Segmentation of Brain Tumors Using a Multi-Modal Segment Anything Model (MSAM) with Missing Modality Adaptation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate glioma segmentation is crucial for treatment planning.
• Existing models struggle with multimodal data integration and missing data scenarios.
• Developing robust segmentation models for diverse clinical data is essential.

Purpose of the Study:

• To introduce a novel multi-modal segment anything model (MSAM) for glioma segmentation.
• To enhance brain tumor segmentation accuracy using multimodal MRI and diffusion tensor imaging data.
• To address the challenge of missing data in clinical settings.

Main Methods:

• Developed a multi-modal feature fusion block for integrating diverse imaging data.
• Implemented a missing modality training method to handle incomplete datasets.
• Compared MSAM against U-Net using Dice Similarity Coefficient and 95% Hausdorff Distance metrics.

Main Results:

• MSAM consistently outperformed U-Net across various modality combinations.
• Performance gains were notable when using structural MRI data alone.
• MSAM demonstrated robustness and adaptability to missing data, including smaller tumor regions.

Conclusions:

• MSAM offers a significant advancement in glioma segmentation, outperforming traditional methods.
• The model's ability to handle missing data makes it suitable for real-world clinical applications.
• Further research should explore MSAM's performance on varying tumor sizes and complexities.