Ding Yiheng, Wei Ziqiang, Wang Chaoyun, Li Xinyue, Li Bingbing, Liu Xueting, Fu Zhijie, Mo Hongwei, Zhang Hong
Eye Hospital, The First Affiliated Hospital of Harbin Medical University, Harbin, China.
School of Intelligent Science and Engineering, Harbin Engineering University, Harbin, China.
Front Cell Dev Biol. 2025 Jul 9;13:1581785. doi: 10.3389/fcell.2025.1581785. eCollection 2025.
As a disease with high global incidence, hypertension is known to cause systemic vasculopathy. Ophthalmic vessels are the only vascular structures that can be directly observed in a non-invasive manner. We aim to investigate the changes in ocular microvessels in hypertension using deep learning on optical coherence tomography angiography (OCTA) images.
The convolutional neural network architecture Xception and multi-Swin transformer were used to screen 422 OCTA images (252 from 136 hypertension subjects; 170 from 85 healthy subjects) for hypertension. Moreover, the separability of the OCTA images based on high-dimensional feature angles was analyzed to better understand how deep learning models distinguish such images with class activation mapping.
Under Xception, the overall average accuracy of 5-fold cross-validation was 76.05% and sensitivity was 85.52%. In contrast, the Swin transformer showed single-model (macular), single-model (optic disk), and multimodel average accuracies of 82.25%, 74.936%, and 85.06%, respectively, for predicting hypertension.
The changes caused by hypertension on the fundus vessels can be observed more accurately and efficiently using OCTA image features through deep learning. These results are expected to assist with screening of hypertension and reducing the risk of its severe complications.
ChiCTR, ChiCTR2000041330. Registered 23 December 2020, https://www.chictr.org.cn/ChiCTR2000041330.
高血压作为一种全球发病率较高的疾病，已知会导致全身血管病变。眼部血管是唯一能够以非侵入性方式直接观察到的血管结构。我们旨在利用深度学习分析光学相干断层扫描血管造影（OCTA）图像，研究高血压患者眼部微血管的变化。
使用卷积神经网络架构Xception和多Swin变压器对422张OCTA图像（136名高血压患者的252张图像；85名健康受试者的170张图像）进行高血压筛查。此外，基于高维特征角度分析OCTA图像的可分离性，以更好地理解深度学习模型如何通过类激活映射区分此类图像。
在Xception架构下，5折交叉验证的总体平均准确率为76.05%，灵敏度为85.52%。相比之下，Swin变压器在预测高血压时，单模型（黄斑）、单模型（视盘）和多模型的平均准确率分别为82.25%、74.936%和85.06%。
通过深度学习利用OCTA图像特征可以更准确、高效地观察高血压对眼底血管造成的变化。这些结果有望有助于高血压的筛查，并降低其严重并发症的风险。
中国临床试验注册中心，ChiCTR2000041330。于2020年12月23日注册，https://www.chictr.org.cn/ChiCTR2000041330。
