Dense Segmentation in Selected Dimensions: Application to Retinal Optical Coherence Tomography

We present a novel convolutional neural network architecture designed for dense segmentation in a subset of the dimensions of the input data. The architecture takes an N-dimensional image as input, and produces a label for every pixel in M output dimensions, where 0 < M < N. Large context is incorporated by an encoder-decoder structure, while funneling shortcut subnetworks provide precise localization. We demonstrate applicability of the architecture on two problems in retinal optical coherence tomography: segmentation of geographic atrophy and segmentation of retinal layers. Performance is compared against two baseline methods, that leave out either the encoderdecoder structure or the shortcut subnetworks. For segmentation of geographic atrophy, an average Dice score of 0.49±0.21 was obtained, compared to 0.46±0.22 and 0.28±0.19 for the baseline methods, respectively. For the layer-segmentation task, the proposed architecture achieved a mean absolute error of 1.305±0.547 pixels compared to 1.967±0.841 and 2.166±0.886 for the baseline methods.