Interpretable deep learning model diagnoses gastrointestinal stromal tumors and lesion characteristics with microprobe endoscopic ultrasonography

Enhancing diagnostic capability of microprobe endoscopic ultrasonography (MEUS) for GISTs is clinical significance. Despite the promise of artificial intelligence (AI) in aiding diagnosis, challenges remain in model interpretability, generalization and diagnostic specificity. To address this, lesion characteristics were integrated into MEUS images to align AI inference with clinical diagnostic workflow, leading to the development of seven deep learning models. The top model, named ECMAI-ME, was assessed in two external validation sets for generalizability and compared with endoscopists in multicenter test for effectiveness. Using 9,229 MEUS images from 873 SELs across five Chinese hospitals, training involved 522 SELs, with 95 and 93 SELs from different sources for external validation, and 163 SELs for a multicenter test. ECMAI-ME achieved an AUC of 0.972 for case classification internally. In two external validation sets, ECMAI-ME maintained consistency diagnostic performance. In multicenter test, ECMAI-ME significantly outperformed endoscopists in accuracy (85.28% vs. 56.44-77.91%, p<0.05) and specificity (89.19% vs. 52.25-72.92%, p<0.001) with comparable sensitivity, and demonstrating high accuracy in distinguishing lesion echogenicity (96.93%), originating layer (88.34%), and echo heterogeneity (77.30%). This interpretable model offers high specificity, adaptability across hospitals and equipment, and strong potential for integration into clinical workflows.