978-1-7281-3975-3/19/$31.00 ©2019 IEEE Segmentation-based Deep Learning Fundus Image Analysis Qian Wu Department of Computer Science Blekinge Institute of Technology SE-371 79, Karlskrona, Sweden qiwu17@student.bth.se Abbas Cheddad, SMIEEE Member Department of Computer Science Blekinge Institute of Technology SE-371 79, Karlskrona, Sweden abbas.cheddad@bth.se Abstract—Diabetic retinopathy is the most common cause of new cases of blindness in people of working age. Early diagnosis is the key to slowing the progression of the disease, thus preventing blindness. Retinal fundus images form an important basis for judging these retinal diseases. To the best of our knowledge, no prior studies have scrutinized the predictive power of the different compositions of retinal images using deep learning. This paper is to investigate whether there exists specific region that could assist in better prediction of the retinopathy disease, meaning to find the best region in fundus images that can boost the prediction power of models for retinopathy classification. To this end, with image segmentation techniques, the fundus image is divided into three different segments, namely, the optic disc, the blood vessels, and the other regions (regions other than blood vessels and optic disk). These regions are then contrasted against the performance of original fundus images. The convolutional neural network as well as transfer deep learning with the state-of-the-art pre-trained models (i.e., AlexNet, GoogleNet, Resnet50, VGG19) are deployed. We report the average of ten runs for each model. Different machine learning evaluation metrics are used. The other regions’ segment reveals more predictive power than the original fundus image especially when using AlexNet/Resnet50. Keywords— Fundus Image, retinopathy, Deep Learning, AlexNet, Image Segmentation I. INTRODUCTION 1 The human eye is a vital sensory organ which enables us the sense of vision. It plays a very important role in our daily lives. We use visual system in almost every activity since the eye allows us to see and interpret objects in the real world by processing the reflected or emitted light from these objects. When light penetrates the eye, it passes through the cornea and the lens and is refracted, focusing an image onto the retina. The retina is a complex transparent tissue composed of several layers covering two-thirds of the eyeball, in which light stimulation occurs, causing visual sensation 2 . The retina can be seen as an extension of the brain, since it is connected to the brain through the optic nerve. Diabetic retinopathy is the most common cause of new cases of blindness in people of working age [1]. Although diabetes disease affects the eye in many ways (for example, the high risk of cataracts), diabetic retinopathy is the most common and the most serious ocular complication [2]. Early diagnosis is the key to slowing the progression of the disease, thus preventing blindness [3]. Screening for diabetic retinopathy aims to detect early sight‐threatening lesions which can then be treated with laser 1 This work is supported by the research project “scalable resource efficient systems for big data analytics”, (Grant: 20140032), funded by the Knowledge Foundation in Sweden. photocoagulation. Therefore, regular examination of the eyes is necessary [4]. Retinal fundus scan images are an important basis for judging these retinal diseases. Diabetic retinopathy displays abnormal features in its onset, such as hemorrhage, microaneurysms and hard and soft exudates [5]. However, the number of professional doctors who can diagnose this lesion is very limited, making it difficult for diabetic patients to receive timely treatment from experts. In what follows, we take the image segmentation approach to verify the importance of each region independently in predicting diabetic retinopathy. Image segmentation means dividing a fundus image into different regions that exhibit unique features which may be helpful for diagnosis. II. RELATED WORK In here, we provide some of the recent work, as the list is quite extensive, we limit that to a few references. In essence, the majority of the work in the literature concentrate on the vasculature present in fundus images. For a literature review, the reader if referred to [6] and [7]. Retinal markers include retinal blood vessel, optic discs, optic cups, macula, and fovea, see Fig.1. Automated detection of these markers facilitates medical analysis of retinal fundus images. Most studies of retinal fundus image analysis are based on segmentation of blood vessels. Matsui et al. were the first to publish a method for retinal image analysis, the method is focused on vessel segmentation [8]. Tyler Coye proposed an algorithm for segmenting retinal blood vessel with an adequate precision. This algorithm uses the principal component analysis (PCA) to convert a given RGB fundus image to grayscale instead of using the green channel as commonly used [9]. The optic disc (OD) is one of the main anatomical structures in the retinal image. It is typically displayed in a normal retinal image as an approximately circular and bright yellow object. In the past few decades, many techniques of automatic OD localisation and segmentation have been investigated [10], and many researchers have proposed methods for automatically diagnosing diabetic retinopathy by segmenting OD. Dharitri Deka et al. [11] proposed the detection of macular and foveal for disease analysis of colour fundus images. Macula represents the region of the retina responsible for colour vision, fovea is the centre of the macula. The macula 2 Retina Anatomy: https://www.britannica.com/science/retina Encyclopædia Britannica.