Introduction: Gastric cancer development can be considered a type of field cancerization (Slaughter DP. Cancer. 1953 Sep;6:963-8.) due to persistent H. pylori infection, and it is recommended that observations during screening esophagogastroduodenoscopy (EGD) should cover the entire stomach. To support comprehensive observation and photodocumentation during screening exams, a new artificial intelligence (AI) function called the Landmark Photo Checker (LMPC) has been integrated into CAD EYE^TM (FUJIFILM. Co, Tokyo). This endoscopic AI function facilitates endoscopic photodocumentation by providing a visual indication of completion when appropriate still images of seven anatomical landmarks in the stomach (Figure 1). Aims & Methods: This study aimed to evaluate the sensitivity of LMPC for these seven landmarks as the primary endpoint, with specificity and accuracy rate as secondary endpoints. EGDs were conducted by expert endoscopists using linked color imaging (LCI). The study involved 121 participants were randomly assigned to two groups: the sensitivity verification group (n=65) and the specificity verification group (n=56). In the sensitivity verification group, the endoscopists captured appropriate still images of the seven landmarks during the examination. Conversely, in the specificity verification group, the endoscopists intentionally captured inappropriate images of the seven landmarks. Sensitivity was defined as the proportion of images marked as "good" by the LMPC in the sensitivity verification group, while specificity was defined as the proportion of images not marked as "good" by the LMPC in the specificity verification group. In the specificity verification group, appropriate images of the landmarks were additionally captured using white light imaging before the endoscope was removed. Results: A total of 808 judgments made by the LMPC during upper gastrointestinal endoscopy using LCI observation light were analyzed statistically. The sensitivity of the landmark, set as the primary endpoint for evaluation of the LMPC, was 87.4%. The secondary endpoints, specificity and accuracy rate, were 75.7% and 81.9%, respectively. The individual sensitivities (%) and specificities (%) for the seven landmarks are as follows: the esophagogastric junction (93.0 and 92.2), the cardia and fundus (75.4 and 42.6), the cardia near field view (80.0 and 55.6), the body greater curvature forward view (89.2 and 87.3), the body lesser curvature retroflex view (94.9 and 79.6), the angulus (88.1 and 83.6), and the pylorus near field view (93.3 and 89.1). The mean examination time for all 121 cases was 7 minutes and 50 seconds. Conclusion: The LMPC demonstrated high sensitivity to anatomical landmarks within the stomach and was found to be effective in supporting the capture of appropriate endoscopic images using the image-enhanced endoscopy technique LCI.

Complete status of the LMPC.

Hiroshi Kawahira

Appearances