BraTS2021脑肿瘤分割实战

Brain Tumor Segmentation (BraTS) Challenge 2021 Homepage

github项目地址 brats-unet: UNet for brain tumor segmentation

BraTS是MICCAI所有比赛中历史最悠久的，到2021年已经连续举办了10年，参赛人数众多，是学习医学图像分割最前沿的平台之一。

1.数据准备

简介：

比赛方提供多机构、多参数多模态核磁共振成像（mpMRI）数据集，包括训练集（1251例）和验证集（219例）以及测试集（530例），一共2000例患者的mpMRI扫描结果。其中训练集包含图像和分割标签，验证集和测试集没有分割标签，验证集被用于公共排行榜，测试集不公开，用作参赛者的最终排名评测。

四种模态数据：flair, t1ce, t1, t2，每个模态的数据大小都为 240 x 240 x 155，且共享分割标签。

分割标签：[0, 1, 2, 4]

• label0：背景（background）
• label1：坏疽（NT, necrotic tumor core）
• label2：浮肿区域（ED,peritumoral edema）
• label4：增强肿瘤区域（ET,enhancing tumor）

​ 本次比赛包括两个任务：

• Task1：mpMRI扫描中分割内在异质性脑胶质母细胞瘤区域
• Task2：预测术前基线扫描中的MGMT启动子甲基化状态

本文从数据处理、评价指标、损失函数、模型训练四个方面介绍Task1的整体实现过程

数据集下载地址：

1.官网：BraTS 2021 Challenge 需要注册和申请（包括训练集和验证集）

2.Kaggle：BRaTS 2021 Task 1 Dataset建议在kaggle上下载，数据集与官网一致（不包括验证集）

数据准备：

下载数据集，解压后如下图所示：

每个病例包含四种模态的MRI图像和分割标签，结构如下：

BraTS2021_00000├── BraTS2021_00000_flair.nii.gz├── BraTS2021_00000_seg.nii.gz├── BraTS2021_00000_t1ce.nii.gz├── BraTS2021_00000_t1.nii.gz└── BraTS2021_00000_t2.nii.gz

建议使用3D Slicer查看图像和标签，直观的了解一下自己要用的数据集。

2.数据预处理

每个病例的四种MRI图像大小为 240 x 240 x 155，且共享标签。

鉴于此，我将四种模态的图像合并为一个4D图像（C x H x W x D , C=4），并且和分割标签一起保存为一个.h5文件，方便后续处理。

import h5pyimport osimport numpy as npimport SimpleITK as sitkfrom tqdm import tqdm# 四种模态的mri图像modalities = ('flair', 't1ce', 't1', 't2')# traintrain_set = {'root': '/data/omnisky/postgraduate/Yb/data_set/BraTS2021/data',# 四个模态数据所在地址'out': '/data/omnisky/postgraduate/Yb/data_set/BraTS2021/dataset/',# 预处理输出地址'flist': 'train.txt',# 训练集名单（有标签）}

