Abstract
OBJECTIVE: This study aims to develop and validate an enhanced computed tomography (CT)-based radiomics model to differentiate gastric schwannomas (GS) from gastrointestinal stromal tumors (GIST) across various risk categories. METHODS: This retrospective analysis was conducted on 26 GS and 82 GIST cases, all confirmed by postoperative pathology. Data was divided into training and validation cohorts at a 7:3 ratio. We collected patient demographics, clinical presentations, and detailed CT imaging characteristics. Through univariable and multivariable logistic regression analyses, we identified independent predictors for discriminating between GS and GIST, facilitating the construction of a conventional model. Radiomic features were extracted and refined through manual 3D segmentation of venous phase thin-slice images to develop a radiomics model. Subsequently, we constructed a comprehensive combined model by integrating selected clinical and radiomics indicators. The diagnostic performances of all models in differentiating GS from GIST and stratifying GISTs according to malignancy risk were evaluated. RESULTS: We identified several key independent variables distinguishing GS from GIST, including tumor location, cystic changes, degree of enhancement in arterial phase, and enhancement uniformity. The conventional model achieved AUCs of 0.939 and 0.869 in the training and validation cohort, respectively. Conversely, the radiomics model, predicated on eight pivotal radiomics features, demonstrated AUCs of 0.949 and 0.839. The combined model, incorporating tumor location, degree of enhancement in arterial phase, enhancement uniformity, and a radiomics model derived rad-score, significantly outperformed the traditional approach, achieving AUCs of 0.989 and 0.964 in the respective cohorts. The combined model showed superior diagnostic accuracy in distinguishing GS from GIST, as well as GS from high or low malignancy potential GISTs, as evidenced by IDI values of 0.2538, 0.2418, and 0.2749 (P<0.05 for all). CONCLUSION: The combined model based on CT imaging features and radiomics features presents a promising non-invasive approach for accurate preoperative differentiation between gastric schwannomas and gastrointestinal stromal tumors.