To test and evaluate the geometric accuracy of delineation of organs at risk (OARs) in head and neck cancer using an atlas-based autosegmentation (ABAS) software.

The atlases for the ABAS software was generated using images from 40 patients with head and neck cancer undergoing intensity-modulated radiotherapy. The software was tested in 40 new patients. Automatic delineation of OARs was carried out on computed tomography images by single-(one to one) and multi-template (ten to one) approaches. In order to evaluate the feasibility of the automatic delineation in clinical application, differences in volume (ΔV%), position (Δx, Δy, and Δz), conformability (sensitivity (S_e), specificity (S_p), and dice similarity coefficient (DSC)), and delineation time were assessed between the automatic and manual delineation. The comparison between the two automatic delineation approaches was made by paried t test.

For all OARs, the multi-template automatic delineation achieved a significantly smaller mean ΔV% value and a significantly larger mean DSC value than the single-template automatic delineation (-0.02%±0.29% vs. -0.16%±0.41%, P<0.05; 0.74±0.16 vs. 0.68±0.20, P<0.05); the position differences between two automatic delineation approaches were less than 0.4 cm in all three directions except for the temporal lobe, lower jaw, and spinal cord; in the receiver operating characteristic curve defined by S_e versus 1-S_p, the data points were all within the first quadrant except for the optic nerve and chiasm; automatic delineation saved 42%-72% of time compared with manual delineation.

The ABAS software achieves satisfactory results of automatic delineation for most of OARs in patients with head and neck cancer. The multi-template automatic delineation, particularly, has better outcomes than the single-template one. In addition, it greatly shortens the time the clinicians spend on delineation of OARs.

Auto-contour; Head and neck neoplasm; Organ at risk