方法：选取2021-01/11在温州医科大学附属眼视光医院杭州院区就诊的145名受试者右眼纳入研究。随机选择其中60名受试者的睑板腺照相用于AI训练。收集睑板腺图像后首先标注出睑板区域和每一根睑板腺腺体。使用残差神经网络(ResNet)结合U-Net模型进行数据训练，获得成熟的AI系统； 85名受试者包括阻塞性MGD患者53名和睑板腺正常的志愿者32名，使用AI系统自动分析其各项睑板腺形态参数。同时观察临床指标包括眼表疾病指数(OSDI)、泪河高度(TMH)、泪膜破裂时间(TBUT)、角膜荧光素染色(CFS)、睑缘评分、睑板腺评分和睑板腺分泌能力评分。分析睑板腺参数与临床指标的相关性。

结果：通过多次版本迭代，最终获得了交并比达92.0%的AI系统。使用该AI系统，发现上眼睑的睑板腺密度与OSDI(r_s=-0.320)、TBUT(r_s=0.484)、睑缘评分(r_s=-0.350)、睑板腺评分(r_s=-0.749)和睑板腺分泌能力评分(r_s=0.425)存在显著相关性(均P<0.05)； 下眼睑的睑板腺密度与OSDI(r_s=-0.420)、TBUT(r_s=0.598)、睑缘评分(r_s=-0.396)、睑板腺评分(r_s=-0.720)和睑板腺分泌能力评分(r_s=0.438)存在显著相关性(均P<0.05)； 总眼睑的睑板腺密度与OSDI(r_s=-0.404)、TBUT(r_s=0.601)、睑缘评分(r_s=-0.416)、睑板腺评分(r_s=-0.805)和睑板腺分泌能力评分(r_s=0.480)存在显著相关性(均P<0.05)。

结论：基于CNN的AI系统是一个准确、高效的睑板腺形态学评价系统，能够方便地采用我们建立的睑板腺密度这一指标对MGD患者的睑板腺形态进行快速准确地评价。睑板腺密度这一指标比目前通用的睑板腺评分更精确，是评价睑板腺萎缩程度的全新定量指标。

METHODS:The right eyes of 145 subjects who were treated at the Hangzhou Branch of the Eye Hospital of Wenzhou Medical University from January to November 2021 were selected for inclusion in the study. Meibography images of 60 of these subjects were randomly selected for AI training. The meibomian region and each MG in meibography were annotated and formed into datasets. The datasets were used for training and obtaining an AI system based on residual neural network(ResNet)combined with the U-Net model. The AI system was used to automatically analyze the MG morphological parameters in 85 subjects, including 53 patients with obstructive MGD and 32 volunteers with normal meibomian glands. The clinical indices including ocular surface disease index(OSDI), tear meniscus height, tear film break-up time(TBUT), corneal fluorescein staining, lid margin score, meiboscore, and meibomian gland expressibility score were also observed. The correlation between MG morphological parameters and clinical indices were analyzed.

RESULTS: After several iterations, we finally obtained an AI system with Intersection over Union of 92.0%. Using this AI system, we found that there was a significant correlation between the MG density in the upper eyelid with OSDI(r_s=-0.320), TBUT(r_s=0.484), lid margin score(r_s=-0.350), meiboscore(r_s=-0.749), and meibum expressibility score(r_s=0.425)(all P<0.05). The MG density in the lower eyelid was significantly correlated with OSDI(r_s=-0.420), TBUT(r_s= 0.598), lid margin score(r_s=-0.396), meiboscore(r_s=-0.720), and meibum expressibility score(r_s=0.438)(all P<0.05). The MG density in the total eyelid was significantly correlated with OSDI(r_s=-0.404), TBUT(r_s=0.601), lid margin score(r_s=-0.416), meiboscore(r_s=-0.805), and meibum expressibility score(r_s=0.480)(all P<0.05).

CONCLUSION:The AI system based on CNN in this study is an accurate and efficient MG morphological evaluation system, which can be conveniently used to evaluate the MG morphology of MGD patients quickly and accurately by using the MG density index established by us. MG density is a new quantitative index to evaluate meibomian gland atrophy, which is more accurate than meiboscore.