Integration of CT radiomics and machine learning for preoperative T staging of esophageal squamous cell carcinoma

Accurate preoperative staging is crucial for esophageal cancer treatment decisions. Traditional clinical T staging via imaging falls short of pathological precision. This study aimed to assess computed tomography (CT) radiomics' accuracy in predicting T stage for esophageal squamous cell carcinoma (ESCC) using machine learning and identify key stable features influencing T staging.

A total of 444 patients with ESCC were retrospectively enrolled. Segmented tumor 3D regions of interest (ROIs) from CT scans were analyzed for radiomics features, with key features selected via Spearman correlation and the least absolute shrinkage and selection operator (LASSO). Five machine learning algorithms including support vector machine (SVM), Gaussian Naive Bayes (NB), random forest (RF), logistic regression (LR), and extreme gradient boosting (XGB) were used to build classification models. Data were randomly divided into a training set (n=355, ~80%) and a testing set (n=89, ~20%). Model performance was evaluated by area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.

A total of 1,223 radiomics features were extracted based on the 3D ROIs of ESCC lesions. Feature selection yielded 14 key features for two-class classification (T1-2 vs. T3-4a) and 18 for three-class classification (T1 vs. T2 vs. T3-4a), with 11 features being consistent across both. Notably, RunEntropy and SurfaceVolumeRatio had high feature coefficients. The best AUC for two-class classification (XGB model) was 0.798 with 95% confidence interval (CI) of 0.701-0.895. The SVM model achieved a mean AUC of 0.740 in three-class classification.

The integration of CT radiomics and machine learning provides a valuable, non-invasive tool for preoperative T staging of ESCC. Specifically, the study demonstrates robust performance in two-class classification and provides preliminary reference for the exploration of three-class classification. Additionally, the features RunEntropy and SurfaceVolumeRatio emerge as key stable radiomic indicators for T staging among resectable ESCC patients.