Abstract
Purpose :
To present performance results of an automatic diabetic retinopathy screening software in data collected at a network of comprehensive diabetic care clinics in Monterrey, Mexico.
Methods :
VisionQuest has developed the Diabetic Retinopathy Risk Analysis Computer Software (DR-RACS) to screen retinal images for the presence of diabetic retinopathy. DR-RACS processed retinal images collected at five clinics that are part of Clinicas del Azucar, Monterrey, Mexico, between February and July, 2015. DR-RACS generated a label for each case: “return in 12 months” for no DR or Mild NPDR, “return in 6 months” for moderate NPDR, “refer to ophthalmology” for severe NPDR or PDR, or “inadequate.” A case is composed of at least one image of the macula and one image of the optic disc from each eye. A case is inadequate if at least one eye does not have a macula centered image with sufficient quality due to small pupils, media opacities, or technical errors.
The images are part of a teleretinal screening system that uses Canon CR-2 non-mydriatic retinal cameras, a cloud-based PACS, and a remote reading center. Images are manually read within 48 hours according to the International Clinical Diabetic Retinopathy Grading Scale. Follow up with ophthalmology is arranged by the clinics.
Performance of the DR-RACS was measured against the ground truth of the readers.
Results :
DR-RACS processed 1,203 cases of which 396 (33%) were rejected for low quality (half due to technical errors) with the results shown in Figure 1.
The sensitivity for detection of moderate NPDR was 81% and the specificity was 69%. The sensitivity for referral of severe NPDR cases or worse was 95%. For suspicion of CSME, DR-RACS labeled 32 of the 38 cases as “refer to ophthalmology” and six as “return in 6 months.”
Conclusions :
The DR-RACS achieved clinically safe and effective performance for the detection of diabetic retinopathy in a clinical setting. Using DR-RACS would reduce reading workload by up to 54% and referrals to ophthalmology by almost 70%. Resource scarcity and limited access to care in places like Mexico make automation a viable alternative to screen their diabetic population.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.