Evaluation of a Deep Learning Reconstruction Technique for Accelerated Brain MRI Scans

Background/Objectives: Magnetic resonance imaging (MRI) is a cornerstone of diagnostic imaging, providing exceptional soft tissue contrast and enabling detailed characterization of neurological, vascular, and musculoskeletal diseases. Despite its clinical utility, prolonged acquisition times remain a primary limitation of MRI. Extended scan durations increase patient discomfort, particularly in vulnerable populations, which elevate the risk of motion artifacts that can degrade image quality and necessitate repeat imaging. Recent advances in machine learning, especially deep learning (DL), promise to ameliorate these limitations by enabling high-quality MRI reconstruction from heavily undersampled data, increasing throughput capacity and improving healthcare delivery efficiency. This work describes an FDA-cleared DL algorithm known as DeepFoqus-Accelerate for the reconstruction of deliberately under-sampled magnetic resonance imaging data. Methods: Fully sampled k-space from 408 brain MRI volumes (108 from 18 healthy volunteers and 300 from the public fastMRI dataset) was retrospectively undersampled. A detailed comparison of reconstructed images using DeepFoqus-Accelerate with corresponding fully sampled clinical images is presented using expert evaluation of overall image quality and quantitative similarity metrics. The only conventional comparator included in the study was GE’s parallel imaging reconstruction (ARC) at matched retrospective undersampling; comparisons with compressed sensing and vendor-specific deep-learning reconstruction methods were outside the study design. Results: This algorithm can allow for a four-fold reduction in brain imaging times relative to a fully sampled acquisition while preserving image quality. No reader rated any reconstruction below the acceptability threshold, and DeepFoqus-Accelerate exceeded ARC on all metrics at 4×. Conclusions: The findings characterize image quality and structural fidelity relative to fully sampled references and support further evaluation of diagnostic performance, with the potential to improve imaging efficiency and patient experience.