Glaucoma is one of the most critical retinal diseases that can cause irreversible blindness if not treated in time. The primary detection of this disease is usually done based on the diameter measurement of optic disc (OD) region in a color fundus image. Many artificial intelligence based expert systems were developed to automatically segment the OD region of a given fundus image. Most of the recent research in this direction employs a variant of convolution neural network based U-Net architecture, which substantially increases the computational load on the training system. Also, these methods are invariant to the pose, orientation, and spatial relationship of OD to other lesions in a given fundus image. In this letter, we propose a novel lightweight W-capsule network, which uses capsules for semantic segmentation through selective routing. The instantaneous vectors of the W-shaped structure at the bottom of the network accurately extract the shape of OD along with its orientation and spatial relationship to other fundus biomarkers, making it unique in compensating for variations due to image acquisition settings. The model was trained and validated on the Indian diabetic retinopathy image dataset, DRISHTI, and RIM-ONE datasets and achieved near state-of-the-art accuracy (acc) of 0.995 and intersection over union score of 0.931 with only 2.68M computational parameters, which is only 8.64% of most widely used U-Net-based segmentation structure.