医学图像预处理----①小白入门

• 1、dicom格式转Niii。
• 2、显示NII格式代码用于整理数据集。（窗体窗位显示的更好）
• 3、制作表格显示主要信息，便于数据集的使用（多套数据集）
• 4、对数据集进行切片处理减少干扰（可利用ITK-SNAP查看）
• 5、重采样得到训练所需体素大小，归一化处理

1、dicom格式转Niii。

这一块比较简单，直接上代码了

import numpy as npimport shutilimport osimport SimpleITK as sitkdef dcm2nii_sitk(path_read, path_save):    reader = sitk.ImageSeriesReader()    seriesIDs = reader.GetGDCMSeriesIDs(path_read)    N = len(seriesIDs)    lens = np.zeros([N])    for i in range(N):        dicom_names = reader.GetGDCMSeriesFileNames(path_read, seriesIDs[i])        lens[i] = len(dicom_names)    N_MAX = np.argmax(lens)    dicom_names = reader.GetGDCMSeriesFileNames(path_read, seriesIDs[N_MAX])    reader.SetFileNames(dicom_names)    image = reader.Execute()    if not os.path.exists(path_save):        os.mkdir(path_save)    sitk.WriteImage(image, path_save+'/data.nii.gz')DICOMpath = r"DATA_MD/Changhai_Int_anonymize/dc4f80590509403da78a763bd7f64baa/0a324d770c46429b9e2e08fee5b88761"   #dicom文件夹路径Midpath = r"SavePng"   #处理中间数据路径Resultpath = r"SaveRaw"    #保存路径cases = os.listdir(DICOMpath)  #获取dicom文件夹路径子文件夹名for c in cases:  #遍历dicom文件夹路径子文件    path_mid = os.path.join(DICOMpath , c)  #获取dicom文件夹下每一套数据的路径    dcm2nii_sitk(path_mid , Midpath )  #将dicom转换为nii，并保存在Midpath中    shutil.copy(os.path.join(Midpath , "data.nii.gz"), os.path.join(Resultpath , c + ".nii.gz"))    #重新对保存后的nii文件名进行命名，并复制到Resultpath下

2、显示NII格式代码用于整理数据集。（窗体窗位显示的更好）

下面展示一些 第一次显示窗位的代码。

import torchio as tioimport numpy as npimport SimpleITK as sitkimport osimport nibabel as nibfrom matplotlib import pyplot as pltold_path="put/P_06"cases = os.listdir(old_path)i=0def window(img):    win_min = -300    win_max = 350    for i in range(img.shape[0]):        img[i] = 255.0 * (img[i] - win_min) / (win_max - win_min)        min_index = img[i] < 0        img[i][min_index] = 0        max_index = img[i] > 255        img[i][max_index] = 255        img[i] = img[i] - img[i].min()        if img[i].max()!=0 :           c = float(255) / img[i].max()           img[i] = img[i] * c    return img.astype(np.uint8)for c in cases:  #遍历p文件子文件    i = i + 1    path_mid = os.path.join(old_path ,c)  #获取dicom文件夹下每一套数据的路径    img = sitk.ReadImage(path_mid)    img = sitk.GetArrayFromImage(img)    img = window(img)    out = sitk.GetImageFromArray(img)    sitk.WriteImage(out, '{0}.nii.gz'.format(i))    t1_img = tio.ScalarImage('{0}.nii.gz'.format(i))    t1_img.plot()    img = nib.load(path_mid)    img_hdr = img.header    voxel = [img_hdr['qoffset_x']]    voxel.append(img_hdr['qoffset_y'])    voxel.append(img_hdr['qoffset_z'])    print("----{0}的头文件信息是,坐标信息是------n{1}n{2}".format(c,img,voxel))print(i)

下面展示一些 优化显示窗位的代码。在调用函数之前写好w_width与w_center

'''调窗'''def adjustMethod1(data_resampled,w_width,w_center):    val_min = w_center - (w_width / 2)    val_max = w_center + (w_width / 2)    data_adjusted = data_resampled.copy()    data_adjusted[data_resampled < val_min] = val_min    data_adjusted[data_resampled > val_max] = val_max    return data_adjusted

