S1582
Physics - Autosegmentation
ESTRO 2026
vendors: RaySearch RayStation (v16.0.0.847), Manteia Mozi (V4.04.3), United Imaging uRT-TPOIS (V R002), Radformation Limbus (v1.8.1), and MediqRT Synaptiq (V1.4.0).3D mean surface distance (3D-MSD), surface Dice similarity coefficient (sDSC) and normalised added path length (nAPL) at 1.0 and 2.0 mm were computed to assess global differences in performance. Furthermore, 3D-LSDMs were used for interactive visualization of local mismatch regions for clinically relevant OOIs. All computations were performed using Inpictura AIQUALIS (v1.1). Non-parametric ANOVA was performed using Kruskal–Wallis test. Pairwise tests were performed using Wilcoxon rank-sum test with Bonferroni correction.Illustrative results are reported for the brainstem and the mandible as depicting informative differences between the different vendors, and having contrasting appearance on the CT. Results: The median values for 3D-MSD for all vendors were in [0.8, 1.9] mm and in [1.2, 1.9] for the mandible and brainstem respectively. The 3D-MSD ANOVA did not show any significant difference in performance (p- value in {0.07, 0.11}). For the brainstem, both the sDSC and nAPL ANOVAs did not show any significant differences for both tolerances (p-values>0.08). For the mandible, it was also the case at 1.0 mm (p- value>0.06. At 2.0 mm however both measures indicated a significant difference (p-value <0.015). Figure 1 shows the mandible measures’ distributions. Figure 2 depicts the 3D-LSDMs for both OOIs indicating differences in regions where corrections are required between the different vendors. Notable variations existed at the base of the brainstem and on the condyles of the mandible.
Conclusion: The quality of the constructed AI-CTVs was inferior to that of the MC-CTVs. Therefore, the auto-segmentation methods require further development before implementation in a clinical setting. We are currently collaborating with MVision to develop an improved, automated version of the AI-CTV, adapted to the contouring guidelines used in our Department. The agreement of the quantitative metrics for the OOIs was overall good, except for the bowel bag. Keywords: rectal cancer, autosegmentation, CTV
Digital Poster 3427
Beyond Dice: 3D Local Surface Distance Maps to Assess the Clinical Reliability of AI Contouring in Head and Neck Radiotherapy Daniel NGUYEN 1 , Sena YOSSI 2 , Djamal BOUKERROUI 3 ,
Jad FARAH 4 , Cristina SPOREA 1 , Oleksandr OGORODNITCHOUK 2 , Mustapha KHODRI 1
1 Physicics, ORLAM, Lyon, France. 2 Radiotherapy, ORLAM, Lyon, France. 3 Inpictura Limited, Inpictura, Abingdon, United Kingdom. 4 Vision rt, Dove House, N3 2JU, United Kingdom Purpose/Objective: The increasing clinical use of AI-based auto-contouring tools requires robust and interpretable assessment methods to ensure patient safety. This study aimed to investigate the use of 3D Local Surface Distance Maps (3D-LSDM) [1,2] to understand spatial discrepancies between clinical and AI-generated contours for head and neck (H&N) Organs-of-interest (OOI) when comparing commercial solutions. Material/Methods: Nineteen OOIs from thirty H&N cancer patients were retrospectively analyzed. OOIs were manually delineated by one expert (Varian Eclipse, V18.0.1.261) and compared with auto-segmentations from five
Made with FlippingBook - Share PDF online