S1604
Physics - Autosegmentation
ESTRO 2026
compare the performance of 8 different commercial DL-based AS solutions for H&N on CT images. Material/Methods: The study included 20 H&N cancer cases. Eight commercial DL-based AS solutions (MiradaDLCExpert (Mirada Medical), MVision (MVision AI), RayStation (Raysearch Laboratories), ART-Plan (Therapanacea), Syngo.Via (Siemens Healthineers), Limbus (Limbus AI), MIM (MIM Software Inc.), AutoContour (RADformation)) were used for delineation of OARs.The analysed OARs were brainstem, oral cavity, spinal canal, parotid glands, mandible, optic nerves, eyes, and submandibular glands. Autocontours were compared with ground truth contours delineated by 5 radiation oncologist experts calculating geometric metrics (DICE coefficient (vDSC), 95% Hausdorff distance (HD95%)). The dosimetric impact of geometric discrepancies was assessed by projecting the reference treatment plan, optimized using manual reference contours, onto the autocontours and comparing Dmean for paired OARs, oral cavity and D0.035cc for mandible, brainstem and spinal canal. Finally, a spearman correlation analysis was performed between geometric and dosimetric differences for each solution, and per OAR. Results: Figure 1 shows the results for the DSC metric.The mean vDSC values exceeded 0.8 across all structures and solutions, except for the optic nerves, oral cavity and parotids. Dose differences are shown in Figure 2. While mean dose differences (ΔDmean and ΔD0.035cc) due to geometric deviations were generally below 0.2 Gy for mandible, optic nerves and eyes, larger deviations were found for brainstem, spinal canal and oral cavity up to ΔD0.035cc 8.9Gy, 6.0 Gy, ΔDmean 9.1Gy, respectively. Only 9% of vDSC- dose datasets and 11% of HD95% - dose datasets showed a significant strong correlation (p < 0.05, 𝝆 > 0.6). Overall, structures closer to the target volumes exhibited stronger correlations, whereas distant structures (e.g., eyes, optic nerves, brainstem) showed weak correlations ( 𝝆 < 0.25).
Made with FlippingBook - Share PDF online