Page 28 - Artificial Intelligence for Computational Modeling of the Heart

Preface xxix

on machine learning methods that scan the image and calculate
a probability for each voxel to belong to the anatomy of interest
(e.g. myocardium). As a result, these methods suffer from inherent
limitations related to the efficiency in scanning high-dimensional
parametric spaces and the learning of representative features for
describing the image content. This chapter introduces new tech-
niques for cardiac image parsing and structure tracking. First, the
marginal space learning framework is introduced, including the
original version of the system that relies on handcrafted steer-
able features, as well as the modern redesign of the framework
based on the latest automatic feature learning technology us-
ing deep learning. We then introduce the concept of intelligent
image parsing based on deep reinforcement learning and scale-
space theory, which overcomes the inherent limitations of the
previous methods. Third, a modern deep image-to-image neu-
ral network architecture for whole heart isolation is presented.
Finally, a review of cardiac structure tracking approaches based
on convolutional neural networks and recurrent neural networks
is provided.
Chapter 4: Data-driven reduction of cardiac models. A common
challenge in physiological modeling is to find the right trade-off
between model fidelity, and hence complexity, and computational
efficiency. Digital twins can be integrated into clinical workflows
only if the data processing is time efficient and user-friendly. In
other words, all complexity needs to be hidden from the user.
In the ideal case, detailed and individualized physics phenom-
ena should be computed in real-time, to allow the user to in-
teract with the virtual heart. Unfortunately, current approaches
to real-time simulations come with significant model simplifica-
tions, which may not accurately reflect the complex physiology
any more. Artificial intelligence could open new ways of simpli-
fying models, while retaining their descriptive power. This chap-
ter describes data-driven approaches for model reduction, with
the aim of retaining the ability of the model to describe complex
physical phenomena, while at the same time drastically reducing
the computational complexity. Three use cases are presented, cov-
ering cardiac electrophysiology, hemodynamics and biomechan-
ics.
Chapter 5: Machine learning methods for robust parameter es-
timation. The second challenge for accurate patient-specific sim-
ulations is data assimilation. In particular, models need to be fit-
ted to patient data in order to reproduce the observed physiology.
In clinical practice, data is often sparse, noisy and incomplete.
For instance, 12-lead electrocardiograms only provide an indirect