ESTRO 2025 - Abstract Book

S1371

Clinical - Lung

ESTRO 2025

Mediastinum nodal involvement was common (N2/N3=70.62%). Median survival was 20 months (95%IC 12.2-26.9). Prediction AUCs were 0.6 for clinical based only model, 0.70 for imaging based model only, 0.75 for multimodal model including clinical and imaging data (see figure 1)

Conclusion: The proposed multimodal method (optimized late fusion ensamble) outperformed conventional unimodal models, bringing significant increase in performance and finding the optimal combination of multimodal models. Multimodal model was able to predict the OS in stage III NSCLC patients population treated with chemoradiation. Further studies are needed to improve accuracy of the model providing more complementary clinical information and adding an external validation.

Keywords: stage III NSCLC, AI prediction, chemoradiation

2708

Digital Poster Detection of pulmonary cancer specific features during radiotherapy using cone-beam CT radiomics Takuya Taniguchi 1,2 , Kosei Adachi 1,2 , Shuto Nakaya 1 , Kaede Kawaguchi 1 , Osamu Tanaka 1 , Masayuki Matsuo 2 1 Radiation Oncology, Asahi University Hospital, Gifu, Japan. 2 Radiology, Gifu University, Gifu, Japan Purpose/Objective: Radiomics analysis is highly accurate in detecting the specific image features of tumors (1). Therefore, during radiotherapy, changes in CT texture features of esophageal cancer have been reported to correlate with radiotherapy response and survival rate (2). The purpose of this study was to detect treatment effects by radiomics analysis of tumor response during radiotherapy using cone-beam computed tomography (CBCT) images obtained during image-guided radiotherapy (IGRT) for primary lung cancer. Material/Methods: This study evaluated 6 cases of primary lung cancer who underwent IGRT with a prescribed dose of 60 Gy / 30 fx. Radiomics analysis was performed retrospectively using daily CBCT images acquired by IGRT. First, the tumor area was manually contoured by a medical physicist from the acquired CBCT images. Next, Radiomics analysis was performed within the delineated tumor to obtain 75 specific features of the image. The obtained features are analyzed to identify parameters that lead to the visualization of treatment effects. Results: Results of the CBCT radiomics analysis showed that in all cases, features related to tumor size, such as voxel count and surface area, decreased over time, while the Coarseness and Strength values of the CBCT texture features showed an increasing trend. Figure 1 shows changes during radiation therapy of lung cancer tumors. Figure 2 shows changes in tumor volume, coarseness and strength of CBCT texture features.