CN109389171B - Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology - Google Patents

Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology Download PDF

Info

Publication number
CN109389171B
CN109389171B CN201811181783.2A CN201811181783A CN109389171B CN 109389171 B CN109389171 B CN 109389171B CN 201811181783 A CN201811181783 A CN 201811181783A CN 109389171 B CN109389171 B CN 109389171B
Authority
CN
China
Prior art keywords
medical image
vector
function
image
noise reduction
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201811181783.2A
Other languages
Chinese (zh)
Other versions
CN109389171A (en
Inventor
杨云
曹丽娟
杨珀
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yunnan University YNU
Original Assignee
Yunnan University YNU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Yunnan University YNU filed Critical Yunnan University YNU
Priority to CN201811181783.2A priority Critical patent/CN109389171B/en
Publication of CN109389171A publication Critical patent/CN109389171A/en
Application granted granted Critical
Publication of CN109389171B publication Critical patent/CN109389171B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing

Abstract

The invention discloses a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder technology, which introduces a convolution neural network on the basis of a traditional automatic encoder; meanwhile, convolution kernels with different sizes are arranged on the same convolution layer, random Gaussian noise is added in a training stage, finally, a deep neural network of three multi-granularity convolution noise reduction automatic encoders is trained on the basis of a greedy layer-by-layer training method, and parameter global optimization is finally achieved after fine adjustment. The invention improves the classification precision, achieves the optimal classification result, has stronger anti-interference capability to the complex environment, and solves the problems of difficult acquisition of data labels, more resource consumption, single extraction characteristic and low anti-interference capability to the complex environment, which cause obvious reduction of generalization in the prior art.

