Skip to main navigation Skip to search Skip to main content

Trustworthy medical concept recalibration through logic-aware multimodal evidence fusion

  • Lei Zhao
  • , Chunzheng Zhu*
  • , Zhan Gao
  • , Pengchen Liang
  • , Xiaoyun Liang
  • , Alex Chit-Yuen Cheng
  • , Chi-Man Pun
  • *Corresponding author for this work

Research output: Journal Publications and Reviews › RGC 21 - Publication in refereed journal › peer-review

Abstract

Concept Bottleneck Models (CBMs) enhance interpretable medical image diagnosis by mapping inputs to human-understandable intermediate concepts. However, existing concept intervention methods rely heavily on labor-intensive manual corrections, incurring prohibitive annotation costs, introducing subjective variability, and limiting clinical scalability. To overcome these limitations, we propose ConChecker, an autonomous concept intervention framework that dynamically recalibrates concept assignments while reducing reliance on human intervention. Formulating initial CBM predictions as primary diagnostic hypotheses, ConChecker conducts targeted self-critique anchored in visual evidence. Specifically, it employs vision-language models to extract rich pathological descriptions from medical images and subsequently verifies each concept via Concept Attribute Fidelity (CAF), a mechanism that models the directional relationships among heterogeneous evidence sources to quantify net concept–text support. ConChecker then derives fine-grained soft interventions through Group Verifiable Concept Policy Optimization (GVCPO), significantly improving diagnostic reliability while preserving concept fidelity. Comprehensive evaluations across multiple medical benchmarks demonstrate that ConChecker substantially reduces the need for human-in-the-loop intervention and achieves state-of-the-art performance. Furthermore, its modular plug-and-play design enables seamless integration into standard CBM architectures without structural modifications. © 2026 Elsevier B.V.
Original languageEnglish
Article number104753
Number of pages13
JournalInformation Fusion
Volume139
Issue numberPart A
Online published7 Sept 2026
DOIs
Publication statusOnline published - 7 Sept 2026

Funding

This work was supported in part by the National Natural Science Foundation of China under Grant 62506124, in part by the Science and Technology Development Fund, Macau SAR, under Grant 0193/2023/RIA3, 0079/2025/AFJ and 0052/2026/RIA1, and the University of Macau under Grant MYRG-GRG2026-00159-FIC.

Research Keywords

  • Concept bottleneck models
  • Interpretable diagnosis
  • Multimodal evidence fusion
  • Trustworthy medical AI

Fingerprint

Dive into the research topics of 'Trustworthy medical concept recalibration through logic-aware multimodal evidence fusion'. Together they form a unique fingerprint.

Cite this