CN109389171A - Medical image classification method based on more granularity convolution noise reduction autocoder technologies - Google Patents
Medical image classification method based on more granularity convolution noise reduction autocoder technologies Download PDFInfo
- Publication number
- CN109389171A CN109389171A CN201811181783.2A CN201811181783A CN109389171A CN 109389171 A CN109389171 A CN 109389171A CN 201811181783 A CN201811181783 A CN 201811181783A CN 109389171 A CN109389171 A CN 109389171A
- Authority
- CN
- China
- Prior art keywords
- medical image
- vector
- image
- noise reduction
- autocoder
- 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.)
- Granted
Links
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 kind of medical image classification methods based on more granularity convolution noise reduction autocoder technologies, by introducing convolutional neural networks on the basis of traditional autocoder；Different size of convolution kernel is set in the same convolutional layer simultaneously, and random Gaussian is added in the training stage, the deep neural network of the noise reduction autocoder of granularity convolution more than three is finally trained to based on greedy successively training method, it is final after fine-tuned to realize that parameter global is optimal.The present invention realizes raising nicety of grading, it is optimal classification results and has complex environment compared with strong anti-interference ability, solves the problems, such as that data category existing in the prior art obtains difficult and expends that resource is more, extraction feature is single low with the anti-interference ability to complex environment causes generalization to be remarkably decreased.
Description
Technical field
The invention belongs to computer vision technical fields, are based on more granularity convolution noise reduction autocodings more particularly to one kind
The medical image classification method of device technology.
Background technique
Image classification is a very active application study direction in computer vision, and the purpose is to various types of other
Image classifies, and the similar image of feature is assigned in same class, and inhomogeneous image be then assigned to it is different
In class.In existing research, image classification can substantially be divided into following steps: (1) image preprocessing；(2) feature mentions
It takes；(3) feature selecting；(4) selection and design of classifier.In these steps, feature extraction is played a key role, certainly
The precision of last classification results and the stability of performance are determined.In past research, the process for carrying out feature extraction is usually
Local feature or global characteristics are extracted, such as: local binary pattern (LBP), histogram of oriented
gradients(HOG),scale invariant feature transform(SIFT).In recent years, with machine learning
Rise researchers sight has been gathered in deep learning with fastest developing speed, this is because existing research show deep learning calculate
Method has the function of simulation human brain structure, the rudimentary feature of bottom successively can be abstracted as advanced features, can be more effective
Express and describe image.It is many both at home and abroad due to the theoretical research with higher of this research field and practical application value
Researcher proposes the technology of more medical image classification, but substantially contains the work of three aspects: feature learning, feature mention
It takes and classifies.Firstly, carrying out feature learning to original medical image data；Secondly, extracting the advanced features of image, most
Afterwards, classification analysis is carried out to image based on the advanced features extracted.
Although researcher has been achieved for some achievements on image classification field, due between actual medical image class
The problem of difference is minimum, and feature distribution is very complicated, and there are also noise jammings, so that the prior art and immature.It is primarily present
The defect of following several respects.(1) data that existing convolutional neural networks need largely to have category；Data class in real life
Target is difficult, and heavy dependence people craft and expertise, this just needs to expend a large amount of resource；(2) existing
The image classification algorithms based on feature extraction in, the feature extracted has unicity, only due to traditional convolutional neural networks
Using a kind of convolution kernel of size, the advanced features extracted is caused not have diversity, the precision that finally will lead to algorithm has
Declined.(3) stability of most sorting algorithm is limited, when encountering some complex environments, such as noise problem
When, the Generalization Capability of algorithm is remarkably decreased.
Therefore, how under conditions of no priori knowledge, different deep learning algorithms is combined reasonably to improve mould
It is a urgent problem to be solved that the robustness of type, which has reached optimal classification results,.
Summary of the invention
The purpose of the present invention is to provide a kind of medical images based on more granularity convolution noise reduction autocoder technologies point
Class method, to realize that one generalization ability of construction is strong, the high medical image classifier of anti-interference, while giving full play to every height
Performance of the component in terms of feature extraction；To solve, data category in the prior art obtains difficulty and consuming resource is more, it is special to extract
Levy single the problem of causing generalization to be remarkably decreased low with anti-interference ability to complex environment.
