Chapter 5 - Segmentation: intracardiac echocardiography contouring

Abstract

In this chapter, we focus on how to design deep neural networks for medical image segmentation. In Section 5.1, we present several design principles, namely how to choose backbone neural networks for image segmentation, how to choose segmentation tasks and learning objectives, and how to combine the image restoration tasks to facilitate the image segmentation. Then in Section 5.2, we provide a case study to show how these design principles are implemented in practice. For the choice of segmentation models, we introduce three families of deep neural networks for segmentation in Section 5.1.1, i.e., FCNs, CNN with graphical models, and encoder–decoder models. In Section 5.2, we proceed with the encoder–decoder model design and provide examples of designing 2D and 3D UNet-like [38] encoder–decoder models for ICE (Intracardiac Echocardiography Contouring) image segmentation. We cover two commonly encountered medical image segmentation tasks in Section 5.1.2, i.e., semantic and instance segmentation. We provide scenarios for which segmentation tasks may be considered in clinical applications. We also introduce different learning objectives for each segmentation task and provide guidance on the choices of learning objectives. In Section 5.2, we give details on how semantic segmentation is applied to delineate the cardiac structures from ICE images. Finally, we note the connection between image restoration and image segmentation in Section 5.1.3 and present three ways of leveraging image restoration to facilitate image segmentation. In Section 5.2, we introduce how 3D ICE image inpainting (one of the image restoration tasks) can be jointly trained with 3D ICE image segmentation. We show that the joint learning provides better 3D understanding of the cardiac structure and, as a result, give better segmentation performance than a 2D only segmentation model.