Classification of Medical Images with Synergic Graph Convolutional Networks

Abstract

Medicine has always been an important area of concern for people's lives. Medical images, as an important basis for doctors to diagnose diseases, has its own particularity. For example, many medical images are often difficult to distinguish due to intra-class variation and inter-class similarity, and medical images have high requirements for processing accuracy. A synergic graph convolutional networks (SGCN) model is proposed for image classification. This model is based on convolutional neural networks on graphs with fast localized spectral filtering. In our model, two graph convolutional networks (GCN) can learn from each other. We choose the Kth-order Chebyshev polynomials of the Laplacian to control K-localized of spectral filters conveniently. Specifically, we concatenate the image representation learned by both GCNs as the input of our synergic deep learning framework to predict whether the pair of input images belong to the same class. The intra-class similarity and inter-class variability of the dataset itself makes the performance of a single graph convolutional neural network better. We evaluated our SGCN model on MNIST and some Brain MRI image classification dataset and achieved advanced performance.