CN112562058B

CN112562058B - Method for quickly establishing intracranial vascular simulation three-dimensional model based on transfer learning

Info

Publication number: CN112562058B
Application number: CN202011322230.1A
Authority: CN
Inventors: 贾广; 李檀平; 张向淮; 郝嘉雪; 黄旭楠; 高敬龙; 张小玲; 汤敏; 谭丽娜; 苗启广; 梁小凤; 王泽�; 张昱; 张艺飞
Original assignee: Xidian University
Current assignee: Xidian University
Priority date: 2020-11-23
Filing date: 2020-11-23
Publication date: 2024-02-23
Anticipated expiration: 2040-11-23
Also published as: CN112562058A

Abstract

The invention discloses a method for quickly establishing an intracranial vascular simulation three-dimensional model based on transfer learning, which comprises the following steps: acquiring a bright blood image group and an enhanced black blood image group of an intracranial vascular part; preprocessing each bright blood image and the corresponding enhanced black blood image to obtain a first bright blood image and a first black blood image; performing image registration on the first bright blood image by using a registration method of mutual information based on Gaussian distribution sampling and an image pyramid; obtaining MIP images in all directions from the registered bright blood image group by using a maximum intensity projection method; taking the MIP image as a target domain, taking the fundus blood vessel image as a source domain, and obtaining a two-dimensional blood vessel segmentation image by using a migration learning method; back-projecting and synthesizing the two-dimensional vessel segmentation graphs in three directions to obtain first three-dimensional vessel volume data; and obtaining an intracranial blood vessel simulation three-dimensional model by using the second three-dimensional blood vessel volume data corresponding to the registered bright blood image group. The invention can obtain the integral state of the intracranial blood vessel simply, conveniently, quickly and intuitively in clinic.

Description

Method for quickly establishing intracranial vascular simulation three-dimensional model based on transfer learning

Technical Field

The invention belongs to the field of image processing, and particularly relates to a method for quickly establishing an intracranial blood vessel simulation three-dimensional model based on transfer learning.

Background

Medical data shows that 34 provincial (including the harbor australia) residents in china have the first cause of death from 1990 to 2017 to be stroke. Cerebral apoplexy is a series of symptoms generated by cerebral tissue necrosis caused by intracranial vascular rupture, stenosis or blockage, including cerebral hemorrhage, cerebral infarction and the like, and if the treatment is not timely, the patient may die; even if the treatment is in time, the patient may be disabled.

Currently, for clinical assessment of the extent of intracranial vascular lesions and the extent of vascular stenosis, methods based on lumen imaging, such as digital subtraction angiography (Digital Subtraction Angiography, DSA), CT vascular imaging (Computed Tomography Angiography, CTA), magnetic resonance vascular imaging (Magnetic Resonance Angiography, MRA), and High-resolution magnetic resonance vascular imaging (High-Resolution Magnetic Resonance Angiography, HRMRA) are commonly used. The intracranial artery blood vessel is connected with the carotid artery and the vertebral artery, forms a ring structure at the bottom of the brain, has special structure form, is in a shape of a curve and has extremely thin wall thickness. The path of the intracranial arterial vessel can be clearly depicted by the magnetic resonance vascular imaging technology.

The magnetic resonance blood vessel imaging technology (MRA or HRMRA) is used as a noninvasive imaging method for a patient, the intracranial blood vessel wall structure can be clearly detected and analyzed, the scanned magnetic resonance image has high resolution for soft tissues, no bone artifact and good image quality, and tissue structures with different imaging characteristics can be obtained by using various sequence scanning, so that the method has obvious superiority in displaying intracranial blood vessels.

The images corresponding to the bright blood sequence and the black blood sequence obtained by the magnetic resonance vascular imaging technology are two-dimensional images, so that the method has limitation. Is not beneficial to obtaining the integral state of the intracranial blood vessel simply, conveniently, quickly and intuitively in clinic.

Disclosure of Invention

To obtain the intracranial blood vessel integral state simply, quickly and intuitively in clinic. The embodiment of the invention provides a method for quickly establishing an intracranial blood vessel simulation three-dimensional model based on transfer learning. Comprising the following steps:

acquiring a bright blood image group and an enhanced black blood image group of an intracranial vascular part; wherein the bright blood image group and the enhanced black blood image group comprise K bright blood images and K enhanced black blood images respectively; the bright blood image group corresponds to the images in the enhanced black blood image group one by one; k is a natural number greater than 2;

Taking each bright blood image and the corresponding enhanced black blood image as an image pair, and preprocessing each image pair to obtain a first bright blood image and a first black blood image of the image pair;

aiming at each first bright blood image, carrying out image registration by using a registration method of image pyramid and mutual information based on Gaussian distribution sampling by taking a corresponding first black blood image as a reference to obtain a registered bright blood image group comprising K registered bright blood images;

projecting the registered bright blood image group in three preset directions by using a maximum intensity projection method to obtain MIP images in all directions;

taking MIP images in all directions as a target domain, taking fundus blood vessel images as a source domain, and obtaining two-dimensional blood vessel segmentation images corresponding to the MIP images in all directions by using a migration learning method;

synthesizing the two-dimensional vascular segmentation maps in three directions by using a back projection method to obtain first three-dimensional vascular volume data; wherein, the voxel value of the blood vessel part in the first three-dimensional blood vessel volume data is 0, and the voxel value of the non-blood vessel part is minus infinity;

and obtaining an intracranial blood vessel simulation three-dimensional model based on the first three-dimensional blood vessel volume data and the second three-dimensional blood vessel volume data corresponding to the registered bright blood image group.

In the scheme provided by the embodiment of the invention, firstly, the image registration is carried out on the bright blood image and the enhanced black blood image obtained by the magnetic resonance blood vessel imaging technology by adopting a registration method of mutual information based on Gaussian distribution sampling and an image pyramid, so that the registration efficiency can be improved, and the registration precision of the images can be improved layer by layer from low resolution to high resolution. According to the method, a MIP image in each direction is obtained through a maximum intensity projection method on the registered bright blood image, and the characteristic that the bright blood sequence MIP image of the intracranial blood is similar to the fundus blood image is utilized, on one hand, a network model for fundus blood vessel segmentation is trained by using labeling samples of the fundus blood image, on the other hand, characteristic MIP images with the same sample distribution as the fundus blood image are obtained through characteristic transformation on the bright blood sequence MIP image of the intracranial blood, and a characteristic migration mode is adopted to migrate the network model pre-trained by a fundus blood vessel segmentation task into the intracranial blood vessel segmentation task, so that a two-dimensional blood vessel segmentation image in each direction corresponding to the bright blood sequence MIP image of the intracranial blood is obtained. According to the embodiment of the invention, the research thought of migration learning is applied to the field of intracranial vessel segmentation, and a relatively accurate vessel segmentation effect can be obtained. And then, obtaining first three-dimensional blood vessel volume data by using a back projection method, and realizing an intracranial blood vessel simulation three-dimensional model by using second three-dimensional blood vessel volume data corresponding to the registered bright blood image group. The intracranial blood vessel simulation three-dimensional model can simulate the morphology of the intracranial blood vessel, realizes the three-dimensional visualization of the intracranial blood vessel, does not need a doctor to restore the vascular tissue structure, the disease characteristics and the like through imagination, can facilitate the doctor to observe and analyze the morphology characteristics of the intracranial blood vessel from any interested angle and hierarchy, can provide the three-dimensional spatial information of the intracranial blood vessel with an image, is convenient for visual observation, and is convenient for positioning and displaying focus areas. The method can obtain the integral state of the intracranial blood vessel simply, conveniently, quickly and intuitively clinically so as to analyze the intracranial vascular lesions.

Of course, it is not necessary for any one product or method of practicing the invention to achieve all of the advantages set forth above at the same time.

The present invention will be described in further detail with reference to the accompanying drawings and examples.

Drawings

In order to more clearly illustrate the embodiments of the invention or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a method for quickly establishing an intracranial vascular simulation three-dimensional model based on transfer learning, which is provided by the embodiment of the invention;

FIG. 2 is a schematic diagram of the transformation of the coordinates of an intracranial vascular magnetic resonance image according to an embodiment of the present invention;

FIG. 3 is a comparison of registration of two search strategies in accordance with an embodiment of the present invention;

FIG. 4 is a graph of the results of pre-registration of intracranial vascular magnetic resonance images in accordance with an embodiment of the present invention;

FIG. 5 is a schematic illustration of the region to be registered of an intracranial vascular magnetic resonance image in accordance with an embodiment of the present invention;

FIG. 6 (a) is a bright blood Gaussian pyramid and a black blood Gaussian pyramid of an intracranial vascular magnetic resonance image according to an embodiment of the invention; FIG. 6 (b) is a diagram of a Laplacian pyramid and a Laplacian pyramid of bright blood and black blood of an intracranial vascular magnetic resonance image, in accordance with an embodiment of the present invention;

FIG. 7 is a registration of Laplacian pyramid images of intracranial vascular magnetic resonance images in accordance with an embodiment of the present invention;

FIG. 8 is a schematic illustration of a Gaussian pyramid image registration step of an intracranial vascular magnetic resonance image based on mutual information in accordance with an embodiment of the present invention;

FIG. 9 is normalized mutual information at different iteration times according to an embodiment of the present invention;

FIG. 10 is a registration result of intracranial vascular magnetic resonance images including multiple registration methods of the mutual information pyramid method;

FIG. 11 is a graph showing the registration results of intracranial vascular magnetic resonance images of a Gaussian distribution sampling-based registration method of mutual information and an image pyramid and a mutual information pyramid method according to an embodiment of the present invention;

fig. 12 is a MIP diagram illustrating an embodiment of the present invention;

fig. 13 is an inverted diagram corresponding to the MIP map and a characteristic MIP map according to an embodiment of the present invention;

FIG. 14 is an effect diagram of an intracranial vascular simulation three-dimensional model according to an embodiment of the present invention.

