Abstract
Multiple-instance learning (MIL) and pathology vision–language models (VLMs) map a whole-slide image (WSI) to a single slide-level output. A pathologist instead follows an ordered chain of questions, in which each answer determines what is examined next, and the REG2026 challenge makes this chain the prediction target. Neither paradigm supervises the intermediate steps: MIL emits one pooled slide-level prediction, and pathology VLMs emit one free-text response. We present PathChain, which casts diagnostic workflow reproduction as visually conditioned sequence generation. The ordered question–answer–next-question steps and the report form one autoregressive target, so each answer is conditioned on those preceding it. A frozen CONCH encoder turns the slide into a bag of patch features whose size varies with the slide. A Perceiver resampler compresses that bag into 256 visual tokens, its latent queries re-reading the bag at every layer. The decoder therefore always receives 256 tokens, so its cost does not grow with slide size. A LoRA-adapted DeepSeek-R1 decoder emits the chain in its native reasoning format, which the evaluation metric parses back into steps without post-processing. Training only 0.6% of the decoder, PathChain attains a Workflow Reasoning Score of 0.721 on a 223-case hold-out (Edge-F1 0.807, Path Validity 0.565). Most failures are single incorrect transitions rather than chain hallucination.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_006.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=VjOShsUFjD
BibTex
@InProceedings{ZiaKha_PathChain_MICCAISAT2026,
author = { Ziaullah, Khan AND Khalid, Abdul Rehman AND Lal, Shrestha Nischal AND Nheng, Vanchhay AND Mozumder, Md Ariful Islam AND Kim, Hee-Cheol},
title = { { PathChain: Reproducing the Pathologist’s Diagnostic Workflow from 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}
}