Clinical Validation of Deep Learning–Accelerated versus Wave-CAIPI Postcontrast 3D T1-MPRAGE for Evaluation of Intracranial Enhancing Lesions

BACKGROUND AND PURPOSE: Deep learning (DL) reconstruction methods have shown promise in accelerating 2D MRI sequences but have yet to be extensively validated for routine 3D volumetric MRI applications. Our purpose was to assess the diagnostic quality of a novel DL-accelerated 3D T1-MPRAGE compared with a state-of-the-art wave-controlled aliasing in parallel imaging (Wave-CAIPI) accelerated 3D T1-MPRAGE for evaluating intracranial enhancing lesions. MATERIALS AND METHODS: This prospective study was approved by the Institutional Review Board. Patients undergoing contrast-enhanced brain MRI in an outpatient setting were scanned on 3T MRI systems. The imaging protocol included a state-of-the-art Wave-CAIPI postcontrast T1-MPRAGE (acceleration factor [R] = 2 × 2, acquisition time [TA] = 2:11 minutes) and a research-based postcontrast DL-T1-MPRAGE (R = 2 × 2, TA = 2:11 minutes). The DL-based reconstruction process involved 2 steps. The first step, inspired by variational networks, involved 6 iterations alternating between data consistency updates and neural network evaluation. The second step applied a super-resolution algorithm for further image enhancement. Two independent neuroradiologists conducted a blinded, randomized head-to-head comparison of the 2 sequences by using a previously published scale across the following criteria: visualization of dural, parenchymal, leptomeningeal, and ependymal enhancement; sharpness; noise; artifacts; and overall diagnostic quality. A third board-certified neuroradiologist adjudicated cases with discrepant ratings. Noninferiority of DL-T1-MPRAGE was tested by using a 15% margin. RESULTS: A total of 115 patients (68 women/47 men, mean age = 54 ± 10 years) were included. The top 3 clinical indications were: neoplasm (52%), vascular lesions (24%), and headache (8%). DL-T1-MPRAGE was noninferior to Wave-CAIPI T1-MPRAGE for delineating enhancing lesions with unanimous agreement in all cases with enhancing pathology. It was also noninferior in terms of noise perception (P < .0001), artifact (P < .0001), sharpness (P = .001), and overall diagnostic quality (P < .0001). DL-T1-MPRAGE provided equivalent visualization of small, subtle enhancing parenchymal, dural, and leptomeningeal lesions. CONCLUSIONS: The highly accelerated postcontrast DL-T1-MPRAGE demonstrated noninferior image quality compared with the clinically validated Wave-CAIPI T1-MPRAGE accelerated sequence while offering enhanced visualization of subtle enhancing lesions.