CN105825509A - Cerebral vessel segmentation method based on 3D convolutional neural network - Google Patents

Abstract

The invention discloses a cerebral vessel segmentation method based on a 3D convolutional neural network, which relates to fields such as machine learning, pattern recognition and image processing. The method comprises steps: firstly, well-marked cerebral vessel angiograms are sequentially stacked into a 3D matrix, patches of 25*25*25 are taken for vessel points, the same number of non-vessel point patches with the same size are randomly taken, and training data are obtained; then, the training data are inputted to the 3D convolutional neural network for training, and a training model is obtained; a patch of 25*25*25 is taken from each pixel point in the actual vessel angiogram sequence image to be inputted to the model, a classification tag is obtained, the classification tag is stretched to be an image with the same size, and the image is the segmented cerebral vessel image. The method has effects of high accuracy and good generic degree.

Description

Cerebrovascular dividing method based on 3D convolutional neural networks

Technical field

The invention belongs to computer vision field, more specifically, relate to a kind of based on 3D convolutional neural networks, for the blood vessel segmentation method of angiographic image.

Background technology

Along with the raising of people's living standard, angiopathy has become one of primary disease of harm health of people.Blood vessel is very important organ in human body, once occurs pathological changes will drastically influence the orthobiosis of people.Therefore, the early prevention of angiopathy, diagnose and treatment highlights important, by means of modern medicine imaging methods, blood vessel checks, analyzes and assists treatment also become the focus of this area research.

At present, clinical technology for angiopathy inspection and diagnosis mainly has: digital subtraction angiography (DigitalSubtractionAngiography based on ray, DSA), based on ultrasonic colored transcranial doppler image (ColorTranscranialDoppler, CTD), magnetic resonance angiography (MagneticResonanceAngiography, MRA), computed tomography angiography (ComputedTomographyAngiography) etc. are above several.

And convolutional neural networks achieves immense success in the classification and segmentation task of natural image, the research that the method that the degree of depth learns is applied to medical image the most in recent years also gets more and more.Convolutional neural networks is the one of artificial neural network, become the study hotspot of current speech analysis and field of image recognition, its weights are shared network structure and are allowed to be more closely similar to biological neural network, reduce the complexity of network model, decrease the quantity of weights.What this advantage showed when the input of network is multidimensional image becomes apparent from, and makes the image can be directly as the input of network, it is to avoid complicated feature extraction and data reconstruction processes in tional identification algorithm.Convolutional network is one multilayer perceptron of particular design for identification two-dimensional shapes, and this network structure has height invariance to the deformation of translation, proportional zoom, inclination or other forms.

Summary of the invention

It is an object of the invention to design a kind of convolutional neural networks that can extract image information from 3 dimensions, for splitting cerebrovascular image, 2 traditional dimension convolutional neural networks have been done 3D extension by the method, it is possible to extract more implicit information in comparable vascular contrastographic picture.

For achieving the above object, a kind of sorting technique based on 3D convolutional neural networks of the present invention, mainly include two stages: training stage and forecast period；Training stage is to input the 3-dimensional cerebral angiography image with label into 3D convolutional neural networks, training network parameter, obtains training pattern.Forecast period is to test data according to this model prediction trained, and splits cerebrovascular.

Training stage is as it is shown in figure 1, concrete techniqueflow is as follows:

Step one: first the puncta vasculosa in cerebral angiography image is carried out label, then the image stack that labelling is good is built up 3-dimensional matrix, the patch that each puncta vasculosa takes 25 × 25 × 25 is positive sample, and the path of the non-vascular point then taking equal number in 3-dimensional matrix at random is negative sample；

Step 2: input training sample, is normalized this sample, then carries out the training of neutral net；

Step 3: arrange 20 convolution kernels at ground floor, each convolution kernel size is 6 × 6 × 6, and uses full connected mode to be connected with input layer to carry out convolution, and obtaining 20 sizes is feature map of 20 × 20 × 20；

Step 4: each feature map of ground floor is carried out down-sampling spatially, and sampling unit is 2, obtaining 20 sizes is feature map of 10 × 10 × 10, is the second layer；