Description

Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
Technical Field
The invention belongs to the technical field of computer views, and particularly relates to a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder technology.
Background
Image classification is a very active application and research direction in computer vision, and aims to classify various types of images, wherein images with similar characteristics are classified into the same class, and images with different classes are classified into different classes. In the existing research, image classification can be roughly divided into the following steps: (1) preprocessing an image; (2) extracting characteristics; (3) selecting characteristics; (4) and (4) selecting and designing a classifier. In these several steps, feature extraction plays a key role, and determines the accuracy of the final classification result and the stability of performance. In past research, the process of feature extraction was typically to extract local or global features, such as: local Binary Pattern (LBP), Histogram of Oriented Graphs (HOG), Scale Innovative Feature Transform (SIFT). In recent years, researchers focus on the deep learning with the fastest development along with the rise of machine learning, because the existing research shows that a deep learning algorithm has the function of simulating the brain structure of a human, low-level features at the bottom layer can be abstracted into high-level features layer by layer, and images can be expressed and described more effectively. Because of the high theoretical research and practical application value of the research field, many researchers at home and abroad propose more medical image classification techniques, but the method roughly comprises three aspects of work: feature learning, feature extraction, and classification. Firstly, performing feature learning on original medical image data; secondly, extracting high-level features of the image, and finally, classifying and analyzing the image based on the extracted high-level features.
Although researchers have achieved some results in the field of image classification, the prior art is not mature due to the minimal differences between the actual medical image classes, the complexity of feature distribution, and the problem of noise interference. There are several major disadvantages. (1) The existing convolutional neural network needs a large amount of data with class marks; the data class targets are difficult to obtain in real life and depend on manual work and expert knowledge seriously, so that a large amount of resources are consumed; (2) in the existing image classification algorithm based on feature extraction, the extracted features are single, and the extracted high-level features are not diverse because the traditional convolutional neural network only uses a convolutional kernel with one size, and finally the accuracy of the algorithm is reduced. (3) Most classification algorithms are limited in stability, and when complex environments such as noise problems are encountered, the generalization performance of the algorithms is significantly reduced.
Therefore, how to reasonably combine different deep learning algorithms to improve the robustness of the model without prior knowledge to achieve the optimal classification result is an urgent problem to be solved.
Disclosure of Invention
The invention aims to provide a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder technology, which is used for constructing a medical image classifier with strong generalization capability and high anti-interference performance and simultaneously fully exerting the performance of each sub-component in the aspect of feature extraction; the method solves the problems that in the prior art, the generalization is obviously reduced due to the fact that the data class labels are difficult to obtain, the consumed resources are large, the extracted features are single, and the anti-interference capability to the complex environment is low.
The technical scheme adopted by the invention is to provide a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder technology, which comprises the following steps:
step 1), medical image preprocessing: calling an OpenCV toolkit, adjusting the size of the medical image to 256 multiplied by 256, carrying out normalization processing on the adjusted image, taking the normalized image as an image data set, and dividing the image data set into a medical image training set and a medical image testing set according to the medical image number proportion of 8: 2;
step 2), establishing a medical image classification model:
a. coding a single medical image by using a network structure of a coding part of a multi-granularity convolution noise reduction automatic coder; the coding part of the multi-granularity convolution noise reduction automatic coder simultaneously adopts convolution kernels with the sizes of 1 × 1, 3 × 3 and 5 × 5, and a single medical image with the size of 256 × 256 after being preprocessed in the step 1), namely an input vector x with the corresponding dimension of 65536 is inputiAdding Gaussian noise to obtain the data characteristics of the damaged medical image, wherein the data characteristics of the damaged medical image are defined as interference input vectors
Figure BDA0001825173640000027
And then linear mapping is carried out to enter a nonlinear activation function to obtain a feature expression vector alpha of the middle layeri：
Figure BDA0001825173640000021
Wherein alpha isiInputting vectors as disturbances
Figure BDA0001825173640000022
Obtaining a middle layer characteristic representation vector after abstraction of the two hidden layers, wherein the corner mark i is the dimension of the image in the medical image training set; x is the number ofiIs an input vector;
Figure BDA0001825173640000023
obtaining a damaged interference input vector, namely an interference input vector, for an original clean input vector after noise interference is added; w1Weight matrix from input layer to intermediate layer for multi-granularity convolution noise-reduction auto-encoder, b1A bias vector from an input layer to an intermediate layer of the auto-encoder;
