Enhancing Diffusion Signal Augmentation Using Spherical Convolutions

Abstract

The application of deep learning in the field of diffusion imaging is becoming increasingly popular. However, correlations of acquired adjacent gradient directions are often ignored. To make use of this information in a neural network, a spherical convolution is necessary. This work evaluates three different ways to include spherical information: 2D projection, local spherical convolution and Fourier space transform. For comparison, all models are designed to have a similar amount of trainable parameters as well as the same network architecture, and are evaluated by considering the example of signal augmentation. Overall, all models achieved comparable good results, improving the reconstruction performance, compared to a reconstruction without augmentation, by ≈ 30% for the fractional anisotropy, ≈ 50% for the mean diffusivity, ≈ 70% for the mean signal kurtosis and ≈ 5% for the diffusion signal itself. Particularly, in comparison to a regular neural network that does not implement a spherical convolution, the average performance for all models that implement a sperical convolution increases slightly for all evaluated measures, where the local spherical convolution shows the most favorable results.