Whole-liver enhanced CT radiomics analysis to predict metachronous liver metastases after rectal cancer surgery

Abstract

Background: To develop a radiomics model based on pretreatment whole-liver portal venous phase (PVP) contrast-enhanced CT (CE-CT) images for predicting metachronous liver metastases (MLM) within 24 months after rectal cancer (RC) surgery.

Methods: This study retrospectively analyzed 112 RC patients without preoperative liver metastases who underwent rectal surgery between January 2015 and December 2017 at our institution. Volume of interest (VOI) segmentation of the whole-liver was performed on the PVP CE-CT images. All 1316 radiomics features were extracted automatically. The maximum-relevance and minimum-redundancy and least absolute shrinkage and selection operator methods were used for features selection and radiomics signature constructing. Three models based on radiomics features (radiomics model), clinical features (clinical model), and radiomics combined with clinical features (combined model) were built by multivariable logistic regression analysis. Receiver operating characteristic (ROC) curves were used to assess the diagnostic performance of models, and calibration curve and the decision curve analysis were performed to evaluate the clinical application value.

Results: In total, 52 patients in the MLM group and 60 patients in the non-MLM group were enrolled in this study. The radscore was built using 16 selected features and the corresponding coefficients. Both the radiomics model and the combined model showed higher diagnostic performance than clinical model (AUCs of training set: radiomics model 0.84 (95% CI, 0.76-0.93), clinical model 0.65 (95% CI, 0.55-0.75), combined model 0.85 (95% CI, 0.77-0.94); AUCs of validation set: radiomics model 0.84 (95% CI, 0.70-0.98), clinical model 0.58 (95% CI, 0.40-0.76), combined model 0.85 (95% CI, 0.71-0.99)). The calibration curves showed great consistency between the predicted value and actual event probability. The DCA showed that both the radiomics and combined models could add a net benefit on a large scale.

Conclusions: The radiomics model based on preoperative whole-liver PVP CE-CT could predict MLM within 24 months after RC surgery. Clinical features could not significantly improve the prediction efficiency of the radiomics model.

Keywords: Computed tomography; Liver metastases; Radiomics; Rectal cancer.

Fig. 1

Schematic diagram of manual segmentation of whole-liver VOI. This was a 70-year-old male patient with RC in the MLM group who developed LM on follow-up images in the 13th month after RC surgery. The red outline in the figure shows the scanning-level area of the liver parenchyma without lesions. Whole-liver VOI without lesions was obtained by sketching the liver layer-by-layer, avoiding the edge of the liver, portal vein, inferior vena cava, and hepatic caudate lobe. A Original PVP CE-CT image of the liver; B manual sketching of one layer; C sketching of one layer was completed; and D schematic diagram after image segmentation

Fig. 2

Flow chart describing the workflow for construction and validation of the radiomics model

Fig. 3

Flowchart of patient enrollment

Fig. 4

Radscore distribution in the training and validation sets. Boxplot showed that there were significant differences in the radscores between the non-MLM (label 0) and MLM (label 1) groups in the training and validation sets (both p < 0.05)

Fig. 5

The ROC curves of the clinical model, radiomics model, and combined model to predict the MLM and non-MLM groups in the training set (A) and validation set (B)

Fig. 6

The nomogram for predicting MLM after RC surgery. The nomogram was composed of the radscore and pT stage

Fig. 7

DCA curves of predictive models. The slash curve (All) represents that all MLM status is positive. The horizontal line (None) represents that all MLM status is negative. The three curves (combined model, clinical model, and radiomics model) represent the clinical value for the prediction of MLM. When the threshold probability was 0.12–0.20, the net benefits of radiomics model was greater than that of combined model. When the threshold probability was 0.20–0.90, the net benefit of the combined model was similar or slightly larger than that of the radiomics model, and both were much larger than that of clinical model