3062 - A Comparative Study of Automated and Manual Contouring in Radiomic Models for Lymph Node Assessment in Esophageal Squamous Cell Carcinoma

05:00pm - 06:00pm PT

Hall F

Screen: 21

Presenter(s)

Junyi He, BS - Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Science, Jinan, China, Jinan, Shandong

J. He¹, X. Zhang², H. Yang³, M. Yang³, and L. Wang⁴; ¹Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong University, Jinan, China, ²Shandong First Medical University, Shandong Cancer Hospital and Institute, Jinan, China, ³United Imaging Research Institute of Intelligent Imaging, Beijing, 100094, China, Beijing, China, ⁴Department of Radiation Oncology, Shandong Cancer Hospital Affiliated to Shandong University, Shandong Academy of Medical Science, Jinan, Shandong, China

Purpose/Objective(s):

To construct a non-invasive radiomic diagnostic model for identifying lymph node involvement in esophageal squamous cell carcinoma (ESCC) and compare the performance of automated versus manual contouring methods in optimizing the predictive accuracy of radiomic models.

Materials/Methods:

This study included patients with ESCC from October 2013 to December 2018. Clinical factors and contrast-enhanced CT data obtained within 2 weeks before surgery were collected. Two experienced clinicians manually contoured six lymph node drainage areas (105, 106, 107, 108, 109, and 110), while corresponding automated contours were generated using a deep learning-based segmentation model. The data were randomly divided into training and validation sets at a ratio of 7:3. Multiple deep learning architectures were evaluated, with optimal models selected for subsequent radiomic model construction. Combined models integrating radiomic and clinical predictors were developed separately for each contouring method. Model performance was evaluated primarily by the area under the receiver operating characteristic curve (AUC), with supplementary metrics including accuracy (ACC), sensitivity (SEN), and specificity (SPE). Delong tests statistically compared performance differences across clinical, radiomic (manual vs. automated), and combined models.

Results:

A total of 877 lymph node CTVs from 346 patients were included for model training. After feature selection, 842 radiomic features were retained. In the radiomic models, manual contouring achieved AUCs of 0.885 (training set) and 0.830 (validation set), while automated contouring yielded comparable AUCs of 0.881 and 0.819, respectively. Upon integrating clinical predictors with radiomic features, combined models utilizing manual contouring attained AUCs of 0.951 (training) and 0.907 (validation), whereas those based on automated contouring achieved comparable AUCs of 0.921 and 0.867, respectively.

Conclusion:

The radiomic-based diagnostic model demonstrated good predictive performance for identifying positive lymph node drainage areas. The combined clinical-radiomic model can be considered as a preferred model. The radiomic model based on automatic segmentation achieved results similar to that based on manual segmentation.