The technical scheme adopted by the invention is that providing a kind of doctor based on more granularity convolution noise reduction autocoder technologies
Treat image classification method, comprising the following steps:
Step 1) medical image pretreatment: calls OpenCV kit, the size of medical image is adjusted to 256
× 256, image adjusted is normalized, and as image data set, will according to the medical image number ratio of 8:2
Described image data set is divided into medical image training set and medical image test set；
Step 2) establishes medical image disaggregated model:
A. individual medical image is compiled with the network structure of more granularity convolution noise reduction autocoder coded portions
Code operation；More granularity convolution noise reduction autocoder coded portions use the volume of 1 × 1,3 × 3,5 × 5 three kinds of sizes simultaneously
Product core, by individual medical image in step 1) after pretreatment having a size of 256 × 256 be corresponding dimension be 65536 it is defeated
Incoming vector xiGaussian noise is added, obtains being damaged medical image data feature, the medical image data feature that is damaged is determined
Justice is exogenous disturbances vectorNonlinear activation primitive is entered by Linear Mapping again and obtains the character representation vector of middle layer
αi:
Wherein, αiFor exogenous disturbances vectorThe middle layer character representation vector obtained after two hidden layers are abstract, angle
Mark the dimension that i is image in medical image training set；xiFor input vector；Noise is being added for original clean input vector
The impaired exogenous disturbances vector obtained after interference, i.e. exogenous disturbances vector；W1For more granularity convolution noise reduction autocodings
Weight matrix of the input layer of device to middle layer, b1For autocoder input layer to middle layer bias vector；Table
The function shown is coding function,The function of expression be neuron activation primitive, usually take sigmoid function or
Person's identity function ReLu；
B. the network structure of decoded portion and coded portion based on more granularity convolution noise reduction autocoders is full symmetric,
It is decoded operation with the network structure of autocoder decoded portion, to advanced features, that is, exogenous disturbances vector of middle layerIt is decoded operation, by the median feature vector α of hidden layeriEnter nonlinear activation primitive, output by Linear Mapping
It is damaged the reconstructed image of medical image, the output vector z after the corresponding reconstruct of the reconstructed image for being damaged medical imagei:
zi=g (αi)=l (W2αi+b2)； (2)
Wherein, ziFor the output vector after reconstruct, i.e. exogenous disturbances vectorOne reconstruct；g(αi) it is decoding functions, W2
For the weight matrix of autocoder middle layer to output layer, b2It is bias vector of the autocoder middle layer to output layer；l
(W2αi+b2) indicate function be neuron activation primitive；
C. cost function is established, the reconstruction property of autocoder is measured: based on abovementioned steps a and b, being established defeated to interfere
Incoming vectorWith output vector ziBetween reconstructed error be variable cost function make to interfere by minimizing cost function
Input vectorWith the output vector z after reconstructiIt is similar；As follows, the penalty values L of cost function is got over cost function L (X, Z)
Small, encoder section feature extraction is more accurate；
Wherein, X represents the medical image training set of input, and Z indicates the reconstruct medical image data of output；
D., formula (3) are carried out to the read group total of penalty values for all medical images in medical image training set, and
Regularization constraint is introduced, overall cost function J of more granularity convolution noise reduction autocoders on medical image training set is obtained
(W, b) is as follows:
Wherein, J is the penalty values of image in entire medical image training set, and W and b are respectively entire more granularity convolution noise reductions
The weight matrix and bias vector of autocoder network,For regularization term, λ is penalty factor, and M is medical image instruction
Practice the quantity for concentrating image；
E. the method based on greedy successively pre-training repeats abovementioned steps a to d three times, obtains the drop of granularity convolution more than three
The deep neural network that autocoder of making an uproar is stacked into is constituted and is formed by stacking by the noise reduction autocoder of granularity convolution more than three
Deep neural network；
F. the decoder section for releasing the more granularity convolution noise reduction autocoders of depth, after remaining deep neural network
