The reliability of a deep learning model in clinical out-of-distribution MRI data: a multicohort study