Page 45 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

Chapter 2 Deep convolutional neural network in medical image processing 33

diseases. The extracted images using different modalities as pre-
sented in Table 2.1 are used as input to the ML/DL model.
Currently, DL, especially CNN has proven to be an effective
advancement of ML for the analysis of medical imaging. There-
fore, in the subsequent section, the architecture of different
CNN followed by its application to the analysis of different dis-
eases encountered in different organs of the human body is
explicitly presented.

3. Convolutional neural network and its
architectures

DL is a subset of ML in which several linear, as well as
nonlinear, processing units are organized in a deep layer design
so that it could model the abstraction in the data [6]. At present,
many DL techniques are in use for the diverse area of applications.
Different DL techniques include Boltzman machines, autoen-
coders, recurrent neural network, deep belief networks, and deep
CNNs. Among these techniques, CNN-based models have more
applications in the domain of medical image processing.
CNNs are named so due to the presence of convolutional
layers in their construction [38]. Detection of certain local fea-
tures in every location of the particular input image is the main
work of the convolutional layers. Each of the nodes in a convolu-
tional layer is linked with only a small subset of spatially con-
nected neurons for the detection of the local features in the
input image. To aid in the search of the same local features all
through the input channels, the weights of the connection are
shared among the nodes of the convolutional layers. Each set of
these shared weights is named as the convolutional kernel or sim-
ply kernel. These convolutional layers along with kernels gradu-
ally learn to detect local features whose strength throughout the
input images is evident in the final feature maps. To achieve a hi-
erarchical set of features and to lessen computational complexity,
each of the convolutional layer sequences is followed by a pooling
layer. The pooling layer is evocative of simple as well as complex
cells in the primary visual cortex [39]. The max-pooling layer plays
a major role in reducing the feature maps' size by choosing the
maximum feature response in local neighborhoods and removing
the exact location of such maximum responses. As a result of this
process, max-pooling further improves translational invariance.
A typical CNN model comprises multiple convolutional and
pooling layer pairs followed by several consecutive fully con-
nected layers as presented in Fig. 2.5. Finally, the network ends