PDF | https://doi.org/10.46613/congastro2023-112
This work is licensed under CC BY 4.0
Tramon J1, Mery D1, Latorre G1, Vela J1, Irrarazabal M1, Silva F1, Espino A1, Roa J1, Riquelme A1, Achurra P1
1Pontificia Universidad Católica de Chile, Santiago, Chile
Introduction: Chronic atrophic gastritis (CAG) is a gastric preneoplastic condition. Endoscopic detection of CAG currently relies on gastric mapping biopsies and histopathologic grading to stratify gastric cancer risk. However, these biopsies involve risks (bleeding and perforation) and costs.
Objective: Todevelop an artificial intelligence (AI) system to evaluate endoscopic images and detect the presence of moderate to severe gastric atrophy.
Methodology: We performed a prospective diagnostic test study. An AI algorithm was created to determine the degree of gastric atrophy on gastroscopy videos. In the first phase, we implemented a gastric segment recognition algorithm using trained convolutional neural network models. In the second phase, we trained deep learning models to classify segmented regions into absent/mild atrophy vs. moderate/severe atrophy, using histopathological diagnosis as the gold standard. Finally, validation was performed in 47 new consecutive videotaped endoscopies from patients and compared the AI prediction to the histopathological diagnosis.
Results: In the study’s first phase, gastric segment recognition showed high sensitivity and specificity to determine the region of the endoscopic images (body, incisura angularis and antrum, Table). In the second phase, the AI system achieved satisfactory overall sensitivity and specificity in detecting moderate to severe gastric atrophy (Table). Finally, in the prospective validation phase, the AI system also outperformed the endoscopist in the diagnosis of moderate to severe gastric atrophy (OLGA II-IV) with a sensitivity and specificity of 92% and 91% vs. 85% and 75%, respectively (Table). This AI approach showed a negative predictive value of 97% and may help to avoid up to 68% of unnecessary gastric biopsies.
Conclusion: The development of this AI system for detecting gastric atrophy on endoscopic images demonstrates a high potential to improve the accuracy and efficiency of clinical diagnosis. However, further prospective external validation is required to confirm its effectiveness and safety.
