Privacy-Preserving Multi-Source Domain Adaptation for Medical Data.

Area of Science:

• Medical imaging analysis
• Deep learning in healthcare
• Neuroscience data science

Background:

• Deep learning significantly advances medical disease diagnosis, but requires large datasets.
• Collecting and labeling extensive medical data is costly and time-consuming for single institutions.
• Sharing data across institutions is hindered by distribution heterogeneity and patient privacy concerns.

Purpose of the Study:

• To propose a novel multi-source, source-free domain adaptation method.
• To address challenges of heterogeneous data distribution and patient privacy in medical data sharing.
• To enable effective deep learning model training using distributed, multi-institutional datasets without direct data transfer.

Main Methods:

• Developed a multi-source source-free domain adaptation technique.
• The method transfers pre-trained source models instead of raw data to align heterogeneous datasets.
• Evaluated on the Autism Brain Imaging Data Exchange (ABIDE) fMRI database and Camelyon17 dataset.

Main Results:

• Achieved an average accuracy of 69.37% on the ABIDE dataset.
• Demonstrated effectiveness in addressing data distribution heterogeneity.
• Validated generalization capabilities and network resource-saving advantages on Camelyon17.

Conclusions:

• The proposed method offers a privacy-preserving and efficient solution for multi-institutional medical data analysis.
• Facilitates collaborative deep learning in healthcare by overcoming data sharing barriers.
• Shows promise for improving diagnostic accuracy and enabling broader application of AI in medicine.