Figure BDA0001825173640000024
the function represented is a coding function,
Figure BDA0001825173640000025
the expressed function is the activation function of the neuron, and a sigmoid function or an identity function ReLu is usually taken;
b. based on the complete symmetry of the network structure of the decoding part and the encoding part of the multi-granularity convolution noise reduction automatic encoder, the network structure of the decoding part of the automatic encoder is used for decoding operation to high-level characteristics of the middle layer, namely interference input vectors
Figure BDA0001825173640000026
Performing decoding operation to the intermediate characteristic vector alpha passing through the hidden layeriOutputting a reconstructed image of the damaged medical image by linear mapping into a nonlinear activation function, wherein the reconstructed image of the damaged medical image corresponds to a reconstructed output vector zi：
zi＝g(αi)＝l(W2αi+b2)； (2)
Wherein z isiFor the reconstructed output vector, i.e. the disturbance input vector
Figure BDA0001825173640000031
One reconstruction of (2); g (. alpha.) (alpha.)i) As a decoding function, W2Weight matrix from the middle layer to the output layer of the automatic encoder, b2Is the offset vector from the middle layer to the output layer of the auto-encoder; l (W)2αi+b2) The expressed function is the activation function of the neuron;
c. establishing a cost function, and measuring the reconstruction performance of the automatic encoder: based on the foregoing steps a and b, an interference-based input vector is established
Figure BDA0001825173640000032
And an output vector ziA cost function with reconstruction error as variable, and minimizing the cost function to make the interference input vector
Figure BDA0001825173640000033
And the reconstructed output vector ziSimilarly; the cost function L (X, Z) is shown as follows, the smaller the loss value L of the cost function is, the more accurate the characteristic extraction of the encoder part is;
Figure BDA0001825173640000034
wherein X represents an input medical image training set and Z represents output reconstructed medical image data;
d. and (3) performing summation calculation of loss values on all medical images in the medical image training set by using a formula (3), and introducing regularization constraint to obtain an overall cost function J (W, b) of the multi-granularity convolution noise reduction automatic encoder on the medical image training set as follows:
Figure BDA0001825173640000035
wherein J is the loss value of the image in the whole medical image training set, W and b are the weight matrix and the offset vector of the whole multi-granularity convolution noise reduction automatic encoder network respectively,
Figure BDA0001825173640000036
the method comprises the following steps of (1) determining the number of images in a medical image training set as a regularization term, wherein lambda is a penalty factor, and M is the number of images in the medical image training set;
e. repeating the steps a to d for three times based on a greedy layer-by-layer pre-training method to obtain a depth neural network formed by stacking three multi-granularity convolution noise reduction automatic encoders, namely forming the depth neural network formed by stacking the three multi-granularity convolution noise reduction automatic encoders;
f. the decoder part of the automatic deep multi-granularity convolution noise reduction encoder is removed, a softmax classifier is connected behind the residual deep neural network, the advanced features of the medical images extracted by the encoder part of the automatic deep multi-granularity convolution encoder are classified, and the functional expression of the softmax classifier is as follows:
Figure BDA0001825173640000041
wherein, wiIs an input vector xiThe weight vector of (1), M is the number of images of the medical image training set, and k is the number of categories of different medical image data sets;
Figure BDA0001825173640000042
the function f (x) is a scoring function and represents the similarity degree of the characteristics of medical image response and a certain disease class, so that the function f (x) is called as normalized probability and represents the probability of predicting the medical image to be confirmed as a certain disease class; e is a numerical constant of an exponential function;
g. fine tuning to obtain the parameter W of deep neural network1、W2、b1、b2An optimal value is reached;
step 3), inputting the images in the medical image test set into the medical image classification model established in the step 2), and judging the disease categories to which the medical images belong according to the classification precision on the test set, wherein the higher the classification precision is, the higher the accuracy of the disease categories to which the medical images belong is, and the purpose of accurate classification is achieved; the expression formula of the classification precision is as follows:
Figure BDA0001825173640000043
wherein f is a medical image based on a multi-granularity convolution noise reduction automatic encoderThe medical image classification method includes D is medical image test set, N is number of images in the medical image test set, and xiFor the ith image in the medical image test set D, yiFor the true class label of the ith image in the medical image test set D, f (x)i) The prediction class mark is a prediction class mark obtained after feature learning is carried out on a medical image by a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder.
The invention has the beneficial effects that:
(1) the feature learning method based on unsupervised learning applies the convolution operation in the convolutional neural network to the network structure of the traditional automatic encoder, has higher classification precision than a single technology, and achieves the effect of sparse connection; meanwhile, the number of network parameters is obviously reduced due to the characteristic of convolution weight sharing, and the calculated amount is reduced; and the convolutional neural network which seriously depends on the labeled data is applied to the field of unsupervised learning, so that the development of the convolutional neural network is widened.
