HYBRID GANNET-PARTICLE SWARM OPTIMIZATION ENHANCED DENSENET WITH CATA-NET WEIGHTS FOR EARLY CATARACT DIAGNOSIS FROM SLIT-LAMP EYE IMAGES

To develop an early and accurate cataract diagnosis system to prevent vision loss and enable timely surgical intervention, particularly in underserved clinical settings.

The study introduces a new ophthalmic dataset, CATA-NET, consisting of 1,480 expert-annotated slit-lamp eye images classified into four categories: Normal, Mild, Moderate, and Severe cataract. This dataset is used to enhance the specificity of deep learning models for ocular features.

We propose a novel deep learning framework, GPSO-DenseNet, which integrates DenseNet121 with a hybrid metaheuristic optimization algorithm combining Gannet Foraging and Particle Swarm Optimization (GPSO). The model utilizes CATA-NET weights, pre-trained on this dataset, to replace generic ImageNet parameters. GPSO jointly tunes hyperparameters (learning rate, batch size) and refines feature extractor weights via adaptive exploration-exploitation strategies. An attention-guided reduction block is embedded within the architecture to preserve spatial information and mitigate gradient vanishing. Comprehensive evaluation techniques, including cross-validation, confusion matrix analysis, Grad-CAM visualization, and ablation studies, are used.

The GPSO-DenseNet model achieved 93.2% accuracy, 0.91 F1-score, and 0.92 recall, outperforming ResNet50, MobileNetV3, and VGG16 by a significant margin.

The results demonstrate that hybrid optimization and domain-specific pretraining enhance cataract classification accuracy, offering an interpretable, accurate, and scalable solution suitable for teleophthalmology and resource-limited environments.