Unsupervised Non-Correspondence Detection in Medical Images Using an Image Registration Convolutional Neural Network

Julia Andresen, Timo Kepp, Jan Ehrhardt, Claus von der Burchard, Johann Roider, Heinz Handels

Keywords: Image registration, Non-correspondence detection, Pathology segmentation

TL;DR: Journal-paper abstract presenting our work on joint image registration and non-correspondence detection with deep learning

Abstract: Missing correspondences in medical images might cause implausible distortions in image registration if they are not taken into account explicitly. Jointly registering images and detecting non-correspondent regions is an ongoing research topic that is transferred to the field of deep learning in this paper. A Y-shaped network architecture is proposed together with a two-step training procedure that allows proper separation of spatial deformation and non-correspondence segmentation. Non-correspondences are considered outliers in the image distance measure and regularized to be small and smoothly bordered. The method is thoroughly validated using two very different datasets and shown to achieve state-of-the-art registration performance while being especially robust against large lesions. Furthermore, it is shown that the proposed method can be used for unsupervised segmentation of lesions, with no annotations of pathologies required for network training.