nnUNetv2训练注意事项

 1.在配置环境的过程中，通常会遇到如下图这个情况，通常是batchgeneratorsv2这个包没有安装成功。

 解决方法：重新创建一个新环境，下载nnunet.zip

pip install -e . 最好用镜像安装，速度快，如下代码：

pip install -e . -i https://pypi.tuna.tsinghua.edu.cn/simple

就可以下载完成。

2.在训练的过程中会出现RuntimeError: One or more background workers are no longer alive. Exiting. Please check the print statements above for the actual error message报错，一般是由于磁盘空间不足，扩容即可。

3. 我的原始数据是png格式，在训练二维图像之前，由于这是2D图片数据，要将其转为3D数据，其实就是z轴为1的3维数据。
所以随意新建一个项目，创建一个2DDataProcessTo3D.py脚本

import os
import random
from tqdm import tqdm
import SimpleITK as sitk
import cv2
import numpy as np


def SplitDataset(img_path, train_percent=0.9):
    data = os.listdir(img_path)
    train_images = []
    test_images = []
    num = len(data)
    train_num = int(num * train_percent)
    indexes = list(range(num))
    train = random.sample(indexes, train_num)
    for i in indexes:
        if i in train:
            train_images.append(data[i])
        else:
            test_images.append(data[i])
    return train_images, test_images


def conver(img_path,  save_dir, mask_path=None, select_condition=None, mode="trian"):
    os.makedirs(save_dir, exist_ok=True)
    if mode == "train":
        savepath_img = os.path.join(save_dir, 'imagesTr')
        savepath_mask = os.path.join(save_dir, 'labelsTr')
    elif mode == "test":
        savepath_img = os.path.join(save_dir, 'imagesTs')
        savepath_mask = os.path.join(save_dir, 'labelsTs')
    os.makedirs(savepath_img, exist_ok=True)
    if mask_path is not None:
        os.makedirs(savepath_mask, exist_ok=True)

    ImgList = os.listdir(img_path)
    with tqdm(ImgList, desc="conver") as pbar:
        for name in pbar:
            if select_condition is not None and name not in select_condition:
                continue
            Img = cv2.imread(os.path.join(img_path, name))
            if mask_path is not None:
                Mask = cv2.imread(os.path.join(mask_path, name), 0)
                Mask = (Mask / 255).astype(np.uint8)
                if Img.shape[:2] != Mask.shape:
                    Mask = cv2.resize(Mask, (Img.shape[1], Img.shape[0]))
            Img_Transposed = np.transpose(Img, (2, 0, 1))
            Img_0 = Img_Transposed[0].reshape(1, Img_Transposed[0].shape[0], Img_Transposed[0].shape[1])
            Img_1 = Img_Transposed[1].reshape(1, Img_Transposed[1].shape[0], Img_Transposed[1].shape[1])
            Img_2 = Img_Transposed[2].reshape(1, Img_Transposed[2].shape[0], Img_Transposed[2].shape[1])
            if mask_path is not None:
                Mask = Mask.reshape(1, Mask.shape[0], Mask.shape[1])
    
            Img_0_name = name.split('.')[0] + '_0000.nii.gz'
            Img_1_name = name.split('.')[0] + '_0001.nii.gz'
            Img_2_name = name.split('.')[0] + '_0002.nii.gz'
            if mask_path is not None:
                Mask_name = name.split('.')[0] + '.nii.gz'
    
            Img_0_nii = sitk.GetImageFromArray(Img_0)
            Img_1_nii = sitk.GetImageFromArray(Img_1)
            Img_2_nii = sitk.GetImageFromArray(Img_2)
            if mask_path is not None:
                Mask_nii = sitk.GetImageFromArray(Mask)
    
            sitk.WriteImage(Img_0_nii, os.path.join(savepath_img, Img_0_name))
            sitk.WriteImage(Img_1_nii, os.path.join(savepath_img, Img_1_name))
            sitk.WriteImage(Img_2_nii, os.path.join(savepath_img, Img_2_name))
            if mask_path is not None:
                sitk.WriteImage(Mask_nii, os.path.join(savepath_mask, Mask_name))


if __name__ == "__main__":
    train_percent = 1
    img_path = r".../img"
    mask_path = r".../mask"
    output_folder = r"./dataset"
    os.makedirs(output_folder, exist_ok=True)
    train_images, test_images = SplitDataset(img_path, train_percent)
    conver(img_path, output_folder, mask_path, train_images, mode="train")
    conver(img_path, output_folder, mask_path, test_images, mode="test")

