SAGCN: Self-adaptive Graph Convolutional Network for pneumonia detection

Pneumonia, due to its high incidence and potential lethality, necessitates rapid and accurate diagnostic methods. Chest X-rays and CT scans are pivotal tools in pneumonia diagnosis. While traditional image analysis techniques heavily depend on the expertise of radiologists, they result in subjectivity and inconsistency. Moreover, these techniques exhibit inefficiency when processing large datasets. Deep learning techniques, especially Convolutional Neural Networks (CNNs), have made significant advances in the field of medical image analysis, improving the accuracy and efficiency of pneumonia detection. However, CNNs face challenges in processing lung images with irregular shape and distribution, and mainly extract local features, with limited performance for global structural information and lesion correlation. Graph Convolutional Networks (GCNs) successfully extend the convolution operation from regular grid data to irregular graph data by using adjacency matrix and node features, and better capture the global correlation in irregular image structures. To address the limitations of the traditional message passing mechanism of GCN, we propose a novel $k$-hop graph construction algorithm that minimizes the introduction of redundant connections in higher-order graphs. We also introduce the Self-Adaptive Graph Convolutional Network (SAGCN), which incorporates an innovative graph convolution method that aggregates information across various hop distances. This method allows the adjustment of the aggregation range by varying the hop $k$ value. Additionally, we integrate a graph attention mechanism to mitigate the impacts of higher-order graph alterations on node connectivity. Moreover, our Node Adaptive Range Fusion (NARF) module enables effective multi-hop feature fusion and eliminates the issues associated with non-interactive nodes. We evaluated the SAGCN on two public pneumatic datasets, where it demonstrated superior performance with accuracies of 98.34% and 97.22%, respectively. These results significantly surpass several state-of-the-art methods, confirming the efficacy of SAGCN in pneumonia detection.