Accurate contouring of target areas affected by cancer is crucial in radiotherapy planning to provide effective treatment for patients. Beyond producing sufficiently precise contours, it is also important to have a reliable measure of the associated uncertainty of where the true anatomical boundaries may be located. Inter-observer variability arising from multiple experts annotating the boundaries of the target area provides a representation of this uncertainty; however, these annotations are labour-intensive to obtain over a large number of images, often making it infeasible in practice. In this study, we evaluate the clinical relevance of predictive auto-contouring uncertainty from training on clinician-defined ranges - a set of annotations from a single expert representing the delineation uncertainty of the target area, which are far more efficient to produce. This aims to bridge the gap in the performance of uncertainty estimates between training on a single contour and a set of contours from multiple annotators. This is achieved by curating a cervical cancer dataset with uncertainty annotations produced on CT scans from patients undergoing radiotherapy. We demonstrate that the resulting uncertainty from a model trained on these clinician-defined ranges is more meaningful compared with training on single contours without an uncertainty range. We also validated its clinical utility with respect to the inter-observer variation on a small hold-out set.