【3D 图像分割】基于 Pytorch 的 VNet 3D 图像分割6（数据预处理）

由于之前哔站作者整理的LUNA16数据处理方式过于的繁琐，于是，本文就对LUNA16数据做一个新的整理，最终得到的数据和形式是差不多的。但是，主要不同的是代码逻辑比较的简单，便于理解。

对于数据集的学习，可以去参考这里：【LIDC-IDRI】 CT 肺结节 XML 标记特征良恶性标签PKL转储（一）

步骤和中心内容，包括一下几个部分：

masks生成：从xml文件中，抽取出对应序列series的结节标记位置坐标（可能一个结节多人多次标注），生成对应的mask数组文件，大小与图像数组大小一致；
肺实质提取操作：从肺区分割的数据中，与原始图像和mask图做乘积操作，非肺区部分进行填充，或者去除操作均可；
resample操作：根据spacing，进行resample操作，可以在zyx三个维度进行resample，也可以仅仅在z方向进行resample操作位1mm（这个我在论文中看到有类似这样做的）；
根据mask，获取结节的zyx中心点坐标，和半径。

至此，我们将收获以下几个文件：

包含ct的图像数据；
对应的mask数据；
记录zyx中心点坐标，和半径的文件。

相比于luna16给出的数据形式，目前的数据就比较好理解，和方便查看了。无论是可视化，还是后续的数据处理和训练，都更加的直观、明了。后面就会针对这部分，一一进行展开。

由于代码量还是比较大，处理的东西，和涉及的文件比较多，可能会几个篇章展开。本篇就先对xml文件进行处理，转出来，以便于查看。这里涉及到xml文件的格式，和处理，就单独开一篇，链接去参考：【医学影像数据处理】 XML 文件格式处理汇总

一、xml文件转储

1.1、认识标注文件xml

对于LIDC-IDRI数据集中，xml文件内各个字段表示什么意思的介绍，可以参考我的另一篇文章，点击这里：【LIDC-IDRI】 CT 肺结节 XML 标记特征良恶性标签PKL转储（一）

在这篇文章里面，着重介绍了这个数据的结构，以及xml各个记录的tag是什么意思。相信你看完，对这个数据集的处理，有更深的理解。

其中大部分代码都是跟上面这个链接介绍和获取的内容是一样的，可以参考这个GitHub：NoduleNet - utils -LIDC

有些内容没有介绍到，简单做个补充

ResponseHeader：这个是头部分，记录了这个病例（也就是单个病人的CT图像）的信息。

为了方便查看，和学习xml文件，可以参考这篇文章：【医学影像数据处理】 XML 文件格式处理汇总。我们就采用其中xml转字典的形式，方便我们查看。下面就展示了转字典后的前后部分内容对比，如下：

原始xml的数据形式，节选了其中的一小段，展示如下：

xml 复制代码

<?xml version="1.0" encoding="UTF-8"?>
<LidcReadMessage uid="1.3.6.1.4.1.14519.5.2.1.6279.6001.1308168927505.0" xmlns="http://www.nih.gov" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xsi:schemaLocation="http://www.nih.gov http://troll.rad.med.umich.edu/lidc/LidcReadMessage.xsd">
  <ResponseHeader>
    <Version>1.7</Version>
    <MessageId>1148851</MessageId>
    <DateRequest>2005-11-03</DateRequest>
    <TimeRequest>12:25:10</TimeRequest>
    <RequestingSite>removed</RequestingSite>
    <ServicingSite>removed</ServicingSite>
    <TaskDescription>Second unblinded read</TaskDescription>
    <CtImageFile>removed</CtImageFile>
    <SeriesInstanceUid>1.3.6.1.4.1.14519.5.2.1.6279.6001.131939324905446238286154504249</SeriesInstanceUid>
    <StudyInstanceUID>1.3.6.1.4.1.14519.5.2.1.6279.6001.303241414168367763244410429787</StudyInstanceUID>
    <DateService>2005-11-03</DateService>
    <TimeService>12:25:40</TimeService>
    <ResponseDescription>1 - Reading complete</ResponseDescription>
    <ResponseComments></ResponseComments>
  </ResponseHeader>