• 将图像保存为32位浮点数（np.float32），标签保存为整数（np.uint8），写入.h5文件
• 对每张图像的灰度进行标准化，但保持背景区域为0

• 上图是预处理后的图像，背景区域为0

def process_h5(path, out_path):""" Save the data with dtype=float32.z-score is used but keep the background with zero! """# SimpleITK读取图像默认是是 DxHxW，这里转为 HxWxDlabel = sitk.GetArrayFromImage(sitk.ReadImage(path + 'seg.nii.gz')).transpose(1,2,0)print(label.shape)# 堆叠四种模态的图像，4 x (H,W,D) -> (4,H,W,D) images = np.stack([sitk.GetArrayFromImage(sitk.ReadImage(path + modal + '.nii.gz')).transpose(1,2,0) for modal in modalities], 0)# [240,240,155]# 数据类型转换label = label.astype(np.uint8)images = images.astype(np.float32)case_name = path.split('/')[-1]# case_name = os.path.split(path)[-1]# windows路径与linux不同path = os.path.join(out_path,case_name)output = path + 'mri_norm2.h5'# 对第一个通道求和，如果四个模态都为0，则标记为背景(False)mask = images.sum(0) > 0for k in range(4):x = images[k,...]#y = x[mask]# 对背景外的区域进行归一化x[mask] -= y.mean()x[mask] /= y.std()images[k,...] = xprint(case_name,images.shape,label.shape)f = h5py.File(output, 'w')f.create_dataset('image', data=images, compression="gzip")f.create_dataset('label', data=label, compression="gzip")f.close()def doit(dset):root, out_path = dset['root'], dset['out']file_list = os.path.join(root, dset['flist'])subjects = open(file_list).read().splitlines()names = ['BraTS2021_' + sub for sub in subjects]paths = [os.path.join(root, name, name + '_') for name in names]for path in tqdm(paths):process_h5(path, out_path)# breakprint('Finished')if __name__ == '__main__':doit(train_set)

数据保存在 mri_norm2.h5 文件中，每个 mri_norm2.h5 相当于一个字典，字典的键为 image 和 label ，值为对应的数组。

处理后的数据，可以用下面的几行代码测试一下，记得修改为你自己的路径

import h5pyimport numpy as npp = '/***/data_set/BraTS2021/all/BraTS2021_00000_mri_norm2.h5'h5f = h5py.File(p, 'r')image = h5f['image'][:]label = h5f['label'][:]print('image shape:',image.shape,'\t','label shape',label.shape)print('label set:',np.unique(label))# image shape: (4, 240, 240, 155)label shape (240, 240, 155)# label set: [0 1 2 4]

将数据集按照 8:1:1随机划分为训练集、验证集和测试集，将划分后的数据名保存为.txt文件

import osfrom sklearn.model_selection import train_test_split# 预处理输出地址data_path = "/***/data_set/BraTS2021/dataset"train_and_test_ids = os.listdir(data_path)train_ids, val_test_ids = train_test_split(train_and_test_ids, test_size=0.2,random_state=21)val_ids, test_ids = train_test_split(val_test_ids, test_size=0.5,random_state=21)print("Using {} images for training, {} images for validation, {} images for testing.".format(len(train_ids),len(val_ids),len(test_ids)))with open('/***/data_set/BraTS2021/train.txt','w') as f:f.write('\n'.join(train_ids))with open('/***/data_set/BraTS2021/valid.txt','w') as f:f.write('\n'.join(val_ids))with open('/***/data_set/BraTS2021/test.txt','w') as f:f.write('\n'.join(test_ids))

划分结果：

Using 1000 images for training, 125 images for validation, 126 images for testing.......BraTS2021_00002_mri_norm2.h5BraTS2021_00003_mri_norm2.h5BraTS2021_00014_mri_norm2.h5......

3.数据增强

下面是我写的Dataset类以及一些数据增强方法

整体架构

import osimport torchfrom torch.utils.data import Datasetimport randomimport numpy as npfrom torchvision.transforms import transformsimport h5pyclass BraTS(Dataset):def __init__(self,data_path, file_path,transform=None):with open(file_path, 'r') as f:self.paths = [os.path.join(data_path, x.strip()) for x in f.readlines()]self.transform = transformdef __getitem__(self, item):h5f = h5py.File(self.paths[item], 'r')image = h5f['image'][:]label = h5f['label'][:]#[0,1,2,4] -> [0,1,2,3]label[label == 4] = 3# print(image.shape)sample = {'image': image, 'label': label}if self.transform:sample = self.transform(sample)return sample['image'], sample['label']def __len__(self):return len(self.paths)def collate(self, batch):return [torch.cat(v) for v in zip(*batch)]if __name__ == '__main__':from torchvision import transformsdata_path = "/***/data_set/BraTS2021/dataset"test_txt = "/***/data_set/BraTS2021/test.txt"test_set = BraTS(data_path,test_txt,transform=transforms.Compose([RandomRotFlip(),RandomCrop((160,160,128)),GaussianNoise(p=0.1),ToTensor()]))d1 = test_set[0]image,label = d1print(image.shape)print(label.shape)print(np.unique(label))

