Interpretable machine learning model using dual-energy CT for predicting adverse histopathological status in gastric cancer: A multicenter study.

[OBJECTIVES] To investigate whether the integration of multiple dual-energy CT (DECT) quantitative parameters using machine learning algorithms could non-invasively decode the adverse histopathological status in gastric cancer (GC).

[METHODS] This observational study included patients with surgically resected GC who underwent DECT at six medical centers. Based on available pathological results, patients were categorized into three groups: T3/T4 stage, lymph node metastasis, and lymphovascular/perineural invasion. Each group included a development cohort (comprising training and test sets) and two external validation cohorts. For each group, five machine learning models were developed using 14 quantitative DECT parameters in the development cohort. The best model in each group was selected and evaluated. Survival analysis was conducted using Cox proportional hazards regression and the Kaplan-Meier method. To enhance model interpretability, SHapley Additive exPlanations (SHAP) were applied. Finally, the generalizability of each best model was assessed across two external validation cohorts.

[RESULTS] In total, there were 259 patients in the T3/T4 stage group, 266 in the LNM group, and 269 in the LVI/PNI group. The area-under-the-curve (AUC) of the best models in predicting adverse histopathological status ranged from 0.825 to 0.907 for training set, and from 0.774 to 0.848 for test set. Random Forest exhibited favorable predictive performance across all study groups. The 140 keV monoenergetic attenuation in the arterial phase was the top-ranked predictor for both stage T3/T4 and lymph node metastasis, whereas the normalized iodine concentration in venous phase was the top feature of lymphovascular/perineural invasion. All best models were associated with the patient's progression-free survival (all p < 0.05) and exhibited favorable generalizability (all AUC > 0.70).

[CONCLUSION] Machine-learning models integrating multiple DECT quantitative parameters could decode the adverse histopathological status and are related to patient progression-free survival.