Enhancing and accelerating brain MRI through deep learning reconstruction using prior subject-specific imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) acquisition is lengthy, impacting cost and patient comfort.
• Deep learning models can leverage prior scans to enhance current MRI reconstruction.
• Integrating prior scans requires time-consuming registration, a bottleneck for clinical use.

Purpose of the Study:

• To develop a novel, efficient deep learning framework for MRI reconstruction using longitudinal data.
• To reduce MRI acquisition and reconstruction times while maintaining or improving image quality.
• To assess the impact of the improved MRI reconstruction on downstream tasks like brain segmentation.

Main Methods:

• A deep learning framework comprising an initial reconstruction network, a deep registration model, and a transformer-based enhancement network was proposed.
• The method was validated on a longitudinal dataset of T1-weighted MRI scans from 18 subjects across multiple acceleration factors.
• Quantitative metrics and downstream brain segmentation accuracy were used for evaluation.

Main Results:

• The proposed deep learning approach demonstrated superior performance over existing methods.
• Significant improvements in reconstruction quality and accuracy were observed.
• The method achieved substantial reductions in total reconstruction time compared to traditional registration techniques.
• Enhanced accuracy and volumetric agreement were noted in the downstream brain segmentation task.

Conclusions:

• The novel deep learning framework offers a faster and more accurate MRI reconstruction solution.
• This approach has the potential for real-time clinical applications due to reduced processing time.
• The method improves both image reconstruction and the performance of subsequent neuroimaging analyses.