Prior-Guided Selective Parameter Fine-Tuning for Source-Free Domain Adaptive Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Source-free domain adaptation (SFDA) is crucial for transferring knowledge to unlabeled medical image domains without accessing private source data.
• Current SFDA methods for medical image segmentation often rely on pseudo-labeling and full model fine-tuning, with limited exploration of underlying principles.
• Generalization error in SFDA is influenced by model complexity and pseudo-label noise, necessitating advanced adaptation strategies.

Purpose of the Study:

• To investigate the principles of SFDA using PAC-Bayesian generalization error bounds.
• To propose a novel selective parameter fine-tuning framework, PATH (PArameter fine-tuning guided by Topological and Historical priors), for SFDA in medical image segmentation.
• To reduce model complexity and pseudo-label noise during adaptation for improved segmentation performance.

Main Methods:

• Investigated SFDA using PAC-Bayesian generalization error bounds to understand constraints from model complexity and pseudo-label noise.
• Developed PATH, a framework that selectively fine-tunes domain-variant and task-distinctive parameters to reduce effective model complexity.
• Integrated topological structure and historical prediction consistency priors to estimate pseudo-label reliability and suppress noise.

Main Results:

• PATH effectively reduces model complexity by sparsely updating identified parameters.
• The framework successfully suppresses pseudo-label noise by leveraging topological and historical priors.
• Achieved state-of-the-art performance on cross-scanner fundus image segmentation and cross-modality abdominal multi-organ segmentation benchmarks.

Conclusions:

• PATH offers a principled approach to SFDA for medical image segmentation by addressing model complexity and pseudo-label noise.
• The selective parameter fine-tuning strategy and noise suppression mechanisms contribute to superior segmentation accuracy.
• The proposed method demonstrates significant improvements over existing SFDA techniques in challenging medical imaging scenarios.