Abstract
The BraTS-Path 2026 challenge asks for the histologic subregion label of a single haematoxylin-and-eosin patch cropped from a digitised glioma tissue section, across ten classes whose frequencies span more than an order of magnitude. We describe a two-stage system pairing three frozen pathology foundation encoders (UNI2-h, Virchow2 and Phikon-v2) with an ensemble of gradient-boosted trees fitted on fixed-width slices of the embedding space, followed by rare-class probability boosting and per-class threshold optimisation. Our main methodological point concerns evaluation rather than architecture. Patches cropped from the same tissue section are not independent observations: splitting them by patch identity leaves nearly every patient represented on both sides of the boundary and drives internal validation accuracy to almost unity, a memorisation score rather than a generalisation estimate. We therefore use patient-grouped stratified five-fold cross-validation, giving every training patch exactly one out-of-fold prediction from a model that never observed its patient. Under this protocol the system reaches 0.762 accuracy, 0.713 Matthews correlation coefficient (MCC) and 0.541 macro F1; on the official validation phase it obtains 0.879 accuracy, 0.832 MCC, 0.924 weighted area under the ROC curve, 0.485 macro recall and 0.450 macro F1. The residual error concentrates in a few rare, morphologically entangled classes, which we quantify with a structured confusion analysis and an early-fusion ablation. Code is available at https://github.com/ujjwalbaid0408/BraTS2026_PATH.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Path_008.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=wiYZFlqgLS
BibTex
@InProceedings{BaiUjj_LeakageFree_MICCAISAT2026,
author = { Baid, Ujjwal AND Pathak, Tilak AND Baheti, Bhakti AND Madabhushi, Anant},
title = { { Leakage-Free Patient-Grouped Evaluation of Frozen Pathology Foundation Models for Long-Tailed Glioma Sub-region Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17252},
month = {pending},
page = {pending}
}