Accelerated Diffusion-Weighted MR Image Reconstruction Using Deep Neural Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Under-sampling in diffusion-weighted imaging (DWI) reduces scan time but introduces artifacts.
• Artifacts include off-resonance effects, geometric distortions, and susceptibility issues.
• Efficient and accurate reconstruction of DWI is crucial for clinical applications.

Purpose of the Study:

• To develop a deep learning-based method (DWI U-Net) for artifact-free diffusion-weighted MR image reconstruction.
• To evaluate the performance of DWI U-Net against conventional Compressed Sensing (CS) reconstruction.
• To identify the optimal optimizer for the DWI U-Net model.

Main Methods:

• Diffusion-weighted MR image reconstruction using a deep learning model named DWI U-Net.
• Training and testing DWI U-Net on variable density highly under-sampled k-space data.
• Comparison with conventional Compressed Sensing (CS) reconstruction.
• Evaluation of optimizers: RMSProp, Adam, Adagrad, and Adadelta.

Main Results:

• DWI U-Net demonstrated significant improvements over conventional CS reconstruction.
• Improvements observed in mean artifact power (AP), mean structural similarity index measure (SSIM), and mean root mean square error (RMSE).
• At an acceleration factor of 6, DWI U-Net with RMSProp showed up to 61.1% improvement in mean AP and 51.4% improvement in mean RMSE compared to CS.
• RMSProp was identified as the best-performing optimizer for DWI U-Net.

Conclusions:

• DWI U-Net is a superior method for reconstructing artifact-free diffusion-weighted MR images from under-sampled data.
• The deep learning approach effectively mitigates under-sampling artifacts.
• RMSProp is the recommended optimizer for DWI U-Net to achieve optimal reconstruction quality.