CN104851101A - Brain tumor automatic segmentation method based on deep learning - Google Patents

Abstract

The invention discloses a brain tumor automatic segmentation method based on deep learning, which comprises the steps of firstly carrying out image block extraction on a magnetic resonance imaging slice containing a brain tumor, using a grayscale sequence of image blocks to act as input of deep learning, carrying out unsupervised pre-training on a network so as to acquire initial network parameters, then carrying out fine tuning on the network parameters reversely by using labeled data, using acquired final parameters to initialize a new network, and carrying out initial classification on images to be tested through the network. When classification is completed, primary division is carried out on points which are preliminarily segmented into the tumor by using a set grayscale value, and post-processing is carried out finally so as to acquire a final segmentation result. Disclosed by the invention is a brain tumor automatic segmentation method, and additional features of the brain tumor are not required to be extracted. The brain tumor automatic segmentation method provides useful information for doctors to formulate a brain tumor operation plan.

Description

A kind of brain tumor automatic division method based on degree of depth study

Technical field

The invention belongs to computer-aided diagnosis field, more specifically say, relate to a kind of brain tumor automatic division method based on degree of depth study.

Background technology

Degree of depth study is the popular domain of current machine Learning Studies, and it carrys out the data such as interpretation of images, sound and text by the process setting up, simulate human brain analytic learning.Degree of depth study is the one of unsupervised learning, and the multilayer perceptron containing many hidden layers is exactly a kind of degree of depth study structure.Degree of depth study forms more abstract high level by combination low-level feature and represents attribute classification or feature, to find that the distributed nature of data represents.Degree of depth study is mainly used in speech recognition, Face datection now, the aspects such as semantic parsing, but is also seldom applied to computer-aided diagnosis field.

At present, the research of degree of depth study in computer-aided diagnosis field is mainly for the classification of alzheimer's disease and mild cognitive impairment.Research proves, alzheimer's disease computer-aided diagnosis system based on degree of depth study can just predict when patient has only suffered from mild cognitive impairment whether this patient following can convert alzheimer's disease to, this provides an objective computer diagnosis result as the second reference to doctor, serves positive effect to diagnosis mild cognitive impairment and alzheimer's disease.

Now, the research of computer aided diagnosis applications on brain tumor now made some progress, large quantifier elimination is in progress.But brain tumor to be united modal disease except cerebrovascular disease ectoneural system, be considered to one of refractory neoplasm the most thorny in neurosurgical treatment, seriously threaten the healthy of the mankind.

One step of the treatment most critical of brain tumor is exactly the excision carrying out performing the operation, and the planning of the segmentation of tumour to pre-operative surgical scheme plays vital effect.If accurately brain tumor can be partitioned into from the nuclear magnetic resonance image of patient, just can provide more valuable reference for operating doctor in the preoperative with in art, tumor resection does not hurt normal structure again as much as possible, the misery that patient can be made to recur away from brain tumor, improves the life quality of patient.At present, the Accurate Segmentation of brain tumor mainly still depends on the subjective experience of doctor, and this not only can the manpower of at substantial, and segmentation result is also uneven because the level of doctor is different.The segmentation realized accurately and fast brain tumor by application modern information technologies will be an important research direction in future computer auxiliary diagnosis field.

Existing brain tumor dividing method mainly contains level-set segmentation, fuzzy clustering, region growing and machine learning, but mostly needs manual intervention.And do not need artificial intervention based on the brain tumor dividing method of degree of depth study, be a kind of unsupervised partitioning algorithm, automatically can be partitioned into brain tumor from the nuclear magnetic resonance image of patient, while ensure that accurate rate, also improve the efficiency of segmentation.

Summary of the invention

The object of the invention is to the deficiency overcoming existing brain tumor dividing method, a kind of brain tumor dividing method based on degree of depth study is provided, by the segmentation of brain tumor being converted to the classification of image block, image block centered by this point is extracted to each point in image, using its gray-scale value as input, the degree of depth is utilized to learn to find the potential feature in image block, for image block is classified, and result is mapped on former figure, obtain preliminary segmentation result, again aftertreatment is carried out to it, obtain final segmentation result.

