Multi-scale hybrid embedding transformer for atrophic gastritis recognition

Atrophic gastritis is a chronic gastric disease that can be identified through gastroscope observation. Automatically identifying atrophic gastritis and its location through endoscopic images can effectively reduce the burden on doctors. However, the similarity of adjacent areas and the less obvious nature of lesions pose significant challenges to existing diagnostic methods. In this paper, we propose a novel method called Multi-scale Hybrid Embedding Transformer (MHET). MHET can capture multi-scale features from images to address this issue, achieving more accurate recognition of atrophic gastritis. Furthermore, we utilize generative adversarial networks (GANs) to synthesize endoscopic images, thereby resolving the problem of data imbalance. We have collected an atrophic gastritis dataset and conducted model training, as well as related experiments. The results indicate that our method achieves high performance on multiple evaluation metrics and its effectiveness is validated through ablation studies.