Eye Disease Classification Using Tetrolet Transform Based Wavemix Architecture: A Comprehensive Multi-Scale Analysis With Deep Learning Integration

Abstract

Background: Eye diseases represent a significant global health burden, affecting over 2.2 billion people worldwide. Early diagnosis through automated classification systems is crucial for preventing vision loss and improving patient outcomes.

Objective: This study proposes an enhanced WaveMix architecture integrating Tetrolet transforms with pre-trained deep learning models for accurate multi-class eye disease classification.

Methods: We developed a novel hybrid approach combining WaveMix architecture with four different transform techniques: Wavelet, Contourlet, Curvelet, and Tetrolet transforms. The framework was evaluated using three pre-trained models (ResNet-18, MobileNetV2, and EfficientNet-B0) on a comprehensive dataset of 9,825 fundus images across six disease categories. Advanced visualization techniques including gradient-weighted class activation mapping (Grad-CAM), confusion matrices, and statistical significance testing were employed for comprehensive evaluation.

Results: The Tetrolet-based WaveMix architecture achieved superior performance with accuracies of 96.95%, 96.69%, and 97.17% for ResNet-18, MobileNetV2, and EfficientNet-B0, respectively. The best-performing model (Tetrolet + EfficientNet-B0) demonstrated exceptional metrics: 97.17% accuracy, 0.7592 sensitivity, 0.9945 specificity, and 0.99 AUC-ROC, with statistical significance (p < 0.001) compared to traditional approaches.

Conclusions: The proposed Tetrolet-based WaveMix architecture significantly outperforms conventional methods, offering a robust, computationally efficient solution for automated eye disease diagnosis with clinical applicability.