Fully Automatic Epiretinal Membrane Segmentation in OCT Scans Using Convolutional Networks

Introduction

Thanks to recent advances in computation architectures and in the development of new and improved artificial intelligence algorithms, Computer-aided Diagnosis (CAD) systems are becoming increasingly relevant in healthcare services, finding application in fields such as audiometry (Fernández, et al., 2018), radiology (Romeny et al., 1998) or encephalography (Hosny et al., 2018). Among the different domains of application of CAD systems, the analysis of medical images stands out.

Deep learning is a subset of machine learning focused on the development of artificial neural network models that have multiple layers in order to progressively extract higher-level features from data (LeCun et al., 2015). Convolutional neural networks are artificial neural networks that use the convolution operation in order to be able to extract visual features from images (Lecun et al., 1998). Thanks to the advent of deep convolutional neural networks, deep learning models are nowadays used for several tasks in the field of image analysis. These tasks can range from image classification (Krizhevsky et al., 2017), in which images are separated in classes according to their content; to image segmentation (Long et al., 2015), where images are partitioned into multiple segments, or every pixel in their content is labelled; to regression (Lv et al., 2014), where a value or a set of values are extracted from the images. With a focus on healthcare, deep convolutional neural networks have successfully been used to analyze and study anatomical structures as well as pathological signs in medical images of several types (Shen et al., 2017; Litjens et al., 2017). In particular, deep learning-based CAD systems have found uses in medical imaging techniques such as magnetic resonance imaging (Kamnitsas et al., 2017), conventional radiography (Lakhani & Sundaram, 2017), ultrasound (Cheng et al., 2016) or computed tomography scans (Setio et al., 2016). These systems have demonstrated that their performance can be on par with, or even exceed that of human experts in different diagnosis-related tasks (Litjens et al., 2017; Lee et al., 2020; Ting et al., 2017; Gulshan et al., 2016).

In ophthalmology, Optical Coherence Tomography (OCT) is an imaging technique that allows the in-depth visualization of tissue (Huang et al., 1991). By shining a beam of low coherence (high bandwidth) light over the tissue and measuring the differences in phase and amplitude in the reflected beam compared to a reference one, a one-dimensional reading or A-Scan can be acquired at every scanned spot. This A-Scan contains depth-wise information about the reflectivity of the scanned tissue. If this beam is swept through the surface of the tissue, these readings can be combined into a two-dimensional reading or B-Scan (Figure 1). These B-Scans can be visualized as high-resolution images which show a cross-sectional view of the tissue, like a tomogram containing the histological information of the scanned tissue. Furthermore, these B-Scans can be laterally combined to produce a volumetric representation of the underlying tissue. This makes OCT a remarkably useful technique for the analysis of healthy or pathological ocular structures since these volumes allow the complete histological visualisation of the retinal tissue in vivo and in a non-invasive manner. For reference, OCT images can be used for the study of the vascular structure of the eye (Kashani et al., 2017; de Moura et al., 2016, 2017a; Spaide et al., 2018); for the diagnosis of glaucoma (Hood, 2017) (Tan et al., 2009; Jaffe & Caprioli, 2004), which is the most common cause of blindness in the developed world for people over 50; exudative macular disease (de Moura et al., 2017b), one of the most common causes of blindness in the developed world; or that of diabetic macular oedema (Hee, 1995; de Moura et al., 2019, 2020; Mookiah et al., 2013), the leading cause of blindness in patients of diabetes mellitus.