NLM IRP Seminar Schedule

Abstract:

Cataract is the leading cause of blindness worldwide in the elderly and accounts for half of global blindness. Its prevalence is predicted to increase due to the aging population in many countries. It forms as an opacity in the crystalline lens that develops slowly and causes visual impairment. In its severe stages, it requires surgical treatment, so early diagnosis is necessary. Its diagnosis requires usually in-person evaluation by an ophthalmologist, which can be difficult. However, color fundus photographs (CFP) are broadly taken outside ophthalmology clinics, which could be a great chance to increase cataract screening through an automated algorithm. We developed DeepOpacityNet to detect cataract and highlight its most relevant features in CFP. We used a balanced dataset of 17,514 CFPs from 2,573 participants obtained from the Age-Related Eye Diseases Study 2 dataset. The ground truth labels were transferred from slit lamp examination and reading center grading of anterior segment photographs for different cataract types. The dataset was split on the participant level into training, validation, and test sets (70%, 10%, and 20% participants, respectively). DeepOpacityNet and other methods were trained and evaluated on these sets. Moreover, 100 random test CFPs were used to compare DeepOpacityNet performance to that of three ophthalmologists and to visually grade the output class activation maps (CAMs). On the test set, DeepOpacityNet outperformed other methods with accuracy of 0.6683 and AUC of 0.6686. On the random test subset, DeepOpacityNet outperformed ophthalmologists with accuracy of 0.6610 and AUC of 0.6612 compared to 0.6025 and 0.5988. The visual grading of output CAMs by ophthalmologists show that DeepOpacityNet highlights more interpretable features compared to other methods. In conclusion, DeepOpacityNet could detect cataract from CFP with interpretable outputs with reasonable performance superior to that of ophthalmologists on such difficult dataset.