Latent Correlation Representation Learning for Brain Tumor Segmentation With Missing MRI Modalities.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor segmentation from Magnetic Resonance Imaging (MRI) is crucial for clinical diagnostics and treatment planning.
• Multi-modal MRI data offers complementary information for enhanced segmentation accuracy.
• Clinical practice often involves missing imaging modalities, posing a challenge for existing segmentation algorithms.

Purpose of the Study:

• To develop a novel brain tumor segmentation algorithm robust to missing modalities in multi-modal MRI.
• To leverage the inherent correlations between multi-modal MRI data to improve segmentation performance.
• To enhance the reliability of brain tumor segmentation in scenarios with incomplete imaging data.

Main Methods:

• A novel correlation model is proposed to represent latent multi-source correlations between imaging modalities.
• Individual representations from each modality encoder are used to estimate modality-independent parameters.
• An attention mechanism fuses cross-modal correlation representations into a shared representation for segmentation.

Main Results:

• The proposed algorithm demonstrates robust performance even when one or more MRI modalities are missing.
• The method outperforms current state-of-the-art techniques on the BraTS 2018 and BraTS 2019 datasets.
• The correlation model effectively captures and utilizes inter-modality relationships for improved segmentation.

Conclusions:

• The developed algorithm provides a robust solution for brain tumor segmentation with missing modalities.
• Leveraging multi-modal correlations significantly enhances segmentation accuracy and reliability.
• This approach has the potential to improve clinical diagnostics and treatment planning in real-world settings.