Radiomics analysis of contrast-enhanced CT predicts lymphovascular invasion and disease outcome in gastric cancer: a preliminary study

Abstract

Background: To determine whether radiomics features based on contrast-enhanced CT (CECT) can preoperatively predict lymphovascular invasion (LVI) and clinical outcome in gastric cancer (GC) patients.

Methods: In total, 160 surgically resected patients were retrospectively analyzed, and seven predictive models were constructed. Three radiomics predictive models were built from radiomics features based on arterial (A), venous (V) and combination of two phase (A + V) images. Then, three Radscores (A-Radscore, V-Radscore and A + V-Radscore) were obtained. Another four predictive models were constructed by the three Radscores and clinical risk factors through multivariate logistic regression. A nomogram was developed to predict LVI by incorporating A + V-Radscore and clinical risk factors. Kaplan-Meier curve and log-rank test were utilized to analyze the outcome of LVI.

Results: Radiomics related to tumor size and intratumoral inhomogeneity were the top-ranked LVI predicting features. The related Radscores showed significant differences according to LVI status (P < 0.01). Univariate logistic analysis identified three clinical features (T stage, N stage and AJCC stage) and three Radscores as LVI predictive factors. The Clinical-Radscore (namely, A + V + C) model that used all these factors showed a higher performance (AUC = 0.856) than the clinical (namely, C, including T stage, N stage and AJCC stage) model (AUC = 0.810) and the A + V-Radscore model (AUC = 0.795) in the train cohort. For patients without LVI and with LVI, the median progression-free survival (PFS) was 11.5 and 8.0 months (P < 0.001),and the median OS was 20.2 and 17.0 months (P = 0.3), respectively. In the Clinical-Radscore-predicted LVI absent and LVI present groups, the median PFS was 11.0 and 8.0 months (P = 0.03), and the median OS was 20.0 and 18.0 months (P = 0.05), respectively. N stage, LVI status and Clinical-Radscore-predicted LVI status were associated with disease-specific recurrence or mortality.

Conclusions: Radiomics features based on CECT may serve as potential markers to successfully predict LVI and PFS, but no evidence was found that these features were related to OS. Considering that it is a single central study, multi-center validation studies will be required in the future to verify its clinical feasibility.

Keywords: Clinical outcome; Gastric cancer; Lymphovascular invasion; Radiomics.

Fig. 1

Flow chart showing the patient selection and exclusion

Fig. 2

Radiomics prediction pipeline for lymphovascular invasion and outcome

Fig. 3

ROC curves of the Radscore, Clinical and Clinical-Radscore for predicting LVI in the train cohort (a) and test cohort (b)

Fig. 4

Clinical-Radscore model presented with a nomogram scaled by the proportional regression coefficient of each predictor

Fig. 5

Calibration curve of the Clinical-Radscore model in the train cohort (a) and test cohort (b)

Fig. 6

Decision-curve analysis for the A + V-Radscore, Clinical and Clinical-Radscore

Fig. 7

Progression-free survival (PFS) curves scaled by histologic LVI status (a) and Clinical-Radscore predicted LVI status (b) with Kaplan-Meier analysis

Fig. 8

Overall survival (OS) curves scaled by histologic LVI status (a) and Clinical-Radscore predicted LVI status (b) with Kaplan-Meier analysis