Iris recognition is considered to be one of the most widely used biometric modality, mainly due to its non-invasive nature and high reliability. However, in the whole process of authentication, segmentation of the iris is the most crucial one as being the second stage of the usual five-stage pipeline, the error introduced gets compounded in the subsequent stages. However, segmentation of the iris in non-ideal conditions is a challenging task owing to numerous artefacts such as occlusion by eyelids, off-angle rotations, irregular reflections, and blurred boundaries. Although the artefacts can be minimised up to a certain extent during the acquisition process, it requires a high level of control over the image capturing environment and also high user cooperation, which is not always feasible. For segmentation, quite a few methods have been put forward, but the ones using classical approaches usually have low generalisability. Over the past decade, various deep learning techniques have been proposed which have given satisfactory results. As the problem at hand is that of an image-to-image generation (the input image and its corresponding segmentation mask), the most common similarity amongst them is the use of a standard encoder-decoder structure called the UNet. In this chapter, we discuss several such techniques and their intricate novelties, and shortcomings, while also throwing some light on the non-deep learning methods so as to get a wholesome comparison. We also discuss briefly about the various publicly available data sets and the artefacts they are riddled with while also discussing about the various metrics that are used by the scientific community to compare their works and establish the state-of-the-art. Lastly, we discuss a short implementation of the UNet done by ourselves on two of the available data sets and conclude this chapter with a thought on the future possibilities of the existing works.