The technical solution adopted for the present invention to solve the technical problems is as follows:

Based on the brain tumor automatic division method of degree of depth study, its flow process as shown in Figure 1, specifically comprises the following steps:

Step one: carry out pre-service to the MRI section that patient comprises brain tumor, cutting edge reduces calculated amount, and promotes contrast;

Step 2: extract its image block to the point in image, is the vector of an one dimension by the gradation conversion of each image block, is arranged in order formation input matrix;

Step 3: as Fig. 2, the number of hidden nodes of set deep learning network, using matrix as input, first carries out pre-training, obtains initial network parameter;

Step 4: as Fig. 3, increases an output layer at the top of network, utilize the data of tape label to finely tune network parameter and obtain final parameter;

Step 5: the sorter network that the parameter initialization one that use obtains is new, to the image needing test, for it extracts onesize image block, and forms same gray matrix, this gray matrix is input in sorter network, obtain initial classification results;

Step 6: result be mapped on original test pattern, utilizes a gray-scale value to divide the point being divided into tumour, finally utilizes an aftertreatment while merging connected region, delete some isolated lung neoplasm regions;

Step 7: net result is mapped on former figure, and statistics segmentation accuracy.

It should be noted that:

For different image data sets, the setting of the model parameter of degree of depth study often has a great difference, needs to be determined by experiment best parameter.

The invention has the beneficial effects as follows:

Degree of depth study introduces in the auto Segmentation of brain tumor by the present invention, proposes and a kind ofly directly from the gray-scale value of Pixel-level, directly can take out high-level characteristic, and accordingly to the transaction module that brain tumor is split; The present invention has the following advantages:

1, compared with traditional brain tumor dividing method, the present invention does not only need, for other features of MRI image zooming-out, also not need artificial intervention, is automatic, a unsupervised brain tumor segmenting system, carrying high-precision while, also substantially increase the efficiency of segmentation;

2, compared with traditional Segmentation by Fuzzy Clustering method, the present invention can distinguish the close brain tumor region of gray-scale value and blood vessel well, does not need manually to determine cluster numbers, and provides the higher accuracy of separation.

Accompanying drawing explanation

Fig. 1 is the brain tumor automatic division method process flow diagram based on degree of depth study provided by the invention;

Fig. 2 is the theory diagram of degree of depth study pre-training;

Fig. 3 is the theory diagram of degree of depth study fine setting.

Specific implementation

The brain tumor sorting technique based on degree of depth study that the present invention proposes specifically comprises the following steps:

Step one: carry out pre-service to the MRI section that patient comprises brain tumor, cutting edge reduces calculated amount, and promotes contrast;

Step 2: 25 × 25 image blocks around it are extracted to the point in image, in order to enable the number of 2 class points at the same order of magnitude, to the point not being tumour, every 4 pixel extraction image block, and to being the point of tumour, each point extracts an image block, and is the row vector of one 625 dimension by the gradation conversion of each image block, obtains altogether N number of row vector x ₁, x ₂, x ₃..., x _n, be arranged in order from top to bottom and form input matrix X:

Step 3: as Fig. 2, set deep learning network only comprises a hidden layer, and the number of hidden nodes is 100, using matrix X as input, pre-training is carried out by noise reduction automatic coding, first random initializtion parameter, wherein W is the weight matrix of 625 × 100, b is the offset vector of 100 dimensions, one is utilized to determine to map input X is mapped as a hiding potential expression Y, the potential expression Y obtained is that Function Mapping returns the reconstruction vector Z become in an input space by a parameter again, meets constraint condition, each training sample x _iall be mapped to as corresponding y _i, reconstruct becomes z again _i, then by optimized reconstruction model parameter, average reconstruction error is minimized:

${\mathrm{\θ}}^{*},{\mathrm{\θ}}^{\′*}=\underset{\mathrm{\θ},{\mathrm{\θ}}^{\′}}{\arg \min }{E}_{q^0\left(X\right)}\left[L_H(X,g_{{\mathrm{\θ}}^{\′}}\left(f_{\mathrm{\θ}}\left(X\right)\right))\right]$

Wherein L _hbe loss function, rebuild cross entropy, that is:

$L_H(x,z)=H\left(B_x\right|\left|B_z\right)=-\underset{k=1}{\overset{d}{\mathrm{\Σ}}}[x_k\log z_k+(1-x_k)\log (1-z_k)]$

Obtain initial network parameter;

Step 4: as Fig. 3, increases an output layer at the top of network, by there being the data of label, utilizing backpropagation and gradient descent method to finely tune the parameter of whole network from top to bottom, obtaining final argument;

Step 5: the sorter network that the parameter initialization one that use obtains is new, to the image needing test, for it extracts onesize image block, and forms same gray matrix, this gray matrix is input in sorter network, obtain initial classification results;

Step 6: result is mapped on original test pattern, a gray-scale value is utilized to divide the point being divided into tumour, if certain any gray scale is greater than this gray-scale value, it is considered to be the point of brain tumor, otherwise not the point of brain tumor, finally utilize an aftertreatment while merging connected region, delete some isolated lung neoplasm regions;

Step 7: net result is mapped on former figure, and statistics segmentation accuracy.

