Harmonization of Diffusion MRI Data Obtained with Multiple Head Coils Using Hybrid CNNs.

In multisite diffusion MRI studies, different acquisition settings can introduce a bias that may overshadow neurological differentiations of the study population. A variety of both classical harmonization methods such as histogram warping as well as deep learning methods have been recently proposed to solve this problem and to enable unbiased multisite studies. However, on our novel dataset—it consists of acquisitions on the same scanner with the same acquisition parameters, but different head coils—available methods were not sufficient for harmonization. Nonetheless, solving the challenge of harmonizing this difficult dataset is relevant for clinical scenarios. For example, a successful harmonization algorithm would allow the continuation of a clinical study even when the employed head coil changes during the course of the study. Even though the differences induced by the change of the head coil are small, they may lead to missed or false associations in clinical studies. We propose a harmonization method based on known operator hybrid 3D convolutional neural networks. The employed neural network utilizes a customized loss that includes diffusion tensor metrics directly into the harmonization of raw diffusion MRI data. It succeeds a preliminary histogram warping harmonization step. The harmonization performance is evaluated with diffusion tensor metrics, multi-shell microstructural estimates, and perception metrics. We further compare the proposed method to a previously published deep learning algorithm and standalone intensity warping. We show that our approach successfully harmonizes the novel dataset, and that it performs significantly better than the previously published algorithms.