Reconstruct incomplete relation for incomplete modality brain tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Multi-modal magnetic resonance imaging (MRI) enhances brain tumor segmentation.
• Clinical settings often lack complete multi-modal MRI data, hindering segmentation performance.
• Existing knowledge transfer methods overlook the structural nature of brain tumor segmentation.

Purpose of the Study:

• To develop a Reconstruct Incomplete Relation Network (RIRN) for brain tumor segmentation with incomplete MRI data.
• To transfer voxel semantic relational knowledge from a full-modality teacher network to an incomplete-modality student network.
• To improve segmentation accuracy by addressing missing tumor-specific information.

Main Methods:

• Proposed RIRN to transfer voxel semantic relational knowledge.
• Introduced Class-relative relations (CRR) and Class-agnostic relations (CAR) to incorporate structural knowledge.
• Utilized adversarial learning to align holistic structural predictions between teacher and student networks.

Main Results:

• RIRN effectively transfers voxel semantic relational knowledge.
• CRR and CAR successfully incorporated structural information into segmentation.
• Adversarial learning aligned structural predictions, enhancing performance.
• Achieved superior performance over state-of-the-art methods on BraTS 2018 and BraTS 2020 datasets.

Conclusions:

• The proposed RIRN method significantly improves brain tumor segmentation with incomplete multi-modal MRI data.
• Incorporating voxel semantic relations and adversarial learning is crucial for handling missing modalities.
• RIRN offers a promising solution for robust brain tumor segmentation in real-world clinical scenarios.