ABSTRACT
The characterization of the heart function and anatomy requires the segmentation of the main regions at both the systole and diastole. This is usually done by means of magnetic resonance image (MRI) cine sequences, usually with short-axis views of the heart.
Convolutional neural networks (cnn) can be employed to achieve the segmentation of the desired regions of interest. This chapter describes how to design and apply convolutional neural networks for the task of segmentation in short-axis cardiac MRI. It covers the main features that characterize the image modality and an overview of the segmentation problem with (cnn). 2The most popular segmentation models are introduced and the most relevant advances done in them are described. A review of the problem at hand is conducted along an exposition of the current state of the art solutions with cnn in cardiac MRI. Several key elements in the development of cnn for segmentation are discussed, including, selection of loss functions, how to tackle the segmentation problem when small datasets are available, the overfitting problem, approaches for better segmentation depending on the target within the images, and best practices for implementation of segmentation architectures in cine cardiac MRI.
