Automated image analysis for Diabetic Retinopathy Screening with iPhone-based fundus camera

Abstract

Purpose : Vision loss from diabetic retinopathy (DR) in ever-increasing population of diabetic patients can be prevented by early screening and diagnosis. To meet this growing need for screening, we present a cost-effective, end-to-end, point-of-care DR screening setup comprising a) iPhone based retinal camera, Ocular Cellscope, and b) analysis software for automated DR screening

Methods : The Ocular CellScope is a retinal imaging device (Figure 1) that easily attaches to an iPhone without needing any modification to the phone hardware. Multiple color fundus images are captured per patient eye using the Ocular CellScope and automatically analyzed to produce a “refer”/”no refer” DR screening recommendation for the patient. A patient is deemed non-referable only if there is mild or no signs of DR and no ME in both eyes, otherwise the patient is deemed to have referable DR.
The image analysis system utilizes several novel multi-scale morphological filter-based image analysis techniques customized for DR screening combined with advanced machine learning techniques for multi-level classification. DR lesions including microaneurysms, hemorrhages, exudates, cotton wool spots, and neo-vascularization are detected and a Refer/No Refer DR screening recommendation is generated.
We evaluate our automated DR screening software on a dataset of 2788 images obtained from 80 patients using the Ocular Cellscope. Each patient case had 3-12 images captured per field and up to 5 fields per eye. Clinical findings from a slit lamp exam by an ophthalmologist are used to generate the reference standard for referable DR.

Results : Our DR screening software achieves a sensitivity of 90.0% (95% CI: 82.0% - 96.8%) at specificity of 45.0% (95% CI: 21.1% – 66.7%) at identifying patients with referable DR. Figure 2 shows the receiver operating characteristic (ROC) curve with the AUROC (area under ROC) being 0.798 (95% CI: 0.685 - 0.895).

Conclusions : Our automated DR screening software achieves good results on retinal images captured using the Ocular Cellscope, proving the feasibility of DR screening using cellphone-based retinal cameras in the real-world.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.

 

Fig1: (A, B) Ocular Cellscope, iphone-based fundus camera: (B, C) Captured retinal image samples.

 

Fig2: Receiver operating characteristic (ROC) curve for automated detection of referable DR on retinal images obtained using the Ocular cellscope.