AI-Powered Three-Category H. pylori Diagnosis via Magnetic Controlled Capsule Endoscopy: A Multicenter Validation of a Vision-Language Model

Shiping Xu^1,2*Xi Sun²Jing Liu²Lili Wu²Xiao Chen²Xiaona Ma²Fei Teng²Ting Zhang³Hui Su⁴Xin Fan⁴Jiaxin Li⁵Peng Jin^4*Hongmei Jiao^5*

• ¹Department of Neurology, First Medical Center, Chinese PLA General Hospital, Beijing, China
• ²Chinese PLA General Hospital Second Medical Center Department of Gastroenterology, Beijing, China
• ³Chinese PLA General Hospital, Beijing, China
• ⁴The Seventh Medical Center of PLA General Hospital, Beijing, China
• ⁵Peking University First Hospital, Beijing, China

Accurate classification of Helicobacter pylori (H. pylori) infection status is critical for gastric cancer risk stratification. Current methods based on traditional convolutional neural networks (CNNs) are limited by their reliance on fragmented single-image analysis and operator-dependent selection variability, impairing diagnostic reliability. To overcome these limitations, we developed MC-CLIP, a vision-language foundation model for fully automated three-categorical H. pylori diagnosis using magnetic-controlled capsule endoscopy (MCCE). The model was pretrained on 2,427,475 MCCE image-text pairs from 123,543 examinations and fine-tuned on 40,695 expertly annotated images from 864 patients. Validated on multicenter internal (n=220) and external (n=208) cohorts, MC-CLIP autonomously selected 30 representative images per case for end-to-end classification, achieving overall accuracies of 89.6% (95% CI: 85.5-93.6%) and 86.6% (80.8-90.3%), respectively. MC-CLIP demonstrated particularly strong performance in detecting H. pylori infection, with sensitivities of 91.4% for current infection and 83.7% for past infection—significantly surpassing both senior endoscopists (84.3% and 71.4%, respectively) and junior endoscopists (74.3% current infection). The model maintained high specificity (>90% across all categories) and excelled at identifying subtle post-eradication mucosal changes, reducing misclassification of past infections as non-infections. By integrating multimodal image-text data through end-to-end analysis, MC-CLIP effectively addresses fundamental limitations of CNN-based approaches and shows strong potential for enhancing MCCE-based gastric cancer screening. Accurate classification of Helicobacter pylori (H. pylori) infection status (current infection, past infection, and non-infection) is crucial for gastric cancer screening, and while traditional convolutional neural networks (CNNs) face limitations in three-categorical diagnosis due to fragmented single-image analysis and operator-dependent variability, this study developed and validated a contrastive language-image pretraining (CLIP) model (MC-CLIP) for automated three-categorical H. pylori diagnosis using magnetic-controlled capsule endoscopy (MCCE), where MC-CLIP was pretrained on 2,427,475 MCCE image-text pairs from 123,543 examinations to establish vision-language alignment and fine-tuned using 40,695 annotated images from 864 patients before being validated on internal (n=220) and external (n=208) cohorts, autonomously selecting 30 representative images per patient for end-to-end classification while demonstrating superior performance with overall accuracies of 89.6% (95% CI: 85.5-93.6%) and 86.6% (80.8-90.3%) in internal and external validation respectively, significantly outperforming