_iou).mean()}')
print(f'mAcc: {np.array(all_acc).mean()}')
print(f'mSens: {np.array(all_sen).mean()}')
print(f'mSpec: {np.array(all_spe).mean()}')
if num_fold == 0:
utils.save_print_score(all_dice,
all_iou,
all_acc,
all_sen,
all_spe,
args.test_result_file,
args.label_names,
drop_non=args.drop_non)
return
return all_dice, all_iou, all_acc, all_sen, all_spe
def test(model, device, args, num_fold=0):
# 导入模型, 选取每一折的最优模型
if os.path.exists(args.val_result_file):
with open(args.val_result_file, "r") as f:
reader = csv.reader(f)
val_result = list(reader)
best_epoch = utils.best_model_in_fold(val_result, num_fold)
else:
best_epoch = args.num_epochs
# 导入模型
model_list = os.listdir(args.checkpoint_dir[num_fold])
model_dir = [x for x in model_list if str(best_epoch) in x][0]
model_dir = os.path.join(args.checkpoint_dir[num_fold], model_dir)
if not os.path.exists(model_dir):
model_dir = os.path.join(args.checkpoint_dir[num_fold],
f'CP_epoch{best_epoch}.pth')
model.load_state_dict(torch.load(model_dir, map_location=device))
print(f'\rtest model loaded: [fold:{num_fold}] [best_epoch:{best_epoch}]')
dataset_test = myDataset(args.data_root,
args.target_root,
args.crop_size,
"test",
k_fold=args.k_fold,
imagefile_csv=args.dataset_file_list,
num_fold=num_fold)
dataloader = DataLoader(dataset_test,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True)
all_dice = []
all_iou = []
all_acc = []
all_sen = []
all_spe = []
model.eval()
with torch.no_grad():
with tqdm(total=len(dataset_test),
desc=f'TEST fold {num_fold}/{args.k_fold}',
unit='img') as pbar:
for batch in dataloader:
image = batch["image"]
label = batch["label"]
file = batch["file"]
assert len(image.size()) == 4
assert len(label.size()) == 3
image = image.to(device, dtype=torch.float32)
label = label.to(device, dtype=torch.long)
outputs = model(image)
pred = outputs["main_out"]
pred = torch.exp(pred).max(dim=1)[1] # 阈值处理
for b in range(image.size()[0]):
hist = utils.fast_hist(label[b, :, :], pred[b, :, :],
args.n_class)
dice, iou, acc, Sensitivity, Specificity = utils.cal_scores(
hist.cpu().numpy(), smooth=0.01)
# 写入每个测试数据的指标
test_result = [file[b], dice.mean()]+list(dice)+[iou.mean()]+list(iou)+[acc] + \
[Sensitivity.mean()]+list(Sensitivity)+[Specificity.mean()]+list(Specificity)
with open(args.test_result_file, "a") as f:
w = csv.writer(f)
w.writerow(test_result)
all_dice.append(list(dice))
all_iou.append(list(iou))
all_acc.append([acc])
all_sen.append(list(Sensitivity))
all_spe.append(list(Specificity))
if args.plot:
file_name, _ = os.path.splitext(file[b])
# save_image(pred[b,:,:].cpu().float().unsqueeze(0), os.path.join(args.plot_save_dir, file_name + f"_pred_{dice.mean():.2f}.png"), normalize=True)
# save_image(image[b,:,:].cpu(), os.path.join(args.plot_save_dir, file[b]))
# save_image(label[b,:,:].cpu().float().unsqueeze(0), os.path.join(args.plot_save_dir, file_name + f"_label.png"), normalize=True)
if "A4" in outputs.keys():
for i in range(0, 25, 5):
proj_map = F.interpolate(
outputs["A1"][b, ...].unsqueeze(0),
size=image.size()[-2:],
mode="bilinear",
align_corners=True).squeeze(0)
save_image(proj_map[i, :, :],
os.path.join(
args.plot_save_dir,
file_name + f"_A1_{i}.png"),
normalize=True)
for i in range(0, 25, 5):
proj_map = F.interpolate(
outputs["A2"][b, ...].unsqueeze(0),
size=image.size()[-2:],
mode="bilinear",
align_corners=True).squeeze(0)
save_image(proj_map[i, :, :],
os.path.join(
args.plot_save_dir,
file_name + f"_A2_{i}.png"),
normalize=True)
for i in range(0, 25, 5):
proj_map = F.interpolate(
outputs["A3"][b, ...].unsqueeze(0),
size=image.size()[-2:],
mode="bilinear",
align_corners=True).squeeze(0)
save_image(proj_map[i, :, :],
os.path.join(
args.plot_save_dir,
file_name + f"_A3_{i}.png"),
normalize=True)
for i in range(0, 25, 5):
proj_map = F.interpolate(
outputs["A4"][b, ...].unsqueeze(0),
size=image.size()[-2:],
mode="bilinear",
align_corners=True).squeeze(0)
save_image(proj_map[i, :, :],
os.path.join(
args.plot_save_dir,
file_name + f"_A4_{i}.png"),
normalize=True)