转换成dictionary字典后的形式。（更便于查看了）

dict 复制代码

{
  "LidcReadMessage": {
    "@uid": "1.3.6.1.4.1.14519.5.2.1.6279.6001.1308168927505.0",
    "@xmlns": "http://www.nih.gov",
    "@xmlns:xsi": "http://www.w3.org/2001/XMLSchema-instance",
    "@xsi:schemaLocation": "http://www.nih.gov http://troll.rad.med.umich.edu/lidc/LidcReadMessage.xsd",
    "ResponseHeader": {
      "Version": "1.7",
      "MessageId": "1148851",
      "DateRequest": "2005-11-03",
      "TimeRequest": "12:25:10",
      "RequestingSite": "removed",
      "ServicingSite": "removed",
      "TaskDescription": "Second unblinded read",
      "CtImageFile": "removed",
      "SeriesInstanceUid": "1.3.6.1.4.1.14519.5.2.1.6279.6001.131939324905446238286154504249",
      "StudyInstanceUID": "1.3.6.1.4.1.14519.5.2.1.6279.6001.303241414168367763244410429787",
      "DateService": "2005-11-03",
      "TimeService": "12:25:40",
      "ResponseDescription": "1 - Reading complete",
      "ResponseComments": null
    },
}

1.2、xml综合记录转为按series的npy文件

LIDC-IDRI有1018个检查，在标记文件夹tcia-lidc-xml中6 个文件夹中，有1318 个xml文件。并且，这些xml文件的名称，和图像的序列名称不是一一对应的。

所以，就需要现将xml文件内标注的信息，给重新整理出来，转为人能轻易看懂和理解的内容。并且，标注文件如果能与图像文件是一一对应的，那么后续的处理也会方便了许多。

这一小节做的事情，就是将xml文件，给抽取出来，留下关心的内容，其他不重要的，不关心的内容暂时不管。

下面是处理的代码，主要的步骤如下概述：

遍历所有的xml文件，一一处理；
对单个xml文件，解析出seriesuid和标注的结节坐标；
存储到以seriesuid命名的npy文件，存储的内容就是一个个结节坐标。

完整代码如下：

python 复制代码

from tqdm import tqdm
import sys
import os
import numpy as np

from pylung.utils import find_all_files
from pylung.annotation import parse

def xml2mask(xml_file):
    header, annos = parse(xml_file)  # get one xml info

    ctr_arrs = []
    for i, reader in enumerate(annos):
        for j, nodule in enumerate(reader.nodules):
            ctr_arr = []
            for k, roi in enumerate(nodule.rois):
                z = roi.z
                for roi_xy in roi.roi_xy:
                    ctr_arr.append([z, roi_xy[1], roi_xy[0]])  # [[[z, y, x], [z, y, x]]]
            ctr_arrs.append(ctr_arr)

    seriesuid = header.series_instance_uid
    return seriesuid, ctr_arrs

def annotation2masks(annos_dir, save_dir):
    # get all xml file path
    files = find_all_files(annos_dir, '.xml')
    for f in tqdm(files, total=len(files)):
        print(f)
        try:
            seriesuid, masks = xml2mask(f)
            np.save(os.path.join(save_dir, '%s' % (seriesuid)), masks)  # save xml 3D coor [[z, y, x], [z, y, x]]
        except:
            print("Unexpected error:", sys.exc_info()[0])


if __name__ == '__main__':
    annos_dir = './LUNA16/annotation/LIDC-XML-only/tcia-lidc-xml'     # .xml
    ctr_arr_save_dir = './LUNA16/annotation/noduleCoor'  # 保存每个注释器解析的中间结节mask的地方

    os.makedirs(ctr_arr_save_dir, exist_ok=True)

