S1559
Physics - Autosegmentation
ESTRO 2026
surface distance (MSD) of 0.37 mm. Blinded expert ratings were not significantly different for manual GTVs vs. prompted GTV auto-segmentations (Mean, 4.2 vs. 4.2, paired p = 0.823).
reconstruction resolution (512x512 vs.1024x1024) on the volumetric and geometric agreement of a deep learning auto-contouring model across a diverse set of thoracic organs-at-risk (OARs). Material/Methods: Ten scans from a photon-counting CT scanner with settings of 120 kVp, 500 mm FOV, and 2 mm slice thickness were reconstructed using standard (SR, 512x512) and ultra-high resolution (UHR, 1024x1024) algorithms. Sixteen OARs were auto-contoured using an AI-based software on both SR and UHR images. The OARs were purposely sampled to represent major anatomical categories. The primary endpoint was the contoured volume of each OAR, compared using a paired, two-tailed t-test. Geometric consistency was quantified using the Dice Similarity Coefficient (DSC), Mean Distance to Agreement (MDA), and 95th percentile Hausdorff Distance (HD95). Results: The AI model's output demonstrated significant, organ-specific resolution-dependence. UHR reconstruction led to statistically significant (p<0.05) volumetric differences for 8 of 16 OARs. Geometric analysis revealed a clear hierarchy of AI performance: Large, homogeneous structures like the Lungs and Liver exhibited excellent overlap and boundary agreement (Mean DICE >0.995, MDA <0.126mm). In contrast, small or complex structures showed markedly poorer agreement. The left anterior descending artery (LAD) (DICE=0.926 ± 0.038, MDA=0.245 ± 0.258mm) and the proximal bronchial tree (PBT) (DICE=0.942 ± 0.106, MDA=1.365 ± 3.830mm) were most affected. Critically, the HD95 metric also revealed severe localized discrepancies for these challenging OARs, with maximum values reaching 10.57mm for the LAD and 78.3mm for the PBT, indicating that the two contour sets could be profoundly misaligned at specific anatomical points. Conclusion: This intra-algorithm analysis conclusively demonstrates that the impact of UHR input is highly organ-specific, and significantly altering the AI model's output for complex or small structures. These volumetric and geometric discrepancies reveal that the model's performance is intrinsically tied to the reconstruction protocol, not universally robust. Thus, integrating UHR-PCCT into the auto contouring workflow does not eliminate uncertainty but shifts it, necessitating enhanced professional oversight. Clinical implementation must be guided by this understanding, ensuring that expert judgment remains the final safeguard for contouring accuracy and patient safety. Keywords: PCCT, Ultra-high Resolution, Auto Contouring
Conclusion: Promptable auto-segmentation models can
substantially increase GTV contouring efficiency, while achieving segmentation quality similar to manual GTV delineation in CT and MRI for multiple treatment sites. These findings demonstrate the potential of this novel model class to improve treatment planning workflows and warrant their systematic evaluation across further imaging modalities and anatomic regions. Integration of promptable auto-segmentation models, like nnInteractive, into treatment planning systems should be considered.
References: Isensee, Fabian, et al. "nninteractive: Redefining 3d promptable segmentation." arXiv preprint arXiv:2503.08373 (2025). Keywords: AI, GTV, Interactive Segmentation Digital Poster 1881 Analysis of resolution impact on AI contouring in photon counting CT Sze Ting Wong, Wing Ki Fung, Ka Fai Cheng, George Chiu Department of Radiotherapy, Hong Kong Sanatorium & Hospital, Hong Kong, China Purpose/Objective: To perform a reference-free, intra-algorithm comparison assessing the exclusive effect of image
Made with FlippingBook - Share PDF online