lisamenzel99
commented
2 months ago Metric options
2D or 3D (but since we already need to compute DICE in 3D, we can do everything in 3D) Overlap-based segmentation metrics
- Dice similarity coefficient (DSC): measures overlap of predicted segmentation and ground truth (Identical to F1-score, more used in medical domain) https://metrics-reloaded.dkfz.de/metric?id=dsc
- Intersection over Union (IoU): penalizes under- and over-segmentation more than the DSC (more used in computer vision community) > often used in combination with DSC https://metrics-reloaded.dkfz.de/metric?id=intersection_over_union
Only 3D Boundary-based segmentation metrics
- Normalized surface distance (NSD): measures the overlap between two boundaries (within a certain distance) to ensure that boundary errors are within a clinically acceptable range > Existence-based penalization (existence of spatial outliers penalized irrespective of their distance to the reference contour) https://metrics-reloaded.dkfz.de/metric?id=normalized_surface_distance
- Hausdorff Distance (HD): measures the distance between ground truth and predicted segmentation >> highly sensitive to outliers though, so maybe symmetric Average Hausdorff Distance (AHD) instead, but maybe NSD is better https://metrics-reloaded.dkfz.de/metric?id=hd Towards a Guideline for Evaluation Metrics in Medical Image Segmentation
Other metrics
- Volumetric similarity (VS): quantifies volumetric overlap (this could be useful to check if the predicted segmentation and ground truth have the same volume which can be useful in the medical domain if you want to use the segmentations to for example determine size of the heart) Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool
The bold ones are the ones I think we could implement first. Please let me know if you have other ideas
SS-Fatima
Also add a justification for each metric