Page 141 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 3 Learning cardiac anatomy 113

used to drive the training. Experiments conducted on a 3D-CT
dataset of 896 cases, using 796 cases for training (10% of these for
validation) and the remaining 100 cases for testing demonstrate
competitive advantages with an average Dice score of 0.93. Fig. 3.8
shows two slices of one test case, with the heart isolation mask as
a transparent overlay.

3.3 Structure tracking

Structure tracking is a fundamental subject in cardiac imaging
research. Efficient and robust tracking on the motion and defor-
mation of the heart serves a large variety of clinical applications
including myocardial strain measurement, anomalous state de-
tection, infarction prediction, real-time image-guided interven-
tion and cardiovascular surgery planning. Recent advances in
imaging technologies allow cardiologists to capture morpholog-
ical and functional information of complex structures and their
dynamic changes effectively. For example, echocardiography cap-
tures the cardiac motion in real time and provides important guid-
ance during surgeries for valve repairments. X-ray angiography
serves as the primary modality in percutaneous coronary inter-
ventions (PCI) and catheter-based electrical physiology (EP) ther-
apies to precisely visualize and target the surgical object [281].
While these imaging techniques have achieved great success in
clinical practice, they create challenges for the research commu-
nity to develop models to extract and process important struc-
tural as well as temporal information from the captured dynamic
images. Such challenges include sophisticated image conditions
(clutters, illumination and noise), complex anatomy motion and
deformation, partial/full object occlusions, and real-time process-
ing requirements.
In computer vision literature, object visual tracking has been
extensively studied. Many different approaches have been pro-
posed in the past decades, see [282–285] for detailed reviews.
Tracking methods can be classified into three categories based on
the representations of the objects: point tracking, kernel-based
tracking and silhouette tracking models. Point tracking models
aim to find correspondences on the detected key points of the tar-
get object at each frame. Examples of key points include object
centroid, critical landmarks and point cloud on object bound-
aries. Kernel-based tracking models mask the target object with
carefully designed kernels and search for representation match-
ings with efficient methods such as mean-shift. Silhouette track-
ing is closely related to kernel tracking. It performs shape or con-
tour matchings of target object cross frames. A substantial limi-