3、制作表格显示主要信息，便于数据集的使用（多套数据集）

4、对数据集进行切片处理减少干扰（可利用ITK-SNAP查看）

下面展示一些 遍历三维图像找到边界值。

import matplotlib.pyplot as pltimport osimport SimpleITK as sitkimport cv2old_path="DATA_MD_01/p_18"cases = os.listdir(old_path)ori_data=sitk.ReadImage(os.path.join(old_path,cases[0]))pic_Size=ori_data.GetSize()data=sitk.GetArrayFromImage(ori_data)y_list = []x_list = []def cv_show(img, name):    cv2.imshow(name, img)    cv2.waitKey()    cv2.destroyAllWindows()def norm_img(image):  # 归一化像素值到（0，1）之间，且将溢出值取边界值    image = (image - MIN_BOUND) / (MAX_BOUND - MIN_BOUND)    image[image > 1] = 1.    image[image < 0] = 0.    return imagefor i in range(0,pic_Size[2]):    image=data[i,:,:]    MIN_BOUND = -1000.0    MAX_BOUND = 700.0    image=norm_img(image)    plt.imshow(image)    plt.show()    draw_img = image.copy()    ret, thresh = cv2.threshold(draw_img, 0.2, 255, cv2.THRESH_BINARY)    for x in range(0,512):        for y in range(0,512):            if thresh[x][y]==255:                y_list.append(x)                break    for y in range(0,512):        for x in range(0,512):            if thresh[x][y]==255:                x_list.append(y)                break    x_list.sort()    y_list.sort()    x_max_list=x_list[len(x_list)-1]    y_max_list=y_list[len(y_list)-1]    x_min=x_list[0]    y_min=y_list[0]print(x_min,x_max_list)print(y_min,y_max_list)

下面展示一些 通过边界值找到切分边界对多余信息切除。

import torchio as tioimport numpy as npimport SimpleITK as sitkimport osdef copy_geometry(image: sitk.Image, ref: sitk.Image):    image.SetOrigin(ref.GetOrigin())    image.SetDirection(ref.GetDirection())    image.SetSpacing(ref.GetSpacing())    return imageold_path="DATA_MD_01/p_03"cases = os.listdir(old_path)print(cases)i=-1for c in cases:    i=i+1    # print(c)    path_img=os.path.join(old_path,c)    img=sitk.ReadImage(path_img)    # img = sitk.GetArrayFromImage(image)    # img = sitk.GetImageFromArray(img)    # img_out_itk = copy_geometry(img, image)    direction = img.GetDirection()    pic_Origin=img.GetOrigin()    pic_Spacing=img.GetSpacing()    pic_Size=img.GetSize()    print("原图位置：{}n原图体素：{}n原图大小{}".format(pic_Origin,pic_Spacing,pic_Size))    data_clip = img[10:510,60:390,:]    data_clip.SetSpacing(pic_Spacing)    data_clip.SetOrigin(pic_Origin)    data_clip.SetDirection(direction)    # print(data_clip)    Resultpath = r"ab"  # 保存路径    sitk.WriteImage(data_clip, os.path.join(Resultpath, c + ".gz"))    t1_img = tio.ScalarImage(os.path.join(Resultpath, c + ".gz"))    t1_img.plot()    print("裁剪后位置：{}n裁剪后体素：{}n裁剪后大小{}".format(pic_Origin,pic_Spacing,pic_Size))

5、重采样得到训练所需体素大小，归一化处理

下面展示一些 遍历切除后生成的.nii格式文件进行重采样并且通过填充的形式进行调节大小格式一致。
对喽，这个地方的重采样貌似有好多方法，我自己就尝试了三种方法，把感觉最好用的拿出来分享一下

