For a study, researchers sought to create deep learning models for automated diagnosis and quantitative categorization of age-related cataracts using pictures of the anterior section. A total of 18,999 images (6,333 triplets) from 1,137 eyes (576 AREDS participants) were included in the study. From 45-degree slit-lamp pictures, deep learning models were trained to identify and quantify nuclear sclerosis (NS; scale 0.9–7.1), cortical lens opacity (CLO; scale 0%–100%), and posterior subcapsular cataract (PSC; scale 0%–100%). The performance of DeepLensNet was compared against 14 ophthalmologists and 24 medical students.
DeepLensNet had a mean MSE of 0.23 (standard deviation [SD], 0.01) for NS, 13.1 (SD, 1.6) for CLO, and 16.6 (SD, 2.4) for PSC on the whole test set. DeepLensNet had a mean MSE of 0.23 (SD, 0.02) for NS, compared to 0.98 (SD, 0.24; P=0.000001) for ophthalmologists and 1.24 (SD, 0.34; P=0.000005) for medical students on a subset of the test set (significantly enriched for positive instances of CLO and PSC). The average MSE for CLO was 53.5 (SD 14.8), compared to 134.9 (SD 89.9) for ophthalmologists and 433.6 (SD 962.1; P=0.0007) for medical students. PSC had a mean MSE of 171.9 (SD 38.9), whereas ophthalmologists had a mean MSE of 176.8 (SD 98.0; P=0.67) and medical students had a mean MSE of 398.2 (SD 645.4; P=0.18). The MSE for DeepSeeNet was 1.27 for NS and 25.5 for PSC in external validation on the Singapore Malay Eye Study (sampled to represent the cataract severity distribution in AREDS).
For all three kinds of age-related cataracts, DeepLensNet conducted automatic and quantitative cataract severity categorization. The accuracy was much better than that of ophthalmologists for the two most frequent kinds (NS and CLO); it was equivalent for the least common type (PSC). DeepLensNet might offer a wide range of therapeutic and scientific uses. Such techniques may improve worldwide cataract evaluation accessibility in the future.
Reference:www.aaojournal.org/article/S0161-6420(21)00967-2/fulltext
