Related Experiment Video
Updated: May 25, 2026
02:09
Multi-modal Pulmonary Imaging: Using Complementary Information from CT and Hyperpolarized 129Xe MRI to Evaluate Lung Structure-Function
Published on: April 12, 2024
Source-free domain adaptation for multi-institutional chest X-ray images
Hyoyi Kim1, Seoyoung Lee2, Seungryong Cho1
1Department of Nuclear and Quantum Engineering, KAIST, Yuseong-gu, Daejeon, South Korea.
Journal of Applied Clinical Medical Physics
|May 23, 2026
Summary
New source-free domain adaptation methods improve tuberculosis (TB) detection from chest X-rays (CXRs) across different hospitals without data sharing. These privacy-preserving techniques enhance model generalization and performance, even with imbalanced datasets.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Public Health
Background:
- Machine learning models for chest X-ray (CXR)-based tuberculosis (TB) detection face performance degradation across institutions due to domain shifts.
- Privacy regulations restrict data sharing, hindering conventional domain adaptation approaches for TB detection models.
Purpose of the Study:
- To enhance TB detection performance in a multi-institutional setting without accessing source data or statistics.
- To address the challenges of source-free domain adaptation (SFDA) specifically for imbalanced binary classification tasks in TB detection.
Main Methods:
- Proposing two SFDA methods: Source HypOthesis Transfer (SHOT) and SFDA via source Distribution Estimation (SFDA-DE).
- Implementing ranking-based pseudo-labeling and selective fine-tuning of network layers to manage class imbalance and prevent overfitting.
Main Results:
- Achieved significant improvements in average F1 score and AUROC on six unseen target domains with high class imbalance.
- Demonstrated effective adaptation without requiring access to source data during the adaptation process.
Conclusions:
- The proposed SFDA methods provide a privacy-preserving and effective strategy for improving the generalization of TB detection models.
- Successfully addressed domain shifts in multi-institutional TB detection while adhering to data-sharing restrictions.
