Learning Articulated Models of Joint Anatomy from Utrasound Images

Parts of a joint anatomy, such as bones or the joint center can be robustly identified in an ultrasound image with the help of an articulated or structural model. Such a model is a structure of parts that represent the bones and skin as polygonal chains and the join as a point, where the parts remain within specified geometric relations. The parts are identified by registration or a match of a structural description derived from the ultrasound image with the articulated model. To account for anatomical differences between the subjects, a library of joint models must be constructed, each model representing a class of joints, where all models together cover the range of possible anatomies. A new method of unsupervised learning is proposed for constructing the library of joint models by clustering structural descriptions computed from image annotations. The clustering method uses an inter-model distance measure defined as a minimum of the objective function that measures a discrepancy between structural descriptions. The objective function is minimized through a search for a best match between two structural descriptions. The method presentation is illustrated with the results of its application to ultrasound images of finger joints.