MVI plays a crucial role in the metastasis and diffusion of various primary tumor tissues, and it serves as a significant prognostic indicator [26, 27, 28, 29]. In addition, with the advancement of modern medicine, MVI has become an essential component in the treatment and prognosis of diseases [11, 12]. In this study, we developed and validated a novel nomogram model utilizing radiomics features to predict MVI, which exhibited excellent predictive performance. The CT model demonstrated an AUC of 0.779, sensitivity of 79.6%, specificity of 69.9%, and accuracy of 72.9%. Moreover, the CT-based prediction model showed significantly superior performance compared to the original model (AUC, 0.779 vs. 0.727, P = 0.01), albeit slightly lower than the nomogram model proposed by Meng et al. (AUC, 0.779 vs. 0.792) [5]. However, our nomogram model was based on three factors, while theirs utilized five, which to some extent helped reduce the potential errors in information collection.
Univariate and multivariate logistic regression analysis revealed significant correlations between MVI and various factors including histological type, Lauren classification, mural stratification, tumor diameter, cT, cN, Alb, CHE, and CA724. Among these factors, Lauren classification, mural stratification, and Alb were identified as independent influencing factors.
Previous studies have demonstrated that the tumor stroma, which is rich in fibroblasts and neovascularization, exhibits significant enhancement following the administration of contrast agents [21, 30]. The interaction between cancer-associated fibroblasts and cancer cells can induce tumor interstitial enhancement and fibrosis, resulting in an unclear boundary between the tumor and the surrounding interstitial tissue on CT images. [24, 31, 32]. In the univariate analysis, our results were consistent with those of Cha et al. [21], as both histology and mural stratification showed significant associations with MVI. However, the multivariate analysis indicated that only mural stratification remained as an independent predictor in the final prediction model, excluding histology. This could be explained by the grade 2 mural stratification, which is characterized by invasive tumors with rich fibrotic stroma, diffuse enhancement, and thickening, leading to unclear boundaries between the tumor and the surrounding stroma. On the other hand, histology (specifically signet ring cell carcinoma) represents enhanced fibrosis and a higher tumor stroma. Therefore, an important CT image feature for assessing the presence of MVI in gastric cancer is the presence of enhanced diffuse interstitial fibrosis with hypertrophy. Li et al. [33] previously reported a close relationship between tumor thickness and lymph node (LN) metastasis. In our correlation analysis based on tumor thickness, we found significant correlations with mural stratification, tumor diameter, cN, cT, and other pathological data. However, in this study, tumor thickness did not show a significant correlation. One possible reason for this discrepancy is that both grade 1 mural stratification (polypoid or cauliflower-like) and grade 2 mural stratification (interstitial hypertrophy) in our study exhibited larger tumor thickness. Furthermore, in our study, the GC Lauren classification was identified as an independent factor affecting MVI and was included in the final prediction model. However, there has been some controversy regarding the association between Lauren classification and MVI risk [34, 35]. This discrepancy may be attributed to the inclusion of MVIs with lymphatic and/or vascular invasion as separate entities in our study. Additionally, low Alb levels were identified as an independent factor contributing to MVI. Although there is currently no literature reference on the association between Alb and MVI, it is undeniable that nutritional indicators of patients (such as BMI, Alb, CRP/Alb ratio, and PNI) play important roles in determining patient prognosis [36, 37].
The main advantages of our research are as follows. Firstly, this is a clinical study that combines CT images with clinicopathological parameters to predict MVI in patients with GC. This approach demonstrates high feasibility for clinical practice and exhibits good predictive ability. Additionally, our model demonstrates a high specificity (86.0%), which can serve as an effective screening tool for non-MVI GC. Secondly, this study reveals a significant correlation between CT mural stratification and multiple tumor characteristics, providing valuable references for establishing other tumor classification and prediction models for different therapeutic effects. However, our research also has some limitations. First and foremost, we lack independent external validation, which results in the absence of reliable external validation data. Nevertheless, we employed bootstrap resampling, considered the best method for internal verification [25]. Secondly, this is a single-center retrospective clinical study with a relatively small sample size. Therefore, it is necessary to conduct a large-scale, multi-center, prospective clinical study for further verification. Lastly, our research is based on previous studies that focused on CT features at the portal vein stage. This approach may not fully represent the entire tumor characteristics and could potentially overlook other tumor details. Hence, further investigation into CT features at different levels is warranted.