Artificial intelligence diagnostic system predicts multiple Lugol-voiding lesions in the esophagus and patients at high risk for esophageal squamous cell carcinoma

概要

Background
When using Lugol's iodine staining, a spotty unstained area is observed in noncancerous epithelium, which we call a Lugol-voiding lesion (LVLf. Multiple LVLs after iodine staining is high-risk for esophageal cancer; however, it is preferable to identify high-risk cases without iodine staining because iodine causes discomfort and prolongs examination times. In this study, we developed an AI diagnostic system to predict multiple LVLI in the esophagus from esophagogastroduodenoscopy (EG伍images without iodine staining. The aim of I the system was to detect multiple LVLs that could not be detected by endoscopists without iodine staining and to identify patients at high risk of esophageal squamous cell carcinomas (ESCCs) and head and neck squamous cell carcinomas (HNSCCs).

Methods
We constructed the deep learning-based AI system to predict multiple LVLs without using Lugol's iodine chromoendoscopy by preparing a training set of 6634 images from white-light (WLI) and narrow-band imaging (NBI) in 595 patients before they underwent endoscopic examination with iodine staining. Diagnostic performance was evaluated on an independent validation dataset (667 images from 72 patients) and compared with that of 10 experienced endoscopists. We also evaluated six endoscopic features that would help to identify multiple LVLs using validation dataset images: 1) few glycogenic acanthosis (<2 per endoscopic image), 2) keratosis, 3) coarse mucosa, 4) invisible mucosal vessels on WLI, 5) reddish background mucosa on WLI, and 6) brownish background mucosa on NBI. I I Moreover, we retrospectively recorded the number of new ESCCs and HNSCCs detected during regular EGD in patients included in the validation dataset.

Results
The sensitivity, specificity, and accuracy of the AI system to predict multiple LVLs were 84.4%, 70.0%, and 76.4%, respectively, compared with 46.9%, 77.5%, and 63.9%, respectively, for the endoscopists. The AI system had significantly higher sensitivity than 910 experienced endoscopists. The findings of few (<2) glycogenic acanthosis per endoscopic image, keratosis, coarse mucosa, reddish background mucosa on WLI, invisible mucosal vessels on WLI, and brownish background mucosa on NBI were significantly more frequent in multiple LVLs patients than in non-multiple LVLs patients. The AI system was more sensitive than all endoscopic findings; among the endoscopic findings, the invisible mucosal vessels on WLI resulted in the highest sensitivity (76.3%) for predicting multiple LVLs ~ Moreover, patients with AI-predicted multiple LVLs had significantly more ESCCs and HNSCCs as newly detected cancers than patients without predicted multiple LVLs.

Conclusions
Our developed AI system could predict multiple LVLs with high sensitivity and identify patients at high risk for cancer using images without iodine staining. The system could enable endoscopists to apply iodine staining more judiciously.