Hippocampus segmentation in MR images: Multiatlas methods and deep learning methods

Abstract

The automatic and accurate segmentation of the hippocampus in brain magnetic resonance (MR) images is important to study various neurological diseases. However, it is a challenging task due to the small structure size, the irregular shape, and the blurred boundaries between the hippocampus and its surrounding structures. In this chapter, we review the two most popularly used types of methods for this task: (1) multiatlas-based methods and (2) learning-based methods. We first review various existing patch-based multiatlas label fusion strategies. Then, we present a supervised metric learning-based label fusion in detail. This method learns a distance metric model from the atlases to keep the image patches of the same structure close to each other and those of different structures distant. For the learning-based methods, we present a multiatlas-based deep learning method and an end-to-end deep learning method in detail. Specifically, the multiatlas-based deep learning method applies a deep learning-based confidence estimation method to alleviate the potential effects of the registration errors in the traditional atlas-based methods. The end-to-end deep learning method directly learns the segmentation maps from the input images, by embedding a dilated dense network in the residual U-net. We present a comprehensive evaluation of the discussed methods compared with the state-of-the-art methods using the public datasets. In the end of the chapter, we include promising directions for related future research.