2301 - A Deep-Learning Segmentation Model for Radiotherapy of Thoracic Lymph Nodes in NSCLC: Precise, Efficient and Clinician-Driven
Presenter(s)
X. Zhang1, J. He2, H. Yang3, M. Yang3, and L. Wang4; 1Shandong First Medical University, Shandong Cancer Hospital and Institute, Jinan, China, 2Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong University, Jinan, China, 3United Imaging Research Institute of Intelligent Imaging, Beijing, 100094, China, Beijing, China, 4Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, Jinan, Shandong, China
Purpose/Objective(s): This study focuses on training and validating deep learning-based auto-segmentation (DLBAS) models for delineating clinical target volumes (CTVs) of thoracic lymph nodes in non-small cell lung cancer (NSCLC) to enhance segmentation support.
Materials/Methods: Contrast-enhanced CT scans from NSCLC patients were collected two weeks before surgery, between October 2018 and December 2021. Two skilled radiation oncologists manually outlined lymph node CTVs across 10 distinct levels (1-10) based on precise anatomical localization. We developed DLBAS models using nnUNet and assessed their performance quantitatively through metrics such as the Dice Similarity Coefficient (DSC) and the 95% Hausdorff Distance (95HD). Additionally, two other radiation oncologists evaluated the automated segmentations in the test sets for clinical applicability, while the time taken for both manual and automated segmentation was recorded.
Results: Our model was trained on data from 1013 patients with 1775 lymph node CTVs. The average DSC for all CTVs surpassed 0.7 (ranging from 0.71 to 0.86), and the mean 95HD ranged from 3.12 mm to 9.76 mm. Most automated segmentations received scores of =2 (indicating minimal or no corrections needed), representing 91.3% of all cases, with no structures rated as 0 (unusable). Furthermore, the DLBAS model significantly decreased the contouring time for individual lymph node CTVs from 267 seconds to 72 seconds compared to manual methods.
Conclusion: The DLBAS model demonstrates significant improvements in the efficiency and accuracy of lymph node CTV segmentation, making it suitable for clinical application and showing promise for CT-guided adaptive radiation therapy.