Automated Screening of Diabetic Retinopathy Using a Smartphone-based Camera

Overview

The prevalence of diabetic retinopathy (DR) in the UK is on the rise. Within 20 years of diabetes diagnosis, nearly all people with type 1 and almost two thirds of people with type 2 diabetes (60%) have some degree of DR. NHS guidelines mandate annual DR screening in all patients aged 12 and above to prevent complications of DR. Screening for DR in England involves labour-intensive manual grading of retinal images through the teleophthalmology platform. Automated retinal image analysis systems with the use of artificial intelligence (AI) may offer an alternative to manual grading. The purpose of this study is to evaluate the performance of a portable, hand-held fundus camera with integrated artificial intelligence for diabetic retinopathy screening by comparing it against the current standard i.e diagnosis provided by trained human graders evaluating the standard photographs/ophthalmologists.

Study will enroll diabetic patients (new and established) meeting the inclusion criteria and who would provide an informed consent to participate in the study. This would include subjects who visit the endocrinology clinic and had either their retina evaluated by a teleophthalmology consult or visited the ophthalmology clinic for an indirect ophthalmoscopic fundus evaluation in the past 3 months. Subjects whose fundus has been recently evaluated in the past 3 months will also be re-called to the endocrinology clinic to get a fundus image captured on the Remidio NM FOP-10 device. Images will be captured by an operator who will be provided training in using the device prior to the study. Non-mydriatic fundus images are captured using a pre-defined imaging protocol. One disc centered and one macula centred 45 deg image is taken for each eye. The images are then submitted to an AI system for analysis that runs locally on the Remidio FOP. An initial image quality assessment is performed by the AI. If the image is of sufficient quality then the AI will give an output of either 'signs of DR detected' (that would include moderate NPDR, severe NPDR, PDR and/or DME present) or no signs of diabetic retinopathy detected (that would include no DR, mild DR, absence of DME). If the image is of insufficient quality, then the operator makes 2 more attempts using the same imaging protocol. The images that are output as insufficient quality will be noted. All subjects who have provided consent for dilation are then instilled with a single drop of 1% tropicamide solution. After 20 mins, mydriatic 2 field (one disc centered and one macula centered) fundus images are captured and submitted to the AI for analysis along the same lines of non-mydriatic imaging. If the image is of insufficient quality then the operator makes 2 more attempts using the same imaging protocol. The images that are output as insufficient quality will be noted. De-identified patient data including age, gender, ethnicity and lens status (Cataract grade, pseudophakia, aphakia) along with the final diagnosis made by a certified grader/ophthalmologist during clinical examination/teleophthalmology are retrieved from the NHS database. This DR diagnosis will act as a reference standard against which the AI output will be compared. Data entry of subjects will be performed by approved trained Site Personnel. The de-identified data is stored onto an electronic data capture system (EDC). Approved Site Personnel will have a user specific log-in name and password to access the electronic data capture (EDC) system in order to enter study data. The principal investigator will ensure that all aspects of the study are complied with. The study will be conducted in accordance with applicable regulatory requirements and established rules for Good Clinical Practice (GCP).