Experiments in the automatic detection of optic disc abnormalities using `code-free' software in practical ophthalmology

Abstract

Background to the problem: The accurate diagnosis of optic disc swelling remains a challenge for non-ophthalmic clinicians and community opticians resulting in a large volume of referrals for tertiary review 1. At present, the demand on Ophthalmology services is increasing considerably and it is now the busiest outpatient specialty within the National Health Service, United Kingdom 2. In 2020, Milea et al. developed a bespoke deep learning model to differentiate between papilloedema, pseudopapilloedema and normal optic discs using colour fundus photographs with impressive results 3. Unfortunately, bespoke deep learning depends upon highly advanced technical expertise as well as significant computational and financial resources. Code-free deep learning allows users to develop deep learning models without any coding expertise. It is also largely cloud-based removing the need for expensive computing resources. The aim of this thesis was to investigate the performance of code-free deep learning in the binary classification of 1) the presence or absence of papilloedema and, 2) normal versus abnormal optic discs. The method: A systematic review of the clinical literature was performed to identify all records investigating the application of deep learning to the defined aims of this study. A bespoke dataset of optical coherence tomography images and colour fundus photographs was curated from the Moorfields Eye Hospital NHS Foundation Trust electronic patient healthcare record. A second dataset was identified from a publicly-available repository and downloaded for additional code-free model development and training. Following completion of model training, all models achieving comparable or superior performance to those cited within the clinical literature were subject to external validation. Results: The systematic review identified seven records meeting the inclusion criteria that investigated the application of deep learning to optic disc abnormalities. A variety of deep learning approaches used within the literature were identified, however, none used code-free deep learning. A dataset of 6980 optical coherence tomography images and 6982 colour fundus photographs was curated from the electronic patient record including normal optic disc, abnormal optic disc and papilloedema labels. A second dataset was identified and downloaded for model training. Sixteen code-free deep learning models were trained in total by a clinician with no coding expertise: fourteen models were trained using the bespoke curated dataset and two were trained using the publicly-available dataset downloaded from the online repository. Four code-free deep learning models achieved performances comparable or superior to the bespoke deep learning models previously described. External validation of these models demonstrated poor performances at all tasks. Conclusion: The application of bespoke deep learning for papilloedema and/or optic disc abnormality identification has demonstrated impressive results within the clinical literature. Based on the results of this thesis, code-free deep learning is not presently a feasible alternative to bespoke deep learning for implementation into the direct patient pathway. Code-free deep learning instead may serve a role within the clinical research environment or as a tool for medical education.