Two-Minute Deep Learning–Powered Brain Quantitative Mapping: Accelerating Clinical Imaging With Synthetic Magnetic Resonance Imaging

Abstract Background Quantitative magnetic resonance imaging (MRI) is an advanced technique that can map the physical properties (T1, T2, and proton density PD) of different tissues, offering crucial insights for disease diagnosis. Nonetheless, the practical application of this technology is indeed constrained by several factors, with the most notable being the protracted scanning duration. Objective This study aimed to explore whether deep learning (DL)–based superresolution reconstruction of ultrafast whole brain synthetic MRI can obtain quantitative T1/T2/PD maps that are closely approximated to those from routine clinical scans, while substantially shortening scan time and preserving diagnostic image quality. Methods A total of 151 healthy adults and 7 individuals with different pathologies were prospectively enrolled. Each individual was examined twice on a 3.0T scanner using routine and fast synthetic MRI protocols. The routine scans (acquisition matrix: 320×256) were interpolated to 512 by 512 for clinical display and served as reference images. The fast scans (acquisition matrix: 192×128) were preprocessed to 256 by 256 and used as inputs to a superresolution generative adversarial network (SRGAN), which reconstructed them to the same 512 by 512 interpolated resolution as the reference. For each quantitative chart, 120 (75.95%) healthy individuals’ images were used for training, and 38 (24.05%) individuals’ images (healthy individuals: n=31, 19.62%; patients: n=7, 4.43%) were used for testing. Agreement was assessed with a paired t test, two 1-sided tests, Bland-Altman analysis, and coefficients of variation. Results DL reconstructed and reference T1/T2/PD values were strongly correlated (T1: R ²=0.98; T2: R ²=0.97; and PD: R ²=0.99). The slopes of the linear regression were near 1.0 both for T1 (0.9418) and PD (0.9946), whereas T2 values were moderate, as the slope of the linear regression was 0.8057. Additionally, the average biases of T1, T2, and PD values were small (0.93%, −0.85%, and 0.31%, respectively). The intra- and intergroup coefficient of variation for most of the brain regions stayed below 5%, especially for PD values, and after DL reconstruction, it still has quantitative accuracy for lesions. Quantitative and qualitative analyses of image quality also indicate that SRGAN markedly suppressed noise and artifacts in fast acquisitions, restoring structural fidelity (structural similarity image measure) and signal fidelity (peak signal-to-noise ratio) close to the level of routine scans while substantially improving perceptual naturalness over fast scans (as measured by the naturalness image quality evaluator), although not yet matching that of routine imaging. Conclusions SRGAN superresolution applied to ultrafast synthetic MRI yields whole brain T1, T2, and PD maps that show strong correlation with routine synthetic MRI while halving acquisition time and maintaining diagnostic image quality. Although T1 and PD values exhibit near-ideal agreement, and T2 values demonstrate a moderate systematic underestimation, this approach represents a promising step toward accelerating clinical deployment of quantitative brain imaging.