CN104794708A - Atherosclerosis plaque composition dividing method based on multi-feature learning - Google Patents

Abstract

A method for segmenting atherosclerotic plaque components based on multi-feature learning, comprising the following steps: 1) acquiring multi-sequence arterial plaque images; 2) image preprocessing; 3) plaque feature extraction and description; 4) classification Trainer training: the classifier model is a support vector machine (SVM); the last four sequences of MRI arterial plaque images generate a total of 4 feature sets, each feature set is trained to generate a classifier, and finally 4 classifiers are obtained, namely the lipid classifier, Fibrous classifier, calcium classifier, hemorrhage classifier; 5) Trainer optimization; 6) For a certain pixel of plaque tissue, use four classifiers to calculate the degree of its belonging to different tissues, and sort according to the obtained degree of membership competition , and take the component with the highest degree of membership as the final result. The invention has good precision and high robustness.

Description

A method for segmentation of atherosclerotic plaque components based on multi-feature learning

technical field

The invention belongs to the field of medical image processing, and in particular relates to the application of computer aided and artificial intelligence technology in medical image analysis.

Background technique

At present, there is no unified framework and technical means for the segmentation of atherosclerotic plaque components. The main reason is that the current imaging technology of vascular plaque is not mature enough, which makes the imaging quality of vascular plaque (such as contrast, sharpness of edges, etc.) is not high. In addition, current vascular images generally have grayscale images or pseudo-color images, which makes it difficult for computers to represent plaque information through quantitative digital quantities. In order to solve these problems, scholars at home and abroad have done a lot of research work, the most representative of which is the plaque segmentation method based on Bayesian theory proposed by F. Liu et al. This method mainly uses the gray level of the plaque The value, the position of the plaque in the blood vessel, and the shape information are used as features, and then a Gaussian model is trained based on these features, and the Bayesian posterior probability formula is used to calculate the probability that the plaque belongs to a certain component to complete the component segmentation. However, this method is very sensitive to image noise, and training a model that can describe plaque characteristics requires a large number of training samples. Some researchers use fuzzy-clustering method to segment plaque components. This method uses the gray value of arterial plaque as a feature and then uses fuzzy clustering to calculate the probability of plaque belonging to different components. To express the information of plaques, secondly, the method is less robust, and the result of fuzzy clustering is easily affected by plaque components. Domestic researcher Qian Huaming proposed a region growing algorithm combined with time domain information. This algorithm uses the gray level difference between images as the growth condition, and combines the gray level information as the growth condition, and then uses the morphological correlation method to The segmentation results are corrected to extract the segmentation boundaries of different components, but this method also has the defects of insufficient expression of plaque features and low robustness of the algorithm.

Contents of the invention

In order to overcome the disadvantages of poor accuracy and low robustness of existing atherosclerotic plaque component segmentation methods, the present invention provides a multi-feature learning-based atherosclerotic model with good accuracy and high robustness. Segmentation of sclerotic plaque components.

The technical solution adopted by the present invention to solve its technical problems is:

A method for segmentation of atherosclerotic plaque components based on multi-feature learning, the segmentation method comprising the following steps:

1) Acquiring multi-sequence arterial plaque images;

2) Image preprocessing

Select the ROI centered on the inner wall of the artery as the research area, adjust the window width and window level of the gray value of the selected ROI, and then denoise the ROI;

3) Plaque feature extraction and description

The characteristics of the plaque are described from the gray value of the plaque image, the Gaussian blur feature, the gradient value, the shortest distance between the plaque and the inner and outer walls of the blood vessel, the Laplace feature of the image and the texture feature of the plaque;

4) Classifier training: The classifier model is a support vector machine SVM, and the training process is as follows:

(4.1) Collection and calibration of training samples

In the training sample, the T1 sequence image is used as the reference image, and the manual calibration of the atherosclerotic plaque components by the doctor is performed on the T1 image, and the plaque components of the other four sequences T1GD, T2, PD, and STIR are mapped according to the results of T1 calibration on it;

(4.2) Sample preprocessing

First, the gray value stretching operation is performed on the sample image so that the gray value distribution range is in the range of [0,255], then the image is denoised, and finally the doctor's calibration result of the plaque components is binarized Processing, that is, storing each component of the plaque calibrated by the doctor in the binary matrix;

(4.3) Further classification and feature quantization processing of training samples

The further classification of training samples is to subdivide the samples into lipid sample set, fibrous sample set, calcium sample set and hemorrhage sample set;

The feature quantization process is to normalize the extracted plaque features, and all the features are normalized to form a feature vector, the feature vector As shown in formula (10).

