Abstract
REG2026 requires systems to generate structured diagnostic reasoning, visually grounded ROI-level answers, and final pathology reports from H\&E whole-slide images (WSIs). We developed a two-part framework for histopathology question answering, visual grounding, and report generation. For Metric A (workflow reasoning), multi-scale tissue patches are extracted from WSIs at both $20\times$ and $10\times$ magnifications and encoded by the UNI2-h foundation model. They are aggregated by an ensemble of ten organ-conditioned, multi-head TransMIL networks that answer structured diagnostic questions. An auxiliary classifier predicts the organ, which conditions the answering heads and selects an organ-specific routing graph; rule-based logic then derives dependent findings, assembles the answered questions into a chain-of-thought reasoning graph, and formats the final pathology report. For Metric B (visual grounding), we fine-tune a compact Qwen3.5-0.8B vision-language model on a dedicated dataset of tissue and background regions of interest curated from public TCGA slides and the challenge data, and pair it with a lightweight rule-based decision gate. This design targets the content-agnostic scoring protocol, which rewards accurate background rejection, invariance under input perturbation, and distinct outputs for tissue and background regions. On the Test Phase1 leaderboard, our submission ranks second among 20 teams with an overall score of 0.8006 and reaches the maximum score on every visual-grounding robustness metric.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_004.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=wJn1cxW5nS
BibTex
@InProceedings{DaiDin_MultiScale_MICCAISAT2026,
author = { Dai, Dingyuan AND Qi, Heli AND Zhou, Zeqi AND Ling, Xitong AND Xuan, Weihao AND Liu, Tianyu},
title = { { Multi-Scale MIL and Multi-Modal Large Language Modeling for Pathology Workflow Reasoning and Visual Grounding } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}