Graph-based retrieval of PET-CT images using vector space embedding

Graph-based content-based image retrieval (CBIR) techniques, which use graphs to represent image features and calculate image similarity using the graph edit distance, achieve high retrieval accuracy. However, such techniques suffer from high computational complexity. In this paper, we present a graph-based CBIR algorithm that achieves improved retrieval efficiency. We compute a vector space embedding for every graph, using their distances from a set of prototype graphs, so that each vector component represents a distortion from a prototype. This process is performed offline. We compare images by computing the Euclidean distance of the vector embeddings, which is a faster process than calculating the graph edit distance. We evaluated our work using 50 combined positron emission tomography and computed tomography (PET-CT) volumes of patients with lung tumours. Our results show that our method is at least 21 times faster than the graph edit distance with a mean average precision difference of less than 4%.