A Line to Align: Deep Dynamic Time Warping for Retinal OCT Segmentation

In order to scan for or monitor retinal diseases, OCT is a useful diagnostic tool that allows to take high-resolution images of the retinal layers. For the aim of fully automated, semantic segmentation of OCT images, both graph based models and deep neural networks have been used so far. Here, we propose to interpret the semantic segmentation of 2D OCT images as a sequence alignment task. Splitting the image into its constituent OCT scanning lines (A-Modes), we align an anatomically justified sequence of labels to these pixel sequences, using dynamic time warping. Combining this dynamic programming approach with learned convolutional filters allows us to leverage the feature extraction capabilities of deep neural networks, while at the same time enforcing explicit guarantees in terms of the anatomical order of layers through the dynamic programming. We investigate both the solitary training of the feature extraction stage, as well as an end-to-end learning of the alignment. The latter makes use of a recently proposed, relaxed formulation of dynamic time warping, that allows us to backpropagate through the dynamic program to enable end-to-end training of the network. Complementing these approaches, a local consistency criterion for the alignment task is investigated, that allows to improve consistency in the alignment of neighbouring A-Modes. We compare this approach to two state of the art methods, showing favourable results.