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

TL;DR

The paper tackles the trade-off between fidelity and efficiency in MRI super-resolution by introducing Efficient Vision Mamba, a 2D-slice SR framework built on multi-head selective state-space modeling (MHSSM) and a lightweight Channel MLP. It employs a hybrid scanning strategy to capture long-range dependencies while maintaining low parameter count and FLOPs. Across 7T brain T1 MP2RAGE and 1.5T prostate T2w data, the method achieves superior quantitative and perceptual metrics with significantly reduced computational demand compared to strong baselines. The approach shows strong potential for clinical translation, though it remains limited to 2D processing and would benefit from broader validation and uncertainty quantification in future work.

Abstract

Background: High-resolution MRI is critical for diagnosis, but long acquisition times limit clinical use. Super-resolution (SR) can enhance resolution post-scan, yet existing deep learning methods face fidelity-efficiency trade-offs. Purpose: To develop a computationally efficient and accurate deep learning framework for MRI SR that preserves anatomical detail for clinical integration. Materials and Methods: We propose a novel SR framework combining multi-head selective state-space models (MHSSM) with a lightweight channel MLP. The model uses 2D patch extraction with hybrid scanning to capture long-range dependencies. Each MambaFormer block integrates MHSSM, depthwise convolutions, and gated channel mixing. Evaluation used 7T brain T1 MP2RAGE maps (n=142) and 1.5T prostate T2w MRI (n=334). Comparisons included Bicubic interpolation, GANs (CycleGAN, Pix2pix, SPSR), transformers (SwinIR), Mamba (MambaIR), and diffusion models (I2SB, Res-SRDiff). Results: Our model achieved superior performance with exceptional efficiency. For 7T brain data: SSIM=0.951+-0.021, PSNR=26.90+-1.41 dB, LPIPS=0.076+-0.022, GMSD=0.083+-0.017, significantly outperforming all baselines (p<0.001). For prostate data: SSIM=0.770+-0.049, PSNR=27.15+-2.19 dB, LPIPS=0.190+-0.095, GMSD=0.087+-0.013. The framework used only 0.9M parameters and 57 GFLOPs, reducing parameters by 99.8% and computation by 97.5% versus Res-SRDiff, while outperforming SwinIR and MambaIR in accuracy and efficiency. Conclusion: The proposed framework provides an efficient, accurate MRI SR solution, delivering enhanced anatomical detail across datasets. Its low computational demand and state-of-the-art performance show strong potential for clinical translation.

Figures (6)

• Figure 1: Selective scanning strategies in Vision Mamba. (a) Horizontal and vertical scanning may separate central pixels from their diagonal neighbors. (b) Diagonal scanning preserves spatial adjacency. (c) Dense extraction increases parameter overhead. (d) Our efficient variant uses depthwise convolutions for preprocessing.
• Figure 2: Proposed Vision Mamba framework for MRI super-resolution. (a) Hybrid selective scanning extracts patch sequences. (b) Each MambaFormer block integrates LN, MHSSM, and Channel MLP with residual connections. (c) Multi-head selective state-space modeling runs $m$ parallel scans with adaptive step sizes.
• Figure 3: Qualitative comparison of super-resolution methods on 7T brain T1 MP2RAGE maps. The first row shows reconstructed images, the second row is the zoomed-in regions, and the last row depicts difference maps with respect to the ground-truth HR image. Arrows highlight cortical and subcortical structures (white: cortical ribbon; black: caudate nucleus and putamen; red: subtle tissue boundaries).
• Figure 4: Quantitative evaluation of super-resolution methods on 7T brain T1 MP2RAGE maps. Bar plots show mean $\pm$ standard deviation of SSIM, PSNR, LPIPS, and GMSD across the test cohort. Higher SSIM/PSNR and lower LPIPS/GMSD indicate better performance.
• Figure 5: Visual comparison of super-resolution approaches on a representative axial prostate T2w MRI slice. Top panels display the ground truth image together with outputs from competing reconstruction methods. The second row provides magnified views from prostate and bladder regions of interest, chosen to emphasize structural detail. Colored arrows indicate specific aspects under evaluation: green arrows indicate edge delineation of fine tissue, white arrows indicate continuity of anatomical features, and red arrows indicate areas sensitive to artifacts. The bottom panels show voxel-wise difference maps relative to the ground truth.