A Hybrid Deep Learning Approach for Eye Disease Detection with Integrating Adaptive CNN Layers for Improved Performance of Image Classification

Abstract

Essential for the proper functioning of eyesight, the eye is a crucial organ. If the retina sustains permanent damage from an eye disease, the patient may have severe blurred vision or perhaps become blind. The use of AI in eye disease categorization (EDC) helps doctors better serves their patients. This study's overarching goal is to construct an EDC model by means of deep learning (DL). This research details a process for detecting eye diseases using DL models that use binary and multi-class classification methods. Leveraging the APTOS 2019 dataset, the Adaptive Hybrid Net-CNN model integrates InceptionV3, MobileNetV3Small, ResNet101, and adaptive CNN layers to enhance feature extraction and classification performance. The preprocessing pipeline involves image auto-cropping, resizing, RGB conversion, and sharpening to ensure high-quality inputs. For binary classification, labels are binary-encoded, achieving accuracies of 99.94% during training and 98.45% during testing, with strong performance metrics, including an F1-score of 0.98. For multi-class classification, labels are consolidated into three categories, and SMOTE is used to balance the dataset also K-fold cross-validation further validated the models. An F1-score of 0.97 and an accuracy of 99.28% on training data and 96.36% on testing data were achieved by the model, which was trained utilising categorical cross-entropy loss. Evaluation is conducted through classification reports, ROC and precision-recall curves, and accuracy/loss graphs, offering a thorough understanding of model performance. The comparative analysis shows that the proposed model works well in comparison to existing (GoogleNet_SVM, Hybrid_SVM, and ResNet-18_SVM) models in terms of all performance measures. The proposed model demonstrates effective and scalable eye disease detection, delivering both binary and multi-class classification solutions. The outcomes show how important the suggested EDC model is for disease classification in complicated fundus images of the eye.