Anatomic-landmark detection using graphical context modelling

Anatomical landmarks in images play an important role in medical practice. This paper presents a graphical model that fully automatically detects such landmarks. The model includes a unary potential using a random forest classifier based on local appearance and binary and ternary potentials encoding geometrical context among different landmarks. The weightings of different potentials are learned in a maximum likelihood manner. The final detection result is formulated as the maximum-a-posteriori estimation jointly over the whole set of landmarks in one image. For validation, the model is applied to detect right-ventricle insert points in cardiac MR images. The result shows that the context modelling is able to substantially improve the overall accuracy.