Softmax classifier is connected, the medical image extracted to the encoder section of the more granularity convolution autocoders of depth is advanced
Feature is classified, and the function expression of softmax classifier is as follows:
Wherein, wiIt is input vector xiWeight vector, M be medical image training set image quantity, k is different medical
The classification number of image data set；For score function, the feature and the phase of a certain disease category of medical image reaction are indicated
Short range degree, therefore f (x) function is known as normalizing probability, indicates that prediction medical image is confirmed as the probability of a certain disease category；E is
The numeric constant of exponential function；
G. it finely tunes, makes the parameter W of deep neural network1、W2、b1、b2It is optimal value；
Step 3), by the image input step 2 in medical image test set) establish medical image disaggregated model in, lead to
The nicety of grading crossed on test set judges the disease category of medical image ownership, and nicety of grading is higher, and medical image ownership is somebody's turn to do
The accuracy of disease category is higher, that is, achievees the purpose that exact classification；The expression formula of nicety of grading are as follows:
Wherein, f is the medical image classification method based on more granularity convolution noise reduction autocoders, and D is medical image survey
Examination collection, N are the quantity of image in medical image test set, xiFor i-th image in medical image test set D, yiFor medical treatment figure
As the true category of i-th image in test set D, f (xi) it is the medical image based on more granularity convolution noise reduction autocoders
Classification method carries out the prediction category after feature learning to medical image.
The beneficial effects of the present invention are:
(1) the present invention is based on the feature learning methods of unsupervised learning, and the convolution operation in convolutional neural networks is used
In the network structure of traditional autocoder, there is nicety of grading more higher than monotechnics, reached partially connected
Effect；The shared characteristic of convolution weight is substantially reduced network parameter quantity simultaneously, reduces calculation amount；And by heavy dependence
The field of unsupervised learning has been applied in the convolutional neural networks for there are category data, has widened the development of convolutional neural networks.
(2) method that more granularity convolution kernels of the invention extract advanced features guarantees on the basis of high efficiency extraction feature
The diversity of the advanced features extracted, can comprehensively characterize the key feature of image, be conducive to the precision for improving classification, reach
Optimal classification result.
(3) present invention has strong anti-interference ability complex environment.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is autocoder network structure.
Fig. 2 is the feature learning algorithm schematic diagram based on more granularity convolution noise reduction autocoders.
Fig. 3 is the more granularity convolution noise reduction autocoder schematic diagrames of depth.
Fig. 4 is classification performance comparison diagram under different noise proportionals.
Fig. 5 is the image under meningitis cerebrospinal fluid microscope.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
On meningitis data set, the medical treatment based on more granularity convolution noise reduction autocoder technologies that the present invention announces is schemed
As classification method has obtained preferable verifying.As shown in figure 5, meningitis is divided into four classes, it is purulent meningitis, Tuberculous respectively
Meningitis, viral meningitis, cryptococcal meningitis.From figure 5 it can be seen that between the imaging class of cerebrospinal fluid under the microscope
Gap is minimum, and the people or the inadequate doctor of clinical experience for lacking professional knowledge are difficult to judge the concrete type of meningitis.One
As in the case of, doctor is according to the comprehensive type for making a definite diagnosis illness of the clinical manifestation of patient and examination of cerebrospinal fluid and then just can be carried out phase
The treatment answered.But doctor will receive personal experience, knowledge, condition etc. when observing the imaging of cerebrospinal fluid under the microscope
Influence；The Medical Devices of image data from Different hospital or the different size of the same race from the same hospital collect
Image data also have certain difference so that the result finally made a definite diagnosis will receive the influence of these many factors.Therefore,
It needs to carry out Accurate classification with classification method of the invention.
