ESTRO 2024 - Abstract Book

S4993

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Purpose/Objective:

Understanding the heterogeneity of lung cancer patients based on lung tissue and tumor characteristics remains a crucial aspect of personalized treatment. This study aims to pioneer an explainable artificial intelligence (AI) model for predicting overall survival (OS) by capturing the intricate, non-linear relationships within radiomics features derived from both lung parenchyma and gross tumor.

Material/Methods:

Locally-advanced non-small cell lung cancer (LA-NSCLC) patients received definitive radiotherapy in three countries were included in this study. A training set comprised 349 patients from China, while test-set-1 and test-set-2 included 229 patients from the Netherlands and 83 patients from the USA, respectively. Planning CT scans and clinical parameters were collected. Semi-automatic segmentations and radiomics feature extraction were performed on normal lung and gross tumor after image resampling. eXtreme Gradient Boosting (XGB) OS prediction models were developed individually for lung (Model-L) and tumor (Model-T), as well as an integrated lung and tumor model (Model LT). Additionally, a combined model incorporating lung, tumor, and clinical features (Model-CLT) was developed. Radiomics feature selection and model explanation were achieved through the Shapley additive interpretation (SHAP) method, highlighting the top 10 important features for radiomics models. A Cox proportional-hazards survival model served as the benchmark. The non-linear relationships and interaction effects between features were elucidated using dependence plots. The discrimination performance of each model was evaluated using the Harrell concordance index (C-index). Furthermore, the models' stratification ability was assessed by categorizing patients into high and low-risk groups based on the median risk score calculated from the model.