CGAM: An end-to-end causality graph attention Mamba network for esophageal pathology grading

Esophageal squamous cell carcinoma (ESCC) is a common squamous epithelial cell carcinoma. The topological interaction between the internal tumor and the tumor microenvironment (TME) leads to the complexity and heterogeneity of the tumor. However, existing mainstream grading models are not well suited for learning, utilizing, or explaining the relationships between tumors and the tumor microenvironment. Inspired by these critical issues, we propose an end-to-end causality graph attention mamba network (CGAM) with three advantages: First, the graph attention mechanism is employed as a graph construction method for topological relationship representation learning to improve the ability of the graph neural network to learn relational representations. Second, the causality of mamba's selective state space mechanism is used to enhance the receptive field of the model and guide the graph neural network to capture the relationship between the tumor and the TME. Third, the entropy constrained decision consistent optimization method is designed to solve the problem of inconsistent two-branch decisions. We compared the grading performance of 19 SOTA models on two pathology datasets, and our grading accuracy was 2.72 % and 2.90 % higher than that of the highest SOTA model. Moreover, our model outperforms the other models in terms of the Matthews correlation coefficient, critical success index and kappa coefficient. Extensive experiments show that CGAM outperforms other SOTA models in terms of pathology grading and interpretability. These findings suggest our CGAM method effectively learns, exploits and explains the topological interaction between tumors and the TME. Finally, these findings that it has superior generalization performance and clinical utility.