S837
Clinical - Lung
ESTRO 2026
Purpose/Objective: New organs at risk (OARs) emerge due to monitoring of new side effects in a constantly evolving radiotherapy (RT) field. Dose-guidance of such OARs requires rapid, robust delineation. This study examines the use of AI auto-segmentation to define emerging organs and retrospectively assess their potential as new OARs. Material/Methods: Planning-CTs (pCTs) were identified from breast/esophagus/lung tumor sites to promote segmentation robustness in which organ-specific guidelines were identified for three recently proposed and outcome-associated organs: base of the heart (HeartBase), normal mediastinal lymph nodes (LNNormal), and thymus. These organs were delineated by four radiation oncologists, and the manual delineations (HeartBase/LNNormal/Thymus: N=30/5/46) were used to train three distinct models using in-house AI segmentation software. Following training, additional pCTs of patients treated for various thoracic tumors (HeartBase/Thymus: N=8/23; LNNormal model retraining pending) were used to validate the model through physician review and editing, which were incorporated into models’ weights. The final algorithm version was invoked on the pCTs of 240 patients treated with concurrent-chemotherapy and intensity-modulated RT to a median of 63 Gy (1.8- 2.0 Gy/fraction) for locally advanced non-small cell lung cancer (LA-NSCLC) in 2004-2014. For each auto- segmented organ, the mean dose was extracted, and organ-specific correlations between the mean dose and overall survival (OS) were examined using Cox proportional hazard regression with bootstrap resampling (significance: p ≤ 0.05). Results: The median OS time among the 240 LA-NSCLC patients was 1.8 (range: 0.8-3.7) years. The population median (interquartile range, IQR) of the mean dose to the auto-segmentations was 22 (12-32) Gy for HeartBase, 49 (41-55) Gy for LNNormal, and 46 (38-53) Gy for Thymus. While higher mean doses to each of these three OARs were numerically associated with worse OS (Hazard Ratios: 1.00-1.01), none was
inspection of boxplots indicates a lower percentage change in primary tumor volume and metabolic parameters (SUVpeak, SUVmax, and SUVmean), with possible prognostic potential for a higher risk of loco- regional failure, see figure 1. Conclusion: For the 93 patients treated with curative intent for LD- SCLC we found a loco-regional failure rate of 38% (LRf and LR+Df) in the follow-up time. No certain prognostic value of initial radiologic and metabolic tumor responses was found, but there were indications of prognostic potential of change in tumor volume and SUV uptake.
Keywords: LD-SCLC, radiologic response, metabolic response
statistically significantly predicting OS (HeartBase/LNNormalThymus: p-values: 0.12/0.50/0.38). Conclusion:
Digital Poster 4229
Emerging thoracic AI auto-segmented normal tissues and their potential role as new organs at risk (OARs) in lung cancer radiotherapy (RT) Maria Thor 1 , Shu Xing 1 , Sharif Elguindi 1 , Aditya Apte 1 , Jeff Meng 2 , Narek Shaverdian 2 , Daphna Gelblum 2 , Jonas Willmann 3 , Joseph O Deasy 1 1 Medical Physics, Memorial Sloan Kettering Cancer Center, NYC, USA. 2 Radiation Oncology, Memorial Sloan Kettering Cancer Center, NYC, USA. 3 Radiation Oncology, University Hospital Zurich, Zurich, Switzerland
Utilizing identified anatomical guidelines and manual delineations, this study has demonstrated that an auto-segmentation algorithm can be successfully trained for three emerging thoracic OARs that are typically not contoured routinely. Continuous model segmentation refinements, and segmentation of these three organs in contemporarily treated patients across thoracic tumor sites will aid untangling an existing relationship between the HeartBase, LNNormal, and
Made with FlippingBook - Share PDF online