图像分割结果的评估---DICE参数

GT（ground truth）图像:我理解的就是含有理论分割结果的图像，用来和结果图像进行比较的参照图像。那么GT图像怎么来，一般通过专家手工勾画出来，得到理论值。当然为避免偶然性，你可以选择多个专家，得到多个GT分割结果，取各项评估参数的平均值。

分割精度：个人理解就是分割准确的面积占GT图像中真实面积的百分比。找了半天找了这么个公式（如下）。

 

不难理解，其中Rs表示专家手工勾画出的分割图像的参考面积，Ts表示算法分割得到的图像的真实面积，|Rs-Ts|表示错误分割的像素点个数。

过分割率：即分割在GT图像参考面积之外的像素点的比率，计算公式如下：

 

含义同上，Os表示本不应该包含在分割结果中的像素点个数，实际却在分割结果中的像素点个数。换句话讲，Os中的像素点出现在实际分割图像中，但不出现在理论分割图像Rs中。

欠分割率：即分割在GT图像参考面积之中欠缺的像素点的比率，计算的公式如下：

 

含义同上，Us表示本应该包含在分割结果中的像素点个数，实际却不在分割结果中的像素点个数。换句话讲，Us中的像素点出现在理论分割图像中，但不出现在实际分割图像中。

代码（由于当时时间紧张，很多可以封装成函数）：

获得肝血管GT图像：

import pydicom
from pylab import *
import os
import PIL.Image as image
from Anisotropicdiffusion import *

PathDicom = "./PATIENT_DICOM20/"  # 与python文件同一个目录下的文件夹
lstFilesDCM1 = []
number1 = 0
for dirName, subdirList, fileList1 in os.walk(PathDicom):
    for filename in fileList1:
        lstFilesDCM1.append(os.path.join(dirName, filename))
        number1 += 1

RefDs = pydicom.read_file(lstFilesDCM1[0])
ConstPixelSpacing = (float(RefDs.PixelSpacing[0]), float(RefDs.PixelSpacing[1]), float(RefDs.SliceThickness))
print(ConstPixelSpacing)

PathDicom = "./liver20/"  # 与python文件同一个目录下的文件夹
lstFilesDCM2 = []
number2 = 0
for dirName, subdirList, fileList2 in os.walk(PathDicom):
    for filename in fileList2:
        lstFilesDCM2.append(os.path.join(dirName, filename))
        number2 += 1

PathDicom = "./artery20/"  # 与python文件同一个目录下的文件夹
lstFilesDCM3 = []
number3 = 0
for dirName, subdirList, fileList3 in os.walk(PathDicom):
    for filename in fileList3:
        lstFilesDCM3.append(os.path.join(dirName, filename))
        number3 += 1

PathDicom = "./portalvein20/"  # 与python文件同一个目录下的文件夹
lstFilesDCM4 = []
number4 = 0
for dirName, subdirList, fileList4 in os.walk(PathDicom):
    for filename in fileList4:
        lstFilesDCM4.append(os.path.join(dirName, filename))
        number4 += 1

PathDicom = "./venacava20/"  # 与python文件同一个目录下的文件夹
lstFilesDCM5 = []
number5 = 0
for dirName, subdirList, fileList5 in os.walk(PathDicom):
    for filename in fileList5:
        lstFilesDCM5.append(os.path.join(dirName, filename))
        number5 += 1

if number1 != number2:
    print("fileerro")

for i in range(number1):
    dcm1 = pydicom.read_file(lstFilesDCM1[i])
    dcm1.image = dcm1.pixel_array * dcm1.RescaleSlope + dcm1.RescaleIntercept
    slices = []
    slices.append(dcm1)
    img1 = slices[int(len(slices) / 2)].image.copy()

    dcm2 = pydicom.read_file(lstFilesDCM2[i])
    dcm2.image = dcm2.pixel_array * dcm2.RescaleSlope + dcm2.RescaleIntercept
    slices = []
    slices.append(dcm2)
    img2 = slices[int(len(slices) / 2)].image.copy()

    dcm3 = pydicom.read_file(lstFilesDCM3[i])
    dcm3.image = dcm3.pixel_array * dcm3.RescaleSlope + dcm3.RescaleIntercept
    slices = []
    slices.append(dcm3)
    img3 = slices[int(len(slices) / 2)].image.copy()

    dcm4 = pydicom.read_file(lstFilesDCM4[i])
    dcm4.image = dcm4.pixel_array * dcm4.RescaleSlope + dcm4.RescaleIntercept
    slices = []
    slices.append(dcm4)
    img4 = slices[int(len(slices) / 2)].image.copy()

    dcm5 = pydicom.read_file(lstFilesDCM5[i])
    dcm5.image = dcm5.pixel_array * dcm5.RescaleSlope + dcm5.RescaleIntercept
    slices = []
    slices.append(dcm5)
    img5 = slices[int(len(slices) / 2)].image.copy()

    img = np.zeros((512, 512))
    img30 = np.zeros((512, 512))
    img40 = np.zeros((512, 512))
    img50 = np.zeros((512, 512))
    height, width = img2.shape
    for k in range(height):
        for j in range(width):
            if img2[k, j] != 0:
                img2[k, j] = 1
            if img3[k, j] != 0:
                img3[k, j] = 1
            if img4[k, j] != 0:
                img4[k, j] = 1
            if img5[k, j] != 0:
                img5[k, j] = 1
            img1[k, j] *= img2[k, j]
            img30[k, j] = img1[k, j]*img3[k, j]
            img40[k, j] = img1[k, j]*img4[k, j]
            img50[k, j] = img1[k, j]*img5[k, j]
            img[k, j] = img30[k, j]+img40[k, j]+img50[k, j]

    img = np.uint8(img)
    img6 = image.fromarray(img)
    filename = "D:/Pychram/liverpicture/mask_" + fileList1[i] + ".png"
    img6.save(filename)
    print(i)

计算分割精度：

import os
import PIL.Image as Image
from Anisotropicdiffusion import *

PathDicom = "C:\\Users\\LCSH\\Desktop\\Pychram\\liverpicture"  # 与python文件同一个目录下的文件夹
lstFilesDCM1 = []
number1 = 0
for dirName, subdirList, fileList1 in os.walk(PathDicom):
    for filename in fileList1:
        lstFilesDCM1.append(os.path.join(dirName, filename))
        number1 += 1

PathDicom = "C:\\Users\\LCSH\\Desktop\\Pychram\\liverpicture1"  # 与python文件同一个目录下的文件夹
lstFilesDCM2 = []
number2 = 0
for dirName, subdirList, fileList2 in os.walk(PathDicom):
    for filename in fileList2:
        lstFilesDCM2.append(os.path.join(dirName, filename))
        number2 += 1

p = 0
for i in range(number1):
    img = Image.open(lstFilesDCM1[i])
    img1 = img.load()
    img = Image.open(lstFilesDCM2[i])
    img2 = img.load()

    f = h = 0
    height, width = img.size
    for k in range(height):
        for j in range(width):
            if img1[k, j] > 0:
               f += 1
            if img1[k, j] > 0 and img2[k, j] == 0:
               h += 1
    g = h/f
    p += g
    print(i)
L = p / number1
print(L)