Abstract
Automated segmentation of brain metastases is complicated by two features of the clinical problem: lesions span three orders of magnitude in volume, and the resection cavity — newly emphasized in the post-treatment setting — is absent from most scans. We submitted a default nnU-Net v2 five-fold ensemble to Task 1 of the BraTS 2026 challenge (team NeuroSeg HFA), obtaining lesion-wise Dice similarity coefficient (DSC) scores of 0.696 (tumour core), 0.673 (enhancing tumour), 0.654 (whole tumour), and 0.446 (resection cavity) on the validation set. Rather than pursuing architectural refinement, we used per-case metrics from cross-validation on 1,036 cases to characterize where this baseline fails. Two findings emerge. First, segmentation quality scales strongly with lesion volume (Spearman rho = 0.39–0.76, p < 10^-13): mean Dice for enhancing tumour falls from 0.820 in the largest-burden cases to 0.485 in the smallest, with complete failure in a quarter of the latter. The challenge’s own instance metrics show that three lesions below the 27 mm3 threshold are missed for every one detected — a systematic under-count of exactly the quantity on which treatment stratification depends, since ASCO-SNO-ASTRO reserves stereotactic radiosurgery for patients with one to four intact metastases. Second, the poor resection-cavity score is not a delineation failure: where a cavity exists, mean Dice is 0.560, but the model predicts a cavity in 38% of the 905 scans that contain none. We argue that aggregate Dice conceals both failure modes, and that size-stratified reporting should accompany challenge results in this domain.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_010.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=VWj8HOqfR6
BibTex
@InProceedings{daMoi_Beyond_MICCAISAT2026,
author = { da Costa Cunha, Moisés AND de Oliveira, Monique M. C. Daemon},
title = { { Beyond Aggregate Dice: Size-Stratified Failure Analysis of an nnU-Net Baseline for Brain Metastasis Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17253},
month = {pending},
page = {pending}
}