Deep learning predicts resistance to neoadjuvant chemotherapy for locally advanced gastric cancer: a multicenter study

Background Accurate pre-treatment prediction of neoadjuvant chemotherapy (NACT) resistance in patients with locally advanced gastric cancer (LAGC) is essential for timely surgeries and optimized treatments. We aim to evaluate the effectiveness of deep learning (DL) on computed tomography (CT) images in predicting NACT resistance in LAGC patients. Methods A total of 633 LAGC patients receiving NACT from three hospitals were included in this retrospective study. The training and internal validation cohorts were randomly selected from center 1, comprising 242 and 104 patients, respectively. The external validation cohort 1 comprised 128 patients from center 2, and the external validation cohort 2 comprised 159 patients from center 3. First, a DL model was developed using ResNet-50 to predict NACT resistance in LAGC patients, and the gradient-weighted class activation mapping (Grad-CAM) was assessed for visualization. Then, an integrated model was constructed by combing the DL signature and clinical characteristics. Finally, the performance was tested in internal and external validation cohorts using area under the receiver operating characteristic (ROC) curves (AUC). Results The DL model achieved AUCs of 0.808 (95% CI 0.724-0.893), 0.755 (95% CI 0.660-0.850), and 0.752 (95% CI 0.678-0.825) in validation cohorts, respectively, which were higher than those of the clinical model. Furthermore, the integrated model performed significantly better than the clinical model (P < 0.05). Conclusions A CT-based model using DL showed promising performance for predicting NACT resistance in LAGC patients, which could provide valuable information in terms of individualized treatment.

Automated summary

This study developed a model using deep learning algorithm and computed tomography (CT) images to predict neoadjuvant chemotherapy (NACT) resistance in patients with locally advanced gastric cancer (LAGC). The model showed promising performance in both internal and external validation cohorts, outperforming the clinical model.