Multiview Volume and Temporal Difference Network for Angle-Closure Glaucoma Screening from AS-OCT Videos

Abstract

Background. Precise and comprehensive characterizations from anterior segment optical coherence tomography (AS-OCT) are of great importance in facilitating the diagnosis of angle-closure glaucoma. Existing automated analysis methods focus on analyzing structural properties identified from the single AS-OCT image, which is limited to comprehensively representing the status of the anterior chamber angle (ACA). Dynamic iris changes are evidenced as a risk factor in primary angle-closure glaucoma. Method. In this work, we focus on detecting the ACA status from AS-OCT videos, which are captured in a dark-bright-dark changing environment. We first propose a multiview volume and temporal difference network (MT-net). Our method integrates the spatial structural information from multiple views of AS-OCT videos and utilizes temporal dynamics of iris regions simultaneously based on image difference. Moreover, to reduce the video jitter caused by eye movement, we employ preprocessing to align the corneal part between video frames. The regions of interest (ROIs) in appearance and dynamics are also automatically detected to intensify the related informative features. Results. In this work, we employ two AS-OCT video datasets captured by two different devices to evaluate the performance, which includes a total of 342 AS-OCT videos. For the Casia dataset, the classification accuracy for our MT-net is 0.866 with a sensitivity of 0.857 and a specificity of 0.875, which achieves superior performance compared with the results of the algorithms based on AS-OCT images with an obvious gap. For the Zeiss AS-OCT video dataset, our method also gets better performance against the methods based on AS-OCT images with a classification accuracy of 0.833, a sensitivity of 0.860, and a specificity of 0.800. Conclusions. The AS-OCT videos captured under changing environments can be a comprehended means for angle-closure classification. The effectiveness of our proposed MT-net is proved by two datasets from different manufacturers