3dunet训练数据集

代码：

GitHub - lee-zq/3DUNet-Pytorch: 3DUNet implemented with pytorch3DUNet implemented with pytorch. Contribute to lee-zq/3DUNet-Pytorch development by creating an account on GitHub.https://github.com/lee-zq/3DUNet-Pytorch

数据集：LITS2017肝脏肿瘤分割挑战数据集 

步骤：预处理 => 训练 => 测试

预处理：根目录下建立文件夹raw_dataset 和fixed_data文件夹

raw_dataset:
    ├── train  # (28~46)
    │   ├── ct
    │   │   ├── volume-27.nii
    │   │   ├── volume-28.nii ...
    │   └── label
    │       ├── segmentation-27.nii
    │       ├── segmentation-28.nii ...
    │       
    ├── test # 0~27和47~131
    │   ├── ct
    │   │   ├── volume-0.nii
    │   │   ├── volume-1.nii ...
    │   └── label
    │       ├── segmentation-0.nii
    │       ├── segmentation-1.nii ...

fixed_data:
    ├── ct  
    │   
    ├── label     

修改preprocess_LiTS.py文件,并运行

if __name__ == '__main__':
    raw_dataset_path = r'G:\B\3DUNet\raw_dataset\train'
    fixed_dataset_path = r'G:\B\3DUNet\fixed_data'

训练：修改config.py文件中 dataset_path的路径

parser.add_argument('--dataset_path',default = r'G:\B\3DUNet\fixed_data',help='fixed trainset root path')

修改train.py的79行

model = torch.nn.DataParallel(model, device_ids=[0]).cuda()    #单GPU
#model = torch.nn.DataParallel(model, device_ids=args.gpu_id)  # multi-GPU

修改models/ResUNet.py 137行--159行中的

nn.Upsample(scale_factor=(k, x, x), mode='trilinear', align_corners=False)

改为nn.Upsample(scale_factor=(k, 0.5x, 0.5x), mode='trilinear', align_corners=False)

运行python train.py --save model_name得到模型best_model.pth保存在experiments/model_name中

预测：

将上述 models/ResUNet.py 137行--159行的(k, 0.5x, 0.5x)修改回去

改config.py测试集路径和模型路径

parser.add_argument('--test_data_path',default = r'G:\B\3DUNet\raw_dataset\test',help='Testset path')
parser.add_argument('--save',default='model_name',help='save path of trained model')

 改test.py50行

model = torch.nn.DataParallel(model, device_ids=[0]).cuda()    #单GPU
#model = torch.nn.DataParallel(model, device_ids=args.gpu_id)  # multi-GPU

运行test.py得到的结果保存在model_name/result中