(2) The method for extracting the advanced features by the multi-granularity convolution kernel ensures the diversity of the extracted advanced features on the basis of efficiently extracting the features, can comprehensively represent the key features of the image, is favorable for improving the classification precision and achieves the optimal classification result.
(3) The invention has stronger anti-interference capability to complex environment.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
Fig. 1 is a diagram of an autoencoder network architecture.
FIG. 2 is a schematic diagram of a feature learning algorithm for an auto-encoder based on multi-granularity convolution noise reduction.
FIG. 3 is a schematic diagram of a depth multi-granularity convolution noise reduction auto-encoder.
Fig. 4 is a comparison graph of classification performance for different noise ratios.
FIG. 5 is an image under a cerebrospinal fluid microscope of meningitis.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
On the meningitis data set, the medical image classification method based on the multi-granularity convolution noise reduction automatic encoder technology disclosed by the invention is well verified. As shown in FIG. 5, meningitis is classified into four groups, i.e., purulent meningitis, tuberculous meningitis, viral meningitis, and cryptococcal meningitis. As can be seen in FIG. 5, the difference between the imaging classes of cerebrospinal fluid under the microscope is very small, and the specific type of meningitis is difficult to judge by people with low expertise or doctors with insufficient clinical experience. Generally, doctors can comprehensively diagnose the type of illness according to the clinical manifestations and cerebrospinal fluid examination of patients and then can carry out corresponding treatment. However, physicians are influenced by personal experience, knowledge, conditions, etc. when observing the imaging of cerebrospinal fluid under a microscope; image data from different hospitals or image data acquired by the same medical equipment with different specifications from the same hospital have certain differences, so that the final diagnosis result can be influenced by the factors. Therefore, the classification method of the present invention needs to be applied for accurate classification.
The specific process is as follows:
1. image preprocessing: the meningitis data set is 3184 x 4230 in original size, so when these image data are obtained, a pre-processing is first performed. In an embodiment of the present invention, the pre-treatment may be divided into three steps: 1) adjusting the size of the image: calling an OpenCV toolkit to perform resizing operation on the meningitis images, wherein in the specific operation, the meningitis images are uniformly processed into 256 × 256 sizes; 2) normalization treatment: in many image processing processes, normalization processing is necessary preprocessing process for the image, namely, the value of the point where the pixel value of the image is distributed between 0 and 255 is distributed between [0 and 1], the convergence speed of model parameters can be accelerated after the normalization processing, and here, normalization operation is also performed on the meningitis image, so that the network parameters are quickly converged in the training process, and the training time is saved; 3) dividing the data set: 148 image meningitis image data sets from a certain three hospitals are randomly divided into a meningitis image training set and a meningitis image testing set according to the image number proportion of 8:2, wherein the meningitis image training set is used for training the medical image classification model, and the meningitis image testing set is used for evaluating the performance of the medical image classification model on a specific meningitis image classification task. In particular, four types of meningitis diseases are uniformly distributed in the meningitis image training set and the meningitis image testing set.
2. Establishing medical image classification modeling: in order to capture the diversified and global characteristics of the medical image and improve the robustness (anti-noise capability) of the model, the invention provides an algorithm model of a multi-granularity convolution noise reduction automatic encoder. The model can be divided into two parts of an encoder and a decoder as a completely symmetrical network structure.
a. Firstly, as shown in fig. 1, an encoder part performs encoding operation on an input meningitis image, and abstracts layer by layer to obtain high-level features of an intermediate layer; namely, a network structure of a coding part of a multi-granularity convolution noise reduction automatic coder is used for coding a single medical image; the coding part of the multi-granularity convolution noise reduction automatic coder simultaneously adopts convolution kernels with the sizes of 1 × 1, 3 × 3 and 5 × 5, and a single medical image with the size of 256 × 256 after being preprocessed in the step 1), namely an input vector x with the corresponding dimension of 65536 is inputiAdding Gaussian noise to obtain the data characteristics of the damaged medical image, wherein the data characteristics of the damaged medical image are defined as interference input vectors
Figure BDA0001825173640000061
And then linear mapping is carried out to enter a nonlinear activation function to obtain a feature expression vector alpha of the middle layeri：
Figure BDA0001825173640000062
Wherein alpha isiInputting vectors as disturbances
Figure BDA0001825173640000063
Obtaining a middle layer characteristic representation vector after abstraction of the two hidden layers, wherein the corner mark i is the dimension of the image in the medical image training set; x is the number ofiIs an input vector;
Figure BDA0001825173640000064
