CHARMS: A CNN-Transformer Hybrid with Attention Regularization for MRI Super-Resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning-based Magnetic Resonance Imaging (MRI) super-resolution (SR) offers high-resolution reconstruction but often requires large, computationally intensive models.
• These demanding models limit practical applications in resource-constrained environments like portable MRI scanners and real-time clinical workflows.

Purpose of the Study:

• To introduce CHARMS, a novel, lightweight convolutional-Transformer hybrid model optimized for MRI super-resolution.
• To achieve high-fidelity image reconstruction with reduced computational cost and inference time for broader clinical accessibility.

Main Methods:

• CHARMS utilizes a Reverse Residual Attention Fusion backbone for local feature extraction and Pixel-Channel/Enhanced Spatial Attention for feature calibration.
• A Multi-Depthwise Dilated Transformer Attention block models long-range dependencies efficiently.
• Novel attention regularization techniques were employed to stabilize training and improve generalization.

Main Results:

• CHARMS demonstrated superior performance over leading lightweight models, achieving 0.1-0.6 dB PSNR and up to 1% SSIM gains at ×2/×4 upscaling.
• The model achieved approximately 40% reduction in inference time compared to existing methods.
• Cross-field fine-tuning enabled 7T-like image quality from 3T MRI data, showing significant PSNR and SSIM improvements.

Conclusions:

• CHARMS provides a significant advancement in MRI super-resolution, balancing high fidelity with computational efficiency.
• Its lightweight design and near-real-time performance make it suitable for clinical workflows, accelerated MRI protocols, and portable MRI systems.
• The model enhances the accessibility and applicability of advanced MRI techniques across various clinical and research settings.