S1570
Physics - Autosegmentation
ESTRO 2026
predictive features in head and neck cancer (HNC) patients. Material/Methods: We retrospectively collected 61 HNC cases with clinically-approved ground-truth contours and treatment plans. Auto-segmentations of 15 OOIs were generated using an in-house nnU-Net model. Edit vectors from each auto-segmentation surface voxel towards the ground truth were clustered using DBSCAN [4] to identify local edit regions. These were then used to simulate realistic local auto-segmentation corrections. Simulated segmentations with a Dmean difference <0.01 Gy relative to the ground truth, were classified as zero difference. Next, spatial and dose features were extracted for each simulated contour alternative (OOI-type, target location, Dmean original contour, mean distance to target and edit volume). Non-collinear features (Variance Inflation Factor (VIF) <5) were kept. Group differences between zero and non-zero edits were tested for statistical significance (Mann-Whitney U test).A gradient boosting classifier was trained to distinguish contour alternatives resulting in zero versus non-zero Dmean differences between original and simulated contours, using an 80/20 patient-disjoint training/test split. Model performance was evaluated using ROC-AUC and a precision- recall curve. Calibration was assessed with the expected calibration error (ECE). Results: Edit vector clustering provided 6902 contour alternatives (median: 4 alternatives per contour). These edits had a median volume of 106 voxels (range: 8-9272 voxels). Of the 6902 contour alternatives, 16.4% (N=1131) resulted in zero Dmean difference. All numeric features showed low multicollinearity (VIF<1.2) and significant differences between the zero and non-zero group (p<0.0001), see Table 1. The gradient boosting model achieved strong predictive performance (ROC-AUC=0.93, average precision=0.89, ECE=0.04), see Figure 2.
Conclusion: A gradient boosting classification model using original dose and spatial features can effectively predict whether a local edit to an auto-segmentation affects OOI Dmean. This capability could support edit prioritisation and serve as a flag for local contour adjustment during re-planning. References: [1] Vaassen, Femke, et al. "Geometric and dosimetric analysis of CT-and MR-based automatic contouring for the EPTN contouring atlas in neuro-oncology." Physica Medica 114 (2023): 103156. [2] González, Patrick J., et al. "Explaining the dosimetric impact of contouring errors in head and neck radiotherapy." Biomedical Physics & Engineering Express 8.5 (2022): 055001.[3] Kieselmann, Jennifer Petra, et al. "Geometric and dosimetric evaluations of atlas-based segmentation methods of MR images in the head and neck region." Physics in Medicine & Biology 63.14 (2018): 145007.[4] Ester, Martin, et al. "A density-based algorithm for discovering clusters in large spatial databases with noise." kdd. Vol. 96. No. 34. 1996. Keywords: Autosegmentation, edit prioritisation, dose impact
Made with FlippingBook - Share PDF online