Comparison of machine learning algorithms for the radiomic classification of gastric subepithelial tumors.

11536 Background: Diagnostic challenges arise in classifying gastric subepithelial tumors (SETs) due to the absence of specific imaging findings and limitations in endoscopic biopsy, particularly for deep-seated tumors. Differentiating SETs that require surgical intervention from those suitable for diagnostic imaging follow-up is crucial. Gastrointestinal stromal tumors (GIST), with malignant potential, necessitate prompt resection, while leiomyoma and schwannoma are benign and can be followed up if asymptomatic. Many radiomic approaches using machine learning (ML) were tried to classify gastric SET. Nevertheless, a comprehensive exploration of current ML algorithms still needs to be explored. Methods: The patients who underwent the surgery for gastric SET in our hospital were considered eligible subjects. All cases were confirmed by pathologic evaluation, and the cases of GIST and leiomyoma were chosen for this study. Manual segmentation and radiologist validation were performed, followed by image resampling to 1 x 1 x 1 mm resolution. Radiomics features were extracted using PyRadiomics 3.0.1, focusing on 32 first-order and shape features for interpretability and robustness. Various models, including K-nearest neighbors (KNN), support vector machines (SVM), logistic regression (LR), lightGBM, random forest (RF), and TabPFN (by Hollman N et al.), underwent evaluation. Nested cross-validation with 10 folds and 10 repetitions ensured fair comparison, and Bayesian hyperparameter optimization was conducted in a 5-fold inner CV, excluding TabPFN. Comparative evaluation included area under the receiver operating characteristics (AUROC), accuracy, balanced accuracy, and Matthew Correlation Coefficient (MCC). Results: Among the 359 eligible patients, resulting in 241 GISTs and 27 leiomyomas post-exclusion. We used 26 pre-isolated cases as a test set of the model. The model test results for various ML models in classifying gastric SETs are presented (Table). Notably, LR exhibited the highest accuracy (0.899), F1 score (0.950), and AUROC (0.653) compared to other models without statistical significance. Conclusions: Our study highlights that LR improves the accuracy of diagnosing gastric SETs. These findings contribute to refining diagnostic approaches and may optimize patient management strategies for gastric SETs.[Table: see text]