ESTRO 2024 - Abstract Book

S3027

Physics - Autosegmentation

ESTRO 2024

[3]: Papaconstadopoulos, P., et al. (2021) An anomaly detector as a clinical decision support system for parotid gland delineations. PMB. https://doi.org/10.1088/1361-6560/abfbf5

732

Digital Poster

Added value of HyperSight CBCT: does it allow for better AI-based autosegmentation?

Judith H. Sluijter, Agustinus J.A.J. van de Schoot, Maartje de Jong, Martine S. van der Knaap-van Dongen, Britt Kunnen, Nienke D. Sijtsema, Joan J. Penninkhof, Kim C. de Vries, Steven F. Petit, Maarten L.P. Dirkx

Erasmus MC, Radiotherapy, Rotterdam, Netherlands

Purpose/Objective:

Online adaptive radiotherapy (ART) demands high efficiency since the patient is immobilized on the treatment couch during the prolonged workflow. Accurate autosegmentation reduces the need for manual correction of contours, which could save valuable time. The recently released HyperSight CBCT imaging solution (Varian Medical Systems) shows improved image quality compared to the conventional Ethos CBCT (Halcyon 3.1, Varian Medical Systems). In this study, we evaluated to what extent the improved image quality of the HyperSight CBCT translates into an increased autosegmentation performance in the pelvic region, requiring less manual contour adjustments, and thereby a faster correction time compared to the conventional Ethos CBCT.

Material/Methods:

Twenty consecutive prostate cancer patients undergoing non-adaptive external beam radiotherapy on an Ethos system (Varian Medical Systems) were included between April and October 2023, after providing informed consent (IRB protocol MEC-2022-0815). Patients with anatomical variations such as prostatectomy, transurethral prostate resection, extracapsular extension, or bladder diverticula, were excluded. On consecutive days, patient treatment was performed on an Ethos system equipped with either HyperSight CBCT or conventional Ethos CBCT imaging. One image pair per patient was included, resulting in a total of 40 CBCT images. All images were blinded and anonymized. Contours of the prostate, seminal vesicles, bladder, rectum and bowel were generated using AI-based autosegmentation in current clinical Ethos software (version 1.1M1, Varian Medical Systems). Three experienced annotators, two radiotherapy technologists and one technical physician, manually corrected these contours, if necessary, in two steps. First, the contours were manually corrected according to our clinical protocol for prostate online ART to evaluate the clinical correction time. According to this protocol, the entire prostate and seminal vesicles are corrected, whereas adjustments for the bladder, rectum, and bowel are limited to a 2 cm range around the prostate and seminal vesicles. The confidence in delineation of each organ was scored from 1 to 5, where 1 indicated very uncertain and 5 completely confident. Then, to evaluate autosegmentation accuracy in the entire pelvic region, the observers also manually corrected bladder, rectum and bowel contours outside the 2 cm range and saved the fully corrected structure set. Autosegmentation accuracy was assessed with the Dice Similarity Coefficient (DSC) and the 95th percentile Hausdorff distance (HD95), quantifying the agreement between the autosegmentations and the fully corrected contours. Per structure and CBCT image, the results for the three observers were averaged. The interobserver agreement was evaluated using DSC and HD95 by comparing the fully corrected contours across observers for each CBCT. We performed Wilcoxon signed-rank tests with a Bonferroni correction for multiple comparisons to assess the significance of differences in correction time, delineation confidence, autosegmentation accuracy, and interobserver agreement between HyperSight CBCT and conventional Ethos CBCT imaging.