A generalized deep learning network for fractional anisotropy reconstruction: Application to epilepsy and multiple sclerosis

Abstract

Fractional anisotropy (FA) is a quantitative map sensitive to microstructural properties of tissues in vivo and it is extensively used to study the healthy and pathological brain. This map is classically calculated by model fitting (standard method) and requires many diffusion weighted (DW) images for data quality and unbiased readings, hence needing the acquisition time of several minutes. Here, we adapted the U-net architecture to be generalized and to obtain good quality FA from DW volumes acquired in 1 minute. Our network requires 10 input DW volumes (hence fast acquisition), is robust to the direction of application of the diffusion gradients (hence generalized), and preserves/improves map quality (hence good quality maps). We trained the network on the human connectome project (HCP) data using the standard model-fitting method on the entire set of DW directions to extract FA (ground truth). We addressed the generalization problem, i.e., we trained the network to be applicable, without retraining, to clinical datasets acquired on different scanners with different DW imaging protocols. The network was applied to two different clinical datasets to assess FA quality and sensitivity to pathology in temporal lobe epilepsy and multiple sclerosis, respectively. For HCP data, when compared to the ground truth FA, the FA obtained from 10 DW volumes using the network was significantly better (p <10−4) than the FA obtained using the standard pipeline. For the clinical datasets, the network FA retained the same microstructural characteristics as the FA calculated with all DW volumes using the standard method. At the subject level, the comparison between white matter (WM) ground truth FA values and network FA showed the same distribution; at the group level, statistical differences of WM values detected in the clinical datasets with the ground truth FA were reproduced when using values from the network FA, i.e., the network retained sensitivity to pathology. In conclusion, the proposed network provides a clinically available method to obtain FA from a generic set of 10 DW volumes acquirable in 1 minute, augmenting data quality compared to direct model fitting, reducing the possibility of bias from sub-sampled data, and retaining FA pathological sensitivity, which is very attractive for clinical applications.