Development and validation of an artificial intelligence classifier for histologic grading of superficial non-ampullary duodenal epithelial tumors.

BACKGROUND AND STUDY AIMS: Superficial non-ampullary duodenal epithelial tumors (SNADETs) are being increasingly diagnosed, and accurate endoscopic diagnosis is essential for determining appropriate treatment strategies. However, no standardized diagnostic method currently exists. We developed a deep learning-based artificial intelligence (AI) classifier to distinguish between low-grade and high-grade neoplasia and evaluated its diagnostic performance.

PATIENTS AND

METHODS

The training set of 85 SNADET cases (low-grade neoplasia, n = 41; high-grade neoplasia, n = 44; white-light images, n = 317) from Yamaguchi University were used to train the AI classifier. Validation was performed using a test set of 69 lesions (low-grade neoplasia, n = 31; high-grade neoplasia, n =38; images, n = 229) from three institutions. Three expert endoscopists also evaluated the test set using an endoscopic scoring system for comparison.

RESULTS

The AI classifier achieved 78.3% accuracy, 92.1% sensitivity, 61.3% specificity, 74.5% positive predictive value (PPV), 86.4% negative predictive value (NPV), and an F1 measure of 0.824. In contrast, endoscopist diagnosis yielded 58.0% accuracy, 50.0% sensitivity, 67.7% specificity, 65.5% PPV, 52.5% NPV, and an F1 measure of 0.567. Diagnostic accuracy and sensitivity of the AI was significantly higher than that of the endoscopists ( P <0.05).

CONCLUSIONS

The AI classifier demonstrated higher sensitivity and overall diagnostic performance in comparison to endoscopists in differentiating high-grade neoplasia, suggesting its potential utility in guiding treatment decisions.