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

In the formula, V represents the gray value of the plaque, G represents the Gaussian blur value of the plaque, GM represents the gray feature after the Gaussian blur of the gray value, L represents the Laplacian feature, and D represents the spatial position feature of the plaque, T represents the texture feature of the plaque.

(4.4) Classifier training, the last four sequences of MRI arterial plaque images generate 4 feature sets in total, each feature set is trained to generate a classifier, and finally 4 classifiers are obtained, namely lipid classifier, fibrous classifier, calcium classifier, hemorrhage classifier;

5) Trainer optimization

Divide the sample set into two parts, the training set and the test set, and then use the samples of the training set to train the classifier, use the obtained classifier to test the recognition rate of the test set, and continue to judge and cycle until the obtained trainer classification accuracy reaches Up to the maximum value, the classifier obtained at this time is considered to be the optimal classifier;

6) For a certain pixel of plaque tissue, four classifiers are used to calculate the degree of its belonging to different tissues, the process is as follows:

(6.1) First extract the features of each pixel of T1GD, T2, PD and STIR plaque tissue images;

(6.2) Put the plaque features extracted in (6.1) into the four classifiers optimized in step (5) respectively, and the four classifiers calculate the degree to which they belong to their components respectively;

(6.3) According to the competitive ranking of the membership degree obtained in (6.2), take the component with the highest membership degree as the final result.

Further, the segmentation method also includes the following steps:

7) Morphological operation of the result, the process is as follows:

(7.1) Store the result of component segmentation in four binary images, where 1 represents the tissue component and 0 represents the background in the binary image, which are the lipid component, fibrous component, calcium component, and hemorrhage component;

(7.2) Then perform corrosion operation on each binary image, and filter out isolated segmentation blocks;

(7.3) Then do the morphological opening operation to fill the holes in the component blocks and connect similar isolated blocks;

(7.4) Finally, image closing operation is performed to obtain atherosclerotic plaque components.

Further, in the step 3),

(a) The gray value is the gray value after the image has been denoised;

(b) Gaussian blur feature

The Gaussian blur feature refers to the gray value of a certain point of the plaque tissue after Gaussian blurring, and the extraction process is to use a 5*5 Gaussian kernel to perform convolution with the plaque tissue, as shown in formula (2).

$\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;\&sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In the formula, σ represents the scale of the Gaussian kernel, x represents the gray value of the image, I and G represent the original vascular plaque image and the image after Gaussian blur respectively, and “*” is the convolution symbol;

(c) Gradient features

Gradient feature GM of plaque tissue:

$\mathrm{<span\; class="notranslate">\; GM</span>}<span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

(d) Laplace feature L:

$\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

(e) Spatial location characteristics of plaques (distance to inner and outer walls)

D _min (I)＝min||C(I(x))-Lumen(I)|| (5)

D _max (I)＝max||C(I(x))-Lumen(I)|| (6)

(f) Texture features

Two types of texture features, the first-order statistics and the gray-level co-occurrence matrix features, are extracted, and there are seven first-order statistics in total, including the gray mean, standard deviation, coefficient of difference, skewness, kurtosis, histogram entropy and Luminance entropy; EtHis as formula (7) and brightness entropy EtBrt as formula (8):

$\mathrm{<span\; class="notranslate">\; Et\; His</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 255</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; EtBrt</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 255</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

Among them, P _i (i=0, 1, . . . , 255) is the probability of grayscale i appearing in the image, ω _i is the pixel intensity that is not 0 after normalization, and N is the total number of pixels in the original image.

The beneficial effects of the present invention are mainly manifested in: using multiple sequence MRI images as an image source, then extracting the characteristics of atherosclerotic plaque from various aspects, using machine learning methods to train classifiers, and finally completing Segmentation of plaque components; good accuracy and high robustness.

Description of drawings

Figure 1 is a flowchart of a method for segmenting atherosclerotic plaque components based on multi-feature learning.

FIG. 2 is a diagram of ROI selection and denoising results, wherein (a) is a schematic diagram of ROI selection, and (b) is a diagram of denoising results.

