Multi-region and multi-image convolutional neural network model for detecting gastric helicobacter pylori infection.

BACKGROUND

Current Artificial Intelligence (AI) models for detecting gastric Helicobacter pylori (HP) infection rely on single-images, lacking integration of multi-regional stomach data. We developed a multi-region, multi-image Convolutional Neural Network (CNN) model to enhance diagnostic accuracy.

METHODS

From Nanfang Hospital of Southern Medical University, 5,169 cases (104,437 images) were split into training (80%) and test (20%) sets. The models-single-image CNN and our multi-region CNN-were trained and tested for HP infection diagnosis. External validation used 696 cases (20,948 images) from three non-training hospitals (Baiyun Branch of Nanfang Hospital of Southern Medical University, Shunde Hospital of Southern Medical University, Shenzhen Second People's Hospital).

RESULTS

(1) Validation results of multi-region and multi-image CNN model and single-image CNN model on the test set of Nanfang Hospital of Southern Medical University are given as follows: The accuracies are 95.1% vs. 93.3%, P < 0.05. The sensitivities are 96.5% vs. 94.4%, P < 0.05. The specificities are 93.4% vs. 92.2%, P > 0.05. The AUC are 99.0% vs. 98.1%. (2) The validation results of the multi-region multi-image CNN model and the single-image CNN model in the data set of non-training data source hospitals (Baiyun Branch of Nanfang Hospital of Southern Medical University, Shunde Hospital of Southern Medical University and Shenzhen Second People's Hospital) are given as follows. The accuracies are 89.7% vs. 77.6%, P < 0.01. The sensitivities are 90.2% vs. 82.4%, P < 0.01. The specificities are 89.1% vs. 72.9%, P < 0.01. The AUC are 92.5% vs. 82.4%.

CONCLUSION

The multi-region, multi-image CNN significantly improves AI's accuracy, sensitivity, specificity, and generalizability in diagnosing gastric HP infection.