Deep Learning-Based Acceleration in MRI: Current Landscape and Clinical Applications in Neuroradiology

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Magnetic resonance imaging (MRI) is crucial for neuroimaging due to its soft-tissue contrast.
• Long MRI acquisition times limit clinical use, causing artifacts, discomfort, and higher costs.
• Traditional acceleration methods like parallel imaging (PI) and compressed sensing (CS) can reduce scan times but may compromise image quality.

Purpose of the Study:

• To review the applications of deep learning-based image reconstruction (DLBIR) in neuroimaging.
• To explore vendor-driven implementations and emerging trends in accelerated MRI.
• To address the challenges and barriers to widespread adoption of DLBIR.

Main Methods:

• Review of current DLBIR techniques and their evolution from 2D to 3D acquisitions.
• Discussion of advancements in self-supervised learning for MRI reconstruction.
• Analysis of vendor-specific DLBIR implementations and their claimed benefits.

Main Results:

• DLBIR offers significant reductions in MRI scan times (up to 85%) while potentially enhancing image quality, noise suppression, and diagnostic accuracy.
• DLBIR techniques show promise in maintaining or improving lesion conspicuity in neuroimaging.
• Advancements in DLBIR are expanding its clinical applicability in neuroimaging.

Conclusions:

• DLBIR presents a significant advancement for accelerated neuroimaging, offering improved efficiency and image quality.
• Challenges include ensuring generalizability across scanners, mitigating artifacts, and validating vendor-specific algorithms.
• Increased transparency and clinical validation are needed to foster end-user trust and adoption of DLBIR in clinical practice.