具体的数据增强方法我列在了下面，包括裁剪、旋转、翻转、高斯噪声、对比度变换和亮度增强的源码，部分代码借鉴了nnUNet的数据增强方法。

随机裁剪

原始图像尺寸为 240 x 240 x 155，但图像周围是有很多黑边的，我将图像裁剪为 160 x 160 x 128

class RandomCrop(object):"""Crop randomly the image in a sampleArgs:output_size (int): Desired output size"""def __init__(self, output_size):self.output_size = output_sizedef __call__(self, sample):image, label = sample['image'], sample['label'](c, w, h, d) = image.shapew1 = np.random.randint(0, w - self.output_size[0])h1 = np.random.randint(0, h - self.output_size[1])d1 = np.random.randint(0, d - self.output_size[2])label = label[w1:w1 + self.output_size[0], h1:h1 + self.output_size[1], d1:d1 + self.output_size[2]]image = image[:,w1:w1 + self.output_size[0], h1:h1 + self.output_size[1], d1:d1 + self.output_size[2]]return {'image': image, 'label': label}

中心裁剪

class CenterCrop(object):def __init__(self, output_size):self.output_size = output_sizedef __call__(self, sample):image, label = sample['image'], sample['label'](c,w, h, d) = image.shapew1 = int(round((w - self.output_size[0]) / 2.))h1 = int(round((h - self.output_size[1]) / 2.))d1 = int(round((d - self.output_size[2]) / 2.))label = label[w1:w1 + self.output_size[0], h1:h1 + self.output_size[1], d1:d1 + self.output_size[2]]image = image[:,w1:w1 + self.output_size[0], h1:h1 + self.output_size[1], d1:d1 + self.output_size[2]]return {'image': image, 'label': label}

随机翻转

旋转可能会导致图像重采样，因为数据集比较充分，我只在{90，180，270}度做一个简单旋转，不涉及重采样。

class RandomRotFlip(object):"""Crop randomly flip the dataset in a sampleArgs:output_size (int): Desired output size"""def __call__(self, sample):image, label = sample['image'], sample['label']k = np.random.randint(0, 4)image = np.stack([np.rot90(x,k) for x in image],axis=0)label = np.rot90(label, k)axis = np.random.randint(1, 4)image = np.flip(image, axis=axis).copy()label = np.flip(label, axis=axis-1).copy()return {'image': image, 'label': label}

高斯噪声

def augment_gaussian_noise(data_sample, noise_variance=(0, 0.1)):if noise_variance[0] == noise_variance[1]:variance = noise_variance[0]else:variance = random.uniform(noise_variance[0], noise_variance[1])data_sample = data_sample + np.random.normal(0.0, variance, size=data_sample.shape)return data_sampleclass GaussianNoise(object):def __init__(self, noise_variance=(0, 0.1), p=0.5):self.prob = pself.noise_variance = noise_variancedef __call__(self, sample):image = sample['image']label = sample['label']if np.random.uniform() < self.prob:image = augment_gaussian_noise(image, self.noise_variance)return {'image': image, 'label': label}

对比度变换

• contrast_range：对比度增强的范围
• preserve_range：是否保留数据的取值范围
• per_channel：是否对每个通道的图像分别进行对比度增强

