Table 6.
DR Detection Tools, Techniques, Methodologies, and Performance Evaluation (I).
| Ref. No | DL Method | Major Focus | Environment | Performance Criteria |
|---|---|---|---|---|
| [14] | CNN | Propose CLEAR-DR CAD system via deep radiomic sequencer |
No | Accuracy = 73.2% |
| [23] | CNN | Propose Siamese-like CNN architecture which accepts input as binocular fundus images |
No | AUC = 95.1%, kappa score = 82.9 |
| [24] | BNCNN | Redesign the LeNet model by adding batch normalization layer with CNN to effectively preventing gradient diffusion to improve model accuracy |
No | Accuracy = 97.56% |
| [25] | DNN | Proposed modification of Inception-V3 model to grade four severity levels of DR |
MXNET | Accuracy = 88.73%, precision = 95.77, Recall = 94.84 |
| [26] | Ensemble CNN | Combine five models; Resnet50, Inceptionv3, Xception, Dense121, and Dense169 |
Keras, Tensorflow |
Accuracy = 80.8%, Precision = 86.72, Recall = 51.5, F-score = 63.85 |
| [27] | Ensemble CNN | Combine three models: inceptionv3, Xception, and inceptionResNetV2 |
Keras | Accuracy = 97.15%, Precision = 0.96, Recall = 0.96, F1-score = 0.96 |
| [28] | WP-CNN | Build various weighted path CNN networks and optimized by backpropagation. WP-CNN105 achieves the highest accuracy. |
No | Accuracy = 94.23%, F1-score = 0.9087 |
| [29] | Ensemble CNN | Five ensemble models VGG-16, ResNet-18, SE-BN-inception, GoogleNet, and DenseNet were used as benchmark for DR grading |
Caffe | Accuracy = 82.84% |
| [30] | OCTD-Net | Develop novel deep network OCTD-NET. Consist of two features one for feature extraction and other for retinal layer information |
Keras | Accuracy = 92% |
| [31] | GoogLeNet | Propose modification of GoogLeNet convolutional neural network |
No | Accuracy = 98% |
| [32] | Ensemble CNN | Ensemble CNN VGG net and ResNet models used as ensemble |
No | AUC = 97.3% |
| [33] | Deep Multi-Instance Learning |
Image patches extracted from the preprocessing step regularly and then fed into CNN based patch level classifier |
MatConvNet | Precision = 86.3, F1-score = 92.1 |
| [34] | Fully connected Network | Construct U-Net based regional segmentation and diagnosis model |
Keras | PM coefficient is 2.55% lower |
| [35] | DCNN | Transfer learning used for initial weight initialization and for feature extraction |
No | Accuracy = 93.6%, 95.8% |
| [36] | DR-Net | Develop DR-Net framework by fully stacked convolution network to reduce overfitting and to improve performance |
imageMagick, OpenCV |
Accuracy = 81.0% |
| [37] | Ensemble CNN | Three models, inception V3, Resnet152, and inception-Resnet-v3 put together that work individually and Adaboost algorithm is used to merge them. |
Ubuntu | Accuracy = 88.21% |
| [38] | DNN | Neural network with 28 convolutional layers, after each layer batch normalization and ReLu applied except the last one Network trained with inception-v3 model |
Tensorflow, Android studio |
Accuracy = 73.3% |
| [39] | CNN | Network consists of range of convolutional layers that converts pixel intensities to local features before converting them to global features |
No | Accuracy = 97.8% |
| [40] | CNN | Propose CNN model with the addition of regression activation map |
No | Accuracy = 94%, 80% |