Detailed process is as follows:
1. image preprocessing: meningitis data set original size is 3184 × 4230, therefore when obtaining these image datas,
It first has to be pre-processed.In an embodiment of the present invention, pretreatment can be divided into three steps: 1) adjusting image size: calling
OpenCV kit, to be adjusted the operation of size to meningitis image, in the concrete operations of this paper, meningitis image
256 × 256 sizes are uniformly processed into size；2) normalized: in many image processing process, image is normalized
Processing is necessary preprocessing process, and so-called normalized is exactly that image pixel value is made to be distributed in the point between 0-255
Distribution value can accelerate the convergent speed of model parameter between [0,1], after processing, and herein, we are also to meningitis image
Operation is normalized, makes network parameter fast convergence in the training process, saves the training time；3) data set is divided: in the future
Meningitis figure is randomly divided into according to the image number ratio of 8:2 from 148 image meningitis image data sets of certain Grade A hospital
As training set and meningitis image measurement collection, wherein meningitis training set of images is used to that medical image of the invention is trained to classify
Model, meningitis image measurement collection are used to assess medical image disaggregated model of the present invention in specific meningitis image classification task
Performance.Particularly, four class meningococcal diseases are all evenly distributed in meningitis training set of images and meningitis image measurement collection
In.
2. establishing medical image classification model construction: in order to capture medical image multifarious and of overall importance feature and
The robustness (anti-noise ability) of model is improved, the invention proposes the algorithm models of more granularity convolution noise reduction autocoders.It should
The model network structure full symmetric as one can be divided into two parts of encoder and decoder.
A. firstly, as shown in Figure 1, encoder section performs the encoding operation the meningitis image of input, successively abstract
To the advanced features of middle layer；I.e. with the network structure of more granularity convolution noise reduction autocoder coded portions to individual medical treatment
Image performs the encoding operation；More granularity convolution noise reduction autocoder coded portions use 1 × 1,3 × 3,5 × 5 three simultaneously
The convolution kernel of kind size, individual medical image in step 1) after pretreatment having a size of 256 × 256, which is corresponded to dimension, is
65536 input vector xiGaussian noise is added, obtains being damaged medical image data feature, it is described to be damaged medical image number
It is exogenous disturbances vector according to characterizing definitionNonlinear activation primitive is entered by Linear Mapping again and obtains the feature of middle layer
Indicate vector αi:
Wherein, αiFor exogenous disturbances vectorThe middle layer character representation vector obtained after two hidden layers are abstract, angle
Mark the dimension that i is image in medical image training set；xiFor input vector；Noise is being added for original clean input vector
The impaired exogenous disturbances vector obtained after interference, i.e. exogenous disturbances vector；W1For more granularity convolution noise reduction autocodings
Weight matrix of the input layer of device to middle layer, b1For autocoder input layer to middle layer bias vector；Table
The function shown is coding function,The function of expression be neuron activation primitive, usually take sigmoid function or
Person's identity function ReLu.
B. as shown in Fig. 2, being decoded operation to the advanced features of middle layer, the reconstruct table of input meningitis image is obtained
Show.It is decoded operation with the network structure of autocoder decoded portion, the advanced features of middle layer are interfered defeated
Incoming vectorIt is decoded operation, by the median feature vector α of hidden layeriEnter nonlinear activation letter by Linear Mapping
Number, output are damaged the reconstructed image of medical image, the output after the corresponding reconstruct of the reconstructed image for being damaged medical image
Vector zi:
zi=g (αi)=l (W2αi+b2)； (2)
Wherein, ziFor the output vector after reconstruct, i.e. exogenous disturbances vectorOne reconstruct；g(αi) it is decoding functions, W2
For the weight matrix of autocoder middle layer to output layer, b2It is bias vector of the autocoder middle layer to output layer；
l(W2αi+b2) indicate function be neuron activation primitive.
