Integrating Tumor and Nodal Radiomics to Predict the Response to Neoadjuvant Chemotherapy and Recurrence Risk for Locally Advanced Gastric Cancer.

Abstract

To develop and evaluate machine learning models using tumor and nodal radiomics features for predicting the response to neoadjuvant chemotherapy (NAC) and recurrence risk in locally advanced gastric cancer (LAGC). Early and accurate response prediction is vital to stratify LAGC patients and select proper candidates for NAC. A total of 218 patients with LAGC undergoing NAC followed by gastrectomy were enrolled in our study and were randomly divided into a training cohort (n = 153) and a validation cohort (n = 65). We extracted 1316 radiomics features from the volume of interest of the primary lesion and maximal lymph node on venous phase CT images. We built 3 radiomics signatures for distinguishing good responders and poor responders based on tumor radiomics (TR), nodal radiomics (NR), and a combination of the two (TNR), respectively. A nomogram was then developed by integrating the radiomics signature and clinical factors. Kaplan- Meier survival curves were used to evaluate the prognostic value of the nomogram. The TNR signature achieved improved predictive value, with AUCs of 0.755 and 0.744 in the training and validation cohorts. Our proposed nomogram model (TNRN) showed a good performance for GR prediction in the prediction efficacy, calibration ability, and clinical benefit, with AUCs of 0.779 and 0.732 in the training and validation cohorts, superior to the clinical model. Moreover, the TNRN could accurately classify the patients into high-risk and low-risk groups in both training and validation cohorts with regard to postoperative recurrence and metastasis. The TNRN performed well in identifying good responders and provided valuable information for predicting progression-free survival time (PFS) in patients with LAGC who underwent NAC.