import numpy as npimport osimport SimpleITK as sitkimport torchio as tio'''调窗'''def adjustMethod1(data_resampled,w_width,w_center):    val_min = w_center - (w_width / 2)    val_max = w_center + (w_width / 2)    data_adjusted = data_resampled.copy()    data_adjusted[data_resampled < val_min] = val_min    data_adjusted[data_resampled > val_max] = val_max    return data_adjusted'''处理-1000——700数据'''def pretreatmentImage(image):    image[image < -1000] = -1000    image[image > 700] = 700    return image'''归一化操作'''def MinMax(image):    if np.max(image) - np.min(image) != 0:        image =(image - np.min(image)) / (np.max(image) - np.min(image))  # 归一化    return image'''重采样'''def img_resmaple(ori_data,new_spacing=[2.5, 2.5, 2.5]):    original_spacing = ori_data.GetSpacing()    original_size = ori_data.GetSize()    transform=sitk.Transform()    transform.SetIdentity()    new_shape = [        int(np.round(original_spacing[0] * original_size[0] / new_spacing[0])),        int(np.round(original_spacing[1] * original_size[1] / new_spacing[1])),        int(np.round(original_spacing[2] * original_size[2] / new_spacing[2])),    ]    # print("新的形状大小为",new_shape)    resmaple = sitk.ResampleImageFilter()#初始化 #初始化一个图像重采样滤波器resampler    resmaple.SetInterpolator(sitk.sitkLinear)    resmaple.SetTransform(transform)    resmaple.SetDefaultPixelValue(0)    resmaple.SetOutputSpacing(new_spacing)    resmaple.SetOutputOrigin(ori_data.GetOrigin())    resmaple.SetOutputDirection(ori_data.GetDirection())    resmaple.SetSize(new_shape)    data = resmaple.Execute(ori_data)    return datadef copy_geometry(image: sitk.Image, ref: sitk.Image):    image.SetOrigin(ref.GetOrigin())    image.SetDirection(ref.GetDirection())    image.SetSpacing(ref.GetSpacing())    return imageif __name__ == "__main__":    '''读取数据'''    old_path = "ab"    cases = os.listdir(old_path)    for c in cases:        ori_data = sitk.ReadImage(os.path.join(old_path, c))        data_res = sitk.GetArrayFromImage(ori_data)        img_pre = pretreatmentImage(data_res)        img_Min = MinMax(img_pre)        img = sitk.GetImageFromArray(img_Min)        img_out_itk = copy_geometry(img, ori_data)        data_pre=img_resmaple(img_out_itk)        data=sitk.GetArrayFromImage(data_pre)        # print(data.shape)        if data.shape[2]<200 or data.shape[1]<160 or data.shape[0]<168:            # print("我进入for循环了")            x_up=int((200-data.shape[2])/2)            if data.shape[2]%2!=0:                x_down=x_up+1            else:                x_down=x_up            y_left=int((160-data.shape[1])/2)            if data.shape[1]%2!=0:                y_right=y_left+1            else:                y_right=x_up            z_front= int((168 - data.shape[0]) / 2)            if data.shape[0] % 2 != 0:                z_after = z_front + 1            else:                z_after = z_front            data_pad=np.pad(data,((z_front,z_after),(y_left,y_right),(x_up,x_down)),'constant', constant_values=(0,0))        # print('------------------------',data_pad.shape)        Resultpath = r"SaveRaw"  # 保存路径        img = sitk.GetImageFromArray(data_pad)        img_out_itk = copy_geometry(img, data_pre)        sitk.WriteImage(img_out_itk, os.path.join(Resultpath,c))        '''验证存储信息是否正确'''        ori = sitk.ReadImage(os.path.join(Resultpath,c ))        pic_Origin=ori.GetOrigin()        pic_Spacing=ori.GetSpacing()        pic_Size=ori.GetSize()        print("--------{}文件预处理重采样后的信息------".format(c))        print("裁剪后位置：{}n裁剪后体素：{}n裁剪后大小{}".format(pic_Origin, pic_Spacing, pic_Size))        img=sitk.GetArrayFromImage(ori)        # print("数组信息为",img)        t1_img = tio.ScalarImage(os.path.join(Resultpath, c))        t1_img.plot()

