Diffeomorphic Autoencoders for LDDMM Atlas Building

Jacob Hinkle, David Womble, Hong-Jun Yoon

Keywords: image registration, atlas-building, deep learning, autoencoder, LDDMM

TL;DR: We train deep neural networks to estimate image registration simultaneously with LDDMM atlas building

Abstract: In this work, we present an example of the integration of conventional global and diffeomorphic image registration methods with deep learning. Our method employs a form of autoencoder in which the encoder network maps an image to a transformation and the decoder interpolates a deformable template to reconstruct the input. This enables image-based registration to occur simultaneously with training of deep neural networks, as opposed to current sequential optimization methods. We apply this approach to atlas creation, showing that a system that jointly estimates an atlas image while training the registration encoder network results in a high quality atlas despite drastic dimension reduction. In addition, the shared parametrization for deformations offered by the neural network enables training the atlas with stochastic gradient descent using minibatches on a single GPU. We demonstrate this approach using affine transformations and diffeomorphisms in the LDDMM vector momentum geodesic shooting formulation using the OASIS-3 dataset.

Code Of Conduct: I have read and accept the code of conduct.