Introduction
Gastrointestinal stromal tumors (GISTs) and leiomyomas represent the most common subepithelial tumors of the stomach. Although tumors originating from the muscularis propria often appear similar during endoscopic evaluation, their clinical management strategies and follow-up regimens differ markedly. This study aimed to develop an automated machine learning (AutoML)-based multimodal model to preoperatively differentiate gastric GISTs from leiomyomas arising from the muscularis propria.
Aims & Methods
This retrospective study included patients treated at the First Affiliated Hospital of Soochow University from January 2015 to June 2024. Eligible patients underwent endoscopic ultrasound (EUS) and contrast-enhanced computed tomography (CT), followed by endoscopic resection and histopathological confirmation. Three separate preoperative prediction models were constructed using AutoML based on: (1) clinical data, including patient baseline characteristics and endoscopic findings from white-light endoscopy (WLE) and EUS; (2) radiomics features derived from CT images; and (3) deep learning features extracted from EUS images using the Xception neural network. Additionally, bimodal and multimodal fusion models were developed to integrate these feature sets. Model performance was assessed through area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, precision, recall, and F1 score. Model interpretability was enhanced using feature importance rankings, SHapley Additive exPlanations (SHAP), and local interpretable model-agnostic explanations (LIME).
Results
A total of 200 patients were enrolled, comprising 137 GIST cases and 63 leiomyoma cases. The dataset was randomly divided into a training set (n = 134) and a validation set (n = 66) at a 70:30 ratio. In the validation set, the clinical and endoscopic features model achieved an AUROC of 0.803, accuracy of 0.848, sensitivity of 0.939, specificity of 0.588, precision of 0.868, recall of 0.939, and F1 score of 0.902. The CT-based radiomics model had an AUROC of 0.785, accuracy of 0.818, sensitivity of 0.878, specificity of 0.647, precision of 0.878, recall of 0.878, and F1 score of 0.878. The EUS-derived deep learning model showed an accuracy of 0.667, precision of 0.755, recall of 0.816, and F1 score of 0.784. Among all models evaluated, the multimodal fusion model exhibited superior performance, with an AUROC of 0.882, accuracy of 0.848, sensitivity of 0.878, specificity of 0.765, precision of 0.915, recall of 0.878, and F1 score of 0.896.
Conclusion
We developed an AutoML-based multimodal fusion model integrating clinical, radiomic, and deep learning features to accurately differentiate gastric GISTs from leiomyomas originating in the muscularis propria. The multimodal fusion approach significantly outperformed single- modality and bimodal models, demonstrating its potential as an effective, non-invasive tool for guiding clinical decision-making.