MGML: A plug-and-play meta-guided multi-modal learning framework for incomplete multimodal brain tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Multimodal Magnetic Resonance Imaging (MRI) is crucial for brain tumor segmentation.
• Clinical MRI data are often incomplete, hindering effective lesion segmentation.
• Maximizing information from incomplete multimodal MRI is a significant research challenge.

Purpose of the Study:

• To present a novel meta-guided multi-modal learning (MGML) framework for brain tumor segmentation.
• To address the challenge of incomplete multimodal MRI data in clinical practice.
• To enhance the utilization of available multimodal information for improved segmentation.

Main Methods:

• Developed a meta-guided multi-modal learning (MGML) framework with two components: meta-parameterized adaptive modality fusion (Meta-AMF) and consistency regularization.
• Meta-AMF integrates information from multiple modalities adaptively under varying input conditions, generating soft-label supervision signals for coherent fusion.
• Consistency regularization enhances segmentation performance, robustness, and generalization without altering the original model architecture.

Main Results:

• The MGML framework achieved superior performance compared to state-of-the-art methods on the BraTS2020 and BraTS2023 datasets.
• On BraTS2020, average Dice scores across fifteen missing modality combinations were 87.55% (WT), 79.36% (TC), and 62.67% (ET).
• The approach integrates seamlessly into training pipelines for end-to-end optimization.

Conclusions:

• The proposed MGML framework effectively leverages incomplete multimodal MRI data for enhanced brain tumor segmentation.
• The method demonstrates robustness and generalization capabilities, outperforming existing approaches.
• The framework offers a practical solution for utilizing incomplete clinical MRI data, with source code publicly available.