Utilising a Mobile Device to Aid in Early Diagnosis of Corneal Abnormalities

The cornea is the clear front surface of the eye that allows light to enter and provides up to 75% of the eyes focusing power. There are many corneal abnormalities, and early diagnosis for many patients is paramount to preserve sight and avoid corneal transplantation. At this stage, diagnosis predominately relies on specialised topographic imaging equipment and evaluation from an experienced ophthalmologist. The hypothesis investigated in this research is that an image taken from a mobile device can be reconstructed into a corneal topographic image and used for automatic diagnosis of abnormalities. To prove the hypothesis, this project differentiated between patient images with and without a common corneal condition called Keratoconus (KC). Through critical analysis of the literature, this research determined machine learning and pre-processing techniques would be adopted. A database of patient images with and without a diagnosis of KC was obtained from the Australian Study of Keratoconus. Pre-processing algorithms and neural networks were developed for this data. The optimum pre-processing techniques and neural network parameters were selected for the final model based on verification with the available database. The final results were promising as it can differentiate between a patient with and without KC 90.9% of the time. This testing was with unique data not used in the learning process. Additionally, a mobile app was developed, which enabled a user to select an advanced topographic image and observe the models prediction. This project has shown promising results, indicative that the research hypothesis may be sustainable. Subject to data limitations, further work would show improvements to this research. This project suggests that additional research with a more extensive database be performed to confirm the results and verify correlations with model predictions and KC severity. Furthermore, if a diverse database of patient images from various topographic imaging equipment is available, more robust classifiers could be created. These classifiers could integrate with mobile topographic technology to further confirm this hypothesis