|
Ref. | Dataset | Used method | Outcomes and metrics | Pros and cons | Research challenges |
|
[6] | MESSIDOR-2 fundus image dataset | CNN for deep learning | Accuracy, AUC, sensitivity, and specificity | Pros: (i) Efficient screening of DR Cons: (i) Not cost effective | Diversified image datasets are not considered |
[7] | Diabetes image dataset + digi fundus Ltd. from Finland | (i) Deep learning CNN based on five-stage diabetic retinopathy screening (ii) Clinical grading system on macular edema | AUC, sensitivity, and specificity | Pros: (i) Ability to increase size of image without modifying structure Cons: (i) Biased results in image feature grading | Inclusion of high-resolution image data but exclusion of low-resolution image data |
[8] | EyePACS dataset | Linear-SVM in VGG-Net | Sensitivity and specificity | Pros: (i) Easier identification of DR Cons: (i) Results lack required satisfaction level | Severity level of the disease is not included |
[9] | OCT (optical coherence tomography) images | Three-layered CNN | Accuracy, specificity, sensitivity, precession, recall, F1 score, and kappa score | Pros: (i) Efficient ocular structure identification Cons: (i) Exclusion of variable eye structure images in the dataset | (i) Use of OCT images alone in implementation (ii) Exclusion of fundus images |
[10] | DIRETDB-1, MESSIDOR-2, and FAZ images dataset | (i) Deep visual features (DVF) + D-color-SIFT GLOH technique (ii) Semisupervised multilayer deep learning algorithm | ROC, sensitivity, specificity, and training errors | Pros: (i) Deep visual features are used for reducing time and image uncertainty problems Cons: (i) Inability to perform multiple image classifications at a time | Inability to classify multiple images at a time |
[11] | Ultrawide field fundus photographs from St. Mary’s hospital in South Korea | ResNet architecture | Accuracy, AUC, sensitivity, and specificity | Pros: (i) Ease in training of huge retinal surface image dataset Cons: (i) Model dependent on single data source for images collection | Segmentation of the huge retinal surface fails to cover mid and far areas of the retina |
[12] | RGB images from public MESSIDOR dataset | CNN models-AlexNet and VGGNet-16 SqueezeNet | Sensitivity, specificity, and accuracy | Pros: (i) Enhanced classification accuracy by the pretrained model Cons: (i) Methodology is much generalized to handle diversified problems | Implementation confined to specific problem, generalized conclusion not appropriate |
[13] | IDRID (Indian association of diabetic retinopathy dataset) and MESSIDOR dataset | Shapley additive explanation (SHAP) multistage transfer learning approach, (EfficientNet-B4, B5, and SE-ResNeXt50) | Quadratic weighted Cohen’s kappa score, sensitivity, and specificity | Pros: (i) Achievement of decreased variation and improved generalization Cons: (i) Lack of stability | Effective hyperparameter optimization method are not included |
[14] | IEEE dataset (IDRID) | Deep-CNN model with lesion detection algorithm | ROC, accuracy, and efficiency | Pros: (i) Minimizes darkness of the retinal lesions Cons: (i) Limitations in the data source | Implementations on larger datasets are not performed |
[15] | Two datasets, Eye-PACS and 13 medical centers in Thailand | Inception-V3 architecture with one-field and three-field | AUC including risk factors | Pros: (i) Ability to identify several hazards relevant to DR prognosis Cons: (i) Does not include features relevant to DR risks | Deep learning model lacks efficiency, risk analysis, and is not cost efficient |
|