S1583
Physics - Autosegmentation
ESTRO 2026
Figure 1: Mandible Boxplots for 3D-MSD, sDSC and nAPL at 2.0 mm tolerance.
become the state-of-the-art in radiotherapy planning, substantially reducing contouring time while maintaining accuracy comparable to expert contours. While AI contouring on planning CTs (CTp) is now well established, its extension to cone-beam CT (CBCT) for daily image guidance remains less explored. Adaptive workflows currently rely on manual contouring or deformable image registration (DIR), both limited by resource demands and geometric uncertainties. This study aimed to evaluate the accuracy and dosimetric implications of AI- and DIR derived CBCT contours for prostate radiotherapy, assessing their suitability for adaptive radiotherapy (ART) workflows. Material/Methods: CBCT datasets from 20 prostate cancer patients (total 140 images) treated with either 42.7Gy in 7 fractions or 60Gy in 20 fractions on Halcyon linear acceleratorswith Hypersight CBCT were retrospectively analyzed. AI contours were generated using Limbus AI v1.8.0, and DIR contours were propagated from planning CTs using Velocity v4.2. Two senior clinicians scored contour accuracy using a four-point Likert scale. Quantitative analysis of the prostate, bladder, and rectum included Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), Mean Surface Distance (MSD), center-of-mass (COM) displacement, and volume variation relative to CTp. Dosimetric evaluation assessed Dmin, Dmean, Dmax, and clinical organ-at- risk dose-volume metrics. Statistical comparisons employed paired t-tests and Wilcoxon signed-rank tests (p < 0.05). Results: AI-generated contours achieved acceptable clinical accuracy in >80% of cases, requiring fewer moderate or severe edits than DIR-derived contours (49% requiring minor corrections). Geometric metrics showed comparable DSC, HD, and MSD values between methods for prostate and rectum, though bladder contours showed larger COM shifts and volume variability in AI compared to DIR. Dosimetrically, prostate dose metrics exhibited significant differences between methods, rectum doses differed consistently, and bladder differences were largely non-significant except at high-dose volumes. Both methods effectively captured daily anatomical variation, suggesting complementary applicability depending on ART strategy.
Figure 2: 3D-LSDMs; Median maps (brainstem, mandible) 10-90% range (mandible) Conclusion: While global assessment measures can indicate differences between vendors, 3D-LSDMs offers a clinically meaningful, spatially resolved evaluation of AI contouring accuracy. This method enhances transparency and patient safety in clinical implementation by highlighting where contouring errors may occur and manual corrections will be needed in routine clinical implementation. References: [1] Brouwer CL, Boukerroui D, Oliveira J, Looney P, Steenbakkers RJHM, Langendijk JA, Both S, Gooding MJ. Assessment of manual adjustment performed in clinical practice following deep learning contouring for head and neck organs at risk in radiotherapy. Phys Imaging Radiat Oncol. 2020 Oct 14;16:54- 60.[2] Vaassen F, Boukerroui D, Looney P, Canters R, Verhoeven K, Peeters S, Lubken I, Mannens J, Gooding MJ, van Elmpt W. Real-world analysis of manual editing of deep learning contouring in the thorax region. Phys Imaging Radiat Oncol. 2022 May 14;22:104-110. Keywords: AI, Autosegmentation, Head&neck
Digital Poster 3450
Evaluation of AI- and Deformable Registration- Based Contours on Hypersight CBCT for Prostate Radiotherapy: Implications for Adaptive Planning. Mark Ashburner, Roger Huang, John Chin, Moamen Aly Radiation Oncology, Te Whatu Ora Health New Zealand, Waikato, New Zealand
Purpose/Objective: Deep learning (DL)-based auto-segmentation has
Made with FlippingBook - Share PDF online