ESTRO 2024 - Abstract Book

S4977

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Material/Methods:

This retrospective multicentric cohort study included patients who received radiation therapy between 2015 and 2019 for locally advanced lung cancer. Patients could receive 3 dimensional conformational radiation therapy (3DCRT) or intensity modulated radiation therapy (IMRT) but stereotactic radiation therapy was not allowed. Patients with small cell lung cancer, with second synchronous cancer or prior thoracic radiation therapy were excluded. Patients who died within 6 months without developing G ≥ 2 RP were also excluded. Dosimetric features were extracted from the volume “lungs – planning target volume”. Baseline computed tomography scanners (CT scans, before chemotherapy or radiation therapy) were segmented to extract radiomic features from the "Lungs - Tumor" volume after image processing. All segmentations were generated semi-automatically using a deformable model and then manually corrected. A second review was performed by a senior radiologist.

The association between the various factors and the occurrence of G ≥ 2 RP was estimated by Wilcoxon-Mann-Whitney test for quantitative variables and Fisher exact test for qualitative variables.

Variables associated with the occurrence of G ≥ 2 RP in the descriptive analysis (p-value < 0.3) were then selected for explanatory and predictive analysis. For radiomics features, only the first order and the texture features were selected. An agglomerative hierarchical clustering (Ward) was performed and then one radiomics feature was selected per cluster. An explanatory analysis was performed by logistic regression with LASSO regularization, followed by predictive analysis with creation of predictive models by logistic regression with LASSO regularization or random forest using repeated stratified cross validation (5x5 folds) with Bayesian optimization (Figure 1).

Figure 1. Workflow

Results: