Preoperative assessment of grade, T stage, and lymph node involvement: machine learning-based CT texture analysis in colon cancer

Abstract

Purpose To investigate whether texture analysis of primary colonic mass in preoperative abdominal computed tomography (CT) scans of patients diagnosed with colon cancer could predict tumor grade, T stage, and lymph node involvement using machine learning (ML) algorithms. Materials and methods This retrospective study included 73 patients diagnosed with colon cancer. Texture features were extracted from contrast-enhanced CT images using LifeX software. First, feature reduction was performed by two radiologists through reproducibility analysis. Using the analysis of variance method, the parameters that best predicted lymph node involvement, grade, and T stage were determined. The predictive performance of these parameters was assessed using Orange software with the k-nearest neighbor (kNN), random forest, gradient boosting, and neural network models, and their area under the curve values were calculated. Results There was excellent reproducibility between the two radiologists in terms of 49 of the 58 texture parameters that were subsequently subject to further analysis. Considering all four ML algorithms, the mean AUC and accuracy ranges were 0.557–0.800 and 47–76%, respectively, for the prediction of lymph node involvement; 0.666–0.846 and 68–77%, respectively, for the prediction of grade; and 0.768–0.962 and 81–88%, respectively, for the prediction of T stage. The best performance was achieved with the random forest model in the prediction of LN involvement, the kNN model for the prediction of grade, and the gradient boosting model for the prediction of T stage. Conclusion The results of this study suggest that the texture analysis of preoperative CT scans obtained for staging purposes in colon cancer can predict the presence of advanced-stage tumors, high tumor grade, and lymph node involvement with moderate specifcity and sensitivity rates when evaluated using ML models.