obtaining a damaged interference input vector, namely an interference input vector, for an original clean input vector after noise interference is added; w1Weight matrix from input layer to intermediate layer for multi-granularity convolution noise-reduction auto-encoder, b1A bias vector from an input layer to an intermediate layer of the auto-encoder;
Figure BDA0001825173640000065
the function represented is a coding function,
Figure BDA0001825173640000071
the expressed function is the activation function of the neuron, and is usually a sigmoid function or an identity function ReLu.
b. As shown in fig. 2, the decoding operation is performed on the high-level features of the middle layer to obtain a reconstructed representation of the input meningitis image. I.e. using the network structure of the decoding part of the automatic encoder to perform decoding operation, and the high-level characteristics of the middle layer, i.e. interference input vector
Figure BDA0001825173640000072
Performing decoding operation to the intermediate characteristic vector alpha passing through the hidden layeriBy linear mappingInputting a nonlinear activation function, and outputting a reconstructed image of the damaged medical image, wherein the reconstructed image of the damaged medical image corresponds to a reconstructed output vector zi：
zi＝g(αi)＝l(W2αi+b2)； (2)
Wherein z isiFor the reconstructed output vector, i.e. the disturbance input vector
Figure BDA0001825173640000073
One reconstruction of (2); g (. alpha.) (alpha.)i) As a decoding function, W2Weight matrix from the middle layer to the output layer of the automatic encoder, b2Is the offset vector from the middle layer to the output layer of the auto-encoder;
l(W2αi+b2) The expressed function is the activation function of the neuron.
c. In order to train the symmetric neural network, a cost function needs to be constructed for one image, and the reconstruction error of the meningitis image is used as the cost function. In actual training, the cost function is made as small as possible, so that the classification model is prompted to select optimal parameters to better capture the high-level features of the meningitis image when learning the meningitis image training set. Thus, based on the modeling processes a and b, the input vector is established to be perturbed
Figure BDA0001825173640000074
And an output vector ziA cost function with reconstruction error as variable, and minimizing the cost function to make the interference input vector
Figure BDA0001825173640000075
And the reconstructed output vector ziSimilarly; the cost function L (X, Z) is shown as follows, and the smaller the loss value L of the minimized cost function is, the more accurate the characteristic extraction of the encoder part is;
Figure BDA0001825173640000076
wherein X represents an input medical image training set and Z represents output reconstructed medical image data;
d. establishing an overall cost function for a plurality of images, namely performing summation calculation of loss values on all medical images in a medical image training set, introducing regularization constraint, and obtaining an overall cost function J (W, b) of the multi-granularity convolution noise reduction automatic encoder on the medical image training set as follows:
Figure BDA0001825173640000077
wherein J is the loss value of the image in the whole medical image training set, W and b are the weight matrix and the offset vector of the whole multi-granularity convolution noise reduction automatic encoder network respectively,
Figure BDA0001825173640000081
the method comprises the following steps of (1) determining the number of images in a medical image training set as a regularization term, wherein lambda is a penalty factor, and M is the number of images in the medical image training set;
e. each multi-granularity convolution noise-reduction automatic encoder comprises convolution kernels with three different sizes, after the first multi-granularity convolution noise-reduction automatic encoder is trained, a greedy layer-by-layer pre-training method is used, the training process is repeated, a deep neural network formed by stacking three multi-granularity convolution noise-reduction automatic encoders is obtained, namely the deep neural network formed by stacking the three multi-granularity convolution noise-reduction automatic encoders is formed, so that the classification accuracy is improved, as shown in figure 3, the first multi-granularity convolution noise-reduction automatic encoder is trained firstly, and after the training is finished, the parameters of the first multi-granularity convolution noise-reduction automatic encoder are fixed
Figure BDA0001825173640000082
And
Figure BDA0001825173640000083
and output the intermediate layer
Figure BDA0001825173640000084
As a second multiple granularity convolution noise reduction autoencoderThen training a second multi-granularity convolution noise reduction automatic encoder, and fixing the parameters of the second multi-granularity convolution noise reduction automatic encoder after the training is finished
Figure BDA0001825173640000085
And
Figure BDA0001825173640000086
output of the middle layer
Figure BDA0001825173640000087
Training a deep stacked neural network formed by three multi-granularity convolution noise reduction automatic encoders by analogy and fixing parameters as the input of a third multi-granularity convolution noise reduction automatic encoder
Figure BDA0001825173640000088
And
Figure BDA0001825173640000089
output of the middle layer
Figure BDA00018251736400000810
And as the input of the softmax classifier, training the classifier, and further completing the whole training. Wherein the content of the first and second substances,
Figure BDA00018251736400000811
respectively corresponding to the weight matrix from the input layer to the middle layer of the first multi-granularity convolution noise reduction automatic encoder, the offset vector and the feature expression vector of the middle layer, and in the same way,
Figure BDA00018251736400000812
and
Figure BDA00018251736400000813
