Adaptive kernel integration in visual geometry group 16 for enhanced classification of diabetic retinopathy stages in retinal images

Abstract

Diabetic retinopathy (DR) is a major cause of vision impairment globally, with early detection remaining a significant challenge. The limitations of current diagnostic methods, particularly in identifying early-stage DR, highlight a pressing need for more accurate diagnostic technologies. In response, our research introduces an innovative model that enhances the visual geometry group 16 (VGG16) architecture with adaptive kernel techniques. Traditionally, the VGG16 model deploys consistent kernel sizes throughout its convolutional layers. In this study, multiple convolutional branches with varying kernel sizes (3×3, 5×5, and 7×7) were seamlessly integrated after the 'block5_conv1' layer of VGG16. These branches were adaptively merged using a softmax-weighted combination, enabling the model to automatically prioritize kernel sizes based on the image's intricate features. To combat the challenge of imbalanced datasets, the synthetic minority over-sampling technique (SMOTE) was employed before training, harmonizing the distribution of the five DR stages. Our results are promising, showing a training accuracy above 94.17% and a validation accuracy over 90.24%, our model significantly outperforms traditional methods. This study represents a significant stride in applying adaptive kernels to deep learning for precise medical imaging tasks. The model's accuracy in classifying DR stages highlights its potential as a valuable diagnostic tool, paving the way for future enhancements in DR detection and management.