Whole-esophageal radiomics and deep learning for predicting the pT and pN categories of resectable esophageal cancer

To investigate the application of whole-esophageal radiomics and deep learning features derived from enhanced arterial phase CT images along with clinical data to differentiate the pT and pN categories in patients with resectable esophageal cancer. Clinical and CT data of 625 patients from Center 1 and 56 patients from Center 2 with esophageal cancer were analyzed. The dataset of Center 1 was randomly divided into a training cohort and an internal testing cohort at an 8:2 ratio. The whole-esophageal ROIs were delineated semi-automatically. The least absolute shrinkage and selection operator algorithm was employed for radiomics and deep learning feature reduction and selection, leading to the development of Radscore and Deepscore. Clinical, radiomics, deep learning, and combined models were constructed. The diagnostic performance of the models was evaluated using the receiver operating characteristic (ROC) curve. In the training cohort for predicting pT categories, the AUC of the combined model was 0.901, which was significantly higher than that of the clinical (0.869), radiomics (0.833), and deep learning (0.830) models. For predicting pN categories, the AUC of the combined model was 0.800, significantly higher than that of the clinical (0.748), radiomics (0.748), and deep learning (0.764) models. Calibration curves also demonstrated good agreement between predicted and actual status, and the decision curve analysis further confirmed its clinical applicability. The combined model offers the most reliable method for predicting thepT and pN categories in patients with resectable esophageal cancer, thereby beneficial to treatment planning.