Utility of deep learning reconstruction to reach the magic triangle in brain MRI

Abstract Background Brain magnetic resonance imaging (MRI) has become the mainstay diagnostic tool for most brain pathologies. However, the relatively long acquisition time restricts its utilization, especially in emergency cases. Deep learning (DL) reconstruction technology has the ability to reduce acquisition times of various MR sequences, while preserving sufficient clinical MR image quality. We aimed to assess the value of implementing DL brain MR image reconstruction for achieving an optimal balance of the three key elements: time, resolution, and signal-to-noise ratio (SNR), and ultimately realizing the brain MRI "magic triangle." Methods This retrospective study included two groups: one underwent brain MRI before implementing DL software using the conventional MR examination, and the other underwent brain MRI using DL reconstruction. Quantitative assessment comparing the two groups included SNR and total scan time. Two readers evaluated MR image quality using a five-point Likert scale. Inter-reader agreement was assessed using the Kappa test. Results DL-reconstructed brain MR images showed significantly higher SNRs than standard and original images across all sequences ( p < 0.005), with DL-reconstructed FLAIR images having the highest SNRs (281.17 ± 92.71, p < 0.001). In addition, DL-reconstructed brain MR images showed significantly higher qualitative image quality across all assessed elements compared to standard and original images ( p < 0.001), with almost perfect inter-observer agreement ( κ > 0.7). Conclusions Our study demonstrated that novel DL-based reconstruction tool significantly reduces scan time without compromising image quality. This supports the potential for reliable and efficient integration of DL reconstruction into clinical practice.