使用数据集：LIDC
代码作用：从xml文件中获取结节坐标并生成mask
备注：没有学过python，都是边写边查，有的地方可能看起来不简洁，请自行优化
21-05-13 Debug 在我训练完Unet之后发现它笨的要死，于是倒回发现数据集有问题！这里有bug。因为不能保证slices和nodule_pos在list中的位置是一一对应的，所以肺实质和肺结节图片会在一个list范围内乱序
效果：
requirements
import os
import SimpleITK
import pydicom
import numpy as np
import glob
from tqdm import tqdm
import bs4
from PIL import Image
from scipy import ndimage as ndi
from skimage.segmentation import clear_border
from skimage.measure import label, regionprops
from skimage.morphology import disk, dilation, binary_erosion, binary_closing
from skimage.filters import roberts, sobel
import cv2
import xml
functions
dataset_size = 300000 # 生成的数据集大小
def is_dicom_file(filename):
'''
判断某文件是否是dicom格式的文件
:param filename: dicom文件的路径
:return:
'''
file_stream = open(filename, 'rb')
file_stream.seek(128)
data = file_stream.read(4)
file_stream.close()
if data == b'DICM':
return True
return False
def load_patient(dir_list,type):
'''
读取某文件夹内的所有dicom文件
:param src_dir: dicom文件夹路径,，！最终文件夹
:return: dicom list
'''
uids=[]
# print(dir_list)
if (type == "nodule"):
uids,nodule_pos = find_nodule_frame(dir_list)
if (uids == None):
return None, None
slices = [0] * len(uids)
file_list = os.listdir(dir_list)
for file in file_list:
if file.endswith("xml"):
continue
file_path = os.path.join(dir_list, file)
ds = pydicom.dcmread(file_path)
#print(ds.SOPInstanceUID)
if (ds.SOPInstanceUID in uids):
#print(file)# 这里的0x0008,0x0018是SOP_Instance_UID
#slices.append(pydicom.read_file(file_path))
slices[uids.index(ds.SOPInstanceUID)]=pydicom.read_file(file_path)
#slices.insert(uids.index(ds.SOPInstanceUID),pydicom.read_file(file_path))
#slices.remove(0)
# print(count)
if (type == "non-nodule"):
uids,nodule_pos = find_nodule_frame(dir_list)
if (uids == None):
return None
file_list = os.listdir(dir_list)
num = 0
around_nodule = 5
for file in file_list:
if file.endswith("xml"):
continue
file_path = os.path.join(dir_list, file)
ds = pydicom.dcmread(file_path)
print(ds.SOPInstanceUID.value)
if (ds.SOPInstanceUID in uids): # 这里的0x0008,0x0018是SOP_Instance_UID
for i in range(around_nodule):
if ((num + i) >= len(file_list) or (num - i) <= 0):
break
file_path = os.path.join(dir_list, file_list[num + i])
ds = pydicom.dcmread(file_path)
if (ds.SOPInstanceUID not in uids):
slices.append(pydicom.read_file(file_path))
else:
continue
file_path = os.path.join(dir_list, file_list[num - i])
ds = pydicom.dcmread(file_path)
if (ds.SOPInstanceUID not in uids):
slices.append(pydicom.read_file(file_path))
else:
continue
else:
continue
num += 1
ol=slices.count(0)
lop=len(slices)
if (slices.count(0) == len(slices)):
return None,None
#莫名其妙！有些CT在xml文件中有但是在dcm里面找不到，可能是我没下完全叭，梯子太抖了
#所以这里做一个筛选
real_slices=[]
real_nodule_pos=[]
for i in range(len(uids) - 1):
if(slices[i] is 0):
continue
else:
real_slices.append(slices[i])
real_nodule_pos.append(nodule_pos[i])
for i in range(len(real_slices)-1):
if(uids[i]!=real_slices[i].SOPInstanceUID):
print("not correct")
print(i)
#os._exit(0)
return None,None
return real_slices,real_nodule_pos
def get_pixels_hu_by_simpleitk(slices):
'''
读取某文件夹内的所有dicom文件,并提取像素值(-4000 ~ 4000)
:param src_dir: dicom文件夹路径
:return: image array
'''
# reader = SimpleITK.ImageSeriesReader()
# dicom_names = reader.GetGDCMSeriesFileNames(dicom_dir)
# reader.SetFileNames(dicom_names)
# image = reader.Execute()
# img_array = SimpleITK.GetArrayFromImage(image)
# img_array[img_array == -2000] = 0
# return img_array
img_array = []
for sli in slices:
img_array.append(np.array(sli.pixel_array))
img_array = np.array(img_array)
img_array[img_array == -2000] = 0
img_array = img_array - 1024 #################################全场最佳！ -1024 不减图片就全是黑的！
return np.array(img_array)
def normalize_hu(image):
'''
将输入图像的像素值(-4000 ~ 4000)归一化到0~1之间
:param image 输入的图像数组
:return: 归一化处理后的图像数组
'''
MIN_BOUND = -1000
MAX_BOUND = 400
image = (image - MIN_BOUND) / (MAX_BOUND - MIN_BOUND)
image[image > 1] = 1.
image[image < 0] = 0.
