SMORE: A Self-Supervised Anti-Aliasing and Super-Resolution Algorithm for MRI Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• High-resolution magnetic resonance (MR) images are crucial but acquiring them with high signal-to-noise ratio (SNR) requires long scan times, leading to patient discomfort, increased costs, and motion artifacts.
• Current 2D and 3D MR imaging protocols often compromise through-plane resolution, resulting in anisotropic voxels and aliasing artifacts in 2D acquisitions, degrading overall image quality.

Purpose of the Study:

• To present SMORE, a novel self-supervised deep convolutional neural network (CNN) approach for enhancing MR image resolution and reducing aliasing artifacts.
• To develop a method that utilizes intrinsic image data for training, eliminating the need for external datasets.
• To provide improved image quality for both 2D and 3D MR datasets.

Main Methods:

• Developed a self-supervised super-resolution (SSR) deep CNN for 3D MR image enhancement using volumetric data.
• Implemented a self-supervised anti-aliasing (SAA) deep CNN preceding the SSR CNN for 2D MR image quality restoration, also trained on volumetric data.
• Evaluated the methods on diverse MR datasets, including downsampled and acquired low-resolution images, using quantitative metrics and visual comparisons.

Main Results:

• The SMORE approach demonstrated significant improvements in image resolution and reduction of aliasing artifacts for both 2D and 3D MR images.
• Quantitative metrics and visual assessments confirmed the superiority of the SMORE method compared to existing techniques.
• The self-supervised nature of the training allowed effective learning directly from the acquired MR data.

Conclusions:

• SMORE offers a powerful and data-efficient solution for enhancing MR image quality, addressing limitations of conventional acquisition protocols.
• The self-supervised learning strategy eliminates the need for paired high-resolution/low-resolution training data, simplifying the application of super-resolution and anti-aliasing techniques.
• This CNN-based approach represents a significant advancement in achieving high-resolution MR imaging suitable for clinical and research applications.