Robust landmark identification for generating subject specific models for biomechanics

Abstract

We propose a robust and accurate method for automatic landmark identification for torso MR images. The method combines cross-correlation and statistical models into a single framework. Principal component analysis is used to generate statistical models of the relative landmark positions, and the template images. Partial least-squares regression is used predict the initial landmark positions, and template images for the landmarks from the characteristics of the unseen MR images. The landmark template images are then cross-correlated with the search regions and the statistical model is used to constrain the search for the maximum combined correlation. The method was trained and tested using MR images from 51 female subjects. The method was able to identify the position of the tracheal bifurcation and jugular notch landmarks with a mean±SD error of 6.1±5.2 mm, with 9.1 % of the errors greater than 10 mm. This result was three times better than the standard template matching method, which gave a mean±SD error of 18.9±21.7 mm, with 33 % of the errors greater than 10 mm.