Abstract:
Retinal diseases are a leading cause of vision impairment globally, necessitating
early and accurate diagnosis. This study proposes an explainable multiclassification
system for retinal fundus images using ensembled transfer learning
models and Gradient-weighted Class Activation Mapping (Grad-CAM). Retinal
images were preprocessed using a pipeline comprising Contrast Limited Adaptive
Histogram Equalization (CLAHE), morphological erosion, and bilateral filtering.
Data augmentation through random transformations was applied for class balance
and model robustness. Four pretrained architectures—ResNet50, VGG19,
EfficientNetB5, and DenseNet201—were evaluated in both baseline and Bayesianoptimized
configurations. Performance was assessed via five-fold cross-validation
using sensitivity, specificity, accuracy, F1-score, and ROC-AUC metrics. The top
three models based on ROC-AUC (EfficientNetB5: 97.36%, EfficientNetB5 Optimized:
96.96%, DenseNet201 Optimized: 96.40%) were ensembled. Among ensemble
strategies, soft voting outperformed hard voting, achieving the highest test
accuracy of 88.00% and a macro F1-score of 87.85%. Grad-CAM visualizations
provided class-specific interpretability by highlighting pathological regions within
fundus images. A Streamlit-based graphical user interface was developed, enabling
users to upload retinal images and receive real-time classifications, class probability
scores, and Grad-CAM heatmaps. The proposed system demonstrates strong
potential as a clinical decision-support tool and educational platform, combining
high classification performance with visual interpretability.
