A Foundation Model for Lesion Segmentation on Brain MRI With Mixture of Modality Experts.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neurological Disease Research

Background:

• Brain lesion segmentation is vital for diagnosing and researching neurological conditions.
• Current methods are often task-specific, requiring predefined lesion types and imaging modalities, limiting real-world application.
• A universal approach is needed to handle diverse brain lesions and MRI modalities efficiently.

Purpose of the Study:

• To develop a universal foundation model for brain lesion segmentation on MRI.
• To enable automatic segmentation of various brain lesion types across different MRI modalities.
• To overcome the limitations of task-specific segmentation models.

Main Methods:

• Proposed a novel Mixture of Modality Experts (MoME) framework with specialized expert networks for different MRI modalities.
• Introduced a hierarchical gating network for combining expert predictions and fostering collaboration.
• Implemented a curriculum learning strategy to maintain expert network specialization.
• Developed MoME+ with a soft dispatch network for flexible input modality routing.

Main Results:

• The proposed MoME and MoME+ models were evaluated on nine diverse brain lesion datasets (five modalities, eight lesion types).
• Achieved superior performance compared to existing state-of-the-art universal brain lesion segmentation models.
• Demonstrated promising generalization capabilities on previously unseen datasets.

Conclusions:

• The universal foundation model effectively segments various brain lesions across multiple MRI modalities.
• The MoME framework offers a robust and adaptable solution for brain lesion segmentation.
• This approach simplifies deployment and enhances the utility of AI in neurological diagnostics and research.