Prediction of lymphovascular invasion of gastric cancer based on contrast-enhanced computed tomography radiomics

Abstract

Background: Lymphovascular invasion (LVI) is a significant risk factor for lymph node metastasis in gastric cancer (GC) and is closely related to the prognosis and recurrence of GC. This study aimed to establish clinical models, radiomics models and combination models for the diagnosis of GC vascular invasion.

Methods: This study enrolled 146 patients with GC proved by pathology and who underwent radical resection of GC. The patients were assigned to the training and validation cohorts. A total of 1,702 radiomic features were extracted from contrast-enhanced computed tomography images of GC. Logistic regression analyses were performed to establish a clinical model, a radiomics model and a combined model. The performance of the predictive models was measured by the receiver operating characteristic (ROC) curve.

Results: In the training cohort, the age of LVI negative (-) patients and LVI positive (+) patients were 62.41 ± 8.41 and 63.76 ± 10.08 years, respectively, and there were more male (n = 63) than female (n = 19) patients in the LVI (+) group. Diameter and differentiation were the independent risk factors for determining LVI (-) and (+). A combined model was found to be relatively highly discriminative based on the area under the ROC curve for both the training (0.853, 95% CI: 0.784-0.920, sensitivity: 0.650 and specificity: 0.907) and the validation cohorts (0.742, 95% CI: 0.559-0.925, sensitivity: 0.736 and specificity: 0.700).

Conclusions: The combined model had the highest diagnostic effectiveness, and the nomogram established by this model had good performance. It can provide a reliable prediction method for individual treatment of LVI in GC before surgery.

Keywords: contrast-enhanced computed tomography; gastric cancer; lymphovascular invasion; oncology; radiomics models.

Figure 1

Flow chart.

Figure 2

(A) LASSO the two vertical dotted lines in figure a represent the logarithm λ of the minimum mean square error (lambda.min) (left dashed line) and the logarithmic λ (right dashed line) of the minimum distance standard error (lambda.1se); lmabda.min is the best value, and lambda.1se is a model with excellent performance and the least number of independent variables. (B) Each curve in b represents the changing trajectory of each independent variable coefficient, the ordinate is the coefficient value, and the Abscissa is the number of non-zero coefficients in the model at this time. With the continuous increase of the parameters, the coefficient is finally compressed to a variable of 0, which shows that it is more important. The dotted line in the graph is the number of features under the lambda.1se parameter. (C) The vertical coordinate of the feature weight diagram represents the feature name, and the Abscissa represents the coefficient value; Coefficients: coefficient. (D) The scale on the right side of the correlation coefficient heat map shows the color depth corresponding to different correlation coefficients, and the higher the color is, the greater the correlation is; as can be seen from the figure, the maximum correlation coefficient between each feature is 0.63, so there is no highly correlated feature pair.

Figure 3

(A) Rad-score Falls Map Abscissa each bar represents a patient, green represents the patient of LVI (-), pink represents the patient of LVI (), and the ordinate represents Rad-score from-1 to 1. (B) ROC curve of the model a ROC curve of each model in the training set; (C) ROC curve of each model in the test set.

Figure 4

(A) The nomogram consists of the maximum diameter of Diameter tumour; the degree of Differentiation; and RS imaging tagging. (B) The calibration curve of the nomogram this chart shows that there is a high consistency between the observation and prediction results of the nomogram. (C) DCA decision curve validation of three model’s clinical decision-making.