Zhang Ye, Zhang Xiaoyue, Zhang Qing, Lv Bin, Hu Man, Lv Chuanfeng, Ni Yuan, Xie Guotong, Li Shuning, Zebardast Nazlee, Shweikh Yusrah, Wang Ningli
Beijing Tongren Eye Center, Beijing Key Laboratory of Ophthalmology & Visual Science, Beijing Tongren Hospital, Capital Medical University, Beijing, China.
Ping an Healthcare Technology, Beijing, China.
Heliyon. 2024 Jul 26;10(15):e35236. doi: 10.1016/j.heliyon.2024.e35236. eCollection 2024 Aug 15.
To develop and validate deep learning algorithms that can identify and classify angle-closure (AC) mechanisms using anterior segment optical coherence tomography (AS-OCT) images.
This cross-sectional study included participants of the Handan Eye Study aged ≥35 years with AC detected via gonioscopy or on the AS-OCT images. These images were classified by human experts into the following to indicate the predominant AC mechanism (ground truth): pupillary block, plateau iris configuration, or thick peripheral iris roll. A deep learning architecture, known as comprehensive mechanism decision net (CMD-Net), was developed to simulate the identification of image-level AC mechanisms by human experts. Cross-validation was performed to optimize and evaluate the model. Human-machine comparisons were conducted using a held-out and separate test sets to establish generalizability.
In total, 11,035 AS-OCT images of 1455 participants (2833 eyes) were included. Among these, 8828 and 2.207 images were included in the cross-validation and held-out test sets, respectively. A separate test was formed comprising 228 images of 35 consecutive patients with AC detected via gonioscopy at our eye center. In the classification of AC mechanisms, CMD-Net achieved a mean area under the receiver operating characteristic curve (AUC) of 0.980, 0.977, and 0.988 in the cross-validation, held-out, and separate test sets, respectively. The best-performing ophthalmologist achieved an AUC of 0.903 and 0.891 in the held-out and separate test sets, respectively. And CMD-Net outperformed glaucoma specialists, achieving an accuracy of 89.9 % and 93.0 % compared to 87.0 % and 86.8 % for the best-performing ophthalmologist in the held-out and separate test sets, respectively.
Our study suggests that CMD-Net has the potential to classify AC mechanisms using AS-OCT images, though further validation is needed.
开发并验证能够使用眼前节光学相干断层扫描（AS-OCT）图像识别和分类闭角（AC）机制的深度学习算法。
这项横断面研究纳入了邯郸眼病研究中年龄≥35岁且通过前房角镜检查或AS-OCT图像检测出AC的参与者。这些图像由人类专家分类为以下几种，以表明主要的AC机制（真实情况）：瞳孔阻滞、高原虹膜构型或周边虹膜肥厚卷缩。开发了一种名为综合机制决策网络（CMD-Net）的深度学习架构，以模拟人类专家对图像级AC机制的识别。进行交叉验证以优化和评估模型。使用保留的独立测试集进行人机比较以确定可推广性。
总共纳入了1455名参与者（2833只眼）的11035张AS-OCT图像。其中，交叉验证集和保留测试集分别纳入了8828张和2207张图像。形成了一个单独的测试集，其中包含在我们眼科中心通过前房角镜检查检测出AC的35例连续患者的228张图像。在AC机制分类中，CMD-Net在交叉验证集、保留测试集和单独测试集中的受试者操作特征曲线（AUC）下的平均面积分别为0.980、0.977和0.988。表现最佳的眼科医生在保留测试集和单独测试集中的AUC分别为0.903和0.891。并且CMD-Net的表现优于青光眼专家，在保留测试集和单独测试集中的准确率分别为89.9%和93.0%，而表现最佳的眼科医生的准确率分别为87.0%和86.8%。
我们的研究表明，CMD-Net有潜力使用AS-OCT图像对AC机制进行分类，不过仍需要进一步验证。
