Artificial intelligence-assisted tear meniscus height measurement: a multicenter study.

BACKGROUND

The tear meniscus height (TMH) is an important index for the diagnosis of dry eye. However, special inspection doctors are required to make rapid TMH measurements during outpatient examinations, which often leads to substantial measurement errors. At the same time, the existing artificial intelligence (AI) model of TMH segmentation has poor generalization because it only uses one mode of TMH pictures and does not include inspection of external verification sets. The purpose of this study was to propose an automatic measurement method for TMH based on convolutional neural networks (CNNs) to handle diverse datasets.

METHODS

This multicenter retrospective study included 3,894 TMH images from five centers across four regions in eastern, southern, and western China. The images were annotated using a gradient information-guided human-computer collaborative method, and an attention-limiting neural network (ALNN) was developed. An internal dataset, consisting of 834 color images and 1,105 infrared images from three centers, was constructed for model development and validation. An external validation set, comprising 996 color images and 959 infrared images from two additional centers, was used to test the model's generalizability. The accuracy of AI segmentation results was compared with the inspection reports of special inspection doctors.

RESULTS

In the test set for the color image modality, the segmentation results showed an average mean intersection over union (MIoU) of 0.9578, a recall rate of 0.9648, a precision of 0.9526, and an F1 score of 0.9576. The TMH results obtained on the test set (r=0.935, P<0.001) and on the external validation set (r=0.957, P<0.001) both showed a high correlation with the ground truth (GT). For the infrared image modality, the test set segmentation results showed an average MIoU of 0.9290, a recall rate of 0.9150, a precision of 0.9388, and an F1 score of 0.9249. The TMH results obtained on the test set (r=0.855, P<0.001) and on the external validation set (r=0.803, P<0.001) both showed a high correlation with the GT.

CONCLUSIONS

This algorithm exhibits strong generalization capabilities, accurately segments key areas, and automatically provides quantitative analysis of the TMH. The measurements obtained using this AI algorithm exhibit high consistency with the GT, surpassing the reliability of special inspection doctors. This provides significant support in the diagnosis of dry eye disease (DED).