Registration of the Spine Using a Physically-Based Image Model for Ultrasound

Despite significant efforts towards various applications of ultrasound-based tissue registration, ultrasound still plays a minor role in clinical computer-assisted intervention. Interpretation of the images is one major obstacle. Towards a robust and accurate approach to automated interpretation, we have developed a probabilistic model representing ultrasonic images in terms of surface shape. The model is derived from a physical description of image formation that incorporates the shape and microstructure of tissue and characteristics of the imaging system. A framework for inference of surface shape is formed by constructing a data likelihood from the probabilistic model. We have used this likelihood with a quasi-Newton optimization algorithm to estimate the pose of a vertebra from a set of three simulated images. In 20 trials, the estimate error was less than 0.2 mm and 0.4 degrees in 15 trials and over 1.0 degrees in only 1 trial. While much work remains to develop clinical utility in any application, these results indicate significant potential for the approach.