Cat-CNN: Human Eye Cataract Detection from Color Fundus Photograph with Deep CNN with Optimized Cascaded Network

Cataracts are clouding of the lens in the eye, leading to loss of vision that can progress to blindness if not treated. This paper proposed a new method for automatic cataract detection using color fundus images and deep learning methods. A dataset of 1,105 color fundus images labeled by expert ophthalmologists was used in this process. We used seven pretrained CNNs (DenseNet121, EfficientNetB0, MobileNetV2, InceptionV3, Xception, ResNet50, VGG16, and VGG19) for feature extraction before reducing the extracted features using PCA. We used the following combination of machine learning classifiers: SVC, RF, Decision Tree, Gaussian Naive Bayes, XGBoost, K-Nearest Neighbors, and Logistic Regression. For evaluating the models' performance, we used accuracy, precision, recall, F1-score, and computational efficiency. For all metrics, MobileNetV2with Random Forest achieved perfect scores: 100% accuracy, precision, recall, and F1-score, with an average processing time of 669 ms ± 28.8 ms. Thus, it can be applied in real-time applications. EfficientNetB0 with SVC gave an average accuracy of 87.33%, with the rest of the precision and recall metrics above 86%. Then, ResNet50, VGG16, and VGG19 followed with high accuracies between the range of 89.64% to 90.50%. It systemizes the proper choice of architectures of CNNs and classifiers, making the system both accurate and computationally efficient. Future work will include augmentation of the dataset, real-time support in the clinical setting, and advanced techniques for image preprocessing, generative adversarial networks. In addition, the development of an automated annotation tool, improvement of explainable AI, will further improve the deployment of robust AI systems in early diagnosis of cataracts, enhancing the outcome for patients.