Probabilistic Brain Extraction in MR Images via Conditional Generative Adversarial Networks

Abstract

Brain extraction, which refers to the task of segmenting brain tissue in an MR image of a subject, forms an essential first step for many quantitative neuroimaging applications. These include quantifying grey and white matter volumes, monitoring neurological diseases like multiple sclerosis (MS) and Alzheimers disease, and estimating brain atrophy. Over the years several algorithms that automate brain extraction have been proposed. More recently, novel image-to-image deep learning methods have been implemented for this task, and have demonstrated significant gains in accuracy and robustness. However, to our knowledge, none of these algorithms account for the uncertainty that is inherent in brain extraction. Motivated by this, we propose a novel, probabilistic deep learning algorithm for brain extraction that recasts this task as a Bayesian inference problem, and then utilizes a conditional generative adversarial network (cGAN) to solve it. The input to the generator network is an MR image of the head and the output is a collection of images of the brain that are drawn from a probability density conditioned on the input image. These images are used to generate a pixel-wise mean image, which serves as the best guess for an image of the brain, and a pixel-wise standard deviation image, which quantifies the uncertainty in the prediction. We test this algorithm on head MR images of fifty subjects and demonstrate that it is more accurate than a commonly used brain extraction tool, and that its performance compares well with the current state of the art in deep learning algorithms. We also demonstrate the utility of the estimates of uncertainty generated by the algorithm.