nnUnet 大模型学习笔记（续）：训练网络（3d_fullres）以及数据集标签的处理

目录

1. 数据集处理

1.1 实现脚本

1.2 json文件

2. 设置读取路径

2.1 设置路径

2.2 数据集转换

2.3 数据集预处理

2.4 训练（3d_fullres)

3. 训练结果展示

---

关于nnUnet 数据集的处理和环境搭建，参考上文：第四章：nnUnet大模型之环境配置、数据集制作-CSDN博客

1. 数据集处理

因为上文数据集的标签有很多问题，虽然处理起来很简单，为了防止后续需要，这里记录下

观察上文发现，数据的标签是19类别，但是mask的绘制不是连续的0 1 2 3，这样在图像分割中是

不允许的，需要做灰度映射。

实际上，在做unet一些列多类别分割的时候，已经介绍过自适应的灰度映射，这里只做简单介绍，具体参考下文：Unet 实战分割项目、多尺度训练、多类别分割_unet实例分割-CSDN博客

如果数据没有问题的话，直接跳到第二章即可！！

1.1 实现脚本

如下

import SimpleITK as sitk
import numpy as np
import os
from tqdm import tqdm
import shutildef main():root = 'labelsTr'images = [os.path.join(root, u) for u in os.listdir(root)]root_ret = 'ret_labelsTr'if os.path.exists(root_ret):shutil.rmtree(root_ret)os.mkdir(root_ret)# 计算灰度cl = []for i in tqdm(images, desc='process'):mask = sitk.ReadImage(i)mask = sitk.GetArrayFromImage(mask)mask = np.unique(mask)for h in mask:if h not in cl:cl.append(h)cl.sort()n = len(cl)print(cl)       # [0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22]print('分割的个数:',n)if n == cl[n-1]:return# 灰度映射for i in tqdm(images, desc='process'):mask = sitk.ReadImage(i)mask = sitk.GetArrayFromImage(mask)for index,h in enumerate(cl):mask[mask==h] = indexmask = sitk.GetImageFromArray(mask)ret_path = i.replace(root,root_ret)sitk.WriteImage(mask,ret_path)# 检查灰度cl_ret = []images = [os.path.join(root_ret, u) for u in os.listdir(root_ret)]for i in tqdm(images, desc='process'):mask = sitk.ReadImage(i)mask = sitk.GetArrayFromImage(mask)mask = np.unique(mask)for h in mask:if h not in cl_ret:cl_ret.append(h)cl_ret.sort()n = len(cl_ret)print(cl_ret)       # [0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22]print('处理后分割的个数:',n)if __name__ == '__main__':main()

