Page 171 - Kaleidoscope Academic Conference Proceedings 2024

Innovation and Digital Transformation for a Sustainable World

Table 3 – Performance parameters of the Modified VGG16 and VGG16

Classifiers Accuracy (%) Precision (%) Recall(%) F1-score (%)
Modified VGG16 98.4 98.2 98.2 97.6
VGG16 97.3 98.1 95.8 97.0

Figure 5 – Confusion matrix of SVM using test data.
Figure 6 – Region of convergence (ROC) curve of SVM using
test data.
Table 4 – Performance comparison of different models

Methods Accuracy (%) adequate but it needs a high performing GPU enabled system
VGG16 97.3 for execution.
ResNet18 95.3
Deshpande et al. [6] 81.6 5. CONCLUSION AND FUTURE SCOPE
Lahmar et al. [3] 93.0
Lahmar et al. [4] 88.8 In conclusion, DR poses a severe threat to vision when left
Kassani et al. [5] 83.0 untreated, necessitating timely detection and intervention.
Proposed method 98.4 The demand for automated DR detection systems is evident
due to the challenges associated with manual examination
features vs. accuracy graph is plotted in Fig. 4. It is observe of fundus images. Leveraging transfer learning in medical
that best accuracy is obtained by following the selecting image analysis, our study presents a refined VGG16 network
features with p-values less than 0.05. It gives feature length tailored for DR detection, integrating a feature selection
of 3964, hence, in the further experiments, 3964 features are technique. Through meticulous parameter tuning and feature
selected. Different machine learning classifiers are tested for selection, our approach achieves a notable performance boost.
classification and the results of validation and testing accuracy Notably, our modified VGG16 network outperforms both
is summarised in the Table 2. It is analysed that best testing the standard VGG16 and ResNet18 networks, attaining an
accuracy of 98.4% is obtained using SVM classifier. Fig. 5 impressive accuracy of 98.4% on a benchmark DR dataset.
and Fig. 5 show confusion matrix and ROC curve using SVM This underscores the potential of our approach in enhancing
for the testing data. In the confusion matrix, x-axis shows automated DR detection systems for clinical use.
predicted class and y-axis shows true class. ROC curve is The future of AI-driven personalized e-services in DR
plotted for true positive rate vs. false positive rate. Area detection holds promise for revolutionizing healthcare
under curve (AUC) value of 0.9984 is obtained. delivery. The boom in telecommunication network and
Table 3 shows the detailed parameter values of modified and enhanced data rate owing to 5G and 6G technology, AI
standard VGG16 network. It is concluded that by integrating can further enhance teleophthalmology services, making
feature selection in the VGG16 network, performance is eye care more accessible, especially in remote areas.
significantly improved. Results are summarised in Table 4 AI-powered tools can enable remote screening for eye
in terms of accuracy. It is observed that proposed method is disorders, allowing patients to upload images of their eyes
outperforming standard VGG16 as well as ResNet18. Results for analysis. This is particularly useful in areas with
are also compared with the state-of-the-art works, such as,[6], limited access to ophthalmologists. AI systems can assist
[3], [4] and [5]. It is observed that with the proposed method in virtual consultations by providing preliminary diagnoses
is giving performing best with the accuracy of 98.4%. and treatment suggestions, which can then be reviewed by
Although the performance of the proposed method is a human specialist. For patients already diagnosed with

– 127 –