Photo Credit: Marina113

The following is a summary of “An artificial intelligence model utilizing endoscopic ultrasonography for differentiating small and micro gastric stromal tumors from gastric leiomyomas,” published in the April 2025 issue of the BMC Gastroenterology by Duan et al.

Gastric submucosal tumors (SMTs), particularly gastric stromal tumors (GSTs) and gastric leiomyomas (GLs), represent the most prevalent benign and malignant subtypes within this category. Accurate preoperative differentiation between GSTs and GLs is critical, as they exhibit divergent biological behavior and require distinct management strategies. While conventional imaging modalities and endoscopic evaluation provide some diagnostic clues, the overlapping features between GSTs and GLs—especially in lesions smaller than 2 cm—pose substantial diagnostic challenges. This study aimed to construct and validate an artificial intelligence model based on endoscopic ultrasonography (EUS) images to improve diagnostic accuracy for small (<2.0 cm) and micro-diameter (<1.0 cm) SMTs.

A total of 358 EUS images of pathologically confirmed GSTs and GLs were used to train AI models employing DenseNet201, ResNet50, and VGG19 convolutional neural network architectures. To evaluate performance, 216 images were divided into two separate validation cohorts: Validation Set 1 included 159 images of micro SMTs (<1.0 cm), and validation Set 2 contained 216 images of small SMTs (<2.0 cm). The diagnostic capacity of each model was assessed both on a per-image and per-tumor basis by aggregating predictions from multiple images representing the same lesion. Model performance was further compared with diagnostic outputs from experienced endoscopists and a clinical feature-based model.

Among the three AI architectures, the ResNet50 model demonstrated superior performance. On a per-image basis, it achieved an AUC values of 0.938 in the training set, 0.832 in the validation set 1, and 0.841 in the validation set 2. When tumor-level predictions were consolidated from multiple images, diagnostic accuracy significantly improved, with AUCs of 0.994, 0.911, and 0.915 in the training, micro SMT, and small SMT sets, respectively. Notably, both clinical signature-based models and human endoscopists showed substantially lower AUCs across all sets, highlighting the enhanced discriminatory capacity of the ResNet50-based AI model in differentiating GSTs from GLs.

The AI diagnostic model built on ResNet50 architecture offers a highly effective and consistent method for distinguishing small and micro-diameter GSTs from GLs using EUS images. It significantly outperforms traditional diagnostic methods, including assessments by experienced endoscopists and clinical scoring models. This tool has strong potential to support more accurate preoperative decision-making and improve individualized treatment planning for patients with SMTs, particularly when conventional diagnostic approaches are limited by tumor size.

Source: bmcgastroenterol.biomedcentral.com/articles/10.1186/s12876-025-03825-y