gmented

【SimpleITK】医疗影像分割结果评价指标计算

⽂章⽬录

OverlapMeasures

jaccard

dice

volume_similarity

fal_negative

fal_positive

其中：

Volumemeasures:

建⽴枚举对象：

fromenumimportEnum

#Uenumerationstoreprentthevariouvaluationmeasures

classOverlapMeasures(Enum):

jaccard,dice,volume_similarity,fal_negative,fal_positive=range(5)

reference_gmentation=reference_gmentation_STAPLE

建⽴空的数组，为了后⾯保存结果：

#Emptynumpyarraystoholdtheresults

overlap_results=((len(gmentations),len(OverlapMeasures.__members__.items())))

overlap_results

array([[0.,0.,0.,0.,0.],

[0.,0.,0.,0.,0.],

[0.,0.,0.,0.,0.]])

计算每个指标：

overlap_measures_filter=verlapMeasuresImageFilter()

fori,ginenumerate(gmentations):

#Overlapmeasures

overlap_measures_e(reference_gmentation,g)

overlap_results[i,]=overlap_measures_cardCoefficient()

overlap_results[i,]=overlap_measures_eCoefficient()

overlap_results[i,_]=overlap_measures_umeSimilarity()

overlap_results[i,_]=overlap_measures_NegativeError()

overlap_results[i,_]=overlap_measures_PositiveError()

overlap_results

结果：

volumeSimilarityv1+v2

2∗(v1−v2)

array([[0.82,0.901,0.052,0.075,0.122],

[0.881,0.937,-0.013,0.069,0.057],

[0.843,0.915,-0.088,0.124,0.044]])

SurfaceDistanceMeasures

hausdorff_distance

mean_surface_distance

median_surface_distance

std_surface_distance

max_surface_distance

建⽴枚举对象：

classSurfaceDistanceMeasures(Enum):

hausdorff_distance,mean_surface_distance,median_surface_distance,std_surface_distance,max_surface_distance=range(5)

reference_gmentation=reference_gmentation_STAPLE

建⽴空的数组，为了后⾯保存结果：

surface_distance_results=((len(gmentations),len(SurfaceDistanceMeasures.__members__.items())))

surface_distance_results

array([[0.,0.,0.,0.,0.],

[0.,0.,0.,0.,0.],

[0.,0.,0.,0.,0.]])

对GT进⾏预处理：

#Utheabsolutevaluesofthedistancemaptocomputethesurfacedistances(distancemapsign,outsideorinside

#relationship,isirrelevant)

label=1

reference_distance_map=(MaurerDistanceMap(reference_gmentation,squaredDistance=Fal))

reference_surface=ontour(reference_gmentation)

statistics_image_filter=ticsImageFilter()

#Getthenumberofpixelsinthereferencesurfacebycountingallpixelsthatare1.

statistics_image_e(reference_surface)

num_reference_surface_pixels=int(statistics_image_())

计算指标：

hausdorff_distance_filter=rffDistanceImageFilter()

fori,ginenumerate(gmentations):

hausdorff_distance_e(reference_gmentation,g)

surface_distance_results[i,rff_]=hausdorff_distance_sdorffDistance()

#Symmetricsurfacedistancemeasures

gmented_distance_map=(MaurerDistanceMap(g,squaredDistance=Fal,uImageSpacing=True))

gmented_surface=ontour(g)

#ultingdistance

#mapscontainnon-zerovaluesonlyonthesurface(theycanalsocontainzeroonthesurface)

g2ref_distance_map=reference_distance_map*(gmented_surface,oat32)

ref2g_distance_map=gmented_distance_map*(reference_surface,oat32)

#Getthenumberofpixelsinthereferencesurfacebycountingallpixelsthatare1.

statistics_image_e(gmented_surface)

num_gmented_surface_pixels=int(statistics_image_())

#Getallnon-zerodistancesandthenaddzerodistancesifrequired.

g2ref_distance_map_arr=ayViewFromImage(g2ref_distance_map)

g2ref_distances=list(g2ref_distance_map_arr[g2ref_distance_map_arr!=0])

g2ref_distances=g2ref_distances+

list((num_gmented_surface_pixels-len(g2ref_distances)))

ref2g_distance_map_arr=ayViewFromImage(ref2g_distance_map)

ref2g_distances=list(ref2g_distance_map_arr[ref2g_distance_map_arr!=0])

ref2g_distances=ref2g_distances+

list((num_reference_surface_pixels-len(ref2g_distances)))

all_surface_distances=g2ref_distances+ref2g_distances

#Themaximumofthesy

#general,itisnotequaltotheHausdorffdistancebetweenallvoxel/pixelpointsofthetwo

#gmentations,thisbelow.

surface_distance_results[i,_surface_]=(all_surface_distances)

surface_distance_results[i,_surface_]=(all_surface_distances)

surface_distance_results[i,_surface_]=(all_surface_distances)

surface_distance_results[i,_surface_]=(all_surface_distances)

print(surface_distance_results)

结果：

[[4.9050.4590.0.854.905]

[3.4690.290.0.6893.469]

[5.2030.4310.0.8315.203]]

为了理解上⾯的代码，我们需要了解豪斯多夫距离（hausdorffdistance）：

豪斯多夫距离

参考,。

在这⾥插⼊图⽚描述

假设有两组集合，，A点集有p个点，⽽B点集有q个点。则这两个点集合之间的Hausdorff距离定义

为:

其中，

为，在A中的每⼀个点，到距离此点最近的B集合中的点的距离，然后对这些距离进⾏排序。

A中有p个点，每个点都有B中距离最近的点。两两之间求距离，需要计算次。

取所有距离中的最⼤值为的值。

someAPI

ontour

Labelsthepixelsontheborderoftheobjectsinalabeledimage.得到mask的3维轮廓。

对reference_gmentation进⾏轮廓提取处理：

reference_surface=ontour(reference_gmentation)

A={a,…,a}1

pB={b,…,b}1

q

H(A,B)=max(h(A,B),h(B,A))

h(A,B)

ai∣a−ib∣i

p∗q

h(A,B)

MaurerDistanceMap

reference_distance_map=(MaurerDistanceMap(reference_gmentation,squaredDistance=Fal,uImageSpacing=True))

ticsImageFilter

statistics_image_filter=ticsImageFilter()

#Getthenumberofpixelsinthereferencesurfacebycountingallpixelsthatare1.

statistics_image_e(reference_surface)

num_reference_surface_pixels=int(statistics_image_())

print(num_reference_surface_pixels)

558

verlapMeasuresImageFilter

overlap_measures_filter=verlapMeasuresImageFilter()

rffDistanceImageFilter

hausdorff_distance_filter=rffDistanceImageFilter()