Abstract
Generating a pathologist-style reasoning chain from a wholeslide image (WSI) is usually framed as free-form generation. We show that, for the REG2026 Pathologist Reasoning-Guided Report Generation Challenge, it is better framed as classification followed by deterministic retrieval. An analysis of the 11,220-case training corpus finds the reasoning target to be highly templated: only 92 canonical questions and 189 directed edges occur, and conditioning on the pair (organ, primary diagnosis) yields ≈86% reasoning-graph purity and ≈92% answer purity. An oracle study bounds the achievable workflow score at 0.889 given perfect fields, and localises essentially all remaining error in fine-grained diagnosis. We therefore reduce the task to predicting (organ, diagnosis) from the slide and emitting the corresponding template graph, answers and report. Tiles are encoded by two frozen pathology foundation models whose 512- d and 1536-d embeddings are concatenated and aggregated by a gatedattention multiple-instance-learning head. On the official test phase, our system scored 0.7707 overall, placing in the top ten; a single-encoder variant scored 0.7449. The gain came entirely from better diagnosis: every diagnosis-driven component of the score rose, while the grounding metrics stayed fixed. We also tried LoRA adaptation, full encoder unfreezing, a third foundation encoder and seed ensembling, and none of them moved the score beyond noise. What limits the system is therefore the diagnostic accuracy of the frozen features, not the capacity of the aggregation head. Code is available at https://github.com/ujjwalbaid0408/REG2026.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_018.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=Cg9Gg9Gx7I
BibTex
@InProceedings{BaiUjj_Reasoning_MICCAISAT2026,
author = { Baid, Ujjwal AND Pathak, Tilak AND Baheti, Bhakti AND Madabhushi, Anant},
title = { { Reasoning as Retrieval: Template-Grounded Chain-of-Thought Report Generation 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}
}