Related Experiment Video
Updated: May 15, 2025
Intraoperative Gastroscopy for Tumor Localization in Laparoscopic Surgery for Gastric Adenocarcinoma
Published on: August 9, 2016
An artificial intelligence model utilizing endoscopic ultrasonography for differentiating small and micro gastric
Ruifeng Duan1, Liwei Duan1, Xin Chen1
1Department of Gastroenterology and Digestive Endoscopy Center, The Second Hospital of Jilin University, Chang Chun, Jilin, China.
Background:
Gastric stromal tumors (GSTs) and gastric leiomyomas (GLs) represent the primary subtypes of gastric submucosal tumors (SMTs) characterized by distinct biological characteristics and treatment modalities. The accurate differentiation between GSTs and GLs poses a significant clinical challenge. Recent advancements in artificial intelligence (AI) leveraging endoscopic ultrasonography (EUS) have demonstrated promising results in the categorization of larger-diameter SMTs (> 2.0 cm). However, the diagnostic capacity of AI models for micro-diameter SMTs (< 1.0 cm) remains uncertain due to limited imaging features. This study seeks to develop a specialized diagnostic model utilizing EUS images to differentiate small and micro GSTs from GLs effectively.
Methods:
In this study, a dataset comprising 358 EUS images of GSTs or GLs was utilized for training the EUS-AI model. Subsequently, 216 EUS images were allocated for validation purposes, with 159 images in validation set 1 (micro SMTs: tumor diameter < 1.0 cm) and 216 images in validation set 2 (small SMTs: tumor diameter < 2.0 cm). The diagnostic performance of the EUS-AI model for individual tumors was assessed by consolidating the diagnostic outcomes of the corresponding images. Comparative analyses were conducted between the diagnostic outcomes of endoscopists, clinical signatures, and those of the EUS-AI models.
Results:
The EUS-AI models were developed using DenseNet201, ResNet50, and VGG19 architectures. Among the three models, the ResNet50 model demonstrated superior performance on EUS images, achieving area under the curve (AUC) values of 0.938, 0.832, and 0.841 in the training set, validation set 1, and validation set 2, respectively. By combining predictions from multiple images for each tumor, the diagnostic efficacy of ResNet50 was further enhanced, resulting in AUCs of 0.994, 0.911, and 0.915 in the aforementioned sets. In comparison, both clinical signatures and endoscopists exhibited notably lower AUC values than those obtained with the EUS-AI model.
Conclusions:
The EUS-AI model utilizing ResNet50 architecture effectively discriminates between micro GSTs and GLs from both image-centric and tumor-centric perspectives. Demonstrating superior diagnostic efficiency compared to clinical models and assessments by endoscopists, the EUS-AI model serves as a valuable tool for clinicians in precisely distinguishing small and micro GSTs from GLs before surgery.
