Backbone extraction aims to reduce network size while preserving key structural features. Many methods exist, but few have been compared systematically. Existing studies often rely on real-world networks, where controlling topological properties is difficult. In this paper, we use synthetic networks generated with the Extended Lancichinetti-Fortunato-Radicchi (LFR) benchmark. This allows us to vary node strength, degree distribution, and network size independently. We evaluate five backbone extraction techniques: Disparity filter, Metric backbone, High Salience Skeleton (HSS), Globally and Locally Adaptive Backbone (GLANB), and Locally Adaptive Network Sparsification (LANS). We assess their performance based on the proportion of preserved nodes and edges. Results show that LANS performs consistently across all scenarios. It preserves all nodes and has linear complexity. Metric also retains all nodes but keeps more edges. GLANB scales well but is more computationally expensive. Disparity and HSS show strong sensitivity to network properties. These findings support informed method selection and scalable application of backbone extraction in diverse settings.