Integrating CT radiomics and morphologic features for preoperative risk stratification of gastrointestinal stromal tumours

Abstract To evaluate the feasibility of combining CT morphologic and radiomics features to predict malignancy risk in GIST patients and develop a multivariate regression model. Ninety-two patients with pathologically confirmed GISTs were enrolled. CT morphologic features were reviewed, and 42 radiomics features were extracted from the tumours on portal venous phase images. In the univariate analysis, each morphological and radiomics characteristic was compared between the low- and moderate/high-risk groups. Three multivariate regression models were performed for morphological, radiomics, and combined features to reveal the best predictor variables and model for malignancy risk prediction. In the multivariate regression model using CT morphologic features, the presence of tumour necrosis and tumour vessels were significant indicators for differentiating low-risk from moderate/high-risk GISTs (AUC = 0.81). In a model using CT radiomics features, skewness, total energy and GLCM_Idmn were significant indicators for differentiating both risk groups (AUC = 0.89). After combining both significant morphological and radiomic features in one model, the presence of tumour vessels and GLCM_Idmn were significant indicators for differentiating both risk groups (AUC = 0.90). Combined CT morphologic and radiomics analysis proved to be a useful method for differentiating low-risk from moderate/high-risk GISTs with high diagnostic performance. This integration has the power to guide comprehensive treatment strategies and determine the eligibility for adjuvant imatinib therapy. To the best of our knowledge, this is the first study to integrate CT morphological and radiomics features into a single predictive model for risk stratification of GISTs.