Conventional approaches to feature extraction based on shape typically focus on using mathematical descriptors such as aspect ratio, rectangularity, area, etc., or the Centroid Contour Distance Curve (CCDC) to describe the shape based upon edge points. Such descriptors are able to achieve modest success when classifying leaf species with example-limited datasets. In this paper, we propose the use of Medial Axis Transformation (MAT) to convert the complex shapes of leaves to graph structures that describe the topological skeleton of the leaves, and utilize the topological skeleton to extract features. From a data set composed of 99 different leaf classes with 10 samples per class, we show that 18 features extracted from the topological skeleton outperforms 64 features obtained from CCDC with `10-fold' cross validation accuracies of 73.84% and 58.08% respectively when using random forest as a classifier. This result suggests that MAT-graph-based features are able to more succinctly distinguish shapes when compared with conventional approaches.