Detailed Description

To obtain the intracranial blood vessel integral state simply, quickly and intuitively in clinic. The embodiment of the invention provides a method for quickly establishing an intracranial blood vessel simulation three-dimensional model based on transfer learning.

It should be noted that, the execution subject of the method for quickly establishing an intracranial vascular simulation three-dimensional model based on transfer learning provided by the embodiment of the present invention may be an apparatus for establishing an intracranial vascular simulation three-dimensional model based on transfer learning, and the apparatus may be operated in an electronic device. The electronic device may be a blood vessel imaging device or an image processing device, but is not limited thereto.

As shown in FIG. 1, the method for quickly establishing the intracranial vascular simulation three-dimensional model based on transfer learning provided by the embodiment of the invention can comprise the following steps:

s1, acquiring a bright blood image group and an enhanced black blood image group of an intracranial vascular part;

the bright blood image group and the enhanced black blood image group comprise K bright blood images and K enhanced black blood images respectively; the images in the bright blood image group and the enhanced black blood image group are in one-to-one correspondence; k is a natural number greater than 2;

the bright blood image group is an image group obtained by carrying out bright blood sequence scanning on an intracranial blood vessel part by using a magnetic resonance blood vessel imaging technology. In particular, the set of bright blood images may be a TOF-MRA sequence. Among them, TOF is one of the bright blood sequence scanning methods, and is called Time of flight (TOF). The enhanced black blood image group is an image group obtained by injecting paramagnetic contrast agent into a patient and then carrying out black blood sequence scanning on intracranial vascular parts by using a magnetic resonance vascular imaging technology. In the embodiment of the invention, the magnetic resonance vascular imaging technology is preferably HRMRA.

The black blood sequence is an effective method for measuring the wall thickness of blood vessels and identifying vascular lesions, and clinically, in order to more clearly reflect the real anatomical morphology of the blood vessel cavity, paramagnetic contrast agents are often injected into patients, and the T1 value of blood is reduced, so that the influence of the blood state on imaging results is reduced, the contrast between the blood and static tissues is enhanced, and the vascular structure obtained by using the black blood sequence scanning can be displayed more clearly. By injecting paramagnetic contrast agent, the enhanced black blood image obtained by using the black blood sequence scanning is clearer in the blood vessel wall structure compared with the black blood image obtained by directly using the black blood sequence scanning, and can clearly reflect the inflammatory reaction and the instability of arterial plaque, thereby being an effective method for measuring the thickness of the blood vessel wall and identifying the pathological characteristics of the blood vessel wall.

The K images in the bright blood image group and the enhanced black blood image group are in one-to-one correspondence, wherein the correspondence is in the same order of the images formed according to the scanning time.

S2, taking each bright blood image and the corresponding enhanced black blood image as an image pair, and preprocessing each image pair to obtain a first bright blood image and a first black blood image of the image pair;

This step can be understood as a preprocessing of the image. In an alternative embodiment, preprocessing each image pair to obtain a first bright blood image and a first black blood image of the image pair may include S21 and S22:

s21, aiming at each image pair, taking the enhanced black blood image as a reference, carrying out coordinate transformation and image interpolation on the bright blood image, using similarity measurement based on mutual information, and adopting a preset search strategy to obtain a first bright blood image after preregistration;

the enhanced black blood image is imaged by coronal plane scanning, the bright blood image is imaged by axial plane scanning, and the difference of the sequence scanning directions leads to different final magnetic resonance imaging layers, so that the magnetic resonance images of different imaging layers need to be observed under a standard reference coordinate system through coordinate transformation.

For blood vessel images, coordinate transformation of the images may be achieved using direction information in a medical image DICOM (Digital Imaging and Communications in Medicine ) file. The DICOM file is an image storage format of a medical device such as CT or nuclear magnetic resonance, and contents stored in the DICOM standard include personal data of a patient, image layer thickness, time stamp, medical device information, and the like, in addition to image information. The DICOM3.0 format image file contains orientation tag information about the imaging direction, which briefly describes the orientation relationship between the patient and the imaging apparatus, from which the exact location information of each pixel in the image can be known.

Specifically, the enhanced black blood image and the bright blood image are images to be registered, the coordinate system of the enhanced black blood image is taken as a reference, the enhanced black blood image is taken as a reference image, the bright blood image is taken as a floating image, the bright blood image is subjected to coordinate transformation, the purpose of rotating the bright blood image to the same coordinate system as the enhanced black blood image is achieved, and the scanning direction of the bright blood image becomes a coronal plane after rotation according to the azimuth label information in the DICOM file of the bright blood image.

In order to facilitate understanding of the method according to the embodiments of the present invention, the following description is briefly provided in connection with an image registration process, and the specific process may be understood by referring to the related art.

For the registration of the two images a and B, it is actually to correspond each coordinate position in the image a to the image B through a mapping relationship. Specific coordinate transformation methods may include rigid body transformation, affine transformation, projective transformation, nonlinear transformation, and the like. Since the intracranial vessel can be regarded as a rigid body, the embodiment of the invention selects the rigid body transformation as the coordinate transformation method.

However, in the coordinate transformation process, the coordinate system of the floating image is stretched or deformed, the pixel coordinates of the image subjected to coordinate transformation are not completely overlapped with the sampling grid of the original image, namely, the pixel coordinate points which are originally integers are not integers after coordinate transformation, so that some areas of the image lose part of pixels, and therefore, in the image coordinate transformation process, the image is required to be subjected to resampling interpolation at the same time to determine the gray values of the pixel coordinate points of the image subjected to coordinate transformation, and the subsequent processing is convenient. The image interpolation method comprises nearest neighbor interpolation, bilinear interpolation, bicubic interpolation and the like. Experiments are carried out on three interpolation methods, 5 evaluation indexes are set in total, namely Root Mean Square Error (RMSE), peak signal-to-noise ratio (PSNR), normalized cross correlation coefficient (NCC), normalized cross information (NMI) and Time consuming Time, wherein registration is more accurate when the RMSE is smaller, and registration is more accurate when the PSNR, NCC and NMI values are higher. From the whole experimental data, the accuracy of bicubic interpolation is obviously better than that of nearest neighbor interpolation and bilinear interpolation, so bicubic interpolation is selected.

After the missing pixel points are subjected to image restoration by using an image interpolation method, similarity of the reference image and the changed floating image is calculated by using a certain similarity measure, then the optimal similarity measure is found by using a search strategy, iteration is repeated repeatedly for optimization until the similarity measure of the two images reaches the optimal, iteration is stopped, and finally, the floating image is subjected to coordinate transformation according to a determined space transformation matrix (rotation matrix) to realize complete registration of the images. After the images to be registered are optimized through an iterative algorithm, the spatial position registration relation of the two images and the registration image can be calculated, so that the similarity between the registered floating image and the reference image is the highest.

The scale for measuring the feature similarity between two images is a similarity measure, and selecting a proper similarity measure can improve the registration accuracy and effectively inhibit noise and the like, and has very important roles in the registration of the images. The common similarity measures are mainly divided into three categories, namely distance measures, correlation measures and information entropy. In the embodiment of the invention, the intracranial blood vessel can be regarded as a rigid body, almost no deformation occurs, organs different from heart or lung and the like can be changed along with the movement of human breath and the like, so that mutual information or normalized mutual information can be selected as similarity measurement for the intracranial blood vessel, and the registration effect is more accurate.

Mutual information and normalized mutual information are one of the information entropies. Mutual information (Mutual Information, MI), which measures the correlation between two images, or the amount of information that is included with each other, is used to interpret whether the two images have reached optimal registration, with a larger value of mutual information indicating that the two images are more similar. Normalized mutual information (Normalization Mutual Information, NMI), which is an improvement of mutual information measurement, is used as similarity measurement when the pixel gray level numbers of two images to be registered are similar, and the obtained registered images are higher in accuracy and more reliable. NMI has a range of values of [0,1], the closer the value is to 1, indicating that the two images are more similar. The concept of normalization mutual information solves the problems that when the overlapped part of two images is smaller or the overlapped area is mostly background information, the image registration based on the mutual information is low in accuracy and poor in registration effect, and the sensitivity of the mutual information to the image overlapped area is reduced.

Image registration is essentially a multiparameter optimization problem in that the similarity measure of two images is ultimately optimized by spatially varying the images using some search strategy that intersects each other during the actual computation. The algorithm idea is to calculate the similarity measure between two images in each iteration, adjust the floating image through the operation of coordinate transformation such as translation or rotation, and interpolate the images until the similarity measure of the two images is maximum. The search strategies commonly used at present include a gradient descent optimizer, an (1+1) -ES based on an evolution strategy (Evolution Strategy, ES), and the like, and the predetermined search strategy in the embodiment of the present invention may be selected as needed.

