S1613
Physics - Autosegmentation
ESTRO 2026
Poland. 5 Centre for Experimental Cardiooncology, Medical University of Gdańsk, Gdańsk, Poland
Purpose/Objective: In the RTOG 0617 trial (NCT00533949) [1], escalating the dose to 74 Gy in stage III NSCLC resulted in shorter survival than the standard 60 Gy [2]. In this study, using custom made nnUNet model for autocontouring, we aimed to analyze the dose distribution in cardiac substructures and assess its potential impact on overall survival. Material/Methods: A custom nnU-Net model [3] delineated cardiac substructures (Table 1), and dosimetric data were extracted from the original plans using MATLAB (2024a). The effect of radiation dose on survival was assessed using a univariable Cox model. To test the stability of predictors and minimize the effect of collinearity, bootstrap resampling for elastic net (EN) and Random Survival Forest (RSF) techniques were applied. Variables that were most significant in those models were incorporated into the multivariable Cox model together with clinical data in accordance with a study by van der Pol et al. [4]. Analyses were performed using Python 3.9 and R 4.4.2. Statistical significance was set at p < 0.05, and for the RSF model, the significance threshold was set at 0.01. Results: In the univariable analysis, significant predictors of survival included CTCAE grade ≥3 (HR 1.66, 95% CI 1.18–2.35, p = 0.0037), mean dose to the left atrium (HR 1.01, 95% CI 1.00–1.02, p = 0.0046), dose to 60% of the right atrium volume (D60%) (HR 1.01, 95% CI 1.00– 1.02, p = 0.0053). In the EN model, independently in the RSF model adjusted for age, sex, and log- transformed PTV, mean left atrial dose consistently remained the most stable predictor of shorter survival. The multivariable Cox model including mean left atrial dose, age, sex, and log-transformed PTV identified mean left atrial dose as the only independent predictor of overall survival (HR 1.012; 95% CI 1.004– 1.020; p = 0.0029). The optimal cut-off point was determined to be 5.64 Gy using rank-based statistics in the training set (50%–50% split). Patients in the validation group receiving ≥ 5.64 Gy had significantly shorter survival compared to those with mean left atrial dose < 5.64 Gy (median survival 20.2 months vs 39.5 months, p = 0.0216) (Figure 1).
Figure 1. Comparison of manual labels (red) with model predictions (green) for two cases (single focal and multi-focal lesions). Magnified views (column D) highlight the high degree of agreement. Conclusion: While the model's segmentation performance requires manual refinement for final treatment planning, its high detection accuracy provides a robust foundation for semi-automatic workflows, enhancing planning robustness through consistent target localization and significantly accelerating the contouring process over manual methods. References: [1] Isensee F, Wald T, Ulrich C, Baumgartner M, Roy S, Maier-Hein K, et al. nnU-Net Revisited: A Call for Rigorous Validation in 3D Medical Image Segmentation 2024.[2] Hatamizadeh A, Nath V, Tang Y, Yang D, Roth HR, Xu D. Swin UNETR: Swin Transformers for Semantic Segmentation of Brain Tumors in MRI Images. Lecture Notes in Computer Science. 2022;12962 LNCS:272–84.[3] Huang Z, Wang H, Deng Z, Ye J, Su Y, Sun H, et al. STU-Net: Scalable and Transferable Medical Image Segmentation Models Empowered by Large-Scale Supervised Pre-training 2023. Keywords: vertebra metastasis, lytic metastasis segmentation Cardiac Substructures Dosimetry and Survival in Stage III Non-Small Cell Lung Cancer: Insights from Automated Analysis of RTOG 0617 Data Kasper J. Kuna 1,2 , Luuk H.G. van der Pol 2 , Konrad Stawiski 1,3 , Martin F. Fast 2 , Bartłomiej Tomasik 4,5 1 Department of Biostatistics and Translational Medicine, Medical University of Łódź, Łódź, Poland. 2 Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands. 3 Department of Radiation Oncology, Dana-Farber Cancer Institute, Boston, USA. 4 Department of Oncology and Radiotherapy, Medical University of Gdańsk, Gdańsk, Digital Poster 4885
Made with FlippingBook - Share PDF online