Radiomics-based ensemble model predicts postoperative recurrence of gastric cancer.

backgroundEarly postoperative recurrence in locally advanced gastric cancer (LAGC) severely compromises patient outcomes, yet current predictive models are inadequate due to limited generalizability and insufficient validation. This study aimed to establish and robustly validate a multimodal Radiomics-Clinical Integrated Risk Stratification Assessment (RSA) model to accurately predict early recurrence.

methodsWe retrospectively analyzed 2,516 LAGC patients from six hospitals across northern and southern China, divided into training, internal validation, and external validation cohorts. Radiomics features were extracted from portal venous-phase abdominal CT images using nnU-Net-based segmentation, followed by a hierarchical feature-selection framework integrating the mRMR algorithm and LASSO regression. A predictive RSA model combining clinical factors and radiomics features was developed. Model robustness was validated internally, externally, prospectively (clinical trial cohort, NCT01516944, n = 569), and further independently tested using a publicly available dataset from The Cancer Imaging Archive (TCIA, n = 41).

resultsThe RSA model exhibited superior performance across all validation cohorts, achieving AUC values of 0.873 (training), 0.871-0.872 (internal validation), 0.870-0.873 (external validation), 0.857 (prospective validation), and 0.850 (TCIA validation). Decision curve analyses confirmed the RSA model provided significant clinical benefits beyond traditional models. Transcriptomic analyses revealed that low-risk patients exhibited enhanced immune infiltration and significant activation of immune-related pathways, including IL6/JAK/STAT3 and interferon signaling. Multivariate Cox regression demonstrated the RSA model independently predicted five-year overall survival across validation cohorts (HR range: 1.830-2.166, all P < 0.001), surpassing standard clinical staging.

conclusionsOur RSA model, integrating robust radiomics methodologies and critical clinical parameters, consistently demonstrated high accuracy and clinical applicability in diverse populations, significantly improving prediction of early postoperative recurrence in LAGC. This approach provides novel insights into tumor-immune microenvironment interactions, paving the way toward personalized postoperative management strategies.