for the weight matrix from the input layer to the middle layer of the second multi-granularity convolution noise reduction automatic encoder, the offset vector and the feature expression vector of the middle layer,
Figure BDA00018251736400000814
and
Figure BDA00018251736400000815
and respectively representing a weight matrix from the input layer to the middle layer, a bias vector and a feature representation vector of the middle layer of the third multi-granularity convolution noise reduction automatic encoder.
As shown in fig. 4, the classification accuracy of the three multi-granularity convolution automatic noise reduction encoders after superposition is obviously improved;
f. as shown in fig. 3, the deep neural network decoder part is removed, the output of the remaining encoder part is the high-level features required by the classification of the medical image classification model, and the softmax classifier is connected behind the remaining deep neural network to classify the high-level features of the meningitis image extracted by the encoder part of the deep multi-granularity convolution automatic encoder; the functional expression of the softmax classifier is as follows:
Figure BDA00018251736400000816
wherein, wiIs an input vector xiThe weight vector of (1), M is the number of images of the medical image training set, and k is the number of categories of different image data sets;
Figure BDA0001825173640000091
the function f (x) is a scoring function and represents the similarity degree of the characteristics of medical image response and a certain disease class, so that the function f (x) is called as normalized probability and represents the probability of predicting the medical image to be confirmed as a certain disease class; e is a numerical constant of an exponential function;
g. parameter W of deep neural network1、W2、b1、b2The local optimal value is achieved after pre-training, however, the classification accuracy of the specific disease category of the meningitis image is not the highest, the network needs to be finely adjusted, and after fine adjustment, the parameter W of the network1、W2、b1、b2The global optimum is reached, and the medical image classification model is aligned with the brain at the momentThe classification accuracy of the disease type to which the membranitis image belongs reaches the highest value.
3. And (3) prediction: in order to evaluate the generalization performance of the medical image classification model of the present invention, experiments need to be performed on a meningitis image test set, and meanwhile, in order to verify the performance difference of the classification model of the present invention and other classification models on the meningitis medical images, the medical image classification method of the present invention and other classification methods are compared with each other on the basis of the meningitis medical image test set. The classification accuracy is expressed by a classification accuracy formula as follows:
Figure BDA0001825173640000092
wherein f is a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder, D is a medical image test set, N is the number of images in the medical image test set, and xiFor the ith image in the medical image test set D, yiFor the true class label of the ith image in the medical image test set D, f (x)i) The prediction class mark is a prediction class mark obtained after feature learning is carried out on a medical image by a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder.
The experimental result shows that the classification accuracy of the method on the meningitis medical image test set reaches 72.83%, which is obviously higher than that of other algorithm models, as shown in table 1. The classification method of the invention has high accuracy in medical image classification.
TABLE 1 Classification accuracy comparison of different algorithms for meningitis medical images
Figure BDA0001825173640000101
In addition, in the experiment of the meningitis medical image data set, the invention adopts a deep neural network formed by stacking three multi-granularity convolution noise reduction automatic encoders. The superposition of the present invention and three single-granularity convolution automatic encoders of different convolution kernel sizes are respectively classified and compared in the meningitis medical image set to obtain different classification accuracy contrast data, as shown in table 2.
TABLE 2 comparison of accuracy of meningitis medical image set tests based on different granularity algorithms
Figure BDA0001825173640000102
As can be seen from Table 2, compared with the convolution automatic encoder with single-granularity convolution kernel, the design of the multi-granularity convolution kernel can capture the comprehensive high-level features of the image, which is beneficial to improving the classification accuracy of the classification model.
In order to improve the robustness of the classification model, Gaussian noise is added in the training process of the classification model, so that the medical image classification model of the invention is forced to learn high-level features with robustness from damaged medical images, and in the following experiment, after the Gaussian noise is added in contrast, classification comparison is carried out on a meningitis image data set by an algorithm, contrast data with different classification accuracies are obtained, and the effectiveness of the Gaussian noise is verified. Where CDAE is a convolution noise reduction auto-encoder with a single granularity convolution kernel.
TABLE 3 test accuracy comparison of different algorithms after noise addition
Figure BDA0001825173640000111
As can be seen from table 3, compared with the case that no gaussian noise is added, after the gaussian noise is added, the classification accuracy of the convolutional automatic encoder on the meningitis image data set is obviously improved, which verifies that the addition of the gaussian noise energy enables the algorithm to extract advanced features with better robustness in the training process of the medical image, so that the classification accuracy is improved.
The above description is only for the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.

