Computer-aided diagnosis of diabetic patients based on color fundus images using machine learning techniques

Diabetic retinopathy (DR) and diabetic macular edema (DME) are regarded as the most common complications of diabetes that, if not treated accordingly, could result in blindness. Early diagnosis and treatment planning can be considered as an essential step in preventing the vision loss, but the large and growing number of diabetic patients coupled with difficulties in screening a high number of patients makes early diagnosis difficult. Additionally, most of the time, a non-trivial inter- and intra-observer variability can be observed, depending on the point in time or the level of experience, different persons or even the same person may outline the anatomical boundaries differently. Computer-assisted diagnosis can be used for checking the retinal condition at different time intervals, providing a fast and reliable way of monitoring patient’s condition during different time frames. However, most of the proposed methods do not contain any grading capabilities and are mostly designed for screening purposes. The proposed computer-assisted diagnosis approach starts with the segmentation of the blood vessels. Then, optic disk and macula regions are located and segmented. Removing vessels, optic disk and macula regions increases accuracy of microaneurysm and exudate segmentation. Finally, retinal images are classified and graded using an AdaBoost classification method based on features extracted utilizing first, second and higher order image features selected by a minimal-redundancy maximal-relevance feature selection approach. Being brighter than the surrounding tissue, optic disk (OD) causes rapid variations in image intensity. This variation can be used for locating the OD region. In our study, OD is located using a variance based approach with OD outline segmented using circular Hough transform. By leveraging the location and the diameter of segmented OD, it is possible to locate the macula region as its position is relatively constant compared to OD. In this study, an exudate segmentation approach based on Kirsch’s Edges method is used with the microaneurysms being segmented using mathematical morphology and thresholding approaches. In this thesis, for each retina image, a feature vector with a fixed size is generated regardless of the position or the number of exudates and microaneurysms, which might not be properly segmented and used in an AdaBoost classifier for screening and grading images with possible signs of diabetic retinopathy and diabetic macular edema. The accuracy of the proposed diabetic grading approaches were comparable to other state of the art methods with an average accuracy of 0.791 and 0.974 in publicly accessible MESSIDOR dataset, respectively. By utilizing computer vision and machine learning concepts, it is possible to increase the DME detection rate considerably as CAD can reduce the workload of the ophthalmologists.

Text FK 2019 54 ir.pdf Download (1MB)

Actions (login required)

View Item