Auto-Contour Guided Prostate Patient Positioning

Abstract

To use contours automatically generated on the pre-treatment cone-beam CT (CBCT) to improve the registration with the plan CT for prostate image-guided radiation therapy (IGRT). 10 patients, each with 10 to 28 daily kV-CBCTs, had their prostate manually contoured on the plan CT and daily CBCTs. Prostate contours were also automatically generated on the CBCTs. Auto-contours are produced by deformably registering prior CBCTs (atlases) to the current CBCT and deforming the prior contours with the resulting deformation field. This set of deformed prostate contours on the current CBCT is then combined using the STAPLE algorithm. The optimal IGRT shift is determined by comparing the centroid of the manual prostate contour in the CBCT and plan CT. This shift is compared both to the clinical shift that was based on qualitative alignment of the images, and to a centroid shift based on auto-contours. The dosimetric impact of the different shift methods is estimated by overlaying the plan prescription isodose line on the CBCT according to the shifts, and then determining the coverage of the manual prostate and PTV contours. On average there is a 2.3±1.5 mm difference in the superior/inferior (SI) direction between the clinical and centroid shifts. The difference is 0.9±0.7 mm in the left/right (LR) and 1.8±1.5 mm in the anterior/posterior (AP) directions. Using the clinical shifts gave on average 98.7±2.6% coverage of the prostate, and 89.6±3.8% coverage of the PTV by the prescription dose. Centroid based shifts gave significantly better coverage of 99.5±1.1% (P=0.0013) and 92.7±2.4% (P<0.00001) for the prostate and PTV respectively. The reduction in standard deviation also implies more consistent coverage day to day with the centroid shifts. The average dice similarity coefficient of prostate auto-contours to manual contours is 0.91, which equals our intra-observer error. The average time for auto-contouring was 90 seconds on a Tesla K20c GPU. Using auto-contours to determine the centroid produces equivalent shifts to the manual contour centroid shifts, with average differences of 0.8±0.6, 0.4±0.3 and 0.8±0.7 mm in SI, LR and AP directions, respectively. This is to be expected given the consistency of the manual and auto-contours. There are >2 mm differences between our new centroid based shifts and the current clinical, image based shifts, with centroid based shifts providing significantly higher prostate coverage. Auto-contouring provides equivalent shifts to manual centroid based shifts. IGRT can thus be fully automated, resulting in improved prostate coverage.