Table 1 Summary of common methods for DR detection.
From: An enhanced diabetic retinopathy detection approach using optimized deep learning technique
Category | Methodology | Pros | Cons | Representative examples |
|---|---|---|---|---|
Traditional Machine Learning (Hand-crafted Features) | Manual extraction of texture, color, morphological, or vessel-based features; classification using SVM, k-NN, Random Forest. | - Simple and interpretable. -Requires smaller datasets. - Low computational cost | -Cannot capture complex patterns -Depends heavily on feature engineering -Limited generalization on diverse datasets | SVM + morphological features; Random Forest + texture features; k-NN + histogram descriptors |
Convolutional Neural Networks (CNNs) | End-to-end deep learning models that automatically learn hierarchical features; includes transfer learning and attention models. | -High accuracy -No manual feature extraction -Strong representational power | -Requires large labeled datasets -High computational demands -Less interpretable Sensitive to image quality | InceptionV3 for DR grading; ResNet-based lesion detection; EfficientNet DR classifiers |
Hybrid (Feature Selection + Classifier) | Features (hand-crafted or deep) reduced using GA, PSO, ACO, GWO, WOA before classification. | -Removes noisy/irrelevant features. - Improves classifier performance - Works with smaller datasets | - Risk of local optima - Requires tuning optimizer parameters - Moderate computation cost | GA + SVM; PSO + Random Forest; ACO + k-NN; GWO + SVM |
Advanced Meta-heuristic Optimization | Uses advanced/chaotic optimizers such as WOA, MVO, SCA, DGOA for high-dimensional feature selection. | - Strong exploration/exploitation - More robust in high-dimensional search - Less premature convergence | - Higher computational complexity - Sensitive to parameters - Variable performance across datasets | WOA + deep features; MVO for DR staging; SCA + RF; DGOA + Ensemble |
Ensemble Learning | Combines multiple classifiers using bagging, boosting, or stacking (RF, AdaBoost, XGBoost, meta-learners). | - High robustness and accuracy - Reduces variance and bias - Better generalization | -Less interpretable - More computationally intensive - Requires careful design of base learners | Random Forest; XGBoost + deep features; Stacking (SVM + RF + CNN features) |