Above the brain tumor automatic division method based on degree of depth study that the present invention proposes is described in detail, but the description of embodiment is only for explaining method of the present invention and core concept thereof, so that the technician of this technology neck understands the present invention, but should be clear, the invention is not restricted to the scope of embodiment, to those skilled in the art, as long as various change to limit and in the spirit and scope of the present invention determined in appended claim, these changes are apparent, all innovation and creation utilizing the present invention to conceive are all at the row of protection.

Claims (3)

1., based on a brain tumor automatic division method for degree of depth study, it is characterized in that, comprise the following steps:

Step one: carry out pre-service to Magnetic resonance imaging (Magnetic ResonanceImaging, the MRI) section that patient comprises brain tumor, cutting edge reduces calculated amount, and promotes contrast;

Step 2: extract its image block to the point in image, is the vector of an one dimension by the gradation conversion of each image block, is arranged in order formation input matrix;

Step 3: the number of hidden nodes of set deep learning network, using matrix as input, first carries out pre-training, obtains initial network parameter;

Step 4: increase an output layer at the top of network, utilizes the data of tape label to finely tune network parameter and obtains final parameter;

Step 5: the sorter network that the parameter initialization one that use obtains is new, to the image needing test, for it extracts onesize image block, and forms same gray matrix, this gray matrix is input in sorter network, obtain initial classification results;

Step 6: result be mapped on original test pattern, utilizes a gray-scale value to divide the point being divided into tumour, finally utilizes an aftertreatment while merging connected region, delete some isolated lung neoplasm regions;

Step 7: net result is mapped on former figure, and statistics segmentation accuracy.

2. the brain tumor automatic division method based on degree of depth study according to claim 1, is characterized in that: described extraction MRI section being carried out to image block in described step 2, and concrete steps are as follows:

Arrange a step-length step, travel through whole MRI and cut into slices, every step pixel chooses a point, extract the image block of 25 × 25 centered by this point, when extracting image block for training set, to the point not being tumour, step=5, and to being the point of tumour, step=1, when extracting image block for test set, step=2, be the row vector of one 625 dimension by the gradation conversion of each image block, obtain altogether N number of row vector x ₁, x ₂, x ₃..., x _n, be arranged in order from top to bottom and form input matrix X:

$X=\left(\begin{array}{c}{x}_1\\ x_2\\ \·\\ \·\\ \·\\ x_N\end{array}\right)={\left(\begin{array}{cccc}{x}_{\mathrm{1,1}}& x_{\mathrm{1,2}}& \·\·\·& x_{\mathrm{1,625}}\\ x_{\mathrm{2,1}}& x_{\mathrm{2,2}}& \·\·\·& x_{\mathrm{2,625}}\\ \·& \·& & \·\\ \·& \·& & \·\\ \·& \·& & \·\\ x_{N,1}& x_{N,2}& \·\·\·& x_{N,625}\end{array}\right)}_{N\×625}$

3. the brain tumor automatic division method based on degree of depth study according to claim 1, it is characterized in that: the described pre-training to network in described step 3, concrete steps are as follows:

(1) set deep learning network only comprises a hidden layer, and the number of hidden nodes is 100, using matrix X as input;

(2) carry out pre-training by noise reduction automatic coding, first random initializtion parameter θ={ W, b}, wherein W is the weight matrix of 625 × 100, and b is the offset vector of 100 dimensions, utilizes one to determine to map Y=f _θ(X) input X is mapped as a hiding potential expression Y by=s (WX, b);

(3) the potential expression Y obtained is θ '={ W ', b ' } function Z=g by a parameter again _{θ '}(Y)=s (W ' Y+b ') maps back the reconstruction vector Z become in an input space, and W ' meets constraint condition W '=W ^t;

(4) each training sample x _iall be mapped to as corresponding y _i, reconstruct becomes z again _i, then by optimized reconstruction model parameter, average reconstruction error is minimized:

${\mathrm{\θ}}^{*},{\mathrm{\θ}\′}^{*}=\underset{\mathrm{\θ},{\mathrm{\θ}}^{\′}}{\arg \min }{E}_{q^o\left(X\right)}\left[L_H(X,g_{{\mathrm{\θ}}^{\′}}\left(f_{\mathrm{\θ}}\left(X\right)\right))\right]$

Wherein L _hbe loss function, rebuild cross entropy, that is:

$L_H(x,z)=H\left(B_x\right|\left|B_x\right)=-\underset{k=1}{\overset{d}{\mathrm{\Σ}}}[x_k\log z_k+(1-x_k)\log (1-z_k)]$

Obtain initial network parameter.