Deep learning for histopathological diagnosis of esophageal squamous cell carcinoma in biopsies: A multicenter analysis

BACKGROUND: Esophageal squamous cell carcinoma (ESCC) is a major cause of cancer mortality in Asia, where histopathological diagnosis of endoscopic biopsy specimens remain challenging. METHODS: To address this challenge, we developed an AI-based ESCC diagnostic system in biopsies (AI-EDS) using 1 104 H&E-stained whole slide images (WSIs) from one center, which were divided into three datasets at the patient level: 515 WSIs (including 226 malignancies) for training, 50 WSIs (22 malignancies) for validation, and 539 WSIs (149 malignancies) for internal testing. Lesion areas were annotated using an iPad-based system, and the DeepLab-v3 model, supported by a ResNet-50 backbone, was trained for analysis. AI-EDS was subsequently applied to 945 esophageal biopsy WSIs (351 malignancies) from four additional hospitals. Finally, we also assessed the model's performance on surgical resection and endoscopic submucosal dissection (ESD) specimens with a total of 173 WSIs (131 malignancies). RESULTS: The AI-EDS demonstrated high accuracy in distinguishing malignant from benign lesions in an internal test dataset, achieving an AUC of 0.986 (95.36% accuracy, 99.49% sensitivity, and 84.56% specificity). In multicenter validation across four external hospitals (945 WSIs), the system maintained strong performance (AUC range: 0.966-0.998). Notably, the AI-EDS matched diagnostic accuracy of junior pathologists (89%) while reducing interpretation time by more than half. Furthermore, the system demonstrated robust generalization capability in surgical resection and endoscopic submucosal dissection specimens (173 WSIs, AUC = 0.827). CONCLUSIONS: This multicenter validation confirms that the AI-EDS serves as a reliable, interpretable tool for computer-aided ESCC diagnosis, particularly valuable in resource-limited regions.