Validation results of a deep learning algorithm for detection of diabetic retinopathy with lesion localization from retinal fundus photographs

Abstract

Purpose : To perform clinical validation tests of a deep learning algorithm (DLA) which detects diabetic retinopathy (DR) from retinal fundus photographs with severity grading and localizing four types of lesions (hard exudates, hemorrhages, microaneurysms, cotton wool spots) related to DR.

Methods : A DLA was trained to detect DR in 120.000 fundus photography images to detect and grade cases according to International Clinical Diabetic Retinopathy Disease Severity Scale classification, in which localization of four types of retinal lesions (hard exudates, hemorrhages, microaneurysms and cotton wool spots) was also performed. Training of the DLA was completed between April and June 2018, and the resultant algorithm was validated in July 2019 using a 98-images set of patients which was graded and lesion-tagged by three experienced ophthalmologists.

Results : Out of 98 fundus images, 10 were found without DR, 28 with moderate DR, 27 with severe DR and 25 with proliferative DR (PDR) by DLS, where 8 images were considered with bad quality. The sensitivity, specificity, accuracy and area under the curve (AUC) of DLA for DR grading were 80.0% (95% CI: 74.6-85.3), 92.1% (95% CI: 90.6-93.6), 87.3% (95% CI: 84.2-90.5) and 0,860 (95% CI: 0,826-0,894) respectively. The AUC values of the algorithm were 0.834 (95% CI: 0.800-0.868) for hard exudates, 0.869 (95% CI: 0.866-0.872) for hemorrhages, 0.805 (95% CI: 0.787-0.823) for microaneurysms and 0.784 (95% CI: 0.751-0.817) for cotton wool spots.

Conclusions : DLA has high sensitivity and specificity to grade diabetic retinopathy with an acceptable lesion localization from fundus photographs. This algorithm can be used as a decision support system for earlier diagnosis of background DR and therefore can improve the clinical managment of DR.

This is a 2020 ARVO Annual Meeting abstract.

 

Severity grading interface of the algorithm.

 

Lesion localization of the algorithm with tagged lesions in color codes for hard exudates (blue), hemorrhages (green), microaneurysms (yellow) and cotton wool spots (cyan).