Robust quality assurance criteria for deformable image registration for adaptive radiotherapy

Purpose or Objective
Plan adaption and accurate dose accumulation in state-of-the-art radiotherapy, in particular in thorax and abdomen, rely increasingly on deformable image registration (DIR). As a consequence, clinical work-flows encompassing DIR necessitate a reliable and fast quality assurance (QA) that enable on-the-fly go/no-go decisions to proceed. The disadvantage of established QA criteria such as the Dice similarity coefficient (DSC) or Hausdorff distance (HDD) for this purpose is thereby that both do not allow complete “hands-free” operation and also frequently display a limited sensitivity with respect to registration errors in soft tissues. Here, we investigate the structural-similarity index (SSI) as an alternative QA criterion, with respect to robustness, specificity and sensitivity for the decision-making process. As the SSI does not need operator input, it also enables fully automated QA for DIR during adaptive procedures with repeated imaging.

Materials and Methods
We have investigated the correlations between a set of QA criteria (including the DSC, the HDD, the mean HDD, the Jaccard index, and the SSI) and the endpoint error of known deformations as the gold standard. All criteria are extensively evaluated using (I) synthetic deformations of 3D MRI prostate patient data sets (3D bTFE with fat suppression) of different amplitudes, (II) simulated biomechanical abdominal and thoracic deformations based on finite-element simulations, and (III) annotated 4D patient data. We furthermore evaluated the QA performance for different DIR algorithms on images of multiple resolutions and signal-to-noise-ratios. The quality assurance criteria are scored in terms of the Pearson correlation coefficient, a linear regression analysis, and a receiving operator characteristic (ROC) where the median endpoint error is used as the cutoff value for decision making.

Results
As shown in table 1, compared to established criteria we find that the SSI has both a higher Pearson correlation coefficient and a higher R2 for linear regression. As shown in figure 1, this leads in turn to a larger area under the curve for the ROC. The ROC-curve shows that the SSI is both more sensitive as well as more specific to detect misregistrations, which is essential for decision making on DIR in clinical radiotherapy work-flows.

Conclusion
The SSI is a QA criterion for DIR with higher and more consistent correlations with the endpoint error than established QA criteria. Due to its higher specificity and sensitivity it is more suited to make decisions on accepting registrations for clinical use. Additionally, the SSI metric does not require operator input and is therefore well suitable in clinical work-flows aiming to reduce operator burden. This also enables fully automated QA for DIR for adaptive radiotherapy treatments.