YOLOv8 Deep Learning Model for Diabetic Retinopathy Fundus Image Segmentation and Disease Classification

Diabetic retinopathy (DR) is a set of retinal and vitreous changes caused by diabetes which causes visual acuity loss in its later stages, and many patients remain undiagnosed even when the disease is already causing damage. Diabetic retinopathy can be diagnosed in its earlier stages by analysis of that it performs close to the state-of-the-art on at least half of the testing dataset but performed worse in a few more complex cases a fundoscopy examination image, in which the ophthalmologist searches for exudates and microaneurysms, which are early lesions caused by DR. Ophthalmologists have several difficulties in providing accurate and reliable diagnoses to all patients examined. Thus, computer vision can assist ophthalmologists in diagnosing this disease using procedures such as image segmentation and classification. This work uses the YOLOv8 deep learning model to segment and classify DR retinal lesions in fundoscopy examination images using the e-ophtha dataset. Four models were trained, one segmenting and classifying exudates, other microaneurysms, and the last two segmenting both lesions together, one trained using the whole dataset, and the other trained using only the lesion images of the dataset without the healthy images. The segmentation results were quantitatively measured using the sensitivity, intersection-over-union, and DICE coefficient metrics. The exudates model obtained the best segmentation results, with 95.37% mean sensitivity and intersectionover-union and 97.62% mean DICE coefficient. The models were able to classify all images correctly between two classes, lesion and healthy, where images with one or more lesions segmented were classified as lesion images, and images without lesions were classified as healthy. The microaneurysms model obtained high median metrics results compared to its mean results, indicating that it performs close to the state-of-the-art on at least half of the testing dataset but performed worse in a few more complex cases.