NREC designed and implemented a medical image registration system to accurately estimate a patient’s position for therapy.
For radiotherapy and other forms of therapy to be successful, patients must be correctly positioned for treatment. This is a challenging problem requiring short registration times and high accuracies. Registration algorithms typically involve trade-offs between speed of execution, accuracy, and ease of application. Image-based registration algorithms, which gather data from large portions of the image in order to increase accuracy, are computationally intensive, and typically suffer performance degradation when the input images contain clutter.
CMU has developed an image-comparison algorithm, Variance-Weighted Sum of Local Normalized Correlation, which greatly decreases the impact of clutter and unrelated objects in the input radiographs. This image comparison approach is combined with hardware-accelerated rendering of simulated X-ray images to permit registration of noisy, cluttered images with sub-millimeter accuracy.
The medical image registration system uses two-dimensional X-rays and three-dimensional CT scans to accurately estimate the position and orientation of a patient’s anatomy with respect to an external coordinate frame.
High-speed computer graphics algorithms render simulated X-rays from CT scan data in a fraction of a second. These high-resolution, simulated radiographs are compared with actual X-ray images using a novel image comparison algorithm. The comparison algorithm ignores noise in the images and clutter from the patient's other anatomy to correctly estimate the patient's position.
The system achieved sub-millimeter registration accuracies in preliminary tests, with total registration times on the order of 50 seconds.