Automatic Cephalometric X-Ray Landmark Detection Challenge 2014: A machine learning tree-based approach

In this paper, we describe the machine learning approach we used in the context of the Automatic Cephalometric X-Ray Landmark Detection Challenge. Our solution is based on the use of ensembles of Extremely Randomized Trees combined with simple pixel-based multi-resolution features. By carefully tuning method parameters with cross-validation, our approach could reach detection rates ≥ 90% at an accuracy of 2.5mm for 8 landmarks. Our experiments show however a high variability between the different landmarks, with some landmarks detected at a much lower rate than others.