Multiscale guided attention network for optic disc segmentation of retinal images

•We propose a deep learning method to segment optic disc in retinal images.•Dual-branch encoder residually learns multiscale features from RGB and LAB images.•Guided attention emphasises OD-specific features over background features.•The method achieves superior performance on clinical optic disc image datasets.•The method achieves highly comparable results on the public fundus image datasets. Optic disc (OD) segmentation from retinal images is crucial for diagnosing, assessing, and tracking the progression of several sight-threatening diseases. This paper presents a deep machine-learning method for semantically segmenting OD from retinal images. The method is named multiscale guided attention network (MSGANet-OD), comprising encoders for extracting multiscale features and decoders for constructing segmentation maps from the extracted features. The decoder also includes a guided attention module that incorporates features related to structural, contextual, and illumination information to segment OD. A custom loss function is proposed to retain the optic disc's geometrical shape (i.e., elliptical) constraint and to alleviate the blood vessels' influence in the overlapping region between the OD and vessels. MSGANet-OD was trained and tested on an in-house clinical color retinal image dataset captured during ophthalmodynamometry as well as on several publicly available color fundus image datasets, e.g., DRISHTI-GS, RIM-ONE-r3, and REFUGE1. Experimental results show that MSGANet-OD achieved superior OD segmentation performance from ophthalmodynamometry images compared to widely used segmentation methods. Our method also achieved competitive results compared to state-of-the-art OD segmentation methods on public datasets. The proposed method can be used in automated systems to quantitatively assess optic nerve head abnormalities (e.g., glaucoma, optic disc neuropathy) and vascular changes in the OD region.