DOES ARTIFICIAL INTELLIGENCE ASSIST ENDOSCOPISTS TO EASIER DIAGNOSE GASTRIC PRECANCEROUS LESIONS AND HELICOBACTER-PYLORI INFECTION? A SYSTEMATIC-REVIEW AND META-ANALYSIS

Aims The endoscopic diagnosis of Helicobacter-pylori(Hp) infection and gastric precancerous lesions(GPL), namely atrophic gastritis and intestinal metaplasia, remains still challenging. Artificial intelligence(AI) may represent a powerful resource for endoscopists, making the endoscopic recognition of these conditions easier. Our study aimed to explore the diagnostic-performance of AI in the endoscopic diagnosis of GPL and Hp infection using AI processed endoscopic images.

Methods A systematic-review of literature, according to PRISMA, was performed searching core databases up to September-2021. Inclusion criteria were studies on the diagnostic-performance of AI-system in the diagnosis of GPL and Hp infection. A meta-analysis was performed on the pooled diagnostic accuracy of all included studies.

Results Overall, 128 studies were found, and four(patients, n=1891) and nine(patients, n=2430) studies exploring AI-system outcomes in GPL and Hp infection, respectively, were finally included. The pooled-accuracy (random effects model) was 89.1%(95%CI 85.7-92.1) and 79.64%(95%CI 66.7-90.0) for detecting GPL and Hp infection, respectively. Heterogeneity among studies, for both GPL and Hp infection, was significant [I²=69.9%(95%CI 13.6-89.5);I²=97.9%(97.2-98.5), respectively]. The Begg’s-test was significant(p=0.0371), indicating publication-bias among studies on the diagnosis of Hp infection, but not in those on GPL. Considering only those studies which used CNN-model(n=5 studies) for the diagnosis of Hp infection, the pooled-accuracy (random effects model) did not substantially change: 74.1%[(95%CI 51.6-91.4);I²=98.9%(95%CI 98.5-99.3)],Begg’s test(p=0.1416).

Conclusions AI-system seems to be a good resource to easier diagnose GPL and Hp infection showing a pooled diagnostic accuracy of 90% and 80%, respectively. Considering the high heterogeneity between studies, these promising data need external-validation by randomized-control-trials and prospective real-time studies.

Publication History

Article published online:
14 April 2022