Table 1.
Review of deep learning applications in diabetic retinopathy and other datasets.
| Reference | Dataset | Method used | Evaluation metrics | Research challenges |
|---|---|---|---|---|
| [19] | Diabetic retinopathy (DR) dataset consisted of 75137 images | 5-Fold cross-validation and data-driven deep learning algorithm | Sensitivity, specificity, and AUC score | The results were not properly evaluated using typical state-of-the-art models |
| [20] | 73 patients (122 eyes) were evaluated, 50.7% men and 49.3% women | RBM-1000, RBM-500, and OPF-1000 | Sensitivity measured, specificity, and accuracy | More in-depth analysis on larger datasets was missing and accuracy may also be improved |
| [21] | 14,186 retinal images and Messidor dataset with 1200 images | Deep learning algorithm | Accuracy, sensitivity, specificity, positive and negative predictive values, and AUC | Dataset is fixed and is not compared with other technique |
| [22] | 128175 retinal images, EyePACS-1 dataset consisted of 9963 images, and Messidor-2 dataset with 1748 images | Deep convolutional neural network | The algorithm had 97.5% and 96.1% sensitivity and 93.4% and 93.9% specificity in the 2 validation sets | Limited dataset, system maybe failed to learn more complex features |
| [23] | Heart disease dataset | Effective heart disease prediction system using enhanced deep genetic algorithm and adaptive Harris hawks optimization-based clustering | Accuracy, precision, recall, specificity, and F-score | Requires more improvement in the learning process |
| [24] | COVID-CT-dataset: 349 and 397 images and CT scans for COVID-19 classification: 4,001 and 9,979 images | Hybrid learning and optimization approach CovH2SD-CovH2SD uses DL. HHO algorithm to optimize the hyperparameters | Accuracy, precision, recall, F1-score, and AUC performance metrics | Not good for multiclass classification |
| [25] | Hand gesture dataset from Kaggle repository | HHO is used for hyperparameter tuning of CNN for enhancing hand gesture recognition | Reduction of the burden on the CNN by reducing the training time and 100% accuracy for hand gesture classification is attained | Requires more improvement in the learning process |