保存至/root/autodl-tmp/nnUNet/nnUNet-master/DATASET/nnUNet_raw/nnUNet_raw_data/Dataset600_Liver路径下

4.在训练2d数据集的过程中，会在最后自动推理并预测出测试集的分割图像以及评估指标，我在评估指标中增加了特异性和敏感性指标。最终的指标会在/nnUNet-master/DATASET/nnUNet_trained_models/Dataset600_Liver/nnUNetTrainer__nnUNetPlans__2d/fold_2/validation/summary.json中。具体改动如下，在/nnUNet-master/nnunetv2/evaluation/evaluate_predictions.py中

更新compute_tp_fp_fn_tn 函数

def compute_tp_fp_fn_tn(mask_ref: np.ndarray, mask_pred: np.ndarray, ignore_mask: np.ndarray = None):
    if ignore_mask is None:
        use_mask = np.ones_like(mask_ref, dtype=bool)
    else:
        use_mask = ~ignore_mask
    tp = np.sum((mask_ref & mask_pred) & use_mask)
    fp = np.sum(((~mask_ref) & mask_pred) & use_mask)
    fn = np.sum((mask_ref & (~mask_pred)) & use_mask)
    tn = np.sum(((~mask_ref) & (~mask_pred)) & use_mask)
    
    # 返回TP, FP, FN, TN
    return tp, fp, fn, tn

更新compute_metrics 函数 

def compute_metrics(reference_file: str, prediction_file: str, image_reader_writer: BaseReaderWriter,
                    labels_or_regions: Union[List[int], List[Union[int, Tuple[int, ...]]]],
                    ignore_label: int = None) -> dict:
    # 读取参考图像和预测图像
    seg_ref, seg_ref_dict = image_reader_writer.read_seg(reference_file)
    seg_pred, seg_pred_dict = image_reader_writer.read_seg(prediction_file)

    ignore_mask = seg_ref == ignore_label if ignore_label is not None else None

    results = {}
    results['reference_file'] = reference_file
    results['prediction_file'] = prediction_file
    results['metrics'] = {}
    
    # 针对每个标签或区域计算指标
    for r in labels_or_regions:
        results['metrics'][r] = {}
        mask_ref = region_or_label_to_mask(seg_ref, r)
        mask_pred = region_or_label_to_mask(seg_pred, r)
        tp, fp, fn, tn = compute_tp_fp_fn_tn(mask_ref, mask_pred, ignore_mask)
        
        # 计算Dice和IoU
        if tp + fp + fn == 0:
            results['metrics'][r]['Dice'] = np.nan
            results['metrics'][r]['IoU'] = np.nan
        else:
            results['metrics'][r]['Dice'] = 2 * tp / (2 * tp + fp + fn)
            results['metrics'][r]['IoU'] = tp / (tp + fp + fn)
        
        # 计算敏感性（Sensitivity）和特异性（Specificity）
        sensitivity = tp / (tp + fn) if (tp + fn) > 0 else np.nan
        specificity = tn / (tn + fp) if (tn + fp) > 0 else np.nan
        
        # 将结果存储到字典中
        results['metrics'][r]['FP'] = fp
        results['metrics'][r]['TP'] = tp
        results['metrics'][r]['FN'] = fn
        results['metrics'][r]['TN'] = tn
        results['metrics'][r]['n_pred'] = fp + tp
        results['metrics'][r]['n_ref'] = fn + tp
        results['metrics'][r]['Sensitivity'] = sensitivity
        results['metrics'][r]['Specificity'] = specificity

    return results

这样，在预测时就可以得到特异性以及敏感性指标了