Fig. 3 is the registration result map of multi-sequence magnetic resonance images, in which (a) is the spatial difference map of T1 (left image) and PD (right image) before registration, (b) is the registration result, T1 (left image) Figure ) and PD (right panel) schematic.

Figure 4 is a flow chart of classifier training.

Figure 5 is a diagram of the classifier optimization process.

Figure 6 is a flowchart of plaque component identification.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to Figures 1 to 6, a method for segmenting atherosclerotic plaque components based on multi-feature learning includes the following steps:

1) Obtain multi-sequence arterial plaque images

Carotid arteries of T1 (T1 weighted imaging), T1GD (T1 contrast agent imaging), T2 (T2 weighted imaging), PD (proton density imaging), STIR (short inversion time inversion recovery imaging) sequences acquired at a magnetic field strength of 1.5T (artery vessel) magnetic resonance image as an example;

2) Image preprocessing

Since the collected image is a cross-sectional image of the entire artery, the location of the artery only accounts for about 2% of the entire image. At the same time, it will increase the amount of computation and reduce the efficiency of algorithm processing. Therefore, the first step in image preprocessing is to select a 60*60 pixel ROI (Region of Interest) centered on the inner wall of the artery as the research area. Adjust the window width and window level of the gray value of the selected ROI so that the gray range is distributed between (0--255) as in formula (1),

$\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; I</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}<span\; class="notranslate">\; \&times;</span>\mathrm{<span\; class="notranslate">\; 255</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Then the ROI is denoised, and the denoising adopted in the present invention is a denoising method based on wavelet decomposition and adaptive threshold adjustment. The processing results are shown in Figure 2. Figure 2(a) represents a T1 sequence image, in which the rectangular frame is the extracted ROI area, and Figure 2(b) is the result of denoising the ROI area.

During the multi-sequence image imaging process, due to the patient's respiration, pulsation, swallowing movements, and subtle tremors, etc., the multi-sequence images will cause rigid and non-rigid deformations. Identification of atherosclerotic plaque components is lethal. For example, a 1mm movement of a patient in real space will cause a shift of ten to dozens of pixels on the image. Therefore, the registration work between the sequence images is necessary. The present invention uses the open source software of elastix V4.7 image registration to complete the registration work, and its registration result is shown in Figure 3, Figure 3 (a) represents the T1 (left) and PD (right) sequence image registration before The difference, where the two irregular circles represent the inner and outer walls of the blood vessel in the reference image, Figure 3(b) shows the result after registration, it can be clearly seen that the inner and outer walls of the blood vessel in T1 and PD sequences are highly consistent;

3) Plaque feature extraction and description

In order to accurately and effectively describe the specificity of atherosclerotic plaques, this method separately analyzes the gray value of the plaque image, Gaussian blur feature, gradient value, the shortest distance from the plaque to the inner and outer walls of the blood vessel, and the pull-out of the image. The characteristics of plaques are described in terms of Plath features and texture features of plaques, respectively. The extraction method of these features will be described below.

(a) gray value

The gray value is the gray value after the image has been denoised. This feature is a direct reflection of the specificity of multi-sequence images for plaques. To a certain extent, it is the most valuable feature. Intuitive and widely used.

(b) Gaussian blur feature

The Gaussian blur feature in this method refers to the gray value of a certain point of the plaque tissue after Gaussian blurring, which eliminates the interference of imaging noise to a certain extent. The extraction process uses a 5*5 Gaussian kernel and plaque The organization performs convolution, as shown in formula (2).

$\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&pi;\&sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; *</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Among them, σ represents the scale of the Gaussian kernel, x represents the gray value of the image, I and G represent the original vascular plaque image and the image after Gaussian blur respectively, and "*" is the convolution symbol;

(c) Gradient features

Gradient feature GM of plaque tissue:

$\mathrm{<span\; class="notranslate">\; GM</span>}<span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; (</span>\frac{<span\; class="notranslate">\; \&PartialD;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

(d) Laplacian feature L:

$\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; G</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; +</span>\frac{{<span\; class="notranslate">\; \&PartialD;</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&PartialD;</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

(e) Spatial location characteristics of plaques (distance to inner and outer walls)

D _min (I)＝min||C(I(x))-Lumen(I)|| (5)

D _max (I)＝max||C(I(x))-Lumen(I)|| (6)

