Abstract
Generating clinically useful pathology reports from gigapixel whole slide images (WSIs) requires computational models that identify diagnostically relevant evidence, follow a valid reasoning path, and pro- duce meaningful reportable conclusions. The MICCAI 2026 REG2 Chal- lenge provides a comprehensive benchmark to facilitate innovation on this space. To this end, we developed two independent chain-of-thought (CoT) systems: i) Base-CoT and ii) R-CoT. Base-CoT is an autore- gressive CoT model combining spatially contextualized CONCH patch embeddings, TITAN slide-level embeddings, and GECKO concept con- ditioning to generate structured diagnostic reasoning traces. R-CoT is a deterministic retrieval-based system decomposing reasoning generation into organ classification, diagnosis classification, and template retrieval conditioned on the predicted (organ, diagnosis) pair. R-CoT uses frozen UNI2-h patch features at 20× and 10× (capped at 512 tissue patches per scale), a linear head organ classifier on mean-pooled features, multi- scale per-organ gated-attention multiple-instance learning diagnosis en- sembles, and a template bank mined from canonical training chains. Our quantitative performance evaluation for Base-CoT yielded a workflow ranking score of 0.445 on the official Test Phase 1 leaderboard, and 0.383 on our internal hold-out validation split, comprising 10% of the training data and stratified by organ with random sampling within each stratum. However, the R-CoT model on internal hold-out validation split yielded a workflow ranking score of 0.683. It achieved Test Phase 2 score of 0.6694 on the hidden test set, providing external validation of the retrieval-based approach.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_009.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=8sQ7MZ2wVN
BibTex
@InProceedings{InnShu_Diagnostic_MICCAISAT2026,
author = { Innani, Shubham AND Thakur, Siddhesh AND Feldman, Michael AND Bakas, Spyridon},
title = { { Diagnostic Pathology Reporting and Reasoning via Autoregressive and Retrieval-Based Chain-Of-Thought } },
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}
}