Hybrid CNN-GCN Model for Tumor Classification: Integrating Spatial Relationships in Medical Imaging

An effective method of classifying medical images, particularly for tumor diagnosis, is proposed in this paper using a hybrid system combining convolutional neural networks (CNNs) and graph convolutional networks (GCNs). A high-dimensional representation of visual patterns and textures is extracted from input images using ResNet-18. By connecting pixels to their top-k nearest neighbours, these features are transformed into a graph structure that models spatial relationships among the pixels. A GCN layer aggregates data from neighbouring nodes to learn both local and global dependencies on a graph. To produce robust, image-level classifications, the model averages node-level predictions. It is designed to capture complex spatial relationships and fine-grained spatial details, which makes it an excellent tool for challenging medical tasks.