A computed tomography-based deep learning radiomics for predicting the response to neoadjuvant chemotherapy combined with immunotherapy in patients with locally advanced esophageal cancer: a multicenter cohort study.

Background: Esophageal squamous cell carcinoma (ESCC) ranks sixth in global cancer mortality and is the main subtype in China; neoadjuvant chemoimmunotherapy for locally advanced ESCC has heterogeneous responses, and there is a lack of non-invasive tools to pretherapeutically identify major pathologic responders (MPRs), leading to unnecessary expenditures and adverse events. The present study seeks to develop a deep learning (DL)-based radiomic nomogram and to prospectively assess its clinical value in pretherapeutically identifying MPR in patients diagnosed with locally advanced ESCC who are scheduled to receive neoadjuvant chemoimmunotherapy. This approach facilitates the reduction of superfluous pharmaceutical expenditures and mitigates the risk of treatment-related adverse events, thereby significantly aiding in personalized therapeutic strategy formulation and prognostic evaluation. Methods: This study comprised 60 patients with a confirmed pathological diagnosis of ESCC. These participants were divided into a training set (n=42) and a testing set (n=18). From arterial-phase computed tomography (CT) images, radiomic features were obtained, while DL features were derived using a ResNet101-based network. Several machine learning classifiers-such as support vector machine, logistic regression, k-nearest neighbors, ExtraTrees, random forest, and XGBoost-were evaluated and compared. Classification performance was examined via receiver operating characteristic (ROC) curves and quantified by the area under the curve (AUC). An integrated model was subsequently developed by combining radiomics and clinical characteristics. The model's predictive ability was evaluated using ROC analysis, and its practical value was further investigated through decision curve analysis. Results: A total of 1,835 radiomics features and 2,048 DL features were extracted from the CT images. Through dimensionality reduction and feature selection, 8 radiomics features and 46 DL features were selected to form the deep learning radiomics (DLR). The combined DLR feature model demonstrated high predictive efficiency and robustness, with an AUC of 0.844 in the testing cohort. The predictive efficiency of different testing models was compared, and XGBoost showed superior predictive performance, achieving an AUC of 0.844 in the testing cohort. Finally, a nomogram was constructed by integrating the selected features with clinical baseline data, which exhibited the best discriminatory ability (AUC, testing cohort: 0.870). Conclusions: Our research successfully constructed and assessed a DLR nomogram for predicting treatment outcomes to neoadjuvant chemoimmunotherapy in individuals diagnosed with locally advanced esophageal carcinoma. This has the potential to promote personalized treatment and improve patient prognosis assessment, providing a non-invasive and effective method for clinical decision-making.