Researchers from Japan have developed a new artificial intelligence (AI)-based system to classify subepithelial lesions (SELs) on endoscopic ultrasonography (EUS) images. This classification system could be useful in distinguishing gastrointestinal stromal tumors (GISTs), the most common SELs, from benign lesions.
“The AI system for classifying upper gastrointestinal SELs showed higher diagnostic performance than that of the experts. It may assist endoscopists in improving the diagnosis of SELs in clinical practice,” the authors said.
To develop the AI-based system, the authors collected 16,110 EUS images from 631 cases from 12 hospitals. These included 5 categories: GISTs, leiomyomas, schwannomas, neuroendocrine tumors, and ectopic pancreas. The images were classified by the AI-based system, endoscopists (n=2), and nonexperts (n=2).
Continue Reading
Hirai et al evaluated the diagnostic performance of the AI-based system not only for each of these 5 categories but also for differentiating GISTs from non-GISTs and GISTs/schwannomas from other SELs.
GISTs registered the highest sensitivity (98.8% for the AI-based system and 63.5%-77.6% for endoscopists), while schwannoma registered the lowest (45.5% for the AI-based system and 0% for endoscopists). Schwannomas were primarily misdiagnosed as GISTs.
Read more about GIST diagnosis
The AI-based system also performed better than endoscopists in terms of accuracy (86.1% for the AI-based system vs 27.0%-68.0% for endoscopists).
When differentiating GISTs from non-GISTs, the AI-based system showed a sensitivity of 98.8%, a specificity of 67.6%, and an accuracy of 89.3%. Positive and negative predictive values were 87.5% and 96.2%, respectively. Both sensitivity and accuracy of the AI-based system were higher than those of endoscopists (98.8% vs 25.9%-77.6% and 89.3% vs 44.3%-71.3%, respectively), whereas specificity was similar (67.6% vs 56.8%-67.6%).
When differentiating GISTs/schwannomas from other SELs, the AI-based system showed a sensitivity of 100.0%, a specificity of 76.96%, and an accuracy of 95.1%. Positive and negative predictive values were 94.1% and 100.0%, respectively. Also, both sensitivity and accuracy of the AI-based system were higher than those of endoscopists (100% vs 35.4%-84.4% and 95.1% vs 42.6%-82.8%, respectively), while specificity was comparable (76.96% vs 76.9%).
Reference
Hirai K, Kuwahara T, Furukawa K, et al. Artificial intelligence-based diagnosis of upper gastrointestinal subepithelial lesions on endoscopic ultrasonography images. Gastric Cancer. Published online November 16, 2021. doi:10.1007/s10120-021-01261-x