(f) Texture features

Extracting and analyzing the texture features of atherosclerotic plaques can help to classify the plaque types objectively, so as to make the related diagnosis more objective and accurate. In order to reflect the statistical law of the gray level distribution inside the plaque, this method extracts two types of texture features, the first-order statistics and the gray level co-occurrence matrix (GLCM) feature. A total of seven first-order statistics, including the mean gray value (Mean), standard deviation (STD), coefficient of variance (CoV), skewness (skewness, Skew), kurtosis (kurtosis, Kurtf ), and histogram entropy (EtHis) such as formula (7), brightness entropy (EtBrt) such as formula (8), and their calculation formulas are:

$\mathrm{<span\; class="notranslate">\; Et\; His</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 255</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; EtBrt</span>}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 255</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; N</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

Wherein, P _i (i=0, 1, ..., 255) is the probability that grayscale i appears in the image, ω _i is the pixel intensity that is not 0 after normalization, and N is the total number of pixels in the original image;

4) Classifier training

The purpose of classifier training is an approximation to the real model of the problem. The classifier model used in this method is support vector machine (SVM). This method is a classic machine learning method and is based on the VC of statistical learning theory. Based on the dimensional theory and the principle of structural risk minimization, according to the limited sample information, the optimal compromise is sought between the complexity of the model (i.e. the learning accuracy for a specific training sample) and the learning ability (i.e. the ability to identify any sample without error). , in order to obtain the best generalization ability (or generalization ability).

The training samples of the classifier are 1/2 randomly selected from the overall samples, and the remaining 1/2 are used to test the classification performance and optimize the classifier.

The training process of the classifier is shown in Figure 4, and the process is mainly divided into the following steps:

(4.1) Collection and calibration of training samples

The training samples mainly use five MRI sequences, including T1 (T1 weighted imaging), T1GD (T1 contrast agent imaging), T2 (T2 weighted imaging), PD (proton density imaging), and STIR (short inversion time inversion recovery imaging). The T1 sequence image is used as a reference image (T1GD, T2, PD, STIR sequence registration uses T1 as a reference). In addition, the manual calibration of atherosclerotic plaque components by doctors is also performed on the T1 image. The other four sequence plaques The components are mapped on it according to the results of T1 calibration.

(4.2) The work to be done in the sample preprocessing is to quantify the sample image and the calibration results of the sample by the doctor. In this process, the gray value stretching operation is first performed on the sample image so that the gray value distribution range is [0,255] Interval range, then denoise the image, and finally binarize the doctor’s calibration results of the plaque components, that is, store each component of the plaque calibrated by the doctor in the binary matrix.

(4.3) Further classification and feature quantization processing of training samples

The further classification of the training samples is to subdivide the samples into a lipid (fat) sample set, a fibrous (fiber) sample set, a calcium (calcification) sample set, and a hemorrhage (thrombus) sample set.

The feature quantization process is to normalize the extracted plaque features. The purpose is to make the distribution range of all extracted features quantized in the [-1,1] interval, because the feature objects of different feature components are different. There is a large difference in magnitude. However, in the linear classification function of SVM, the large feature component has a greater influence than the small feature component, but this does not mean that the large feature is more important, so it is necessary to normalize and unify the features. As shown in formula (9).

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

Among them, x _i is the characteristic component value, Represents the average value of the feature component x, N is the number of training samples;

All features are normalized to form a feature vector, the feature vector As shown in formula (10).

$\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; G</span>}}\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; ,</span>\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

In the formula, V represents the gray value of the plaque, G represents the Gaussian blur value of the plaque, GM represents the gray feature after the Gaussian blur of the gray value, L represents the Laplacian feature, D represents the spatial position feature of the plaque, and T Represents the texture features of the plaque.

(4.4) Classifier training, the last four kinds of sequence MRI arterial plaque images generate 4 feature sets altogether, each feature set can be trained to generate a classifier, finally get 4 classifiers, i.e. lipid classifier, fibrous classifier, calcium classifier, hemorrhage classifier.

The training of the classifier is actually a convex optimization problem, and a hyperplane H is optimized through training samples:

ω·x+b=0 (11)

Among them, ω is a weight vector, x is a feature vector, b is a parameter, and the hyperplane distinguishes positive and negative samples in the form of the largest boundary.

5) Classifier optimization

The four classifiers (lipid, fibrous, calcium, and hemorrhage) obtained during the classifier optimization process are optimized, and the optimization process is shown in Figure 5. The optimization process of the lipid classifier is used as an example to illustrate the optimization process.

