GAN-Enhanced Hybrid Deep Learning with Explainable AI for Automated Cataract Diagnosis

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Cataracts are a leading cause of vision impairment, necessitating accurate and timely diagnosis.
• Current automated cataract detection systems face challenges with data variety, interpretability, and real-world generalization.
• Explainable AI (XAI) and Generative AI (GenAI) offer potential solutions to these limitations.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for enhanced cataract detection.
• To improve the interpretability and generalization capabilities of automated diagnostic systems.
• To leverage GenAI for data augmentation and XAI for clinical transparency.

Main Methods:

• A fine-tuned InceptionResNetV2 model was trained on a hybrid dataset, including real and Generative Adversarial Network (GAN)-generated synthetic images.
• Class weights and stratified K-Fold cross-validation were employed to manage data imbalance and ensure robust model evaluation.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was used for graphical interpretation of model predictions.

Main Results:

• The model achieved a mean K-Fold accuracy of 97.58% (SD=0.0040).
• On an external dataset, the model attained 97% overall accuracy, 0.9944 AUC, and high precision (96%), recall (94%), and F1-score (95%) for the cataract class.
• The system demonstrated superior performance and interpretability compared to existing methods.

Conclusions:

• The proposed deep learning approach, integrating GenAI and XAI, significantly enhances cataract detection accuracy and reliability.
• The method addresses data limitations and improves model interpretability, fostering greater clinical trust and facilitating expert validation.
• This novel system offers a promising advancement over current automated diagnostic tools for cataracts.