Interpretable Machine Learning Model Based on CT Semantic Features and Radiomics Features to Preoperatively Predict Ki-67 Expression in Gastrointestinal Stromal Tumors

To develop and validate a machine learning (ML) model which combined computed tomography (CT) semantic and radiomics features to preoperatively predict Ki-67 expression in gastrointestinal stromal tumors (GISTs) patients. We retrospectively collected the clinical, imaging and pathological data of 149 GISTs patients. We randomly assigned the patients in a ratio of 7:3 to a training set (104 cases) and a validation (45 cases) set. We divided the patients into low and high Ki-67 expression group according to postoperative pathology. CT semantic features were analyzed from preoperative enhancement CT images and radiomics features were extracted from venous phase-enhanced images. We used intraclass correlation coefficient, maximal relevance and minimal redundancy and least absolute shrinkage and selection operator method to screen radiomics features and build radiomics label. 6 ML models were used for model construction. Receiver operating characteristic curves were used to evaluate the predictive efficiency of ML models. SHAP analysis was used to explain the contribution of different variables and their risk threshold. AUC of radscores in predicting Ki-67 expression of GIST patients were 0.749 and 0.729 in training and validation set. Among the 6 ML models, SVM exhibited best prediction accuracy. AUC of SVM model in predicting Ki-67 expression of GIST patients were 0.840, 0.767 and 0.832 in training, validation and test set. SHAP analysis showed that radscores and tumor diameter had highly positive contribution to the model. Therefore, the interpretable SVM model can predict Ki-67 expression of GISTs patients individually before surgery, which can provide reliable imaging biomarkers for clinical treatment decisions.