Validation of an accelerated ‘demons’ algorithm for deformable image registration in radiation therapy

A greyscale-based fully automatic deformable image registration algorithm, originally known as the 'demons' algorithm, was implemented for CT image-guided radiotherapy. We accelerated the algorithm by introducing an 'active force' along with an adaptive force strength adjustment during the iterative process. These improvements led to a 40% speed improvement over the original algorithm and a high tolerance of large organ deformations. We used three methods to evaluate the accuracy of the algorithm. First, we created a set of mathematical transformations for a series of patient's CT images. This provides a 'ground truth' solution for quantitatively validating the deformable image registration algorithm. Second, we used a physically deformable pelvic phantom, which can measure deformed objects under different conditions. The results of these two tests allowed us to quantify the accuracy of the deformable registration. Validation results showed that more than 96% of the voxels were within 2 mm of their intended shifts for a prostate and a head-and-neck patient case. The mean errors and standard deviations were 0.5 mm ± 1.5 mm and 0.2 mm ± 0.6 mm, respectively. Using the deformable pelvis phantom, the result showed a tracking accuracy of better than 1.5 mm for 23 seeds implanted in a phantom prostate that was deformed by inflation of a rectal balloon. Third, physician-drawn contours outlining the tumour volumes and certain anatomical structures in the original CT images were deformed along with the CT images acquired during subsequent treatments or during a different respiratory phase for a lung cancer case. Visual inspection of the positions and shapes of these deformed contours agreed well with human judgment. Together, these results suggest that the accelerated demons algorithm has significant potential for delineating and tracking doses in targets and critical structures during CT-guided radiotherapy.