Abstract
Generating a pathology report from a whole-slide image (WSI) is increasingly framed as a reasoning problem: a model should not only produce the report but also expose the diagnostic reasoning behind it. Free-form generated reasoning is difficult to verify and can diverge from the reported findings. We present the Grounded Ontology Walker (GOW), a discriminative pipeline in which the reasoning is an explicit walk over a diagnostic ontology mined from expert reasoning chains, and the report is a deterministic render of that walk. Every reasoning step is a valid ontology transition, so the chain is auditable, and because the report is derived from the chain it is faithful to the reasoning by construction. Each diagnostic question is answered from the specific image regions that a question-conditioned selector attends to, over patch features from Virchow2, a frozen pathology foundation model, which grounds and localizes every decision. To remain safe on organs that are absent from training, GOW includes a per-organ distribution-shift detector that flags unseen organs and routes them through the nearest in-ontology topology with open-vocabulary naming, and we show it generalizes to novel organs on public external slides. On the blind Test Phase 1 evaluation GOW attains an overall score of 0.791: the visual-grounding metric (background rejection, input sensitivity, and cross-region consistency) is maximal at 1.00, the reasoning graph is recovered at edge F1 0.852 with matchededge answer similarity 0.772, and the report scores 0.572; the final Test Phase 2 evaluation gives an overall score of 0.733 with the same maximal visual-grounding score. On the held-out development split, edge F1 is 0.922 and organ recognition 0.993. We report performance per metric and evaluation setting and analyze the contribution of each component.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_007.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=m8GFe07dbD
BibTex
@InProceedings{LalDev_Grounded_MICCAISAT2026,
author = { Lalwani, Devansh AND Bhat, Swapnil},
title = { { Grounded Ontology Walker: Interpretable Diagnostic Reasoning and Report Generation for Whole Slide Images } },
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}
}