A task-specific contour tracker for ultrasound

One way of dealing with the problems of computerized analysis of ultrasound images is to develop task-specific systems, which allows assumptions about the structures under consideration. Here, the authors introduce a new ultrasound contour tracker, tonTrak, for the mid-sagital and coronal ultrasound image sequences of the human tongue. tonTrak makes assumptions about the ultrasound modality and possible positions of the tongue inside the vocal tract. It also provides intuitive and user friendly ways of interaction with the user, a crucial feature for a medical software system. tonTrak is fundamentally different from the previous tongue tracking systems. It is based on the minimization of a deformable dual-mesh energy using a novel gradient descent technique that provides a way of efficiently imposing both spatial and temporal constraints at the same time. This new energy optimization method does not need contour initialization and it is robust against local minima. Numerous experiments and user feedback verified the efficiency and accuracy of the system. It is currently being used by speech and swallowing scientists.