Abstract
Reliable automated pathology reporting should yield not only an accurate final report but also a transparent, auditable trace of the diagnostic decisions behind it. We present RECAP (REtrieval-based Chain-of-thought Answering for Pathology), a training-free method that maps an H\&E whole-slide image (WSI) to a retrieved reasoning graph consisting of canonical diagnostic question-answer steps that terminates in a structured pathology report. Each slide is encoded with the frozen TITAN foundation model, and the canonical reasoning graph is completed by $k$-nearest-neighbour ($k$-NN) retrieval over TITAN slide embeddings, so that the final report, case-variable answers, and reasoning-graph edges are all drawn from the training distribution. No generative language model is used during inference. Rather than assembling a prediction from independently retrieved components, a report-first stage commits to the single real case whose findings best match the draft and adopts its consensus reasoning chain, ensuring consistency between the retrieved reasoning graph and the final report. Because RECAP requires no task-specific training and encodes only the test slide at inference, it is inexpensive to deploy in resource-constrained settings and benefits directly from improvements to the underlying foundation models. On a held-out split of the released training data, RECAP attains a Metric~A Reg2026-Score of ${\approx}0.88$.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_016.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=4fsvlWlcVD
BibTex
@InProceedings{BueLuk_Reporting_MICCAISAT2026,
author = { Buess, Lukas AND Maier, Andreas},
title = { { Reporting Without Generating: Retrieval-Based Chain-of-Thought Answering for Pathology } },
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}
}