Learning iterative optimisation for deformable image registration with recurrent convolutional networks

Fenja Falta, Lasse Hansen, Mattias P Heinrich

Keywords: Optimisation, Deep learning, Recurrent network

TL;DR: We propose a fully deep learning-based approach for deformable image registration, that aims to emulate the structure of gradient-based optimisation using a recurrent network and dynamic cost sampling.

Abstract: We propose a fully deep learning-based approach for deformable image registration, that aims to emulate the structure of gradient-based optimisation as used in conventional registration and thus learns how to optimise. Our architecture consists of recurrent updates on a convolutional network with deep supervision and uses dynamic sampling of the cost function and hidden states to mimic gradient-based optimization, requiring fewer iterations than traditional techniques. Without pre-registration, our method achieves 2\,mm TRE on the DIR-Lab COPD dataset and outperforms Adam optimisation. Our code is publicly available at: https://anonymous.4open.science/r/Learn2Optimise-BF1F/.