Evaluation Metrics

[Noza23]

We need some evaluation metrics to evaluate the performance of our predictions.

Myotube:

Pairwise measures

• [x] Boundary IoU (varying parameter 'd') https://metrics-reloaded.dkfz.de/metric?id=boundary_intersection_over_union

  Calculates the overlap between the predicted and true borders of the segmented object. This metric is especially relevant in tasks where the precision of object boundaries is important. Higher Border IoU values indicate better performance, as they signify that the model's predicted border aligns closely with the ground truth.

• [x] Normalized Surface Distance (NSD) (varying parameter tau) https://metrics-reloaded.dkfz.de/metric?id=normalized_surface_distance

  Measures DSC (2*iou/(1+iou)) on boundary pixels. It is particularly useful in medical imaging where precise boundary delineation is crucial.

Processes

• [x] Panoptic Quality (takes both detection and segmentation quality into account) Quality https://metrics-reloaded.dkfz.de/metric?id=panoptic_quality

  Defined as the production between detection quality and segmentation quality.

• [x] Average Precision https://metrics-reloaded.dkfz.de/metric?id=average_precision

  AP might be useful because it balances the need for the model to detect as many objects as possible (high recall) while maintaining that the detections are accurate (high precision).

Nuclei:

Pairwise Measures

• [x] IOR (reason: 1 prediction for 2 references) https://metrics-reloaded.dkfz.de/metric?id=intersection_over_reference

Processes

• [x] Panoptoc Quality (takes both detection and segmentation quality into account) Quality https://metrics-reloaded.dkfz.de/metric?id=panoptic_quality

• [x] Average Precision https://metrics-reloaded.dkfz.de/metric?id=average_precision

[Noza23]

24