Fifty-seventh annual meeting of the American association of physicists in medicine
SU-D-207-02: Automated Respiratory Motion Tracking with Machine Vision
Current systems for obtaining respiratory information at time of treatment typically require the use of additional equipment and add to patient setup time. In this study, we develop a fully automated method to track respiratory motion by tracking the motion of the diaphragm surface, making use of only current imaging technology and without the need for additional mechanical, pneumatic, or optical devices.
A statistical algorithm, RANdom SAmple Consensus (RANSAC), was used for direct tracking of the diaphragm surface through kV CBCT projection images. A third order polynomial was used to model the shape of the diaphragm. Model constraints based on the diaphragm shape as well as a small motion assumption (between adjacent projections) were incorporated to constrain the model. The changing position of the diaphragm along the Superior-Inferior direction as the patient respires then gives the respiratory trace.
On the patients studied, the respiratory trace can be obtained with between 5,000 and 1,000,000 iterations of the RANSAC algorithm, per projection. Visual comparison with the projection images in a cine like mode shows the computed respiratory trace matches well with the true respiratory cycle. With a low number of iterations (<10,000 per projection), the diaphragm location computation time is similar to projection image acquisition and therefore the trace can be computed in real time.
The RANSAC algorithm is a viable method for obtaining a respiratory trace with the patient on the treatment couch. Implementing the RANSAC algorithm on a GPU should allow for a significantly larger number of iterations while maintaining the ability to acquire the respiratory trace in real time. This method of obtaining a respiratory trace should reduce costs for technology such as 4DCBCT as well as reduce patient setup times and improve patient comfort.