Abstract
Purpose :
Advances in automatic diabetic retinopathy (DR) screening using artificial intelligence and fundus photography have been rapidly growing. Such systems showed robust diagnostic performance for DR screening, performing as well as the best retina experts. However, these techniques have one major drawback: they can only detect DR, and possibly other common pathologies, such as glaucoma and macular degeneration. In this study, we propose a more general model aiming to automatically detect abnormal fundus images.
Methods :
The proposed approach was developed and evaluated on a selected dataset extracted from OPHDIAT, a telemedical network for DR screening in Il-de-France, consisting of more than 160,000 exams performed between 2004 and 2017. All screening exams were analyzed by one of seven certified ophthalmologists. The structured report included the grading of DR in each eye. It also indicated the presence or suspicion of presence of other pathologies in each eye. For the purpose of this study, 77,812 normal or abnormal fundus images were selected, and annotations were verified by another expert. Abnormal fundus images have at least one of the 41 conditions identified.
Results :
We evaluated our model using 7,855 fundus photographs from a study population with mean age of 55 years (minimum: 9 years, maximum: 99 years and standard deviation: 15 years) and comprising 42% of females and 58% of males. The results show that retinal pathologies were correctly identified with an area under the ROC curve of 0.9557. The algorithm can detect 90% of pathological cases with a specificity 88%. Additionally, the algorithm takes less than two seconds to analyze a retinal image.
Conclusions :
We showed that the proposed AI can automatically differentiate between normal and abnormal fundus images with a good precision. This solution is more general than existing methods, which only focus on a specific pathology such as DR. While guaranteeing a very high sensitivity, this system could be used for automatic triage, referring only the pathological fundus images to ophthalmologists. This model is very promising because it is paving the way towards an ideal screening system, managing all pathologies visible in fundus photographs. In future works, we will investigate its use for mass eye pathology screening in the general population.
This is a 2020 ARVO Annual Meeting abstract.