Specific experimental results are shown below, referring to fig. 2, fig. 2 is a schematic diagram of coordinate transformation of an intracranial vascular magnetic resonance image in an embodiment of the present invention, where a first line is an enhanced black blood image and a bright blood image, and a second line is an enhanced black blood image and a bright blood image after coordinate transformation, and it is seen that after coordinate transformation, the bright blood image and the enhanced black blood image have consistent scanning directions and are all in a coronal plane.

The gradient descent optimizer and the (1+1) -ES are used for registering 160 bright blood images with 160 enhanced black blood images of corresponding scanning layers respectively, wherein the enhanced black blood images are reference images, the bright blood images are floating images, registration results are shown in figure 3, and figure 3 is a registration comparison result of the two search strategies in the embodiment of the invention; the left plot in fig. 3 is the two images pair-wise displayed without the optimizer registration, the middle plot is the images pair-wise displayed with the gradient descent optimizer registration, and the right plot is the images pair-wise displayed with the (1+1) -ES optimizer registration. The right image display adopts a montage effect, a pseudo-color transparent processing is used for enhancing a black blood image and a bright blood image, purple is used for enhancing the black blood image, and green is used for enhancing the bright blood image (the image of the drawing is an image after original image gray scale processing, and the color is not shown). As can be seen from the figure, in the image which is not registered by using the optimizer, the enhanced black blood image and the bright blood image are not overlapped and have more shadows; when the gradient descent optimizer is used for registering images, although the registration effect is better than that of a left image, obvious misalignment phenomenon still occurs at the grey brain matter; in the image using the (1+1) -ES optimizer, the registration result is accurate, and the misaligned shadow part in the image completely disappears. The data shown in table 1 are 3 evaluation indexes of the registration result, which are normalized mutual information NMI, normalized cross correlation coefficient NCC and algorithm Time consuming Time respectively. From the experimental result graph, the registration image effect of the (1+1) -ES is displayed more clearly and is superior to that of the gradient descent optimizer; from experimental data, all three evaluation indexes show good performance of the (1+1) -ES optimizer, so in the embodiment of the invention, the predetermined search strategy is preferably (1+1) -ES.

TABLE 1 analysis of results under different search strategies

^a The value in (2) is the mean value of the evaluation index based on the registration of 160 bright blood images and 160 enhanced black blood images ± mean square error

Referring to fig. 4, fig. 4 is a graph showing the result of preregistration of intracranial vascular magnetic resonance images according to an embodiment of the present invention. The left image is a first bright blood image after preregistration, wherein the interpolation method adopts bicubic interpolation; the middle image is an enhanced black blood image, both images are visible as coronal planes, the right image is an effect image obtained by directly superposing the images, and the right image can be used for observing a bright blood image and an enhanced black blood image under the current imaging layer under the same coronal plane although pre-registration is carried out, but the bright blood image and the enhanced black blood image still have a non-superposition phenomenon, so that the image fine registration is required to be carried out subsequently.

The pre-registration of the step can preliminarily realize the contrast of the magnetic resonance images of the same scanning level under the same coordinate system, but because the time of the scanning of the bright blood sequence and the black blood sequence is different and the patient possibly moves slightly before and after the scanning, the operation is only a rough coordinate transformation, the complete registration of the multi-mode magnetic resonance images can not be realized only by the pre-registration, but the step can omit unnecessary processing procedures for the subsequent accurate registration link, and the processing speed is improved.

S22, extracting the area content which is the same as the scanning range of the first bright blood image from the enhanced black blood image to form the first black blood image.

Because the scanning ranges of the blood vessel imaging in different magnetic resonance sequences are different, after the bright blood image is subjected to image coordinate transformation, the information of the coronal surface is not rich in the information of the enhanced black blood image, so that the same region under the two sequences can be registered more quickly and accurately, the same scanning region can be extracted from the enhanced black blood image according to the scanning region of the first bright blood image, and the registration range of the subsequent image is reduced.

Optionally, S22 may include the steps of:

1. obtaining edge contour information of a blood vessel in a first bright blood image;

specifically, the edge contour information can be obtained by using a Sobel edge detection method and other methods. The edge profile information contains coordinate values of the respective edge points.

2. Extracting minimum and maximum values of an abscissa and an ordinate in the edge profile information, and determining an initial extraction frame based on the obtained four coordinate values;

namely, in the edge profile information, the minimum abscissa value, the maximum abscissa value, the minimum ordinate value and the maximum ordinate value are extracted, and four vertexes of the square frame are determined by utilizing the four coordinate values, so that an initial extracted frame is obtained;

3. Respectively expanding the initial extraction frame by a preset number of pixels in four directions within the size range of the first bright blood image to obtain a final extraction frame;

wherein the four directions are the positive and negative directions of the horizontal and vertical coordinates respectively; the preset number is reasonably selected according to the type of the blood vessel image, so as to ensure that the expanded final extraction frame does not exceed the size range of the first bright blood image, for example, the preset number can be 20.

4. And extracting the content of the corresponding region in the final extraction frame from the enhanced black blood image to form a first black blood image.

And extracting the content of the corresponding region in the enhanced black blood image according to the coordinate range defined by the final extracted frame, and forming a first black blood image from the extracted content. The common scanning range of the magnetic resonance images under the two modes is acquired by extracting the region to be registered, so that the subsequent rapid registration is facilitated.

Referring to fig. 5, fig. 5 is a schematic diagram of a region to be registered of an intracranial vascular magnetic resonance image according to an embodiment of the present invention, wherein a left image is a first bright blood image after pre-registration, a right image is an enhanced black blood image, and a box is a region to be extracted in the enhanced black blood image. This region includes the common scan range of the bright blood sequence and the dark blood sequence in the intracranial vascular magnetic resonance image, and useful information can be focused more quickly by determining the region to be extracted.

The two-step preprocessing process of the embodiment of the invention plays an important role, the preprocessed image can pay more attention to useful information, extraneous information is eliminated, and in actual use, the image preprocessing can improve the reliability of intracranial blood vessel image registration and recognition.

S3, aiming at each first bright blood image, carrying out image registration by using a registration method of image pyramid and mutual information based on Gaussian distribution sampling by taking a corresponding first black blood image as a reference to obtain a registered bright blood image group comprising K registered bright blood images;

in an alternative embodiment, S3 may include S31 to S34:

s31, sampling and selecting part of preprocessed image pairs by Gaussian distribution as test image pairs;

the test image pair of the embodiment of the invention is an image pair to be registered, and the random selection of the image pair to be registered adopts Gaussian distribution sampling. The method is characterized in that the scanning directions of the bright blood image and the enhanced black blood image are different, and in the image preprocessing process, in order to observe the bright blood image and the enhanced black blood image on the same imaging layer, the bright blood image is subjected to coordinate transformation and interpolation, so that each bright blood image corresponds to the enhanced black blood image of the current layer; meanwhile, the data of the edge layer of the bright blood image may not be complete because the scanning range of the bright blood image is different from that of the enhanced black blood image. In summary, the bright blood data and the enhanced black blood data of the scanning middle layer are most abundant, so that the Gaussian average value mu is half of the total number of images to be registered, and the probability of the Gaussian random selection to the middle layer image registration is maximum.

S32, performing image registration on the first bright blood image and the first black blood image in each test image pair by adopting a registration method based on mutual information and an image pyramid to obtain a rotation matrix corresponding to the first bright blood image in the test image pair after registration; in an alternative embodiment, S32 may specifically include steps S321 to S324:

s321, obtaining a bright blood Gaussian pyramid from a first bright blood image and obtaining a black blood Gaussian pyramid from a first black blood image based on downsampling for each test image pair; the light blood Gaussian pyramid and the black blood Gaussian pyramid comprise m images with sequentially smaller resolutions from bottom to top; m is a natural number greater than 3;

in order to improve the accuracy of image registration, the problem that the image converges to a local maximum value in the registration process can be solved by using a multi-resolution strategy, and meanwhile, the multi-resolution strategy can improve the algorithm execution speed and the robustness under the condition that the image registration accuracy is met. Constructing an image pyramid is an effective way to increase the registration accuracy and speed by increasing the complexity of the model, that is, in the registration process, the image of low resolution is registered first in the order from coarse registration to fine registration, and then the image of high resolution is registered on the basis that the registration of the image of low resolution is completed.

In an alternative embodiment, the step S321 may include:

acquiring an input image of an ith layer, filtering the input image of the ith layer by using a Gaussian kernel, deleting even lines and even columns of the filtered image to obtain an ith layer image G of a Gaussian pyramid _i And image G of the ith layer _i As the input image of the (i+1) -th layer, an (i+1) -th layer image G of the Gaussian pyramid is obtained _i+1 ；

Wherein i=1, 2, …, m-1; when the Gaussian pyramid is a bright blood Gaussian pyramid, the input image of the 1 st layer is a first bright blood image, and when the Gaussian pyramid is a black blood Gaussian pyramid, the input image of the 1 st layer is a first black blood image.

