EndoDSR: Automatic Detection, Segmentation and Restoration of Artifacts in Helicobacter pylori Infection Using Endoscopic Images

Objective: To develop EndoDSR, an AI-based model for the automatic detection, segmentation, and restoration of artifacts in endoscopic images. This model aims to enhance the diagnostic accuracy of Helicobacter Pylori(H. pylori), ultimately contributing to better clinical outcomes in the detection of gastric cancer. Methods and procedures: The proposed EndoDSR model makes use of the YOLOv8 architecture for artifact detection and segmentation, followed by an image restoration method, utilizing interpolation and extrapolation techniques to remove artifacts while retaining the endoscopic image features. The dataset comprises 170 endoscopic images, categorized as normal(85 images) and H. pylori positive(85 images) with classifications confirmed by biopsy reports. Results: The EndoDSR model for our endoscopic images is YOLOv8m, which has a precision of 0.92 and 0.93, a recall value of 0.89 and 0.87, a precision recall value of 0.772 and 0.749 for the boundary box and mask, respectively. These performance metrics indicate the capacity of the model to identify artifacts. Additionally, the mean Average Precision(mAP)@50 was 0.772 and the mean Average Precision(mAP)@50-95 was 0.7685, highlighting the robustness of YOLOv8m in endoscopic artifact detection. The restoration model depicted an average SSIM value of 0.879 and the Cov.PSNR is 0.078. Conclusion: The EndoDSR model tackles the important challenge in artifact removal and restoration of endoscopic images, providing a robust solution to enhance the diagnostic reliability of infection. The integration of artifact detection, segmentation, and restoration will help in real-time applications in clinical settings. Future work will focus on expanding the dataset, improving generalizability, and exploring advanced image restoration techniques to optimize gastroenterology diagnosis.