C. the symmetrical neural network is trained, needs to construct a cost function for an image, the present invention is by meninx
The reconstructed error of scorching image is as cost function.In hands-on, keep cost function small as far as possible, to promote disaggregated model to exist
Learn the advanced features for selecting optimal parameter preferably to capture meningitis image when meningitis training set of images.Cause
This, is based on modeling process a and b, establishes with exogenous disturbances vectorWith output vector ziBetween reconstructed error be variable cost
Function makes exogenous disturbances vector by minimizing cost functionWith the output vector z after reconstructiIt is similar；Cost function L
(X, Z) as follows, the penalty values L for minimizing cost function is smaller, and encoder section feature extraction is more accurate；
Wherein, X represents the medical image training set of input, and Z indicates the reconstruct medical image data of output；
D. overall cost function is established to multiple images, i.e., is carried out for all medical images in medical image training set
The read group total of penalty values, and regularization constraint is introduced, more granularity convolution noise reduction autocoders are obtained in medical image training
Overall cost function J (W, b) on collection is as follows:
Wherein, J is the penalty values of image in entire medical image training set, and W and b are respectively entire more granularity convolution noise reductions
The weight matrix and bias vector of autocoder network,For regularization term, λ is penalty factor, and M is medical image instruction
Practice the quantity for concentrating image；
E. various sizes of convolution kernel there are three types of containing in each more granularity convolution noise reduction autocoders, at more than first
After spending the training of convolution noise reduction autocoder, based on the method for greedy successively pre-training, the above training process is repeated, is obtained
The deep neural network that the noise reduction autocoder of granularity convolution more than three is stacked into is constituted by the noise reduction of granularity convolution more than three certainly
The deep neural network that dynamic encoder is formed by stacking, so that classification precision is mentioned, as shown in figure 3, training more than first first
Granularity convolution noise reduction autocoder fixes its parameter after trainingWithAnd by the output of wherein interbedAs
Granularity convolution noise reduction autocoder more than second, instruction are then trained in the input of the noise reduction autocoder of granularity convolution more than two
After white silk, its parameter is fixedWithBy the output of wherein interbedAs the more granularity convolution noise reduction autocodings of third
The input of device, and so on, train the depth stack nerve net being made of the noise reduction autocoder of granularity convolution more than three
Network, preset parameterWithBy the output of wherein interbedAs the input of softmax classifier, the classifier is trained, into
And complete entire training.Wherein,Respectively correspond the input layer of the noise reduction autocoder of granularity convolution more than first
To the character representation vector of the weight matrix of middle layer, bias vector and middle layer, similarly,AndIt is second
Feature of the input layer of a more granularity convolution noise reduction autocoders to the weight matrix of middle layer, bias vector and middle layer
Indicate vector,WithRespectively power of the input layer of the more granularity convolution noise reduction autocoders of third to middle layer
The character representation vector of value matrix, bias vector and middle layer.
As shown in figure 4, the classification precision of the automatic noise reduction codes device of granularity convolution more than three after superimposing significantly improves；
F. as shown in figure 3, removing the deep neural network decoder section, the output of remaining encoder section is exactly to cure
It treats image classification model and carries out required advanced features of classifying, be followed by upper softmax classification in remaining deep neural network
Device, the meningitis image advanced features extracted to the encoder section of the more granularity convolution autocoders of depth are classified；
The function expression of softmax classifier is as follows:
Wherein, wiIt is input vector xiWeight vector, M be medical image training set image quantity, k is different images
The classification number of data set；For score function, the feature and the phase short range of a certain disease category of medical image reaction are indicated
Degree, therefore f (x) function is known as normalizing probability, indicates that prediction medical image is confirmed as the probability of a certain disease category；E is index
The numeric constant of function；
G. the parameter W of deep neural network1、W2、b1、b2All reaching local optimum after pre-training, however this
When be not to the classification accuracy rate of the disease specific classification of meningitis image it is highest, finally need to the network carry out it is micro-
It adjusts, after fine tuning, the parameter W of the network1、W2、b1、b2Global optimum is reached, medical image disaggregated model is to meninx at this time
The classification accuracy rate of disease type belonging to scorching image reaches peak.
3. prediction: the Generalization Capability in order to assess medical image disaggregated model of the present invention is needed in meningitis image measurement
It is tested on collection, at the same time, in order to verify disaggregated model and other disaggregated models of the invention in meningitis medical image
Medical image classification method and other classification methods of the invention are based on meningitis medical treatment herein by the performance difference of upper classification
The comparison of image measurement collection progress classification accuracy rate.Classification accuracy rate is expressed as follows by nicety of grading formula:
Wherein, f is the medical image classification method based on more granularity convolution noise reduction autocoders, and D is medical image survey
Examination collection, N are the quantity of image in medical image test set, xiFor i-th image in medical image test set D, yiFor medical treatment figure
As the true category of i-th image in test set D, f (xi) it is the medical image based on more granularity convolution noise reduction autocoders
Classification method carries out the prediction category after feature learning to medical image.