First, the sample set is divided into two parts, the training set and the test set. The selection of these two parts is random, each accounting for 50%. Then the classifier is trained with the samples of the training set, and the test set is tested with the obtained classifier. The recognition rate is judged successively until the classification accuracy rate of the obtained lipid trainer reaches the maximum value. At this time, the obtained lipid classifier is considered to be the optimal classifier of lipid. Other component classifiers are optimized in the same way.

6) Plaque component identification and morphological manipulation

Plaque component identification is the core part of this method. In this method, four classifiers are used to calculate the degree of belonging to different tissues for a certain pixel of plaque tissue. The process is shown in Figure 6.

(6.1) First extract the features of each pixel in the T1, T1GD, PD, STIR plaque tissue images, for example, the texture features corresponding to the pixel point P include four sequences of P points in T1, T1GD, PD, STIR The texture features in T(P)=[T ₁ (P), T ₂ (P), T ₃ (P), T ₄ (P)] are extracted respectively. The other five features are extracted in the same way.

(6.2) Put the plaque features extracted in (6.1) into the four classifiers optimized in step (5), and the four classifiers respectively calculate the degree to which they belong to their components.

(6.3) According to the competitive ranking of the membership degree obtained in (6.2), take the component with the highest membership degree as the final result.

7) Morphological operation of the result

The purpose is to eliminate the influence of noise on the segmentation results, to exclude isolated tissue components and holes in the segmentation results. The specific operation is as follows:

(7.1) The result of component segmentation is stored in four binary images. In the binary image, 1 represents the tissue component, and 0 represents the background, which are the lipid component, fibrous component, calcium component, and hemorrhage component.

(7.2) Then perform corrosion operations on each binary image, and filter out isolated segmentation blocks (the number of pixels is less than or equal to ten), because too small component blocks are a kind of interference to the doctor's diagnosis, and this component is extremely harmful. May be caused by image noise,

(7.3) Then do the morphological opening operation. The purpose of the opening operation is to fill the holes in the component blocks and connect similar isolated blocks to increase their connectivity.

(7.4) Finally, the image closing operation is performed, which is to offset the expansion effect of the opening operation on the tissue components.

Claims (3)

1., based on an atherosclerotic plaque composition dividing method for multiple features study, it is characterized in that: described dividing method comprises the following steps:

1) multisequencing arteries patch image is obtained;

2) Image semantic classification

Choose ROI centered by arteries inwall as survey region, to window width, the adjustment of window position that the ROI chosen does gray-scale value, then denoising is done to ROI;

3) Patch properties extracts and describes

The feature of patch is described respectively from the gray-scale value of patch image, Gaussian Blur feature, Grad, patch to the minimum distance of Ink vessel transfusing outer wall, Laplce's feature of image and the textural characteristics of patch;

4) sorter training: sorter model is support vector machines, and training process is as follows:

(4.1) collection of training sample and demarcation

In training sample, T1 sequence image is as reference image, and it is operate on T1 image that doctor manually demarcates atherosclerotic plaque composition, and the plaque component of other four kinds of sequence T1GD, T2, PD and STIR is that the result of demarcating according to T1 maps thereon;

(4.2) sample preprocessing

First gray-scale value stretched operation is done to sample image, make its grey value profile scope [0,255] interval range, then be to image denoising sonication, finally that doctor done binary conversion treatment to the calibration result of plaque component, in the two values matrix that each composition of the patch namely doctor demarcated is stored in;

(4.3) training sample is classified and characteristic quantification process further

The further classification of training sample is that sample is subdivided into lipid sample set, fibrous sample set, calcium sample set and hemorrhage sample set;

Characteristic quantification process is that the Patch properties extracted is done normalized, and all features form a proper vector after normalization, this proper vector shown in (10):

$\stackrel{\&RightArrow;}{F}=[\stackrel{\&RightArrow;}{V},\stackrel{\&RightArrow;}{G},\stackrel{\&OverBar;}{G}\stackrel{\&RightArrow;}{M},\stackrel{\&RightArrow;}{L},\stackrel{\&RightArrow;}{T}]---\left(10\right)$

In formula, V represents the gray-scale value of patch, and G represents the Gaussian Blur value of patch, and GM represents the gray feature after gray-scale value Gaussian Blur, and L represents Laplce's feature, and D represents patch locus feature, and T represents the textural characteristics of patch.；

