Abstract
PURPOSE
To develop and evaluate deep learning models to perform automated diagnosis and quantitative classification of age-related cataract, including all three anatomical types, from anterior segment photographs.
DESIGN
Application of deep learning models to Age-Related Eye Disease Study (AREDS) dataset.
PARTICIPANTS
18,999 photographs (6,333 triplets) from longitudinal follow-up of 1,137 eyes (576 AREDS participants).
METHODS
Deep learning models were trained to detect and quantify nuclear cataract (NS; scale 0.9-7.1) from 45-degree slit-lamp photographs and cortical (CLO; scale 0-100%) and posterior subcapsular (PSC; scale 0-100%) cataract from retroillumination photographs. Model performance was compared with that of 14 ophthalmologists and 24 medical students. The ground truth labels were from reading center grading.
MAIN OUTCOME MEASURES
Mean squared error (MSE).
RESULTS
On the full test set, mean MSE values for the deep learning models were: 0.23 (SD 0.01) for NS, 13.1 (SD 1.6) for CLO, and 16.6 (SD 2.4) for PSC. On a subset of the test set (substantially enriched for positive cases of CLO and PSC), for NS, mean MSE for the models was 0.23 (SD 0.02), compared to 0.98 (SD 0.23; p=0.000001) for the ophthalmologists, and 1.24 (SD 0.33; p=0.000005) for the medical students. For CLO, mean MSE values were 53.5 (SD 14.8), compared to 134.9 (SD 89.9; p=0.003) and 422.0 (SD 944.4; p=0.0007), respectively. For PSC, mean MSE values were 171.9 (SD 38.9), compared to 176.8 (SD 98.0; p=0.67) and 395.2 (SD 632.5; p=0.18), respectively. In external validation on the Singapore Malay Eye Study (sampled to reflect the distribution of cataract severity in AREDS), MSE was 1.27 for NS and 25.5 for PSC.
CONCLUSIONS
A deep learning framework was able to perform automated and quantitative classification of cataract severity for all three types of age-related cataract. For the two most common types (NS and CLO), the accuracy was significantly superior to that of ophthalmologists; for the least common type (PSC), the accuracy was similar. The framework may have wide potential applications in both clinical and research domains. In the future, such approaches may increase the accessibility of cataract assessment globally. The code and models are publicly available at https://XXX.
Collapse