Segmentation Aided Multiclass Tumor Classification in Ultrasound Images using Graph Neural Network

The merit in adapting a Graph Neural Network (GNN) for image analysis is that it can capture long-range dependencies between distant parts of the image. This is particularly important in domains such as medical imaging, where it is crucial to distinguish between organs and/or normal and abnormal regions for a disease diagnosis. While GNNs offer significant benefits, they necessitate preprocessing techniques to effectively represent images as graphs. Several techniques are proposed in the literature to address this; however, their reliance on human intervention limits their applications. Therefore, this work proposes a data aided technique that complements the model with prior knowledge of the abnormal region location within the image. Specifically, we divide an image into patches and use a deep learning-based segmentation model to extract the mask of abnormal regions to learn patch, mask, and position embeddings for graph construction. This graph is fed into a GNN for multiclass classification of breast cancer in ultrasound images. Simulation analysis shows that the proposed segmentation-aided GNN model achieved better classification performance in terms of various evaluation metrics compared to existing models. For example, compared to existing GNN models that do not require additional data, our model achieved 4% better accuracy score.