Real‐time artificial intelligence for endoscopic diagnosis of early esophageal squamous cell cancer (with video). Issue 7 (24th January 2021)

• Record Type: Journal Article
• Authors: Yang, Xiao‐Xiao; Li, Zhen; Shao, Xue‐Jun; Ji, Rui; Qu, Jun‐Yan; Zheng, Meng‐Qi; Sun, Yi‐Ning; Zhou, Ru‐Chen; You, Hang; Li, Li‐Xiang; Feng, Jian; Yang, Xiao‐Yun; Li, Yan‐Qing; Zuo, Xiu‐Li
• Is Part Of: Digestive endoscopy. Volume 33:Issue 7(2021)
• Page Start: 1075
• Page End: 1084
• Publication Date: 2021-01-24
• Subjects: artificial intelligence -- esophageal squamous cell cancer -- magnifying endoscopy; Digestive organs -- Diseases -- Periodicals; Digestive organs -- Diseases -- Diagnosis -- Periodicals; Endoscopy -- Periodicals; Digestive System Diseases -- diagnosis -- Periodicals; Digestive System Diseases -- therapy -- Periodicals; Endoscopy -- Periodicals; 616.3
• DOI: 10.1111/den.13908
• Languages: English
• ISSNs: 0915-5635

Abstract

Abstract : Background and aims: Endoscopic diagnosis of early esophageal squamous cell cancer (ESCC) is complicated and dependent on operators' experience. This study aimed to develop an artificial intelligence (AI) model for automatic diagnosis of early ESCC. Methods: Non‐magnifying and magnifying endoscopic images of normal/noncancerous lesions, early ESCC, and advanced esophageal cancer (AEC) were retrospectively obtained from Qilu Hospital of Shandong University. A total of 10, 988 images from 5075 cases were chosen for training and validation. Another 2309 images from 1055 cases were collected for testing. One hundred and four real‐time videos were also collected to evaluate the diagnostic performance of the AI model. The diagnostic performance of the AI model was compared with endoscopists by magnifying images and the assistant efficiency of the AI model for novices was evaluated. Results: The AI diagnosis for non‐magnifying images showed a per‐patient accuracy, sensitivity, and specificity of 99.5%, 100%, 99.5% for white light imaging, and 97.0%, 97.2%, 96.4% for optical enhancement/iodine straining images. Regarding diagnosis for magnifying images, the per‐patient accuracy, sensitivity, and specificity were 88.1%, 90.9%, and 85.0%. The diagnostic accuracy of the AI model was similar to experts (84.5%, P = 0.205) and superior to novices (68.5%, P = 0.005). The diagnostic performance of novices was significantly improved by AI assistance. When it comes to the diagnosis for real‐time videos, the AI model showed acceptable performance as well. Conclusions: The AI model could accurately recognize early ESCC among noncancerous mucosa and AEC. It could be a potential assistant for endoscopists, especially for novices.