return image
def load_patient_images(src_dir, wildcard="*.*", exclude_wildcards=[]):
'''
读取一个病例的所有png图像，返回值为一个三维图像数组
:param image 输入的一系列png图像
:return: 三维图像数组
'''
src_img_paths = glob.glob(src_dir + wildcard)
for exclude_wildcard in exclude_wildcards:
exclude_img_paths = glob.glob(src_dir + exclude_wildcard)
src_img_paths = [im for im in src_img_paths if im not in exclude_img_paths]
src_img_paths.sort()
images = [cv2.imread(img_path, cv2.IMREAD_GRAYSCALE) for img_path in src_img_paths]
images = [im.reshape((1,) + im.shape) for im in images]
res = np.vstack(images)
return res
def get_segmented_lungs(im):
'''
对输入的图像进行肺部区域分割，提取有效的肺部区域，用于模型训练
:param 输入的图像
:return: 返回分割结果
'''
binary = im < -400 # Step 1: 转换为二值化图像
cleared = clear_border(binary) # Step 2: 清除图像边界的小块区域
label_image = label(cleared) # Step 3: 分割图像
areas = [r.area for r in regionprops(label_image)] # Step 4: 保留2个最大的连通区域
areas.sort()
if len(areas) > 2:
for region in regionprops(label_image):
if region.area < areas[-2]:
for coordinates in region.coords:
label_image[coordinates[0], coordinates[1]] = 0
binary = label_image > 0
selem = disk(2) # Step 5: 图像腐蚀操作,将结节与血管剥离
binary = binary_erosion(binary, selem)
selem = disk(10) # Step 6: 图像闭环操作,保留贴近肺壁的结节
binary = binary_closing(binary, selem)
edges = roberts(binary) # Step 7: 进一步将肺区残余小孔区域填充
binary = ndi.binary_fill_holes(edges)
get_high_vals = binary == 0 # Step 8: 将二值化图像叠加到输入图像上
im[get_high_vals] = -2000
print('lung segmentation complete.')
return im, binary
# 找出所有有结节的frame的UID
def find_nodule_frame(folder_path):
'''
folder_path:最终存放dcm文件的文件夹路径
--folder_path
-xx.dcm
-xxx.dcm
...
'''
file_list = os.listdir(folder_path)
xml_path=folder_path
for file in file_list:
if file.endswith("xml"):
file_path = os.path.join(folder_path, file)
xml_path = file_path
if(xml_path==folder_path):
return None,None
uids = []
nodule_pos = []
with open(xml_path, 'r') as xml_file:
markup = xml_file.read()
xml = bs4.BeautifulSoup(markup, features="xml")
reading_session = xml.LidcReadMessage.find_all("readingSession")
for readingSession in reading_session:
nodule_info = readingSession.find_all("unblindedReadNodule")
for nodule_info in nodule_info:
nodule_roi = nodule_info.find_all("roi")
for nodule_roi in nodule_roi:
edge_map=nodule_roi.find_all("edgeMap")
one_nodule=[]
frame_uid = nodule_roi.find_all("imageSOP_UID")
if(frame_uid[0].string in uids): # 因为有多个医生的诊断信息 所以会出现重复UID
continue
else:
uids.append(frame_uid[0].string)
for edge in edge_map:
one_nodule.append([edge.find("xCoord").text, edge.find("yCoord").text])
nodule_pos.append(one_nodule)
#nodule_pos里面装的是edgeMap，与uid是一一对应的
return uids,nodule_pos
main
if __name__ == '__main__':
dicom_dir=r"H:\manifest-1600709154662\LIDC-IDRI\rest"
#dicom_dir=r"./dicom_demo"
count=7631
dir_list = []
for root, dirs, files in os.walk(dicom_dir):
# 10是根据数据集特点写的，因为一个病患有两种ct文件，一个是胸部横截面，一个是正面CT，正面的只有两张
if (len(files) > 10 and dataset_size > 0):
dataset_size -= len(files)
dir_list.append(root)
#dir_list=[r'H:\manifest-1600709154662\LIDC-IDRI\LIDC-IDRI-0659\01-01-2000-29838\3000020.000000-74972']
for file_dir in dir_list:
print(file_dir)
slices,edgeMap = load_patient(file_dir,"nodule")
#print(len(slices))
#print(len(edgeMap))
if(slices==None ):
continue
# 提取dicom文件中的像素值
image = get_pixels_hu_by_simpleitk(slices)
# print("")
for i in tqdm(range(image.shape[0])):
if(len(edgeMap[i])<10): #控制结点大小
continue
lung_mask_path="./dataset3.0/lung_masks/lungMask_"+ str(count) + ".png"
lung_parenchyma_path="./dataset3.0/lung_parenchyma/lungparen_"+ str(count) + ".png"
lung_image_path="./dataset3.0/lung_images/lungImg_"+ str(count) + ".png"
lung_parenchyma_white_path = "./dataset3.0/lung_parenchyma_white/lungparen_" + str(count) + ".png"
org_img = normalize_hu(image[i])
# 保存图像数组为灰度图(.png)
cv2.imwrite(lung_image_path, org_img * 255)
im, binary = get_segmented_lungs(image[i]) # 肺部区域分割 这一句不隐藏 保存的就是肺实质图片
# 将像素值归一化到[0,1]区间
org_img = normalize_hu(image[i])
# 保存图像数组为灰度图(.png)
cv2.imwrite(lung_parenchyma_path, org_img * 255)
cv2.imwrite(lung_parenchyma_white_path, im * 255)
cv2.imwrite(lung_mask_path, binary * 255)
nodule_mask_path = "./dataset3.0/nodule_masks/noduleMask_" + str(count) + ".png"
img = np.zeros([512, 512], np.uint8)
edge = np.int32([edgeMap[i]])
cv2.polylines(img, edge, 1, 1)
cv2.fillPoly(img, edge, 1)
cv2.imwrite(nodule_mask_path, img * 255)
count += 1
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
