Couinaud segmenation, which divides the liver into functional regions, is the most widely used functional anatomy of the liver and is important for surgical planning and lesion monitoring. Couinaud segmentation can be a time-consuming task, thereby necessitating automated methods. In this study, we propose a deep learning approach for automatically defining Couinaud segments 2 to 8, based on a novel application of automatic landmark localisation. We utilise a heatmap regression CNN to predict landmark locations in the liver, which can subsequently be used to derive the planes that divide the liver into Couinaud segments. A novel postprocessing step for reducing false-positive peaks in heatmaps and/or aiding quality control is also presented. We apply our approach to non-contrast T1-weighted MRI data and compare the accuracy of the derived segments to those obtained directly from a semantic segmentation network. We show that the approach we propose can match and potentially outperform the direct segmentation approach, and thus can be a good alternative option for automatic Couinaud segmentation.