Visualized radio-clinical biomarker for predicting neoadjuvant chemotherapy response and prognosis from pretreatment oversampled CT images of LAGC patients: A multicenter study

Background: The early noninvasive screening of patients suitable for neoadjuvant chemotherapy (NCT) is essential for personalized treatment in locally advanced gastric cancer (LAGC). The aim of this study was to develop and visualized a radio-clinical biomarker from pretreatment oversampled CT images to predict the response and prognosis to NCT in LAGC patients.Methods: 1060 LAGC patients were retrospectively recruited from six hospitals from January 2008 to December 2021. The training (TC) and internal validation cohort (IVC) were randomly selected from center I. The external validation cohort (EVC) comprised 265 patients from 5 other centers. An SE-ResNet50-based chemotherapy response predicting system (DL signature) was developed from pretreatment CT images preprocessed with imaging oversampling method (i.e. DeepSMOTE). Then, DL signature and clinic-based features were fed into the deep learning radio-clinical signature (DLCS). The predictive performance was evaluated according to discrimination, calibration and clinical usefulness. Model for OS prediction were built to further explore the survival benefit of the proposed DL signatures and clinicopathological characteristic. Result: DLCS showed perfect performance in predicting the response to NCT in the IVC (AUC, 0.86) and EVC (AUC, 0.82), with good calibration in all cohorts (p > 0.05). In addition, the performance of DLCS was better than that of the clinical model (p＜0.05). Finally, we found that the DL signature could also serve as an independent factor for prognosis (HR, 0.828, p = 0.004). The C-index, iAUC, IBS for the OS model were 0.64, 1.24 and 0.71 in the test set.Conclusion: We proposed the DLCS that links the imaging features to clinical risk factors to generate high accuracy classification of tumor response and risk identification of OS in LAGC patients prior to NCT that then can be used for guiding personalized treatment plans with the help of the visualization of computerized tumor-level characterization.