Angle-closure assessment in anterior segment OCT images via deep learning

•This work is the first attempt to classify anterior chamber angles into open, appositional- and synechial- angle-closure, by Anterior Segment Optical Coherence Tomography (AS-OCT), rather than the conventional gonioscopic examination.•We introduced a Multi-Sequence Deep Network (MSDN), which learns to identify discriminative representations from a sequence of AS-OCT images, especially with a view to improving performance in separating appositional-angle from its occludable angle forms.•We have constructed a AS-OCT dataset for which the AS-OCT of each eye were acquired under both dark and bright illumination conditions. We take advantage of the resulting changes in pupil size to simulate the pressure of the goniolens, which can push the angle open and help determine the true angle configuration. [Display omitted] Precise characterization and analysis of anterior chamber angle (ACA) are of great importance in facilitating clinical examination and diagnosis of angle-closure disease. Currently, the gold standard for diagnostic angle assessment is observation of ACA by gonioscopy. However, gonioscopy requires direct contact between the gonioscope and patients’ eye, which is uncomfortable for patients and may deform the ACA, leading to false results. To this end, in this paper, we explore a potential way for grading ACAs into open-, appositional- and synechial angles by Anterior Segment Optical Coherence Tomography (AS-OCT), rather than the conventional gonioscopic examination. The proposed classification schema can be beneficial to clinicians who seek to better understand the progression of the spectrum of angle-closure disease types, so as to further assist the assessment and required treatment at different stages of angle-closure disease. To be more specific, we first use an image alignment method to generate sequences of AS-OCT images. The ACA region is then localized automatically by segmenting an important biomarker - the iris - as this is a primary structural cue in identifying angle-closure disease. Finally, the AS-OCT images acquired in both dark and bright illumination conditions are fed into our Multi-Sequence Deep Network (MSDN) architecture, in which a convolutional neural network (CNN) module is applied to extract feature representations, and a novel ConvLSTM-TC module is employed to study the spatial state of these representations. In addition, a novel time-weighted cross-entropy loss (TC) is proposed to optimize the output of the ConvLSTM, and t