Federated Spatial Prior-Based Source-Free Domain Adaptation for White Matter Hyperintensities Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• White matter hyperintensities (WMH) are key indicators of cerebral small vessel disease.
• Accurate WMH segmentation is vital for brain health assessment and diagnosis.
• Cross-domain segmentation faces challenges due to data privacy and limited labels.

Purpose of the Study:

• To develop a robust and privacy-preserving framework for automatic WMH segmentation.
• To enhance generalization and accuracy in cross-domain WMH segmentation.
• To improve the detection of small WMH lesions and boundary delineation.

Main Methods:

• A source-free domain adaptation (SFDA) framework was developed, incorporating federated spatial prior modeling.
• A dual-path pseudo-label generator leveraged spatial priors for improved boundary accuracy and small lesion detection.
• Federated learning optimized spatial priors across sites without raw data sharing, followed by pseudo-label fine-tuning.

Main Results:

• The proposed method outperformed state-of-the-art UDA and SFDA techniques, showing 3-10% DSC improvement across multiple datasets.
• Demonstrated superior performance in detecting small lesions and accurately delineating WMH boundaries.
• Achieved robust generalization and maintained data privacy.

Conclusions:

• The developed framework offers a privacy-preserving and effective solution for WMH segmentation.
• This approach significantly supports the early diagnosis and risk assessment of cerebrovascular diseases.
• Federated spatial prior modeling enhances model generalization in challenging clinical settings.