Non-invasive prediction of Ki-67 expression in gastric cancer using AI-based dual-energy CT: a multicenter study.

OBJECTIVE

To develop and validate a machine learning model based on quantitative parameters of dual-energy CT (DECT) virtual monoenergetic images (VMIs) for the noninvasive preoperative prediction of Ki-67 expression status in gastric cancer.

METHODS

A total of 367 patients with pathologically confirmed gastric adenocarcinoma were enrolled and divided into a training cohort, two external validation cohorts, and a cross-platform validation cohort. Patients were classified into high or low Ki-67 expression groups based on a 70% cutoff. Quantitative parameters of VMI were measured and incorporated into machine learning algorithms to construct the DECT model. The optimal imaging model was combined with independent clinical predictors to develop a nomogram. Model performance was evaluated using ROC analysis, calibration curves, DCA, and Kaplan-Meier survival analysis.

RESULTS

The logistic regression model was identified as the optimal DECT model. Its combination with clinical features yielded AUC values of 0.788 and 0.777 in the two DECT external validation cohorts, respectively. In the cross-platform validation cohort, the combined model achieved an AUC of 0.668. Calibration curves and DCA demonstrated good fitting and clinical usefulness of the integrated model in DECT cohorts. Stratified analysis confirmed that the model's performance was stable across different clinical characteristics. Furthermore, Kaplan-Meier analysis indicated that the combined model effectively stratified patients into high- and low-risk groups regarding progression-free survival (PFS).

CONCLUSION

The individualized model based on DECT virtual monoenergetic images effectively predicts Ki-67 expression status and provides valuable prognostic risk stratification for gastric cancer patients.