A comparison of deformable registration techniques for pre and post-treatment cholangiocarcinoma CT images

Purpose: Response assessment of radiotherapy in the treatment of cholangiocarcinomas across longitudinal images is challenging due to liver deformations caused by tumor response and patient positioning. To address this, advanced deformable image registration (DIR) techniques are required. Here we compare four commercially used DIR techniques. Methods: Fifty-one locally treated choloangiocarcinoma patients were retrospectively obtained. Liver segmentations were automatically generated on pre and post-radiotherapy CT images, using a convolutional neural network. The post-treatment image was then registered to the pre-treatment image using four DIR techniques that are available in a commercial treatment planning system: (1) intensity-based registration (IR); (2) intensity and structure-based hybrid registration targeted on the liver (HRL); (3) a finite-element-based biomechanical registration targeted on the liver (BRL) and (4) BRL with a single manually-contoured internal structure as an additional target (BRLI). Accuracy of registration was evaluated using target registration error (TRE). The results were statistically analyzed with a one-way ANOVA. This was followed by Tukey multiple-comparisons to test individual differences. To compensate for different image resolutions, the TREs were standardized for the ANOVA by dividing them by the corresponding voxel diagonal. Results: We have analyzed twenty patients to date. Out of the DIR techniques, BRLI had the best mean TRE (5.6mm, SD 2.0mm) and resulted in the best accuracy in nine of the twenty cases. Moreover, for 60% of the patients the performance was within the image resolution. The statistical analysis found DIR technique to be a significant factor (p-value=0.003). However, the multiple-comparisons showed that only IR was statistically different (poorer performance) from the other methods. Conclusion: Our study demonstrates the limitations of standard intensity-based DIR as well as the potential utility of biomechanical DIR. Automated segmentation of the entire vasculature is expected to further improve the biomechanical technique with internal boundary conditions and enable a fully automated workflow