DGCNet: a clinical knowledge-driven model for the segmentation of reflux esophagitis in endoscopic images

Digestive endoscopy is vital for diagnosing reflux esophagitis (RE) and is crucial for assessing disease severity. Although some existing pixel-level precision segmentation models can achieve fine-grained lesion region recognition, artificial intelligence-assisted diagnosis of RE is still in its infancy. The mutable shape, uneven distribution, and indistinct inflammatory texture boundaries inherent to RE endoscopic features pose formidable challenges for accurate lesion recognition and segmentation. This study proposes a classification and segmentation model to improve the Los Angeles grade (LA-grade) accuracy and enable precise lesion identification and quantification.

In this study, we compiled the largest RE endoscopic image dataset, conducted a detailed classification study on it, and tested current mainstream and classic models. Based on RE endoscopic features, we introduced a clinical knowledge-driven RE segmentation model: the Deep Graph Convolution Network (DGCNet). We devised two novel modules: the Point Preservation Module to enhance model adaptability to lesion edge features and the Edge Graph Extraction Module for multi-scale fusion of edge and internal lesion features.

For classification results, we tested four classic models. The Swin Transformer model achieved the best performance for various evaluation indicators, with a classification accuracy of 86.2% ± 2.1% and an F1 Score of 82.9% ± 2.3%. This study proposed a DGCNet framework and compared it with six classic segmentation networks. The results showed that our proposed model is significantly better than the other models, with mean Dice reaching 87.8% ± 1.4% and mean IoU of 79.2% ± 1.6%.

DGCNet demonstrates exceptional segmentation performance, showing promise for enhancing the accuracy of endoscopic LA-grade classification and precise recognition and segmentation of lesions.