Page 143 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 3 Learning cardiac anatomy 115

directional RNN and uses it to improve correlation filters for de-
tection cross frames. Ondrúška et al. [302] leveraged RNNs as a
feature encoder to reveal object occlusion by learning from raw
sensor data. Kahou et al. [303]and Ganetal.[304] used RNNs and
attention schemes to predict the distribution of the object’s loca-
tions. Binary classifiers are constructed to find the final location.
Similarly, Ning et al. [305] proposed a spatially supervised recur-
rent convolutional neural network to exploit the history of object
locations for object tracking.
Tracking with Reinforcement Learning. Besides landmark
detection and image registration, deep reinforcement learning
has also been applied in visual tracking. Zhang et al. [306]pro-
posed to learn tracking policies with a detection-memorization
CNN-LSTM scheme. The reward is calculated by the detection er-
ror at each frame. Similar to the similarity learning models with
Siamese networks, Choi et al. [307] introduced a reinforcement
learning scheme for tracking based on a template selection strat-
egy. After extracting features of candidate regions and template
via a Siamese network, a policy network is attached to deter-
mine the similarities between them. Instead of taking the output
directly from the last layer of the policy network for decision,
Huang et al. [308] further improved the robustness and speed up
achieved by such model scheme by adding an agent to choose the
decision layer selectively. Yun [309] proposed an Action-Decision
Network to use pre-trained networks to control actions of the
agents and add online fine-tuning during tracking.
Other methods such as correlation filtering [297,310–312]and
the more recent adversarial learning [313] have also demonstrated
great value in object tracking. While these models have been suc-
cessfully applied in object tracking in natural scenes, how to ef-
fectively adapt these models in the medical domain, especially for
cardiac activity tracking, is still an open question.

3.4 Summary

This chapter presented several methods for cardiac anatomy
detection in medical images. This includes classification-based
solutions, such as the marginal space (deep) learning frameworks,
and a reinforcement learning-based solution referred to as in-
telligent multi-scale image navigation. Through extensive experi-
ments, we demonstrated the power of deep learning architectures
in capturing the image information and performing prediction.
We also analyzed the significant advantages of approaching the
detection problem as a reinforcement learning task, highlighting
the improved precision, robustness and in particular detection