Room: Exhibit Hall | Forum 9
Purpose: In the adaptive radiotherapy workflow, manual contouring on CBCT images is unpractical and an automatic approach is needed. The go-to approach has been mapping of structures via deformable image registration (DIR). Deep learning segmentation (DLS) has so far mainly been used on CTs due to the poor image quality inherent to CBCTs. This works compares the segmentation performance of a pre-trained CT DLS algorithm used on enhanced CBCTs with that of deformably mapping structures.
Methods: Two prostate cases were used in the study, one with small (P1) and one with larger (P2) anatomical differences between the CTs and the CBCTs. Ground truth segmentations were manually contoured on the CTs and CBCTs. The CBCT image quality was enhanced by a novel analytic correction and conversion algorithm. A deep learning segmentation algorithm, pre-trained on 45 CTs (no CBCTs in the training set), was used on the enhanced CBCT. The deformable registration was intensity-based and performed with the original non-enhanced CBCTs. Both a structure similarity comparison with the Dice Similarity Coefficient (DSC) and a comparison of dose statistics from clinical VMAT treatment plans were performed. The DIR, CBCT enhancement and DLS were all performed in a research version of the commercial treatment planning system RayStation 8B.
Results: In P1, the two methods showed similar DSC agreement to the ground truth. In P2, CBCT correction with DLS outperformed deformable mapping. In the dose statistics comparison, the results follow a similar trend. However, the dose statistics for the prostate were approximately the same in both cases due to robust plans.
Conclusion: The results suggest that DLS on enhanced CBCTs is the preferred approach in the adaptive workflow, as it utilizes the up-to-date patient information directly and has no need of a high-quality DIR. DIR may provide inaccurate dose statistics for organs-at-risk, demonstrated by P2.
Funding Support, Disclosures, and Conflict of Interest: Both authors are full-time employees at RaySearch Laboratories.