Abstract
This paper describes our contribution to the BraTS-Path 2026 challenge, in which each H&E-stained tissue patch has to be assigned to one of ten glioblastoma sub-regions. Two properties make this difficult: the sub-regions overlap considerably in appearance, and they follow a long-tailed class distribution. Most approaches take the representations of a pathology foundation model as they are and fit a lightweight classifier on top. This leaves the pretrained representation intact, but caps performance at what those features already make decodable. We quantify this ceiling using linear probes on eight foundation encoders: UNI, UNI2-h, Virchow, Virchow2, Prov-GigaPath, neuroFM-HE20x, H-optimus-0 and H-optimus-1. All eight lie within 0.05 macro-F1 of one another, with the two H-optimus ViT-g/14 models leading. Block expansion in the style of LLaMA Pro then moves past this limit: four identity-initialised transformer blocks are interleaved among the encoder’s frozen blocks and trained with a linear head, end-to-end on raw patches. Because the inserted blocks compute the exact identity at initialisation, training starts at the frozen encoder’s linear-probe solution and the pretrained information is preserved rather than overwritten. Training only these 113M parameters (≈ 10% of the encoder) raises macro-F1 from 0.901 to 0.959 on our internal split, with every sub-region but one above 0.94 F1. On the organisers’ Validation-Set the gain is larger still, from 0.550 for the frozen probe to 0.752. The code is available on GitHub (https://github.com/M4rt1nM4yr/BraTS26-Path_Team-FAU).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BraTS_Path_005.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=aFum4MYmdI
BibTex
@InProceedings{MayMar_Beyond_MICCAISAT2026,
author = { Mayr, Martin AND Rivera Monroy, Luis Carlos AND Maier, Andreas AND Gupta, Pooja AND Köstler, Harald AND Wellein, Gerhard},
title = { { Beyond Frozen Features: Block Expansion of Pathology Foundation Models for Glioblastoma 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}
}