Step 5: using the 3D convolution kernel of 5 × 5 × 5 sizes to carry out 3D convolution each feature map of the second layer, being output as 40 sizes is feature map of 6 × 6 × 6, and connected mode uses full connection, and this is third layer；

Step 6: each feature map of third layer is carried out down-sampling spatially, and sampling unit is 2, obtaining 40 sizes is feature map of 3 × 3 × 3, is the 4th layer；

Step 7: each feature map of the 4th layer uses the 3D convolution kernel of 3 × 3 × 3 sizes carry out 3D convolution, and being output as 80 sizes is feature map of 1 × 1 × 1, connected mode uses full connection, and this is layer 5；

Step 8: each feature map of layer 5 is drawn into the characteristic vector that dimension is 80, then the random parameter matrix to this feature vector premultiplication one 128 × 80, obtains the characteristic vector of one 128 dimension, and this is layer 6；

Step 9: the characteristic vector obtained by layer 6 inputs into a LogisticRegression grader, is output as the floating number of 0 to 1, represents that input center of a sample is the probability of puncta vasculosa, and this is layer 7；

Step 10: be adjusted the calculating parameter of each layer by BP (back propagation) algorithm, makes finally to predict that the error function of label and training label is minimum, when error meets the condition of convergence, iteration terminates, and obtains training pattern.

Forecast period

Step 11: each pixel testing image carries out the operation of the patch taking 25 × 25 × 25 sizes in step one, obtains test sample, test sample inputted into training pattern, obtains predicting label；

Step 12 a: threshold value (such as 0.8) is set, will be greater than being set to 1 equal to the prediction label of threshold value, prediction label less than threshold value is set to 0, and then being replied according to correspondence position by all of prediction label is script image size, is the blood vessel segmentation image obtained.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of 3D convolution；

Fig. 2 is the overall construction drawing of 3D convolutional neural networks.

Below in conjunction with the accompanying drawings the detailed description of the invention of the present invention is described, in order to those skilled in the art is more fully understood that the present invention.Requiring particular attention is that, in the following description, the detailed description of the known function and design that may desalinate main contents of the present invention will be left in the basket.

In the present embodiment, a kind of cerebrovascular dividing method based on 3D convolutional neural networks of the present invention mainly includes following link: 1. propagated forward, 2. back propagation.

Wherein during propagated forward, 3D convolution operation realizes equation below:

$v_{ij}^{xyz}=\tanh (b_{ij}+\underset{m}{\Σ}\underset{p=0}{\overset{P_i-1}{\Σ}}\underset{q=0}{\overset{Q_i-1}{\Σ}}\underset{r=0}{\overset{R_i-1}{\Σ}}{\mathrm{\ω}}_{ijm}^{pqr}{v}_{(i-1)m}^{(x+p)(y+q)(z+r)})$

Wherein P_i,Q_i,R_iFor the size of convolution kernel,It is that convolution kernel is connected in front layer m-th feature map coordinate for (i, j, parameter m).

Back propagation updates weights and uses BP algorithm, the BP algorithm that 3D convolutional neural networks uses is different from traditional BP algorithm, and due in convolutional neural networks convolutional layer and down-sampling layer be alternately present, therefore the calculating of the error penalty term δ of convolutional layer and down-sampling layer is different；

Error penalty term δ for output layer neuron is:

δ^L=f ' (u^L)ο(yⁿ-tⁿ)

Wherein, yⁿRepresent the output vector that neutral net is actual, tⁿFor the physical tags that sample is corresponding, L represents last layer of classification layer, and ο represents dot product.

The error penalty term of l layer is as follows:

δ^l=(W^l+1)^Tδ^l+1ο f ' (u^l)

Then the gradient of weighting parameter is calculated by calculated error penalty term.Vector pattern can be obtained by the inner product calculating input vector:

$\frac{\∂E}{\∂W^l}=x^{l-1}{\left({\mathrm{\δ}}^l\right)}^T$

${\mathrm{\ΔW}}^l=-\mathrm{\η}\frac{\∂E}{\∂W^l}$

It can thus be seen that the error of preceding layer depends on the error of later layer, i.e. calculate gradient and progressively calculated to front layer by rear layer.

