Hybrid Deep Learning and Random Forest-Based EyeSmart Framework for Enhanced Retinal Disease Detection

In this paper, a novel hybrid approach called the EyeSmart Framework that combines deep learning-based feature extraction with an optimized random forest classifier for the multi disease retinal analysis is presented. The framework uses pre trained convolution neural networks (CNNs), like ResNet50, to extract retinal images features at a high level, and classify them using an improved random forest model at an enhanced random forest model using hyperparameter. The approach being proposed here is to address the challenges of high accuracy, interpretability, and computational efficiency for diagnosis of retinal disease. Our framework was extensively experimented on publicly available retinal disease datasets, and performance was compared against traditional and supervised deep learning methods. We show that the EyeSmart Framework results in an accuracy of 92.1% which is a substantial improvement over a CNN only baseline model (85.6<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>) and a conventional random forest classifier (82.3%). Further, the framework reports a precision of 91.5%, recall of 90.8% and F 1 score of 91.2% for diabetic retinopathy, glaucoma, and age related macular degeneration thus making it robust to detect all three diseases. The AUC-ROC score of 0.94 is further validation of its capacity to differentiate between diseased and non diseased cases in comparison to CNN with score of 0.89 and random forest with 0.87. The impact of various CNN architectures and random forest hyperparameters on performance is ablated in the study. Through Grad-CAM visualizations, we show that the critical regions for classification are located on the retina, improving the interpretability of the model. Based on the results, the EyeSmart Framework serves as a viable and interpretable approach to automated retinal disease diagnosis and can be extended to real world clinical applications.