Abstract

This study aims to create and assess machine learning models for predicting lymph node metastases following neoadjuvant treatment in advanced gastric cancer (AGC) using baseline and restaging computed tomography (CT). We evaluated CT images and pathological data from 158 patients with resected stomach cancer from two institutions in this retrospective analysis. Patients were eligible for inclusion if they had histologically proven gastric cancer. They had received neoadjuvant chemotherapy, with at least 15 lymph nodes removed. All patients received baseline and preoperative abdominal CT and had complete clinicopathological reports. They were divided into two cohorts: (a) the primary cohort (n = 125) for model creation and (b) the testing cohort (n = 33) for evaluating models' capacity to predict the existence of lymph node metastases. The diagnostic ability of the radiomics-model for lymph node metastasis was compared to traditional CT morphological diagnosis by radiologist. The radiomics model based on the baseline and preoperative CT images produced encouraging results in the training group (AUC 0.846) and testing cohort (AUC 0.843). In the training cohort, the sensitivity and specificity were 81.3% and 77.8%, respectively, whereas in the testing cohort, they were 84% and 75%. The diagnostic sensitivity and specificity of the radiologist were 70% and 42.2% (using baseline CT) and 46.3% and 62.2% (using preoperative CT). In particular, the specificity of radiomics model was higher than that of conventional CT in diagnosing N0 cases (no lymph node metastasis). The CT-based radiomics model could assess lymph node metastasis more accurately than traditional CT imaging in AGC patients following neoadjuvant chemotherapy.