Abstract
Purpose :
Diagnosing cataract usually requires in-person evaluation by an ophthalmologist. However, many individuals undergo color fundus photography (CFP) outside ophthalmology clinics, which represents an opportunity towards more accessible diagnosis by developing automated cataract detection methods. Therefore, the purpose of this study was to develop an explainable deep learning network that can detect cataract from color fundus photographs (CFPs) obtained from the Age-Related Eye Diseases Study 2 (AREDS2).
Methods :
A dataset of 17, 514 CFPs was obtained from 2,573 AREDS2 participants. The dataset comprised 8,681 CFPs from eyes with cataract and 8,833 CFPs from eyes without cataract. The ground truth labels were transferred from slit lamp exams of nuclear cataracts conducted by ophthalmologists and from reading center gradings of the anterior segment photos (red reflex photos) for evaluating cortical and posterior subcapsular cataracts. The dataset was divided into training, validation, and testing datasets (70%, 20%, and 10% participants, respectively). A deep learning network was developed using separable convolutions and residual connections. The network performance was compared to that of three ophthalmologists. The network was visualized using Gradient Class Activation Maps (Grad-CAMs) where areas of high signal correspond to the CFP pixels that contributed most to the network decision.
Results :
The network achieved performance scores (with 95% CI) of 0.6683 (0.6531, 0.6838), 0.6691 (0.6536, 0.685), 0.6686 (0.6533, 0.6842), 0.6686 (0.6533, 0.6842) compared to that of the combined ophthalmologists of 0.6025 (0.5100, 0.7000), 0.6119 (0.5085, 0.7123), 0.5988 (0.5074, 0.6895), and 0.5988 (0.5074, 0.6895) for accuracy, precision, recall, and AUC, respectively. Qualitative analysis of the Grad-CAMs, of correctly graded CFPs, showed that areas of high signal in CFPs without cataract and areas of low signal in CFPs with cataract corresponded to retinal blood vessels, as shown in Fig. 1.
Conclusions :
The proposed network outperformed three ophthalmologists in the detection of cataract from CFP, which may increase the accessibility of cataract diagnosis. The detections were generally explainable where the retinal blood vessels seemed to be an important diagnostic feature.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.