Evaluation of deformable image registration accuracy for CT images of the thorax region

Evaluate the performance of three commercial deformable image registration (DIR) solutions on computed tomography (CT) image-series of the thorax.

DIRs were performed on CT image-series of a thorax phantom with tumor inserts and on six 4-dimensional patient CT image-series of the thorax. The center of mass shift (CMS), dice similarity coefficient (DSC) and dose-volume-histogram (DVH) parameters were used to evaluate the accuracy. Dose calculations on deformed patient images were compared to calculations on un-deformed images for the gross tumor volume (GTV) (D_mean, D_98%), lung (V_20Gy, V_12Gy), heart and spinal cord (D_2%).

Phantom structures with constant volume and shifts ≤30 mm were reproduced with visually acceptable accuracy (DSC ≥ 0.91, CMS ≤ 0.9 mm) for all software solutions. Deformations including volume changes were less accurate with 9/12 DIRs considered visually unacceptable. In patients, organs were reproduced with DSC ≥ 0.83. GTV shifts ≤1.6 cm were reproduced with visually acceptable accuracy by all software while larger shifts resulted in failures for at least one of the software. In total, the best software succeeded in 18/25 DIRs while the worst succeeded in 12/25 DIRs. Visually acceptable DIRs resulted in deviations ≤3.0% of the prescribed dose and ≤3.6% of the total structure volume in the evaluated DVH-parameters.

The take home message from the results of this study is the importance to have a visually acceptable registration. DSC and CMS are not predictive of the associated dose deviation. Visually acceptable DIRs implied dose deviations ≤3.0%.