Specifically, the plurality of images in the gaussian pyramid are images corresponding to the same original image with different resolutions. The gaussian pyramid acquires an image by gaussian filtering and downsampling, and each layer of construction steps can be divided into two steps: firstly, smoothing filtering is carried out on an image by using Gaussian filtering, namely, gaussian kernel filtering is adopted; and deleting even lines and even columns of the filtered image, namely reducing the width and height of the lower image by half to obtain the current image, so that the current image is one-fourth of the size of the lower image, and finally obtaining the Gaussian pyramid by continuously iterating the steps.

Gaussian filtering is in fact a low-pass filter, in which the image frequency range in a gaussian pyramid is very wide, with the cut-off frequency of the lower layer image being 2 times the cut-off frequency of the higher layer image. The gaussian filtering is firstly calculated by using a gaussian function to obtain a weight matrix, and then the weight matrix is used for carrying out convolution operation on the original image, and generally, a two-dimensional gaussian template can be used for carrying out the processing. Although the effect of blurring an image can be achieved by using a two-dimensional Gaussian template, when one point is at the boundary and enough points are not around, the edge image is lost due to the relation of a weight matrix, so that the two-dimensional Gaussian template is optimized. The two-dimensional Gaussian filter can be split into two independent one-dimensional Gaussian filters, and the two-dimensional Gaussian filters are used for image filtering in the transverse direction and the longitudinal direction respectively. The Gaussian functions are separated, so that edges generated by the two-dimensional Gaussian templates can be eliminated, and the running speed of a program can be greatly increased. Compared with other fuzzy filters, the Gaussian filter can achieve the fuzzy effect of the image and better reserve the marginal effect.

In the step, the first bright blood image and the first black blood image after pretreatment are subjected to the treatment, so that a bright blood Gaussian pyramid and a black blood Gaussian pyramid can be obtained. As shown in fig. 6 (a), fig. 6 (a) shows a bright blood gaussian pyramid and a black blood gaussian pyramid of an intracranial vascular magnetic resonance image according to an embodiment of the present invention.

These progressively lower resolutions, resulting from the combination of different resolutions of the same image, are arranged to resemble a pyramid, and are therefore referred to as image pyramids, with the highest resolution image at the bottom of the pyramid and the lowest resolution image at the top of the pyramid. Images with different resolutions under computer vision simulate an image observed by human eyes at different distances, and in the aspect of image information processing, compared with the traditional single-resolution image, the multi-resolution image is easier to acquire the essential characteristics of the image.

S322, obtaining a bright blood Laplacian pyramid by using a bright blood Gaussian pyramid and obtaining a black blood Laplacian pyramid by using a black blood Gaussian pyramid based on up-sampling processing; the method comprises the steps that a bright blood Laplacian pyramid and a black blood Laplacian pyramid comprise m-1 images with sequentially smaller resolutions from bottom to top;

since the gaussian pyramid is downsampled, i.e. the image is scaled down, a part of the data of the image is lost. Therefore, in order to avoid data loss of the image in the zooming process and restore detail data, the embodiment of the invention uses the Laplacian pyramid to realize image reconstruction together with the Gaussian pyramid, and the detail is highlighted on the basis of the Gaussian pyramid image.

In an alternative embodiment, the step S322 may include:

for the (i+1) -th layer image G of Gaussian pyramid _i+1 Up-sampling is carried out, and the newly added rows and columns are filled with data 0, so that a filled image is obtained;

convolving the filling image by using a Gaussian kernel to obtain an approximate value of the filling pixel, and obtaining an amplified image;

image G of the ith layer of Gaussian pyramid _i Subtracting the amplified image to obtain an ith layer image L of the Laplacian pyramid _i ；

When the Gaussian pyramid is a bright blood Gaussian pyramid, the Laplacian pyramid is a bright blood Laplacian pyramid, and when the Gaussian pyramid is a black blood Gaussian pyramid, the Laplacian pyramid is a black blood Laplacian pyramid.

Since the Laplacian pyramid is the residual error between the image and the original image after the downsampling operation, the Laplacian pyramid is compared from bottom to top, and the Laplacian pyramid is one layer of high-layer image less than the Gaussian pyramid structure.

Specifically, the mathematical formula for generating the Laplacian pyramid structure is shown in (1), wherein L _i Represents an i-th Laplacian pyramid (either a Leucasian Laplacian pyramid or a Black Laplacian pyramid), G _i Representing an i-th Gaussian pyramid (either a bright blood Gaussian pyramid or a black blood Gaussian pyramid), while UP operates to upsample the magnified image, symbolized Is a convolution symbol +.>Is the gaussian kernel used in constructing the gaussian pyramid. The formula shows that the Laplacian pyramid is formed by subtracting residual data of an image which is firstly reduced and then enlarged from an original image, is a residual prediction pyramid, and has the core idea that the Laplacian pyramid is used for storing differences between the image and the original image after the downsampling operation and retaining high-frequency information of the image so as to completely recover the image before the downsampling operation of each level. Since a part of the information lost in the previous downsampling operation cannot be completely recovered by upsampling, that is, the downsampling is irreversible, the display effect of the image after downsampling and then upsampling is blurred compared with the original image. Through residual errors between the stored image and the original image after downsampling operation, details can be added to images of different frequency layers on the basis of the Gaussian pyramid image, and the details and the like are highlighted.

Corresponding to the Gaussian pyramid with 4 layers, the step can obtain the bright blood Laplacian pyramid and the black blood Laplacian pyramid with the image layer number of 3. As shown in fig. 6 (b), fig. 6 (b) shows a bright blood laplacian pyramid and a black blood laplacian pyramid of an intracranial vascular magnetic resonance image according to an embodiment of the present invention. The image display uses gamma correction to achieve a clearer effect, with a gamma value of 0.5.

S323, registering images of corresponding layers in the bright blood Laplacian pyramid and the black blood Laplacian pyramid to obtain registered bright blood Laplacian pyramid;

in an alternative embodiment, the step S323 may include:

aiming at each layer in the Laplacian pyramid and the Laplacian pyramid, taking a Laplacian image corresponding to the layer as a reference image, taking a Laplacian image corresponding to the layer as a floating image, using similarity measurement based on mutual information, and adopting a preset search strategy to realize image registration, so as to obtain the Laplacian image of the layer after registration;

forming a registered Laplacian pyramid from bottom to top according to the sequence of sequentially decreasing resolution of the registered multilayer Laplacian images;

the black blood Laplacian image is an image in a black blood Laplacian pyramid, and the bright blood Laplacian image is an image in a bright blood Laplacian pyramid.

The registration process in the step is similar to the pre-registration process, and the registration of the images is realized by carrying out coordinate transformation and image interpolation on the bright blood Laplace images and using similarity measurement and a preset search strategy, so that the registered bright blood Laplace images can be obtained. The coordinate transformation, image interpolation, similarity measurement and predetermined search strategy are not described in detail.

As shown in fig. 7, fig. 7 is a registration result of a laplacian pyramid image of an intracranial vascular magnetic resonance image according to an embodiment of the present invention, wherein the left image is a reference image in a black laplacian pyramid, the middle image is a registered image in a bright laplacian pyramid, the right image is an effect image obtained by directly overlapping the left image and the middle image, the overlapped image is a montage effect, a pseudo-color transparent process is used to enhance the black blood image and the bright blood image, wherein purple is an enhanced black laplacian pyramid image, and green is a bright laplacian pyramid image (the image of the drawing is an image of which the original image is subjected to gray processing, and the color is not shown).

S324, using the registered Laplacian pyramid as superposition information, performing top-down registration on each layer of images in the Laplacian pyramid and the Laplacian pyramid to obtain a registered Laplacian pyramid, and obtaining a rotation matrix corresponding to the first Laplacian image in the pair of registered test images.

In the registration of the step, the registration of the images with different resolutions in the Gaussian pyramid is needed, and as the registration of the images with low resolution can be easier to hold the essential features of the images, the embodiment of the invention registers the images with high resolution on the basis of the registration of the images with low resolution, namely, the registration of the images of the Gaussian pyramid from top to bottom, and the registration result of the images of the upper layer is used as the input of the registration of the images of the lower layer.

In an optional implementation manner, in S324, using the registered laplacian pyramid of bright blood as superposition information, performing top-down registration on each layer of images in the gaussian pyramid of bright blood and the gaussian pyramid of black blood to obtain a registered gaussian pyramid of bright blood, which may include:

for the j-th layer from top to bottom in the bright blood Gaussian pyramid and the black blood Gaussian pyramid, taking a black blood Gaussian image corresponding to the layer as a reference image, taking the bright blood Gaussian image corresponding to the layer as a floating image, using similarity measurement based on mutual information, and adopting a preset search strategy to realize image registration to obtain a registered j-th bright blood Gaussian image;

performing up-sampling operation on the registered jth layer bright blood Gaussian image, adding the registered jth layer bright blood Gaussian image with the registered corresponding layer bright blood Laplacian image, and replacing the j+1th layer bright blood Gaussian image in the bright blood Gaussian pyramid by using the added image;

taking the black blood Gaussian image of the j+1th layer as a reference image, taking the replaced bright blood Gaussian image of the j+1th layer as a floating image, and realizing image registration by using a preset similarity measure and a preset search strategy to obtain the registered bright blood Gaussian image of the j+1th layer;

Where j=1, 2, …, m-1, the black blood gaussian image is an image in a black blood gaussian pyramid and the bright blood gaussian image is an image in a bright blood gaussian pyramid.

