Treder Maximilian, Lauermann Jost Lennart, Alnawaiseh Maged, Eter Nicole
Department of Ophthalmology, University of Muenster Medical Center, Muenster, Germany.
Cornea. 2019 Feb;38(2):157-161. doi: 10.1097/ICO.0000000000001776.
To evaluate a deep learning-based method to automatically detect graft detachment (GD) after Descemet membrane endothelial keratoplasty (DMEK) in anterior segment optical coherence tomography (AS-OCT).
In this study, a total of 1172 AS-OCT images (609: attached graft; 563: detached graft) were used to train and test a deep convolutional neural network to automatically detect GD after DMEK surgery in AS-OCT images. GD was defined as a not completely attached graft. After training with 1072 of these images (559: attached graft; 513: detached graft), the created classifier was tested with the remaining 100 AS-OCT scans (50: attached graft; 50 detached: graft). Hereby, a probability score for GD (GD score) was determined for each of the tested OCT images.
The mean GD score was 0.88 ± 0.2 in the GD group and 0.08 ± 0.13 in the group with an attached graft. The differences between both groups were highly significant (P < 0.001). The sensitivity of the classifier was 98%, the specificity 94%, and the accuracy 96%. The coefficient of variation was 3.28 ± 6.90% for the GD group and 2.82 ± 3.81% for the graft attachment group.
With the presented deep learning-based classifier, reliable automated detection of GD after DMEK is possible. Further work is needed to incorporate information about the size and position of GD and to develop a standardized approach regarding when rebubbling may be needed.
评估一种基于深度学习的方法，用于在前节光学相干断层扫描（AS-OCT）中自动检测Descemet膜内皮角膜移植术（DMEK）后移植片脱离（GD）情况。
在本研究中，共使用1172幅AS-OCT图像（609幅：移植片附着；563幅：移植片脱离）来训练和测试深度卷积神经网络，以自动检测AS-OCT图像中DMEK手术后的GD情况。GD定义为移植片未完全附着。使用其中1072幅图像（559幅：移植片附着；513幅：移植片脱离）进行训练后，用其余100幅AS-OCT扫描图像（50幅：移植片附着；50幅：移植片脱离）对创建的分类器进行测试。据此，为每幅测试的OCT图像确定一个GD概率评分（GD评分）。
GD组的平均GD评分为0.88±0.2，移植片附着组为0.08±0.13。两组之间的差异具有高度显著性（P<0.001）。分类器的灵敏度为98%，特异性为94%，准确率为96%。GD组的变异系数为3.28±6.90%，移植片附着组为2.82±3.81%。
使用所提出的基于深度学习的分类器，可以可靠地自动检测DMEK后的GD情况。需要进一步开展工作，纳入有关GD大小和位置的信息，并制定关于何时可能需要再次注入气泡的标准化方法。
