Related Experiment Video
Updated: Jul 1, 2026
05:47
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
RE-LIG: A Faithfulness-Driven Layer Integrated Gradients Framework for Explainable Medical Visual Question Answering
Esra Balık1, İrfan Aygün2, Mehmet Kaya3
1Department of Software Engineering, Bandırma Onyedi Eylül University, Bandırma, Türkiye.
Journal of Imaging Informatics in Medicine
|June 29, 2026
Summary
This study introduces a high-resolution framework (RE-LIG) to improve the explainability and reliability of Medical Visual Question Answering (Med-VQA) systems, enhancing clinical decision-making.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Explainable AI
Background:
- Medical Visual Question Answering (Med-VQA) systems aid clinical decisions but suffer from limited transparency and low-resolution issues.
- Existing 'black-box' models hinder clinical applicability due to a lack of explainability.
Purpose of the Study:
- To propose a high-resolution holistic framework, Robust and Efficient Layer-Integrated Gradients (RE-LIG), to enhance reliability and explainability in Med-VQA.
- To improve the transparency and trustworthiness of Med-VQA systems for clinical use.
Main Methods:
- Developed a high-resolution framework integrating PubMedCLIP for visual encoding and BioLinkBERT for semantic fusion via coattention.
- Incorporated the RE-LIG algorithm, combining noise tunneling and layer-based integration, to overcome limitations of traditional gradient-based explainability methods.
- Utilized dynamic positional embedding interpolation for high-resolution visual encoding and a coattention mechanism for multimodal semantic fusion.
Main Results:
- The RE-LIG framework significantly increased model faithfulness, achieving +28.9% higher explanation fidelity compared to standard gradient approaches (RE-LIG AOPC=0.3180 vs. Vanilla IG=0.2467).
- The system maintained competitive performance with state-of-the-art models, showing 80.77% overall accuracy, 87.61% closed-ended, and 77.34% open-ended accuracy.
- Ablation studies confirmed that noise reduction mechanisms improved focus on pathological boundaries, demonstrating explainability as a verifiable requirement.
Conclusions:
- The proposed RE-LIG framework enhances explainability and reliability in Med-VQA systems without compromising diagnostic performance.
- Explainability in Med-VQA is a measurable and verifiable requirement, crucial for clinical confidence and decision-making.
- The RE-LIG framework offers a promising solution for transparent and trustworthy AI in medical image interpretation.
