Abstract
Identification and assessment of optic disc (OD) are essential in any retinal imaging analysis. Retinal imaging allows ophthalmologists to diagnose vision-threatening diseases such as glaucoma. Diagnose of glaucoma mainly depends on OD assessment. Cup disc ratio (CDR) is a common measurement to evaluate glaucoma. Only very few works have been done so far on CDR and OD assessment, with limited scalability and applicability. In this research, we proposed multiple machine learning and deep learning approaches for OD identification and assessment. This not only has outperformed the current algorithms in terms of scalability and efficiency but has filled the gap in the literature for OD abnormality assessment. First, we proposed two pre-analysis autoencoder-based models to filter out irrelevant images and to identify retinal images from left or right eyes. The retinal filtering model achieved high accuracy (99.94%). The novelty of the left and right eye classifier on utilising the OD region allowed the model to classify images where other existing approaches are proved either inefficient or inapplicable. Second, we proposed an OD localiser based on multiple classifiers and convolutional neural network (CNN). The evaluation of this model has been carried out extensively on multiple databases with high accuracy (95.87%). This model also successfully detected OD on video clips captured through a mobile device, and to the best of our knowledge, this is the first work on this type of retinal data. Third, we explored CNN architecture in multiple dimensions for CDR estimation. The proposed CNN ensemble-based model demonstrated not only human level accuracy but also its ability to measure CDR on images captured from a low-cost mobile camera. Finally, we developed CNN-based classification and regression models for OD abnormality analysis. The later has demonstrated its effectiveness with high specificity (99.96%) and sensitivity (98.67%). These results are very promising and demonstrate the potential of the proposed approaches to support remote healthcare in many regions where ophthalmologists are scarce.