MONAI
MONAI copied to clipboard
Implementation of the centre line Dice score / topology precision / topology sensitivity
Is your feature request related to a problem? Please describe. Need for metrics to evaluate connectivity - particularly useful for evaluation of vascular segmentations
Describe the solution you'd like Implementation of the three main topology metrics relying on skeletonisation of reference and binary predictions. Notation for indication in the below formula: $S_{Pred}$ - Skeleton of the prediction $S_{Ref}$ - Skeleton of the reference $Pred$ - Prediction $Ref$ - Reference
Topology precision: $T_{prec}(S_{Pred},Ref) = \dfrac{\vert S_{Pred} \bigcap Ref \vert}{\vert S_{Pred}\vert}$ Topology sensitivity $T_{sens}(S_{Ref}, Pred) = \dfrac{\vert S_{Ref} \bigcap Pred\vert}{\vert S_{Ref} \vert}$ The centre line DSC uses the skeletonisation of a binary segmentation. It is expressed as a function of the topology precision and the topology sensitivity
$clDSC = \dfrac{2*T_{prec}(S_{Pred},Ref) * T_{sens}(S_{Ref},Pred)}{T_{prec}(S_{Pred},Ref) + T_{sens}(S_{Ref},Pred)}$