Super-resolution head and neck MRA using deep machine learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Nonenhanced MR angiography (MRA) is crucial for head and neck imaging.
• Improving spatial resolution in MRA can enhance diagnostic accuracy.
• Deep learning offers potential for image reconstruction advancements.

Purpose of the Study:

• To evaluate the feasibility of deep learning-based super-resolution (SR) for nonenhanced head and neck MRA.
• To assess if SR reconstruction can maintain or improve image quality compared to lower-resolution data.

Main Methods:

• High-resolution 3D quiescent interval slice-selective (QISS) MRA data were acquired.
• Ground-truth data were downsampled to create lower-resolution inputs.
• Four deep neural network (DNN) models, including U-Net architectures, were used for SR reconstruction.
• Quantitative metrics (DSC, SSIM, diameter, sharpness) and qualitative review by neuroradiologists were employed.

Main Results:

• DNN-based SR significantly improved image quality metrics (DSC, SSIM, diameter, sharpness) compared to lower-resolution inputs.
• Three-dimensional (3D) DNN SR models outperformed 2D models.
• Neuroradiologists confirmed consistent image quality improvements with 3D DNN SR.

Conclusions:

• Deep learning-based SR is a feasible method for head and neck MRA.
• This technique allows for potential reduction in MRA acquisition time (up to fourfold) without compromising spatial resolution.
• DNN-based SR holds promise for efficient and high-quality nonenhanced MRA of the head and neck.