Found out by experimental result, classification accuracy rate of the present invention on meningitis medical image test set reaches
72.83%, hence it is evident that be higher than other algorithm models, as shown in table 1.Classification method of the invention is sufficiently demonstrated in medical image point
Has higher accuracy in class.
The classification accuracy rate of 1 meningitis medical image algorithms of different of table compares
In addition, the present invention is using granularity convolution noise reduction more than three in the experiment of meningitis medical image data sets
Deep neural network made of autocoder stacks.The present invention is automatic from three simple grain degree convolution of different convolution kernel sizes
The superposition of encoder carries out classification comparison in meningitis medical image concentration respectively, obtains pair of respectively different classification accuracies
Than data, as shown in table 2.
Accuracy of the table 2 based on varigrained algorithm in meningitis medical image integrated test compares
From Table 2, it can be seen that compared to the convolution autocoder with simple grain degree convolution kernel, more granularity convolution kernels are set
Meter can capture the comprehensive advanced features of image, be conducive to the nicety of grading for improving disaggregated model.
In order to improve the robustness of disaggregated model, the present invention joined Gaussian noise in the training process of model, force
Medical image disaggregated model of the invention from be damaged from medical image in study to more with robustness advanced features,
It tests for comparison to be added after Gaussian noise below and classification comparison is carried out on meningitis image data set to algorithm, obtain respectively
The validity of Gaussian noise is added in the correlation data of different classification accuracies, verifying.Wherein CDAE is to roll up with Monosized powder
The convolution noise reduction autocoder of product core.
The test accuracy comparison of algorithms of different after noise is added in table 3
From table 3 it can be seen that being added after Gaussian noise, convolution autocoder exists compared to that Gaussian noise is not added
Classification accuracy rate on meningitis image data set significantly improves, this, which demonstrates addition Gaussian noise, to make algorithm in training process
In extract medical image more there are the advanced features of robustness, so as to improve classification accuracy rate.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the scope of the present invention.It is all
Any modification, equivalent replacement, improvement and so within the spirit and principles in the present invention, are all contained in protection scope of the present invention
It is interior.
Claims (1)
1. a kind of medical image classification method based on more granularity convolution noise reduction autocoder technologies, which is characterized in that including
Following steps:
Step 1), medical image pretreatment: calling OpenCV kit, the size of medical image is adjusted to 256 ×
256, image adjusted is normalized, and as image data set, according to the medical image number ratio of 8:2 by institute
It states image data set and is divided into medical image training set and medical image test set；
Step 2) establishes medical image disaggregated model:
A. coding behaviour is carried out to individual medical image with the network structure of more granularity convolution noise reduction autocoder coded portions
Make；More granularity convolution noise reduction autocoder coded portions use the convolution of 1 × 1,3 × 3,5 × 5 three kinds of sizes simultaneously
Individual medical image in step 1) after pretreatment having a size of 256 × 256 is the input that corresponding dimension is 65536 by core
Vector xiGaussian noise is added, obtains being damaged medical image data feature, it is described to be damaged medical image data characterizing definition
For exogenous disturbances vectorNonlinear activation primitive is entered by Linear Mapping again and obtains the character representation vector α of middle layeri:
Wherein, αiFor exogenous disturbances vectorThe middle layer character representation vector obtained after two hidden layers are abstract, footmark i
For the dimension of image in medical image training set；xiFor input vector；Noise jamming is being added for original clean input vector
The impaired exogenous disturbances vector obtained later, i.e. exogenous disturbances vector；W1For more granularity convolution noise reduction autocoders
Weight matrix of the input layer to middle layer, b1For autocoder input layer to middle layer bias vector；It indicates
