Randomization-Driven Hybrid Deep Learning for Diabetic Retinopathy Detection

Diabetic retinopathy (DR), a severe consequence of diabetes, dramatically enhances the likelihood of experiencing vision impairment. Timely identification is crucial for efficient intervention, as untreated diabetic retinopathy can progress to irreversible vision loss. Despite advances, existing diagnostic methods face challenges such as resource dependency, variability in accuracy, and limited accessibility, especially in underserved regions. This study pioneers an innovative framework, using Multi-Scale Discriminative Robust Local Binary Pattern (MS-DRLBP) features, combined with a hybrid Convolutional Neural Network-Radial Basis Function (CNN-RBF) classifier, to enhance the detection of DR. Inspired by principles of randomization-based learning, our approach incorporates elements of stochastic modeling within the CNN-RBF architecture to optimize feature extraction and classification, mirroring the efficiency of non-iterative training processes. We enhance the model’s diagnostic capability through complex image preprocessing techniques, such as improved noise reduction and morphological approaches. Additionally, we use Otsu’s thresholding method to segment blood vessels accurately. Our methodology demonstrates superior performance in DR screening, significantly exceeding traditional diagnostic methods. Specifically, our precision reached 96.10%, sensitivity was 95.35%, specificity achieved 97.06%, and accuracy was 96.10%. This research enhances the precision of DR diagnosis by applying it to different publicly accessible datasets. It contributes to the broader discourse on the potential of hybrid, randomization-inspired neural networks in medical imaging. This fusion of deep learning innovation with the principles of randomization-based algorithms opens new avenues for developing accessible, accurate diagnostic tools, potentially alleviating the global impact of diabetic vision loss.