Early Hypertensive Retinopathy Detection Using Effective Segmentation and Fine-Tuned Cnn Classification With Stratified Training Protocol

Hypertensive Retinopathy (HR) early detection is basically figuring out retinal changes that are very subtle and come from high blood pressure, even before there is significant vascular damage. To overcome this Hybrid Retinal Efficient Smart Fundus Classification (HRESFC) model is designed, which uses deep learning to automatically identify HR from retinal fundus images. It analyses 9,170 images containing 3,410 HR and 5,760 non-HR cases, all verified by ophthalmologists. Resizing, augmentation, and balancing are part of the pre-processing. The optic disc and cup are extracted by extracting features and using an active contour algorithm. The MobileNetV2-based model employs dense layers activated by Rectified Linear Unit (ReLU), and dropout layers for regularisation. Stratified sampling with the Adam optimiser, L2 regularisation, and early stopping is used for training. The team measures the system's performance by accuracy (99%), precision 98%, recall 97% to 76%, specificity, F1-score 78% and 71%, AUC-ROC 85-92%, and other statistical metrics.