def augment_contrast(data_sample, contrast_range=(0.75, 1.25), preserve_range=True, per_channel=True):if not per_channel:mn = data_sample.mean()if preserve_range:minm = data_sample.min()maxm = data_sample.max()if np.random.random() < 0.5 and contrast_range[0] < 1:factor = np.random.uniform(contrast_range[0], 1)else:factor = np.random.uniform(max(contrast_range[0], 1), contrast_range[1])data_sample = (data_sample - mn) * factor + mnif preserve_range:data_sample[data_sample < minm] = minmdata_sample[data_sample > maxm] = maxmelse:for c in range(data_sample.shape[0]):mn = data_sample[c].mean()if preserve_range:minm = data_sample[c].min()maxm = data_sample[c].max()if np.random.random() < 0.5 and contrast_range[0] < 1:factor = np.random.uniform(contrast_range[0], 1)else:factor = np.random.uniform(max(contrast_range[0], 1), contrast_range[1])data_sample[c] = (data_sample[c] - mn) * factor + mnif preserve_range:data_sample[c][data_sample[c] < minm] = minmdata_sample[c][data_sample[c] > maxm] = maxmreturn data_sampleclass ContrastAugmentationTransform(object):def __init__(self, contrast_range=(0.75, 1.25), preserve_range=True, per_channel=True,p_per_sample=1.):self.p_per_sample = p_per_sampleself.contrast_range = contrast_rangeself.preserve_range = preserve_rangeself.per_channel = per_channeldef __call__(self, sample):image = sample['image']label = sample['label']for b in range(len(image)):if np.random.uniform() < self.p_per_sample:image[b] = augment_contrast(image[b], contrast_range=self.contrast_range,preserve_range=self.preserve_range, per_channel=self.per_channel)return {'image': image, 'label': label}

亮度变换

附加亮度从具有μ和σ的高斯分布中采样

def augment_brightness_additive(data_sample, mu:float, sigma:float , per_channel:bool=True, p_per_channel:float=1.):if not per_channel:rnd_nb = np.random.normal(mu, sigma)for c in range(data_sample.shape[0]):if np.random.uniform() <= p_per_channel:data_sample[c] += rnd_nbelse:for c in range(data_sample.shape[0]):if np.random.uniform() <= p_per_channel:rnd_nb = np.random.normal(mu, sigma)data_sample[c] += rnd_nbreturn data_sampleclass BrightnessTransform(object):def __init__(self, mu, sigma, per_channel=True, p_per_sample=1., p_per_channel=1.):self.p_per_sample = p_per_sampleself.mu = muself.sigma = sigmaself.per_channel = per_channelself.p_per_channel = p_per_channeldef __call__(self, sample):data, label = sample['image'], sample['label']for b in range(data.shape[0]):if np.random.uniform() < self.p_per_sample:data[b] = augment_brightness_additive(data[b], self.mu, self.sigma, self.per_channel,p_per_channel=self.p_per_channel)return {'image': data, 'label': label}

数据类型转换

将Numpy数组转为Tensor

class ToTensor(object):"""Convert ndarrays in sample to Tensors."""def __call__(self, sample):image = sample['image']label = sample['label']image = torch.from_numpy(image).float()label = torch.from_numpy(label).long()return {'image': image, 'label': label}

相比其他医学影像数据集，BraTS2021是非常高质量的，对数据增强方法并不是很敏感。

4.评价损失

损失函数：

combination of dice and crossentropy loss

dice loss

• μ是网络的softmax输出
• v是分割标签的one-hot编码

其实就是将计算dice时的torch.argmax替换为了torch.softmax

import torch.nn.functional as Fimport torch.nn as nnimport torchfrom einops import rearrangeclass Loss(nn.Module):def __init__(self, n_classes, weight=None, alpha=0.5):"dice_loss_plus_cetr_weighted"super(Loss, self).__init__()self.n_classes = n_classesself.weight = weight.cuda()# self.weight = weightself.alpha = alphadef forward(self, input, target):smooth = 0.01# 防止分母为0input1 = F.softmax(input, dim=1)target1 = F.one_hot(target,self.n_classes)input1 = rearrange(input1,'b n h w s -> b n (h w s)')target1 = rearrange(target1,'b h w s n -> b n (h w s)')input1 = input1[:, 1:, :]target1 = target1[:, 1:, :].float()# 以batch为单位计算loss和dice_loss，据说训练更稳定，和上面的公式有出入# 注意，这里的dice不是真正的dice，叫做soft_dice更贴切inter = torch.sum(input1 * target1)union = torch.sum(input1) + torch.sum(target1) + smoothdice = 2.0 * inter / unionloss = F.cross_entropy(input,target, weight=self.weight)total_loss = (1 - self.alpha) * loss + (1 - dice) * self.alphareturn total_lossif __name__ == '__main__':torch.manual_seed(3)device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')losser = Loss(n_classes=4, weight=torch.tensor([0.2, 0.3, 0.25, 0.25])).to(device)x = torch.randn((2, 4, 16, 16, 16)).to(device)y = torch.randint(0, 4, (2, 16, 16, 16)).to(device)print(losser(x, y))