    # xml信息，转储npy（临时文件）
    annotation2masks(annos_dir, ctr_arr_save_dir)

下面打开一个·npy·文件进行查看，记录的内容如下，是所有医生对这个序列标注的所有结节的polygon坐标点：

[list([[-299.8, 206, 42], [-299.8, 207, 41], [-299.8, 208, 41], [-299.8, 209, 40], [-299.8, 210, 40], [-299.8, 211, 41], [-299.8, 212, 41], [-299.8, 213, 42], [-299.8, 214, 42], [-299.8, 215, 43], [-299.8, 216, 44], [-299.8, 216, 45], [-299.8, 215, 46], [-299.8, 215, 47], [-299.8, 215, 48], [-299.8, 214, 49], [-299.8, 213, 49], [-299.8, 212, 49], [-299.8, 211, 49], [-299.8, 210, 49], [-299.8, 209, 49], [-299.8, 208, 48], [-299.8, 207, 47], [-299.8, 207, 46], [-299.8, 206, 45], [-299.8, 206, 44], [-299.8, 206, 43], [-299.8, 206, 42], [-298.0, 206, 46], [-298.0, 207, 45], [-298.0, 207, 44], [-298.0, 208, 43], [-298.0, 209, 42], [-298.0, 209, 41], [-298.0, 210, 40], [-298.0, 211, 40], [-298.0, 212, 39], [-298.0, 213, 40], [-298.0, 214, 41], [-298.0, 215, 42], [-298.0, 215, 43], [-298.0, 216, 44], [-298.0, 216, 45], [-298.0, 216, 46], [-298.0, 216, 47], [-298.0, 215, 48], [-298.0, 214, 48], [-298.0, 213, 48], [-298.0, 212, 48], [-298.0, 211, 48], [-298.0, 210, 48], [-298.0, 209, 48], [-298.0, 208, 48], [-298.0, 207, 47], [-298.0, 206, 46], [-296.2, 209, 42], [-296.2, 210, 41], [-296.2, 211, 40], [-296.2, 212, 40], [-296.2, 213, 41], [-296.2, 214, 42], [-296.2, 215, 43], [-296.2, 216, 44], [-296.2, 216, 45], [-296.2, 216, 46], [-296.2, 216, 47], [-296.2, 216, 48], [-296.2, 215, 49], [-296.2, 214, 49], [-296.2, 213, 49], [-296.2, 212, 49], [-296.2, 211, 48], [-296.2, 210, 47], [-296.2, 209, 46], [-296.2, 209, 45], [-296.2, 209, 44], [-296.2, 209, 43], [-296.2, 209, 42]])
 list([[-227.8, 151, 405], [-227.8, 152, 404], [-227.8, 153, 403], [-227.8, 154, 402], [-227.8, 155, 402], [-227.8, 156, 402], [-227.8, 157, 403], [-227.8, 157, 404], [-227.8, 157, 405], [-227.8, 158, 406], [-227.8, 158, 407], [-227.8, 158, 408], [-227.8, 157, 409], [-227.8, 156, 409], [-227.8, 155, 409], [-227.8, 154, 408], [-227.8, 153, 408], [-227.8, 152, 407], [-227.8, 151, 406], [-227.8, 151, 405], [-226.0, 152, 405], [-226.0, 153, 404], [-226.0, 154, 404], [-226.0, 155, 403], [-226.0, 156, 404], [-226.0, 157, 405], [-226.0, 157, 406], [-226.0, 157, 407], [-226.0, 156, 408], [-226.0, 155, 408], [-226.0, 154, 408], [-226.0, 153, 408], [-226.0, 152, 407], [-226.0, 152, 406], [-226.0, 152, 405]])
 list([[-226.0, 158, 407], [-226.0, 157, 408], [-226.0, 156, 409], [-226.0, 155, 409], [-226.0, 154, 409], [-226.0, 153, 409], [-226.0, 152, 408], [-226.0, 151, 407], [-226.0, 152, 406], [-226.0, 153, 405], [-226.0, 153, 404], [-226.0, 154, 403], [-226.0, 155, 402], [-226.0, 156, 402], [-226.0, 157, 403], [-226.0, 158, 404], [-226.0, 158, 405], [-226.0, 158, 406], [-226.0, 158, 407], [-227.8, 159, 407], [-227.8, 158, 408], [-227.8, 157, 409], [-227.8, 156, 410], [-227.8, 155, 410], [-227.8, 154, 410], [-227.8, 153, 409], [-227.8, 152, 408], [-227.8, 151, 407], [-227.8, 151, 406], [-227.8, 151, 405], [-227.8, 152, 404], [-227.8, 153, 403], [-227.8, 154, 402], [-227.8, 155, 402], [-227.8, 156, 402], [-227.8, 157, 403], [-227.8, 158, 404], [-227.8, 158, 405], [-227.8, 158, 406], [-227.8, 159, 407]])
 list([[-296.2, 214, 46], [-296.2, 213, 47], [-296.2, 212, 47], [-296.2, 211, 47], [-296.2, 210, 46], [-296.2, 209, 45], [-296.2, 208, 44], [-296.2, 208, 43], [-296.2, 208, 42], [-296.2, 209, 41], [-296.2, 210, 42], [-296.2, 211, 42], [-296.2, 212, 43], [-296.2, 213, 44], [-296.2, 214, 45], [-296.2, 214, 46], [-298.0, 216, 47], [-298.0, 215, 48], [-298.0, 214, 49], [-298.0, 213, 49], [-298.0, 212, 49], [-298.0, 211, 49], [-298.0, 210, 49], [-298.0, 209, 48], [-298.0, 208, 47], [-298.0, 207, 46], [-298.0, 207, 45], [-298.0, 207, 44], [-298.0, 208, 43], [-298.0, 208, 42], [-298.0, 209, 41], [-298.0, 210, 41], [-298.0, 211, 41], [-298.0, 212, 41], [-298.0, 213, 41], [-298.0, 214, 42], [-298.0, 215, 43], [-298.0, 216, 44], [-298.0, 216, 45], [-298.0, 216, 46], [-298.0, 216, 47], [-299.8, 216, 50], [-299.8, 215, 51], [-299.8, 214, 51], [-299.8, 213, 50], [-299.8, 212, 50], [-299.8, 211, 50], [-299.8, 210, 49], [-299.8, 209, 48], [-299.8, 208, 47], [-299.8, 207, 46], [-299.8, 207, 45], [-299.8, 207, 44], [-299.8, 208, 43], [-299.8, 209, 42], [-299.8, 210, 42], [-299.8, 211, 41], [-299.8, 212, 41], [-299.8, 213, 42], [-299.8, 214, 42], [-299.8, 215, 43], [-299.8, 216, 44], [-299.8, 216, 45], [-299.8, 216, 46], [-299.8, 216, 47], [-299.8, 216, 48], [-299.8, 216, 49], [-299.8, 216, 50]])
 list([[-226.0, 158, 407], [-226.0, 157, 408], [-226.0, 156, 409], [-226.0, 155, 409], [-226.0, 154, 409], [-226.0, 153, 409], [-226.0, 152, 409], [-226.0, 151, 409], [-226.0, 151, 408], [-226.0, 151, 407], [-226.0, 151, 406], [-226.0, 151, 405], [-226.0, 152, 404], [-226.0, 152, 403], [-226.0, 153, 403], [-226.0, 154, 402], [-226.0, 154, 401], [-226.0, 155, 401], [-226.0, 156, 401], [-226.0, 157, 401], [-226.0, 157, 402], [-226.0, 158, 403], [-226.0, 158, 404], [-226.0, 158, 405], [-226.0, 158, 406], [-226.0, 158, 407], [-227.8, 159, 407], [-227.8, 158, 408], [-227.8, 158, 409], [-227.8, 157, 409], [-227.8, 156, 410], [-227.8, 155, 410], [-227.8, 154, 409], [-227.8, 153, 409], [-227.8, 152, 409], [-227.8, 151, 408], [-227.8, 151, 407], [-227.8, 151, 406], [-227.8, 151, 405], [-227.8, 151, 404], [-227.8, 152, 403], [-227.8, 152, 402], [-227.8, 153, 401], [-227.8, 154, 401], [-227.8, 155, 401], [-227.8, 156, 401], [-227.8, 157, 401], [-227.8, 158, 402], [-227.8, 158, 403], [-227.8, 159, 404], [-227.8, 159, 405], [-227.8, 159, 406], [-227.8, 159, 407]])
 list([[-296.2, 215, 47], [-296.2, 214, 48], [-296.2, 213, 48], [-296.2, 212, 48], [-296.2, 211, 48], [-296.2, 210, 47], [-296.2, 209, 47], [-296.2, 208, 46], [-296.2, 208, 45], [-296.2, 207, 44], [-296.2, 207, 43], [-296.2, 208, 42], [-296.2, 209, 42], [-296.2, 210, 42], [-296.2, 211, 42], [-296.2, 212, 43], [-296.2, 213, 43], [-296.2, 214, 44], [-296.2, 215, 45], [-296.2, 215, 46], [-296.2, 215, 47], [-298.0, 216, 47], [-298.0, 215, 48], [-298.0, 214, 49], [-298.0, 214, 50], [-298.0, 213, 50], [-298.0, 212, 50], [-298.0, 211, 49], [-298.0, 210, 49], [-298.0, 209, 48], [-298.0, 208, 48], [-298.0, 207, 47], [-298.0, 207, 46], [-298.0, 207, 45], [-298.0, 207, 44], [-298.0, 207, 43], [-298.0, 207, 42], [-298.0, 207, 41], [-298.0, 208, 41], [-298.0, 209, 41], [-298.0, 210, 41], [-298.0, 211, 41], [-298.0, 212, 41], [-298.0, 213, 41], [-298.0, 214, 41], [-298.0, 215, 42], [-298.0, 215, 43], [-298.0, 216, 44], [-298.0, 216, 45], [-298.0, 216, 46], [-298.0, 216, 47], [-299.8, 217, 46], [-299.8, 216, 47], [-299.8, 216, 48], [-299.8, 215, 49], [-299.8, 214, 50], [-299.8, 213, 50], [-299.8, 212, 50], [-299.8, 211, 50], [-299.8, 210, 50], [-299.8, 209, 49], [-299.8, 208, 48], [-299.8, 208, 47], [-299.8, 207, 46], [-299.8, 207, 45], [-299.8, 207, 44], [-299.8, 208, 43], [-299.8, 209, 42], [-299.8, 209, 41], [-299.8, 210, 41], [-299.8, 211, 41], [-299.8, 212, 41], [-299.8, 213, 41], [-299.8, 214, 42], [-299.8, 215, 42], [-299.8, 215, 43], [-299.8, 216, 44], [-299.8, 217, 45], [-299.8, 217, 46], [-301.6, 214, 45], [-301.6, 213, 46], [-301.6, 212, 47], [-301.6, 211, 47], [-301.6, 210, 46], [-301.6, 209, 45], [-301.6, 210, 44], [-301.6, 211, 43], [-301.6, 212, 43], [-301.6, 213, 44], [-301.6, 214, 45]])
 list([[-296.2, 209, 43], [-296.2, 209, 44], [-296.2, 210, 45], [-296.2, 211, 46], [-296.2, 212, 47], [-296.2, 212, 48], [-296.2, 213, 48], [-296.2, 214, 48], [-296.2, 215, 47], [-296.2, 215, 46], [-296.2, 215, 45], [-296.2, 214, 44], [-296.2, 213, 43], [-296.2, 212, 43], [-296.2, 211, 43], [-296.2, 210, 43], [-296.2, 209, 43], [-298.0, 208, 42], [-298.0, 208, 43], [-298.0, 208, 44], [-298.0, 208, 45], [-298.0, 208, 46], [-298.0, 208, 47], [-298.0, 209, 47], [-298.0, 210, 48], [-298.0, 211, 48], [-298.0, 211, 49], [-298.0, 212, 49], [-298.0, 213, 48], [-298.0, 214, 48], [-298.0, 215, 47], [-298.0, 216, 46], [-298.0, 216, 45], [-298.0, 216, 44], [-298.0, 215, 43], [-298.0, 214, 43], [-298.0, 213, 42], [-298.0, 212, 42], [-298.0, 212, 41], [-298.0, 211, 41], [-298.0, 210, 41], [-298.0, 209, 42], [-298.0, 208, 42], [-299.8, 210, 43], [-299.8, 209, 43], [-299.8, 208, 44], [-299.8, 207, 44], [-299.8, 207, 45], [-299.8, 207, 46], [-299.8, 208, 47], [-299.8, 209, 48], [-299.8, 210, 49], [-299.8, 211, 49], [-299.8, 212, 49], [-299.8, 213, 50], [-299.8, 214, 49], [-299.8, 215, 48], [-299.8, 215, 47], [-299.8, 216, 46], [-299.8, 216, 45], [-299.8, 215, 44], [-299.8, 215, 43], [-299.8, 214, 43], [-299.8, 214, 42], [-299.8, 213, 42], [-299.8, 212, 41], [-299.8, 211, 41], [-299.8, 210, 42], [-299.8, 210, 43]])] <class 'numpy.ndarray'>

二、标记次数和mask数组生成

生成npy文件并不是此次标注信息的最终结果，有以下几个原因：

xml文件内标注的结节坐标是多个医生分别标注的，所以会存在标注上的重叠（也就是一个结节被多个医生重复标注，很多是背靠背标注，也不知道其他医生标注了什么）。所以需要对多人标注的内容进行处理，留下最终的结节坐标；
只是坐标点，还需要生成和image一样shape，相互对应的mask文件。

根据上面几个原因，生成最终mask文件，就需要经历以下几个步骤：

标记的结节坐标点，需要将hu z到instanceNum处理，对应的图像上；
对多个医生标注的结节，进行处理，根据iou重叠规则，留下最终的结节；
留下的结节坐标，绘制到mask上，存储下来。

实现代码如下：

python 复制代码

import nrrd
import SimpleITK as sitk
import cv2
import os
import numpy as np

def load_itk_image(filename):
    """
    Return img array and [z,y,x]-ordered origin and spacing
    """
    # sitk.ReadImage返回的image的shape是x、y、z
    itkimage = sitk.ReadImage(filename)
    numpyImage = sitk.GetArrayFromImage(itkimage)

    numpyOrigin = np.array(list(reversed(itkimage.GetOrigin())))
    numpySpacing = np.array(list(reversed(itkimage.GetSpacing())))

    return numpyImage, numpyOrigin, numpySpacing


def arrs2mask(img_dir, ctr_arr_dir, save_dir):
    cnt = 0
    consensus = {1: 0, 2: 0, 3: 0, 4: 0}  # 一致意见

    # generate save document
    for k in consensus.keys():
        if not os.path.exists(os.path.join(save_dir, str(k))):
            os.makedirs(os.path.join(save_dir, str(k)))

    for f in os.listdir(img_dir):
        if f.endswith('.mhd'):
            pid = f[:-4]
            print('pid:', pid)
            # ct
            img, origin, spacing = load_itk_image(os.path.join(img_dir, '%s.mhd' % (pid)))

            # mask coor npy
            ctr_arrs = np.load(os.path.join(ctr_arr_dir, '%s.npy' % (pid)), allow_pickle=True)
            cnt += len(ctr_arrs)

            nodule_masks = []
            # 依次标注结节处理
            for ctr_arr in ctr_arrs:
                z_origin = origin[0]
                z_spacing = spacing[0]

                ctr_arr = np.array(ctr_arr)
                # ctr_arr[:, 0] z轴方向值，由hu z到instanceNum  [-50, -40, -30]-->[2, 3, 4]
                ctr_arr[:, 0] = np.absolute(ctr_arr[:, 0] - z_origin) / z_spacing  # 对数组中的每一个元素求其绝对值。np.abs是这个函数的简写
                ctr_arr = ctr_arr.astype(np.int32)
                print(ctr_arr)

                # 每一个标注的结节，都会新临时生成一个与img一样大小的mask文件
                mask = np.zeros(img.shape)
                # 遍历标注层的 z 轴序列
                for z in np.unique(ctr_arr[:, 0]):  # 去除其中重复的元素 ，并按元素 由小到大排序
                    ctr = ctr_arr[ctr_arr[:, 0] == z][:, [2, 1]]
                    ctr = np.array([ctr], dtype=np.int32)
                    mask[z] = cv2.fillPoly(mask[z], ctr, color=(1,))
                nodule_masks.append(mask)

            i = 0
            visited = []
            d = {}
            masks = []
            while i < len(nodule_masks):
                # If mached before, then no need to create new mask
                if i in visited:
                    i += 1
                    continue
                same_nodules = []
                mask1 = nodule_masks[i]
                same_nodules.append(mask1)
                d[i] = {}
                d[i]['count'] = 1
                d[i]['iou'] = []

                # Find annotations pointing to the same nodule
                # 当前结节mask[i]，与其后面的所有结节，依次求iou
                for j in range(i + 1, len(nodule_masks)):
                    # if not overlapped with previous added nodules
                    if j in visited:
                        continue
                    mask2 = nodule_masks[j]
                    iou = float(np.logical_and(mask1, mask2).sum()) / np.logical_or(mask1, mask2).sum()

                    # 如果iou超过阈值，则当前第i个mask记为被重复标记一次
                    if iou > 0.4:
                        visited.append(j)
                        same_nodules.append(mask2)
                        d[i]['count'] += 1
                        d[i]['iou'].append(iou)

                masks.append(same_nodules)
                i += 1

            print(visited)
            exit()
            # only 4 people, check up 4 data
            for k, v in d.items():
                if v['count'] > 4:
                    print('WARNING:  %s: %dth nodule, iou: %s' % (pid, k, str(v['iou'])))
                    v['count'] = 4
                consensus[v['count']] += 1

            # number of consensus
            num = np.array([len(m) for m in masks])
            num[num > 4] = 4  # 最多4次，超过4次重复标记的，记为4次

            if len(num) == 0:
                continue
            # Iterate from the nodules with most consensus
            for n in range(num.max(), 0, -1):
                mask = np.zeros(img.shape, dtype=np.uint8)

                for i, index in enumerate(np.where(num >= n)[0]):
                    same_nodules = masks[index]
                    m = np.logical_or.reduce(same_nodules)
                    mask[m] = i + 1  # 区分不同的结节，不同的结节给与不同的数值，依次增加（如果是分割，可以直接都给1，或者最后统一处理为1也可以）
                nrrd.write(os.path.join(save_dir, str(n), pid+'.nrrd'), mask)  # mask

    print(consensus)
    print(cnt)

if __name__ == '__main__':
    img_dir = r'./LUNA16/image_combined'        # data

    ctr_arr_save_dir = r'./LUNA16/annotation/noduleCoor'  # 保存每个注释器解析的中间结节mask的地方
    noduleMask_save_dir = r'./LUNA16/nodule_masks'  # 保存合并结节掩码的文件夹

    # 对转储的临时文件，生成mask
    arrs2mask(img_dir, ctr_arr_save_dir, noduleMask_save_dir)

至此，和image一样shape的mask是生成了。下面用itk-snap打开查看处理后的结果，如下所示：

属于分别打开image和mask的nrrd图像，mhd格式的image，转nrrd，可以参考下面的代码：

python 复制代码

nii_path = os.path.join(r'./LUNA16/image_combined', '1.3.6.1.4.1.14519.5.2.1.6279.6001.184412674007117333405073397832.mhd')
image = itk.array_from_image(itk.imread(nii_path))

nrrd.write(r'./image.nrrd', image)

三、总结

lidc-idri的数据集内的数据格式，都是我们不常遇到的数据形式，尤其是mhd文件的raw文件，同时表示一个数据的两个不同部分，也是很少遇到的。

但是对于初学者来说，理解这种数据形式，还是有些陌生，这部分相信通过本系列可以有较深的理解。与此同时，本篇还存储为nrrd文件，这是我比较喜欢的数组存储格式，理解的好理解和简单。

到这里，你就收获了一个新的一一对应关系。这样比你看xml文件，理解起来会简单很多。下一节，我们就对初步得到的image和mask，与肺区分割结合，进一步进行精细化处理。resample操作，调整数据到统一的尺度。