Habitat and Peritumoral Radiomics with Clinical Variables for Predicting Lymphovascular Invasion in Gastric Cancer: A Cross-Sectional Study.

RATIONALE AND

OBJECTIVES

To construct radiomics models based on various spatial feature strategies for the preoperative prediction of Lymphovascular invasion (LVI) in gastric cancer, to compare the predictive performance of intratumoral, peritumoral, and habitat-based modeling approaches, and to explore the incremental value of integrating radiomics models with clinical factors.

MATERIALS AND

METHODS

We analyzed 247 patients with pathologically confirmed gastric adenocarcinoma who underwent gastrectomy between 2018 and 2025. Arterial-phase contrast-enhanced Computed tomography (CT) images were used to delineate intratumoral and peritumoral regions. Habitat radiomics was defined using unsupervised K-means clustering to capture intratumoral spatial heterogeneity. Based on different feature construction and model-fusion strategies, we established multiple models, including intratumoral, peritumoral (with various expansion ranges), data-level fusion, feature-level fusion, and habitat models. Model performance was evaluated using the area under the receiver operating characteristic curve Area under the curve (AUC), calibration curves, and decision curve analysis (DCA). The optimal radiomics model was then integrated with the clinical model, and a nomogram was developed to predict the risk of LVI in gastric cancer.

RESULTS

A total of 68 predictive models were developed, encompassing intratumoral, 2-mm, and 3-mm peritumoral regions, habitat radiomics features, and various fusion strategies. Multivariate analysis identified independent risk factors for LVI, and the resulting clinical model showed stable validation performance (AUC = 0.846, 95% CI 0.745-0.946; Accuracy = 0.716). Among radiomics models, peritumoral features notably enhanced predictive ability; the 2-mm peritumoral model outperformed the 3-mm model (validation AUC = 0.738 vs 0.660), suggesting that a smaller peritumoral range better reflects the LVI-related microenvironment. The habitat radiomics model captured intratumoral heterogeneity, with a difference of ΔAUC = 0.108 between the training (AUC = 0.828) and validation (AUC = 0.720) cohorts, indicating a slight performance decline in the validation set. Feature-level fusion outperformed simple concatenation, with the intratumoral-peritumoral feature-level fusion model (IntraPeri) achieving AUCs of 0.853 (training) and 0.791 (validation). Integrating the optimal radiomics model with clinical variables yielded the best-performing combined model, achieving an AUC of 0.908 (95% CI: 0.843-0.974) in the validation cohort, demonstrating strong generalizability and clinical utility.

CONCLUSION

This study analyzed independent risk factors for lymphovascular invasion in gastric cancer and constructed multiple models, including clinical, intratumoral, peritumoral, and habitat-based models, which were subsequently integrated using various fusion strategies. The results showed that the feature-level fusion model combining peritumoral 2-mm features with clinical variables achieved the best performance, with the AUC in the test cohort improving to 0.908 (95% CI: 0.843-0.974). A nomogram was then developed based on this combined model.