DwinTMIL: An Innovative Approach for Helicobacter Pylori Infection Classification

This study introduces a novel approach that combines a dynamic window-sized Swin Transformer with Multi-Instance Learning (MIL) for classifying Helicobacter pylori (Hp) infections in gastroscopy images. By incorporating the U2-Net model for image segmentation, the method enhances the consistency of image quality and the robustness of the model. Experimental results indicate that this approach has distinct advantages in handling patient-labeled data, particularly demonstrating strong adaptability across images acquired from different devices. Hp infection, a global health threat, has significantly impacted human gastrointestinal health. Traditional Hp detection methods have several limitations, so developing a non-invasive, efficient, and accurate Hp infection classification method is clinically and practically significant. This study innovatively integrates a dynamic window-sized Swin Transformer and MIL to address the challenge of patient-labeled data. The Swin Transformer, with its unique hierarchical feature extraction mechanisms, effectively captures both local and global image features. MIL, through an attention mechanism, weighs and aggregates multiple image instances to generate patient-level feature representations. Moreover, the study introduces the U2-Net model for image segmentation, aiming to improve the consistency of image quality. The proposed method outperforms traditional models (e.g., Inception V3 and Efficient B7) in classification accuracy, especially showing strong adaptability and robustness when processing images from different devices. However, some challenges remain, such as the impact of data quality on model performance and the risk of overfitting in small-sample scenarios. Future research will focus on accumulating richer datasets, employing more diverse data augmentation techniques, and exploring additional model optimization methods to further enhance classification performance. This study not only provides a new technical approach for the precision diagnosis of Hp infections but also offers a novel research direction for other medical image classification tasks that rely on weakly supervised data.