Abstract
Purpose :
E-health is entering the medical retina field and referrals based on automated grading of retinal images made by local optometrists is likely to become routine practice in the future. This study aimed to evaluate a robust algorithm for referable age-related macular degeneration (AMD) diagnosed on color fundus photographs.
Methods :
Referable AMD was considered present on color fundus photographs when the patient was eligible for treatment, e.g. AREDS supplements or anti-VEGF, corresponding to AREDS severity scale of 2 and higher. The evaluated algorithm (RetCAD v1.3.0) used deep-learning based on convolutional neural networks and was trained by ±100.000 images with various AMD lesions graded by at least one experienced grader on image level. RetCAD produced a score between 0 and 100, indicating the likelihood of referable AMD. The algorithm was validated in a for AMD enriched set of persons with follow up from the Rotterdam Study (n=1772). The validation set was graded by two experienced graders on lesion level. The performance of RetCAD was compared to that of the human graders by unpaired t-tests, ANOVA, and area under the receiver operating characteristic curve (AUC).
Results :
The AUC for correspondence between RetCAD and human graders was 0.89 for referable AMD.
RetCAD scores increased significantly per AREDS severity class (P trend <0.0001). RetCAD scores were 37.0±1.3 and 59.0±2.4 for AREDS classes 2 and 3, respectively, and 63.8±4.6 for AREDS class 4. Scores did not differ between advanced AMD subtype (GA, CNV, mixed AMD). Of all AMD lesions, RetCAD showed the most prominent trend with drusen type.
Conclusions :
This deep learning algorithm can help discriminate patients with clinically significant AMD from those without. Our study adds to the growing evidence that automated grading systems of retinal images can be used reliably in first line settings, and may accelerate the digital solution for an increasing public health problem.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.