Application of deep learning image assessment software VeriSee™ for diabetic retinopathy screening

Abstract

Purpose: To develop a deep learning image assessment software VeriSee™ and to validate its accuracy in grading the severity of diabetic retinopathy (DR).

Methods: Diabetic patients who underwent single-field, nonmydriatic, 45-degree color retinal fundus photography at National Taiwan University Hospital between July 2007 and June 2017 were retrospectively recruited. A total of 7524 judgeable color fundus images were collected and were graded for the severity of DR by ophthalmologists. Among these pictures, 5649 along with another 31,612 color fundus images from the EyePACS dataset were used for model training of VeriSee™. The other 1875 images were used for validation and were graded for the severity of DR by VeriSee™, ophthalmologists, and internal physicians. Area under the receiver operating characteristic curve (AUC) for VeriSee™, and the sensitivities and specificities for VeriSee™, ophthalmologists, and internal physicians in diagnosing DR were calculated.

Results: The AUCs for VeriSee™ in diagnosing any DR, referable DR and proliferative diabetic retinopathy (PDR) were 0.955, 0.955 and 0.984, respectively. VeriSee™ had better sensitivities in diagnosing any DR and PDR (92.2% and 90.9%, respectively) than internal physicians (64.3% and 20.6%, respectively) (P < 0.001 for both). VeriSee™ also had better sensitivities in diagnosing any DR and referable DR (92.2% and 89.2%, respectively) than ophthalmologists (86.9% and 71.1%, respectively) (P < 0.001 for both), while ophthalmologists had better specificities.

Conclusion: VeriSee™ had good sensitivity and specificity in grading the severity of DR from color fundus images. It may offer clinical assistance to non-ophthalmologists in DR screening with nonmydriatic retinal fundus photography.

Keywords: Artificial intelligence; Convolutional neural network; Deep learning; Diabetic retinopathy; Retinal fundus photography.