S3162 Unlocking Microscopic Insights: AI-Driven Analysis of Magnified Narrow Band Imaging for H. pylori

Introduction: Helicobacter pylori (H. pylori) is well known for causing gastritis and gastric malignancy, necessitating accurate and timely diagnosis for effective management. While endoscopy allows direct visualization, advanced techniques like narrow band imaging (NBI) with optical magnification enhance mucosal details associated with H. pylori infection. These subtle patterns, however, require expert interpretation, potentially leading to variability. This study aims to develop an artificial intelligence (AI)–based system capable of accurately detecting and identifying H. pylori infection from endoscopic images obtained through NBI. Methods: This prospective study was conducted at Aster MIMS Kannur, India, over a 1-year period from March 2023 to March 2024. Sixty patients undergoing upper gastrointestinal endoscopy were recruited for the study after meeting the inclusion. The study cohort comprised 2 equal groups: 30 patients diagnosed as H. pylori-negative and 30 patients diagnosed as H. pylori-positive based on rapid urease test. High-resolution 4K NBI images with optical magnification were acquired and transferred for feature extraction at National Institute of Technology Calicut. Features were extracted using Gabor bank 2 filters combined with pixel values, excluding black pixels, resulting in a large feature set. Eight AI models (ANN, CNN, SVM on Gabor 1, Random Forest on Gabor Bank 2) were developed and their validation accuracy was assessed. Results: This study was a prospective investigation conducted at Aster MIMS Kannur from March 2023 to March 2024, involving 60 patients (30 H. pylori-positive and 30 H. pylori-negative). High-resolution 4K NBI images with optical magnification were acquired and transferred for feature extraction at National Institute of Technology Calicut. Features were extracted using Gabor bank 2 filters combined with pixel values, excluding black pixels, resulting in a large feature set. Eight AI models (ANN, CNN, SVM on Gabor 1, Random Forest on Gabor Bank 2) were developed and their validation accuracy was assessed. Conclusion: This study identifies the potential of AI, particularly CNN and Random Forest, for H. pylori diagnosis using magnified NBI. The superior performance of the Random Forest on NBI images with Gabor bank 2 features emphasizes the importance of both imaging modality and feature extraction. This write-up used help from the Gemini AI tool.