Abstract
We describe the NeuroForge submission to the BraTS 2026 pathology challenge (BraTS-Path), which asks for patch-level classification of digitized glioma tissue sections into nine histologic tissue classes plus a “None of the above” (NOTA) class. The task couples two difficulties that interact badly: extreme class imbalance with patient-level concentration of several classes — the NOTA class (15.3% of patches) lives in only 5 of the 126 labeled patients, presence of lymphocytes (PL) in 7, and leptomeningeal infiltration (LI) in 10 — and a distribution shift between the labeled training data and the official validation data, which makes any locally constructed validation split unrepresentative by construction. Our final system is a seven-model ensemble of patch classifiers built on near-frozen pathology foundation model backbones (backbone learning rate 100× lower than the head): two ResNet-34 networks initialized from ImageNet, two phikon ViT-B/16 models (self-supervised iBOT pretraining on 40M histology tiles) and three phikon-v2 ViT-L/16 models (DINOv2 pretraining on 460M tiles), combined by softmax averaging with flip test-time augmentation over a whole-patch input view. Two findings mattered more than any architecture choice. First, a routine patient-level holdout had silently swept 98.8% of one class out of the training pool, driving its recall to exactly zero; retraining on all 126 patients with true-prior class weighting restored the dead class with no collateral damage. Second, architecture diversity decorrelates ensemble errors far more than seed diversity (90–96% versus 98.6% agreement with the ensemble), which justified spending the ensemble budget on heterogeneous backbones. On the public validation leaderboard the ensemble reached an accuracy of 0.8755 (rank 15 of 20 at the July 2026 snapshot). We also report a series of negative results — native-resolution multi-crop evaluation and expectation-maximization prior correction — that follow directly from the distribution shift, and we argue that on this task the leaderboard is the only unbiased validator available to participants.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Path_011.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=n7mlkYpvmu
BibTex
@InProceedings{ZhaJia_Pathology_MICCAISAT2026,
author = { Zhang, Jianfeng AND Hou, Zhiwei AND Guan, Mengshan AND Huang, Xinping AND Zhang, Hui},
title = { { Pathology Foundation Models Under Class Starvation and Distribution Shift: NeuroForge at the BraTS 2026 Pathology Challenge } },
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}
}