Spatial-channel relation learning for brain tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Brain tumor segmentation from MRI data faces challenges due to ambiguous patterns and semantic gaps in feature fusion.
• Current methods struggle to fully leverage spatial and channel similarities and their inter-correlations for improved volumetric segmentation.

Purpose of the Study:

• To design a mechanism that fully utilizes spatial and channel similarities and their correlations to enhance brain tumor segmentation.
• To address the limitations of linear fusion and semantic gaps in multi-modal medical imaging data.

Main Methods:

• A revised cascade structure network incorporating a context exploitation module.
• Dual attention mechanisms within subnetworks to learn spatial and channel information.
• Space interaction learning to model relationships between spatial and channel features.

Main Results:

• Improved Dice Coefficient (DC) by 2.1% for whole tumor (WT), 2.0% for tumor core (TC), and 1.4% for enhancing tumor (ET) on the BraTS19 dataset.
• Achieved competitive results compared to state-of-the-art brain tumor segmentation approaches.
• Demonstrated robustness across different modalities.

Conclusions:

• Context exploitation effectively models dependencies in semantic features and bridges the semantic gap in multi-modal data.
• The proposed method enhances brain tumor segmentation by leveraging intraspace and interspace relations.
• The approach shows robustness and improved performance in volumetric segmentation tasks.