Abstract
Brain metastases are the most common intracranial tumors in adults, and typically multifocal, with most lesions under a centimeter across at diagnosis. Because radiosurgery plans and doses each lesion as a separate target, a metastasis that is not delineated is a metastasis that is not treated. BraTS-MET 2026 reflects this by scoring each lesion independently rather than the segmentation as a whole, which puts detection and delineation in tension: suppressing false positives improves overlap while discarding lesions the system would otherwise report. We describe the system we submit and the choices that follow. Its body is a Primus vision transformer tokenized finely enough that a small metastasis occupies a meaningful fraction of a token, trained in the nnU-Net framework and finetuned with an instance-balanced rather than volume-weighted objective. A second transformer, a dedicated small-lesion detector, contributes only to cases carrying a high lesion burden, and its detections are kept only where the body’s folds corroborate them. Holding the pipeline together is Semantic Spatial Uncertainty, which compares an ensemble prediction against the union of its cross-validation folds and therefore needs nothing a cross-validated pipeline does not already produce. At inference it acts as failure detection and recovery, flagging lesions the averaged ensemble under-segments and repairing them by fold majority. Over the training set, the same signal ranks annotation errors for expert review, and we release it as a quality-control tool that turns any five-fold ensemble into a way of auditing the labels it was trained on.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_METS_015.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=GXSNOeNRbR
BibTex
@InProceedings{NitFra_ATransformer_MICCAISAT2026,
author = { Nitti, Francesco AND Müller, Michael AND Ermiş, Ekin AND Reyes, Mauricio},
title = { { A Transformer Pipeline with Uncertainty-Guided Curation and Failure Recovery for 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}
}
