Abstract
Brain metastases segmentation in the pre- and post-treatment
setting requires delineating four regions: non-enhancing tumor core, surrounding FLAIR hyperintensity, enhancing tumor, and, in post-treatment
cases, the resection cavity (RC). RC is by far the hardest: it appears in
only 12.9% of training cases and is frequently abandoned during training. We address BraTS 2026 Task 1 with a residual-encoder nnU-Net
and attack RC at three levels. Two independent observations show that
adequate RC exposure is not sufficient for RC to be learned: a fold stratified to guarantee exposure produced zero RC voxels, and a baseline fold
learned RC by epoch 27 and lost it at epoch 29, remaining at exactly
zero for 471 further epochs with its data unchanged. An architecture-level
probe diverged before convergence and is inconclusive. An optimizationlevel intervention (foreground oversampling with a class-weighted crossentropy term on RC) raises RC lesion-wise Dice from 0.394 to 0.461, but
only when paired one-to-one with a strong RC-competent fold; diluted
into a four-model average the gain disappears, and control pairings of two
baseline folds do not reproduce it. Instance-level analysis shows the residual error is false-positive dominated: the true-positive count is unchanged
across every ensemble built on the specialist. A connected-component filter removing sub-100 mm3 RC components cuts false positives from 128
to 11, raising RC Dice to 0.518, although with RC scored on only 22
of 179 cases this is not statistically significant. Our final configuration
attains lesion-wise Dice of 0.650, 0.681, 0.649, and 0.518 for enhancing
tumor, tumor core, whole tumor, and resection cavity. For rare regions,
ensemble composition matters more than ensemble size, and the number
of scored cases limits what any single comparison can establish.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_006.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{GraAnd_RareRegion_MICCAISAT2026,
author = { Grajales, Andres Felipe Romero},
title = { { Rare-Region Optimization and Ensemble Composition for Pre- and Post-Treatment Brain Metastases 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}
}