Self-Supervised Super-Resolution of 2D Pre-clinical MRI Acquisitions

Area of Science:

• Biomedical Imaging
• Veterinary Medicine
• Artificial Intelligence in Medicine

Background:

• Animal models are crucial for disease research and therapeutic development.
• Magnetic Resonance Imaging (MRI) facilitates longitudinal studies in animal models, but 2D acquisitions are limited by resolution and scan time.
• Current 3D MRI techniques face challenges with extended scan durations and animal sedation.

Purpose of the Study:

• To evaluate SMORE, a self-supervised deep learning super-resolution method, for enhancing through-plane resolution in anisotropic 2D animal MRI scans.
• To achieve isotropic resolutions from existing 2D MRI data without external training sets.
• To assess the potential of pre-training to reduce processing time.

Main Methods:

• SMORE utilizes self-supervised learning, training on high-resolution in-plane data to infer through-plane information.
• The method was tested on mouse MRI scans with varying through-plane resolutions.
• Performance was compared against traditional interpolation techniques.

Main Results:

• SMORE significantly outperformed traditional interpolation methods in enhancing MRI resolution.
• The self-supervised approach successfully generated isotropic resolutions from anisotropic 2D scans.
• Pre-training SMORE demonstrated a reduction in processing time without sacrificing performance.

Conclusions:

• SMORE is an effective deep learning approach for improving the resolution of animal MRI scans.
• This technology can enhance the translational value of animal research by providing higher quality imaging data.
• Further development, including pre-training strategies, can optimize processing efficiency for broader clinical application.