Background Glaucoma is an irreversible blinding eye disease, and research on the characteristics of glaucoma fundus images and establishment of a diagnostic model using artificial intelligent technology based on professional clinical databases can help to achieve a rapid and objective judgment of the fundus images of glaucoma patients.
Objective To evaluate the performance of a glaucoma lesion diagnosis system based on artificial intelligent fundus analysis technology, and use the system to explore the development of the cup-to-disk ratio of glaucoma in disease progression.
Methods Fundus images were collected from 4 000 patients who completed fundus photography in the Third Medical Center of Chinese PLA General Hospital from March 2020 to April 2021, and then the images were used for segmenting the optic cup, the optic disc, and the retinal nerve fiber layer defect and labeling the multi-classification of lesions and the stage of glaucoma to obtain the fundus image-glaucoma dataset. An artificial intelligence glaucoma lesion diagnosis system was established based on deep learning model, and fundus images were randomly selected for internal validation to evaluate the performance of this system. The artificial intelligence system was used to identify and calculate the cup-to-disc ratio of glaucoma fundus images in different stages, in order to analyze the difference in the distribution of cup-to-disc ratio with the progression of glaucoma.
Results A total of 6 837 fundus images were included in this study, among which 60% (4 102) were used as training set and 40% (2 735) were used as validation set. In the validation set, the artificial intelligence glaucoma lesion diagnosis system had a mean recall rate of 0.837, a mean accuracy of 0.814, a mean intersection over union of 0.816, and a mean AUC of 0.874 in predicting the segmentation of the optic disc and the optic cup; this system had a mean recall rate of 0.928, a mean accuracy of 0.926, a mean intersection over union of 0.916, and a mean AUC of 0.941 in predicting optic disc segmentation; in predicting the segmentation of retinal nerve fiber layer defect, this system had a mean recall rate of 0.653, a mean accuracy of 0.612, a mean intersection over union of 0.480, and a mean AUC of 0.749. In the validation set, the artificial intelligence glaucoma lesion diagnosis system had a mean accuracy of 0.890, a mean sensitivity of 0.896, a mean specificity of 0.638, and a mean AUC of 0.893 in predicting focal RNFLD lesion; in predicting diffuse RNFLD lesion, the system had a mean accuracy of 0.950, a mean sensitivity of 0.744, a mean specificity of 0.961, and a mean AUC of 0.901; in predicting optic disc hemorrhage, the system had a mean accuracy of 0.966, a mean sensitivity of 0.650, a mean specificity of 0.967, and a mean AUC of 0.969; in predicting optic cup fissure, the system had a mean accuracy of 0.951, a mean sensitivity of 0.794, a mean specificity of 0.957, and a mean AUC of 0.892. The cup-to-disc ratio was identified and calculated for the fundus images of glaucoma in different stages, and the results showed that the cup-to-disc ratio gradually increased with the progression of glaucoma optic neuropathy.
Conclusion The application of artificial intelligent fundus analysis technology in the glaucoma lesion diagnosis system can provide ideas for realizing glaucoma screening.
DownLoad: