Table 7.

DR Detection Tools, Techniques, Methodologies, and Performance Evaluation (II).

Ref. No	DL Method	Major Focus	Environment	Performance Criteria
[41]	Graph-NN	Propose GNN model which consists of two features. One is to extract region-of-interest focusing only regions to remove noise while preprocessing and others in applying GNN for classification	No	Accuracy = 80%
[42]	CNN	Constructed a model in which artificial neurons are organized in a hierarchical manner which are able to learn multiple level of abstraction	No	Accuracy = 79.3%
[43]	Deep CNN	Use VGG-16 DCNN to automatically detect local features and to generate a classification model	No	Specificity 97%, Sensitivity 96.7%
[44]	CNN	Demonstrates the potential of CNN to classify DR fundus images based of severity in real times	No	AUC 96.6%, Specificity = 97.2 %, Sensitivity = 94.7%
[45]	Ensemble DNN	Build high quality medical imaging dataset of DR also propose a grading and identification system called DeepDR and evaluate the model using nine validity matrices	No	Specificity = 92.29%, Sensitivity = 80.28%
[46]	Modified Hopfield NN	Propose Modify Hopfield neural network to handling drawbacks of conventional HNN where weigh values changed based on the output values in order to avoid the local minima	Keras	AUC = 90.1 %, Sensitivity = 84.6, 90.6, Specificity = 79.9, 90 %
[47]	Deep CNN	DCNN pooling layer is replaced with fractional max pooling to drive more discriminative features for classification also use SVM to classify underlying boundary of distribution. Furthermore build an app called Deep Retina	No	Accuracy = 99.25%, Specificity = 99.0%
[48]	DCNN	Data driven features learned from deep learning network through dataset and then these deep features were propagated into a tree based classification model that output a final diagnostic disease	No	Accuracy = 86.71, Specificity = 90.89, Sensitivity = 89.30
[49]	DNN	Propose Alex Net DNN with caffeNet model to extract multi-dimensional features at fully connected DNN layers and use SVM for optimal five class DR classification	No	AUC = 97%, Sensitivity = 94, Specificity = 98
[50]	CNN	Build an automated system to detect DR IDX-DR X2.1 composed of client software and analysis software. The device applied a set of CNN based detectors to examine each image	No	Accuracy = 97.93
[51]	DCNN	Proposed a systematic computation model using DCNN for DR classification and assessed performance on non-open dataset and found that model achieves better results with only a small fraction of training set images	No	AUC = 98.0%, Sensitivity = 96.8, Specificity = 87.0
[52]	CNN	Employ LSTM, CNN, and their combination for extracting complex features to input into heart rate variability dataset.	No	Sensitivity = 88.3, Specificity = 98.0