For Task 1, is detection F1 computed per-region (WT/TC/ET/RC) as in config_mets.yaml (in code and current leaderboard), or on the combined ET+NETC+RC lesion as stated in the proposal (https://zenodo.org/records/19714728)?

Created by Soumya Kundu aymuos
Thank you so much for the detailed responses and guidance. Best, Soumya
@aymuos Please proceed with the exact implementation available in the BraTs Evaluation pipeline (config_mets.yaml). In practice, a single case may have multiple lesions, and not all of them contain all the subregions (so the model may detect ET but not NETC). Therefore, for a practical implementation and more realistic evaluation, these metrics should be computed per subregion individually. But I agree with you that the description in the design document is not aligned. Thank you for your attention. Best regards, /Mehdi
Page 23 of design document: "We apply the detection evaluation for every single lesion within one MRI study. **For the detection arm of the challenge we define a lesion as a collective term including the Enhancing Tumor, the Non-enhancing Tumor Core, and the Resection cavity."** This is different from what is defined in the config_mets.yaml . This actually means RC as a class is not existing at all for detection?
Hi @aymuos, Thanks for raising this concern. Specifically, DSC, NSD, and F1 scores are the metrics for the final leaderboard ranking, as described in the [challenge description](https://www.synapse.org/Synapse:syn74274097/wiki/639602). In accordance with the [designed document](https://zenodo.org/records/19714728), the [config_mets.yaml](https://github.com/BraTS/BraTS_evaluation/blob/main/brats_evaluation/configs/config_mets.yaml) implements the subregions of interest, including ET, TC, RC, and WT. Therefore, the mentioned metrics will be quantified over the mentioned regions, followed by statistical analysis for the final ranking. Best regards, /Mehdi
Hello, thank you for your response. In my opinion this is something which should be explicitly clear as currently there is ambiguity regarding the evaluation of the exact label. Unless I am missing something, I do not see a clear merit in not specifiying exactly which labels will be involved in the final ranking?
Hi @aymuos, Thanks for reaching out. Please note that the evaluation pipeline computes all the combinations and metrics to help participants tune their solutions. Nevertheless, not all of them will be used for ranking. Best regards, /Mehdi

Computing RC class Scores in BraTS-MeTS 2026 page is loading…