To enable diagnosis of gastric premalignant conditions from endoscopy images in time and solve the problem of invasion, high cost, and time-consuming of biopsies, we propose a precise diagnosis method for gastric premalignant conditions based on dual deep learning models. This AI-assisted approach helps physicians make faster and more accurate assessments of gastric cancer risk while avoiding the bleeding risk associated with biopsy procedures, ultimately achieving the goal of precision health.

The dual deep learning models contain the CGI classification and GIM segmentation modules. The CGI module is composed by the gastric section correlation network, which is the first method to fuse different gastric section features based on medical priori knowledge for CGI. The GIM module is composed by the mask focal modulation network, which integrates multi-level focal attention features for GIM segmentation. The results are further fused to assess the gastric cancer risk.

The proposed AI methods can be integrated into existing endoscope systems in medical institutions or health check-up centers. We have developed both stand-alone and cloud-based versions, which can assist endoscopists in identifying high-risk patients to arrange surveillance gastroscopy, enabling early stomach cancer diagnosis, reducing mortality, and advancing precision health. It also helps manufacturers tap into the growing stomach cancer diagnostic market with an 8.7% annual growth rate.