Complex deep neural networks for denoising ultra-fast submillimeter T2*-weighted imaging and quantitative susceptibility mapping

The 2024 revisions of the McDonald diagnostic criteria for multiple sclerosis (MS) have incorporated susceptibility-based magnetic resonance imaging (MRI) biomarkers to improve diagnostic sensitivity and specificity. However, the addition of imaging sequences used to visualize these biomarkers (Time of Acquisition, TA: ~6minutes) increases the overall patient scan time. This study addresses this by combining parallel imaging (TA: ~2minutes) with our proposed deep learning-based image denoising method, complex-valued denoising convolutional neural network ( C DnCNN), to generate high quality data. Network layers used in the denoising convolutional neural network (DnCNN) were extended to the complex domain for learning complex-valued MR image features in the image domain. Four real-valued and complex-valued versions of this network (2D DnCNN, 2D C DnCNN, 3D DnCNN, and 3D C DnCNN) were developed for testing on a simulated noise testing set and a real-world noise testing set. In the simulated noise testing set, 3D C DnCNN outperformed the other approaches for denoising magnitude and complex MRI data across all levels of simulated noise. In the real-world noise testing set, 2D DnCNN yielded the highest increases for NRMSE, PSNR and CNR measures, while 3D DnCNN yielded the highest improvement for SSIM when denoising T2*-weighted magnitude images at the highest acceleration factor. For the complex-valued data, 3D C DnCNN outperformed 2D C DnCNN to produce high quality magnitude and phase data at all acceleration factors in all image quality metrics except for CNR measures. Overall, our study demonstrates the capability of deep learning-based image denoising methods to efficiently denoise ultra-fast submillimeter isotropic data. Our proposed C DnCNN is able to denoise complex-valued MRI data which further enables the generation of high-quality quantitative susceptibility mapping (QSM). Besides that, our experimental results using different denoising models indicate that CNNs should be designed to learn 3-dimensional image features in the complex domain to achieve optimal performance on MRI data. Purpose To develop and evaluate a complex-valued deep learning-based denoising convolutional neural network ( C DnCNN) for improving image quality of ultra-fast submillimeter isotropic T2*-weighted data acquired using a 3-dimensional echo-planar-imaging (3D-EPI) sequence. Materials and Methods This retrospective analysis of a prospective study included 56 participants (female-to-male ratio, 7:1; mean age, 40.5 years) imaged at 3 sites (Cedars-Sinai Medical Center, Cleveland Clinic, University of Pennsylvania). T2*-weighted magnitude and phase images at 0.65mm isotropic resolution were acquired from a 3-dimensional echo-planar-imaging (3D-EPI) sequence, without (Time of Acquisition, TA: ~4minutes) and with CAIPIRINHA undersampling at three acceleration factors: R=2, (TA: ~4minutes), R=3 (TA: ~3minutes) and R=4 (TA: ~2minutes). Four real-valued and complex-valued denoising convolutional neural networks (2D DnCNN, 2D C DnCNN, 3D DnCNN, and 3D C DnCNN) were developed for testing on a simulated noise testing set and a real-world noise testing set. Image quality was evaluated using NRMSE, PSNR, SSIM and CNR measures and statistical testing was done using repeated measures ANOVA, paired t-tests and paired two one-sided t-tests. Results Results from the simulated noise testing set showed that 3D C DnCNN outperformed the other denoising approaches for both magnitude and complex MRI data across all levels of added noise (P < 0.001). In the real-world noise testing set, 2D DnCNN yielded the highest increases for NRMSE (+21%), PSNR (+10%) and CNR measures, while 3D DnCNN yielded the highest improvement for SSIM (+14%) when denoising T2*-weighted magnitude images at the highest acceleration factor (R=4). For the complex-valued data, 3D C DnCNN outperformed 2D C DnCNN at all acceleration factors in all image quality metrics except for CNR measures (P < 0.001). Paired two one-sided t-tests also showed equivalence of the denoised QSM data to the reference QSM data with and without denoising by the complex-valued networks. Conclusion Our study demonstrates the capability of deep learning-based image denoising methods to efficiently denoise ultra-fast submillimeter isotropic data. This allows for generation of high quality T2*-weighted magnitude, phase images, and QSM maps for the detection of MS biomarkers.