Knowledge Bank

MRI uses magnetic fields and radiofrequency pulses to generate high-contrast soft tissue images. Acquisition process in MRI is inherently slow, leading to noisy images or unreasonably long acquisition times. Low-field MRI, which is becoming increasingly common, in particular, produces noisy images. Certain iterative image reconstruction methods such as compressed sensing (CS) can suppress noise but are inherently slow. Deep learning (DL) models using convolutional neural networks (CNN) have proven more effective than CS in improving the quality of medical images. Supervised DL methods use a large dataset of clean and noisy image pairs to train the CNN to remove noise from images. Supervised models suffer from lack of generalizability and a need for large training datasets. In self-supervised methods, in contrast, the entire training can be performed using the original noisy images, requiring no training sets other than the images to be denoised. While several denoisers have been identified for different medical imaging applications, there is no comparative study for self-supervised DL denoisers for MRI. In this study, we have searched rapidly evolving literature on self-supervised denoisers and identified candidates that can benefit MRI. Using a small set of existing MRI data from the fastMRI dataset, we systematically optimized hyperparameters within the deep learning models. Then, we extensively evaluated the performance of the optimized models on a larger subset of fastMRI images using metrics such as peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), structural similarity index measure (SSIM), and training time. Through this work, we show that self-supervised deep learning methods can outperform wavelet based denoising, a traditional image denoising method.