Development of a cataract screening model using an open dataset and deep machine learning algorithms

Abstract

Relevance. Untreated cataract is the cause of permanent blindness. The main factors of untimely surgical treatment are the lack of patient's awareness about the need for surgical treatment (36.1%) and work or household employment (25.3%). Thus, regular cataract screening is an effective way to prevent blindness and identify patients in need of surgery. Purpose. Development of a cataract screening system based on an open data set, as well as its validation on clinical data. Material and methods. An open dataset (No. 1) of 9668 smartphone camera images, of which 4514 were cataracts and 5154 were normal eyes. The set for external validation (No. 2) was obtained under clinical conditions in the diagnostic department of the Krasnodar branch of the The S. Fyodorov Eye Microsurgery Federal State Institution. The set contained 51 cataract and normal images. To create a machine learning model, we used a convolutional neural network (CNN). Results. The data classification accuracy value was 0.97 for the internal validation set and 0.75 for the external one. The predictive value was low for cataract at the change in data set №2 and was only 0.54, as well as for sensitivity (0.87) and specificity (0.69) metrics. The area under the ROC curve was 0.99 (for dataset No. 1) and 0.78 (for dataset No. 2). Conclusion. These results indicate that it is necessary to fine-tune the model and provide the necessary levels of performance metrics for this scenario. Keywords: cataract, artificial intelligence, machine learning, screening, open datasets