Achieving accurate automated image registration for prostate radiotherapy

Prostate cancer is the second most common type of cancer in men in the United States affecting about 0.4% of men. If detected early, prostate cancer is treatable. Three dimensional (3D) conformal radiation therapy, which is employed for the treatment of localized cancers, allows radiation dose escalation and improved local tumor control while sparing surrounding healthy tissue. 3D conformal radiation therapy is susceptible, however, to geometric (targeting) uncertainties in treatment delivery. Geometric uncertainties can be reduced by matching low resolution two-dimensional (2D) portal images acquired during treatment to a higher resolution 3D planning computed tomography (CT) image to determine deviations from the intended patient setup.; The objective of this study was to develop an automated 3D to 2D image registration framework to determine patient setup deviations in radiation therapy. The underlying assumption of this study is that CT and portal images are related modalities. Three registration methods, line-based correlation, intensity-based correlation and maximum likelihood matching, were formulated based on the imaging physics and different assumptions on signal and noise characteristics of the images.; The methods were validated on accurately collected kilovoltage cone beam CT (CBCT) images and kilovoltage radiographs and megavoltage portal images of an anthropomorphic phantom of the pelvis. Additional validation of the line-based algorithm was performed on phantom, simulated and patient data acquired with conventional CT scanners. Estimates of setup deviations obtained with the three methods were accurate to at least 0.25mm and 0.1° on the CBCT and kilovoltage radiographs and at least 0.7mm and 0.2° on the CBCT and portal images. The accuracy of the line-based correlation algorithm on the additional data tested was better than 1mm and 1°.; This study achieved setup verification accuracy that far exceeds current clinical expectations. The results obtained support the assumptions and arguments made. Namely, the relationship between kilovoltage and megavoltage images can be approximated to be linear, that useful signal is present at the higher spatial frequencies of the images employed, and that the noise in the images can be assumed to be independent and Gaussian.