MCA-GAN: A lightweight Multi-scale Context-Aware Generative Adversarial Network for MRI reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) offers high-resolution, non-invasive medical imaging but suffers from long acquisition times.
• Traditional compressed sensing (CS) methods for MRI acceleration degrade reconstruction quality at high undersampling rates.
• Existing deep learning models (CNNs, GANs) struggle with long-range dependencies and high computational complexity, limiting practical MRI acceleration.

Purpose of the Study:

• To develop a lightweight deep learning model for accelerated MRI reconstruction with improved quality and efficiency.
• To address the limitations of local receptive fields and high computational costs in current deep learning-based MRI acceleration methods.

Main Methods:

• Proposed a lightweight Multi-scale Context-Aware Generative Adversarial Network (MCA-GAN) utilizing dual-domain generators for k-space and image-space optimization.
• Integrated novel lightweight modules: Depthwise Separable Local Attention (DWLA), Adaptive Group Rearrangement Block (AGRB), Multi-Scale Spatial Context Modulation Bridge (MSCMB), and Channel-Spatial Multi-Scale Self-Attention (CSMS).
• Employed extensive experiments on IXI, MICCAI 2013, and MRNet datasets to evaluate reconstruction performance (PSNR, SSIM) and efficiency.

Main Results:

• MCA-GAN consistently outperformed existing methods in PSNR and SSIM across multiple datasets.
• Achieved a 27.3% reduction in parameter size and 19.6% reduction in computational complexity compared to SepGAN, with the shortest reconstruction time.
• Demonstrated robust performance across various undersampling masks and acceleration rates, with strong cross-dataset generalization capabilities.

Conclusions:

• MCA-GAN offers an efficient and accurate solution for accelerated MRI by enhancing reconstruction quality and significantly reducing computational cost.
• The lightweight architecture and multi-scale feature fusion effectively capture global context and improve reconstruction performance.
• The model's strong generalization potential makes it suitable for diverse clinical MRI scenarios requiring accelerated acquisition.