Function is coding function,The function of expression is the activation primitive of neuron, takes sigmoid function or identical letter
Number ReLu；
B. the network structure of decoded portion and coded portion based on more granularity convolution noise reduction autocoders is full symmetric, uses
The network structure of autocoder decoded portion is decoded operation, to advanced features, that is, exogenous disturbances vector of middle layerInto
Row decoding operate, by the median feature vector α of hidden layeriEnter nonlinear activation primitive by Linear Mapping, exports quilt
Damage the reconstructed image of medical image, the output vector z after the corresponding reconstruct of the reconstructed image for being damaged medical imagei:
zi=g (αi)=l (W2αi+b2)； (2)
Wherein, ziFor the output vector after reconstruct, i.e. exogenous disturbances vectorOne reconstruct；g(αi) it is decoding functions, W2For certainly
Dynamic weight matrix of the encoder middle layer to output layer, b2It is bias vector of the autocoder middle layer to output layer；l(W2αi
+b2) indicate function be neuron activation primitive；
C. establish cost function, measure the reconstruction property of autocoder: based on abovementioned steps a and b, establish with exogenous disturbances to
AmountWith output vector ziBetween reconstructed error be variable cost function make exogenous disturbances by minimizing cost function
VectorWith the output vector z after reconstructiIt is similar；As follows, the penalty values L of cost function is smaller by cost function L (X, Z), compiles
It is more accurate that code device Partial Feature is extracted；
Wherein, X represents the medical image training set of input, and Z indicates the reconstruct medical image data of output；
D., formula (3) are carried out to the read group total of penalty values for all medical images in medical image training set, and are introduced
Regularization constraint obtains overall cost function J (W, b) of more granularity convolution noise reduction autocoders on medical image training set
It is as follows:
Wherein, J is the penalty values of image in entire medical image training set, and W and b are respectively that entire more granularity convolution noise reductions are automatic
The weight matrix and bias vector of encoder network,For regularization term, λ is penalty factor, and M is medical image training set
The quantity of middle image；
E. the method based on greedy successively pre-training repeats abovementioned steps a to d three times, obtains the noise reduction of granularity convolution more than three certainly
The deep neural network that dynamic encoder is stacked into constitutes the depth being formed by stacking by the noise reduction autocoder of granularity convolution more than three
Spend neural network；
F. the decoder section for releasing the more granularity convolution noise reduction autocoders of depth, is followed by remaining deep neural network
Softmax classifier, the medical image advanced features that the encoder section of the more granularity convolution autocoders of depth is extracted
Classify, the function expression of softmax classifier is as follows:
Wherein, wiIt is input vector xiWeight vector, M be medical image training set image quantity, k be different images data
The classification number of collection；For score function, the feature of medical image reaction and the close degree of a certain disease category are indicated, therefore
F (x) function is known as normalizing probability, indicates that prediction medical image is confirmed as the probability of a certain disease category；E is exponential function
Numeric constant；
G. it finely tunes, makes the parameter W of deep neural network1、W2、b1、b2It is optimal value；
Step 3), by the image input step 2 in medical image test set) establish medical image disaggregated model in, pass through survey
Nicety of grading on examination collection judges the disease category of medical image ownership, and nicety of grading is higher, the disease of medical image ownership
The accuracy of classification is higher, that is, achievees the purpose that exact classification；The expression formula of nicety of grading are as follows:
Wherein, f is the medical image classification method based on more granularity convolution noise reduction autocoders, and D is medical image test set,
N is the quantity of image in medical image test set, xiFor i-th image in medical image test set D, yiFor medical image survey
The true category of i-th image, f (x in examination collection Di) it is the medical image classification based on more granularity convolution noise reduction autocoders
Method carries out the prediction category after feature learning to medical image.
