RUN-UP: Accelerated multishot diffusion-weighted MRI reconstruction using an unrolled network with U-Net as priors

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Multishot diffusion-weighted imaging (DWI) is crucial for various clinical applications.
• Accelerating DWI reconstruction is vital for improving patient comfort and throughput.
• Current reconstruction methods can be time-consuming and may compromise image quality.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for accelerating multishot DWI reconstruction.
• To improve image quality and reduce reconstruction time compared to existing techniques.
• To enable near real-time DWI reconstruction for enhanced clinical feasibility.

Main Methods:

• An unrolled deep learning pipeline incorporating model-based gradient updates and neural networks was designed.
• The network was trained using single-direction data to predict jointly reconstructed multidirection data.
• Shot-to-shot phase correction was integrated into the reconstruction pipeline.
• In vivo brain and breast DWI data were used for evaluation.

Main Results:

• The proposed method achieved a reconstruction time of 0.1 second per image, over 100-fold faster than traditional methods.
• Reconstructed image quality was comparable to joint reconstruction, with a PSNR of 35.3 dB, NRMSE of 0.0177, and SSIM of 0.944.
• Significant improvements were observed compared to locally low-rank reconstruction, including higher PSNR and SSIM, and lower NRMSE.
• The method demonstrated generalization from brain to breast DWI data, with fine-tuning reducing artifacts.

Conclusions:

• A data-driven approach enables near real-time multishot DWI reconstruction with high image quality.
• The developed method significantly accelerates reconstruction while maintaining diagnostic image quality.
• This advancement enhances the feasibility of multishot DWI in diverse clinical and neuroscientific studies.