评价指标：

dice计算方法：
2 ( A ∩ B )A+B 2{(A \cap B)}\over{A + B} A+B2(A∩B)​

def Dice(output, target, eps=1e-3):inter = torch.sum(output * target,dim=(1,2,3)) + epsunion = torch.sum(output,dim=(1,2,3)) + torch.sum(target,dim=(1,2,3)) + eps * 2x = 2 * inter / uniondice = torch.mean(x)return dice

• output: (b, num_class, d, h, w) target: (b, d, h, w)
• dice1(ET):label4
• dice2(TC):label1 + label4
• dice3(WT): label1 + label2 + label4
• 注意，这里的label4已经被替换为3

def cal_dice(output, target):output = torch.argmax(output,dim=1)dice1 = Dice((output == 3).float(), (target == 3).float())dice2 = Dice(((output == 1) | (output == 3)).float(), ((target == 1) | (target == 3)).float())dice3 = Dice((output != 0).float(), (target != 0).float())return dice1, dice2, dice3

5.模型训练

以UNet为例，我把完整代码放在了下面

module:

import torchimport torch.nn as nnclass InConv(nn.Module):def __init__(self, in_ch, out_ch):super(InConv, self).__init__()self.conv = DoubleConv(in_ch, out_ch)def forward(self, x):x = self.conv(x)return xclass Down(nn.Module):def __init__(self, in_ch, out_ch):super(Down, self).__init__()self.mpconv = nn.Sequential(nn.MaxPool3d(2, 2),DoubleConv(in_ch, out_ch))def forward(self, x):x = self.mpconv(x)return xclass OutConv(nn.Module):def __init__(self, in_ch, out_ch):super(OutConv, self).__init__()self.conv = nn.Conv3d(in_ch, out_ch, 1)# self.sigmoid = nn.Sigmoid()def forward(self, x):x = self.conv(x)# x = self.sigmoid(x)return xclass DoubleConv(nn.Module):def __init__(self, in_ch, out_ch):super(DoubleConv, self).__init__()self.conv = nn.Sequential(nn.Conv3d(in_ch, out_ch, kernel_size=3, stride=1, padding=1),nn.BatchNorm3d(out_ch),nn.ReLU(inplace=True),nn.Conv3d(out_ch, out_ch, kernel_size=3, stride=1, padding=1),nn.BatchNorm3d(out_ch),nn.ReLU(inplace=True))def forward(self, x):x = self.conv(x)return xclass Up(nn.Module):def __init__(self, in_ch, skip_ch,out_ch):super(Up, self).__init__()self.up = nn.ConvTranspose3d(in_ch, in_ch, kernel_size=2, stride=2)self.conv = DoubleConv(in_ch+skip_ch, out_ch)def forward(self, x1, x2):x1 = self.up(x1)x = torch.cat([x2, x1], dim=1)x = self.conv(x)return x

model:

class UNet(nn.Module):def __init__(self, in_channels, num_classes):super(UNet, self).__init__()features = [32,64,128,256]self.inc = InConv(in_channels, features[0])self.down1 = Down(features[0], features[1])self.down2 = Down(features[1], features[2])self.down3 = Down(features[2], features[3])self.down4 = Down(features[3], features[3])self.up1 = Up(features[3], features[3], features[2])self.up2 = Up(features[2], features[2], features[1])self.up3 = Up(features[1], features[1], features[0])self.up4 = Up(features[0], features[0], features[0])self.outc = OutConv(features[0], num_classes)def forward(self, x):x1 = self.inc(x)x2 = self.down1(x1)x3 = self.down2(x2)x4 = self.down3(x3)x5 = self.down4(x4)x = self.up1(x5, x4)x = self.up2(x, x3)x = self.up3(x, x2)x = self.up4(x, x1)x = self.outc(x)return xif __name__ == '__main__':x = torch.randn(1, 4, 160, 160, 128)net = UNet(in_channels=4, num_classes=4)y = net(x)print("params: ", sum(p.numel() for p in net.parameters()))print(y.shape)

