MedMAP: Promoting Incomplete Multi-Modal Brain Tumor Segmentation With Alignment.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation commonly relies on multi-modal Magnetic Resonance Imaging (MRI).
• Clinical practice often faces scenarios with missing MRI modalities, posing significant segmentation challenges.
• Existing methods like Knowledge Distillation and Domain Adaptation struggle to bridge modality gaps and learn invariant features.

Purpose of the Study:

• To address the limitations of current methods in handling missing MRI modalities for brain tumor segmentation.
• To propose a novel training paradigm that effectively aligns latent features across different MRI modalities.
• To theoretically certify the effectiveness of the proposed alignment paradigm.

Main Methods:

• Proposed a novel paradigm aligning latent features of involved MRI modalities to a distribution anchor.
• Utilized this alignment as a substitute for pre-trained models, which are scarce in brain tumor segmentation.
• Theoretically proved that the training paradigm ensures a tight evidence lower bound for effectiveness.

Main Results:

• The proposed paradigm enables the learning of invariant feature representations across different MRI modalities.
• Demonstrated a significant narrowing of modality gaps in brain tumor segmentation models.
• Achieved superior performance on BraTS2018, BraTS2020, and Brain Metastasis datasets.

Conclusions:

• The novel alignment paradigm effectively mitigates performance degradation caused by missing MRI modalities.
• The method provides a robust solution for brain tumor segmentation in challenging clinical scenarios.
• The approach offers a promising direction for developing more generalized and accurate medical image segmentation models.