Angle-closure glaucoma is one of the major causes of blindness in Asia. In this paper, we present a new approach for the classification of the anterior chamber angles into open, narrowed, and closure, in anterior segment optical coherence tomography (AS-OCT), by learning the manual annotations from gonioscopy, so as to further assist the assessment of angle-closure glaucoma. The proposed framework firstly localizes the anterior chamber angle region automatically, which is the primary structural image cue for clinically identifying glaucoma. Then three scales of cropped chamber angle images are fed into our Multi-Scale Regions Convolutional Neural Networks (MSRCNN) architecture, in which three parallel convolutional neural networks are applied to extract feature representations. Finally, the representations are stacked to fully-connected layer for glaucoma type classification. The proposed method is evaluated across a dataset of 9728 anterior chamber angle images, and the experimental results show that the proposed method outperforms existing state-of-the-art methods in applicability, effectiveness, and accuracy.