Feng Li, Zhijian Li, Wen Wu, Chunxiao Lai, Xigang Zhang, Ruiya Zhang, Lizhi Yi, Hui Yang, Lin Qiu, Yu Lin, Guiming Liu, Quansheng Guan, Zhenyu Wang, Zhifang Zhao, Shunhui He, Fang Xie, Mengnan Zhang, Weimin Lu, Qimei Luo, Lu Yang, Junping Wang, Jie Dai, Yang Bai
This study demonstrates that a CNN-based system can achieve superior performance to human endoscopists in detecting Hp infection in atrophic gastritis, while maintaining consistent performance across varying mucosal conditions. These findings support the potential role of CNN-based artificial intelligence as a reliable tool for Hp screening in high-risk, precancerous populations. In addition, the system may facilitate targeted biopsy by providing visual guidance, thereby potentially improving the diagnostic reliability of histopathology in this setting.
BACKGROUND: Convolutional neural network (CNN)-based artificial intelligence systems have demonstrated promise in detecting Helicobacter pylori (Hp) infection in patients with preserved gastric mucosa; however, their performance in atrophic gastritis (AG), a precancerous condition with complex mucosal changes that can obscure typical endoscopic features, remains unclear.
METHODS: This retrospective multicenter cross-sectional study included 22,986 patients (464,480 endoscopic images) from four clinical centers. A CNN model was trained and optimized using a dedicated training and validation set. The model was subsequently evaluated on an independent test set and compared directly with endoscopists to assess its performance in detecting Hp infection in patients with atrophic and non-atrophic gastritis.
RESULTS: (1) Senior endoscopists exhibited a notable decline in diagnostic performance for Hp infection in AG compared to non-atrophic gastritis (NAG), with accuracy (69.9% vs. 78.0%), sensitivity (82.7% vs. 79.3%), and specificity (50.0% vs. 76.8%). Junior endoscopists showed consistently lower performance across both conditions. (2) In contrast, the CNN model maintained consistently high performance in both AG and NAG, achieving accuracy (95.9% vs. 94.9%), sensitivity (95.2% vs. 97.3%), and specificity (97.0% vs. 92.7%). (3) In AG, the CNN model significantly outperformed both senior and junior endoscopists across all diagnostic metrics (each P < 0.01).
CONCLUSION: This study demonstrates that a CNN-based system can achieve superior performance to human endoscopists in detecting Hp infection in atrophic gastritis, while maintaining consistent performance across varying mucosal conditions. These findings support the potential role of CNN-based artificial intelligence as a reliable tool for Hp screening in high-risk, precancerous populations. In addition, the system may facilitate targeted biopsy by providing visual guidance, thereby potentially improving the diagnostic reliability of histopathology in this setting.