A convolutional neural network for the quality control of MRI defacing
Large scale neuroimaging datasets present unique challenges for automated processing pipelines. Motivated by a large-scale clinical trials dataset of Multiple Sclerosis (MS) with over 235,000 magnetic resonance imaging (MRI) scans, we consider the challenge of defacing - anonymisation to remove identifying features on the face and the ears. The defacing process must undergo quality control (QC) checks to ensure that the facial features have been adequately anonymised and that the brain tissue is left completely intact. Visual QC checks - particularly on a project of this scale - are time-consuming and can cause delays in preparing data for research. In this study, we have developed a convolutional neural network (CNN) that can assist with the QC of MRI defacing. Our CNN is able to distinguish between scans that are correctly defaced, and three sub-types of failures with high test accuracy (77\%). Through applying visualisation techniques, we are able to verify that the CNN uses the same anatomical features as human scorers when selecting classifications. Due to the sensitive nature of the data, strict thresholds are applied so that only classifications with high confidence are accepted, and scans that are passed by the CNN undergo a time-efficient verification check. Integration of the network into the anonymisation pipeline has led to nearly half of all scans being classified by the CNN, resulting in a considerable reduction in the amount of time needed for manual QC checks, while maintaining high QC standards to protect patient identities.