ARDGen: Augmentation Regularization for Domain-Generalized Medical Report Generation
Ahsan, Syed Bilal Zaheer, Muhammad Zaigham
Ahsan, Syed Bilal
Zaheer, Muhammad Zaigham
Supervisor
Department
Computer Vision
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Automated medical report generation from chest radiographs is pivotal for clinical decision support, yet existing systems suffer from performance degradation due to domain shifts across diverse imaging sources. In this work, we propose a multi-modal framework that robustly generates clinically relevant diagnostic reports by integrating visual and textual modalities. ARDGen comprises an image classification branch employing a pre-trained ResNet-based encoder with advanced image augmentation and consistency regularization, and a report generation branch featuring BERT-based dual decoders. The primary text decoder produces the diagnostic narrative while our proposed Augmentation Regularization Decoder (ARD), used exclusively during training, serves as a regularizer to enhance the model's adaptability. We further enforce text-level consistency through augmentation-driven losses. Extensive experiments conducted on the MIMIC-CXR and IU-Xray datasets demonstrate that our approach significantly outperforms existing methods, achieving superior generalization and improved report quality on unseen data. This framework offers a robust solution for reliable automated diagnosis, bridging the gap between visual evidence and accurate clinical narratives.
Citation
S. B. Ahsan, M. Ikhalas, M. M. Khan, S. Ullah and M. Z. Zaheer, "ARDGen: Augmentation Regularization for Domain-Generalized Medical Report Generation," 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Nashville, TN, USA, 2025, pp. 6526-6535, doi: 10.1109/CVPRW67362.2025.00649.
Source
IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops
Conference
2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2025
Keywords
Subjects
Source
2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2025
Publisher
IEEE