DeB5-XNet: An explainable ensemble model for ocular disease classification using feature extraction and Grad-CAM

Abstract

Background and Objective:

Vision serves as a window to the world, enabling individuals to fully appreciate various dimensions of everyday life. Some eye diseases can lead to irreversible loss of vision. Developing an algorithm for a clinical decision support system that explains its predictions is essential to assist the limited number of ophthalmologists in managing the increasing patient load with severe ocular diseases. In contrast to earlier models that concentrated on single-disease classification without providing insights into predictions, this approach introduces DeB5-XNet, a novel explainable ensemble model for multi-categorical classification of ocular conditions.

Methods:

This study presents an ensemble model developed to categorize images into glaucoma (G), cataract (C), diabetic retinopathy (DR), and healthy condition labeled as Normal(N). This proposal operates on three levels: First, the images are enhanced using CLAHE in LAB color space, which improves the model’s predictive capability. Second, an ensemble model is constructed by concatenating features derived from pairs of seven pre-trained models, utilizing their diverse architectures to capture complex characteristics essential for accurate diagnosis. These extracted features are then fine-tuned using a consistent classifier. Third, it has been observed that trust in any diagnostic method is dependent on explainability. Therefore, the selected approach was validated, and its effectiveness was demonstrated using Grad-CAM. The performance of this diagnostic model was evaluated using recall, precision, F1-score, and accuracy metrics.

Results:

The ensemble models outperformed the individual models. DeB5-XNet, an ensemble model that extracted features from DenseNet121 and EfficientNetB5, achieved the highest test accuracy of 95%, notably reducing false negatives compared to standalone models. Remarkably, the model further demonstrated an F1-score of 97% for cataract, 100% for diabetic retinopathy, 90% for glaucoma, and 91% for normal cases.

Conclusion:

The proposed ensemble model, DeB5-XNet shows an improvement over the individual pre-trained models. The Grad-CAM technique demonstrates that the features used by the ensemble model for classification closely align with those identified by ophthalmologists for diagnostic purposes. This alignment strengthens the model’s reliability and potential usefulness in clinical settings.