# if "x_proj_1" in outputs.keys():
# for i in range(7):
# proj_map = F.interpolate(outputs["x_proj_1"][b, ...].unsqueeze(0), size=image.size()[-2:],
# mode="bilinear", align_corners=True).squeeze(0)
# save_image(proj_map[i, :, :], os.path.join(args.plot_save_dir, file_name + f"_A1_{i}.png"),
# normalize=True)
# for i in range(7):
# proj_map = F.interpolate(outputs["x_proj_2"][b, ...].unsqueeze(0), size=image.size()[-2:],
# mode="bilinear", align_corners=True).squeeze(0)
# save_image(proj_map[i, :, :], os.path.join(args.plot_save_dir, file_name + f"_A2_{i}.png"),
# normalize=True)
save_image(image[b, :, :].cpu(),
os.path.join(args.plot_save_dir, file[b]))
pred_image = pred[b, :, :].unsqueeze(0)
true_mask = label[b, :, :].unsqueeze(0)
result_image = torch.cat(
(pred_image, true_mask,
torch.zeros_like(pred_image)),
dim=0).permute(1, 2, 0).cpu().numpy()
result_image = Image.fromarray(
np.uint8(result_image) * 255)
result_image.save(
args.plot_save_dir +
f"/{file_name}({dice.mean():.2f}).png")
pbar.update(image.size()[0])
print(f"\r---------Fold {num_fold} Test Result---------")
print(f'mDice: {np.array(all_dice).mean()}')
print(f'mIoU: {np.array(all_iou).mean()}')
print(f'mAcc: {np.array(all_acc).mean()}')
print(f'mSens: {np.array(all_sen).mean()}')
print(f'mSpec: {np.array(all_spe).mean()}')
if num_fold == 0:
utils.save_print_score(all_dice, all_iou, all_acc, all_sen, all_spe,
args.test_result_file, args.label_names)
return
return all_dice, all_iou, all_acc, all_sen, all_spe
def test(model, device, args, num_fold=0):
# 导入模型, 选取每一折的最优模型
if os.path.exists(args.val_result_file):
with open(args.val_result_file, "r") as f:
reader = csv.reader(f)
val_result = list(reader)
best_epoch = utils.best_model_in_fold(val_result, num_fold)
else:
best_epoch = args.num_epochs
# 导入模型
model_dir = os.path.join(args.checkpoint_dir[num_fold], f'CP_epoch{best_epoch}.pth')
model.load_state_dict(torch.load(model_dir, map_location=device))
print(f'\rtest model loaded: [fold:{num_fold}] [best_epoch:{best_epoch}]')
dataset_test = myDataset(args.data_root, args.target_root, args.crop_size, "test",
k_fold=args.k_fold, imagefile_csv=args.dataset_file_list, num_fold=num_fold)
dataloader = DataLoader(dataset_test, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True)
all_dice = []
all_iou = []
all_acc = []
all_sen = []
all_spe = []
model.eval()
with torch.no_grad():
with tqdm(total=len(dataset_test), desc=f'TEST fold {num_fold}/{args.k_fold}', unit='img') as pbar:
for batch in dataloader:
image = batch["image"]
label = batch["label"]
file = batch["file"]
assert len(image.size()) == 4
assert len(label.size()) == 3
image = image.to(device, dtype=torch.float32)
label = label.to(device, dtype=torch.long)
outputs = model(image)
pred = outputs["main_out"]
pred = torch.exp(pred).max(dim=1)[1] # 阈值处理
for b in range(image.size()[0]):
hist = utils.fast_hist(label[b,:,:], pred[b,:,:], args.n_class)
dice, iou, acc, Sensitivity, Specificity = utils.cal_scores(hist.cpu().numpy(), smooth=0.01)
# 写入每个测试数据的指标
test_result = [file[b], dice.mean()]+list(dice)+[iou.mean()]+list(iou)+[acc] + \
[Sensitivity.mean()]+list(Sensitivity)+[Specificity.mean()]+list(Specificity)
with open(args.test_result_file, "a") as f:
w = csv.writer(f)
w.writerow(test_result)
all_dice.append(list(dice))
all_iou.append(list(iou))
all_acc.append([acc])
all_sen.append(list(Sensitivity))
all_spe.append(list(Specificity))
if args.plot:
file_name, _ = os.path.splitext(file[b])
save_image(pred[b,:,:].cpu().float().unsqueeze(0), os.path.join(args.plot_save_dir, file_name + f"_pred_{dice.mean():.2f}.png"), normalize=True)
save_image(image[b,:,:].cpu(), os.path.join(args.plot_save_dir, file[b]))
save_image(label[b,:,:].cpu().float().unsqueeze(0), os.path.join(args.plot_save_dir, file_name + f"_label.png"), normalize=True)
pbar.update(image.size()[0])
print(f"\r---------Fold {num_fold} Test Result---------")
print(f'mDice: {np.array(all_dice).mean()}')
print(f'mIoU: {np.array(all_iou).mean()}')
print(f'mAcc: {np.array(all_acc).mean()}')
print(f'mSens: {np.array(all_sen).mean()}')
print(f'mSpec: {np.array(all_spe).mean()}')
if num_fold == 0:
utils.save_print_score(all_dice, all_iou, all_acc, all_sen, all_spe, args.test_result_file, args.label_names)
return
return all_dice, all_iou, all_acc, all_sen, all_spe