Abstract
212
Purpose: To investigate the usefulness of radiomics with machine learning using FDG-PET/CT in patients with rectal cancer. Materials and Methods: A total of 103 patients (male:female = 74:29, age: 65.4±12.1 years) with rectal cancer who had been treated with surgery (± neoadjuvant therapy) were retrospectively reviewed and allocated to training and test data sets (2:1 ratio). The volume of interest of the primary tumor was semi-automatically defined with a threshold of 40% of the maximum standardized uptake value, and radiomic features including global, local, and regional textural features were extracted. A random survival forest (RSF) model for predicting overall survival (OS) was trained with 1) radiomic features and 2) clinical profiles. The performance of RSF model was evaluated with Kaplan-Meier analysis with log rank test, and integrated area under the receiver operating characteristic curve (iAUC).
Results: The median follow-up of the patients was 943 days. The radiomic RSF model appropriately stratified patients from the test set into low-risk and high-risk groups of poor prognoses (log-rank p=0.007, hazard ratio: 10.2). The GLCM_ Dissimilarity was the most relevant radiomic feature by the variable-hunting algorithm of the RSF model. The RSF model of radiomics were successfully validated on the test set respectively (iAUC=0.67), and the combination of radiomics and clinical RSF model showed higher survival prediction (iAUC=0.78).
Conclusions: Radiomics with machine learning using FDG-PET/CT have a potential for predicting overall survival in rectal cancer, which diagnostic value can be increased by integration with clinical profiles.