Using deep learning to screening for hypertension in OCTA image to reduce the risk of serious complications

Yiheng Ding¹Ziqiang Wei²Chaoyun Wnag²Xinyue Li¹Bingbing Li²Xueting Liu¹Zhijie Fu²Hong Zhang^1*Hongwei Mo^2*

• ¹First Affiliated Hospital of Harbin Medical University, Harbin, China
• ²Harbin Engineering University, Harbin, Heilongjiang Province, China

Background: As a global high incidence disease, hypertension causes systemic vasculopathy. Ophthalmic vessels are the only vascular structures that can be directly observed in vivo in a non-invasive manner. We aim to investigate the changes in ocular microvessels in hypertension by optical coherence tomography angiography (OCTA) images with deep learning method.: Xception and Multi-Swin Transformer were used to screen hypertension with 422 OCTA images(252 from 136 hypertension subjects, 170 from 85 healthy subjects). Besides, the separability of OCTA images from high-dimensional feature angles was analyzed to better understand how deep learning models distinguish OCTA images with Class Activation Mapping. Results: Under CNN Xception, the overall average accuracy of 5-fold cross-validation is 76.05%, the sensitivity is 85.52%. As a contrast by Swin Transformer, the average accuracy of Single-model (macular), Single-model (optic disk) and Multi-model in prediction of hypertension are 82.25%, 74.936% and 85.06%, respectively. Conclusions: The changes of hypertension on fundus vessels can be more accurately and efficiently observed by recognizing OCTA image features through deep learning. The results would contribute to screening hypertension and reducing the risk of severe complications of hypertension.