Deep Learning-based Longitudinal Intra-subject Registration of Pediatric Brain MR Images

Andjela Dimitrijevic, Vincent Noblet, Benjamin De Leener

Keywords: Learning-based image registration, pediatric, MRI

TL;DR: Validating deep learning-based combined with conventional registration frameworks for a pediatric population to accomplish longitudinal intra-subject deformable registration

Abstract: Deep learning (DL) techniques have the potential of allowing fast deformable registration tasks. Studies around registration often focus on adult populations, while there is a need for pediatric research where less data and studies are being produced. In this work, we investigate the potential of unsupervised DL-based registration in the context of longitudinal intra-subject registration on 434 pairs of publicly available Calgary Preschool dataset of children aged 2-7 years. This deformable registration task was implemented using the DeepReg toolkit. It was tested in terms of input spatial image resolution (1.5 vs 2.0 mm isotropic) and three pre-alignement strategies: without (NR), with rigid (RR) and with rigid-affine (RAR) initializations. The evaluation compares regions of overlap between warped and original tissue segmentations using the Dice score. As expected, RAR with an input spatial resolution of 1.5 mm shows the best performances. Indeed, RAR has an average Dice score of of 0.937$\pm$0.034 for white matter (WM) and 0.959$\pm$0.020 for gray matter (GM) as well as showing small median percentages of negative Jacobian determinant (JD) values. Hence, this shows promising performances in the pediatric context including potential neurodevelopmental studies.