Automatic detection of exudates and hemorrhages in low-contrast color fundus images using multi semantic convolutional neural network

Diabetic retinopathy (DR) is a pathology occurring in the optic nerve due to an excessive blood sugar level in human body. It is one of the major reasons for visual impairment in the developed and developing countries. Patients with DR usually suffer from visual damages due to a high blood sugar level in retinal blood vessel walls. These damages may also leak into other retinal layers of the eye within time. As a result of these leakages and nutritional disorders, a number of lesions such as excudate, edema, microaneurysm, and hemorrhage may occur. In this respect, an accurate and effective detection of these lesions in earlier stages of DR plays an important role in the progression of the disease. In the proposed study, exudate and hemorrhages, which are important clinical findings for DR, were automatically detected from low contrast colored fundus images. Exudate and hemorrhages are lesions with different characteristics. However, in this study, high performance was achieved by making a three-class semantic segmentation. In addition, a color space transformation was performed and the classical U-Net algorithm was provided to achieve stable high performance in low contrast images. Finally, lesion images which were manually detected by a physician were matched with automatically segmented excudate and hemorrhage images using the proposed method. Thus, both segmentation and lesion detection performances of the proposed method were measured. The findings demonstrated that Dice and Jaccard similarity indexes were calculated nearly as 0.95 for the segmentation performance. A sensitivity of 98% and specificity value of 91% were measured for detection performance. It can be inferred from these figures that the proposed method can be effectively used as a supporting system by physicians for the detection and classification of lesions in the color fundus images for the diagnosis of DR.