Purpose: To evaluate the clinical potential of deformable image registration (DIR)-based automatic propagation of physician-drawn contours from a planning CT to midtreatment CT images for head and neck (H&N) adaptive radiotherapy.
Methods: Ten H&N patients, each with a planning CT (CT1) and a subsequent CT (CT2) taken approximately 3-4 week into treatment, were considered retrospectively. Clinically relevant organs and targets were manually delineated by a radiation oncologist on both sets of images. Four commercial DIR algorithms, two B-spline-based and two Demons-based, were used to deform CT1 and the relevant contour sets onto corresponding CT2 images. Agreement of the propagated contours with manually drawn contours on CT2 was visually rated by four radiation oncologists in a scale from 1 to 5, the volume overlap was quantified using Dice coefficients, and a distance analysis was done using center of mass (CoM) displacements and Hausdorff distances (HDs). Performance of these four commercial algorithms was validated using a parameter-optimized Elastix DIR algorithm.
Results: All algorithms attained Dice coefficients of >0.85 for organs with clear boundaries and those with volumes >9 cm(3). Organs with volumes <3 cm(3) and/or those with poorly defined boundaries showed Dice coefficients of ∼ 0.5-0.6. For the propagation of small organs (<3 cm(3)), the B-spline-based algorithms showed higher mean Dice values (Dice = 0.60) than the Demons-based algorithms (Dice = 0.54). For the gross and planning target volumes, the respective mean Dice coefficients were 0.8 and 0.9. There was no statistically significant difference in the Dice coefficients, CoM, or HD among investigated DIR algorithms. The mean radiation oncologist visual scores of the four algorithms ranged from 3.2 to 3.8, which indicated that the quality of transferred contours was "clinically acceptable with minor modification or major modification in a small number of contours."
Conclusions: Use of DIR-based contour propagation in the routine clinical setting is expected to increase the efficiency of H&N replanning, reducing the amount of time needed for manual target and organ delineations.