Abstract
The REG2026 challenge asks a system, given a gigapixel whole-slide image (WSI), to emit a pathologist-style chain-of-thought reasoning trajectory that terminates in a free-text pathology report, and separately to answer a visual-context question about a tissue region of interest. Rather than generate the trajectory freely, we decouple the task into diagnosis and expression. A per-organ attention-based multiple-instance classifier over frozen pathology foundation-model patch features predicts the diagnosis; the diagnosis then indexes a canonical reasoning trajectory and a visually retrieved report from a training-built library. Against free generation by the same LoRA-adapted language model, this retrieval-based decoder gains 0.076 Metric A (95\ the official metric further shape the submitted system: the reasoning target is a directed acyclic graph (DAG) rather than a linear path, and the multi-step evaluation score is monotone non-decreasing in the emitted edge set. We exploited both with a confidence-gated edge-set augmentation that adds a second candidate’s uncovered edges instead of replacing the trajectory, and whose benefit is conditional on the diagnosis-error rate (break-even p* ≈ 0.35). We further report a practical finding: a cluster of our lowest-scoring cases were not diagnostic failures but input-decoding failures on striped Adobe-Deflate BigTIFF slides that the default WSI reader cannot open, and a bounded tifffile reader decodes all of them. On Test Phase 1 our system attains an overall score of 0.7738 (Metric A 0.6769) and on Test Phase 2 0.7214 (Metric A 0.6021), with a maximal region-metric score of 1.0 (B1 = B2 = B3 = 1.0) in both phases. Held-out analysis localizes the remaining error to diagnosis rather than to expression.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_013.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/REG2026_013_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=tuBdBPVVBS
BibTex
@InProceedings{JanSun_DiagnosisGrounded_MICCAISAT2026,
author = { Jang, Sungyong AND Shin, Dongjin AND Kang, Hyunseo AND Jung, Seung-Won},
title = { { Diagnosis-Grounded Structured Decoding for Pathologist Reasoning-Trajectory and Report Generation } },
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}
}