摆放如下：脚本会将labelsTr的标签自动映射成0 1 2 3连续的，并且保存在新生成的ret下

运行如下：

可以看到mask的灰度已经进行了映射

通过itk打开，可以发现mask并没有改变，只是里面的数字变了，这样颜色显示也就变了

源标签：

处理完的：

1.2 json文件

更改如下：当然新的json文件可以用上文的脚本生成

{"labels": {"0": "background","1": "L1","2": "L2","3": "L3","4": "L4","5": "L5","6": "L6","7": "L7","8": "L8","9": "L9","10": "L10","11": "L11","12": "L12","13": "L13","14": "L14","15": "L15","16": "L16","17": "L17","18": "L18"},"modality": {"0": "CT"},"numTest": 0,"numTraining": 40,"tensorImageSize": "3D","test": [],"training": [{"image": "./imagesTr/spine_001.nii.gz","label": "./labelsTr/spine_001.nii.gz"},{"image": "./imagesTr/spine_002.nii.gz","label": "./labelsTr/spine_002.nii.gz"},{"image": "./imagesTr/spine_003.nii.gz","label": "./labelsTr/spine_003.nii.gz"},{"image": "./imagesTr/spine_004.nii.gz","label": "./labelsTr/spine_004.nii.gz"},{"image": "./imagesTr/spine_005.nii.gz","label": "./labelsTr/spine_005.nii.gz"},{"image": "./imagesTr/spine_006.nii.gz","label": "./labelsTr/spine_006.nii.gz"},{"image": "./imagesTr/spine_007.nii.gz","label": "./labelsTr/spine_007.nii.gz"},{"image": "./imagesTr/spine_008.nii.gz","label": "./labelsTr/spine_008.nii.gz"},{"image": "./imagesTr/spine_009.nii.gz","label": "./labelsTr/spine_009.nii.gz"},{"image": "./imagesTr/spine_010.nii.gz","label": "./labelsTr/spine_010.nii.gz"},{"image": "./imagesTr/spine_011.nii.gz","label": "./labelsTr/spine_011.nii.gz"},{"image": "./imagesTr/spine_012.nii.gz","label": "./labelsTr/spine_012.nii.gz"},{"image": "./imagesTr/spine_013.nii.gz","label": "./labelsTr/spine_013.nii.gz"},{"image": "./imagesTr/spine_014.nii.gz","label": "./labelsTr/spine_014.nii.gz"},{"image": "./imagesTr/spine_015.nii.gz","label": "./labelsTr/spine_015.nii.gz"},{"image": "./imagesTr/spine_016.nii.gz","label": "./labelsTr/spine_016.nii.gz"},{"image": "./imagesTr/spine_017.nii.gz","label": "./labelsTr/spine_017.nii.gz"},{"image": "./imagesTr/spine_018.nii.gz","label": "./labelsTr/spine_018.nii.gz"},{"image": "./imagesTr/spine_019.nii.gz","label": "./labelsTr/spine_019.nii.gz"},{"image": "./imagesTr/spine_020.nii.gz","label": "./labelsTr/spine_020.nii.gz"},{"image": "./imagesTr/spine_021.nii.gz","label": "./labelsTr/spine_021.nii.gz"},{"image": "./imagesTr/spine_022.nii.gz","label": "./labelsTr/spine_022.nii.gz"},{"image": "./imagesTr/spine_023.nii.gz","label": "./labelsTr/spine_023.nii.gz"},{"image": "./imagesTr/spine_024.nii.gz","label": "./labelsTr/spine_024.nii.gz"},{"image": "./imagesTr/spine_025.nii.gz","label": "./labelsTr/spine_025.nii.gz"},{"image": "./imagesTr/spine_026.nii.gz","label": "./labelsTr/spine_026.nii.gz"},{"image": "./imagesTr/spine_027.nii.gz","label": "./labelsTr/spine_027.nii.gz"},{"image": "./imagesTr/spine_028.nii.gz","label": "./labelsTr/spine_028.nii.gz"},{"image": "./imagesTr/spine_029.nii.gz","label": "./labelsTr/spine_029.nii.gz"},{"image": "./imagesTr/spine_030.nii.gz","label": "./labelsTr/spine_030.nii.gz"},{"image": "./imagesTr/spine_031.nii.gz","label": "./labelsTr/spine_031.nii.gz"},{"image": "./imagesTr/spine_032.nii.gz","label": "./labelsTr/spine_032.nii.gz"},{"image": "./imagesTr/spine_033.nii.gz","label": "./labelsTr/spine_033.nii.gz"},{"image": "./imagesTr/spine_034.nii.gz","label": "./labelsTr/spine_034.nii.gz"},{"image": "./imagesTr/spine_035.nii.gz","label": "./labelsTr/spine_035.nii.gz"},{"image": "./imagesTr/spine_036.nii.gz","label": "./labelsTr/spine_036.nii.gz"},{"image": "./imagesTr/spine_037.nii.gz","label": "./labelsTr/spine_037.nii.gz"},{"image": "./imagesTr/spine_038.nii.gz","label": "./labelsTr/spine_038.nii.gz"},{"image": "./imagesTr/spine_039.nii.gz","label": "./labelsTr/spine_039.nii.gz"},{"image": "./imagesTr/spine_040.nii.gz","label": "./labelsTr/spine_040.nii.gz"}]
}

2. 设置读取路径

回到正文，这里的Task下有如下数据，source nnunet/bin/activate 激活nnunet环境

Tips：这里的 labelsTr和dataset.json是第一节处理后的

任务名称为Task01_Spine

2.1 设置路径

这里设置为绝对路径，除了DATASET后面的，前面部分需要根据不同机器设定

在这里更改 vim .bashrc（vim ~/.bashrc 末尾最后面）

export nnUNet_raw_data_base="/*/DATASET/nnUNet_raw"
export nnUNet_preprocessed="/*/DATASET/nnUNet_preprocessed"
export RESULTS_FOLDER="/*/DATASET/nnUNet_trained_models"

这里设置后，如果想要训练其他模型，不需要在进行更改

添加完成后保存， source ~/.bashrc 更新环境变量，可以通过echo $RESULTS_FOLDER 检查是否修改成功

2.2 数据集转换

下面命令都是在environments 目录里进行操作

转换命令为：

nnUNet_convert_decathlon_task -i DATASET/nnUNet_raw/nnUNet_raw_data/Task01_Spine/

转换完的数据在：

图像可能具有多种模态，nnU-Net通过其后缀（文件名末尾的四位整数）识别成像模态。因此，图像文件必须遵循以下命名约定：case_identifier_XXXX.nii.gz。

这里，XXXX是模态标识符。dataset.json文件中指定了这些标识符所属的模态。

标签文件保存为case_identifier.nii.gz

例如：BrainTumor。每个图像有四种模态：FLAIR（0000）、T1w（0001）、T1gd（0002）和T2w（0003）

2.3 数据集预处理

命令如下：（这里只会做训练集进行预处理，测试集不会处理）

nnUNet_plan_and_preprocess -t 1

只需要一行命令，因为 Task_id是1，所以这里的数字就是1。这个过程会消耗很多的时间，速度慢的原因在于对要进行插值等各种操作。

生成的数据在crop和precocessed里面查看

2.4 训练（3d_fullres)

命令如下：

nnUNet_train 3d_fullres nnUNetTrainerV2 1 0

1 指的是Task标号，5 指定训练的是5倍交叉验证的哪一倍。

会实时生成如下结果：在这里 nnUNet_trained_models

3. 训练结果展示

RTX 3090跑一个epoch大概100s，1000个epoch估计要一两天，等跑完下篇文章在贴训练结果吧