Convolutional Neural Networks for Diabetic Retinopathy Grading from iPhone Fundus Images

Diabetic eye diseases is a major issue in Europe and the prevalence of visual impairment and blindness caused by Diabetic Retinopathy (DR) has significantly increased in the last decades. Efficient screening and early diagnose of DR by family physicians would help to reduce costs in health systems and shorten waiting lists, thus decreasing patients' emotional stress. In this sense, the use of portable image devices (e.g., a mobile phone with a specific fundus image capturing device attach to it) combined with AI-based systems arise as a powerful tool to address this problem. This paper develops 2 well-known pre-trained Convolutional Neural Networks and fine-tune them on a local Spanish cohort and 3 more publicly available fundus image dataset for DR grading. The models trained were evaluated on fundus images captured using an iPhone mobile within the local Spanish cohort. The results of the analysis showed how in one of the settings tested, one of the models was able to surpass human-level performance achieving an AUC of 0.679 in comparison to an AUC of 0.667 achieved by ophthalmologists when diagnosing the grade of DR on the same iPhone fundus images, although further work and improvements need to take place in order to consider it for a realistic deployment in the daily clinical practice.