Validation of the deformable image registration system elastix in the head and neck region

Purpose/Objective: Deformable registrations of medical images are becoming widely used for both radiotherapy optimization, but also for treatment response evaluation, e.g. defining relapses. However, the accuracy of deformable image registration (DIR) is rarely validated. The current study evaluates the accuracy of the open source deformable registration tool elastix when used for registration of different organ structures on planning CT and relapse CT scans of head and neck patients. Materials and Methods: Twenty patients treated with definitive IMRT for oral cavity, oropharynx and hypopharynx cancer in 2010-2012 were retrospectively and randomly included in the study. For each patient a planning CT and a diagnostic relapse CT scan were used (CT voxels were 3 mm in the cranio-caudal (CC) direction and 1 mm in the in-plane directions). One observer manually contoured 8 organs (spinal cord, mandible, right/left parotid and submandibular glands, thyroid gland and vertebrae C3-5) on planning CT (pCT), relapse CT (rCT) and re-delineated again on the planning CT (reCT). The contouring on the relapse CT was mapped to the planning CT using elastix (http://elastix.isi.uu.nl/). Spatial analysis of the difference in ROI delineation was performed by measuring the minimum distance between the two ROI's on each surface point of the original ROI. For each ROI the combined uncertainty of DIR and delineation was calculated as the standard deviation of the minimum surface distances. Organ type specific uncertainties were calculated as Root- Mean-Square (RMS) of the patient individual standard deviations. Re-delineation was used as a baseline for the obtainable uncertainty. Spearman's correlation between uncertainties of re-delineation and deformed delineation was performed to check whether delineation uncertainty influenced the outcome of the deformable results. Results: Surface plot of the spatial resolved deviation showed that the largest deviations where mainly in the CC direction (areas where the ROI surface normal primarily is in the CC direction) (left submandibular gland as an example, picture). RMS values for re-delineation and DIR are shown in table. A significant but low correlation exists between the ROI uncertainty of individual re-delineation and deformable contours (R=0.23, p=0.004). Thus a part of the deformable uncertainties are related to uncertainty of the manual ROI delineation. Significant correlations within single organs were not found. Conclusions: Deformable registration of head and neck CT images using elastix resulted in a combined delineation and deformation uncertainty of approximately twice the uncertainty related to the manual delineation performed on one CT. These uncertainties should be considered when using image deformations and registrations for clinical and research purposes.