AbstractRefractive error is the most common of visual impairments and impacts millions of people globally. Regular vision screening is the recommended best strategy to ensure timely diagnosis and treatment, however, many people do not have access to optometric care and a comprehensive vision examination is inaccessible to many people. There is therefore a need for fast, lowcost and easilyoperate vision screen ing approaches. In this thesis, we aim to investigate the possibility of conducting photorefraction, a common vision screening procedure, on the mobile platform, to address the challenge.
Our approach exploits machine learning algorithms and computer vision techniques. Starting from principles from optometry and prior studies, we create several handcrafted features corresponding detection methods. The experiment results indicate that our detection methods outperform contemporary approaches, leading to a better performance of refractive error measurement and amblyopia risk factor detection. We then move on to pretrained features extracted by convolutional neural networks (CNN). We employ the convolutional layers from multiple pretrained CNN models to encode features and train machine learning models to predict the refractive error. The experiments show promising results, even though the CNN models were not trained on photorefraction datasets.
Given these encouraging results, we further investigate the possibility of data augmentation. One of our challenges is that it is not possible to collect a large amount of data which is enough to train a well performing CNN model from scratch. Therefore, we investigate the use of synthetic data for augmentation. We develop a model of the eye based on the principle of photorefraction, and use it to generate synthetic pupil images with predetermined refractive errors. Evaluation results show that models trained on these synthetic pupil images can achieve similar performance as real images on multiple experiments, which provides solid evidence for the correctness of our photorefraction model.
We finally apply transfer learning to solve the insufficient data issue. CNN models pretrained on large scale public image datasets are finetuned with photorefraction images and the experiments results show large improvement. The CNN models are then trained on more than 10,000 images of synthetic eyes generated via our eye model, and finetuned using real images, achieving performances that outperform all of the previous models. These results support the feasibility of the proposed photorefraction model, and provides a novel direction to obtain training data, which may be extensible to other similar domains.
|Date of Award||2022|
|Supervisor||Grace Ngai (Supervisor)|