For 3D convolutional neural networks, being calculated as follows of the error penalty term of its concrete convolutional layer:

Wherein, C is constant, represents the yardstick of down-sampling, and up is up-sampling function, each dimension of matrix will extend C times.

Being calculated as follows of the error penalty term of down-sampling layer:

Wherein, conv3 represents 3-dimensional convolution, and rot180 represents convolution kernel revolves turnback, and full represents the mode on convolution border.

After obtaining the error penalty term of each layer, it is possible to the gradient of calculating parameter:

$\frac{\∂E}{\∂k_{ij}^l}=rot180\left(conv3\left(x_i^{l-1},rot180\left({\mathrm{\δ}}_j^l\right){,}^{\′}{\mathrm{valid}}^{\′}\right)\right)$

$\frac{\∂E}{\∂b_j}=\underset{u,v}{\Σ}{\left({\mathrm{\δ}}_j^l\right)}_{uv}$

Wherein, valid is the mode on convolution border, represents and border is not done any process.

Compared to other traditional cerebrovascular dividing method such as Threshold segmentation, region growing, actively profile, our method can be extracted the three-dimensional feature of cerebrovascular image, can reach more preferable segmentation effect.And owing to neutral net comprises quantity of parameters, therefore the training pattern obtained has good generalization ability, to various cerebral angiography images (such as CTA, MRA) all with higher segmentation accuracy rate.

Claims (3)

1. a kind of sorting technique based on 3D convolutional neural networks of the present invention, mainly includes two stages: training stage and forecast period.

Step one: first the puncta vasculosa in cerebral angiography image is carried out label, then builds up 3-dimensional matrix by the image stack that labelling is good, takes the patch of 25 × 25 × 25 as sample；

Step 2: input training sample, is normalized this sample, then carries out the training of neutral net；

Step 3: arrange 20 convolution kernels at ground floor, each convolution kernel size is 6 × 6 × 6, and uses full connected mode to be connected with input layer to carry out convolution, and obtaining 20 sizes is feature map of 20 × 20 × 20；

Step 4: each feature map of ground floor is carried out down-sampling spatially, and sampling unit is 2, obtaining 20 sizes is feature map of 10 × 10 × 10, is the second layer；

Step 5: using the 3D convolution kernel of 5 × 5 × 5 sizes to carry out 3D convolution each feature map of the second layer, being output as 40 sizes is feature map of 6 × 6 × 6, and connected mode uses full connection, and this is third layer；

Step 6: each feature map of third layer is carried out down-sampling spatially, and sampling unit is 2, obtaining 40 sizes is feature map of 3 × 3 × 3, is the 4th layer；

Step 7: each feature map of the 4th layer uses the 3D convolution kernel of 3 × 3 × 3 sizes carry out 3D convolution, and being output as 80 sizes is feature map of 1 × 1 × 1, connected mode uses full connection, and this is layer 5；

Step 8: each feature map of layer 5 is drawn into the characteristic vector that dimension is 80, then the random parameter matrix to this feature vector premultiplication one 128 × 80, obtains the characteristic vector of one 128 dimension, and this is layer 6；

Step 9: the characteristic vector obtained by layer 6 inputs into a LogisticRegression grader, is output as the floating number of 0 to 1, represents that input center of a sample is the probability of puncta vasculosa, and this is layer 7；

Step 10: be adjusted the calculating parameter of each layer by BP (back propagation) algorithm, is made finally to predict that the error function of label and training label is minimum, obtains training pattern.

2. the training pattern obtained according to step 10, carries out sample predictions:

Step 11: each pixel testing image carries out the operation of the patch taking 25 × 25 × 25 sizes in step one, obtains test sample, test sample inputted into training pattern, obtains predicting label；

Step 12: arrange a threshold value, will be greater than being set to 1 equal to the prediction label of threshold value, the prediction label less than threshold value is set to 0, and then being replied according to correspondence position by all of prediction label is script image size, is the blood vessel segmentation image obtained.

3. sorting technique based on 3D convolutional neural networks, mainly convolutional layer and the interactive stacking of pond layer as claimed in claim 1, and by BP Algorithm for Training parameter, obtain feature and state for classifying.