Background
Compressed sensing (CS) is available for clinical 1.5T scanners with limited reduction factor due to a small channel of receiver coil. Typically, low-resolution (LR) with CS is able to further reduce scan time. However, LR images may be insufficient for clinical diagnosis. Recently, Deep Learning (DL) approaches have demonstrated the ability to generate high-resolution (HR) MR images from LR images. Thus, this study investigated the possibility of using LR-CS and DL-based super-resolution technique for brain volume measurement (BVM) application at 1.5T MRI.
Materials and Methods
For model training, human brain volumes acquired with 3D-TFE-T1W were incorporated, in which LR images were generated from the original HR images with 2x sub-sampled strategy. After data augmentation, pairs of LR and HR images were used for training 3D Residual Dense Net (RDN). For model testing, LR CS-3D-TFE-T1W images were acquired using 1.5T MRI with one minute scan time. Normalized Root-Mean-Square Error (NRMSE), Peak Signal to Noise Ratio (PSNR), and Structural Similarity (SSIM) were used for model evaluation. BVMs were performed using Freesurfer’s software. Wilcoxon signed rank test, Pearson’s correlation, and effect size were used for statistical analysis.
Results
The results showed that DL-SR model is able to synthesize HR images from LR images, in which no significant differences between DL-SR and actual HR (p < 0.01) reported by NRMSE (0.051 vs 0.059), PSNR (25.885 vs 24.679), SSIM (0.961 vs 0.951). For volumetric assessments, there were no significant differences between DL-SR and actual HR images (p > 0.01, Pearson’s correlation > 0.90) at seven subcortical regions.
Conclusions
The combination of LR CS-MRI and DL-SR can effectively address the issue of prolonged scan time in 3D MRI scans while preserving the image quality and the accuracy of brain volume measurements.