Repeating the above operation until high resolution registration of the bottom Gaussian pyramid image is completed, and obtaining a registered bright blood Gaussian pyramid. In the registration process, the pre-registration process is similar to that described previously. The coordinate transformation, image interpolation, similarity measurement and predetermined search strategy are not described in detail.

Specific steps of Gaussian pyramid image registration based on mutual information are shown in FIG. 8, and FIG. 8 is a schematic diagram of steps of Gaussian pyramid image registration based on mutual information of intracranial vascular magnetic resonance images according to an embodiment of the invention. Firstly, registering a low-resolution black blood Gaussian image of a top layer and a low-resolution bright blood Gaussian image of the top layer based on mutual information; then, carrying out up-sampling operation on the registered bright blood Gaussian image, adding the registered bright blood Gaussian image with the bright blood Laplacian image of the corresponding layer which retains high-frequency information and is registered according to the operation, and taking the bright blood Laplacian image as a bright blood Gaussian image of the next layer; and then taking the bright blood Gaussian image obtained by the operation as an input image, registering with the black blood Gaussian image of the corresponding layer, and repeating the operation until high-resolution registration of the Gaussian pyramid image of the bottom layer is completed.

In the Gaussian pyramid image registration based on mutual information, registration is required to be carried out on each layer of bright blood Gaussian image and black blood Gaussian image by taking normalized mutual information as similarity measurement, and NMI of the two images is calculated through loop iteration until NMI reaches the maximum. When the iteration number is too small, accurate registration of the images cannot be completed, but when the iteration number is too large, the calculated amount is increased sharply, fig. 9 is normalized mutual information under different iteration numbers in the embodiment of the invention, and when the registration of the first layer image, that is, the bottom layer image with the highest resolution in the gaussian pyramid reaches the maximum NMI value and the data is stable, the iteration is stopped.

The registration Gaussian pyramid with bright blood is obtained, the coordinate system of the bright blood image is consistent with the coordinate system of the enhanced black blood image, and the images have higher similarity, so that the vessel image registration process of the embodiment of the invention can be completed. And a rotation matrix corresponding to the first bright blood image in the pair of the test images after registration can be obtained after registration.

In order to verify the effectiveness and practicality of the registration method based on mutual information and an image pyramid (called a mutual information pyramid method for short), a comparison experiment is also carried out, and intracranial vascular magnetic resonance images of five patients are used, wherein the reinforced black blood image and the bright blood image of the patient A, B, C, D are 160 respectively, and the reinforced black blood image and the bright blood image of the patient E are 150 respectively; and simultaneously selecting an algorithm for registering by only using DICOM image azimuth label information and a registration algorithm based on mutual information measurement, and comparing the algorithm with the mutual information pyramid method in the embodiment of the invention, wherein the algorithm based on mutual information measurement is to search the optimal transformation between a reference image and a floating image by a multi-parameter optimization method so that the mutual information value of the two images is maximum and the image pyramid algorithm is not used.

The experimental platform was Matlab R2016b. For the image registration result of the experiment, a combination of qualitative analysis and quantitative analysis is adopted. In the aspect of qualitative analysis, because larger gray scale difference exists between the multi-mode medical images, the difference image obtained by subtracting the registration image from the reference image cannot effectively reflect the registration result of the multi-mode medical image, therefore, the embodiment of the invention obtains the color overlapped image capable of reflecting the alignment degree of the registration image and the reference image by overlapping the registration image and carries out qualitative analysis on the registration effect of the multi-mode registration algorithm through the color overlapped image, fig. 10 shows the registration result of the multi-mode intracranial vascular magnetic resonance image, and fig. 10 shows the registration result of the intracranial vascular magnetic resonance image containing multiple registration methods of the mutual information pyramid method. Wherein (a) is a reference image; (b) is a floating image; (c) is an overlaid image based on the image orientation tag information; (d) is an overlay image based on mutual information metrics; (e) Is an overlapping image of the mutual information pyramid method of the embodiment of the invention. Wherein the drawings are gray level drawings of original drawings, and the color is not shown. In the aspect of quantitative analysis, since the evaluation index root mean square error RMSE and the peak signal-to-noise ratio PSNR are not suitable for evaluating an image with larger gray scale variation, in order to better evaluate the registration result of the multi-mode medical image, the normalized cross-correlation coefficient NCC and the normalized mutual information NMI are adopted as the evaluation index, when the values of the normalized cross-correlation coefficient NCC and the normalized mutual information NMI are larger, the higher the image registration precision is represented, and the evaluation index result analysis of different registration algorithms is shown in table 2.

Table 2 analysis of results for different registration methods

^a The value in (a) is the mean value of evaluation indexes based on registration of a plurality of images of a patient ± mean square error

Qualitative analysis: as is evident from the overlapping images of fig. 10, the method based on the mutual information metric exhibits a large registration shift, and the analysis reason is likely because the method based on the mutual information metric is only used to easily fall into a local optimum, not a global optimum; the registration effect based on the image azimuth label information is also poor, and partial non-overlapping condition of the images occurs; the registration image effect of the mutual information pyramid method is good, the image display is clearer, and the images are almost completely overlapped.

Quantitative analysis: from table 2, compared with the registration algorithm using only the azimuth label information of DICOM images and the registration algorithm based on mutual information measurement, the mutual information pyramid method of the embodiment of the invention improves the registration accuracy, which indicates that the registration method based on the mutual information and the image pyramid provided by the embodiment of the invention can well process the registration of the multi-mode intracranial vascular magnetic resonance images.

S33, obtaining the average value of the rotation matrix of all the test image pairs;

In the last step, a rotation matrix corresponding to the first bright blood image after registration in each test image pair can be obtained, and then the average value of the rotation matrices of all the test image pairs can be calculated.

S34, carrying out coordinate transformation on the first bright blood image in the rest preprocessed image pairs except the test image pair by utilizing the mean value of the rotation matrix, and finishing image registration to obtain a plurality of registered image pairs.

Considering that when a patient scans by using a magnetic resonance bright blood sequence, if slight movement occurs, the coordinate position of an intracranial blood vessel image obtained by the bright blood sequence scanning slightly changes, and then each bright blood image needs to be subjected to a space coordinate transformation operation to keep the same coordinate position as that of an enhanced black blood image. In the registration process of the registration method based on the mutual information and the image pyramid, the mutual information of every two images to be registered needs to be calculated repeatedly, when the size and the number of the images are large, the time consumption is too large, and if the registration method based on the mutual information and the image pyramid is used for all the image pairs to be registered, the calculation speed is slow. The inventor considers that the intracranial blood vessel can be regarded as a rigid body, and organs such as heart or lung are different from organs such as heart or lung, and the organs are changed along with the respiratory motion of a person, so that the space coordinate transformation operation of each bright blood image is almost consistent, namely, almost the same rotation matrix is used. Referring to table 3, table 3 shows the mean and mean square error of the rotation matrix calculated by the mutual information pyramid method, and the mean square error of the rotation matrix obtained by registering 160 400×400 enhanced black blood images and 160 400×400 bright blood images of the patient a through the mutual information pyramid method is small. Then, the rotation matrix obtained by registration calculation of a few bright blood images of the patient can be used to perform the same spatial coordinate transformation on all bright blood images, without solving the rotation matrix for each bright blood image, so that the image registration process is accelerated.

TABLE 3 mean and mean squared error of rotation matrices

In the step, the first bright blood image in the rest preprocessed image pairs is subjected to coordinate transformation by using the mean value of the rotation matrix, namely a new rotation matrix, so that the registration of all the images can be rapidly completed, and the registration speed is greatly improved. The process of implementing coordinate transformation of the bright blood image by using the rotation matrix, image interpolation, using similarity measurement based on mutual information, and implementing image registration by adopting a predetermined search strategy can be referred to above, and can also be understood in combination with the related prior art, and will not be described herein.

After the step, a plurality of registered bright blood images can be obtained, and the bright blood images and the corresponding enhanced black blood images are in the same coordinate system and have higher similarity.

