SXX
Corrigendum
ESTRO 2026
square tests. Statistical significance was set at p<0.05. Results: The median follow-up was 31 months (range:7–54), and the median treatment interval was 12 months (range:6–40). Patient, tumor, and treatment characteristics are summarized in Table-1. Fifteen patients (94%) initially received SBRT (50–60 Gy/4–8 fx), and one underwent hypofractionated RT (60 Gy/15 fx). The median EQD2 dose ( α / β =8) of the first RT was 93 Gy (range:72–102.5). Re-irradiation was performed for out-of-field LRR: 10 patients (62%) received definitive chemoradiotherapy, 5 (31%) SBRT, and 1 (6%) hypofractionated RT. The median re-RT EQD2 dose was 60 Gy (range:54–138) with a median fraction size of 2 Gy (range:2–15). All cases showed ≥ 50% isodose overlap, and 9 patients (56%) had direct PTV overlap. During follow-up, in-field recurrence occurred in 3 patients (18.8%) with a median LRR-free interval of 11 months (SE±2.19). The 1-, 2-, and 3-year overall survival rates were 87.5%, 74.5%, and 45%; corresponding cancer-specific survival rates were 93%, 85%, and 75%. No significant associations were observed between survival and histology, gender, comorbidity, or re-irradiation modality. Although higher cumulative BED ₁₀ showed a non-significant trend toward improved OS (p=0.068). Acute AE included grade 1 in 3 patients (18.8%) and grade 2 in 2 (12.5%). Late AE consisted of grade 1 pneumonitis in one and grade 3 in another, with no ≥ grade 4 events.
with advanced image-guided techniques after prior definitive RT achieved favorable local control and survival with limited AE. These findings support its feasibility and safety in selected early-stage cases with careful dose planning and modern delivery approaches. Keywords: re-irradiation, early stage lung cancer Category: Physics: Autosegmentation Digital Poster 3697 Detecting Automation Bias in Radiotherapy Contouring: The Role of Similarity Measures Mark J Gooding 1,2 , Djamal Boukerroui 1 1 -, Inpictura Ltd, Abingdon, United Kingdom. 2 Division of Cancer Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Manchester, United Kingdom
Purpose/Objective: Recent publications [1-4] have explored using
continuous monitoring of editing of AI auto-contouring using statistical process control (SPC) as a method of detecting automation bias. While some studies have reported indications of automation bias [1-3] in their clinical workflow, others have not [4]. The objective of this study was to understand the impact that the choice of contour similarity measure(s) has in detecting changes in the clinical process.
Material/Methods: Starting from a public dataset of 30 Head and Neck cases with ‘ground truth’ contours [5], 350 synthetic
cases were generated using deformable augmentation. Subsequently, these were
automatically contoured (Mediq, Synapiq, Romania). Clinical editing was simulated by correcting the auto- contours, slice-by-slice, using the ‘ground truth’ where the deviation between the two was outside a predefined ‘acceptable’ tolerance. This tolerance was kept constant for the first 50 cases. Automation bias was simulated thereafter by linearly relaxing this tolerance using two approaches: one according to the local Euclidian distance (LED) between the contours, and one following the slice-wise intersection-over- union (IOU). Figure 1 illustrates this process. SPC (moving average window: 31 cases) was used to determine when bias would be detected at 99% confidence for seven measures: Mean Surface Distance (3D-MSD), 3D 95% Hausdorff distance (3D- 95%-HD), Normalized Added Path Length (NAPL), Surface Dice Similarity Coefficient (SDSC), Percentage
Conclusion: In this predominantly elderly and comorbid cohort with early-stage lung cancer, re-irradiation delivered
Made with FlippingBook - Share PDF online