Endoscopic ultrasound image-based nomogram combining radiomics, deep learning, and clinical features for differentiating gastric gastrointestinal stromal tumors from leiomyomas
Year: 2026
Type: article
Study design: Observational Study
Abstract: Accurately differentiating between gastric gastrointestinal stromal tumors (GISTs) and leiomyomas before surgery holds great importance due to their divergent malignant potential; however, conventional endoscopic ultrasonography (EUS) has inherent limitations, especially for lesions under 2 cm. In this retrospective study involving 408 patients (198 GISTs and 210 leiomyomas) with gastric submucosal tumors (< 2 cm), a novel deep learning-radiomics (DLR) signature was developed via feature-level early fusion, concatenating compressed deep features from a pre-trained MobileNet_v2 with handcrafted radiomics features, followed by joint feature selection. This DLR signature was integrated with independently identified clinical factors to construct a combined nomogram. The proposed combined nomogram exhibited better diagnostic performance, and the test AUC was 0.901 (95% CI 0.846–0.956), while accuracy, sensitivity, and specificity were 0.846, 0.833, and 0.857, respectively, significantly outperforming the standalone radiomics (AUC = 0.822), deep transfer learning (DTL; AUC = 0.820), DLR (AUC = 0.860), and clinical (AUC = 0.724) signatures (DeLong test; all p < 0.05). The model showed excellent calibration and the greatest net clinical benefit among various threshold probabilities on decision curve analysis. This multimodal AI-driven nomogram provides a non-invasive tool for accurately differentiating gastric GISTs from leiomyomas in lesions smaller than 2 cm, with the ability to guide personalized management and reduce unnecessary invasive procedures.
Source: Scientific Reports
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225913087