(4.4) sorter training, the symbiosis of last four kinds of sequence MRI artery plaque images becomes 4 feature sets, each feature set training generation sorter, finally obtain 4 sorters, i.e. lipid sorter, fibrous sorter, calcium sorter, hemorrhage sorter;

5) training aids optimization

Sample set is divided into training set and test set two parts, then the sample training sorter of training set is used, with the discrimination of the sorter verification test collection obtained, judge that circulation is gone down successively, until the training aids classification accuracy rate obtained reaches maximal value, optimum classifier thought by the sorter now obtained;

6) for some pixels of plaque, by four classifier calculated, it is under the jurisdiction of the degree of different tissues respectively, and process is as follows:

(6.1) feature extraction first T1GD, T2, PD and STIR plaque image pointed out in each pixel out;

(6.2) Patch properties extracted in (6.1) is put in four sorters after optimizing in step (5) respectively, four kinds of sorters calculate its degree being under the jurisdiction of its composition respectively;

(6.3) according to the degree of membership competition sequence in (6.2) that obtain, using composition the highest for degree of membership as net result.

2. as claimed in claim 1 a kind of based on multiple features study atherosclerotic plaque composition dividing method, it is characterized in that: described dividing method is further comprising the steps of:

7) morphological operation of result, process is as follows:

(7.1) result of composition segmentation is stored in 4 bianry images, in bianry image, 1 represents structural constituent, and 0 represents background, is lipid composition, fibrous composition, calcium composition respectively, hemorrhage composition;

(7.2) then erosion operation is done to each bianry image, isolated block filtering;

(7.3) then do morphology opening operation, fill up the cavity in component block and be communicated with close isolated block;

(7.4) finally do closing operation, obtain atherosclerotic plaque composition.

3. as claimed in claim 1 a kind of based on multiple features study atherosclerotic plaque composition dividing method, it is characterized in that: described step 3) in,

A () gray-scale value is at the gray-scale value of image after denoising;

(b) Gaussian Blur feature

Described Gaussian Blur feature refers to the gray-scale value of plaque a bit after Gaussian Blur, and leaching process is that the gaussian kernel of an employing 5*5 and plaque do convolution, shown in (2);

$G(x,\mathrm{\&sigma;})=\frac{1}{\sqrt{2\mathrm{\&pi;\&sigma;}}}{e}^{-\frac{{\left|\right|x\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}}*I---\left(2\right)$

In formula, σ represents the yardstick of gaussian kernel, and x represents the gray-scale value of image, and I, G distinguish the image after representative image primitive vessel patch image and Gaussian Blur, and " * " is convolution symbol;

(c) Gradient Features

The Gradient Features GM of plaque:

$\mathrm{GM}=\sqrt{{\left(\frac{\&PartialD;\left(G(x,\mathrm{\&sigma;})\right)}{\&PartialD;x}\right)}^2+{\left(\frac{\&PartialD;\left(G(x,\mathrm{\&sigma;})\right)}{\&PartialD;y}\right)}^2}---\left(3\right)$

(d) Laplce feature L:

$L=\frac{{\&PartialD;}^2\left(G(x,\mathrm{\&sigma;})\right)}{\&PartialD;x^2}+\frac{{\&PartialD;}^2(I,(x,\mathrm{\&sigma;}))}{\&PartialD;x^2}---\left(4\right)$

(e) patch locus feature (distance between inside and outside wall)

D _min(I)＝min||C(I(x))-Lumen(I)|| (5)

D _max(I)＝max||C(I(x))-Lumen(I)|| (6)

(f) textural characteristics

Be extracted first order statistic and gray level co-occurrence matrixes feature two class textural characteristics, totally seven first order statistic, comprising patch interior intensity average, standard deviation, coefficient of variation, the degree of bias, kurtosis, entropy of histogram and brightness entropy; EtHis such as formula (7) and brightness entropy EtBrt such as formula (8):

$\mathrm{EtHis}=-\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}{p}_i{\log }_2\left(p_i\right)---\left(7\right)$

$\mathrm{EtBrt}=-\underset{i=0}{\overset{255}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&omega;}}_i{\log }_2{\mathrm{\&omega;}}_i}{{\log }_{2}N}---\left(8\right)$

Wherein, P _i(i=0,1 ..., 255) be the probability that GTG i occurs in the picture, ω _ibe after normalization be not 0 image pixel intensities, N is the sum of all pixels of original image.