M-MSSEU: source-free domain adaptation for multi-modal stroke lesion segmentation using shadowed sets and evidential...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Unsupervised domain adaptation faces challenges due to unavailable source data.
• Source-free domain adaptation (SFDA) is crucial for leveraging pre-trained models without source data.
• Multi-modal medical imaging offers rich information for stroke lesion segmentation.

Purpose of the Study:

• To develop a novel SFDA method for multi-modal stroke lesion segmentation.
• To effectively utilize multi-modal information within the SFDA framework.
• To improve the reliability and accuracy of segmentation using evidential deep learning.

Main Methods:

• A multi-modal opinion fusion module employing Dempster-Shafer evidence theory for cross-modality decision fusion.
• Pseudo-label learning utilizing a pre-trained source model for domain adaptation.
• Pseudo-label filtering via shadowed sets theory and refinement using evidential uncertainty to enhance label quality.

Main Results:

• The proposed method demonstrated superior performance compared to existing state-of-the-art SFDA techniques.
• Experimental validation on two multi-modal stroke lesion datasets confirmed the method's effectiveness.
• The developed schemes for pseudo-label quality improvement were efficient and effective.

Conclusions:

• The novel SFDA method effectively segments multi-modal stroke lesions.
• Evidential deep learning and advanced pseudo-labeling strategies significantly improve adaptation performance.
• This work offers a promising direction for medical image analysis in unsupervised settings.