Abstract
Pathology report generation from whole-slide images (WSIs) requires both accurate evidence selection and faithful diagnostic reasoning. We propose Graph-Constrained Visual Grounding for Pathology Reasoning, a compact framework for the REG2026 reasoning and grounding tasks. Annotated chain-of-thought questions are canonicalized into a directed diagnostic graph, where graph-constrained routing guides intermediate question answering and report generation. Frozen CONCH tile embeddings, TITAN slide embeddings, and PubMedBioBERT question embeddings are integrated through question-conditioned attention for evidence retrieval, while a LoRA-adapted BioGPT decoder generates answers and reports. For ROI-level grounding, the same model is reused without additional training as a single-step 5x ROI VQA system with sufficiency-aware prompting. On the held-out validation split, the framework achieved a workflow reasoning score of 0.759. On Challenge Test Phases 1 and 2, it achieved workflow reasoning scores of 0.620 and 0.538, and visual grounding scores of 0.892 and 0.850, respectively, indicating strong alignment between generated responses and visual evidence while limiting unsupported diagnostic claims. These results demonstrate the potential of graph-constrained routing for structured pathology reasoning, warranting further evaluation across broader clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_014.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/REG2026_014_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=dbE1rnxGnZ
BibTex
@InProceedings{KotSuj_GraphConstrained_MICCAISAT2026,
author = { Kotte, Sujatha AND Sinha, Animesh Sanjay AND Nandagopal, Dhandapani AND Walia, Vidushi AND Sivakumar, Megha AND Kulasekhar, Pranesh AND Saipradeep, Vangala G. AND Sivadasan, Naveen},
title = { { Graph-Constrained Visual Grounding for Pathology Reasoning } },
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}
}