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 true CN109389171A (en) | 2019-02-26 |
| CN109389171B 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) |
Cited By (8)
| 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 |
| CN110456332A (en) * | 2019-08-12 | 2019-11-15 | 哈尔滨工程大学 | A kind of underwater sound signal Enhancement Method based on autocoder |
| CN110738168A (en) * | 2019-10-14 | 2020-01-31 | 长安大学 | distributed strain micro crack detection system and method based on stacked convolution self-encoder |
| CN111311507A (en) * | 2020-01-21 | 2020-06-19 | 山西大学 | Ultra-low light imaging method based on multi-granularity cooperative network |
| CN112183561A (en) * | 2020-11-09 | 2021-01-05 | 山东中医药大学 | Joint fusion subtraction automatic encoder algorithm for image feature extraction |
| CN112446415A (en) * | 2020-10-09 | 2021-03-05 | 山东中医药大学 | Fusion subtraction automatic encoder algorithm for image feature extraction |
| CN113590902A (en) * | 2021-08-13 | 2021-11-02 | 郑州大学 | Big data-based personalized information support system for hematological malignancy |
Citations (4)
| 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 |
-
2018
- 2018-10-11 CN CN201811181783.2A patent/CN109389171B/en active Active
Patent Citations (4)
| 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)
| Title |
|---|
| MUNMI GOGOI ET AL: "Image Classification using Deep Autoencoders", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTATIONAL INTELLIGENCE AND COMPUTING RESEARCH》 * |
| 刘洋洋 等: "基于多尺度单层自编码器的医学图像分类", 《科教导刊》 * |
Cited By (8)
| 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 |
| CN110456332A (en) * | 2019-08-12 | 2019-11-15 | 哈尔滨工程大学 | A kind of underwater sound signal Enhancement Method based on autocoder |
| CN110738168A (en) * | 2019-10-14 | 2020-01-31 | 长安大学 | distributed strain micro crack detection system and method based on stacked convolution self-encoder |
| CN111311507A (en) * | 2020-01-21 | 2020-06-19 | 山西大学 | Ultra-low light imaging method based on multi-granularity cooperative network |
| CN112446415A (en) * | 2020-10-09 | 2021-03-05 | 山东中医药大学 | Fusion subtraction automatic encoder algorithm for image feature extraction |
| CN112183561A (en) * | 2020-11-09 | 2021-01-05 | 山东中医药大学 | Joint fusion subtraction automatic encoder algorithm for image feature extraction |
| CN113590902A (en) * | 2021-08-13 | 2021-11-02 | 郑州大学 | Big data-based personalized information support system for hematological malignancy |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109389171B (en) | 2021-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109389171A (en) | Medical image classification method based on more granularity convolution noise reduction autocoder technologies | |
| CN108717568B (en) | A kind of image characteristics extraction and training method based on Three dimensional convolution neural network | |
| CN106202997B (en) | A kind of cell division detection method based on deep learning | |
| CN105138993B (en) | Establish the method and device of human face recognition model | |
| CN105046277B (en) | Robust mechanism study method of the feature significance in image quality evaluation | |
| CN107977932A (en) | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method | |
| CN106446942A (en) | Crop disease identification method based on incremental learning | |
| CN108595602A (en) | The question sentence file classification method combined with depth model based on shallow Model | |
| CN107085716A (en) | Across the visual angle gait recognition method of confrontation network is generated based on multitask | |
| CN110309861A (en) | A kind of multi-modal mankind's activity recognition methods based on generation confrontation network | |
| Zhou et al. | Multi-classification of skin diseases for dermoscopy images using deep learning | |
| CN108665005A (en) | A method of it is improved based on CNN image recognition performances using DCGAN | |
| CN106779064A (en) | Deep neural network self-training method based on data characteristics | |
| CN110033440A (en) | Biological cell method of counting based on convolutional neural networks and Fusion Features | |
| Paupamah et al. | Quantisation and pruning for neural network compression and regularisation | |
| CN110399821A (en) | Customer satisfaction acquisition methods based on facial expression recognition | |
| CN108921233A (en) | A kind of Raman spectrum data classification method based on autoencoder network | |
| CN109410114A (en) | Compressed sensing image reconstruction algorithm based on deep learning | |
| CN109086802A (en) | A kind of image classification method based on biquaternion convolutional neural networks | |
| CN107679572A (en) | A kind of image discriminating method, storage device and mobile terminal | |
| CN106485259A (en) | A kind of image classification method based on high constraint high dispersive principal component analysiss network | |
| CN106127240A (en) | A kind of classifying identification method of plant image collection based on nonlinear reconstruction model | |
| CN109740734A (en) | A kind of method of neuron spatial arrangement in optimization convolutional neural networks | |
| CN111127490A (en) | Medical image segmentation method based on cyclic residual U-Net network | |
| CN110096976A (en) | Human behavior micro-Doppler classification method based on sparse migration 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 |