Page 293 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

284 Chapter 10 Deep neural network in medical image processing

time, whereas performance of deep learning algorithms increases
over time as we feed them more and more data. Pioneers of deep
learning Benjio et al. explain in their very famous review paper
that deep learning can be successfully applied to the field of im-
age processing.
• CNNs or ConvNets are designed in a way to process data that
come in the form of multiple arrays, for example, a complex
medical image. Much like machine learning network, it con-
sists of different layers of neurons where each neuron performs
a simple mathematical function and passes it onto the next
layer. ConvNets are very suitable for image processing tasks.
• 3D CNNs and normal 2D convolutional networks as
mentioned in the earlier paragraph hit a roadblock when
dealing with 3D images such as MRI data or a CT scan. In
this type of network architecture, image is processed as
four-dimensional array.
• Fully convolutional neural network is basically a CNN without
any fully connected layers. In this type of network, any layer is
used for learning filters, even the end of the network decision-
making layers.
• A recurrent neural network is a type of ANN, which can take
input with no predetermined limit; basically, it remembers
the past decisions. It involves a hidden state vector based on
the prior input/output.
• A deep belief network is a type where layers are formed with
variable with random probability distribution, i.e., stochastic.

3.1 Deep learning architectures
Various types of deep learning architectures used in image
processing applications in medical domain are as follows:
• Torch [35] is an old machine learning library developed by
Collobert et al. It was initially developed in Lua programming
language although having an internal core written in C lan-
guage. However, a Python wrapper was released in 2019, which
supports deep learning algorithms.
• Keras is a fully functional deep learning framework for Python
language.The USP of Keras is that it uses the same code for
GPU and CPU. It was developed by FranÃxois Chollet from
Google.
• Theano [39] is an open-source deep learning library developed
in 2007 by the University of Montreal. It supports both GPU
and CPU operations.
• Mxnet [37] is an existing new open-source deep learning
library developed by Apache. It is accessible from many