附加另一种重采样方法zoom或者resize；具体实现自己搞一下很简单

class Resize_img(Base):    def __init__(self, shape):        self.shape = shape    def tf(self, img, k=0):        if k == 1:            img = resize(img, (img.shape[0], self.shape[0], self.shape[1], self.shape[2]),                         anti_aliasing=False, order=0)        else:            img = resize(img, (img.shape[0], self.shape[0], self.shape[1], self.shape[2]),                         anti_aliasing=False, order=3)        return img

class Resize_img(Base):    def __init__(self, shape):        self.shape = shape    def tf(self, img, k=0):        if img.shape[-1] == 3:            img = zoom(img, (1, (self.shape[0] / img.shape[1]), (self.shape[1]/img.shape[2]), 1), order=3)        else:            img = zoom(img, (1, (self.shape[0] / img.shape[1]), (self.shape[1] / img.shape[2])), order=3)        #resize(img, (img.shape[0], self.shape[0], self.shape[1]), anti_aliasing=False, order=3)        return img

良心点我也给大家附上吧

import numpy as npimport osimport SimpleITK as sitkfrom scipy.ndimage import zoomimport collectionsfrom skimage.transform import rescale, resizefrom scipy import ndimage, miscimport torchio as tio'''处理-1000——700数据'''def pretreatmentImage(image):    image[image < -1000] = -1000    image[image > 700] = 700    return image'''归一化操作'''def MinMax(image):    if np.max(image) - np.min(image) != 0:        image = (image - np.min(image)) / (np.max(image) - np.min(image))  # 归一化    return imagedef adjustMethod1(data_resampled,w_width,w_center):    val_min = w_center - (w_width / 2)    val_max = w_center + (w_width / 2)    data_adjusted = data_resampled.copy()    data_adjusted[data_resampled < val_min] = val_min    data_adjusted[data_resampled > val_max] = val_max    return data_adjusted'''重采样'''def resize_01(img):    original_spacing = ori_data.GetSpacing()    img =zoom(img,(original_spacing[2]/2.5, original_spacing[1]/2.5, original_spacing[0]/2.5),order=3)    print(img.shape)    return imgdef copy_geometry(image: sitk.Image, ref: sitk.Image):    image.SetOrigin(ref.GetOrigin())    image.SetDirection(ref.GetDirection())    image.SetSpacing([2.5,2.5,2.5])    return imageif __name__ == "__main__":    w_width = 400    w_center = 40    old_path = "ab"    cases = os.listdir(old_path)    ori_data = sitk.ReadImage(os.path.join(old_path, cases[0]))    data = sitk.GetArrayFromImage(ori_data)    img_pre = pretreatmentImage(data)    img_Min = MinMax(img_pre)    img_re=resize_01(img_Min)    img = sitk.GetImageFromArray(img_re)    img_out_itk = copy_geometry(img,ori_data)    Resultpath = r"SaveRaw"  # 保存路径    sitk.WriteImage(img_out_itk, os.path.join(Resultpath,cases[0]))    '''验证存储信息是否正确'''    ori = sitk.ReadImage(os.path.join(Resultpath, cases[0]))    print('保存的图像信息', ori)    img = sitk.GetArrayFromImage(ori)    # print("数组信息为",img)    image = adjustMethod1(img, w_width, w_center)    t1_img = tio.ScalarImage(os.path.join(Resultpath, cases[0]))    t1_img.plot()

总结一下子，老师是真的牛我是真的菜，一会说一下我检索的来源与出处

如果感兴趣来自哪里就搜一下吧

• 我没用到这个，但是他给我提供了下处理流程的思路： 医学nii图像 预处理——图像裁剪 重采样 灰度区域 归一化 修改图像尺寸
• 我主要用这个调的窗体窗位Python CT图像预处理——nii格式读取、重采样、窗宽窗位设置
• 这是我模仿的这个哥哥的整体思路 CT医学图像的预处理（重采样）
• 这个是主要做输出用0填充背景让大小一致的numpy.pad()函数使用详解
  重采样方法之二ZOOM（当时没看明白多搜了几个自己看看吧）
• scipy.ndimage.zoom矩阵放缩
• Python scipy.ndimage.zoom用法及代码示例
• Python scipy ndimage.zoom()函数
• CT图像预处理之重采样

行了，要是有不对的记得留言告诉我毕竟我也是刚学，那我走了，时间2023年5月8日，当然我会跟着项目尽量到训练模型跑出整个项目出来啊。毕竟论文还是要写的