ISSN: 2578-465X
Authors: Carolyn YTW*, Timing L, Tin LW and Henry HWL
Purpose: Myopia and glaucoma are major causes of vision impairment that are expected to rise in incidence in East Asia. Artificial intelligence (AI)-aassisted mass screenings may help reduce disease burden and improve long-term prognosis. CFDL is a deep learning (DL) subtype that facilitates non-AI-experts to derive their own AI models. With its resource-saving nature and ease of development, it may benefit resource-limited screening settings. This study evaluated the performance of CFDL in performing pathological myopia (PM) and glaucoma screening on colour fundus photographs (CFP)s. Methods: We used labelled CFPs from the ODIR-K dataset to develop our CFDL algorithm using Google’s CFDL platform i.e. Vertex AI. 3374 normal, PM and glaucoma CFPs were identified and uploaded to Vertex AI. The uploaded images were split into 8-1-1 for training, validation, and testing. Our model’s performance was later compared to the state-of-the-art DL models identified through our targeted literature search. External validation of the model was performed on an independent cohort of CFPs retrieved from other datasets. Results: At the 0.5 confidence threshold, our CFDL model achieved an area under receiver operator curve (AUROC) of 0.998, accuracy of 90.74% and recall of 90.74%. The sensitivity ranged from 94.44% (PM detection) to 40% (glaucoma detection). When externally validated, the model had a lower AUROC (0.863), accuracy (77.78%), and recall (77.78%) at the 0.5 confidence threshold. Conclusion: The study demonstrated the feasibility of a highly accurate CFDL model for PM and glaucoma screenings on CFPs.
Keywords: Artificial Intelligence; Image Segmentation; Retinal Blood Vessel; Colour Fundus Photograph
