Consistently denoising 3D MR images using 2D neural networks

Motivation: Denoising of 3-dimensional MR images with 2D CNNs introduces unwanted inter-slice fluctuations. High-dimensional CNNs do not introduce these artifacts but place high demands on the underlying computer hardware, which might not be met on today's preclinical systems. Goal(s): Investigating a method for consistent denoising of high-dimensional datasets using 2-dimensional convolutional neural networks. Approach: Denoising of a high-dimensional dataset is divided into multiple 2D processing steps. Data resulting from these steps is then weighted and combined in k-space. Results: A large 3-dimensional dataset was denoised on a desktop workstation using 2D convolutional neural networks, avoiding inter-slice fluctuations in the final dataset. Impact: The presented method enables consistent high-dimensional denoising using 2-dimensional convolutional networks. Thus, processing of large MRI datasets becomes possible on standard workstations without the need of expensive computer hardware or connection to a remote server.