Diabetic eye disease in computerised tomography of feature extraction and classification in hybrid neural network

Figures & data

Figure 1. A suggested block diagram.

Figure 2. Sample images.

Figure 3. Fundus image dataset of histogram equalization.

Figure 4. Suggested IR_ODCNN model for detecting Diabetic Retinopathy.

Figure 5. InceptionV3 model architecture.

Figure 6. Basic ResNet 50 model architecture.

Figure 7. DeepConvolutional neural network.

Table 1. Hyper parameters the suggested model’s value.

Hyper parameter	Value
Loss function	Categorical Crossentropy
No. of. Epochs	70
Batch size	30
Dropout rate	0.9
Learning rate	0.005
Tuning Technique	Learning Rate Scheduler
Convolutional layers	17, 22, 35 with size 6×6
Max pooling layer	3×3
Fully connected layers	700 and 200 units

Figure 8. Sample input images from image net dataset.

Figure 9. Pre-processed image.

Figure 10. Confusion matrix.

Figure 11. Comparison of each class’s performance.

Table 2. Performance metrics determined by observing the confusion matrix and calculating the percentage.

	Accuracy	Precision	Recall	Specificity	F1- score
Normal	99.2	98	98	99.5	98
Mild DR	99.4	98	99	99.5	98.4
Moderate DR	99	97	98	99.2	97.49
Severe DR	99.2	98.9	97	99.75	97.94
PDR	99.2	98	98	99.5	98

Figure 12. ROC curve.

Table 3. The suggested hybrid architecture, ResNet50, and InceptionV3 performance.

Method	Accuracy	Recall	Specificity	Precision	F1 score
Inceptionv3	81.97	93.71	95.12	93.26	92.79
ResNet 50	83.15	91.58	88.29	89.47	92.67
IR_ODCNN	99.2	98	99.49	97.98	97.96

Table 4. Computation time.

Classifier	Computation time
SVM	101.23 s
KNN	251.53 s
Proposed	650 s

Table 5. Comparison between the suggested model and alternative DL models.

Method	Accuracy	Recall	Specificity	Precision	F1 score
SVM	83	61.9	71.5	-	-
NN-CNN	93.23	93	90	-	90
KNN	87	66	91	-	-
GoogleNet	93	96	-	96.5	96.24
AlexNet	87.8	82.6	-	58.46	68.46
Emsemble classifier	95.46	98.11	-	95.29	96.67
Custom CNN	91	91.1	-	88.4	89.72
VGGNet	95.68	86.47	-	88.22	87.33
Proposed	99.2	98	99.49	97.98	97.96