الملحق والملخص

الصفحة 195

Abstract Introduction: Keratoconus is a common disease, and its early detection is very important. This is done by carefully reading topographic images to distinguish between keratoconus, suspected and normal corneas. Many studies attempted to discover criteria and indicators that assist in early detection, with varying values of the accuracy indices. But all of the above was based on a tiny portion of the topographic information. After the qualitative leap in computer science, especially the field of artificial intelligence and artificial neural networks, and the emergence of medical applications for reading and analysing medical images, it was necessary to study this technique to detect suspected and keratoconus corneas. Aim of the study: Study the efficiency of using computer vision technology and deep learning algorithms to distinguish between topographic maps images of normal, keratoconus and suspected corneas. Materials and methods: Our study consists of two parts, the first section is a retrospective study of patients' records and images to collect a training sample which consists of 987 eyes (300 cones, 610 natural, 77 suspects). The second section is a cross-section study to collect the test sample, which consists of 422 eyes (13 cones, 366 normal, 43 suspicious). The artificial intelligence system consists of ten Artificial neural networks. Each one is responsible for reading one of the maps (anterior and posterior tangential maps, anterior and posterior sagital maps, anterior and posterior elevation maps, anterior, posterior and equivalent refractive maps, thickness) and predicting the correct class, and its outputs constitute inputs to a final neural network, which is responsible for the final decision of the system. Results: The AI system achieved accuracy and overall Weighted F-score values on the test group (92.2% -91.2%), and there is no statistically significant difference between it and both the program attached to the SIRIUS device and the physician(p> 5%). The physician, with the help of the AI system, achieved the best result (96.2% -95.9%) compared to all other models, and there was a statistically significant difference (p <5%). Conclusions: We recommend that the AI system is adopted as an aid tool to the physician in reading topographic maps.

Keywords: keratoconus, artificial intelligence, deep learning, neural networks

195

الصفحة 196

Syrian Arab Republic Damascus University Faculty of Medicine Ophthalmology Department

Usage of Computer Vision technique and Deep Learning Algorithms to differentiate between Topography Images of normal, keratoconus and suspected corneas

A dissertation submitted in partial fulfillment of the requirements for the degree of Ph.D in Ophthalmology

By Dr.Mohammad zafr allah Askar

Supervisor: Yosra Haddeh Professor of ophthalmology at faculty of medicine, Damascus University

Co-Supervisor: Madhat Alsoos Artificial Intelligence Department, Faculty of Informatics Technology Engineering, Damascus University

2021-2020

196