Abstract
Automatic medical image segmentation normally is a difficult task because medical images are complex in nature therefore many researchers have studied a lot of approaches to analyze patterns of images. In which, the crucial applications of deep learning in medicine are growing trends, especially Convolutional Neural Networks (CNNs) in the field of Computer Vision, yielding many remarkable results. In this paper, we propose a method to apply graph-based signal processing to CNNs architecture for medical image segmentation application. In particular, the processed architecture is based on the graph convolution to extract features in the image instead of the traditional convolution in DSP (Digital Signal Processing). The proposed solution is effective in learning neighboring links. We also introduce a back-propagation algorithm that optimizes the weights of the graph filter and finds the adjacency matrix that fits the training data. Then, the network model is applied on the dataset of medical images to help detect abnormal areas.
