Evaluation¶

The evaluation implementation is available in the TopAneu-26 Repository for your reference, more details on the implementation and methodology can be found in the respective READMEs.

Classification Metrics¶

These metrics apply to both task 1 and task 2.

Precision is computed for every class using: Prec = TP/(TP+FP)

Recall is computed for every class using: Rec = TP/(TP+FN)

MCC is computed for every class using: MCC = (TN*TP - FN*FP)/SQRT((TP+FP)*(TP+FN)*(TN+FP)*(TN+FN))

At evaluation time all TP, FP, FN and TN are accumulated for each individual class over every prediction. TNs are counted as the number instances in the ground truth minus the number of TP and FN. In the case of task 1 a TP is defined by mere presence, in the case of task 2 a TP is defined by non-zero overlap between the predicted and ground truth masks. The TP, FP, TN, FN are accumulated per-class across all samples and the metrics are computed per-class from the resulting values. The metrics are averaged across all classes to obtain the final metrics for the submission.

Segmentation Metrics¶

These metrics apply only to task 2.

DSC is computed for every class using: DSC = 2*(A&B)/(abs(A)+abs(B))

VS is computed for every class using: VS = 1-(abs(abs(A)-abs(B))/(abs(A)+abs(B)))

HD95 is computed for every class as the 95th percentile of the longest shortest bidirectional distance between two objects' surfaces.

All metrics are computed for all cases and all classes. The per class metrics are accumulated across cases and normalized by the number of TP, FP and FN to compute the per-class metrics of the submission. The metrics are averaged across all classes to obtain the final metrics for the submission.

Ranking¶

Submissions are ranked in all individual metrics independently, the final ranking is computed as the mean rank of the individual metrics.

Last Updated: 06 Aug 2026