No modality left behind: Adapting to missing modalities via knowledge distillation for brain tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor segmentation is critical for treatment planning.
• Multi-modal MRI provides complementary tumor information.
• Missing MRI modalities challenge current deep learning segmentation models.

Purpose of the Study:

• To develop a robust multi-modal brain tumor segmentation framework for missing modality scenarios.
• To enhance the adaptability and generalizability of segmentation models under data scarcity.
• To address limitations of existing methods relying on complete multi-modal MRI inputs.

Main Methods:

• Proposed AdaMM: a multi-modal segmentation framework using knowledge distillation.
• Incorporated Graph-guided Adaptive Refinement, Bi-Bottleneck Distillation, and Lesion-Presence-Guided Reliability modules.
• Evaluated knowledge distillation and missing-modality strategies.

Main Results:

• AdaMM demonstrated superior segmentation accuracy and robustness compared to existing methods.
• The framework performed exceptionally well in single-modality and weak-modality configurations.
• Systematic evaluation confirmed the superiority of knowledge distillation for missing modalities.

Conclusions:

• AdaMM offers a robust solution for brain tumor segmentation with missing multi-modal MRI data.
• Knowledge distillation is a highly effective strategy for handling missing modalities.
• The study provides practical guidance for future research in medical image segmentation.