3D Deep Convolutional Neural Network for Detection of Anomalous Rigid and Affine Registrations in Big Data Brain MRI

Abstract

Registration to a reference image is an important step during preprocessing of structural brain magnetic resonance imaging (MRI) in group studies for disease diagnosis and prognosis. The manual quality control (QC) of these many images is time-consuming, tedious, and requires prior expertise. Owing to the availability of free MRI datasets and the recent advances in computational infrastructure and deep learning frameworks, it is now feasible to train a deep learning model on larger datasets. To facilitate fully automatic QC in large-scale MRI studies, we proposed 3D deep convolutional neural network models for checking rigid and affine registrations of T1-weighted and T2-weighted MRI data. Because it is a supervised learning approach, five artificially misaligned images are generated for each image type and registration type. The proposed models were cross-validated and tested on the dataset from IXI consisting of 580 T1w and 576 T2w images, where 80 percent of them are used for cross-validation and remaining for testing. Performance metrics such as accuracy, F1-score, recall, precision, and specificity demonstrated a value greater than or equal to 0.99. Therefore, the models could be deployed during fully automatic QC of rigid and affine registrations in the bigdata structural MRI processing pipeline.