Hello BraTS 2026 organizers, Could you please confirm three details of the official Task 3 (BraTS-GoAT) evaluation configuration so participants can reproduce the leaderboard scoring accurately? 1. **Ranking metrics:** The Task 3 evaluation page and the populated `global_*` leaderboard columns indicate that ranking uses global DSC and global NSD for ET, TC, and WT. Is HD95 excluded from both the validation leaderboard and the final Task 3 ranking? 2. **NSD configuration:** `config_GoAT.yaml` in `BraTS-evaluation` does not appear to specify the NSD tolerance explicitly, so local evaluation inherits the Panoptica default. Could you confirm the exact NSD tolerance, its units, and the Panoptica version used by the official scorer? 3. **Aggregation and missing values:** Is the official aggregation a per-subject arithmetic mean with `NaN` values skipped, evaluator-error subjects skipped, and genuinely missing subjects assigned zero? Should participants therefore expect the realized denominator to differ across regions or metrics? Thank you for your clarification.

Created by Nathan Chen NathanC
Thank you for the clarification. For reproducibility, could you also confirm: 1. The exact Panoptica version used for final Task 3 scoring. 2. Whether `tau=1` for global NSD is interpreted in voxel units or physical distance (millimetres), and whether the scored images have standardized 1 mm isotropic spacing. 3. How `tau=1` is supplied to global NSD in the official scorer. In `BraTS-evaluation` 0.0.8 with Panoptica 2.1.4, the global metric path appears to call NSD without an explicit threshold or voxel spacing, and therefore falls back to 0.5. 4. Whether evaluator errors other than an absent prediction are also assigned DSC=NSD=0, and whether NaN values are excluded independently from each region's mean denominator. Thank you.
HI @NathanC, Thanks for your questions. 1- Please note HD95 metric will not be considered for the final ranking. All the metrics, including lesion-wise, hd95, TP, FP, FN, etc are provided for the participants to better evaluate the developed models, though. 2- The default value of NSD in Panoptica is 0.5, which leads to slight underestimation of the results. For the final ranking, we will use tau=1. However, this will be applied to all the submitted jobs and will be a fair and objective quantification. 3. If a certain tumor subregion did not exist, then the metric values will be NaN. However, if a submitted model fails to predict a case, the penalty score (DSC, NSD=0) will be used. Best regards, /Mehdi

Task 3 — confirmation of official DSC/NSD scoring configuration page is loading…