Claims (1)

1. A medical image classification method based on a multi-granularity convolution noise reduction automatic encoder technology is characterized by comprising the following steps:
step 1), medical image preprocessing: calling an OpenCV toolkit, adjusting the size of the medical image to 256 multiplied by 256, carrying out normalization processing on the adjusted image, taking the normalized image as an image data set, and dividing the image data set into a medical image training set and a medical image testing set according to the medical image number proportion of 8: 2;
step 2), establishing a medical image classification model:
firstly, coding a single medical image by using a network structure of a coding part of a multi-granularity convolution noise reduction automatic coder; the coding part of the multi-granularity convolution noise reduction automatic coder simultaneously adopts convolution kernels with the sizes of 1 × 1, 3 × 3 and 5 × 5, and a single medical image with the size of 256 × 256 after being preprocessed in the step 1), namely an input vector x with the corresponding dimension of 65536 is inputiAdding Gaussian noise to obtain the data characteristics of the damaged medical image, wherein the data characteristics of the damaged medical image are defined as interference input vectors
Figure FDA0003012853440000011
And then linear mapping is carried out to enter a nonlinear activation function to obtain a feature expression vector alpha of the middle layeri：
Figure FDA0003012853440000012
Wherein alpha isiInputting vectors as disturbances
Figure FDA0003012853440000013
Obtaining a middle layer characteristic representation vector after abstraction of the two hidden layers, wherein the corner mark i is the dimension of the image in the medical image training set; x is the number ofiIs an input vector;
Figure FDA0003012853440000014
obtaining a damaged interference input vector, namely an interference input vector, for an original clean input vector after noise interference is added; w1Weight matrix from input layer to intermediate layer for multi-granularity convolution noise-reduction auto-encoder, b1A bias vector from an input layer to an intermediate layer of the auto-encoder;
Figure FDA0003012853440000015
the function represented is a coding function,
Figure FDA0003012853440000016
the expressed function is an activation function of the neuron, and a sigmoid function or an identity function ReLu is taken;
secondly, the network structure of the decoding part and the encoding part of the multi-granularity convolution noise reduction automatic encoder is completely symmetrical, the network structure of the decoding part of the automatic encoder is used for decoding operation, and the high-level characteristics of the middle layer, namely the interference input vector, are subjected to decoding operation
Figure FDA0003012853440000017
Performing decoding operation to the intermediate characteristic vector alpha passing through the hidden layeriOutputting a reconstructed image of the damaged medical image by linear mapping into a nonlinear activation function, wherein the reconstructed image of the damaged medical image corresponds to a reconstructed output vector zi：
zi＝g(αi)＝l(W2αi+b2)； (2)
Wherein z isiFor the reconstructed output vector, i.e. the disturbance input vector
Figure FDA0003012853440000018
One reconstruction of (2); g (. alpha.) (alpha.)i) As a decoding function, W2Weight matrix from the middle layer to the output layer of the automatic encoder, b2Is the offset vector from the middle layer to the output layer of the auto-encoder; l (a)W2αi+b2) The expressed function is the activation function of the neuron;
establishing a cost function to measure the reconstruction performance of the automatic encoder: based on the first and second steps, an interference input vector is established
Figure FDA0003012853440000021
And an output vector ziA cost function with reconstruction error as variable, and minimizing the cost function to make the interference input vector
Figure FDA0003012853440000022
And the reconstructed output vector ziSimilarly; the cost function L (X, Z) is shown as follows, the smaller the loss value L of the cost function is, the more accurate the characteristic extraction of the encoder part is;
Figure FDA0003012853440000023
wherein X represents an input medical image training set and Z represents output reconstructed medical image data;
and fourthly, carrying out summation calculation on loss values of all medical images in the medical image training set by using a formula (3), and introducing regularization constraint to obtain an overall cost function J (W, b) of the multi-granularity convolution noise reduction automatic encoder on the medical image training set, wherein the overall cost function J (W, b) is as follows:
Figure FDA0003012853440000024
wherein J is the loss value of the image in the whole medical image training set, W and b are the weight matrix and the offset vector of the whole multi-granularity convolution noise reduction automatic encoder network respectively,
Figure FDA0003012853440000025
the method comprises the following steps of (1) determining the number of images in a medical image training set as a regularization term, wherein lambda is a penalty factor, and M is the number of images in the medical image training set;
thirdly, repeating the first step to the fourth step for three times based on a greedy layer-by-layer pre-training method to obtain a deep neural network formed by stacking three multi-granularity convolution noise-reduction automatic encoders, namely forming the deep neural network formed by stacking the three multi-granularity convolution noise-reduction automatic encoders;
sixthly, removing a decoder part of the deep multi-granularity convolution noise reduction automatic encoder, connecting a softmax classifier behind the residual deep neural network, classifying the advanced features of the medical images extracted by the encoder part of the deep multi-granularity convolution automatic encoder, wherein the functional expression of the softmax classifier is as follows:
Figure FDA0003012853440000031
wherein, wiIs an input vector xiThe weight vector of (1), M is the number of images of the medical image training set, and k is the number of categories of different image data sets;
Figure FDA0003012853440000032
the function f (x) is a scoring function and represents the similarity degree of the characteristics of medical image response and a certain disease class, so that the function f (x) is called as normalized probability and represents the probability of predicting the medical image to be confirmed as a certain disease class; e is a numerical constant of an exponential function;
fine tuning to make parameter W of deep neural network1、W2、b1、b2An optimal value is reached;
step 3), inputting the images in the medical image test set into the medical image classification model established in the step 2), and judging the disease categories to which the medical images belong according to the classification precision on the test set, wherein the higher the classification precision is, the higher the accuracy of the disease categories to which the medical images belong is, and the purpose of accurate classification is achieved; the expression formula of the classification precision is as follows:
Figure FDA0003012853440000033
wherein f is a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder, D is a medical image test set, N is the number of images in the medical image test set, and xiFor the ith image in the medical image test set D, yiFor the true class label of the ith image in the medical image test set D, f (x)i) The prediction class mark is a prediction class mark obtained after feature learning is carried out on a medical image by a medical image classification method based on a multi-granularity convolution noise reduction automatic encoder.
CN201811181783.2A 2018-10-11 2018-10-11 Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology Active CN109389171B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811181783.2A CN109389171B (en) 2018-10-11 2018-10-11 Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811181783.2A CN109389171B (en) 2018-10-11 2018-10-11 Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology

