ESTRO 2024 - Abstract Book

S3751

Physics - Image acquisition and processing

ESTRO 2024

Distinguishing and labeling vertebrae on CBCT is a challenging task due to the limited field of view, combined with reduced image quality compared to CT. Standard CBCT to CT registration software such as XVI (Elekta AB, Stockholm, Sweden) can result in an incorrect registration with an offset of a full vertebra, possibly resulting in a missed target. Similarly, automatic delineation algorithms delineate vertebrae on CBCT with high accuracy, as they are clearly visible, but often result in poor labeling, with offsets of a full vertebra. We show that providing prior information in the form of a previously acquired CT eliminates vertebra labeling errors and the associated full-vertebra positioning errors. Our methodology was validated retrospectively using 273 conventional palliative vertebra treatments. Solving this issue of vertebra labeling enables safe fully automated palliative vertebra irradiation.

Material/Methods:

Our dataset consists of 273 patients treated palliatively on the vertebrae (28 cervical, 135 thoracic and 110 lumbar) using conventional CBCT-guidance. For each patient, we have a CBCT (Elekta AB, Stockholm, Sweden), a CT, the treatment plan, and the parameters for the clinical CBCT-CT registration. To delineate and label the vertebrae on the CBCT, a deep learning algorithm is applied [1]. This algorithm was developed for and trained on regular CTs, and its delineations on CBCT are satisfactory, but the vertebra labeling is incorrect in 40% of cases, mostly off-by-one.

To ensure the correctness of labeling in clinical use, we register the CBCT to a prior scan of the patient, for which the labeling can be performed accurately automatically [1] and verified beforehand.

Our proposed methodology searches for the global optimum of the registration cost function by starting the gradient descent from different points, offset by a full vertebra, to avoid local minima. The resulting registration is used to determine the correspondence between the vertebrae seen on the CBCT (with labels to be verified) and the vertebrae on the prior scan, with reliable labels. To further quantify the results of this investigation, we determine the distance between the global optimum of our registration and the original clinical registration. A mutual information registration cost function was used, with 32x32 joint histogram bins, and six degrees of freedom (rotations and translations) for the registration (ITK/Elastics). In contrast to the clinical registration process, where only a region around the target is used to calculate the cost function, the registration cost function is calculated over the entire overlap of the scans. This takes the surrounding anatomy into account, making the registration robust against off-by-one-vertebra errors.