Evaluation of a deformable image registration quality assurance tool for head and neck cancer patients

A challenge in implementing deformable image registration (DIR) in radiation therapy planning is effectively communicating registration accuracy to the radiation oncologist. This study aimed to evaluate the MIM® quality assurance (QA) tool for rating DIR accuracy.

Retrospective DIR was performed on CT images for 35 head and neck cancer patients. The QA tool was used to rate DIR accuracy as good, fair or bad. Thirty registered patient images were assessed independently by three RTs and a further five patients assessed by five RTs. Ratings were evaluated by comparison of Hausdorff Distance (HD), Mean Distance to Agreement (MDA), Dice Similarity Coefficients (DSC) and Jacobian determinants for parotid and mandible subregions on the two CTs post-DIR. Inter-operator reliability was assessed using Krippendorff's alpha coefficient (KALPA). Rating time and volume measures for each rating were also calculated.

Quantitative metrics calculated for most anatomical subregions reflected the expected trend by registration accuracy, with good obtaining the most ideal values on average (HD = 7.50 ± 3.18, MDA = 0.64 ± 0.47, DSC = 0.90 ± 0.07, Jacobian = 0.95 ± 0.06). Highest inter-operator reliability was observed for good ratings and within the parotids (KALPA 0.66-0.93), whilst ratings varied the most in regions of dental artefact. Overall, average rating time was 33 minutes and the least commonly applied rating by volume was fair.

Results from qualitative and quantitative data, operator rating differences and rating time suggest highlighting only bad regions of DIR accuracy and implementing clinical guidelines and RT training for consistent and efficient use of the QA tool.