Automatic Quality Assessment and Segmentation of Diabetic Retinopathy Images

Publication Type:
Conference Proceeding
Proceedings of the 2016 IEEE Region 10 Conference (TENCON), 2016, pp. 997 - 1000 (4)
Issue Date:
Full metadata record
Files in This Item:
Filename Description Size
07848155.pdfPublished version256.58 kB
Adobe PDF
Diabetes, often referred to as diabetes mellitus, describes a group of metabolic diseases whose patients are diagnosed with having a high glucose level in blood. One form of diabetes commonly found in Thailand is Diabetic Retinopathy, in which prevalent symptoms of diabetes are very likely to induce vision impairment or even blindness. Thus, the ability to detect early symptoms of diabetes through retinal images is proven vital for proper medical treatments to reduce the risk of the loss of sight. Although recent researches have proposed several methodologies for assessing the severity level according to the condition of retinal components from retinal images, issues over the accuracy and precision of the evaluation model compel the improvement for a better model to be practically implemented with more favorable outcome. Earlier researches have introduced several methods such as the image quality assessment based on the histogram back projection [1], or the image segmentation for the diagnosis of the diabetic retinopathy severity level via the construction of Echo State Neural Network (ESNN). This paper proposes methods to assess the image quality and to segment the image components for analyzing the disease severity level of a retinal image. Four major information constituting the image quality, namely color, contrast, focus, and illumination, are extracted in order to evaluate the overall image gradability. By classifying images according to their image quality factors, the Principal Component Analysis (PCA) technique is applied to reduce the data dimensionality and vividly project discriminant features over minor misleading details that could hamper the accuracy of the evaluation result. Then, the k-nearest neighbor is applied as a classification tool. Moreover, regarding the identification of components such as hard exudates in a retinal image [2] [3], image segmentation techniques are applied in order to deduce the representation of each retinal component for further assessment on the severity level of Diabetic Retinopathy. In this paper, the segmentation is performed based on the thresholding, component dilation, and grabcut techniques. The datasets used in this paper are DIABRETDB1 [4] and DRIVE [5]. They provide retinal images with the corresponding quality label and medical examination of the severity. The aim of this work is to develop a program that can effectively assist medical personals in a classification of the retinal image gradability based on the image's quality, and provide a preliminary medical diagnosis on the level of severity of the disease. The rest of this paper is organized as follows. The detailed explanation of the retinal image quality assessment is in section II. The retinal image segmentation and preliminary disease's severity diagnosis are proposed in section III. Experimental results are shown in section IV, and conclusions are drawn in section V.
Please use this identifier to cite or link to this item: