医学影像数据集DICOM文件(影像数据+轮廓数据)转为标注可识别 python代码

1.dcm转jpg

def convert_from_dicom_to_jpg(img, low_window, high_window, save_path):
    lungwin = np.array([low_window * 1., high_window * 1.])
    newimg = (img - lungwin[0]) / (lungwin[1] - lungwin[0])
    newimg = (newimg * 255).astype('uint8')
    cv2.imwrite(save_path, newimg, [int(cv2.IMWRITE_JPEG_QUALITY), 100])


def dcm2jpg():
    count = 1
    path = r"C:\data\dcm"
    filename = os.listdir(path)

    for i in filename:
        document = os.path.join(path, i)
        outputpath = r"C:\jpg"

        index = i.rfind('.')
        countname = i[:index]
        countfullname = countname + '.jpg'
        output_jpg_path = os.path.join(outputpath, countfullname)

        ds_array = sitk.ReadImage(document)
        img_array = sitk.GetArrayFromImage(ds_array)
        shape = img_array.shape  # name.shape
        img_array = np.reshape(img_array, (shape[1], shape[2]))
        high = np.max(img_array)
        low = np.min(img_array)
        convert_from_dicom_to_jpg(img_array, low, high, output_jpg_path)
        print('FINISHED')

        count = count + 1

当时找了好多个代码，感觉转换出来都没那么亮，就随便贴一个吧。

2.dcom轮廓数据转标注可识别

这里遇到的问题是 轮廓数据里基本找到属性为：(1)引用到原数据的ruid，(2)ROIName，(3)世界坐标，然后根据这些数据 转换坐标，然后 对每一个原图找到它用到的ROI，加到字典里，后续新建json存入。

dcm后缀 数据+轮廓数据 转为eiseg标注可识别工具代码。

可根据代码内函数自定义改为labelme或者其他可识别的标签文件。

---

做记录使用，没有做太多详细注释，当时debug然后根据属性获取 转换的。希望对大家有帮助。

import os
import numpy as np
import pydicom
from os.path import splitext, basename
import random
import json

uid2file = {}
file2uid = {}

# 世界坐标转图像坐标
def worldToVoxelCoord(worldCoord, origin, spacing):
    stretchedVoxelCoord = np.absolute(worldCoord - origin)
    voxelCoord = stretchedVoxelCoord / spacing
    return voxelCoord

# 
def dcm2coord():
    data = dict()
    dcms_data = dict()
    data_dir = r'C:\data'
    dcm_dir = os.path.join(data_dir, 'dcm')
    rt_dir = os.path.join(data_dir, 'rt')
    for dcm_file in os.listdir(dcm_dir):  # 遍历读取数据
        dcm_file_path = os.path.join(dcm_dir, dcm_file)
        dcm_name = splitext(basename(dcm_file_path))[0]
        ds = pydicom.dcmread(dcm_file_path)
        array = ds.pixel_array
        origin = ds.ImagePositionPatient  # 网格原点在世界坐标系的位置
        spacing = ds.PixelSpacing  # 采样间隔
        uid = ds.SOPInstanceUID
        uid2file[uid] = dcm_name
        file2uid[dcm_name] = uid
        dcms_data[uid] = {'dcmSpacing': spacing, 'dcmOrigin': origin, 'array': array}
    data['dcms_data'] = dcms_data  # 存放图像信息，后续根据该信息获取 ruid对应关系

    # rt data
    rt_data = dict()  # 以{RUID:label_data}形式返回结果
    rt_file_path = os.path.join(rt_dir, os.listdir(rt_dir)[0])

    ds = pydicom.dcmread(rt_file_path)
    seq_lists = []
    ROI_name_lists = []
    for i in range(0, len(ds.ROIContourSequence)):
        sequences = ds.ROIContourSequence[i].ContourSequence
        ROI_name = ds.ROIContourSequence[i].ReferencedROINumber
        seq_lists.append(sequences)
        ROI_name_lists.append(ROI_name)  # seq name 顺序一一对应
    # 0是 RIO31, 有38个对应image
    all_rt_data = []
    for i, sequences in enumerate(seq_lists):
        tmp_rt_datas = dict()
        tmp_rt_data = dict()
        ruids = []
        dup = 1
        for sequence in sequences:
            ruid = sequence.ContourImageSequence[0].ReferencedSOPInstanceUID
            array = sequence.ContourData
            num = sequence.NumberOfContourPoints
            if ruid in ruids:
                dep_ruid = ruid + '-' + str(dup)
                rt_data[dep_ruid] = {'ROIname': ROI_name_lists[i], 'pointNumber': num, 'array': array}
                tmp_rt_data[dep_ruid] = {'ROIname': ROI_name_lists[i], 'pointNumber': num, 'array': array}
                dup += 1
                continue  # 跳出本次循环进入下一循环
            rt_data[ruid] = {'ROIname': ROI_name_lists[i], 'pointNumber': num, 'array': array}
            tmp_rt_data[ruid] = {'ROIname': ROI_name_lists[i], 'pointNumber': num, 'array': array}
            ruids.append(ruid)
        data['rt_data'] = rt_data  # 得到与之对应的image的ruid和 轮廓点坐标  会覆盖
        tmp_rt_datas['rt_data'] = tmp_rt_data
        all_rt_data.append(tmp_rt_datas['rt_data'])  # 共12个元素
        '''
        00  ROI31   元素分别为 对应到CT Image的 ruid 和 点坐标
        01  ROI1
        02  ROI3
        03  ROI4
        04  ROI11
        05  ROI12
        06  ROI13
        07  ROI14
        08  ROI15
        09  ROI16
        10  ROI17
        11  ROI46
        '''

    ROI_cood_dic = {}
    for index, ROI in enumerate(all_rt_data):  # 遍历所有的ROI 即每个ROI的 rt_data
        point_data = {}
        for uid, value in ROI.items():  # 遍历当前ROI的 对应CT Image的ruid和 轮廓坐标
            num = value['pointNumber']
            ROI_name = value['ROIname']
            true_uid = uid.split('-')[0] if uid.rfind('-') != -1 else uid
            label_data = value['array']
            dcm_origin = data['dcms_data'][true_uid]['dcmOrigin']
            dcm_spacing = data['dcms_data'][true_uid]['dcmSpacing']

            point = []  # 坐标[(x1,y1),(...),...]
            for i in range(0, len(label_data), 3):
                x = label_data[i]  # 轮廓世界坐标系
                y = label_data[i + 1]
                X = worldToVoxelCoord(float(x), float(dcm_origin[0]), float(dcm_spacing[0]))
                Y = worldToVoxelCoord(float(y), float(dcm_origin[1]), float(dcm_spacing[1]))
                point.append([X, Y])
            point_data[uid] = point
        ROI_cood_dic[ROI_name_lists[index]] = point_data

    return ROI_cood_dic


# 由原本的 ROI 对应 Image  转换为 Image 下所有ROI轮廓
def ROI_cood_to_img_coord(ROI_cood_dic):
    '''
    对某一个img名字，把旗下所有的ROI标签拿到，对应ROI标签坐标拿到。
    Args:
        ROI_cood_dic:

    Returns:

    '''
    img_dir = r'C:\data\jpg'
    over_lists = {}
    for img in os.listdir(img_dir):  # 遍历img

        img_name = splitext(img)[0]
        img2uid = file2uid[img_name]  # 该图片对应的ruid
        labels_lists = []
        for k, v in ROI_cood_dic.items():  # 遍历12个ROI
            ROIname = k
            uid_coords = v
            color = [random.randint(16, 255), random.randint(32, 255), random.randint(64, 255)]
            for uid, vals in uid_coords.items():  # 遍历每个ROI对应的image
                labels = dict()
                if uid.find(img2uid) != -1:  # 当前图片 有这个label
                    labels["name"] = ROIname
                    labels["labelIdx"] = int(ROIname)
                    labels["color"] = color
                    labels["points"] = vals
                if labels != {}:
                    labels_lists.append(labels)
        over_lists[img] = labels_lists
    return over_lists


def toeiseg(over_lists):
    '''
    字典中存的 key是图片名 value是 不同标签的 dict。 将这些个label遍历 然后依次添加到对应的list中。
    Args:
        over_lists:

    Returns:

    '''
    desc_dir = r'C:\data\labels'
    for img, lists in over_lists.items():
        img_name = splitext(img)[0]
        img2json = img_name + '.json'
        desc_path = os.path.join(desc_dir, img2json)
        if not desc_path:
            os.mkdir(desc_path)
        with open(desc_path, 'w') as file_obj:
            json.dump(lists, file_obj, indent=2)


if __name__ == "__main__":
    ROI_cood_dic = dcm2coord()
    over_lists= ROI_cood_to_img_coord(ROI_cood_dic)
    toeiseg(over_lists)