Regularized CNN with Geodesic Active Contour and Edge Predictor for Image Segmentation

In order to exploit effectively the benefits of classical variational methods with good interpretability and high generalization performance, this paper proposes a novel regularized convolutional neural network (CNN) based on geodesic active contour (GAC) and edge predictor (EP) for image segmentation. The main idea is to establish a variational problem which integrates the Heaviside function such that the GAC prior is easily added into the problem. Furthermore, an edge predictor module is designed to predict the edges of target objects and an edge predictor function (EPF) is generated instead of the traditional edge indicator function in the GAC. Besides, an iterative convolution soft thresholding module (ICSTM) is developed to numerically solve the GAC and EPF based variational problem, and merged into an existing CNN to generate our new end-to-end network. It is also proved that the ICSTM algorithm is unconditionally stable. Finally, experimental results on synthetic, MRI and CT images show that the proposed method is quite competitive with the other state-of-the-art segmentation methods especially in segmenting noisy images with low contrast.