Source-Free Active Domain Adaptation for Brain Tumor Segmentation via Mamba and Region-Level Uncertainty.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate brain tumor segmentation from MRI is vital for diagnosis but hindered by domain shifts, data privacy, and high annotation costs.
• Existing source-free active domain adaptation (SFADA) methods often neglect the structural complexity of tumor regions.

Purpose of the Study:

• To develop a novel SFADA framework that addresses limitations in current brain tumor segmentation techniques.
• To improve segmentation accuracy and boundary delineation while minimizing annotation requirements and respecting data privacy.

Main Methods:

• Proposed a Region-level Uncertainty-Guided Sample Selection (RUGS) strategy for efficient target-domain sample identification.
• Introduced SFADA-Net, a Mamba-driven model with dual-path multi-kernel convolution and a structure-aware prompted Mamba module for enhanced feature extraction.

Main Results:

• Demonstrated superior adaptability and consistent high segmentation accuracy across diverse datasets (BraTS-2021, BraTS-SSA, BraTS-PED, BraTS-MEN 2023).
• Achieved state-of-the-art performance with only a 5% annotation budget, approaching fully supervised learning accuracy.
• Showcased robust segmentation across domains, outperforming existing methods.

Conclusions:

• The framework achieves high accuracy in brain tumor segmentation and boundary delineation with limited annotations.
• Effectively mitigates domain shift and ensures data privacy, reducing manual annotation burdens.
• Accelerates the deployment of accurate diagnostic tools for cross-center clinical application.