Artificial intelligence-assisted confocal laser endomicroscopy for predicting invasion depth of superficial esophageal mucosal lesions: a cohort study.

INTRODUCTION: Accurate assessment of invasion depth in esophageal squamous cell carcinoma (ESCC) is essential for optimal treatment. Probe-based confocal laser endomicroscopy (pCLE) enables real-time in vivo imaging, but its interpretation depends heavily on endoscopist expertise. This study aimed to develop and validate an artificial intelligence-assisted pCLE (AI-pCLE) system for the differential diagnosis of low-grade intraepithelial neoplasia (LGIN) from high-grade intraepithelial neoplasia with submucosal invasion (HGIN-SM1). METHODS: In this retrospective single-center study, 1234 pCLE images were used to develop and validate the AI model. Histopathology from resected specimens served as the reference standard. Six deep learning algorithms were evaluated using accuracy, sensitivity, specificity, area under the curve, positive predictive value (PPV), and negative predictive value (NPV). The best-performing model was selected and compared with 10 endoscopists. RESULTS: For the differentiation between LGIN and HGIN-SM1 lesions, the AI-pCLE system exhibited a sensitivity of 97.4%, specificity of 92.6%, accuracy of 95.3%, PPV of 94.17%, and NPV of 96.70%, respectively. In comparison, the mean sensitivity, specificity, accuracy, PPV, and NPV of the endoscopists were 83.79%, 85.26%, 84.46%, 88.14%, and 82.61%, respectively. With AI-pCLE assistance, the accuracy, sensitivity, and NPV of the endoscopists increased to 94.41% (P=0.001), 96.98% (P=0.006), and 96.24% (P=0.001), respectively. The diagnostic performance of the AI-assisted system was comparable to that of expert endoscopists. CONCLUSIONS: The AI-pCLE system demonstrated robust diagnostic performance in differentiating LGIN from HGIN-SM1 lesions, indicating its potential as a reliable tool for evaluating the invasion depth of ESCC.