| Objective:Quantitative analysis of images and tear film break-up time related data obtained from non-invasive eye surface comprehensive analyzer using artificial intelligence learning algorithm,early and accurate diagnosis of meibomian gland dysfunction(MGD),and establishment of standardized artificial intelligence diagnostic system.Methods:This study included 1354 images of meibomian gland and tear film break-up time in 677 eyes of 677 subjects who visited the First Affiliated Hospital of Nanchang University.Three ophthalmologists from the First Affiliated Hospital of Nanchang University manually labeled the area of the tarsal plate and each gland.The labeled meibomian gland images are randomly divided into three groups for training,validation,and testing of the nn U-Net segmentation model.Using manually segmented images as the gold standard,the Dice coefficient and intersection to union ratio(Io U)are used to calculate the similarity between intelligent segmentation and manual segmentation,and to evaluate the performance of the nn U-Net model.In addition,the accuracy of the intelligent segmentation of the meibomian gland score was compared with the meibomian gland score estimated by ophthalmologists with the naked eye.Calculate the correlation between the degree of meibomian gland loss,age,and tear film break-up time,and construct a prediction model for tear film break-up time.The first tear film break-up time and average tear film break-up time prediction models use ridge regression method,while the tear film break-up time grading prediction uses multi classification models.The mean absolute error(MAE)and root mean square error(RMSE)are used as the evaluation indicators of the regression model,and the evaluation accuracy rate(ACC)of the multi classification model is used.Results:The average Dice coefficient of the nn U-Net model for identifying tarsus is93.45%,the average Io U is 88.08%,the average sensitivity is 94.31%,and the average specificity is 96.23%;The average Dice coefficient for identifying meibomian glands is 69.92%,the average Io U is 54.36%,the average sensitivity is73.97%,and the average specificity is 96.34%.The accuracy of the ophthalmologist’s naked eye evaluation of the meibomian gland score is 40.11%,with an average time of 0.92 seconds per image.The accuracy of the intelligent model is 70.37%,with an average time of 0.04 seconds per image.The degree and age of upper meibomian gland deficiency were negatively correlated with first tear film break-up time(P<0.05),and positively correlated with the grading of tear film break-up time(P<0.05).There was no significant correlation between the degree of lower meibomian gland deficiency and the parameters related to tear film break-up time(P>0.05).In the regression model,the performance of the prediction model with age factors is improved compared to the model without age.In predicting the first tear film rupture time,the MAE index of the model is 3.66,and the RMSE is 4.83.In predicting the average tear film rupture time,the MAE index of the model is 4.25,and the RMSE is 5.27.In the classification model,the accuracy of the model also improved after adding age factors.The ACC of the prediction model based on the degree of upper eyelid gland deficiency was 0.44,while the ACC of the prediction model based on the degree of lower eyelid gland deficiency and the comprehensive degree of upper and lower eyelid deficiency were both 0.43.Conclusion:Artificial intelligence algorithms can quickly and accurately quantify the degree of meibomian gland atrophy,providing the possibility of establishing a standardized and automated artificial intelligence diagnostic system.Using only the degree of meibomian gland loss and age is not enough to accurately predict the time of tear film rupture.In the future,we will add other relevant data to improve the performance of the MGD prediction model,further improve the early accurate diagnosis system of meibomian gland,and provide a basis for early prevention and treatment of MGD. |
PDF Full Text Request