CNN. All nodes of each input layer and all nodes of the hidden layer are connected to
each other. All nodes of the hidden layer and all nodes of the output layer are also
connected to each other [10].
Moreover, when calculating the output of the neural network, the scale of the
matrix to be calculated is relatively large. When the scale of the matrix is large to a
certain extent, the convergence time is too long, and it is easy to fall into the local
extreme value. In response to these shortcomings, the convolutional neural network
(CNN) uses local perceptual fields and weight sharing to shorten training time and
improve training accuracy.
Improved convolution neural network method. Local perception field is one of
the methods to effectively reduce the number of weights and the number of parameters
without reducing the accuracy [11]. It is generally believed that the human brain’s
perception of the real world is a process from the local to the global. For the discussion
of images, you can follow this process. The condition in the image that best describes
the relationship of pixel points is spatial location of the image.
In this research, we proposed an improved method in convolutional neural network,
for extending the local perception field by using multi-scale convolution kernels as
shown in Fig. 4.
The improved kernels are based on shapes of 1 Â 1, 3 Â 3, 5 Â 5. Later two are
similar to those working in CNN, and the 1 Â 1 kernel works most importantly. The
1 Â 1 convolution kernel mainly performs the function of full-join calculation through
the convolution operation, thereby extracting features at a higher scale. When a convolutional layer enters a large number of features, convolution of this input will generate a huge amount of computation; if the input is first performed, the number of
convolutions will be significantly reduced after reducing the number of features. Then,
we need to solve the matter of information combination. Different characters are
produced from different scales kernel, these branches are convolved or merged with
convolution kernels of different sizes and finally stitched together in the feature
dimension. More feature richness also means more accurate judgment in the final
classification [12] which is shown in Fig. 5.
64
64
52
52
20
26
26
20
22
22
20
11
11
20
20
20
Convolution
Max
Pooling
Convolution
Max
Pooling
Max
Pooling
Convolution
Fully
connected
8
8
Fig. 3. Architecture of lung nodules diagnosis based on CNN
Research of Lung Cancer-Assisted Diagnosis Algorithm Based …
501
each other. All nodes of the hidden layer and all nodes of the output layer are also
connected to each other [10].
Moreover, when calculating the output of the neural network, the scale of the
matrix to be calculated is relatively large. When the scale of the matrix is large to a
certain extent, the convergence time is too long, and it is easy to fall into the local
extreme value. In response to these shortcomings, the convolutional neural network
(CNN) uses local perceptual fields and weight sharing to shorten training time and
improve training accuracy.
Improved convolution neural network method. Local perception field is one of
the methods to effectively reduce the number of weights and the number of parameters
without reducing the accuracy [11]. It is generally believed that the human brain’s
perception of the real world is a process from the local to the global. For the discussion
of images, you can follow this process. The condition in the image that best describes
the relationship of pixel points is spatial location of the image.
In this research, we proposed an improved method in convolutional neural network,
for extending the local perception field by using multi-scale convolution kernels as
shown in Fig. 4.
The improved kernels are based on shapes of 1 Â 1, 3 Â 3, 5 Â 5. Later two are
similar to those working in CNN, and the 1 Â 1 kernel works most importantly. The
1 Â 1 convolution kernel mainly performs the function of full-join calculation through
the convolution operation, thereby extracting features at a higher scale. When a convolutional layer enters a large number of features, convolution of this input will generate a huge amount of computation; if the input is first performed, the number of
convolutions will be significantly reduced after reducing the number of features. Then,
we need to solve the matter of information combination. Different characters are
produced from different scales kernel, these branches are convolved or merged with
convolution kernels of different sizes and finally stitched together in the feature
dimension. More feature richness also means more accurate judgment in the final
classification [12] which is shown in Fig. 5.
64
64
52
52
20
26
26
20
22
22
20
11
11
20
20
20
Convolution
Max
Pooling
Convolution
Max
Pooling
Max
Pooling
Convolution
Fully
connected
8
8
Fig. 3. Architecture of lung nodules diagnosis based on CNN
Research of Lung Cancer-Assisted Diagnosis Algorithm Based …
501