Train:

下面是我写的训练函数，具体细节见代码注释

• 优化器：optim.SGD(model.parameters(),momentum=0.9, lr=0, weight_decay=5e-4)
• 学习率余弦衰减：最大学习率0.004，最小学习率0.002，预热10个epoch
• 优化策略可参考我的另一篇博客nnUnet代码解读–优化策略

import osimport argparsefrom torch.utils.data import DataLoaderimport torchimport torch.optim as optimfrom tqdm import tqdmfrom BraTS import *from networks.Unet import UNetfrom utils import Loss,cal_dice,cosine_schedulerdef train_loop(model,optimizer,scheduler,criterion,train_loader,device,epoch):model.train()running_loss = 0dice1_train = 0dice2_train = 0dice3_train = 0pbar = tqdm(train_loader)for it,(images,masks) in enumerate(pbar):# update learning rate according to the scheduleit = len(train_loader) * epoch + itparam_group = optimizer.param_groups[0]param_group['lr'] = scheduler[it]# print(scheduler[it])# [b,4,128,128,128] , [b,128,128,128]images, masks = images.to(device),masks.to(device)# [b,4,128,128,128], 4分割outputs = model(images)# outputs = torch.softmax(outputs,dim=1)loss = criterion(outputs, masks)dice1, dice2, dice3 = cal_dice(outputs,masks)pbar.desc = "loss: {:.3f} ".format(loss.item())running_loss += loss.item()dice1_train += dice1.item()dice2_train += dice2.item()dice3_train += dice3.item()optimizer.zero_grad()loss.backward()optimizer.step()loss = running_loss / len(train_loader)dice1 = dice1_train / len(train_loader)dice2 = dice2_train / len(train_loader)dice3 = dice3_train / len(train_loader)return {'loss':loss,'dice1':dice1,'dice2':dice2,'dice3':dice3}def val_loop(model,criterion,val_loader,device):model.eval()running_loss = 0dice1_val = 0dice2_val = 0dice3_val = 0pbar = tqdm(val_loader)with torch.no_grad():for images, masks in pbar:images, masks = images.to(device), masks.to(device)outputs = model(images)# outputs = torch.softmax(outputs,dim=1)loss = criterion(outputs, masks)dice1, dice2, dice3 = cal_dice(outputs, masks)running_loss += loss.item()dice1_val += dice1.item()dice2_val += dice2.item()dice3_val += dice3.item()# pbar.desc = "loss:{:.3f} dice1:{:.3f} dice2:{:.3f} dice3:{:.3f} ".format(loss,dice1,dice2,dice3)loss = running_loss / len(val_loader)dice1 = dice1_val / len(val_loader)dice2 = dice2_val / len(val_loader)dice3 = dice3_val / len(val_loader)return {'loss':loss,'dice1':dice1,'dice2':dice2,'dice3':dice3}def train(model,optimizer,scheduler,criterion,train_loader,val_loader,epochs,device,train_log,valid_loss_min=999.0):for e in range(epochs):# train for epochtrain_metrics = train_loop(model,optimizer,scheduler,criterion,train_loader,device,e)# eval for epochval_metrics = val_loop(model,criterion,val_loader,device)info1 = "Epoch:[{}/{}] train_loss: {:.3f} valid_loss: {:.3f} ".format(e+1,epochs,train_metrics["loss"],val_metrics["loss"])info2 = "Train--ET: {:.3f} TC: {:.3f} WT: {:.3f} ".format(train_metrics['dice1'],train_metrics['dice2'],train_metrics['dice3'])info3 = "Valid--ET: {:.3f} TC: {:.3f} WT: {:.3f} ".format(val_metrics['dice1'],val_metrics['dice2'],val_metrics['dice3'])print(info1)print(info2)print(info3)with open(train_log,'a') as f:f.write(info1 + '\n' + info2 + ' ' + info3 + '\n')if not os.path.exists(args.save_path):os.makedirs(args.save_path)save_file = {"model": model.state_dict(), "optimizer": optimizer.state_dict()}if val_metrics['loss'] < valid_loss_min:valid_loss_min = val_metrics['loss']torch.save(save_file, 'results/UNet.pth')else:torch.save(save_file,os.path.join(args.save_path,'checkpoint{}.pth'.format(e+1)))print("Finished Training!")def main(args):torch.manual_seed(args.seed)# 为CPU设置种子用于生成随机数，以使得结果是确定的torch.cuda.manual_seed_all(args.seed)# 为所有的GPU设置种子，以使得结果是确定的torch.backends.cudnn.deterministic = Truetorch.backends.cudnn.benchmark = Trueos.environ['CUDA_VISIBLE_DEVICES'] = '0'device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')# data infopatch_size = (160,160,128)train_dataset = BraTS(args.data_path,args.train_txt,transform=transforms.Compose([RandomRotFlip(),RandomCrop(patch_size),GaussianNoise(p=0.1),ToTensor()]))val_dataset = BraTS(args.data_path,args.valid_txt,transform=transforms.Compose([CenterCrop(patch_size),ToTensor()]))test_dataset = BraTS(args.data_path,args.test_txt,transform=transforms.Compose([CenterCrop(patch_size),ToTensor()]))train_loader = DataLoader(dataset=train_dataset, batch_size=args.batch_size, num_workers=12, # num_worker=4shuffle=True, pin_memory=True)val_loader = DataLoader(dataset=val_dataset, batch_size=args.batch_size, num_workers=12, shuffle=False,pin_memory=True)test_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, num_workers=12, shuffle=False, pin_memory=True)print("using {} device.".format(device))print("using {} images for training, {} images for validation.".format(len(train_dataset), len(val_dataset)))# img,label = train_dataset[0]# 1-坏疽(NT,necrotic tumor core),2-浮肿区域(ED,peritumoral edema),4-增强肿瘤区域(ET,enhancing tumor)# 评价指标：ET(label4),TC(label1+label4),WT(label1+label2+label4)model = UNet(in_channels=4,num_classes=4).to(device)criterion = Loss(n_classes=4, weight=torch.tensor([0.2, 0.3, 0.25, 0.25])).to(device)optimizer = optim.SGD(model.parameters(),momentum=0.9, lr=0, weight_decay=5e-4)scheduler = cosine_scheduler(base_value=args.lr,final_value=args.min_lr,epochs=args.epochs, niter_per_ep=len(train_loader),warmup_epochs=args.warmup_epochs,start_warmup_value=5e-4)# 加载训练模型if os.path.exists(args.weights):weight_dict = torch.load(args.weights, map_location=device)model.load_state_dict(weight_dict['model'])optimizer.load_state_dict(weight_dict['optimizer'])print('Successfully loading checkpoint.')train(model,optimizer,scheduler,criterion,train_loader,val_loader,args.epochs,device,train_log=args.train_log)# metrics1 = val_loop(model, criterion, train_loader, device)metrics2 = val_loop(model, criterion, val_loader, device)metrics3 = val_loop(model, criterion, test_loader, device)# 最后再评价一遍所有数据，注意，这里使用的是训练结束的模型参数# print("Train -- loss: {:.3f} ET: {:.3f} TC: {:.3f} WT: {:.3f}".format(metrics1['loss'], metrics1['dice1'],metrics1['dice2'], metrics1['dice3']))print("Valid -- loss: {:.3f} ET: {:.3f} TC: {:.3f} WT: {:.3f}".format(metrics2['loss'], metrics2['dice1'], metrics2['dice2'], metrics2['dice3']))print("Test-- loss: {:.3f} ET: {:.3f} TC: {:.3f} WT: {:.3f}".format(metrics3['loss'], metrics3['dice1'], metrics3['dice2'], metrics3['dice3']))if __name__ == '__main__':parser = argparse.ArgumentParser()parser.add_argument('--num_classes', type=int, default=4)parser.add_argument('--seed', type=int, default=21)parser.add_argument('--epochs', type=int, default=60)parser.add_argument('--warmup_epochs', type=int, default=10)parser.add_argument('--batch_size', type=int, default=1)parser.add_argument('--lr', type=float, default=0.004)parser.add_argument('--min_lr', type=float, default=0.002)parser.add_argument('--data_path', type=str, default='/***/data_set/BraTS2021/dataset')parser.add_argument('--train_txt', type=str, default='/***/data_set/BraTS2021/train.txt')parser.add_argument('--valid_txt', type=str, default='/***/data_set/BraTS2021/valid.txt')parser.add_argument('--test_txt', type=str, default='/***/data_set/BraTS2021/test.txt')parser.add_argument('--train_log', type=str, default='results/UNet.txt')parser.add_argument('--weights', type=str, default='results/UNet.pth')parser.add_argument('--save_path', type=str, default='checkpoint/UNet')args = parser.parse_args()main(args)

