2198 - Whether Planning Target or Gross Tumor Volume is the Key Zone for Developing a Precise Prediction Model for Overall Survival of NSCLC Patients

Purpose/Objective(s): Planning target volume (PTV) represents an expansion beyond gross tumor volume (GTV) to account for inter/intrafraction motion. Additionally, in stereotactic body radiation therapy (SBRT) technique the PTV, by default, includes the extension of microscopic disease outside of the GTV. There have been numerous initiatives to use radiomic features extracted from the target volumes for developing artificial intelligence (AI) based models for predicting various clinical outcomes including diagnosis, treatment outcomes, etc. However, it is unclear which target volume (GTV or PTV) is more effective for extracting radiomic radiomic features for these AI-based prediction modeling. This study aims to identify the most influential region of interest (ROI) for radiomic feature extraction in predicting overall survival (OS) after two years, following SBRT.

Materials/Methods: SBRT (50-60 Gy in 3-5 fractions) planning CTs form 155 non-small cell lung cancer (NSCLC patients were used for this study. A total of 47 radiomic features were extracted from planning CT within each tumor volume zone (GTV and PTV) using histogram analysis, gray level co-occurrence matrix, gray level run length matrix, and gray level size zone matrix techniques. Dimension of the data frame was reduced based on the Point-biserial correlation matrix and a total of 16 radiomic features were selected as most influential features to develop a compact OS prediction model. Finally, an ensemble model incorporating Support Vector Machine, Gradient Boosting, and Random Forest algorithms, was developed for the OS prediction. The designed model was trained using 70% of the data and the rest of the data was used for testing purpose.

Results: Median age of the patient cohort was 75 years (range 52-91 years); T-stage T1-T2b; 54% were male. The ensemble model with 16 radiomic features extracted from GTV achieved performance metrics of ROC-AUC: 0.71, sensitivity: 0.64, and specificity: 0.79. The model utilizing PTV based features yielded much lower performance, with ROC-AUC of 0.61, sensitivity of 0.73, and specificity of 0.50.

Conclusion: This study suggests that the developed prediction model with GTV zone-based radiomic features produced superior performance with a 14% improved ROC-AUC score, despite the PTV presents additional radiomic features. In future OS prediction models, GTV may be the preferred choice. An ongoing multi-institutional study including a larger cohort of patients will aim to validate the findings of the current study.