Bridging the Gap Between Conventional and CNN-Based Registration: Application to Mouse Brain Morphometry

Deformable image registration is a fundamental step in medical imaging. With the advent of deep learning, con-volutional neural networks (CNN) are investigated to per-form image registration. Despite enabling faster inference time, CNN-based registration methods still do not signifi-cantly surpass conventional approaches in accuracy. Here, we formulate the link between conventional and CNN-based registration methods. We highlight the interest in considering CNN-based registration methods as a conventional reg-istration approach for groupwise registration. Although the accuracy is not significantly improved compared to conventional registration, the regularization effect induced by the CNN-based formulation appears to be of great interest for morphometric group study.