Purchase this article with an account.
Natalia Porporato, Baskaran Mani, Xu Yanwu, Tin A Tun, Sameer Trikha, Damon W K Wong, Tin Aung; Automated grading of anterior segment Swept Source OCT images: A validation study for assessment of angle closure. Invest. Ophthalmol. Vis. Sci. 2017;58(8):2086.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the performance of an automated machine-learning algorithm for detecting angle closure in a single frame anterior segment swept source optical coherence tomography (SSOCT, CASIA, Tomey Corp, Japan) image, compared to gonioscopic angle closure in a community based sample.
1900 subjects from a community screening study underwent gonioscopy and CASIA 360 degree SSOCT scans. Of these, horizontal scans from 578 subjects (957 eyes) underwent a machine-learning algorithm for detecting angle closure (based on histogram of oriented gradients and linear support vector machine techniques) and manual grading of images by a clinician (masked to gonioscopy). Manual grading and machine-learning assessment were compared to gonioscopic angle closure definitions using receiver operating characteristic curves.
Of the 957 eyes, 2 quadrant gonioscopic angle closure was seen in 32 (3.3%) eyes and 3 quadrant closure was seen in 26 (2.7%) eyes. The AUC (area under the curve) for manual grading of images was 0.78 (95% CI 0.75, 0.8) and 0.83 (95% CI 0.8, 0.85) for 2 and 3 quadrants goniosopic closure. The AUC for the automated machine learning algorithm was 0.71 (95% CI 0.69, 0.75) and 0.75 (95% CI 0.72, 0.78) respectively. The AUC for the automated machine learning algorithm was 0.83 (95% CI 0.8, 0.85) for detecting a closed-angle in images graded by the clinician. With a fixed specificity of 0.9, false positive rate was 0.1 and sensitivity was 0.74.
The algorithm for automated assessment of angle closure in SSOCT images showed moderate performance when compared to gonioscopy. We suggest further enhancement of the algorithm using feature extraction methods and deep learning in future experiments.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only