Quantitative magnetic resonance imaging enables the mapping of tissue relaxation times and proton density, offering objective parameters to support clinical assessment. Synthetic MRI allows multiple quantitative maps and contrasts to be generated within a single acquisition, yet whole brain protocols still require several minutes, potentially affecting patient comfort and workflow efficiency. Efforts to accelerate acquisition often compromise image quality or quantitative reliability. A deep learning–based superresolution approach has been evaluated to address this trade-off in whole brain synthetic MRI. By applying a superresolution generative adversarial network to ultrafast acquisitions, quantitative T1, T2 and proton density maps were reconstructed and compared with routine clinical scans. The objective was to determine whether acquisition time could be substantially reduced while preserving quantitative agreement and maintaining diagnostic image quality suitable for clinical neuroimaging.
Accelerated Acquisition and Network Design
A total of 151 healthy adults and 7 individuals with white matter hyperintensities, cerebral infarcts and encephalomalacia were prospectively enrolled. All participants underwent both routine and fast synthetic MRI examinations on a 3.0 T scanner using a multidynamic multiecho sequence. Routine scans required 4 minutes 55 seconds, followed by approximately 1 minute of vendor postprocessing. Fast scans were completed in 1:52 by reducing acquisition matrix size and modifying the acceleration factor, while other parameters remained unchanged.
Must Read:VLMs Lag Behind Neuroradiologists
Quantitative maps from routine scans were reconstructed at 512x512 resolution and served as reference images. Fast scan maps were reconstructed at 256x256 resolution and used as low-resolution inputs to a superresolution generative adversarial network. The generator incorporated convolutional layers and residual blocks with rectified linear unit activation, followed by subpixel convolution for upsampling. The discriminator comprised convolutional layers with Leaky ReLU activation and batch normalisation, producing a probability of real versus generated images. A pretrained VGG-19 network was used for feature extraction within the loss function. Separate models were trained for T1, T2 and proton density maps using identical hyperparameters. Following reconstruction, 512x512 quantitative maps were generated in approximately 1 second, reducing total acquisition and reconstruction time from 5:55 to 1:53.
Quantitative Agreement Across Tissue and Regions
Quantitative accuracy was assessed using paired statistical testing, two one-sided tests for equivalence, linear regression and Bland-Altman analysis. Deep learning–reconstructed T1, T2 and proton density values showed strong correlation with routine scans, with coefficients of determination of 0.98, 0.97 and 0.99, respectively. Linear regression slopes were near unity for T1 and proton density, whereas T2 demonstrated a lower slope of 0.8057. Average percentage biases were small at 0.93% for T1, −0.85% for T2 and 0.31% for proton density.
When clinically acceptable limits were defined as ±5% of routine mean values, the 90% confidence intervals for T1, T2 and proton density in both grey matter and white matter were contained within predefined margins. In grey matter, no significant differences were observed between reconstructed and routine scans for T1 and proton density, although T2 remained significantly different. In white matter, T2 values after reconstruction did not differ significantly from routine scans.
Region-based analysis was performed using anatomical automatic labelling and Johns Hopkins University atlases. Intragroup coefficients of variation were generally low, particularly for proton density, and intergroup coefficients of variation were lower than intragroup values across all regions. Significant differences in T2 coefficients of variation were observed in several grey matter regions, including frontal, temporal, parietal and occipital cortices, insula and hippocampus. In lesion regions, quantitative values reconstructed by deep learning did not differ significantly from routine scans for T1, T2 or proton density.
Image Quality and Clinical Implications
Objective image quality assessment included peak signal-to-noise ratio and structural similarity index as full-reference metrics, alongside the naturalness image quality evaluator as a no-reference measure. Reconstructed maps demonstrated preservation of quantitative accuracy and consistently outperformed fast acquisitions across objective metrics. For T1, T2 and proton density maps, structural similarity values were 0.82, 0.93 and 0.99, respectively, with corresponding peak signal-to-noise ratios of 21.94, 29.14 and 58.01.
Naturalness scores showed substantial improvement over fast scans, indicating reduced distortion, although routine scans retained slightly lower scores, reflecting residual perceptual differences. Visual inspection confirmed restoration of structural fidelity and suppression of noise and artefacts relative to fast acquisitions.
Analysis of lesion regions indicated that white matter hyperintensities and acute cerebral infarcts exhibited similar quantitative profiles, whereas encephalomalacia demonstrated markedly higher T1 values approaching those of cerebrospinal fluid. Although no significant quantitative differences were detected between reconstructed and routine lesion measurements, several reconstructed T2 values were lower than reference values in white matter hyperintensities and infarcts, indicating a degree of systematic underestimation.
Deep learning–based superresolution applied to ultrafast synthetic MRI enables whole brain T1, T2 and proton density mapping with acquisition time reduced to under 2 minutes. Reconstructed T1 and proton density values demonstrate excellent agreement with routine scans, while T2 values show moderate systematic underestimation. Strong correlations, low average biases and preserved lesion quantification support the feasibility of accelerated quantitative neuroimaging. Image quality metrics confirm substantial improvement over fast acquisitions, with near-routine structural fidelity. These findings indicate that superresolution reconstruction represents a promising approach to accelerating clinical deployment of quantitative brain MRI while maintaining diagnostic reliability.
Source: JMIR Medical Informatics
Image Credit: iStock
References:
Liu Y, Yin H, Zheng Z et al. (2026) Two-Minute Deep Learning–Powered Brain Quantitative Mapping: Accelerating Clinical Imaging With Synthetic Magnetic Resonance Imaging. JMIR Med Inform;14:e79389.
