Joint Cross-Attention Network With Deep Modality Prior for Fast MRI Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accelerated Magnetic Resonance Imaging (MRI) reconstruction using deep learning often relies on Convolutional Neural Networks (CNNs) for subsampled k-space data.
• Existing models face limitations in coil sensitivity estimation, structural prior utilization, and CNN inductive bias, hindering performance.

Purpose of the Study:

• To propose an unrolling-based joint Cross-Attention Network (jCAN) for accelerated multi-coil MRI reconstruction.
• To enhance the accuracy of coil sensitivity map (SM) estimation and improve the model's representation capabilities.
• To leverage intra-subject data as a reference modality for guided reconstruction.

Main Methods:

• jCAN simultaneously optimizes the latent MR image and sensitivity map (SM), incorporating Gating and Gaussian layers to refine SM estimation.
• The model integrates Vision Transformer (ViT) in the image domain and CNN in the k-space domain for enhanced representation.
• A self- and cross-attention mechanism utilizes pre-acquired intra-subject scans to guide the reconstruction of subsampled target modalities.

Main Results:

• jCAN demonstrated superior performance compared to state-of-the-art methods on public knee and in-house brain datasets.
• Significant improvements were observed in Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) across various acceleration factors and sampling masks.
• The proposed method effectively addresses challenges in coil sensitivity estimation and representation learning for accelerated MRI.

Conclusions:

• The developed jCAN model offers a robust and effective solution for accelerated multi-coil MRI reconstruction.
• The joint optimization strategy and the integration of ViT and attention mechanisms significantly advance MRI reconstruction quality.
• The publicly available code facilitates further research and application in the field of fast MRI.