Please note that this is a pre-print version provided for personal use only. DOI: 10.1109/IC4ME247184.2019.9036614 Diabetic Retinopathy Classification with a Light Convolutional Neural Network Md. Sanaullah Chowdhury Electronics & Communication Engineering Discipline Khulna University Khulna, Bangladesh sanaullahashfat@gmail.com Faozia Rashid Taimy Electronics & Communication Engineering Discipline Khulna University Khulna, Bangladesh faozia.rashid.taimy@gmail.com Niloy Sikder Computer Science & Engineering Discipline Khulna University Khulna, Bangladesh niloysikder333@gmail.com Abdullah-Al Nahid Electronics & Communication Engineering Discipline Khulna University Khulna, Bangladesh nahid.ece.ku@gmail.com Abstract—The number of diabetic patients is increasing rapidly every year all around the world, and the worst fact is that these patients suffer from a wide range of physical conditions directly associated with long-term diabetes. Diabetic Retinopathy (DR) is a perfect example which affects the eyes of more than 50% of all diabetes patients to some degree. Starting from blurred vision, the effects of DR can extend to permanent blindness; and in most of the cases, victims fail to report any early symptoms. The traditional detection process of DR involves a trained clinician who takes enhanced pictures of the retina and looks for the presence of lesions and vascular abnormalities within them, which by description is a time-consuming and error-prone procedure. Alternatively, we can employ machine learning techniques that will automate the detection process as well as provide fast and more importantly, reliable results. Using a deep learning technique this paper determines the presence and severity of DR in diabetic individuals by analyzing the pictures of their retina. The CNN-based models are potent enough to carry out their tasks with accuracy up to 89.07%, even when the images are captured or provided in very low resolutions. Keywords—diabetic retinopathy detection, deep CNN, coye filter, image classification, machine learning I. INTRODUCTION Diabetic Retinopathy (DR) is the leading complication of blindness in adults. Statistics show that at present about 4.2 million adults had DR and 655,000 had vision-loss due to DR. The number of people who are affected by vision loss due to DR increased in every year [1]. It is known to all that DR becomes a common complication of Diabetes Mellitus (DM). The treatment of DR is not easy as there is no symptom shown at the earlier-stages of DR and patients rarely notice a vision loss [2]. Most of them could not realize that they have DR until the disease started to affect their vision, which usually occurs in the last stage. As a result, they might not go through treatment promptly. Therefore, a coordinated management strategy is crucial to optimally address the clinical challenges of DR and limiting its progression. Early identification and classification of disorder of the retinal images always being a serious concern to the research community. There are a few works that have been performed to classify retinal image malignancy utilizing conventional shallow learning techniques. However, Deep Learning (DL) techniques have gained tremendous success in solving visual related problems [3], [4]. One of the variants of this method known as Convolutional Neural Network (CNN) has gained a tremendous momentum after the famous model "AlexNet" which is proposed in 2012 [5]. Initially, this method used to classify natural images. However, in later, several variants of this model are used to solve many real lives vision-related problems, such as identification of malignancy of retinal images [6], [7], [8]. Image classifications rate might be degraded due to different issues such as contrast, illumination etc. [9], which can be improved by various image pre- processing techniques. Chen et al. utilized linear un-sharp masking filter to enhance the edge and detailed information, then, the enhanced image is feed to a CNN method named as SI2DRNet-v1 for retinal image classification [7]. V. Raman et al. classified a set of retinal images whereas a pre- processing technique, Contrast-Limited Adaptive Histogram Equalization (CLAHE) method is utilized to improve the visibility of images [10]. D. K. Prasad et al. classify a set of retinal images (DIARETDB1 dataset) for detecting the early status of the DR and those images are preprocessed by adaptive histogram equalization techniques to improve the image contrast [1]. M. U. Akram et al. classify retinal lesions utilizing a hybrid classifier where they utilized Gabor filter as a pre-processing tool [8]. From the above discussion it turns out that, an image preprocessing stage before the classification stage allows a classifier to perform better, Coye filter has been used as an image preprocessor which enhances the contrast information of the corresponding image [11]. Then a noble light CNN model is hyper-tuned by these enhanced images to classify a set of retinal images into their respective classes. To do so, this paper starts with Section I stating the background of DR and its severity throughout the world. Various aspects of the proposed scheme especially the entire methodology has been step-wise illustrated in section II. A brief analysis in CNN to classify DR has been incorporated in Section III along with our proposed a light noble CNN model. In Section IV, the overall analysis and comparison of the results have been presented. This paper concludes with Section V providing a discussion on overall research. II. METHODOLOGY This experiment has performed in a supervised manner as Fig. 1. In a general principle, supervised learning method trained based on a training dataset which is properly labeled. This labeled data is used to select probable best parameters for the classifier. Lastly, the selected or targeted data is fit with the model which provide relatively the best decisions. This experiment is performed on a publicly available dataset named as EyePACS [12]. This dataset contains five classes provided on a scale of 0 to 4. Table I shows the statistic information about the different classes of the dataset such as The table shows that almost 78 to 80 percent of the total images belong to class 0. However, another four classes of the data cover almost 20% of the total data. This indicates that this dataset is imbalanced.