Predicting early recurrence in locally advanced gastric cancer after gastrectomy Using CT-based deep learning model: a multicenter study

Background: Early recurrence in patients with locally advanced gastric cancer (LAGC) portends aggressive biological characteristics and a dismal prognosis. Prediction of early recurrence may help determine treatment strategies for LAGC. To develop a deep learning model for early recurrence prediction (DLER) based on preoperative multiphase computed tomography (CT) images and further explore the underlying biological basis of the proposed model. Materials and methods: In this retrospective study, 620 LAGC patients from January 2015 to March 2023 were included in three medical centres and The Cancer Image Archive (TCIA). The DLER model was developed using DenseNet169 and multiphase 2.5D CT images, and then crucial clinical factors of early recurrence were integrated into the multilayer perceptron classifier (MLP) model (DLER MLP ). The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity and specificity were applied to measure the performance of different models. The log-rank test was used to analyze survival outcomes. The genetic analysis was performed using RNA-sequencing data from TCIA. Results: Using the MLP classifier combined with clinical factors, DLR MLP showed higher performance than DLER and clinical models in predicting early recurrence in internal validation set (AUC: 0.891 vs 0.797, 0.752), two external test set1 (0.814 vs. 0.666, 0.808) and external test2 (0.834 vs. 0.756, 0.766). Early recurrence-free survival, disease-free survival, and overall survival can be stratified using the DLER MLP (all P < .001). High DLER MLP score is associated with upregulated tumour proliferation pathways (WNT, MYC, and KRAS signalling) and immune cell infiltration in the tumour microenvironment. Conclusion: The DLER MLP based on CT images was able to predict early recurrence of patients with LAGC and served as a useful tool for optimizing treatment strategies and monitoring.