Truls Svensson: Simultaneous Registration and Reconstruction: Deep Learning for Indirect Registration in 4D Spatio-Temporal CT using the LDDMM Framework

Abstract: A novel deep learning based framework is presented for performing indirect registration in 4D spatio-temporal CT. To address the challenge of the combined problem of simultaneous registration and reconstruction, a new operator-based unrolled neural network architecture is developed - merging ideas from tomographic reconstruction and image registration into a joint framework. This results in a Learned Primal Dual-inspired network architecture with dedicated registration blocks that estimate velocity fields, which are integrated into a diffeomorphic flow according to the Large Deformation Diffeomorphic Metric Mapping framework to obtain a temporal sequence of registered images by deforming a reconstructed template. The model is trained and tested on the 4D lung data set, and the results demonstrate that the model successfully recovers complex respiratory motion from a time sequence of simulated CT data.