Page 125 - Artificial Intelligence for Computational Modeling of the Heart

P. 125

Learning cardiac anatomy

Florin C. Ghesu, Bogdan Georgescu, Yue Zhang, Sasa Grbic,
Dorin Comaniciu
Siemens Healthineers, Princeton, NJ, United States

3.1 Introduction

Parsing of cardiac and vascular structures is an essential pre-
requisite for cardiac image analysis and represents the ﬁrst step in
constructing a comprehensive heart model. In practice, parsing
refers to the detection, segmentation and tracking of anatomical
structures. This information enables not only the direct deriva-
tion of measurements (distances, volumes, ejection fraction, etc.)
for quantitative assessment, but also the estimation of physiolog-
ical heart models or the mechanical simulation of blood ﬂow. The
features extracted during the parsing step are often critical for the
timely diagnosis of acute conditions, serving as input for artiﬁcial
intelligence methods for diagnosis or interventional guidance.
In this chapter, we present several state-of-the-art methods for
cardiac image parsing. We cover the marginal space learning [31]
and marginal space deep learning [257,258] frameworks, demon-
strating their performance at detecting and segmenting various
structures, such as heart chambers and valves based on 3D com-
puted tomography (CT) and 3D ultrasound (US) images. We also
present the concept of multi-scale image navigation, as an efﬁ-
cient alternative to exhaustive image scanning. The accuracy and
speed of this new framework at detecting various cardiac land-
marks, as well as vascular landmarks in the body, are reported
on several 2D magnetic resonance (MR), 2D-US and 3D-CT im-
age datasets. We also present a modern deep image-to-image fully
convolutional segmentation network and report its performance
at segmenting the entire heart. Finally, we discuss the structure
tracking problem in cardiac modeling and review the state-of-the-
art deep learning based methods.

120 121 122 123 124 125 126 127 128 129 130