S1584
Physics - Autosegmentation
ESTRO 2026
a model CTV. Each parameter quantifies the extent of disease in a specific brain structure. Parameter m defines the overall tumor extent that requires treatment. Parameters 𝜌 brainstem and 𝜌 optic modulate the CTV margins in the brainstem and optic structures, accounting for the reduced likelihood of tumor infiltration or preferential sparing. Parameters dventricle and dfalx increase the CTV margins at the ventricles and falx cerebri, accounting for positional uncertainties of those structures over the course of treatment. Except for the gross tumor volume (GTV), all brain structures were auto-segmented using an in- house developed CT segmentation model. The parameters were fitted across 93 glioma cases (80/20 train/test), collected retrospectively from three physicians at our institution, by matching the model CTV to the manually-delineated CTVs. The clinical quality of the model CTV was assessed by two neuro- radiation oncologists who blindly rated 40 manual versus model CTVs on a scale of 0–3 (0: unusable, 1: major edits, 2: minor edits, 3: acceptable). Results: Figure 1 shows how fitting parameters m, 𝜌 brainstem, and dventricle can accurately replicate the model CTV to the manual CTV in the brain, brainstem and ventricles, respectively. Figure 2 shows the similarity of the model CTV versus the manual CTVs for the training versus test set. The median Dice, 95% Hausdorff and surface DSC for the training set was >0.95, <4.1 mm, >0.83, and for the test set >0.93, <6.4 mm, >0.71 Physician quality assessments showed no statistically significant difference in quality ratings between model-generated and manually drawn CTVs (p > 0.05), with both manual and model CTVs achieving an average quality rating of 2.3.
Conclusion: This study highlights organ-specific differences in contouring accuracy and dosimetric impact when comparing AI- and DIR-derived contours on CBCT images for adaptive radiotherapy. AI-generated CBCT contours demonstrate high geometric and clinical accuracy with reduced need for expert editing compared to DIR, supporting their integration into prostate ART workflows. However, observed dosimetric discrepancies, particularly for rectum and high-dose bladder regions, highlight the necessity of expert review prior to clinical use. The findings support AI-based contouring as a viable method to streamline adaptive workflows and reduce clinical workload while maintaining safety and treatment quality. Keywords: AI Contours, Deformable Image Registration, ART Proffered Paper 3500 Parametric delineation: introducing a new paradigm for clinical target volume delineation of glioma Gregory Buti 1 , Ali Ajdari 1 , Christopher P Bridge 2 , Ariel E Marciscano 1 , Gregory C Sharp 1 , Fredrik Löfman 3 , Helen A Shih 1 , Thomas R Bortfeld 1 1 Radiation Oncology, Mass General Brigham, Boston, USA. 2 Radiology, Mass General Brigham, Charlestown, USA. 3 Machine Learning, RaySearch Laboratories AB, Stockholm, Sweden Purpose/Objective: Despite the emergence of AI-contouring methods, delineating the clinical target volume (CTV) remains challenging. Variability exists given physicians' clinical experience, institutional standards, and the guidelines being followed. We propose a new approach to delineating the CTV that replaces manual edits with physics-based models that quantify tumor extent and its uncertainties. We analyze the generalizability of the parametric method based on its ability to replicate different physician-delineated CTVs within our institution. Material/Methods: Based on in-house expert knowledge and international consensus guidelines, we propose five delineation parameters for glioma tumors that define
Made with FlippingBook - Share PDF online