Pourjavan Sayeh, Gouverneur François, Macq Benoit, Van Drooghenbroeck Thomas, De Potter Patrick, Boschi Antonella, El Maftouhi Adil
Department of Ophthalmology, Cliniques Universitaires Saint Luc, UCL, Brussels, Belgium.
Institute for Information and Communication Technologies, Electronics, and Applied Mathematics (ICTEAM), Louvain School of Engineering, UCLouvain, Louvain-la-Neuve, Belgium.
Clin Ophthalmol. 2024 Nov 26;18:3493-3502. doi: 10.2147/OPTH.S472231. eCollection 2024.
To evaluate the discriminative power of optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) images, identifying the best image combination for differentiating glaucoma from healthy eyes using deep learning (DL) with a convolutional neural network (CNN).
This cross-sectional study included 157 subjects contributing 1,106 eye scans. We used en-face images of the superficial and choroid layers for OCTA-based vessel density and OCT-based structural thickness of the macula (M) and optic disc (D). Images were preprocessed, resized, and normalized for CNN analysis. The CNN architecture had two components: one extracted features for each image type (OCT-D, OCT-M, OCTA-D, OCTA-M), while the second combined these features to classify eyes as healthy or glaucomatous. Performance was measured by accuracy, sensitivity, specificity, and area under the curve (AUC).
For OCT images, the D+M combination outperformed disc (D) or macula (M) alone in three of the four metrics. For OCTA images, D+M also performed better than D or M alone, with D+M outperforming disc (D) in all criteria. Across all metrics for combined OCT+OCTA images, D+M performed better than D or M alone, and the macula (M) outperformed the disc (D). In disc (D) imaging, OCTA outperformed both OCT and OCT+OCTA in accuracy, sensitivity, and specificity, while OCT+OCTA had a higher AUC. OCTA consistently outperformed OCT and OCT+OCTA across all metrics for combined D+M images.
The OCTA D+M combination performed best, followed by the OCT+OCTA D+M combination. When both en-face images are available, OCTA is preferred. Always include both disc and macula images for optimal diagnosis.
评估光学相干断层扫描（OCT）和光学相干断层扫描血管造影（OCTA）图像的鉴别能力，使用卷积神经网络（CNN）的深度学习（DL）确定区分青光眼与健康眼睛的最佳图像组合。
这项横断面研究纳入了157名受试者，共进行了1106次眼部扫描。我们使用了基于OCTA的黄斑（M）和视盘（D）的表层和脉络膜层的正面图像，以获取血管密度以及基于OCT的结构厚度。对图像进行预处理、调整大小并归一化，以便进行CNN分析。CNN架构有两个部分：一个用于提取每种图像类型（OCT-D、OCT-M、OCTA-D、OCTA-M）的特征，另一个将这些特征组合起来，将眼睛分类为健康或青光眼性。通过准确性、敏感性、特异性和曲线下面积（AUC）来衡量性能。
对于OCT图像，在四个指标中的三个指标上，D+M组合的表现优于单独的视盘（D）或黄斑（M）。对于OCTA图像，D+M的表现也优于单独的D或M，在所有标准中D+M均优于视盘（D）。对于联合的OCT+OCTA图像的所有指标，D+M的表现优于单独的D或M，并且黄斑（M）的表现优于视盘（D）。在视盘（D）成像中，OCTA在准确性、敏感性和特异性方面优于OCT和OCT+OCTA，而OCT+OCTA的AUC更高。在联合的D+M图像的所有指标上，OCTA始终优于OCT和OCT+OCTA。
OCTA D+M组合表现最佳，其次是OCT+OCTA D+M组合。当两种正面图像都可用时，首选OCTA。为了获得最佳诊断，始终要同时包括视盘和黄斑图像。
