Background Treatment approaches for gastrointestinal stromal tumors (GISTs) of varying risk levels diverge. However, the accuracy of preoperative prediction of their risk level classification remains to be enhanced. This study is aimed at developing an integrated deep learning framework for preoperative risk stratification of GISTs using contrast‐enhanced CT. By integrating multiphase CT information and patient‐level aggregation strategies, we sought to improve the accuracy of preoperative risk stratification to support personalized treatment planning. Material and Methods A total of 229 GIST patients were retrospectively analyzed, and CT images from various phases (noncontrast, arterial, portal venous, and delayed) were processed using a 2.5D imaging technique. The deep learning model, referred to as the GIST network, incorporated transformer‐based feature extraction, multi‐instance learning (MIL), and ensemble methods for robust feature fusion. Various clinical characteristics such as tumor size and necrosis were also incorporated into the model. The model′s performance was validated using internal and external datasets. Results The ResNet34 architecture demonstrated strong predictive performance with AUC values of 0.881 in the training cohort, 0.864 in the internal validation cohort, and 0.804 in the external validation cohort. The MIL‐based approach achieved an AUC of 0.987 in the training cohort, outperforming the ensemble and transformer models. When combining clinical features with MIL, ensemble methods, and transformer‐based models, the hybrid model achieved an AUC of 0.998 in the training set and 0.963 in internal validation. The combined model′s performance was robust across all cohorts, with an AUC of 0.881 in the external validation test. Conclusion An integrated deep learning framework combining MIL, ensemble strategies, and transformer‐based feature fusion may provide a useful tool for preoperative GIST risk stratification. By incorporating both imaging and clinical features, the model enhances decision‐making in the clinical management of GISTs, offering significant improvements in predictive accuracy and generalizability across diverse patient populations.
Duan et al. (Thu,) studied this question.