Related Experiment Video
Updated: Jan 14, 2026
08:20
Author Spotlight: AI-Driven Trypanosome Species Detection from Microscopic Images
Published on: October 27, 2023
2.5K
Improving personalized federated learning to optimize site-specific performance in computer-aided
Aiki Yamada1,2, Shouhei Hanaoka2, Tomomi Takenaga2
1Chiba University, Graduate School of Science and Engineering, Department of Medical Engineering, Chiba, Japan.
Journal of Medical Imaging (Bellingham, Wash.)
|October 24, 2025
Summary
Improved Ditto enhances personalized federated learning (PFL) for medical imaging AI. This method stabilizes model updates and reduces tuning costs for computer-aided detection/diagnosis (CAD) software.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Medical Imaging
Background:
- Personalized federated learning (PFL) addresses data heterogeneity and privacy in AI.
- Ditto, a PFL method, faces challenges with unstable updates and high tuning costs.
- PFL applications in computer-aided detection/diagnosis (CAD) software are under investigation.
Purpose of the Study:
- To introduce Improved Ditto, a novel PFL method.
- To enhance stability and reduce hyperparameter tuning costs in PFL.
- To apply PFL to computer-aided detection/diagnosis (CAD) software.
Main Methods:
- Developed a personalized model update rule for Improved Ditto.
- Dynamically adjusted global model weights based on L2-norm of gradient and global model terms.
- Evaluated Improved Ditto on three CAD tasks: aneurysm detection, metastasis detection, and liver lesion classification.
Main Results:
- Improved Ditto demonstrated competitive performance against Ditto and other federated learning methods in two of three CAD tasks.
- Achieved a reduced hyperparameter search space, lowering tuning costs.
- Enhanced the stability of personalized model updates, indicating improved adaptability.
Conclusions:
- Dynamically adjusting global model weights improves PFL stability and adaptability.
- Improved Ditto offers reduced hyperparameter tuning costs for PFL.
- The proposed method shows potential benefits for computer-aided detection/diagnosis (CAD) software.
