Detection and classification of hypertensive retinopathy based on retinal image analysis using a deep learning approach

• Early detection models are essential to improve treatment accuracy and prevent heart attacks and strokes. • This study proposes a DCNN model for detecting and classifying hypertensive retinopathy disease. • The study presents a new dataset of 9 HR classes from the relabeling of the Messidor database, by adding a borderline class. • The model testing achieved Specificity, Recall, Accuracy, and F-Score levels of 90%, 81.82%, 90%, and 90%, respectively. The issue is that most heart attacks and strokes happen unexpectedly to people who have signs of high blood pressure that are not identified in time for treatment. These gap factors make the research on hypertensive retinopathy urgent since it requires an early detection model to improve treatment accuracy and prevent heart attacks and strokes before they happen. This research utilizes secondary data, specifically a retinal image dataset from the open-source Messidor database. This database comprises 1200 retinal images, each measuring 1440 × 940 pixels. The dataset is divided into 60 % training and 40 % validation data. The next step is the image analysis process, which involves extracting retinal blood vessels using the Otsu segmentation algorithm. A Morphological Approach is used to obtain comprehensive features of the blood vessels around the Optic Disc (OD). This stage aims to extract and sample the comparison between the width of the artery and vein (AVR). This research uses a Deep Convolutional Neural Network (DCNN) classification model with cross-validation training using the Leave-one-out method. The results of testing the model with nine output classes, the features extracted in each convolutional layer, the second layer successfully extracts the retina and eye blood vessels, the third layer extracts the retinal image texture, and the fourth layer extracts hard exudates, hemorrhages, and cotton wool spots. Meanwhile, the Specificity, Recall, Accuracy, and F-Score results are 90 %, 81.82 %, 90 %, and 90 %, respectively. This research's findings first include applying the AVR calculation algorithm to build a new dataset with 9 class categories. Second, the architectural specifications of the CNN model are determined, and the input size, depth, and number of nodes for each layer, as well as the transfer function, learning rate, and number of epochs, are set by adjusting hyperparameters.