Fundus photograph-based cataract evaluation network using deep learning

Abstract

Background: Our study aims to develop an artificial intelligence-based high-precision cataract classification and grading evaluation network using fundus images.Methods: We utilized 1,340 color fundus photographs from 875 participants (aged 50–91 years at image capture) from the Beijing Eye Study 2011. Four experienced and trained ophthalmologists performed the classification of these cases based on slit-lamp and retro-illuminated images. Cataracts were classified into three types based on the location of the lens opacity: cortical cataract, nuclear cataract, and posterior subcapsular cataract. We developed a Dual-Stream Cataract Evaluation Network (DCEN) that uses color photographs of cataract fundus to achieve simultaneous cataract type classification and severity grading. The accuracy of severity grading was enhanced by incorporating the results of type classification.Results: The DCEN method achieved an accuracy of 0.9762, a sensitivity of 0.9820, an F1 score of 0.9401, and a kappa coefficient of 0.8618 in the cataract classification task. By incorporating type features, the grading of cataract severity can be improved with an accuracy of 0.9703, a sensitivity of 0.9344, an F1 score of 0.9555, and a kappa coefficient of 0.9111. We utilized Grad-CAM visualization technology to analyze and summarize the fundus image features of different types of cataracts, and we verified our conclusions by examining the information entropy of the retinal vascular region.Conclusion: The proposed DCEN provides a reliable ability to comprehensively evaluate the condition of cataracts from fundus images. Applying deep learning to clinical cataract assessment has the advantages of simplicity, speed, and efficiency.