Page 46 - Handbook of Deep Learning in Biomedical Engineering Techniques and Applications

34 Chapter 2 Deep convolutional neural network in medical image processing

Pooled Feature
Feature Maps Maps Fully Connected 1
Feature Maps Pooled Feature ( PY X )
Maps

Input

Pooling 1
Pooling 2
Convolution Convolution
Layer 1 Layer 2
Figure 2.5 General architecture of CNN. CNN, convolutional neural network.

with a regression or softmax layer so as to generate the desired re-
sults. In modern architecture for more computational efficiency,
the pooling layer is replaced by a convolutional layer with a stride
larger than one. The major advantage of CNN is the backpropaga-
tion of the error signal attained by the loss function, so as to
improve the feature learning, and thus, the CNN has better repre-
sentation. Another strength of the CNN model is that the initial
layers capture blobs, edges, and some local structures, whereas
the nodes in the higher layers give more importance to different
portions of individual organs of the human body, and few neurons
in the end layers consider the whole organ [40].

3.1 Architectures of deep convolutional neural
network
Given the popularity of CNNs application in medical imaging,
the most commonly used CNN architectures are elaborated in
the following.

3.1.1 General classification architectures
LeNet [41] and AlexNet [42] proposed within a decade gap are
very similar to each other. Both of these networks were compara-
tively shallow in which LeNet comprises two and AlexNet com-
prises five convolutional layers. Both of the designs used kernels
having large receptive fields in initial layers and smaller kernels
in the final layers. Instead of using hyperbolic tangent as the acti-
vation function, AlexNet incorporated activation function such as
rectified linear units.