Effectiveness for the Artificial Intelligence Diagnosis Model for Gastric Cancer in Gastric Cancer Surveillance Program

Abstract

　Aim: Endoscopic examination has been started since 2016 for the national gastric cancer surveillance program. The national gastric cancer surveillance program included 1^st and 2^nd endoscopic image reading. The aim of our study was to evaluate the effectiveness of artificial intelligence gastric cancer diagnosis in 2^nd endoscopic image reading.

　Methods: We retrospectively collected endoscopic images from endoscopic examinations performed in Nagasaki Prefecture. We analyzed these endoscopic images using our object detection AI model for detecting gastric cancer. Our AI model evaluated endoscopic images classified as anomaly or normal images. We evaluated sensitivity and specificity for gastric cancer detection per patient level and per image level, respectively.

　Results: A total of 65 patients were analyzed. The mean number of endoscopic images per patient was 66.7 images (standard deviation 9.4 images). The sensitivity and specificity for gastric cancer per patient level were 100% (6/6 patients) and 20.3% (12/59 patient), respectively. The sensitivity and specificity for gastric cancer per image level were 60.6% (54/89 images) and 95.7% (4,067/4,252 images), respectively. The mean number of ‘anomaly' images was 3.6 images (standard deviation 4.1 images). The mean number of ‘normal' images was 63.1 images (standard deviation 8.8 images).

　Conclusion: Our gastric cancer endoscopic diagnosis AI model may be useful for 2^nd reading in gastric cancer screening programs.