Comparison of 12 deformable registration strategies in adaptive radiation therapy for the treatment of head and neck tumors.

BACKGROUND AND PURPOSE Weight loss, tumor shrinkage, and tissue edema induce substantial modification of patient's anatomy during head and neck (HN) radiotherapy (RT) or chemo-radiotherapy. These modifications may impact on the dose distribution to both target volumes (TVs) and organs at risk (OARs). Adaptive radiotherapy (ART) where patients are re-imaged and re-planned several times during the treatment is a possible strategy to improve treatment delivery. It however requires the use of specific deformable registration (DR) algorithms that requires proper validation on a clinical material. MATERIALS AND METHODS Twelve voxel-based DR strategies were compared with a dataset of 5 patients imaged with computed tomography (CT) before and once during RT (on average after a mean dose of 36.8Gy): level-set (LS), level-set implemented in multi-resolution (LS(MR)), Demons' algorithm implemented in multi-resolution (D(MR)), D(MR) followed by LS (D(MR)-LS), fast free-form deformable registration via calculus of variations (F3CV) and F3CV followed by LS (F3CV-LS). The use of an edge-preserving denoising filter called "local M-smoothers" applied to the registered images and combined to all the aforesaid strategies was also tested (fLS, fLS(MR), fD(MR), fD(MR)-LS, fF3CV, fF3CV-LS). All these strategies were compared to a rigid registration based on mutual information (MI, fMI). Chronological and anti-chronological registrations were also studied. The various DR strategies were evaluated using a volume-based criterion (i.e. Dice similarity index, DSI) and a voxel-intensity criterion (i.e. correlation coefficient, CC) on a total of 18 different manually contoured volumes. RESULTS For the DSI analysis, the best three strategies were D(MR), fD(MR)-LS, and fD(MR), with the median values of 0.86, 0.85 and 0.85, respectively; corresponding inter-quartile range (IQR) reached 9.6%, 10% and 10.2%. For the CC analysis, the best three strategies were fD(MR)-LS, D(MR)-LS and D(MR) with the median values of 0.97, 0.96 and 0.94, respectively; corresponding IQR reached 11%, 9% and 15%. Concerning the time-sequence analysis, the anti-chronological registration (all deformable strategies pooled) showed a better median DSI value (0.84 vs 0.83, p<0.001) and IQR (11.2% vs 12.4%). For CC, the anti-chronological registration (all deformable strategies pooled) had a slightly lower median value (0.91 vs 0.912, p<0.001) but a better IQR (16.4% vs 21%). CONCLUSIONS The use of fD(MR)-LS is a good registration strategy for HN-ART as it is the best compromise in terms of median and IQR for both DSI and CC. Even though less robust in terms of CC, D(MR) is a good alternative. None of the time-sequence appears superior.