ESTRO 2025 - Abstract Book

S2489

Physics - Autosegmentation

ESTRO 2025

Conclusion: Geometric data perturbation using DIR is a feasible approach to identify contour uncertainty in a clinical setting. Further research will investigate the impact of perturbation method on the estimate of uncertainty, whether this uncertainty relates to regions of lower quality contouring, and how the uncertainty compares to that with direct access to the model of the black box.

Keywords: Uncertainty, Black-box, Head and Neck

References: 1.

Huet-Dastarac M et al. Quantifying and visualising uncertainty in deep learning-based segmentation for radiation therapy treatment planning: What do radiation oncologists and therapists want?. Radiotherapy and Oncology. 2024 Dec 1;201:110545. 2. Nikolov S et al. Clinically applicable segmentation of head and neck anatomy for radiotherapy: deep learning algorithm development and validation study. Journal of medical Internet research. 2021 Jul 12;23(7):e26151. 3. Brouwer CL et al. 3D Variation in delineation of head and neck organs at risk. Radiation Oncology. 2012 Dec;7:1 10.

2707

Digital Poster Estimating rectal dose in prostate radiotherapy using AI segmentation on CBCT Banashree Kalita 1 , Aodh Mac Gairbhith 2 , John Jones 3 , Brendan McClean 4,2

1 Medical Physics, University College Dublin, Dublin, Ireland. 2 Medical Physics, St. Lukes Radiation Oncology Network at St. James Hospital, Dublin, Ireland. 3 Radiation Oncology, St. Lukes Radiation Oncology Network at St. James Hospital, Dublin, Ireland. 4 School of Physics, University College Dublin, Dublin, Ireland Purpose/Objective: Accurately estimating the dose delivered to the rectum during prostate radiotherapy can be time-consuming with manual contouring. While AI-based segmentation is becoming increasingly common for contouring on CT, its application to cone beam CT (CBCT) remains underexplored. This study aims to assess the feasibility of using an AI segmentation model developed for CT imaging to estimate rectal dose in prostate radiotherapy using CBCT images. The objectives are: i) Evaluate the accuracy of AI-generated rectal contours compared to manual contours on CBCT ii) Compare the estimated rectal dose with the planned dose Material/Methods: MVision AI GBS™ was used to generate rectal contours on CBCT images (Varian Truebeam) of 10 patients treated with 60Gy in 20 fractions to the prostate (70 CBCTs total). AI-generated contours were manually edited by an oncologist. Contour accuracy was evaluated using volume differences, Dice Similarity Coefficient (DSC), and dose metrics based on clinical dose-volume constraints (DVCs). The rectal dose to the CBCT contours was derived from the original dose distribution and the average dose metrics were used to estimate the delivered dose to the rectum. Results: The average DSC between AI and manual contours was 0.95 ± 0.03. AI-generated volumes were, on average, 7.9 cc (± 6.1) smaller than manual contours, with a statistically significant volume difference (p < 0.001, Wilcoxon signed rank test). The largest discrepancies were observed at the superior rectum boundary with the sigmoid. Notably, the smaller AI contour volumes will affect relative dose metrics, despite being outside the high-dose region. While the AI contours showed a 1.6% ± 2.8% higher average dose volume metric, this increase was not shown to be statistically significant (Figure 1). The largest patient-specific discrepancy observed was 6.5%, identified as both quantitatively and qualitatively an outlier.