Development of a multimodal predictive model using PET/CT radiomics and clinical data for preoperative assessment of lymphovascular invasion in gastric cancer

Abstract

Objectives: To develop and validate a multimodal predictive model combining positron emission tomography/computed tomography (PET/CT) radiomic features with clinical data for the preoperative assessment of lymphovascular invasion (LVI) in patients with gastric cancer (GC).

Methods: Between December 2017 and December 2022, 325 GC patients with pathologically confirmed LVI status were retrospectively enrolled. PET/CT scans were performed according to standard protocols, and 1,057 radiomic features were extracted from both imaging modalities following appropriate preprocessing. LASSO regression was used to select informative features for separate CT, PET, and combined PET/CT models. Key clinical variables - including age, maximum standardized uptake value, total lesion glycolysis, lymph node metastasis, and tumor grade - were integrated using multivariate logistic regression to construct a comprehensive predictive model. Model performance was assessed using ROC curve analysis. Diagnostic metrics - including AUC, sensitivity, specificity, accuracy, and Net Reclassification Improvement (NRI) - were calculated for each model.

Results: The CT, PET, and combined PET/CT models achieved AUCs of 0.823, 0.761, and 0.861, respectively. The final multimodal model integrating PET/CT radiomics with clinical data demonstrated superior performance, with an AUC of 0.904, specificity of 91.91% and sensitivity of 74.07%. Independent predictors of LVI included age, SUVmax, TLG, and lymph node metastasis. NRI analysis showed a 10.35% improvement in risk classification compared to the PET/CT radiomic model alone.

Conclusions: The multimodal predictive model demonstrated excellent diagnostic accuracy for preoperative assessment of LVI in GC patients and may support individualized treatment planning and risk stratification. Prospective multicenter studies are needed to further validate its clinical utility.

Keywords: Gastric cancer; lymphovascular invasion; multimodal model; positron emission tomography/computed tomography; radiomics.

Figure 1

Selection process of feature variables for CT, PET, and CT + PET models based on Lasso regression. A. CT Model Lasso Regression Selection Process: The left graph shows the variation in cross-validation error, and the right graph displays the path of regression coefficients for each variable, resulting in the selection of 28 feature variables. B. PET Model Lasso Regression Selection Process: The left graph shows the variation in cross-validation error, and the right graph displays the path of regression coefficients for each variable, resulting in the selection of 14 feature variables. C. CT + PET Combined Model Lasso Regression Selection Process: The left graph shows the variation in cross-validation error, and the right graph displays the path of regression coefficients for each variable, resulting in the selection of 33 feature variables. Note: CT: Computed Tomography, PET: Positron Emission Tomography, CT + PET: Combined Computed Tomography and Positron Emission Tomography, LASSO: Least Absolute Shrinkage and Selection Operator.

Figure 2

ROC curves and AUC comparison of various predictive models. Note: ROC: Receiver Operating Characteristic, AUC: Area Under the Curve, CT: Computed Tomography, PET: Positron Emission Tomography, PET/CT: Combined Computed Tomography and Positron Emission Tomography. Clinical Data (Age, SUVmax, TLG, Lymph Node Metastasis).

Figure 3

Risk reclassification performance of the new and old models. Note: The scatter plot illustrates the distribution of classification probabilities for the new model (PET/CT + clinical data) and the old model (PET/CT) in both case and control groups. The new model demonstrates more pronounced risk reclassification improvements in the control group.