ABSTRACT
Segmenting the whole heart from cardiovascular volumetric scanning is critical for many clinical applications. However, manually delineating the boundary is expert-dependent, time-consuming, and has low reproducibility. Variations in image quality, heart scale, shape, and pose make it a non-trivial task for automated solutions. The surge of deep neural networks (DNNs) has brought about profound changes to this segmentation problem. Convolutional neural network (CNN), especially fully convolutional network (FCN), is one of the most popular architectures, which has witnessed many breakthroughs. In this chapter, we will introduce our recent effort in systematically designing DNNs tailored for the whole heart segmentation. We concentrate on five problems that prove to be closely related to the network performance, including how the DNNs decompose volumetric data, how to improve the training efficacy for deep DNNs, which kind of DNN architecture is more superior and computationally economic, how to properly initialize the network to avoid local minima during training, and how to design loss functions to guide the training procedure and alleviate bias. Our modified designs are verified not only in segmenting the blood pool and myocardium of the heart, but also on the more challenging task of simultaneously partitioning the whole heart into seven fine-grained substructures. Our proposed designs are general for the whole heart segmentation tasks across different modalities, such as CT and MR. All the implementations and models are online and available to encourage future investigations in this field.
