Using Synthetic Training Data In Neural Networks for the Estimation of Fiber Orientation Distribution Functions From Single Shell Data

Several studies have investigated the possibility of predicting the fiber orientation distribution function (fODF), which is obtained using the very accurate multi-shell multi-tissue constrained spherical deconvolution (MT-CSD) from single-shell or low angular resolution multi-shell diffusion magnetic resonance imaging (dMRI) data using deep learning. While all these approaches show promising results, the vast majority have in common that they require multi-shell high angular resolution diffusion imaging (HARDI) data to calculate the ground truth fODF using the MT-CSD for training their networks. This data, however, is difficult to acquire in a clinical context and it is yet unclear how well networks trained on data acquired on a certain scanner with a certain protocol would generalize to different data. In this work, we address these shortcomings and present a method that can estimate an accurate fODF from single-shell diffusion data without the need for multi-shell data for training. This is achieved by generating patient-, acquisition- and scanner-specific synthetic single voxel diffusion signals with a known ground truth fODF from single shell data that can be used to train the neural network. The trained network will then be applied to the real patient data to predict the fODF with a quality standard close to that of an MT-CSD and the ability to determine if white matter (WM) is present in the underlying voxel. The approach is evaluated on 20 subjects from the Human Connectome Project (HCP) for all three shells (b=1000, 2000, 3000 s/mm^2). When comparing both this approach and a single shell CSD to the results of the MT-CSD, this work outperforms the single shell CSD in terms of the angular correlation coefficient and root mean squared error on all three shells.