LIU Zhanpeng, WANG Xiaoxiao, LIU Bowen, PENG Chen, LU Chao, WANG Zhixuan, PAN Donggang, ZHOU Yueyuan, SHAN Xiuhong
2026, 36(01): 65-74.
Objective: To explore the feasibility of preoperative prediction of perineural invasion (PNI) and lymphovascular invasion (LVI) in advanced gastric cancer using dual-energy computed tomography (DECT) venous phase imaging features and spectral parameters of a hypotonic water-filled stomach, along with clinical laboratory information.
Methods: A retrospective analysis was conducted on 161 cases of advanced gastric cancer that underwent DECT imaging within one week before surgery, and the cases were randomly divided into training sets and test sets in a 7∶3 ratio. Based on postoperative pathological assessment, 115 cases demonstrated LVI and/or PNI positivity, whereas 46 cases were negative for both. A predictive model for LVI/PNI was developed using venous phase imaging of a hypotonic water-filled stomach, DECT parameters [including the slope of the spectral Hounsfeld unit curve (between 40 keV and 100 keV), normalized iodine concentration (NIC), and effective atomic number], and clinical laboratory data (inflammatory and tumor markers). The predictive performance of the model was evaluated using the area under the ROC curve (AUC), and its clinical utility was assessed using decision curve analysis. Results: The AUC values of the radiomics model (Rad-score) in the training sets and test sets were 0.776 (95%
CI: 0.653-0.821) and
0.781(95%CI:0.582~0.847), respectively. The independent predictors for the DECT parametric model was NIC, with AUC values of
0.729(95%CI:0.615~0.790 in the training sets and
0.771(95%CI:0.604~0.864)in the test sets. For the clinical information predictive model, the independent predictor was lymphocyte percentage, with AUC values of 0.693 (95%CI: 0.638-0.805) in the training sets and 0.502 (95%CI: 0.352-0.648) in the test sets. The combined model integrating the Rad-score, DECT parameters, and clinical information had independent predictors including Rad-score, NIC, and lymphocyte percentage. The AUC values for this combined model were 0.880 (95%CI: 0.701-0.859) in the training sets and 0830 (95%CI: 0.602-0.857) in the test sets, demonstrating superior performance compared to the radiomics model, DECT parametric model, and clinical model. The DeLong test showed that the AUC of the combined model was significantly higher than that of the radiomics model, DECT parametric model, and clinical information model in the training sets (Z=1.979, P=0.048; Z=3.199, P=0.001; Z=3.053, P=0.001). In the test sets, the AUC of the combined model was also significantly higher than that of the clinical information model (Z=2.417, P=0.015). Decision curve analysis revealed that when the risk threshold ranges from 0.15 to 0.96, adopting the combined model for treatment guidance yielded a higher clinical net benefit rate. Conclusion: The integrated model, incorporating radiomics, NIC, and lymphocyte percentage, serves as a comprehensive predictive model for assessing lymphovascular invasion and perineural invasion status in advanced gastric cancer.