Mitigating site effects in covariance for machine learning in neuroimaging data