Page 25 - Artificial Intelligence for Computational Modeling of the Heart

xxvi Preface

rhythmias [6,8,9,19–21], valve surgery planning [22,23]orcongen-
ital heart disease treatments [24,25]. Yet, physiological models are
still built and used by modelers and scientists, outside of the clin-
ical environment. They are also still computationally demanding
and complex to operate, making their large scale, prospective clin-
ical validation difficult. Finally, the effects of data noise, parameter
uncertainty and model assumptions still need to be properly as-
sessed [26,27].

The new age of artificial intelligence
The concept of intelligent machines can be traced back to the
origin of human made machines. However, the term “artificial
intelligence” (AI) was first introduced by John McCarthy, one of
the founding father of AI, during the Dartmouth conference in
1956 [28]. Since then, the AI community has gone through several
waves of hype and winter periods. Nowadays, AI is a core technol-
ogy in many fields, powered by what is called deep learning. Deep
learning methods rely on artificial neural networks (ANN) with a
very large number of layers. While ANNs were first proposed in
1943 [29], their power could be unlocked only recently thanks to
the availability of large amount of data along with the tremendous
computational power offered by high-performance computing in-
frastructures. In computer vision for instance, one could consider
the 2012 AlexNet convolutional neural network [30]asone ofthe
main trigger of modern deep learning, along with the ImageNet
database (2009), which facilitated an exponential development in
computer vision technologies.
AI did not wait long to penetrate medical imaging. Machine
learning methods were already explored in the 2000’s for organ
detection, contouring and tracking [31,32]. As in computer vision
deep learning has provided a significant boost in performance for
each of these core image processing tasks, but also image recon-
struction [33]orevenscannerautomation.Asanexample,anAI
algorithm has been shown to be able to estimate an avatar of a
patient laying on the table of a computed tomography scanner.
The avatar is then used by the system to perform automatic iso-
centering or scan range estimation to image a specific organ [34,
35]. As large databases with outcome data are becoming avail-
able, researchers are investigating how AI could help estimating
new risk scores with significantly better sensitivity and specificity
than the state of the art. For instance, in [36], the authors show
that an AI algorithm can calculate a risk score for radiotherapy re-
sponse that is 45% more discriminative than tumor size only, the
current criterion for clinical decision making. With AI, workflows