Landmark Detection on Cephalometric X-rays Using Particle Swarm Optimisation

Locating special points of interest, known as landmarks, on X-rays of human heads is a time consuming manual process in the medical field known as cephalometry. We automate this task using the evolutionary computing approach of particle swarm optimisation (PSO). Particularly, we represent several existing programming solutions produced by genetic programming as linear function optimisation tasks. Seven experiments are performed for landmarks with varying detection difficulties. The detection accuracies, required evaluations and detector sizes were compared. Our results show that PSO is up to 14% more accurate at testing on the hardest landmarks. It also matched the solutions of genetic programming with 43% to 74% less training times, and 33% to 78% smaller program sizes. It is observed that PSO finds cephalometric landmark detectors with greater success than the existing genetic programming approaches, in both detection accuracy and computational efficiency.