Prediction of T2/T3 Staging in Patients with Volume-Equivalent Esophageal Squamous Cell Carcinoma on the Basis of PET/CT Radiomics.

Objectives: To develop and externally validate PET/CT-based radiomic models for predicting tumor invasion depth (≤T2 vs. ≥T3) in patients with esophageal squamous cell carcinoma (ESCC) with volume-matched tumors.

Methods: Semiautomatic segmentation was performed on F-FDG PET images, and radiomic features were extracted from PET and coregistered CT scans. Feature reproducibility was evaluated with the intraclass correlation coefficient (ICC), which showed excellent agreement (ICC > 0.95). Propensity score matching (PSM) was applied to control for tumor volume and patient demographics. Dimensionality reduction was conducted using principal component analysis (PCA), followed by feature selection via LASSO and MRMR. Logistic regression (LR), linear discriminant analysis (LDA), support vector machine (SVM), and random forest (RF) models were constructed. Model performance was assessed on internal and independent external cohorts.

Results: In the internal validation, the PET and combined PET/CT radiomic models outperformed the CT-alone models, with the LDA and LR classifiers achieving area under the ROC curve (AUC) greater than 0.97. In the external validation, only the models based on PET features maintained good predictive performance (LR AUC 0.8438, accuracy 81.25 %; LDA AUC 0.8281, accuracy 87.5 %). Models built on CT or combined PET/CT features failed to produce valid results, defaulting to single-class predictions. PET feature-based models demonstrated stable generalizability across datasets.

Conclusions: Radiomic models based on PET features and LDA or LR classifiers can accurately predict tumor invasion depth in patients with volume-equivalent ESCC and show strong external generalizability. CT or combined feature models may not be reliable under stringent tumor volume constraints.