Efficient vision mamba for MRI super-resolution via hybrid selective scanning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• High-resolution MRI is crucial for diagnostics but limited by long scan times.
• Existing deep learning super-resolution (SR) methods face trade-offs between accuracy and computational cost.
• This limits their clinical applicability and integration into workflows.

Purpose of the Study:

• To develop an efficient and accurate deep learning framework for MRI SR.
• The goal is to preserve fine anatomical details while minimizing computational overhead.
• Enable practical integration into clinical imaging workflows.

Main Methods:

• A novel SR framework utilizing multi-head selective state-space models (MHSSM) and a lightweight channel MLP.
• Employs 2D patch extraction with hybrid scanning strategies for capturing long-range dependencies.
• Evaluated on 7T brain and 1.5T prostate MRI datasets, compared against multiple baseline methods.

Main Results:

• The proposed model achieved superior performance in SSIM, PSNR, LPIPS, and GMSD across both datasets.
• Demonstrated exceptional computational efficiency with significantly fewer parameters and GFLOPs than advanced methods.
• Statistically significant improvements were observed over all compared baselines.

Conclusions:

• The developed framework offers a computationally efficient yet accurate solution for MRI SR.
• It delivers well-defined anatomical details and improved perceptual fidelity.
• The model shows strong potential for clinical translation and scalable integration into imaging workflows.