2.2 Machine Learning in CAD
In the field of computer-aided diagnosis on lung nodule detection, there are many good
methods on extraction of features and classification for benign and malignant nodules. In this chapter, we propose to select representative eigenvalues to classify pulmonary nodules based on the common characteristic parameters of pulmonary nodules.
This process directly determines the data integrity and accuracy of benign and
malignant discrimination [5].
The support vector machine (SVM) is based on linear separability and is gradually extended to a linearly inseparable case. SVM is essentially a nonlinear mapping,
which transforms the nonlinear problem in the original sample space into a linear
separable problem in the feature space [6]. The SVM applies the kernel function
expansion theorem without having to know the explicit expression of the nonlinear
mapping. With the help of the kernel function, the linear learning machine in the highdimensional feature space does not increase the computational complexity.
The extreme learning machine (ELM) is a fast single hidden layer neural network
(SLFN) training algorithm. When using the ELM algorithm, only the number of hidden
layer neurons needs to be set, and the output weight needs to solve the Moore-Penrose
generalized inverse problem of a matrix. According to this feature, no iteration is
required to determine the parameters [7], thus reducing the overall tuning time.
Therefore, compared with the traditional training methods of machine learning, ELM is
fast and generalized.
2.3 Deep Learning in CAD
Convolutional neural network (CNN) is a neural network with a multilayer structure,
typically model shown as Fig. 3. Each level of CNN contains one or more neurons, and
the relationships between these neurons are independent of each other. The performance of the CNN algorithm depends on two factors, one is the performance of the
computer hardware, and the other is a large number of image data sets. When using the
traditional method for the diagnosis of pulmonary nodules [8], the chest CT object
needs to be analyzed to find the CT image of the chest corresponding to the pulmonary
nodules, and the lung nodule area of the chest CT image is marked; then, the characteristics of the lung nodule area are extracted, and the extracted features have the
degree of solidity, and the center distance, texture, and density are used to model the
features. Finally, a large number of feature models are used to train the classifier to
obtain the final classification result. The use of CNN in these situations typically
produces a more accurate classification than the handcrafted features of a standard
classifier, and CNN learns the most meaningful features of a given task itself. The
convolutional layer in the CNN may also be referred to as a feature extraction layer,
that is, a C layer [9]. The pooling layer in the CNN may also be referred to as a
downsampling layer, an S layer. CNN network has strong earning ability, mainly
reflected in its multiple hidden layers. The inclusion of a large number of neurons in
multiple hidden layers enables the CNN to process a large number of data sets, and the
downsampling operation can assist the CNN in reducing the complexity of training.
The structure of the traditional neural network is different from the structure of the
500
Y. Li et al.
In the field of computer-aided diagnosis on lung nodule detection, there are many good
methods on extraction of features and classification for benign and malignant nodules. In this chapter, we propose to select representative eigenvalues to classify pulmonary nodules based on the common characteristic parameters of pulmonary nodules.
This process directly determines the data integrity and accuracy of benign and
malignant discrimination [5].
The support vector machine (SVM) is based on linear separability and is gradually extended to a linearly inseparable case. SVM is essentially a nonlinear mapping,
which transforms the nonlinear problem in the original sample space into a linear
separable problem in the feature space [6]. The SVM applies the kernel function
expansion theorem without having to know the explicit expression of the nonlinear
mapping. With the help of the kernel function, the linear learning machine in the highdimensional feature space does not increase the computational complexity.
The extreme learning machine (ELM) is a fast single hidden layer neural network
(SLFN) training algorithm. When using the ELM algorithm, only the number of hidden
layer neurons needs to be set, and the output weight needs to solve the Moore-Penrose
generalized inverse problem of a matrix. According to this feature, no iteration is
required to determine the parameters [7], thus reducing the overall tuning time.
Therefore, compared with the traditional training methods of machine learning, ELM is
fast and generalized.
2.3 Deep Learning in CAD
Convolutional neural network (CNN) is a neural network with a multilayer structure,
typically model shown as Fig. 3. Each level of CNN contains one or more neurons, and
the relationships between these neurons are independent of each other. The performance of the CNN algorithm depends on two factors, one is the performance of the
computer hardware, and the other is a large number of image data sets. When using the
traditional method for the diagnosis of pulmonary nodules [8], the chest CT object
needs to be analyzed to find the CT image of the chest corresponding to the pulmonary
nodules, and the lung nodule area of the chest CT image is marked; then, the characteristics of the lung nodule area are extracted, and the extracted features have the
degree of solidity, and the center distance, texture, and density are used to model the
features. Finally, a large number of feature models are used to train the classifier to
obtain the final classification result. The use of CNN in these situations typically
produces a more accurate classification than the handcrafted features of a standard
classifier, and CNN learns the most meaningful features of a given task itself. The
convolutional layer in the CNN may also be referred to as a feature extraction layer,
that is, a C layer [9]. The pooling layer in the CNN may also be referred to as a
downsampling layer, an S layer. CNN network has strong earning ability, mainly
reflected in its multiple hidden layers. The inclusion of a large number of neurons in
multiple hidden layers enables the CNN to process a large number of data sets, and the
downsampling operation can assist the CNN in reducing the complexity of training.
The structure of the traditional neural network is different from the structure of the
500
Y. Li et al.