MR image prediction at high field strength from MR images taken at low field strength using multi-to-one translation

Abstract

Patients with implants may need to undergo Magnetic Resonance Imaging (MRI) at lower field strengths to avoid negative impacts from the strong electromagnetic fields. However, the quality of low-field MRI images may be inferior, potentially leading to inaccurate clinical diagnoses. To overcome this limitation, our study proposes a convolutional neural network developed by using U-Net to generate high-field MR images from low-field ones. The proposed model employs multiple MRI contrasts at lower field strength to generate MR images in one or several contrasts at higher field strength. This method overcomes the limitations of previous research which only utilized a single contrast (one contrast-to-one contrast translation) or multiple contrasts including MR image at high (target) field strength as input. After creating a dataset for multi contrast-toone contrast translation, the model was optimized using techniques such as data augmentation and selection of the best model with minimum validation loss. The generated MR images were evaluated using metrics such as Mean Squared Error (MSE), Pearson Correlation Coefficient (Corr), and Peak Signal-to-Noise Ratio (PSNR). The results indicate that for predicting T1- and PD-weighted MR images at high field strength, the average range of MSE and PSNR over the test dataset (1392 images) did not result in improvements compared to one-to-one translation, while Corr shows improvement in PD-weighted MR image prediction. Also, the reported results for the average range of MSE and PSNR suggest improvements in high-field T2-weighted MR image prediction using multi-to-one transla-
tion.