Evaluating Retinal Disease Diagnosis with an Interpretable Lightweight CNN Model Resistant to Adversarial Attacks

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Manual diagnosis of retinal diseases from Optical Coherence Tomography (OCT) images is labor-intensive.
• There is a need for automated, efficient, and accurate diagnostic tools for retinal anomalies.
• Convolutional Neural Networks (CNNs) show promise for image classification tasks in medical diagnostics.

Purpose of the Study:

• To develop and evaluate a CNN model for classifying OCT images into categories: Choroidal NeoVascularization (CNV), Diabetic Macular Edema (DME), Drusen, and Normal.
• To assess the model's robustness against adversarial attacks using the Fast Gradient Sign Method (FGSM).
• To enhance model interpretability using Explainable AI (XAI) techniques like LIME and SHAP.

Main Methods:

• A lightweight CNN model was trained and validated on an OCT dataset using 10-fold cross-validation.
• The Fast Gradient Sign Method (FGSM) was employed to test the model's resilience to noise, with varying epsilon values.
• Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) were utilized for model interpretability.
• Supplementary datasets (COVID-19, Kidney Stone) were incorporated to improve model robustness.

Main Results:

• The model achieved an average k-fold test accuracy of 94.29% and validation accuracy of 94.12%.
• The CNN model demonstrated robustness, correctly handling noise levels below 0.1 epsilon using FGSM.
• Explainable AI methods provided interpretable insights into the model's decision-making process for retinal images.
• The model's performance was comparable to state-of-the-art methodologies, even after incorporating additional datasets.

Conclusions:

• The proposed CNN model offers an efficient and accurate solution for automated OCT image-based diagnosis of retinal diseases.
• The integration of XAI enhances trust and understanding of the AI model in clinical settings.
• The study highlights the potential of AI in advancing medical diagnostics and supporting Internet-of-Medical-Things (IoMT) applications.