Abstract
Structured pathology report generation emits, from a whole-slide image, a chain-of-thought of (question, answer) steps forming a directed acyclic graph (DAG) that ends in a free-text report. An all-learned generator strong in-distribution collapses at a new institution – its path validity drops over 40% – because a frozen encoder’s decision thresholds do not transfer across sites; on our data institution identity is linearly decodable from the very embeddings those thresholds sit on (AUROC 1.000, nine organ-site pairs). We decouple structure from content. The structure (question DAG and answer grammar) is data-mined from the training ground truth and rebuilt by a deterministic per-organ router; it is institution-invariant by construction and, under oracle recognition, reproduces the ground-truth CoT edge-for-edge on every held-out case (path validity 1.000, 1,665/1,665). The content is read by a single frozen CONCH-v1 vision-language encoder with no training: each attribute is a zero-shot text-prototype margin between disease anchors written in diagnostic language, so a node is tuned by editing a prompt, not by retraining. On the challenge’s blind new-institution test the system scores Overall 0.7646 (Metric A 0.685), against 0.7522 for the all-learned stack it replaced on the same test. Fixing the structure lets the residual error be located, and three metrics with unrelated failure modes agree on where: the first ground-truth edge a failing case omits is always a recognition node, never a construction step; reports whose recognized diagnosis is correct score 0.9910, so 87% of the lost text score is diagnosis rather than phrasing; and the 43 missed cancers concentrate in grade – prostate sensitivity 0.706 against 1.000 in three organs. With structure and report assembly both accounted for, the encoder is the only remaining constraint.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/REG2026_002.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/REG2026_002_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=RDx435LpoO
BibTex
@InProceedings{NguHoa_Decoupling_MICCAISAT2026,
author = { Nguyen, Hoang-Vinh},
title = { { Decoupling Structure from Perception: A Deterministic-DAG Report Generator with Leak-Fair Cross-Site Recognition for Computational Pathology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}
