Accurate landmark-based segmentation by incorporating landmark misdetections

We propose a novel approach to landmark detection, where the appearance of landmarks is modeled by Haar-like features and a random forest classifier, and spatial relationships among landmarks are modeled by Gaussian distributions augmented by shape-based random forest classifiers for identification of landmark misdetections. The proposed approach was evaluated on segmentation of lung fields from a publicly available database of chest radiographs, and the resulting segmentation performance was 1.43 ± 0.85 mm in terms of boundary-to-boundary distance and 95.3 ± 2.0% in terms of area overlap coefficient. By objective comparison to existing state-of-the-art approaches, the proposed approach proved superior in terms of segmentation results and computational efficiency.