文献速递角膜与眼表疾病

基于深度学习的睑板腺成像测量与评估方法

[An advanced imaging method for measuring and assessing meibomian glands based on deep learning].
PubMed 文献 · Zhonghua Yan Ke Za Zhi · 2020-01-01

模型标记的睑板腺比例与人工标记差异无统计学意义(t=1.935,P>0.05),且每幅图像评估仅需0.499秒。结论认为该模型可提高检查准确性并节省时间,可用于睑板腺功能障碍相关疾病的临床辅助诊断和筛查。

要点

  • 深度学习模型标记睑板腺比例53.24%±11.09%,与人工标记52.13%±13.38%无统计学差异。
  • 模型对睑结膜标记mAP>0.976、验证集损失值<0.35,对睑板腺标记mAP>0.922、验证集损失值<1.0。
  • 模型评估每幅图像仅需0.499秒,临床医师平均用时超过10秒。

核心结论

  • 基于深度学习的睑板腺图像处理模型可提高检查准确性并节省时间,可用于睑板腺功能障碍相关疾病的临床辅助诊断和筛查。

研究设计与样本

  • 图像由2名临床医师标记,利用深度学习算法建立模型,采用平均精度均值(mAP)及验证集损失值评价模型对特征区域识别的准确性。
  • 随机选取数据库以外64幅睑板腺图像,由7名医师独立评估后与模型评估结果进行配对t检验。

主要发现

  • 模型标记的睑板腺比例为53.24%±11.09%,人工标记为52.13%±13.38%,差异无统计学意义(t=1.935,P>0.05)。
  • 模型对睑结膜标记的mAP>0.976、验证集损失值<0.35,对睑板腺标记的mAP>0.922、验证集损失值<1.0。
  • 模型评价每幅图像仅需0.499秒,而临床医师用时平均超过10秒。

关键数值与推荐

  • 模型对睑结膜标记的mAP>0.976、验证集损失值<0.35。
  • 模型对睑板腺标记的mAP>0.922、验证集损失值<1.0。
  • 模型与人工标记睑板腺比例差异无统计学意义(t=1.935,P>0.05)。

临床怎么用

  • 可将该深度学习模型用于睑板腺功能障碍相关疾病的临床辅助诊断和筛查,以提高检查效率。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

Objective: To evaluate the application value of a deep-learning-based imaging method for rapid measurement and evaluation of meibomian glands. Methods: Diagnostic evaluation study. From January 2017 to December 2018, 2 304 meibomian gland images of 576 dry eye patients who were treated at the Eye Center of Wuhan University People's Hospital with an average age of (40.03±11.46) years were collected to build a meibomian gland image database. These images were labeled by 2 clinicians, and a deep learning algorithm was used to build a model and detect the accuracy of the model in identifying and labeling the meibomian glands and calculating the rate of meibomian gland loss. Mean average precision (mAP) and validation loss were used to assess the accuracy of the model in identifying feature areas. Sixty-four meibomian gland images apart from the database were randomly selected and evaluated by 7 clinicians independently. The results were analyzed with paired t-test. Results: This model marked the meibomian conjunctiva (mAP>0.976, validation loss<0.35) and the meibomian gland (mAP>0.922, validation loss<1.0), respectively, thereby achieving high accuracy to calculate the area and ratio of meibomian gland loss. The proportion of meibomian glands marked by the model was 53.24%±11.09%, and the artificial marking was 52.13%±13.38%. There was no statistically significant difference (t=1.935, P>0.05). In addition, the model took only 0.499 second to evaluate each image, while the average time for clinicians was more than 10 seconds. Conclusion: The deep-learning-based imaging model can improve the accuracy of the examination and save time and be used for clinical auxiliary diagnosis and screening of diseases related to meibomian gland dysfunction.(Chin J Ophthalmol, 2020, 56: 774-779). 目的: 探讨睑板腺图像深度处理分析方法的临床应用价值。 方法: 诊断评价研究。采集2017年1月至2018年12月就诊于武汉大学人民医院眼科中心年龄(40.03±11.46)岁的干眼患者的2 304幅睑板腺图像构建睑板腺图像数据库,由2名临床医师对图像进行标记,利用深度学习算法建立模型,检测模型对睑板腺识别及标注的准确性并计算睑板腺缺失率。采用平均精度均值(mAP)及验证集损失值评价模型对特征区域识别的准确性。并随机选取64幅数据库以外的睑板腺图像,由7名受试医师独立评估后与模型评估结果进行统计性t检验。 结果: 模型对睑结膜进行标记的mAP>0.976,验证集损失值<0.35;对睑板腺标记的mAP>0.922,验证集损失值<1.0。模型标记的睑板腺比例为53.24%±11.09%,人工标记为52.13%±13.38%,差异无统计学意义(t=1.935,P>0.05)。模型评价每幅图像仅需0.499 s,而临床医师用时平均超过10 s。 结论: 该睑板腺图像深度处理方法可提高临床检查结果的准确性,提高诊断效率,可用于睑板腺功能障碍相关疾病的临床辅助诊断和筛查。(中华眼科杂志,2020,56:774-779).

来源:Zhonghua Yan Ke Za Zhi,查看原文。版权归原作者所有。

引用

  1. Zhou YW, Yu Y, Zhou YB, et al. [An advanced imaging method for measuring and assessing meibomian glands based on deep learning]. Zhonghua Yan Ke Za Zhi. 2020. PMID: 33059421. doi:10.3760/cma.j.cn112142-20200415-00272

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。