MIC-DKFZ/nnUNet
选择Linux环境运行该项目,Windows环境需要更改较多的参数,暂不支持。
安装环境
数据处理
2D数据转成伪3D数据,即z轴层数为1或3(灰度图为1,彩图为3)。
参考:nnUNet/nnunet/dataset_conversion/Task120_Massachusetts_RoadSegm.py
直接转换数据,将数据存放到nnUNet_raw_data_base/nnUNet_raw_data/Task120_MassRoadsSeg
-
文件夹目录
└─Task120_MassRoadsSeg │ dataset.json │ ├─imagesTr │ img-1_0000.nii.gz │ img-1_0001.nii.gz │ img-1_0002.nii.gz │ img-2_0000.nii.gz │ img-2_0001.nii.gz │ img-2_0002.nii.gz │ img-3_0000.nii.gz │ img-3_0001.nii.gz │ img-3_0002.nii.gz │ img-4_0000.nii.gz │ img-4_0001.nii.gz │ img-4_0002.nii.gz │ img-5_0000.nii.gz │ img-5_0001.nii.gz │ img-5_0002.nii.gz │ ... ├─imagesTs │ img-10.nii.gz │ img-11.nii.gz │ ... └─labelsTr img-1.nii.gz img-2.nii.gz img-3.nii.gz img-4.nii.gz img-5.nii.gz ...
-
json文件信息
nnUNet/nnunet/dataset_conversion/utils.py
里面的函数generate_dataset_json
可以生成相应任务的json
文件。{ "name": "MassRoadsSeg", "description": "MassRoadsSeg", "reference": "https://www.kaggle.com/insaff/massachusetts-roads-dataset", "licence":'hands off!', "release":"1.0 06/08/2018", "tensorImageSize": "4D", "modality": { "0": "Red", "1": "Green", "2": "Blue" }, "labels": { "0": "background", "1": "street" }, "numTraining": 804, "numTest": 13, "training":[{"image":"./imagesTr/img-1.nii.gz","label":"./labelsTr/img-1.nii.gz"},{"image":"./imagesTr/img-2.nii.gz","label":"./labelsTr/img-2.nii.gz"},{"image":"./imagesTr/img-3.nii.gz","label":"./labelsTr/img-3.nii.gz"},{"image":"./imagesTr/img-4.nii.gz","label":"./labelsTr/img-4.nii.gz"},{"image":"./imagesTr/img-5.nii.gz","label":"./labelsTr/img-5.nii.gz"},...], "test":["./imagesTs/img-10.nii.gz","./imagesTs/img-11.nii.gz",...] }
预处理
# 只进行2d预处理,不进行3d预处理
nnUNet_plan_and_preprocess -t 120 -pl3d None
训练
nnUNet_train 2d nnUNetTrainerV2 120 0
测试
nnUNet_predict -i /root/nnUNet_raw_data_base/nnUNet_raw_data/Task120_MassRoadsSeg/imagesTs -o /root/nnUNet_trained_models/nnUNet/2d/Task120_MassRoadsSeg/nnUNetTrainerV2__nnUNetPlansv2.1/fold_0/infer -t 120 -m 2d