Abstract
Accurate breast ultrasound diagnosis benefits from integrating multiple complementary image types. Training multimodal large language models (MLLMs) for this setting is challenging because breast ultrasound reports rarely contain step-by-step diagnostic reasoning, and collecting such rationales at scale requires substantial expert effort. Here we present BUThink, a three-stage framework for guideline-based reasoning while reducing reliance on manually curated expert explanations. First, we automatically convert clinical guidelines into executable logical rules and use them to synthesize stepwise diagnostic rationales, forming chain-of-thought (CoT) style supervision. Second, we introduce an adaptive prompting scheme for supervised fine-tuning that conditions the model on multiparametric ultrasound inputs, maps modality-specific observations to clinical questions, and resolves conflicting cues by prioritizing higher-risk interpretations. Third, we apply Group Relative Policy Optimization (GRPO) with rule-based rewards to refine diagnostic accuracy and report quality without relying on human preference labels. We train BUThink on 49,145 images, and evaluate it on a private cohort and the public dataset US3M. This framework achieves SOTA performance on US3M, with 92.34% accuracy and 95.65% recall. By replacing manual rationale annotation with guideline-derived supervision and rule-guided optimization, BUThink provides a practical approach for building more reliable medical MLLMs.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ASMUS_027.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/ASMUS_027_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=NfvfdVz7Rt
BibTex
@InProceedings{XuChe_BUThink_MICCAISAT2026,
author = { Xu, Chenqi AND Lv, Ang AND Liu, Yanbo AND Liu, Yong AND Liu, Ting AND Sun, Guanghao AND Cong, Longfei AND Hou, Rui AND An, Xing},
title = { { BUThink: Aligning Multiparametric Breast Ultrasound with Clinical Expert Logic via Rule-Constrained GRPO } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}