Publications (2)

Publication Number Publication Date
CN109389171A CN109389171A (en) 2019-02-26
CN109389171B true CN109389171B (en) 2021-06-25

Family

ID=65426966

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811181783.2A Active CN109389171B (en) 2018-10-11 2018-10-11 Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology

Country Status (1)

Country Link
CN (1) CN109389171B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110175560A (en) * 2019-05-24 2019-08-27 哈尔滨工程大学 A kind of radar signal intra-pulse modulation recognition methods
CN110414989A (en) * 2019-07-29 2019-11-05 中国工商银行股份有限公司 Method for detecting abnormality and device, electronic equipment and computer readable storage medium
CN110456332B (en) * 2019-08-12 2022-08-05 哈尔滨工程大学 Underwater acoustic signal enhancement method based on automatic encoder
CN110738168B (en) * 2019-10-14 2023-02-14 长安大学 Distributed strain micro crack detection system and method based on stacked convolution self-encoder
CN111311507B (en) * 2020-01-21 2022-09-23 山西大学 Ultra-low light imaging method based on multi-granularity cooperative network
CN113590902B (en) * 2021-08-13 2022-07-12 郑州大学 Big data-based personalized information support system for hematological malignancy

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105654117A (en) * 2015-12-25 2016-06-08 西北工业大学 Hyperspectral image spectral-spatial cooperative classification method based on SAE depth network
CN107145836A (en) * 2017-04-13 2017-09-08 西安电子科技大学 Hyperspectral image classification method based on stack boundary discrimination self-encoding encoder
CN108416389A (en) * 2018-03-15 2018-08-17 盐城师范学院 The image classification method sampled based on the sparse autocoder of noise reduction and density space
CN108537233A (en) * 2018-03-15 2018-09-14 南京师范大学 A kind of pathology brain image sorting technique based on the sparse self-encoding encoder of depth stack

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105654117A (en) * 2015-12-25 2016-06-08 西北工业大学 Hyperspectral image spectral-spatial cooperative classification method based on SAE depth network
CN107145836A (en) * 2017-04-13 2017-09-08 西安电子科技大学 Hyperspectral image classification method based on stack boundary discrimination self-encoding encoder
CN108416389A (en) * 2018-03-15 2018-08-17 盐城师范学院 The image classification method sampled based on the sparse autocoder of noise reduction and density space
CN108537233A (en) * 2018-03-15 2018-09-14 南京师范大学 A kind of pathology brain image sorting technique based on the sparse self-encoding encoder of depth stack

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Image Classification using Deep Autoencoders;Munmi Gogoi et al;《2017 IEEE International Conference on Computational Intelligence and Computing Research》;20171231;第1-5页 *
基于多尺度单层自编码器的医学图像分类;刘洋洋 等;《科教导刊》;20170515(第15期);第20-21、192页 *

Also Published As

Publication number Publication date
CN109389171A (en) 2019-02-26

Similar Documents

Publication Publication Date Title
CN109389171B (en) Medical image classification method based on multi-granularity convolution noise reduction automatic encoder technology
CN107516110B (en) Medical question-answer semantic clustering method based on integrated convolutional coding
CN110532900B (en) Facial expression recognition method based on U-Net and LS-CNN
CN110335261B (en) CT lymph node detection system based on space-time circulation attention mechanism
CN107229914B (en) Handwritten digit recognition method based on deep Q learning strategy
CN111738363B (en) Alzheimer disease classification method based on improved 3D CNN network
WO2022127500A1 (en) Multiple neural networks-based mri image segmentation method and apparatus, and device
CN113688869B (en) Photovoltaic data missing reconstruction method based on generation countermeasure network
CN112766283B (en) Two-phase flow pattern identification method based on multi-scale convolution network
CN110851654A (en) Industrial equipment fault detection and classification method based on tensor data dimension reduction
CN114048468A (en) Intrusion detection method, intrusion detection model training method, device and medium
Wu et al. Optimized deep learning framework for water distribution data-driven modeling
CN114169377A (en) G-MSCNN-based fault diagnosis method for rolling bearing in noisy environment
CN111145145A (en) Image surface defect detection method based on MobileNet
CN111259735A (en) Single-person attitude estimation method based on multi-stage prediction feature enhanced convolutional neural network
CN114565594A (en) Image anomaly detection method based on soft mask contrast loss
Kong et al. Semi-supervised classification of wafer map based on ladder network
CN111695611A (en) Bee colony optimization kernel extreme learning and sparse representation mechanical fault identification method
CN115290326A (en) Rolling bearing fault intelligent diagnosis method
CN113887559A (en) Brain-computer information fusion classification method and system for brain off-loop application
CN111798455A (en) Thyroid nodule real-time segmentation method based on full convolution dense cavity network
CN111275109A (en) Power equipment state data characteristic optimization method and system based on self-encoder
Termritthikun et al. Neural architecture search and multi-objective evolutionary algorithms for anomaly detection
CN109782156B (en) Analog circuit fault diagnosis method based on artificial immune diagnosis network
CN113935413A (en) Distribution network wave recording file waveform identification method based on convolutional neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant