[深度学习]胶质瘤病灶分割技术文档

[深度学习]胶质瘤病灶分割技术文档（不断更新）

参考：

• https://blog.csdn.net/weixin_42338058/article/details/84190420（python中用于读取医学图像的常用库）

• https://imageio.readthedocs.io/en/stable/userapi.html （imageioAPI， 用于nii转化为jpg）

1. 数据预处理

1.1 数据处理

原始数据集为nii格式，常用网络需要jpg格式的输入，故首先需要将数据转化为jpg。python用于处理医疗影像图片的库有两个

• SimpleITK

• nibabel

  SimpleITK 和 Nibabel 的区别：

  • SimpleITK 加载数据是通道在前，如（18，512，512）；

  • Nibabel 是 通道在后，如（512，512，18），其中18是图像通道数，即18张图像，可以把nii看成二维图像，也可以看成三维。

  • nibabel加载出来的图像被旋转了90度，横过来了; SimpleITK加载的图片旋转了180°

选用SimpleITK提取图片

1.2 遇到的问题

• 问题1 scipy.misc.imsave()弃用

  解决方式，scipy.misc.imsave()在高版本被取代改用imageio.imwrite（）

• 问题2转化jpg出现ValueError: Max value == min value, ambiguous given dtype

  解决方式，针对此类异常进行修正

  if mi == ma:    DataSlice = np.ones_like(DataSlice, dtype=np.uint8) * mi
  

• 问题3 报错 sitk::ERROR: Unable to determine ImageIO reader for “…/…/RAWDATA/NEW_MultiCenter_Lesion/MS\HS\HSMS003\2D-T2FLAIR_labelnii”

  解决方式：修改该文件后缀

• 问题4 读取出来的图片的有8种分辨率不一致

  resolution	(256,204)	(256,228)	(256,256)	(320,270)	(320,320)	(512,408)	(512,512)	(640,640)
  number	20	48	1928	521	269	100	2609	80

  解决方式：将8种分辨率置于不同文件夹，在后期决定如何训练。

• 问题5： 读取的label像素值为0-255，需要将有病灶的区域置于1，其它区域置为0.

  解决方式：分3种情况保存

      if sum(sum(ImgArray)) == 0:#全是背景，直接保存imageio.imwrite(SavePath, ImgArray)else:if ImgArray.max() == 1:# 如果图片本身已经是正确标注，直接保存imageio.imwrite(SavePath, ImgArray)else:# label标签为0-255时，设定阈值50ImgArray[ImgArray[:] < 50] = 0ImgArray[ImgArray[:] >= 50] = 1  #将label有病灶置为1，没有病灶置为0imageio.imwrite(SavePath, ImgArray)
  

  

1.3 数据预处理模块代码

数据预处理模块由4部分构成

1. 数据读取

   • 将nii文件转化为jpg格式数据并储存

   • 数据命名格式：CQMS005_0000.jpg 最后4位为切片的维度。data和label名称一致，但放在不同的文件夹

     # coding=utf-8
     import nibabel as nib
     import SimpleITK as sitk
     import os
     import numpy as np
     import matplotlib.pyplot as pltDataRootPath = r'../../RAWDATA/NEW_MultiCenter_Lesion/MS'
     SaveRootPath = r'../../RAWDATA/RAWDATA_jpg'
     # 将MS所有图像转化为jpgdef read_img(path):img = sitk.ReadImage(path)data = sitk.GetArrayFromImage(img)return data# 显示一个系列图
     def show_img(data):for i in range(data.shape[0]):plt.imshow(data[i, :, :],  cmap='gray')print(i)plt.show()def get_shape(DataPath, SavePath, SaveName, FileName):import SimpleITK as sitkimg = sitk.ReadImage(DataPath)DataVolumn = sitk.GetArrayFromImage(img)for channel in range(DataVolumn.shape[0]):SaveNameOutput = SaveName + '_%04d.jpg' % channelOutputPath = os.path.join(SavePath, SaveNameOutput)DataSlice = DataVolumn[channel, :, :]ImageShape = (DataSlice.shape[0], DataSlice.shape[1])with open('%s.txt'%FileName, 'a') as f:f.write(OutputPath +  '    ' +  str(ImageShape) + '\n')def nii2jpg(DataPath, SavePath, SaveName):import SimpleITK as sitkimport imageioimport numpy as npimg = sitk.ReadImage(DataPath)DataVolumn = sitk.GetArrayFromImage(img)for channel in range(DataVolumn.shape[0]):SaveNameOutput = SaveName + '_%04d.jpg'%channelOutputPath = os.path.join(SavePath, SaveNameOutput)print(OutputPath)DataSlice = DataVolumn[channel,:,:]mi = np.nanmin(DataSlice)ma = np.nanmax(DataSlice)if mi == ma:DataSlice = np.ones_like(DataSlice, dtype=np.uint8) * miimageio.imsave(OutputPath, DataSlice)if __name__ == '__main__':PersonNames = os.listdir(DataRootPath)  # iteration_onefor PersonName in PersonNames:  # how many sickerif len(PersonName) >= 8:  # 有一个乱码长度大于8continuePersonNamePath = os.path.join(DataRootPath, PersonName)PersonPhotoTimes = os.listdir(PersonNamePath)  # iteration_twofor PersonPhotoTime in PersonPhotoTimes:if len(PersonPhotoTime) >= 10:  # 避免乱码continuePersonPerTimeImagePath = os.path.join(PersonNamePath, PersonPhotoTime)FileNames = os.listdir(PersonPerTimeImagePath)for FileName in FileNames:if 'nii' not in FileName:# 跳过非数据文件continue# print(FileName)DataPath = os.path.join(PersonPerTimeImagePath, FileName)SaveRootName = PersonPhotoTimeif 'label' not in FileName:SavePath = os.path.join(SaveRootPath, 'data')nii2jpg(DataPath=DataPath, SavePath=SavePath, SaveName=SaveRootName)get_shape(DataPath=DataPath, SavePath=SavePath, SaveName=SaveRootName, FileName='ImageShape_data')else:SavePath = os.path.join(SaveRootPath, 'label')nii2jpg(DataPath=DataPath, SavePath=SavePath, SaveName=SaveRootName)get_shape(DataPath=DataPath, SavePath=SavePath, SaveName=SaveRootName, FileName='ImageShape_label')

2. 检查异常项（主要检查同一data和label维度是否匹配的情况）

   import shutil
   import osAbnormalDataRootPath = r'../../RAWDATA/AbnormalData'with open('ImageShape_data.txt', 'r') as f1:with open('ImageShape_label.txt', 'r') as f2:i = 0for DataLine, LabelLine in zip(f1.readlines(), f2.readlines()):i += 1DataLineSplit = DataLine.split('    ')LabelLineSplit = LabelLine.split('    ')DataLineShape =DataLineSplit[1][:-1]LabelLineShape = LabelLineSplit[1][:-1]print(DataLineShape, LabelLineShape)if DataLineShape != LabelLineShape:shutil.move(DataLineSplit[0], os.path.join(AbnormalDataRootPath, 'data'))shutil.move(LabelLineSplit[0], os.path.join(AbnormalDataRootPath, 'label'))

3. resolution classify 将jpg图片按照不同分辨率置于不同的文件夹

   # import pandas as pd
   import os
   import shutildef main():ImageShapeClass = []with open('ImageShape_data.txt', 'r') as f1:with open('ImageShape_label.txt', 'r') as f2:i = 0for DataLine, LabelLine in zip(f1.readlines(), f2.readlines()):i += 1DataLineSplit = DataLine.split('    ')LabelLineSplit = LabelLine.split('    ')DataLineShape = DataLineSplit[1][:-1]LabelLineShape = LabelLineSplit[1][:-1]DataLinePath = os.path.split(DataLineSplit[0])[0]LabelLinePath = os.path.split(LabelLineSplit[0])[0]FileName = os.path.split(DataLineSplit[0])[1]if LabelLineShape != DataLineShape:  # if data's shape != label's shape skipcontinueelse:# create file path for different resolution imageif LabelLineShape not in ImageShapeClass:ImageShapeClass.append(LabelLineShape)NewPathFile = 'Resolution_' + str(LabelLineShape)DataNewPath = os.path.join(DataLinePath, NewPathFile)LabelNewPath = os.path.join(LabelLinePath, NewPathFile)DataIsExists = os.path.exists(DataNewPath)LabelIsExists = os.path.exists(LabelNewPath)if not DataIsExists:os.mkdir(DataNewPath)if not LabelIsExists:os.mkdir(LabelNewPath)DataMovePath = os.path.join(DataLinePath, 'Resolution_' + str(LabelLineShape))LabelMovePath = os.path.join(LabelLinePath, 'Resolution_' + str(LabelLineShape))DataRawPath = DataLineSplit[0]LabelRawPath = LabelLineSplit[0]shutil.move(DataRawPath, DataMovePath)shutil.move(LabelRawPath, LabelMovePath)if __name__ == '__main__':main()

4. 将jpg格式的label转化为baseline的png格式，并对label的像素值进行处理，有病灶的区域置为1，其他区域置为0。

   import imageio
   import osImageRootPath = r'..\..\RAWDATA\RAWDATA_jpg\label\Resolution_(640, 640)'
   SaveRootPath = r'..\..\RAWDATA\RAWDATA_jpg\label\Resolution_(640, 640)\label_png'def jpg2png(ImagePath, SavePath):ImgArray = imageio.imread(ImagePath)if sum(sum(ImgArray)) == 0:#全是背景，直接保存imageio.imwrite(SavePath, ImgArray)else:if ImgArray.max() == 1:# 如果图片本身已经是正确标注，直接保存imageio.imwrite(SavePath, ImgArray)else:# label标签为0-255时，设定阈值50ImgArray[ImgArray[:] < 50] = 0 #将label有病灶置为1，没有病灶置为0ImgArray[ImgArray[:] >= 50] = 1imageio.imwrite(SavePath, ImgArray)if __name__ == '__main__':ImageNames = os.listdir(ImageRootPath)print('{:_^50}'.format('start transform'))for ImageName in ImageNames:if 'jpg' not in ImageName:continueImagePath = os.path.join(ImageRootPath, ImageName)ImageNameNew = ImageName.split('.')[0] + '.png' # png格式命名SavePath = os.path.join(SaveRootPath, ImageNameNew)jpg2png(ImagePath=ImagePath, SavePath=SavePath)print('{:_^4}'.format('Successful!'))

2 模型训练

2.1 baseline

采用MIT开源全卷积神经网络（FCN）进行baseline训练。程序已跑通，下周使用公司GPU完整运行。

MIT开源全卷积神经网络（FCN）源码地址：

• https://github.com/shekkizh/FCN.tensorflow

[外链图片转存失败(img-yprT6Xr3-1563698300993)(C:\Users\eveadam\AppData\Roaming\Typora\typora-user-images\1563697418017.png)]

未完待续。。。

本文来自互联网用户投稿，文章观点仅代表作者本人，不代表本站立场，不承担相关法律责任。如若转载，请注明出处。 如若内容造成侵权/违法违规/事实不符，请点击【内容举报】进行投诉反馈！