LiteMamba-Synth: lightweight state space models for efficient 3T-to-7T MRI translation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• High acquisition costs and limited availability of 7T MRI hinder clinical utility.
• Current deep learning models for 3T-to-7T MRI synthesis are computationally intensive and prone to overfitting.
• There is a need for efficient and scalable solutions for advanced MRI synthesis.

Purpose of the Study:

• To propose LiteMamba-Synth, a streamlined deep learning framework for efficient high-fidelity 7T MRI synthesis from 3T data.
• To reduce computational redundancy and the risk of overfitting in MRI translation models.
• To enable the deployment of advanced MRI synthesis in resource-constrained clinical settings.

Main Methods:

• Developed LiteMamba-Synth, a state space framework incorporating the ConvMamba block for efficient spatial dependency capture.
• Integrated Wavelet-Enhanced Skip connections (WES) for multi-scale frequency-domain feature fusion to preserve anatomical details.
• Employed a lightweight Convolutional Block Attention Module (CBAM) for adaptive feature recalibration.

Main Results:

• LiteMamba-Synth achieved a Peak Signal-to-Noise Ratio (PSNR) of 20.82 dB and a Structural Similarity Index Measure (SSIM) of 0.711 on the UNC T1w dataset.
• The model has a compact footprint of only 2.15 million parameters, significantly less than existing baselines.
• Demonstrated high-fidelity MRI translation with substantial reductions in model complexity.

Conclusions:

• LiteMamba-Synth provides a practical and scalable solution for 7T MRI synthesis.
• The model's efficiency and high-fidelity results make advanced MRI accessible in resource-limited environments.
• This framework addresses the limitations of current deep learning approaches in medical imaging.