The classification of corneal arcus images by using improved convolutional neural network

• Main Author: Abdul Halim, Nur Farahin
• Format: Thesis
• Language: English; English
• Published: 2024
• Online Access: http://eprints.utem.edu.my/id/eprint/28325/1/The%20classification%20of%20corneal%20arcus%20images%20by%20using%20improved%20convolutional%20neural%20network.pdf; http://eprints.utem.edu.my/id/eprint/28325/2/The%20classification%20of%20corneal%20arcus%20images%20by%20using%20improved%20convolutional%20neural%20network.pdf; http://eprints.utem.edu.my/id/eprint/28325/; https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=124261

Corneal arcus (CA), also known as arcus senilis or simply arcus, refers to a condition characterized by the accumulation of lipids, particularly cholesterol, in the cornea of the eye. The presence of CA may be an indicator of underlying systemic conditions such as high cholesterol in human body. When CA is detected during an eye examination, it may prompt further investigation into the patient's lipid profile and overall cardiovascular health. Individuals with CA may be advised to undergo blood tests to measure cholesterol and triglyceride levels. The most common test is known as a lipid panel or a cholesterol blood test. Therefore, leveraging image processing offers a non-invasive and painless alternative to traditional blood tests for detecting corneal arcus. This research is about the implementation of convolutional neural network (CNN) in detecting cholesterols presence by classifying normal and CA images. A dataset of 459 images comprising 237 for normal and 222 for CA images were formed. There are three different CNN models were proposed for feature extraction and classifying the normal and CA images which are CNN, Resnet-50 and VGG-19. From the parameter evaluation, it can be concluded that batch size of 20 and learning rate of 0.0001 suit with Resnet-50 and CNN model, while VGG-19 suit with batch size of 10 and learning rate of 0.00001 to classify with the normal and CA dataset. The best result was exhibited by Resnet-50 with 10-fold cross-validation producing high average detection in terms of sensitivity, specificity, and accuracy of 100%. Thus, deeper networks implementation is recommended in the future to further improve CA localisation in cholesterol detection.