Oct Segmentation Using Convolutional Neural Network

Abstract

Optical coherence tomography (OCT) is a powerful tool for diagnosing many ophthalmic diseases that causes variations to the structure of the eyes. The size of edema and thickness of choroid layers can be ascertained by proper segmentation of OCT images of retina. This paper proposes a model using Convolutional Neural Network (CNN) for segmenting edema and choroid layers in OCT images. Our CNN model is basically an encoder-decoder architecture designed to extract pixel wise information of images to delineate boundaries. For enabling this, a CNN is trained to derive pixel wise labels for the region of interest and its exterior. The pixel labels are then converted into binary segments using morphological operations followed by edge detection. Our algorithm for edema segmentation showed superior accuracy and consistency with an average BF score of 0.91. Results obtained for choroid segmentation are also in agreement with expert findings and proved robust both for images with retinal pathologies and images sourced from different machines.