MD-SA2: optimizing Segment Anything 2 for multimodal, depth-aware brain tumor segmentation in sub-Saharan populations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Machine learning aids radiologists in medical image segmentation, but challenges remain with complex, multimodal, and variable-quality MRI scans for brain tumor segmentation.
• Existing methods, including conventional machine learning and Segment Anything (SA)-based models, struggle with the intricacies of brain tumor imaging.

Purpose of the Study:

• To address limitations in current brain tumor segmentation techniques for MRI.
• To introduce MD-SA2, an adaptation of Segment Anything 2 (SA2) enhanced with a U-Net aggregator for improved medical image segmentation.

Main Methods:

• Customized and fine-tuned SA2 for enhanced efficiency and accuracy.
• Concatenated multi-modal MRI slices to improve tumor subtype delineation.
• Integrated a lightweight U-Net aggregator with SA2 to incorporate depth awareness.
• Evaluated performance on the 2023 BraTS-Africa dataset, comprising low-resolution MRI from sub-Saharan patients.

Main Results:

• MD-SA2 achieved a tenfold statistically significant improvement over current methods, with a Dice coefficient of 0.7893.
• Demonstrated superior Intersection over Union (IoU) and lower 95% Hausdorff distance metrics.
• An ablation study confirmed the effectiveness of key MD-SA2 components.

Conclusions:

• MD-SA2 shows strong potential for aiding brain tumor diagnosis and treatment planning.
• The approach may help reduce health inequities, particularly in medically underserved areas facing data limitations.