Abstract
PURPOSE: To propose a model based on computed tomography (CT) subregional radiomics to predict the preoperative mitotic index of 2-5 cm Gastrointestinal Stromal Tumors (GISTs) of the stomach. MATERIALS AND METHODS: This retrospective study enrolled a total of 368 patients with GISTs from two institutions: Center 1 comprised 239 patients (122 M, 117 F; mean age 61.66 ± 10.86 years), and Center 2 comprised 129 patients (51 M, 78 F; mean age 60.28 ± 9.72 years). Radiomics features were extracted from the entire tumor. Concurrently, k-means clustering was applied to imaging features to define three distinct tumor subregions, from which radiomics features were subsequently extracted. The Recursive Feature Addition method was used to identify features correlated with the mitotic index in patients with 2-5 cm gastric GISTs. Using the selected features from each subregion and the whole tumor, logistic regression (LR) was employed to construct subregion-based radiomics models and conventional whole-tumor-based radiomics models, respectively. RESULTS: Better performance was observed for unenhanced CT subregions 1, 2, and 3 compared with the conventional radiomics model. The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity of the model for subregion 3 in the training set were 0.98, 0.97, 0.98, and 0.90, respectively. In the validation and external test sets, the AUC values were 0.874 and 0.804, respectively. The conventional whole-tumor radiomics model based on venous phase CT demonstrated superior performance compared to all subregion-based models, achieving an AUC of 0.956 in the training set, with accuracy, sensitivity, and specificity of 0.94, 0.97, and 0.83, respectively. In the validation and external test sets, it attained AUC values of 0.892 and 0.805, respectively. CONCLUSION: Subregional CT radiomics may be used to predict the mitotic index of patients with 2-5 cm gastric GIST before surgery. In particular, subregional radiomics models based on unenhanced CT showed excellent predictive performance.