Deep learning based multiplexed sensitivity-encoding (DL-MUSE) for high-resolution multi-shot DWI.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence in Medicine

Background:

• Multi-shot echo-planar imaging (MSH-EPI) diffusion-weighted imaging (DWI) is crucial for high-resolution imaging.
• Conventional phase correction methods struggle with acquisition imperfections, leading to unreliable phase estimation and artifacts.
• Deep learning offers a potential solution for more robust phase correction in MSH-EPI DWI.

Purpose of the Study:

• To propose and validate a deep learning-based phase correction method for high-resolution MSH-EPI DWI.
• To assess the efficacy and generalization capability of the proposed method in healthy volunteers and patients.

Main Methods:

• A deep learning multiplexed sensitivity-encoding (DL-MUSE) framework utilizing a convolutional neural network (CNN) was developed for phase estimation.
• A dual-channel U-net architecture was employed for phase estimation in MSH-EPI.
• The network was trained on aliasing-free single-shot (SSH) DW images and tested on MSH-DWI datasets from healthy subjects and patients with varying shot numbers.

Main Results:

• DL-MUSE demonstrated high and robust performance in correcting inter-shot phase errors.
• Compared to conventional MUSE, DL-MUSE reduced distortion, noise, and signal loss in high b-value DWIs, with improvements increasing with shot number.
• The method enhanced white matter orientation consistency and fiber delineation compared to SSH-EPI, and showed preliminary generalizability across different acquisition parameters and sites.

Conclusions:

• A deep learning-based reconstruction algorithm (DL-MUSE) was successfully developed for MSH-EPI DWI.
• The method significantly improves image quality and demonstrates feasibility and generalizability in both healthy volunteers and patients.
• DL-MUSE holds promise for routine clinical applications and neuroimaging research.