训练集1000张，验证集125张，测试集126张。保存在验证集上损失最小的模型。

6.实验结果

训练30轮的loss曲线如上图所示，下面是我用不同的模型训练60轮，在测试集上的评价指标：

3D MRI Brain Tumor Segmentation（BraTS2021）

• Attention UNet在UNet的基础上，在上采样模块引入像素注意力。

7.滑动推理

加载训练好的权重，采用滑动窗口法进行推理，代码见inference.py

def test_single_case(net, image, stride_xy, stride_z, patch_size, num_classes=1):# print(image.shape)c, ww, hh, dd = image.shapesx = math.ceil((ww - patch_size[0]) / stride_xy) + 1sy = math.ceil((hh - patch_size[1]) / stride_xy) + 1sz = math.ceil((dd - patch_size[2]) / stride_z) + 1# print("{}, {}, {}".format(sx, sy, sz))score_map = np.zeros((num_classes, ) + image.shape[1:]).astype(np.float32)cnt = np.zeros(image.shape[1:]).astype(np.float32)for x in range(0, sx):xs = min(stride_xy*x, ww-patch_size[0])for y in range(0, sy):ys = min(stride_xy * y,hh-patch_size[1])for z in range(0, sz):zs = min(stride_z * z, dd-patch_size[2])test_patch = image[:,xs:xs+patch_size[0], ys:ys+patch_size[1], zs:zs+patch_size[2]]test_patch = np.expand_dims(test_patch,axis=0).astype(np.float32)test_patch = torch.from_numpy(test_patch).cuda()y1 = net(test_patch)y = F.softmax(y1, dim=1)y = y.cpu().data.numpy()y = y[0,:,:,:,:]score_map[:, xs:xs+patch_size[0], ys:ys+patch_size[1], zs:zs+patch_size[2]] \= score_map[:, xs:xs+patch_size[0], ys:ys+patch_size[1], zs:zs+patch_size[2]] + ycnt[xs:xs+patch_size[0], ys:ys+patch_size[1], zs:zs+patch_size[2]] \= cnt[xs:xs+patch_size[0], ys:ys+patch_size[1], zs:zs+patch_size[2]] + 1score_map = score_map/np.expand_dims(cnt,axis=0)label_map = np.argmax(score_map, axis = 0)return label_map, score_map

以标签1（NT, necrotic tumor core）为例，上图中红色的是金标签，蓝色的是UNet预测结果

确实，脑肿瘤分割相比其他三维分割任务，结果要好太多了，是一个非常适合练手的项目。感兴趣的同学可以按照我的步骤复现一下，效果也不会差。

代码我都放在上面了，码字不易，有用的话还请点个赞，后续也会更新图像分割和深度学习方面的内容，欢迎交流讨论。