ESTRO 2024 - Abstract Book

S3752

Physics - Image acquisition and processing

ESTRO 2024

Results:

We find for all 273 of our patients all vertebrae in the CBCTs were correctly labeled: In every single case, we find that the global optimum of our full scan cost function puts the CBCT vertebrae on the correct corresponding CT vertebra. The distance found between the clinical registration and the global cost function optimum is always less than 10mm, and within 5mm in 90% of cases.

Conclusion:

We conclude that a simple, rigid registration between CBCT and a prior CT with verified vertebra labels, with a cost function calculated over the full scan, and initiated from multiple points, is sufficient to obtain correct labeling of vertebra segmentation on CBCT. This enables safe “one-stop-shop” treatments, using the CBCT to delineate and plan the treatment on. It can also be used as a safeguard against mistreatment in more conventional treatments, by avoiding the risk of a misregistration that is off by a full vertebra, which clinical registration algorithm settings using a bone match on a limited field of view are prone to.

Keywords: Vertebra, labeling, CBCT

References:

[1] Lessmann N, van Ginneken B, de Jong PA, Išgum I. Iterative fully convolutional neural networks for automatic vertebra segmentation and identification. Med Image Anal. 2019 Apr;53:142-155. doi: 10.1016/j.media.2019.02.005. Epub 2019 Feb 12. PMID: 30771712.