Purpose
To introduce ACAI, a system to automatically grade anterior-chamber angles for angle closure glaucoma detection with three dimensional anterior segment optical coherence tomography (3D AS-OCT, CASIA SS2000, Tomey, Japan) images.
Methods
An automatic system is proposed to detect angle closure via anterior chamber angle assessment with 3D AS-OCT image sequences.<br /> The system is based on image processing and reconstruction, including image realighment and a reconstruction-based angle grading module. ACAI differs from previous works that rely on anterior chamber angle assessment metrics used clinically, such as AOD (angle-opening distance) [Leung, et al, Arch. Ophthalmol. 2006], TIA (trabecular-iris angle) [Leung, et al, Invest. Ophthalmol. Vis. Sci. 2008], TISA (trabecular-iris space area) [Tian, et al, TBME 2011]. It also outperforms other image processing and pattern recognition based approaches, such as HOG (histogram of gradients) [Xu, et al, EMBC 2012] and HEP (heistogram equalized pixels) [Xu, et al, EMBC 2013] based SVM classifiers.<br /> The system was tested on 30 AS-OCT circular scan videos from open and closed angle subjects detected by gonioscopy. The anterior chamber angle in a quadrant was classified as closed if the posterior trabecular meshwork could not be seen. An eye was classified as having angle closure if there were 2 or more quadrants of closure. Any irido-corneal contact beyond sclera spur (SS) was considered as angle closure in 3D ASOCT scans.
Results
The 30 circular scan videos comprise of 3840 AS-OCT frames from 30 patient eyes, 15 of them with primary angle closure and other 15 with open angles from normal subjects. Each video contains 128 frames, with two angles per frame. We adopted the leave-one-out (LOO) method. For each testing round, 6 videos (728 images) from 3 angle closure and 3 open angle patients were used for testing while others were used as references (training samples) for reconstruction, thus 5 rounds were performed to test all scans. The proposed automated system achieved an AUC value of 0.91±0.05 and a balanced accuracy of 80.3±7.4, in terms of per angle assessment.
Conclusions
Using an innovative image processing and reconstruction approach, the system achieves good performance in angle closure assessment from 3D AS-OCT images, and the results are promising, with potential to be expanded into an automated screening tool for primary angle closure glaucoma (PACG) detection.