6533b839fe1ef96bd12a6cc1

RESEARCH PRODUCT

Automated Diagnostics of Retinal Pathologies Using OCT Volumes

subject

description

The leading cause of blindness in the population could mostly be the degeneration of the retina caused by the diabetic-related problems and the aging issue. Diabetic retinopathy (DR) and diabetic macular edema (DME) are the main direct causes of vision problems in the labor age citizens of most advanced countries. The elevated number of diabetic people globally indicates that DME and DR will remain to be the principal factor to partial or total vision loss, which affects the lives quality of patients for many years to come and threaten their lives. Therefore, early detection followed by fast treatment procedures of persons with diabetic-related diseases is significant in preventing optical problems and can decrease the risk of blindness. In addition, people above 50 are exposed to age-related macular degeneration (AMD) disease that hits the retina. Therefore, researchers over the world have attracted to the differences related to several retinal diseases. Several automated methods using Artificial Intelligence (AI) (varying from traditional computer vision to advanced machine learning algorithms) have been applied for the detection and examination of retinal diseases. Unluckily, these models are able to be mistaken with computational inability, which necessitates additional interference from specialist personal. This thesis presents an automatic method - based on deep learning neural networks algorithms - to detect DME and DR, which allows overstepping the subjective handy evaluation of ophthalmologists. Based on Convolutional Neural Network, a proposed model is presented with a soft-max classifier and fully trained from scratch for the automatic classification of Optical Coherence Tomography (OCT) retinal images where OCT screening techniques are applied as the current dependable assessment and measurement method to discover the existence of swallow in the retina. This model has the ability to detect patterns for DR and DME using these retinal images with improved accuracy and sensitivity. Moreover, a pre-trained model has been fined-tuned and re-trained using a dataset that has been augmented using Generative Adversarial Networks (GANs). In opposite to manual retinal disease diagnosis based on personal clinical examination and analysis of OCT images, this method showed the capability to automatically predict DME diseased cases versus healthy cases. The experiments have been evaluated over several datasets provided by different institutions. The model, compared to other CNN end-to-end or transfer learned models, shows effective extracting features, with less time consumption, based on an efficient data pre-processing stage. The experimental results showed a higher accuracy of classification which is promising in the field of early detection of diabetic diseases to aid ophthalmologists in biomedical technologies.