Improving GNNs for Image Classification: Addressing Homophily Challenges

Graph Neural Networks (GNNs) are rapidly becoming essential tools in deep learning, but their effectiveness when applied to images is often limited by challenges in graph representation. Traditional image-to-graph conversions often result in structures with low homophily (dissimilar connected nodes), hindering GNN performance. This issue is particularly acute in medical imaging, where subtle structural variations can signify crucial diagnostic information, but it also affects a wide range of other image analysis tasks. This article introduces a novel GNN architecture designed to address these challenges broadly. Our model incorporates a learnable dictionary to capture representative node features and dynamically group similar nodes into subgraphs, enabling effective feature aggregation and promoting homophily. Coupled with attention-based pooling, this approach allows the model to learn the underlying structure of the image graph, capturing relationships between nodes and their spatial context. We demonstrate the effectiveness of our method on diverse datasets, including medical image datasets like MedMNIST and HAM10000, alongside general graph and image datasets such as TUDataset, CIFAR-10, and PascalVOC, achieving a substantial increase in accuracy and AUC relative to traditional GNNs. Our findings demonstrate a crucial step towards overcoming the limitations of applying GNNs to complex image data, with significant implications for medical image analysis and beyond.