In order to verify the feasibility of the registration method of the mutual information and the image pyramid based on Gaussian distribution sampling, five intracranial vascular magnetic resonance images of the patient are used for registration, wherein the reinforced black blood image and the bright blood image of the patient A, B, C, D are 160 respectively, and the reinforced black blood image and the bright blood image of the patient E are 150 respectively. Aiming at the image registration results of experiments, as the acquisition principles of the multi-mode magnetic resonance images are different, the presentation information is different, no unified gold standard exists at the present stage to evaluate which registration algorithm is best, and the quality of the registration results should be evaluated from specific registration objects and application purposes, so that qualitative analysis and quantitative analysis are combined. In terms of qualitative analysis, the registration algorithm results are qualitatively analyzed by overlaying the color images that reflect the alignment of the registration image and the reference image, and fig. 11 shows a comparison of the registration results for the multimodality intracranial vascular magnetic resonance image. Fig. 11 is a graph showing the registration results of intracranial vascular magnetic resonance images by the registration method of the mutual information and image pyramid based on gaussian distribution sampling and the mutual information pyramid method according to an embodiment of the present invention. Wherein (a) is a reference image; (b) is a floating image; (c) overlapping images of the mutual information pyramid method; (d) Taking 1 for the standard deviation sigma of the overlapped images of the method; (e) Taking 2 standard deviation sigma for the overlapped image of the method; (f) Taking 3 standard deviation sigma for the overlapped image of the method; (g) Taking 4 standard deviation sigma for the overlapped image of the method; (h) Taking 5 standard deviation sigma for the overlapped image of the method; (i) For the superimposed image of the method of the invention, the standard deviation σ is taken as 6. Each image is a processed gray scale image, and the color is not shown. In the aspect of quantitative analysis, normalized cross-correlation coefficient NCC is adopted, normalized cross-information NMI is used as an evaluation index, and the higher the NCC and NMI values, the higher the image registration accuracy is. Table 4 shows the registration results (the rest of patient data is not displayed because of space limitation) of the registration method based on the mutual information and the image pyramid (the method is called a mutual information pyramid method for short) and the registration method based on the gaussian distribution sampling (the method is called a method for short) for each patient all image pair, and the experimental platform is Matlab R2016b. Because the method does not need to register all the images, only a small number of images are selected at random for registration, the Gaussian distribution mean value mu is set to be half of the total number of the images to be registered through experiments, and 20 enhanced black blood images are randomly extracted for registration with corresponding bright blood images when standard deviation sigma is 1,2,3,4,5 and 6 respectively.

Table 4 analysis of the results of different registration algorithms

The value in a is the mean value of the evaluation index based on the registration of 160 bright blood images and 160 enhanced black blood images ± mean square error

^a The value in (2) is the mean value of the evaluation index based on registration of 150 bright blood images and 150 enhanced black blood images ± mean square error

As is evident from the overlapping images in fig. 11, the registered images of the mutual information pyramid method perform well with the registered images of the method of the present invention, and the images are almost completely overlapped; from table 4, it can be seen that, from two evaluation indexes, NCC and NMI, although the accuracy of the method of the present invention is lower than that of the mutual information pyramid method, the two evaluation indexes are not much different under different gaussian distribution function settings. When the size of the images participating in registration is larger and the number of the images is larger, the calculated amount of the mutual information pyramid method is larger. The method improves the algorithm aiming at the rigid space transformation characteristics of the intracranial blood vessels, accelerates the image registration by utilizing the similarity of the transformation matrix, and experimental results prove that the time consumption of the improved algorithm registration is only one fifth of the registration time of the mutual information pyramid method, thereby greatly improving the registration speed and well realizing the registration of the multi-mode intracranial blood vessel magnetic resonance images.

In the registration scheme provided by the embodiment of the invention, the intracranial blood vessel can be regarded as a rigid body, and the space coordinate transformation operation of each bright blood image is almost consistent, so that the same rotation matrix can be used. Therefore, on one hand, a small number of image pairs are selected to perform image registration of the bright blood images, the bright blood images of the rest image pairs are subjected to the same space coordinate transformation by utilizing the mean value of the rotation matrix of the registered small number of bright blood images, and the rotation matrix is not required to be calculated for each rest bright blood image, so that the image registration process can be accelerated. On the other hand, in the image registration process, based on mutual information as similarity measurement, and an image pyramid algorithm is adopted to increase the complexity of the model, so that the registration accuracy and speed can be improved. Compared with the prior art, the observation of the bright blood image and the black blood image of the intracranial blood vessel requires a doctor to depend on space imagination and subjective experience. The embodiment of the invention can unify the bright blood image and the enhanced black blood image under the same coordinate system by adopting the image registration method, can facilitate doctors to understand the black blood sequence and intracranial blood vessel images corresponding to the bright blood sequence, simply, conveniently and quickly obtain comprehensive information required by diagnosis, and provides accurate and reliable reference information for subsequent medical diagnosis, making operation plans, radiotherapy plans and the like. The registration scheme provided by the embodiment of the invention can provide a better reference mode for the registration of other medical images and has great clinical application value.

S4, projecting the registered bright blood image group in three preset directions by using a maximum intensity projection method to obtain MIP images in all directions;

the maximum intensity projection method (maximum intensity projection, MIP) is one of the CT three-dimensional image reconstruction techniques, denoted MIP. Which passes through the volume data series along a preselected view angle using a set of projection lines, the highest CT value on each projection line being encoded to form a two-dimensional projection image. Is one method of producing a two-dimensional image by computing the maximum density of pixels encountered along each ray of the scanned object. Specifically, when the optical fiber bundle passes through an original image of a section of tissue, the pixels with the greatest density in the image are reserved and projected onto a two-dimensional plane, so as to form an MIP reconstructed image (referred to as an MIP map in the embodiment of the present invention). MIP can reflect X-ray attenuation values of corresponding pixels, small density changes can be displayed on MIP images, and stenosis, dilation and filling defects of blood vessels can be well displayed, and calcification on blood vessel walls and contrast agents in blood vessel cavities can be distinguished.

As can be appreciated by those skilled in the art, the registered bright blood image set is actually three-dimensional volume data, and the three-dimensional volume data can be projected in three preset directions by using the above MIP method to obtain a two-dimensional MIP map of each direction, where the three preset directions include: axial, coronal, and sagittal.

For the MIP method, please refer to the related description of the prior art, and the description is omitted herein, referring to fig. 12, fig. 12 is an MIP diagram illustrating an embodiment of the present invention.

S5, taking the MIP graph in each direction as a target domain, taking the fundus blood vessel graph as a source domain, and obtaining a two-dimensional blood vessel segmentation graph corresponding to the MIP graph in each direction by using a migration learning method.

The inventor finds through research that the MIP diagram of the intracranial blood vessel brightening sequence has the distribution of a blood vessel tree similar to that of a fundus blood vessel, in the MIP diagram of the intracranial blood vessel brightening sequence, a common carotid artery and a vertebral artery are the main trunk of the blood vessel tree, the intracranial blood vessel is a branch of the blood vessel tree, and a tiny blood vessel is a tiny branch of the blood vessel tree. Because the MIP graph of the intracranial blood vessel bright blood sequence has a certain similarity with the fundus blood vessel image, namely has the same characteristics, the inventor considers that a model pre-trained by the fundus blood vessel (source domain) segmentation task is migrated into the intracranial blood vessel segmentation task by means of a migration learning method, particularly by adopting a characteristic migration mode. Feature based TL is to transform the features of the source domain and the target domain into the same space by Feature transformation, so that the source domain data and the target domain data have the same distribution of data in the space, and then perform conventional machine learning, assuming that the source domain and the target domain have some common cross features.

For S5, in an alternative embodiment, S51 to S53 may be included:

s51, obtaining a target neural network pre-trained for fundus blood vessel graph segmentation tasks;

the target neural network is obtained by training in advance according to a fundus blood vessel map data set and an improved U-net network model.

As described above, the embodiment of the present invention desirably migrates the model pre-trained for the fundus blood vessel (source domain) segmentation task into the intracranial blood vessel segmentation task by means of the migration learning method of feature migration. Thus, a sophisticated network model for vessel segmentation of fundus vessel images needs to be obtained first. Specifically, the obtaining of the target neural network may be performed in the following steps:

step 1, obtaining an original network model;

in the embodiment of the invention, the structure of the existing U-net network model can be improved, and each sub-module is replaced by a residual module with a residual connection form, so that the improved U-net network model is obtained. According to the embodiment of the invention, the residual error module is introduced into the U-net network model, so that the problem that the training error is not reduced and reversely increased due to gradient disappearance caused by deepening of the layer number of the neural network can be effectively solved.

Step 2, obtaining sample data of a fundus blood vessel map;

The embodiment of the invention acquires a fundus blood vessel map data set-a DRIVE data set, wherein the data set is already marked.

And step 3, training an original network model by using sample data of the fundus blood vessel map to obtain a trained target neural network.

The method specifically comprises the following steps:

1) And taking the annotation data corresponding to each sample image in the DRIVE data set as the true value corresponding to the sample image, and training each sample image and the corresponding true value through the improved U-net network model to obtain the training result of each sample image.

2) And comparing the training result of each sample image with the true value corresponding to the sample image to obtain the output result corresponding to the sample image.

3) And calculating the loss value of the network according to the output result corresponding to each sample image.

4) And (3) adjusting parameters of the network according to the loss value, and repeating the steps 1) -3) until the loss value of the network reaches a certain convergence condition, namely the loss value reaches the minimum, wherein the training result of each sample image is consistent with the true value corresponding to the sample road image, so that the network training is completed, and the target neural network is obtained.

The following summary describes some of the parametric features of the target neural network of embodiments of the present invention:

The improved U-net network model in the embodiment of the invention has 5 levels, and forms a trapezoid network with 2.5M parameters. In each residual module, a dropoff rate of 0.25 is used (dropoff refers to that in the training process of the deep learning network, for a neural network unit, the neural network unit is temporarily discarded from the network according to a certain probability; and using batch normalization (Batch Normalization, BN), varying the variance size and mean position with optimization so that the new distribution is more tailored to the true distribution of the data, thereby guaranteeing the nonlinear expression capability of the model. The activation function adopts a LeakyRelu; the last layer of the network model is activated using Softmax.

