Abstract
We describe our submission to the REG2 (Pathologist Reasoning-Guided Report Generation) Challenge, which pairs each whole-slide image (WSI) with a step-by-step diagnostic chain-of-thought (CoT) and a structured pathology report (Interface 1, Metric A), alongside free-text answers about regions of interest that must correctly reject non-informative background (Interface 0, Metric B). Our pipeline encodes stratified, artifact-filtered tissue patches with a frozen foundation model, aggregates them with a five-fold ensemble of attention-based multiple-instance-learning heads, selects a diagnostic-template graph with a second ensembled classifier, and generates answers and the final report with a LoRA-adapted medical vision-language model, optionally conditioned on retrieved similar training cases. Because the platform allows exactly one non-revocable final submission, every trainable component is validated with five-fold cross-validation rather than trained on the full corpus, preserving a held-out check even for the deployed model. On the platform Test Phase 1, our submission scored an Overall Score of 0.7504 (Metric A 0.6435, Metric B 1.0000), obtained through a one-time reopened submission that does not count toward the official leaderboard or ranking. Our official, final Test Phase 2 result (Overall Score 0.6775, Metric A 0.5393, Metric B 1.0000) has since been released. We report both results alongside clean cross-validation metrics and a LoRA/RAG ablation on a held-out set, explicit throughout about which numbers are platform-measured, which are local proxies, and where known evaluation leakage inflates a figure.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_015.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/REG2026_015_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=7PJIuUZDqD
BibTex
@InProceedings{OleMik_Ensembled_MICCAISAT2026,
author = { Olesiński, Mikołaj},
title = { { Ensembled Foundation-Model Pipeline for Reasoning-Guided Pathology Report Generation: A REG2 Challenge Submission } },
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}
}