Deep Learning-Accelerated 3D T2 CUBE MRI of the Cochlea at 1.5T and 3T: Sonic DL With AIR Recon DL Versus AIR Recon DL Alone.
Authors
Affiliations (2)
Affiliations (2)
- Ear Harvard Medical School (K.L.R., S.H., I.J.), Harvard Medical School, GE HealthCare (E.M., M.T.), Boston, MA, USA; GE HealthCare (A.G.), Waukesha, WI, USA; GE HealthCare (D.R.), Rochester, MN, USA; Wentworth-Douglass Hospital (I.J.), Enterprise Service Group - Radiology, Mass General Brigham, Dover, NH, USA and Department of Radiology (I.J.), University of Turku, Turku, Finland.
- Ear Harvard Medical School (K.L.R., S.H., I.J.), Harvard Medical School, GE HealthCare (E.M., M.T.), Boston, MA, USA; GE HealthCare (A.G.), Waukesha, WI, USA; GE HealthCare (D.R.), Rochester, MN, USA; Wentworth-Douglass Hospital (I.J.), Enterprise Service Group - Radiology, Mass General Brigham, Dover, NH, USA and Department of Radiology (I.J.), University of Turku, Turku, Finland. [email protected].
Abstract
Three-dimensional fast spin-echo MR cisternography (CUBE) of the internal auditory canal (IAC) and cerebellopontine angle (CPA) requires longer acquisition times that can limit clinical throughput. We evaluated whether a deep learning (DL)-based highly accelerated reconstruction technique, Sonic DL`™ (SDL), combined with a commercially available DL denoising reconstruction (AIR™ Recon DL, ARDL), preserves or improves subjective image quality and cochlear anatomy visualization relative to standard CUBE-ARDL while shortening scan time. Consecutive patients underwent brain MRI for suspected IAC pathology at 1.5T and 3T with paired acquisition of standard 3D CUBE-ARDL and an accelerated 3D CUBE-SDL sequence. Both used the same voxel size at 1.5T (0.6x0.6x1.2 mm<sup>3</sup>) and 3T (0.6x0.6x0.6 mm<sup>3</sup>). Scan times for CUBE-ARDL versus CUBE-SDL were 2:33 versus 1:41 min at 1.5T and 1:55 versus 1:11 min at 3T. Two independent neuroradiologists blinded to MRI acquisition method scored randomized, de-identified images on a 5-point Likert scale across five domains (overall image quality, noise, sharpness, artifacts, cochlear anatomy visualization). Reader-averaged scores for CUBE-ARDL and CUBE-SDL were compared with paired Wilcoxon signed-rank tests, with reader-specific and cumulative link mixed model (CLMM) sensitivity analyses and false discovery rate (FDR) adjustment; intraclass correlation coefficients (ICC) quantified inter-reader reliability for each technique; Gwet's AC2 and weighted (Conger's) kappa with linear ordinal weights quantified pooled inter-reader agreement. After exclusion of 5 training cases per field strength, 49 (1.5T) and 34 (3T) patients were analyzed. CUBE-SDL scored significantly better than CUBE-ARDL for all five domains at both field strengths (all FDR-adjusted P < .001), consistently for each reader and in the CLMM. Inter-reader reliability (ICC) for CUBE-SDL was moderate-to-excellent at 3T (0.63-0.92) and moderate at 1.5T (0.65-0.74), generally exceeding CUBE-ARDL (3T: 0.35-0.61; 1.5T: 0.50-0.71). Pooled inter-reader agreement (Gwet's AC2) ranged from 0.56 to 0.89. A combined Sonic DL acceleration with ARDL-denoised reconstruction for 3D T2 CUBE produced significantly higher subjective image quality at both 1.5T and 3T, and, at 3T, higher inter-reader reliability than standard CUBE-ARDL, while shortening acquisition time by 34%-38%. These findings support further evaluation, including diagnostic accuracy studies, of CUBE-SDL as an alternative to CUBE-ARDL.