Moreover, because of the problem of uneven foreground and background distribution of the medical image sample, the loss function uses a Dice coefficient (Dice coefficient) loss function commonly used for medical image segmentation, and specifically uses an improved Dice loss function to solve the problem of unstable training of the Dice loss function.

In the aspect of neural network optimization, an Adam optimization algorithm and default parameters are adopted, and the batch size is 256. 250 epochs were trained using a "reduced learning rate" strategy, with the learning rates on epochs 0, 20 and 150 set to 0.01, 0.001, 0.0001, respectively, and the total learning rate set to 250. And the data enhancement is performed by using a random clipping mode, and the training samples in the DRIVE data set are enlarged 20000 times.

The above briefly describes the process of obtaining the target neural network, and the trained target neural network can realize the blood vessel segmentation of the fundus blood vessel map to obtain a corresponding two-dimensional blood vessel segmentation map.

S52, respectively carrying out gray level inversion processing and contrast enhancement processing on the MIP images in all directions to obtain corresponding characteristic MIP images;

the realization of feature transfer learning requires that the source domain (fundus blood vessel image) and the target domain (intracranial blood vessel bright blood sequence MIP map) have high similarity, and the same data distribution is realized.

Therefore, in step S52, the MIP map is subjected to the gradation reversal processing and the contrast enhancement processing, and a characteristic MIP map is obtained so that the characteristic MIP map is closer to the fundus blood vessel image.

In an alternative embodiment, S52 may include S521 and S522:

s521, performing pixel transformation on the MIP map by using a gray level inversion formula to obtain an inversion map;

wherein, the gray inversion formula is T (x) =255-x, x is the pixel value in the MIP map, and T (x) is the pixel value in the inversion map;

this step can be understood as a gray inversion process colloquially, and because the pixel range of the MIP map is between 0 and 255, the original brighter region can be darkened and the original darker region can be darkened by this step, concretely, the pixel inversion can be implemented by the above gray inversion formula, and the obtained inversion map is referred to the left map in fig. 13, and the left map in fig. 13 is the inversion map corresponding to the MIP map in the embodiment of the present invention.

S522, enhancing the contrast of the inversion diagram by using a contrast-limiting self-adaptive histogram equalization method to obtain a characteristic MIP diagram.

The main purpose of this step is to enhance the contrast of the inversion chart to show a clearer vascularity. With respect to the manner in which contrast is enhanced, any of the prior art implementations may be employed, and in an alternative embodiment, this step may employ limited contrast adaptive histogram equalization (Contrast Limited Adaptive Histogram Equalization, CLAHE) to enhance contrast. For the CLAHE method, please refer to the prior art, and the description is omitted here. The obtained characteristic MIP map is shown in the right diagram in fig. 13, and the right diagram in fig. 13 is the characteristic MIP map corresponding to the MIP map in the embodiment of the present invention. It can be seen that the contrast is significantly enhanced and the vessels are more clear for the characteristic MIP map compared to the inverse map.

After S522, corresponding characteristic MIP maps can be obtained for the MIP maps in each direction.

In the embodiment of the invention, the intracranial blood vessel bright blood sequence MIP graph and the fundus blood vessel image are considered to have cross characteristics, so that the MIP image characteristic is mapped to the fundus blood vessel image by adopting a migration learning method of characteristic migration, and the intracranial blood vessel input sample and the fundus blood vessel input sample corresponding to the target neural network have the same sample distribution. Wherein, S51 and S52 may be out of order.

S53, respectively inputting the characteristic MIP graphs in all directions into a target neural network to obtain corresponding two-dimensional vessel segmentation graphs;

and respectively inputting the characteristic MIP graphs in all directions into a target neural network to obtain two-dimensional vessel segmentation graphs corresponding to all directions, wherein the two-dimensional vessel segmentation graphs are binary graphs, namely pixels only have 0 and 255, white represents vessels, and black represents a background.

S6, synthesizing the two-dimensional vascular segmentation graphs in three directions by using a back projection method to obtain first three-dimensional vascular volume data;

the principle of the back projection method is that the measured projection values are distributed to each passing point evenly according to the original projection path, the projection values in all directions are back projected, and the back projection images of all angles are accumulated to deduce the original image. By synthesizing the two-dimensional vascular segmentation maps in three directions by using a back projection method, three-dimensional volume data can be obtained, and the embodiment of the invention is called first three-dimensional vascular volume data. The back projection method in the embodiment of the present invention may be a direct back projection method, a filtered back projection method, a convolution back projection method, or the like, which is not limited herein.

In the embodiment of the present invention, the voxel value of the blood vessel part in the obtained first three-dimensional blood vessel volume data is 0 and the voxel value of the non-blood vessel part is minus infinity by the pixel control of the back projection method.

And S7, obtaining an intracranial blood vessel simulation three-dimensional model based on the first three-dimensional blood vessel volume data and the second three-dimensional blood vessel volume data corresponding to the registered bright blood image group.

In an alternative embodiment, S7 may include S71 and S72:

s71, adding the first three-dimensional blood vessel volume data and the second three-dimensional blood vessel volume data to obtain third three-dimensional blood vessel volume data;

the three-dimensional blood vessel volume data can be obtained by directly adding the voxel values in the first three-dimensional blood vessel volume data and the second three-dimensional blood vessel volume data, and by the step, the cerebral spinal fluid and fat signals with the intracranial blood vessel signal intensity being the same can be eliminated.

S72, processing the third three-dimensional blood vessel volume data by using a threshold segmentation method to obtain the intracranial blood vessel simulation three-dimensional model.

The threshold segmentation method is an area-based image segmentation technique, and the principle is to divide image pixel points into several classes. The purpose of thresholding the image is to divide the pixel sets into a subset of regions corresponding to the real scene according to gray levels, each region having a consistent attribute within it, and adjacent regions not having such consistent attribute. The threshold value is defined as a limit, and the threshold segmentation is a method of processing an image into a high-contrast, easily identifiable image with an appropriate pixel value as a limit.

The threshold segmentation method adopted by the embodiment of the invention comprises a maximum inter-class variance method, a maximum entropy, an iteration method, an adaptive threshold value, a manual method, an iteration method, a basic global threshold value method and the like. In an alternative implementation, the embodiment of the present invention may employ a maximum inter-class variance method.

The maximum inter-class variance method (or OTSU for short) is a method for automatically obtaining a threshold value from a condition suitable for double peaks, and uses the idea of clustering, wherein the gray level of an image is divided into two parts according to gray levels, so that the gray value difference between the two parts is maximum, the gray level difference between each part is minimum, and a proper gray level is found for division through calculation of variance. Therefore, the threshold value can be automatically selected for binarization by adopting an OTSU algorithm during binarization. The method is characterized in that the image is divided into a background part and a target part according to the gray characteristic of the image. The larger the inter-class variance between the background and the object, the larger the difference between the two parts that make up the image. The OTSU algorithm is considered as the optimal algorithm for selecting the threshold value in image segmentation, and is simple to calculate and is not influenced by the brightness and contrast of the image. Thus, a segmentation that maximizes the inter-class variance means that the probability of misclassification is minimal. S72 using the OTSU may include the steps of:

Firstly, calculating a first threshold value corresponding to centered fourth three-dimensional blood vessel volume data in third three-dimensional blood vessel volume data by using OTSU;

in this step, one threshold value corresponding to a plurality of images in a small cube (referred to as fourth three-dimensional blood vessel volume data) located near the middle portion among the large three-dimensional cubes of the third three-dimensional blood vessel volume data is obtained as a first threshold value by using the OTSU method. Because blood information is basically concentrated in the middle of the image in the third three-dimensional blood vessel volume data, a first threshold value is determined by selecting small centered cube data (fourth three-dimensional blood vessel volume data) in the third three-dimensional blood vessel volume data, so that the threshold value calculation amount can be reduced, the calculation speed can be improved, and the first threshold value can be accurately applied to all blood information in the third three-dimensional blood vessel volume data.

For the size of the fourth three-dimensional blood vessel volume data, the center point of the third three-dimensional blood vessel volume data can be determined first, and then the third three-dimensional blood vessel volume data extends in six directions corresponding to the cube with preset side lengths, so that the size of the fourth three-dimensional blood vessel volume data is determined; the preset side length may be determined according to an empirical value including a Willis loop, for example, 1/4 of the side length of the cube, which is the third three-dimensional blood vessel volume data, etc. The Willis loop is the most important side branch circulation path in the cranium, and connects the two hemispheres with the anterior and posterior circulation.

And then, utilizing the first threshold value to realize threshold segmentation of the third three-dimensional blood vessel body data, and obtaining the intracranial blood vessel simulation three-dimensional model.

It will be appreciated by those skilled in the art that by thresholding, the gray level of the point on the image corresponding to the third three-dimensional vessel volume data can be set to 0 or 255, i.e., the entire image is rendered as a distinct black and white effect, the blood information is highlighted as white, and the extraneous information is displayed as black. For the processing procedure of threshold segmentation, please refer to the prior art, and will not be described herein. Referring to fig. 14, fig. 14 is an effect diagram of the intracranial vascular simulation three-dimensional model according to an embodiment of the present invention. The figure is gray-scale processed, the color not being shown, and in practice, the blood vessel region may be displayed in a color such as red.

