ESTRO 2025 - Abstract Book

S1345

Clinical - Lung

ESTRO 2025

Material/Methods: Five anonymised cases of locally advanced NSCLC (N2) with 2 to 5 involved nodal levels were analysed. An experienced consultant radiation oncologist manually contoured the ATS nodal levels on each patient's peak inspiration (PI), end expiration (EE), and maximum intensity projection (MIP) datasets to create internal target volume (ITV) structures (ground truth). After a time gap to mitigate recall bias, the same oncologist reviewed and amended the AI-generated nodal contours (MVision AITM). Both manual and AI-corrected contouring sessions were timed. Contour comparisons utilized metrics including Surface DICE (sDSC) at 2-3mm tolerances, Hausdorff distances, and average surface distance (AvSD) to assess consistency and potential clinical impact. Unedited AI generated nodal volumes were also compared to the ground truth. Results: The average per-nodal-level AI-corrected contouring reduced time was reduced by 40.6% or 6.8 minutes (EE: 64.7% saving or 3.8 minutes; PI: 26.1% saving or 1.5 minutes; MIP: 28.8% saving or 1.5 minutes). AI-corrected sDSC at 2 and 3 mm tolerances were 0.70, and 0.85, respectively, comparable to raw AI values of 0.72, and 0.86. Hausdorff distances improved from 7.6mm (raw AI) to 5.5mm (corrected AI), and AvSDs were reduced from 1.5mm to 1.2mm. These results demonstrate close spatial agreement between AI-corrected and manual contours, though significant outliers in unedited AI volumes highlight the need for human evaluation.