Segmentation evaluation metrics

CERR provides functions to compute similarity metrics between segmentations, viz., Dice similarity coefficient (DSC), Hausdorff distance (HD), deviance, and added path length (APL)^[1].

DSC

global planC
%Provide structure nos. for comparison
structNum1 = 1;
structNum2 = 2;
dice = calc_DiceMetric(structNum1,structNum2,planC);

95^th percentile HD (HD₉₅)

global planC
%Provide structure nos. for comparison
structNum1 = 1;
structNum2 = 2;
hd95 = calc_HausdorffMetric(structNum1,structNum2,planC); %Returns HD95 in cm

Deviance

Deviance is the ratio of the volume of the true structure in excess of the test structure + the excess volume of the test structure over the true structure to the volume of the true structure.

• Margin of tolerance: The true structure is expanded and contracted (in 3D) by a user-specified margin (cm) while computing the excess volume of the true (test) structure over the test(true) structure.

global planC
%Provide structure nos. for comparison
trueStructNum = 1;
testStructNum = 2;
%Define tolerance margin in cm
tolMargin = 0.1;
deviance = calc_deviance(trueStructNum,testStructNum,tolMargin,planC)

Surface dice similarity metric^[1]

Additionally, wrappers to the DeepMind library for computing surface distance-based performance metrics for segmentation tasks are provided.

global planC
%Provide names of structures for comparison
struct1 = 'strName1'; 
struct2 = 'strName2'; 
%Define tolerance margin in cm
tol_cm = 1;  
surfDice = calcSurfaceDice(struct1, struct2, tol_cm, planC);

References

1. Vaassen F, Hazelaar C, Vaniqui A, Gooding M, van der Heyden B, Canters R, van Elmpt W. Evaluation of measures for assessing time-saving of automatic organ-at-risk segmentation in radiotherapy. Physics and Imaging in Radiation Oncology. 2020 Jan 1;13:1-6.