An Efficient and Accurate Super-Resolution Approach to Low-Field MRI via U-Net Architecture With Logarithmic Loss and L2 Regularization
Low-field (LF) MRI scanners have the power to revolutionize medical imaging by provid- 27 ing a portable and cheaper alternative to high-field MRI scanners. However, such scanners are usu- 28 ally significantly noisier and lower quality than their high-field counterparts. This prevents them 29 from appealing to global markets. The aim of this paper is to improve the SNR and overall image quality of low-field MRI scans (called super-resolution) to improve diagnostic capability and, as a result, make it more accessible. To address this issue, we propose a Nested U-Net neural network architecture super-resolution algorithm that outperforms previously suggested super-resolution deep learning methods with an average PSNR of 78.83 ± 0.01 and SSIM of 0.9551 ± 0.01. Our ANOVA paired t-test and Post-Hoc Tukey test demonstrate significance with a p-value < 0.0001 and no other network demonstrating significance higher than 0.1. We tested our network on artificial noisy downsampled synthetic data from 1500 T1 weighted MRI images through the dataset called the T1- mix. Four board-certified radiologists scored 25 images (100 image ratings total) on the Likert scale (1-5) assessing overall image quality, anatomical structure, and diagnostic confidence across our architecture and other published works (SR DenseNet, Generator Block, SRCNN, etc.). Our algo- rithm outperformed all other works with the highest MOS, 4.4 ± 0.3. We also introduce a new type of loss function called natural log mean squared error (NLMSE), outperforming MSE, MAE, and MSLE on this specific SR task. Additionally, we ran inference on actual Hyperfine scan images with successful qualitative results using a Generator RRDB block. In conclusion, we present a more ac- curate deep learning method for single image super-resolution applied to low-field MRI via a 45 Nested U-Net architecture.