Integrated Imaging Classification Framework for Diagnosing Diabetic Retinopathy

Abstract

One of the primary causes of impaired vision in adults and middle-aged people is diabetic newlineretinopathy (DR), a drastic severe complication of diabetes that causes changes in retinal newlinemicrovascular structures. DR can cause complete vision loss if left untreated. Therefore, it newlineis recommended that people with diabetes be screened regularly and treated as soon as newlinepossible. There has been a marked increase in the detection of retinal lesions using fundus newlineimages with contrast features. Fundus imaging gives a more detailed description, and newlinecharacterization of the retinal lesions, such as microaneurysms, hemorrhages, and soft and newlinehard exudates associated with DR. However, fundus images with extreme contrast and newlinebrightness variations often hinder diagnosis performance. In addition, rural areas cannot newlineuse the current DR detection methods, and the existing screening capabilities cannot keep newlineup with the increasing number of diabetics in the population. As a result, automatic newlinedetection of DR is highly desirable. newlineThe ultimate purpose of the research study reported in this thesis was to suggest a unique newlineCAD system that can facilitate on-demand pre-processing operations and a learning-driven newlineautomated classification system for detecting the severity of DR. On-demand preprocessing newlineframework in single windowing operation is proposed to aid DR classification newlinefrom the perspective of the human visual system, where various image treatment newlineprocedures are discussed to address image geometrical related problems, suppression of newlinerandom noise, and correction of the intensity variations in the fundus image. On the other newlinehand, the second framework introduces a simplified feature representation mechanism newlinebased on Gaussian blurring and Sobel edge detection. This operation provides primary newlinelevel enhancement and is then subjected to an adaptive filter resulting in a noise-free and newlineenhanced features set. Further, a multi-class classification system using a convolution newlineneural network is designed to identify different stages of DR. newlineA MATLAB computing tool is used in the design and develop of the proposed preprocessing newlineframework. The proposed feature enhancement and DR classification system is newlinedeveloped using Python programming language and validated on the APTOS blindness newlinedetection dataset. The outcome suggests that the DR screening can be conducted efficiently newlineix newlinewith the proposed system, which is reliable, adaptive, cost-efficient, and user-friendly. This newlinewill help ophthalmologists to connect with patients more effectively. newline