Radiomics signature for the preoperative assessment of stage in advanced colon cancer

Abstract

The purpose of this study was to develop a radiomics signature for distinguishing stage in advanced colon cancer (CC). 195 colon cancer patients were enrolled in this study (stage III, n = 146 vs. stage IV, n = 49) and divided into training cohort (n = 136) and validation cohort (n = 59). A total of 286 radiomic features were extracted from tumor and LN images. A radiomics signature was generated using the least absolute shrinkage and selection operator (LASSO) technique. The relationship between radiomics signature and CC staging was explored using a kernel-based support vector machine (SVM) classifier model. The classification performance was assessed by accuracy and the receiver operating characteristics (ROC) curve. A total of 5 features (2 for tumor and 3 for LN) were selected among 286 features. Radiomics signature built from extracted features successfully differentiated stage III from stage IV CC with no known distant metastases on imaging preoperatively. Furthermore, the SVM classifier model generated using tumor and LN images together achieved better performance than the tumor alone, with accuracies of 86.03% vs. 78.68% and 83.05% vs. 76.27% in training and validation cohorts, respectively. In ROC analysis, the model showed a significant improvement for training (AUC 89.16% vs. 69.5%) and validation cohorts (AUC 75.15% vs. 55%) in comparison with the combined analysis and the tumor alone. In conclusion, the radiomics signature based on preoperative CT may distinguish stage III from stage IV CC with no known distant metastases. In addition, the radiomic features from combined images achieved better classification performance than tumor alone.

Keywords: Colon cancer; computed tomography; metastatic lymph node; radiomics; staging.

Figure 1

Recruitment pathway for patients in this study.

Figure 2

The workflow of necessary steps in the current study. Image segmentation is performed on CT images. Experienced radiologists delineate ROI covering the whole tumor and metastatic LN by stacking up the region of interest slice by slice. The quantitative features are extracted from the tumor and LN regions on the CT images, including texture, structure and wavelet representation. The significant features were identified using LASSO regression. The performance of the classifier was evaluated measuring accuracy, sensitivity, specificity, F1-score, and ROC analysis. The SVM classifier model was used to predict the stage of CC patients using radiomics features obtained from conventional CT image.

Figure 3

An example of manual segmentation of primary tumor (green regions) and lymph node (red regions) on colon cancer CT images. (A, C) and (B, D) are raw and marked CT image of stage III and IV patients.

Figure 4

Selection of a subset of features for tumor and lymph node tissues using LASSO regularization. The features were selected among the experiments at the location that minimum error was obtained (argmin (MSE)). 2 features (RF₃₅: variance of horizontal detail wavelet representation image, RF₁₄₁: gray level non-uniformity of gradient image) for tumor (A) and 3 features (RF₁: mean, RF₃₈: variance of diagonal detail wavelet representation image, RF₈₀: contrast of vertical detail wavelet representation image) for lymph node (B) were selected among 286 features that calculated using eight different feature extraction approaches.

Figure 5

ROC curves in the training and validation cohort. A. For initial experiments, an SVM classifier was generated using only tumor texture which resulted in 78.68% and 76.27% for training and validation cohort. B. Classifier model utilizing features combined tumor and lymph node with an accuracy of 86.03% and 83.05% for training and validation cohort, respectively.