Efficient Vision Mamba for MRI Super-Resolution via Hybrid Selective Scanning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• High-resolution Magnetic Resonance Imaging (MRI) is crucial for diagnostics but limited by long acquisition times.
• Deep learning-based super-resolution (SR) methods offer potential but often involve a trade-off between image fidelity and computational efficiency.

Purpose of the Study:

• To develop a computationally efficient and accurate deep learning framework for MRI SR.
• The framework aims to preserve anatomical detail for seamless clinical integration.

Main Methods:

• A novel SR framework integrating multi-head selective state-space models (MHSSM) with a lightweight channel MLP was proposed.
• The model employed 2D patch extraction and hybrid scanning for long-range dependency capture, utilizing MambaFormer blocks with MHSSM, depthwise convolutions, and gated channel mixing.
• Evaluations were conducted on 7T brain T1 MP2RAGE and 1.5T prostate T2w MRI datasets, comparing against Bicubic, GANs, transformers, Mamba, and diffusion models.

Main Results:

• The proposed framework demonstrated superior performance and exceptional efficiency across both brain and prostate datasets.
• Achieved high SSIM and PSNR values (e.g., 0.951 SSIM, 26.90 dB PSNR for brain data), significantly outperforming all compared baselines (p<0.001).
• Required only 0.9M parameters and 57 GFLOPs, representing a 99.8% reduction in parameters and 97.5% reduction in computation compared to Res-SRDiff, while surpassing SwinIR and MambaIR in accuracy and efficiency.

Conclusions:

• The developed framework offers an efficient and accurate solution for MRI SR, enhancing anatomical detail in diverse datasets.
• Its low computational requirements and state-of-the-art performance indicate strong potential for clinical translation and adoption.