Fully automatic cephalometric evaluation using Random Forest regression-voting

Cephalometric analysis is commonly used as a standard tool for orthodontic diagnosis and treatment planning. The identification of cephalometric landmarks on images of the skull allows the quantification and classification of anatomical abnormalities. In clinical practice, the landmarks are placed manually which is time-consuming and subjective. This work investigates the application of Random Forest regression-voting to fully automatically detect cephalometric landmarks, and to use the identified positions for automatic cephalometric evaluation. Validation experiments on two sets of 150 images show that we achieve an average mean error of 1.6mm 1.7mm and a successful detection rate of 75% 85% for a 2mm precision range, and that the accuracy of our automatic cephalometric evaluation is 77% 79%. This work shows great promise for application to computer-assisted cephalometric treatment and surgery planning.