Application of artificial intelligence for detection of Helicobacter pylori infection by magnetically controlled...

Application of artificial intelligence (AI) in magnetically controlled capsule endoscopy (MCCE) is increasing. The aim of this study was to develop an AI system that can automatically detect Helicobacter pylori (H. pylori) infection in MCCE images and to evaluate the diagnostic performance of the system.

This study prospectively enrolled subjects with known H. pylori infection status for MCCE examination between April 2022 and March 2023. The MCCE images were collected and prepared for the data set (80% for training, 10% for validation and 10% for testing). Four convolutional neural network models (i.e., MobileNetV2, DenseNet264, ShuffleNetV2 and ResNet50) were applied in the AI system respectively. We evaluated their diagnostic performance and identify the optimal model by calculating the accuracy, sensitivity, specificity, and average reading time for per image.

A total of 142 subjects were registered including 71 H. pylori-positive and 71 H. pylori-negative. The numbers of images in the training set, validation set and testing set were 25,985 (114 patients), 2767 (14 patients), and 3027 (14 patients), respectively. The accuracy, sensitivity, specificity, and average reading time for per image of each CNN model were as follows: MobileNetV2 model, 95.7, 98.1, 93.3% and 0.01176 s; DenseNet264 model, 95.3, 98.5, 92.2% and 0.04572 s; ShuffleNetV2 model, 95.7, 98.5, 93.1% and 0.00588 s; ResNet50 model, 95.1, 95.8, 94.5% and 0.01110 s, respectively.

The AI system based on the ShuffleNetV2 model achieved better performance, highlighting its potential for future application to help clinicians detect H. pylori infection in MCCE examination.