Objective To investigate the application of combined deep learning and radiomic features derived from enhanced arterial phase CT imaging with clinical data to differentiate between T2 and T3 staging in patients with esophageal cancer.

Methods A retrospective study was conducted using clinical and CT data from 388 patients with pathologically confirmed esophageal cancer treated at The First Affiliated Hospital of Wannan Medical College between May 2015 and April 2024. The dataset was randomly divided into a training set (271 cases) and validation set (117 cases) in a 7:3 ratio. Radiomic and deep learning features were extracted from enhanced arterial phase CT images. The least absolute shrinkage and selection operator algorithm was employed for feature reduction and selection, leading to the development of radiomic (Radscore) and deep learning (Deepscore) scores. Univariate and multivariate Logistic regression analyses were conducted to identify independent risk factors, and clinical, radiomic, deep learning, and combined models were constructed. A nomogram was generated for the combined model. The diagnostic performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC) and compared using the DeLong test. Clinical net benefit was assessed through decision curve analysis, and model calibration was evaluated using calibration curves.

Results Nine radiomicand 12 deep learning features were selected after dimensionality reduction. Multivariate Logistic regression identified tumor length, boundary, Radscore, and Deepscore as independent risk factors for distinguishing between T2 and T3 staging. In the training set, the AUC of the combined model was 0.867, which was significantly higher than that of the clinical (0.774, P<0.001), radiomic (0.795, P<0.001), and deep learning (0.821, P=0.001) models. In the validation set, the AUC of the combined model was 0.810, which was significantly higher than that of the clinical (0.653, P=0.002), radiomic (0.719, P=0.033), and deep learning (0.750, P=0.009) models. The decision curve analysis indicated that the combined model provided the highest clinical benefit in both datasets. The calibration curves demonstrated a good fit for both datasets (P=0.084, 0.053).

Conclusion The integration of deep learning and radiomic features obtained from enhanced arterial phase CT images with clinical data offers a reliable method for accurately distinguishing between preoperative T2 and T3 staging in esophageal cancer, thereby supporting clinical decision-making for treatment planning.