Deep learning radio-clinical signature for predicting neoadjuvant chemotherapy response and prognosis from pretreatment CT images of locally advanced gastric cancer patients

BACKGROUND: Early noninvasive screening of patients who would benefit from neoadjuvant chemotherapy (NCT) is essential for personalized treatment of locally advanced gastric cancer (LAGC). The aim of this study was to identify radio-clinical signatures from pretreatment oversampled computed tomography (CT) images to predict the response to NCT and prognosis of LAGC patients. METHODS: LAGC patients were retrospectively recruited from six hospitals from January 2008 to December 2021. An SE-ResNet50-based chemotherapy response prediction system was developed from pretreatment CT images preprocessed with an imaging oversampling method (i.e. DeepSMOTE). Then, the deep learning (DL) signature and clinic-based features were fed into the deep learning radio-clinical signature (DLCS). The predictive performance of the model was evaluated based on discrimination, calibration, and clinical usefulness. An additional model was built to predict overall survival (OS) and explore the survival benefit of the proposed DL signature and clinicopathological characteristics. RESULTS: A total of 1060 LAGC patients were recruited from six hospitals; the training cohort (TC) and internal validation cohort (IVC) patients were randomly selected from center I. An external validation cohort (EVC) of 265 patients from five other centers was also included. The DLCS exhibited excellent performance in predicting the response to NCT in the IVC [area under the curve (AUC), 0.86] and EVC (AUC, 0.82), with good calibration in all cohorts ( P >0.05). Moreover, the DLCS model outperformed the clinical model ( P <0.05). Additionally, we found that the DL signature could serve as an independent factor for prognosis [hazard ratio (HR), 0.828, P =0.004]. The concordance index (C-index), integrated area under the time-dependent ROC curve (iAUC), and integrated Brier score (IBS) for the OS model were 0.64, 1.24, and 0.71 in the test set. CONCLUSION: The authors proposed a DLCS model that combined imaging features with clinical risk factors to accurately predict tumor response and identify the risk of OS in LAGC patients prior to NCT, which can then be used to guide personalized treatment plans with the help of computerized tumor-level characterization.