A multiclass deep learning framework for retinal disease detection using VGG16-HCCNN

Abstract Fundus imaging is a crucial diagnostic tool in ophthalmology, as it enables the detection of early signs of various ocular diseases such as diabetic retinopathy, glaucoma, and hypertensive retinopathy. If left untreated, these conditions can lead to vision loss or blindness. In this work, a combined VGG16 and Hybrid Centric Convolutional Neural Network (HCCNN) model is suggested that provides a balanced trade-off between computational efficiency and performance, providing a fast and precise method of classifying retinal images into one of the four classes: normal, diabetic retinopathy, glaucoma, and hypertensive retinopathy. This hybrid model seeks to improve feature extraction by integrating VGG16 feature maps with the adaptability of HCCNN, which is designed to collect domain-specific features in images. To improve the performance, this model uses regularization and dropout method that reduces complexity, and prevent overfitting. The proposed technique was assessed using a diverse dataset sourced from eight publicly available datasets (HRF, SYSU, SRM, PAPILA, DRISHTI_GS, RIM-ONE, INSPIREAVR, AVRDB), attaining an average classification accuracy of 96.93%, sensitivity of 94%, specificity of 97.96% and precision of 93.95% in detecting multiclass retinal disease. This represents a significant improvement over state-of-the-art models like VGG16, InceptionV3, and ResNet101, outperforming them by 5.46% in accuracy, 10.95% in sensitivity, and 3.65% in specificity. The model’s efficiency was also assessed using a technique called 5-fold cross-validation. This method enhances precision and is a significant resource for medical professionals.