A 3D Cross-Modality Feature Interaction Network With Volumetric Feature Alignment for Brain Tumor and Tissue...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Neuroscience computational methods

Background:

• Accurate volumetric segmentation of brain tumors and tissues from multi-modal Magnetic Resonance (MR) images is crucial for quantitative analysis and disease identification.
• Existing 3D Fully Convolutional Networks (3D FCNs) struggle with complex cross-modality relationships and feature misalignment, limiting their effectiveness.
• 3D FCNs efficiently model local image features but are less adept at capturing global spatial context in volumetric data.

Purpose of the Study:

• To introduce the Aligned Cross-Modality Interaction Network (ACMINet) for enhanced segmentation of brain tumors and tissues in multi-modal MR images.
• To address limitations in current 3D FCNs regarding cross-modality fusion, feature alignment, and global context modeling.
• To improve the accuracy and robustness of automated segmentation for various brain pathologies.

Main Methods:

• Development of an Aligned Cross-Modality Interaction Network (ACMINet) incorporating specialized modules.
• Implementation of a cross-modality feature interaction module for adaptive fusion and refinement of multi-modal features.
• Introduction of a volumetric feature alignment module using a learnable deformation field for dynamic feature alignment.
• Integration of a volumetric dual interaction graph reasoning module for graph-based global context modeling in spatial and channel dimensions.

Main Results:

• ACMINet demonstrated state-of-the-art segmentation performance across four benchmark datasets: BraTS2018, BraTS2020, Vestibular Schwannoma, and iSeg-2017.
• The method achieved the highest Dice Similarity Coefficient (DSC) score for the hard-segmented enhanced tumor region on the BraTS2020 validation leaderboard.
• Experimental results confirm the network's efficacy in segmenting brain glioma, vestibular schwannoma, and general brain tissues.

Conclusions:

• The proposed ACMINet effectively overcomes the limitations of traditional 3D FCNs in multi-modal brain image segmentation.
• ACMINet's integrated approach to cross-modality interaction, feature alignment, and global context modeling significantly enhances segmentation accuracy.
• The network represents a significant advancement in automated brain tumor and tissue segmentation, offering superior performance on diverse clinical datasets.