Classifying handedness with MRI

When aggregating neuroimaging data across many subjects, an important consideration is establishing some group-level uniformity prior to further statistical analysis. Spatial normalization and motion correction are two important preprocessing steps that help achieve this goal. Researchers have also often excluded left-handed subjects due to presumptions about variable asymmetries relating to both brain structure and function, which may interfere with achieving a desired level of group homogeneity. It is well-known, however, that hand-preference is not a binary attribute and is not a perfect representation of structural asymmetry or hemispheric specialization. In an effort to demonstrate a more objective, data-driven approach for quantifying asymmetries across handedness, we tested the reliability of single-subject classification of handedness using data obtained from structural MRI in extant samples. We utilized data from deformation fields created during the spatial normalization process within a priori regions of interest (ROIs), including the motor and somatosensory cortex, and Broca's and Wernicke's areas. Using these deformation fields as features in machine learning classifiers, we achieved classification accuracies greater than 75% across two independent datasets (i.e., a sample of incarcerated adult offenders and a sample of community adults from the Netherlands). These results demonstrate reliability of morphological features attributable to handedness as represented in neuroimaging data and further suggest that application of data-driven techniques may be a principled approach for addressing asymmetries in group analysis. •We demonstrate a method for quantifying morphological asymmetries in the brain and extend these to classify handedness.•Deformation fields in ROIs were captured from spatial normalization of MRI and included in classification models.•Models utilizing deformation field values achieve classification accuracy over 80% for subject's hand preference.•Inclusion of deformation fields may improve classification models in samples with both right and left-handed participants.