Computed Tomography-Based Radiomics Model to Preoperatively Predict Microsatellite Instability Status in Colorectal Cancer: A Multicenter Study

Abstract

Objectives: To establish and validate a combined radiomics model based on radiomics features and clinical characteristics, and to predict microsatellite instability (MSI) status in colorectal cancer (CRC) patients preoperatively.

Methods: A total of 368 patients from four hospitals, who underwent preoperative contrast-enhanced CT examination, were included in this study. The data of 226 patients from a single hospital were used as the training dataset. The data of 142 patients from the other three hospitals were used as an independent validation dataset. The regions of interest were drawn on the portal venous phase of contrast-enhanced CT images. The filtered radiomics features and clinical characteristics were combined. A total of 15 different discrimination models were constructed based on a feature selection strategy from a pool of 3 feature selection methods and a classifier from a pool of 5 classification algorithms. The generalization capability of each model was evaluated in an external validation set. The model with high area under the curve (AUC) value from the training set and without a significant decrease in the external validation set was final selected. The Brier score (BS) was used to quantify overall performance of the selected model.

Results: The logistic regression model using the mutual information (MI) dimensionality reduction method was final selected with an AUC value of 0.79 for the training set and 0.73 for the external validation set to predicting MSI. The BS value of the model was 0.12 in the training set and 0.19 in the validation set.

Conclusion: The established combined radiomics model has the potential to predict MSI status in CRC patients preoperatively.

Keywords: AUC; colorectal cancer; computed tomography; logistic regression; microsatellite instability.

Figure 1

Schematic shows workflow for this study.

Figure 2

One patient with descending coloncancer, male, 52years old. The area inside the red line represents the ROI for the tumor.

Figure 3

The ROC curves of the radiomics signature in the training set.

Figure 4

The ROC curves of the radiomics signature in the external validation set.

Figure 5

Plot of regression coefficients for features F1: energy; F2: location; F3: correlation; F4: coarseness; F5: gray level nonuniformity; F6: dissimilarity; F7: compactness; F8: cluster tendendcy; F9: number of voxels.