Optimal artificial intelligence model based on gastrointestinal filling contrast-enhanced ultrasound: Risk stratification of gastric gastrointestinal stromal tumors.

BACKGROUND: Gastrointestinal stromal tumors (GISTs) are the most common mesenchymal tumors in the gastrointestinal tract. Imaging examinations are of great significance in the preoperative auxiliary diagnosis, postoperative monitoring of therapeutic effect and follow-up process of GISTs. Gastrointestinal filling contrast-enhanced ultrasound, as an emerging imaging technique, has the advantages of being non-invasive, radiation-free and easily tolerated by patients. It has shown a high accuracy rate in the screening and risk classification assessment of GISTs. However, the results of contrast-enhanced ultrasound examination are easily influenced by the operator's experience and subjective judgment. Therefore, it is particularly necessary to introduce an objective auxiliary evaluation technique. In recent years, predictive models based on imaging images have made significant progress in research related to GISTs, especially showing obvious advantages in disease screening and diagnosis. Based on this, the goal of this study is to develop a deep learning model based on gastrointestinal filling contrast-enhanced ultrasound images to achieve early screening, auxiliary diagnosis and risk classification assessment for patients with GISTs. It is beneficial for the follow-up and timely treatment of patients with early-stage GISTs.

METHODS: A total of 121 patients with primary gastric GISTs from July 2019 to August 2024 were enrolled and randomly assigned to a training cohort (TC) and an internal validation cohort (IVC). Ultrasound contrast - enhanced images were trained and tested using four deep - learning models to evaluate their predictive performance.

RESULTS: Between the low-risk and high-risk groups, tumor diameter, heterogeneity, and growth pattern differed significantly (P < 0.05). Among the ResNet, CNN, ViT, and EfficientNet models for gastrointestinal filling contrast-enhanced ultrasound image prediction, ResNet showed the best performance (AUC = 0.896).

CONCLUSIONS: Based on the M-NIH standard, the prediction model of gastrointestinal filling contrast-enhanced ultrasound images was successfully constructed. This model can effectively assist in the screening of patients with low-risk primary gastric stromal tumors and achieve individualized risk classification prediction.