Representation learning for ultra-low-field brain MRI super-resolution

Motivation: High, isotropic resolution is desirable for lesion detection and biomarkers extraction for cognitive disorders. However, ULF MRI severely suffers from low spatial resolution and SNR. Goal(s): To introduce a new deep learning (DL) architecture for improving image quality of noisy, low-resolution 3D ULF data. Approach: Implemented new DL super-resolution method with UNet-based encoder and a novel decoder based on KAN to learn and decode concise representation of noisy, low-resolution 3D ULF isotropic images. Results: Proposed method improves resolution, suppresses noise and artifacts, and produces images similar to those obtained from high-field MRI scanners in terms of overall appearance. Impact: Enhancing image resolution and fidelity for ULF brain imaging at 0.055T using data-driven 3D deep learning approach. Potentially enable portable and point-of-care diagnosis.