S1616
Physics - Autosegmentation
ESTRO 2026
for detection and segmentation performance using lesion-wise metrics, including precision, recall, and Dice similarity coefficient. Results: The model achieved strong detection performance, with a mean precision of 0.814 (median 0.882), and mean recall of 0.960 (median 1.000). Across all image datasets, the model showed a lesion-wise false negative rate of 5.7%. The mean Dice score was 0.85, indicating accurate delineation of detected lesions. Figure 1 shows an example of model results.Figure 1. 3D representation of DL auto-segmentation model results (green) and GT (orange).
Digital Poster 5001 Deep learning-based auto-contouring of thoracic organs-at-risk and tumors for radiotherapy planning Hamdi Yalın Yalıç 1 , Levent Karacan 2 , Durmuş Etiz 3 , Kerem Duruer 3 , Alpay Levent 4 , Alaettin Uçan 5 , Ali Yaşar Yiğit 5 , Adem Ali Yılmaz 5 1 R&D, Caria Health, Ankara, Turkey. 2 Computer Engineering, Gaziantep University, Gaziantep, Turkey. 3 Radiation Oncology, Eskisehir Osmangazi University, Eskişehir, Turkey. 4 Radiotherapy Solutions, Prokal Medikal, İstanbul, Turkey. 5 R&D, TIGA Information Technologies, Ankara, Turkey Purpose/Objective: This study aims to develop and evaluate a deep learning–based auto-contouring model for accurate delineation of thoracic organs-at-risk (OARs) and gross tumor volumes (GTV) on CT images. By integrating a pre-trained MedSAM encoder with a 3D U-Net decoder, the model seeks to enhance geometric accuracy and contour consistency in radiotherapy planning. The objective is to reduce interobserver variability and manual workload while maintaining clinically acceptable segmentation performance across both typical and tumor-bearing lung structures. Material/Methods: This study was conducted on a dataset of 70 thoracic CT scans from clinical cases, manually annotated by experts, including tumor-bearing lung regions. All images were anonymized and approved for research use. The scans were preprocessed by resampling to an isotropic voxel size of 1 mm³ and by intensity clipping to the range [−1000, 400] HU to enhance lung tissue contrast. Each volume was normalized to zero mean and unit variance.The dataset was randomly divided into training (50 cases), validation (7 cases), and testing (13 cases) subsets. The proposed segmentation model employs a pre-trained MedSAM-based encoder [1] integrated with a 3D U-Net-style decoder [2]. The encoder captures multi-scale semantic and structural information from axial slices, while the decoder reconstructs the volumetric tumor mask using skip connections and feature fusion.Training was performed for 300 epochs using the AdamW optimizer (learning rate = 1×10 ⁻ ⁴, weight decay = 1×10 ⁻ ²). The MedSAM encoder was kept frozen during training. Model performance was evaluated using the Dice coefficient and the Hausdorff-95 distance. Results: The proposed model achieved an overall mean Dice score of 0.833 ± 0.054 and a mean HD95 of 5.69 ± 2.01 mm across all structures. Among OARs, the right lung and left lung obtained the highest Dice scores (0.977 ± 0.005 and 0.971 ± 0.007, respectively), while the esophagus showed the lowest (0.773 ± 0.048). The GTV
Conclusion: The evaluated DL auto-segmentation model demonstrated strong performance for intracranial metastases delineation on contrast-enhanced T1- weighted MRI. Its high sensitivity and contour precision suggest significant potential to streamline contouring workflows, and support stereotactic radiotherapy strategies in clinical practice. To more accurately evaluate its performance across different clinical environments and to assess real-world benefits, further analysis is required. References: Achrol, A.S., Rennert, R.C., Anders, C. et al. Brain metastases. Nat Rev Dis Primers5, 5 (2019). https://doi.org/10.1038/s41572-018-0055-yLamba, Nayan et al. “Epidemiology of brain metastases and leptomeningeal disease.” Neuro-oncology vol. 23,9 (2021): 1447-1456. doi:10.1093/neuonc/noab101Vogelbaum, Michael A et al. “Treatment for Brain Metastases: ASCO-SNO-ASTRO Guideline.” Journal of clinical oncology : official journal of the American Society of Clinical Oncology vol. 40,5 (2022): 492-516. doi:10.1200/JCO.21.02314Aizer, Ayal A et al. “Brain metastases: A Society for Neuro-Oncology (SNO) consensus review on current management and future directions.” Neuro-oncology vol. 24,10 (2022): 1613-1646. doi:10.1093/neuonc/noac118 Keywords: brain metastases, deep-learning, MRI
Made with FlippingBook - Share PDF online