In the scheme provided by the embodiment of the invention, firstly, the image registration is carried out on the bright blood image and the enhanced black blood image obtained by the magnetic resonance blood vessel imaging technology by adopting a registration method of mutual information based on Gaussian distribution sampling and an image pyramid, so that the registration efficiency can be improved, and the registration precision of the images can be improved layer by layer from low resolution to high resolution. Through the image registration, the bright blood image and the enhanced black blood image can be unified under the same coordinate system, so that subsequent unified observation is facilitated. In the field, the registered bright blood images are two-dimensional images, and although the registered bright blood images and the corresponding enhanced black blood images are in the same coordinate system, the difficulty of observation of doctors is reduced, but the two-dimensional images have limitations, the doctors need to combine a plurality of two-dimensional images to imagine the specific form of the blood vessel, and the overall state of the intracranial blood vessel cannot be obtained simply, conveniently, quickly and intuitively in clinic. According to the embodiment of the invention, the MIP image in each direction is obtained by utilizing the maximum intensity projection method on the registered bright blood image, and the characteristic that the bright blood sequence MIP image of the intracranial blood is similar to the fundus blood image is utilized, on one hand, a network model for fundus blood segmentation is trained by utilizing a labeling sample of the fundus blood image, on the other hand, characteristic MIP images with the same sample distribution as the fundus blood image are obtained by carrying out characteristic transformation on the bright blood sequence MIP image of the intracranial blood, and the characteristic MIP image with the same sample distribution as the fundus blood image is obtained, and the pre-trained network model of the fundus blood segmentation task is migrated to the intracranial blood segmentation task in a characteristic migration mode, so that a two-dimensional blood segmentation image in each direction corresponding to the bright blood sequence MIP image of the intracranial blood is obtained. According to the embodiment of the invention, the research thought of migration learning is applied to the field of intracranial vessel segmentation, and a relatively accurate vessel segmentation effect can be obtained. And then, obtaining first three-dimensional blood vessel volume data by using a back projection method, and realizing an intracranial blood vessel simulation three-dimensional model by using second three-dimensional blood vessel volume data corresponding to the registered bright blood image group. The intracranial blood vessel simulation three-dimensional model can simulate the morphology of the intracranial blood vessel, realizes the three-dimensional visualization of the intracranial blood vessel, does not need a doctor to restore the vascular tissue structure, the disease characteristics and the like through imagination, can facilitate the doctor to observe and analyze the morphology characteristics of the intracranial blood vessel from any interested angle and hierarchy, can provide the three-dimensional spatial information of the intracranial blood vessel with an image, is convenient for visual observation, and is convenient for positioning and displaying focus areas. The method can obtain the integral state of the intracranial blood vessel simply, conveniently, quickly and intuitively clinically so as to analyze the intracranial vascular lesions.

In an alternative embodiment, after S7, the method may further include:

displaying the intracranial vascular simulation three-dimensional model, in particular, displaying the intracranial vascular simulation three-dimensional model on a display screen of a computer and other equipment so as to facilitate observation of doctors; and, it is reasonable that both the bright blood image and the enhanced black blood image can be displayed simultaneously.

Note that: the experimental data of the patients in the embodiment of the invention are all from the Shaanxi province people hospital, and the images can be used as general scientific researches.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises an element.

In this specification, each embodiment is described in a related manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, as relevant to see a section of the description of method embodiments.

The foregoing is merely illustrative of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for quickly establishing an intracranial vascular simulation three-dimensional model based on transfer learning is characterized by comprising the following steps:

Aiming at each first bright blood image, carrying out image registration by using a registration method of image pyramid and mutual information based on Gaussian distribution sampling by taking a corresponding first black blood image as a reference to obtain a registered bright blood image group comprising K registered bright blood images; comprising the following steps: selecting the preprocessed partial image pair as a test image pair by Gaussian distribution sampling; performing image registration on the first bright blood image and the first black blood image in each test image pair by adopting a registration method based on mutual information and an image pyramid to obtain a rotation matrix corresponding to the first bright blood image in the test image pair after registration; obtaining the average value of the rotation matrix of all the test image pairs; performing coordinate transformation on the first bright blood images in the rest preprocessed image pairs except for the test image pair by using the mean value of the rotation matrix to finish image registration, and obtaining a registered bright blood image group comprising K registered bright blood images; the image registration of the first bright blood image and the first black blood image in each test image pair is performed by adopting a registration method based on mutual information and an image pyramid, so as to obtain a rotation matrix corresponding to the first bright blood image in the test image pair after registration, and the method comprises the following steps: for each test image pair, based on downsampling, obtaining a bright blood Gaussian pyramid from the first bright blood image and obtaining a black blood Gaussian pyramid from the first black blood image; the bright blood Gaussian pyramid and the black blood Gaussian pyramid comprise m images with sequentially smaller resolutions from bottom to top; m is a natural number greater than 3; based on up-sampling processing, obtaining a bright blood Laplacian pyramid by using the bright blood Gaussian pyramid, and obtaining a black blood Laplacian pyramid by using the black blood Gaussian pyramid; the bright blood Laplacian pyramid and the black blood Laplacian pyramid comprise m-1 images with sequentially smaller resolutions from bottom to top; registering images of corresponding layers in the bright blood Laplacian pyramid and the black blood Laplacian pyramid to obtain registered bright blood Laplacian pyramid; using the registered Laplacian pyramid as superposition information, performing top-down registration on each layer of images in the Laplacian pyramid and the Gaussian black pyramid to obtain a registered Laplacian pyramid, and obtaining a rotation matrix corresponding to the first Laplacian image in the pair of registered test images;

taking MIP images in all directions as a target domain, taking fundus blood vessel images as a source domain, and obtaining two-dimensional blood vessel segmentation images corresponding to the MIP images in all directions by using a migration learning method; comprising the following steps: obtaining a target neural network pre-trained for fundus blood vessel graph segmentation tasks; the target neural network is obtained by training in advance according to a fundus blood vessel map data set and an improved U-net network model; respectively carrying out gray level inversion processing and contrast enhancement processing on MIP images in all directions to obtain corresponding characteristic MIP images; wherein the characteristic MIP map and the fundus blood vessel map have the same sample distribution; respectively inputting the characteristic MIP maps in all directions into the target neural network to obtain corresponding two-dimensional vessel segmentation maps;

2. The method of claim 1, wherein preprocessing each image pair to obtain a first bright blood image and a first dark blood image of the image pair comprises:

for each image pair, carrying out coordinate transformation and image interpolation on the bright blood image by taking the enhanced black blood image as a reference, using similarity measurement based on mutual information, and adopting a preset searching strategy to obtain a first bright blood image after preregistration;

and extracting the area content which is the same as the scanning range of the first bright blood image from the enhanced black blood image to form a first black blood image.

3. The method of claim 2, wherein the obtaining a bright blood gaussian pyramid from the first bright blood image and a black blood gaussian pyramid from the first black blood image based on the downsampling process comprises:

4. A method according to claim 3, wherein the obtaining a bright blood laplacian pyramid using the bright blood gaussian pyramid and obtaining a black blood laplacian pyramid using the black blood gaussian pyramid based on the upsampling process comprises:

5. The method of claim 4, wherein using the registered laplacian bright blood pyramid as superposition information, performing top-down registration on each layer of images in the gaussian bright blood pyramid and the gaussian black blood pyramid to obtain a registered gaussian bright blood pyramid, comprises:

for the j-th layer from top to bottom in the bright blood Gaussian pyramid and the black blood Gaussian pyramid, taking a black blood Gaussian image corresponding to the layer as a reference image, taking a bright blood Gaussian image corresponding to the layer as a floating image, using similarity measurement based on mutual information, and adopting a preset search strategy to realize image registration to obtain a registered j-th layer bright blood Gaussian image;

performing up-sampling operation on the j-th layer of bright blood Gaussian image after registration, adding the j-th layer of bright blood Gaussian image with the corresponding layer of bright blood Laplacian image after registration, and replacing the j+1-th layer of bright blood Gaussian image in the bright blood Gaussian pyramid by using the added image;

Where j=1, 2, …, m-1, the black blood gaussian image is an image in the black blood gaussian pyramid, and the bright blood gaussian image is an image in the bright blood gaussian pyramid.

6. The method according to claim 5, wherein the performing gray inversion processing and contrast enhancement processing on the MIP map in each direction to obtain a corresponding characteristic MIP map includes:

performing pixel transformation on the MIP map by using a gray level inversion formula to obtain an inversion map; wherein, the gray inversion formula is T (x) =255-x, x is a pixel value in the MIP map, and T (x) is a pixel value in the inversion map;

and enhancing the contrast of the inversion diagram by using a contrast-limiting self-adaptive histogram equalization method to obtain a characteristic MIP diagram.

7. The method according to claim 1 or 6, wherein obtaining the intracranial vessel simulation three-dimensional model based on the first three-dimensional vessel volume data and the second three-dimensional vessel volume data corresponding to the registered bright blood image set comprises:

adding the first three-dimensional blood vessel volume data and the second three-dimensional blood vessel volume data to obtain third three-dimensional blood vessel volume data;

and processing the third three-dimensional blood vessel body data by using a threshold segmentation method to obtain an intracranial blood vessel simulation three-dimensional model.