Dual consistent pseudo label generation for multi-source domain adaptation without source data for medical image...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation (UDA) enables models to generalize to new domains without labeled target data.
• Medical image segmentation faces challenges due to diverse data distributions and privacy concerns limiting data sharing.
• Existing UDA methods often require access to source data, which is not feasible for sensitive medical information.

Purpose of the Study:

• To propose a novel multi-source and source-free (MSSF) domain adaptation framework for medical image segmentation.
• To develop a method that adapts models using only source domain segmentation models, without accessing source data.
• To address privacy issues associated with medical image data sharing.

Main Methods:

• Introduced a dual consistency constraint (domain-intra and domain-inter) for high-quality pseudo-label generation.
• Developed a progressive entropy loss minimization to enhance feature consistency across domains.
• Implemented a framework that trains without direct access to source medical images.

Main Results:

• Achieved state-of-the-art performance in retinal vessel segmentation under MSSF conditions.
• Demonstrated the highest sensitivity metric compared to existing methods by a significant margin.
• Validated the effectiveness of the proposed dual consistency and entropy minimization techniques.

Conclusions:

• The proposed MSSF framework is effective for medical image segmentation, particularly retinal vessel segmentation.
• This approach successfully overcomes the limitations of data accessibility and privacy in medical AI.
• Future work should focus on balancing high sensitivity with overall accuracy.