Learning Modality-Aware Representations: Adaptive Group-Wise Interaction Network for Multimodal MRI Synthesis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Multimodal Magnetic Resonance Imaging (MRI) synthesis aims to generate missing image modalities from available ones.
• Current image-to-image translation methods struggle with precise feature alignment across modalities, leading to suboptimal synthesis.
• Challenges include effectively fusing information and mapping features across different MRI sequences.

Purpose of the Study:

• To propose an Adaptive Group-wise Interaction Network (AGI-Net) for improved multimodal MR image synthesis.
• To explicitly model inter-modality and intra-modality relationships for enhanced feature and semantic alignment.
• To improve the representational capacity and fusion capabilities in multimodal MRI synthesis.

Main Methods:

• Developed AGI-Net, partitioning feature channels into groups and applying an adaptive rolling mechanism to convolutional kernels.
• Introduced a cross-group attention module for effective feature fusion across groups.
• Validated the network on the IXI and BraTS2023 datasets for multimodal MR image synthesis.

Main Results:

• AGI-Net achieved state-of-the-art performance in multimodal MR image synthesis tasks.
• Demonstrated superior feature and semantic correspondence capture between different MRI modalities.
• Confirmed the effectiveness of the proposed modality-aware interaction design.

Conclusions:

• AGI-Net significantly advances multimodal MR image synthesis by addressing feature alignment challenges.
• The proposed network architecture effectively models complex relationships within and across MRI modalities.
• The approach offers a robust solution for generating high-quality synthetic MRI data.