Brain Image Parcellation Using Multi-Atlas Guided Adversarial Fully Convolutional Network

Abstract

Brain image parcellation is an essential procedure in neural image analysis. Recently, deep learning methods, such as fully convolutional network (FCN), have shown high computational efficiency and good performance in brain image parcellation. In this paper, a new multi-atlas guided adversarial FCN is proposed to enhance the parcellation quality. The generative model in our method is an improved FCN which is integrated with brain atlases information and multi-level feature skip connection. The discriminative model is a convolutional neural network (CNN) with multi-scale l1 loss function. Comparing to most existing deep learning based brain image parcellation methods, which use voxel-wise loss function only (e.g., cross entropy), the discriminative model in our method considers multi-scale deep features to guide the parcellation. In the experiment, two public MR brain image datasets LONI LPBA40 and NIREP-NAO are used to evaluate our method. Evaluation results demonstrate that our method outperforms the state-of-the-art methods in both datasets.