Purpose : Quantitative synthetic MRI enables simultaneous quantification of T1 and T2 relaxation times and proton density, as well as acquisition of various contrast-weighted images1). However, MR angiography (MRA) scans could not be obtained using synthetic MRI. Thus, the present study aimed to develop a deep-learning algorithm for synthetic MRI to acquire MRA scans without additional scanning time.
Materials and Methods : Ten healthy volunteers and three patients with intracranial aneurysms were included in this study. All participants underwent the time-of-flight (TOF) MRA sequence and three-dimensional (3D) synthetic MRI sequence, namely 3D-quantification using an interleaved Look-Locker acquisition sequence with T2 preparation pulse (3D-QALAS)2). The five raw images of 3D-QALAS were used as inputs to a deep-learning network. The proposed network was carefully designed to minimize the risk for generating false objects; each pixel value of the output image was generated by a combination of each corresponding pixel of the input images. A simple linear combination model was prepared for comparison3). Subsequently, a three-fold cross-validation was performed. Peak signal-to-noise ratio and normalized root mean square error were calculated for deep-learning MRA (DL-MRA) and linear-combination MRA (linear-MRA) against TOF-MRA and compared between DL-MRA and linear-MRA using the non-parametric Wilcoxon signed-rank test. Overall image quality and branch visualization on a 5-point Likert scale were evaluated by a neuroradiologist blinded to the type of sequence. These visual scores were compared between DL-MRA, linear-MRA, and TOF-MRA using the pairwise Dunn-Bonferroni post-hoc test.
Results : DL-MRA scans were successfully obtained in all participants (Figs. 2, 3). The peak signal-to-noise ratio and normalized root mean square error were significantly higher and lower, respectively, in DL-MRA and linear-MRA (both Ps<.05). The overall image quality score of DL-MRA was 4.6±0.4. Branch visualizations were comparable between DL-MRA and TOF-MRA. Linear-MRA failed to depict intracranial aneurysms, whereas DL-MRA successfully depicted intracranial aneurysms in all three patients (Fig. 4).
Conclusion : Deep learning for 3D synthetic MRI enables visualization of major intracranial arteries as efficiently as does TOF-MRA, with coaligned quantitative maps and various contrast-weighted images.
