Daniel Kostro, Ahmed Abdulkadir, Alexandra Durr, Raymund Roos, Blair R. Leavitt, Hans Johnson, David Cash, Sarah J. Tabrizi, Rachael I. Scahill, Olaf Ronneberger, Stefan Klöppel, Track-HD Investigators

Automated analysis of structural magnetic resonance images is a promising way to improve early detection of neurodegenerative brain diseases. Clinical applications of such methods involve multiple scanners with potentially different hardware and/or acquisition sequences and demographically heterogeneous groups. To improve classification performance, we propose to correct effects of subject-specific covariates (such as age, total intracranial volume, and sex) as well as effects of scanner by using a non-linear Gaussian process model. To test the efficacy of the correction, we performed classification of carriers of the genetic mutation leading to Huntington's disease (HD) versus healthy controls. Half of the HD carriers were free of typical HD symptoms and had an estimated 5 to 20years before onset of clinical symptoms, thus providing a model for preclinical diagnosis of a neurodegenerative disease. Structural magnetic resonance brain images were acquired at four sites with pairs of sites which had the identical scanner type, equipment, and acquisition parameters. For automatic classification, we used spatially normalized probabilistic maps of gray matter, then removed confounding effects by Gaussian process regression, and then performed classification with a support vector machine. Voxel-based morphometry of gray matter maps showed disease effects that were spatially wider spread than effects of scanner, but no significant interactions between scanner and disease were found. A model trained with data from a single scanner generalized well to data from a different scanner. When confounding diagnostics groups and scanner during training, e.g. by using controls from one scanner and gene carriers from another, classification accuracy dropped significantly in many cases. By regressing out confounds with Gaussian process regression, the performance levels were comparable to those obtained in scenarios without confound. We conclude that models trained on data acquired with a single scanner generalized well to data acquired with a different same-generation scanner even when the vendor differed. When confounding grouping and scanner during training is unavoidable to gather training data, regressing out inter-scanner and between-subject variability can reduce the loss in accuracy due to the confound.