Improving the accessibility of deep learning-based denoising for accelerated brain MRI using self-supervised learning and/or transfer learning.

The quality of human brain magnetic resonance imaging (MRI) is often compromised by noise. Deep learning-based image denoising methods using convolutional neural networks (CNNs) have proven to be more effective compared to conventional denoising methods. Nonetheless, most CNN-based denoising methods are not accessible in practice because they require high signal-to-noise ratio (SNR) reference data acquired on many subjects for supervising the training of CNNs and may be further hampered by long training time and lack of access to adequate computing resources. This study seeks to address these challenges using transfer learning and/or self-supervised learning. Here, we demonstrate the efficacy of the proposed approaches on denoising highly accelerated (R=3 × 3) three-dimensional T1-weighted MRI data used for brain morphometry. Specifically, an extension of the "Self2Self" denoising method for volumetric brain data with "average masking" (entitled "Self2Self-AM") is proposed to remove the need for additional high-SNR data by training the CNN on the noisy image volume itself. Transfer learning implemented by fine-tuning parameter values of the pre-trained CNN enables supervised denoising using high SNR data from only a single subject with short training time, and substantially reduces the training time of Self2Self-AM. The denoising performance is systematically and quantitatively evaluated and compared in terms of image quality and morphometric quantification accuracy by comparing to reference images acquired with 9-fold longer scan time. Supervised denoising with fine-tuning and Self2Self-AM with subject-specific training achieved the best denoising performance, as quantified by measures of image similarity (structural similarity index measures of 0.943 and 0.934), gray-white boundary sharpness (13.32% and 12.92%), cortical thickness estimation (whole-brain average discrepancy of 0.16 mm and 0.17 mm) and spatial overlap in brain segmentation (Dice coefficients of 0.986 and 0.985, respectively), significantly improving upon the raw images and outperforming conventional benchmark denoising methods. We present these methods as examples of techniques that may help to promote the wider adoption and practical application of CNN-based denoising methods for brain MRI, thereby benefiting a broader range of clinical and neuroscientific applications.