This paper presents a novel hybrid preprocessing method for improving noise resilience and improved computational efficiency of image skeletonisation. Common techniques for extracting the topological skeleton of a shape include distance transforms, thinning, and geometric analysis. All of these methods are sensitive to noise, and can suffer from instability and unwanted spurs. In the case of needing to match skeletons from different images, instability can be a significant problem. Skeleton stability using the proposed preprocessing method is evaluated for a range of existing medial axis transforms. It is shown to be more effective for suppressing unwanted spurs and improving stability against other preprocessing techniques.