Fast brain region contour recommendation method based on dense sift features

Abstract

The present application relates to a kind of brain region profile fast recommendation method based on DenseSIFT feature, comprising: obtaining 2D sequence biological brain image, image texture is enhanced;Image is marked with brain region contour line, and contour point is sampled by contour line;Calculate the coordinates of multiple pixel points in the normal direction of brain region contour point;Calculate DenseSIFT feature;Calculate Euclidean distance, obtain initial profile recommended point set S;Initial profile recommended point screening is carried out, and the corrected profile recommended point is obtained;The recommended contour line is obtained by fitting the corrected profile recommended point through Bezier curve.It is compared with the popular machine learning method, and the present application can recommend contour without training data set, so as to solve the problem of lacking artificial annotation data set for new data or new brain region, the method is very strong in universality, greatly reduces the problem of uneven recommended contour and slow speed due to point concentration, high accuracy, short time consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of 2D image processing, and particularly relates to a brain region contour rapid recommendation method based on DenseSIFT features. BACKGROUND

[0002] To understand how the brain works, researchers need to identify various cells and determine their distribution in various regions of the brain, as well as the molecular expression pattern of each cell type, which requires researchers to establish a cell type atlas. At present, the international neuroscience research community mostly uses rodents (mice and rats) as animal models to study the neural mechanisms of brain function under physiological and pathological conditions. However, it is now found that non-human primates may be more suitable experimental animal models for understanding human higher cognitive functions (such as thinking and consciousness). China has rich resources of macaques and is rapidly developing in the research of using macaques to establish models of human diseases, which makes China possibly make a unique contribution in the research of high-level cognitive functions such as empathy, consciousness and language, and the pathological mechanisms and intervention means of brain diseases.

[0003] The size of the monkey brain magnetic resonance (MRI) atlas data is 400x530x340 pixels. In 2021, a domestic scientific research team realized micron-level resolution three-dimensional analysis of the macaque brain through self-developed high-throughput three-dimensional fluorescence imaging VISoR technology and primate brain mapping SMART process. The resolution of the existing VISoR monkey brain atlas reaches 1x1x2.5 microns, and the entire raw data volume exceeds 1PB. The original size is 100 times that of the MRI data. Even after downsampling, the size of the VISoR data is more than 1,000. In the case of unlabeled data, researchers will spend months or even a year to draw the contour of each brain region on each image, which is a very large workload. In addition, for many biological brain images, the extraction method is through slice imaging. Because the original data volume is large, and many brain images do not have corresponding contour markers, only professional people can manually draw the brain region contour on the 2D slice according to experience and observation. Moreover, because the amount of data with markers is small, the segmentation or tracking method of deep learning is difficult to achieve high accuracy. For some regions without markers, it is also impossible to automatically identify or generate new brain region contours. Manual marking of contours faces the problems of slow speed and low efficiency. The Allen Institute for Brain Science in the United States spent 3 years to complete the manual annotation of the mouse whole brain reference atlas. Now the size of the macaque brain image data imaged by the VISoR technology is several thousand times that of the mouse brain. Even if the brain region drawing work is carried out on the downsampled image, it will take months or even a year. Therefore, a method for assisting manual drawing is necessary.

[0004] Therefore, how to quickly draw the brain region contour under the premise of ensuring the accuracy of the result is a problem that needs to be solved at present. SUMMARY

[0005] In order to solve the problem that a large amount of time and energy is needed for manual brain region contour drawing in the initial stage of the experiment without label data, the purpose of the present application is to provide a brain region contour fast recommendation method based on DenseSIFT features, which can recommend the brain region contour of subsequent images based on the brain region contour of an image drawn by researchers, so that the researchers only need to fine-tune and confirm the recommended brain region to complete the drawing of the brain region, and the working efficiency of the brain region drawing can be greatly improved.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme: a brain region contour fast recommendation method based on DenseSIFT features, which comprises the following steps in sequence:

[0007] (1) obtaining 2D sequence biological brain images and performing image texture enhancement: sorting the 3D images of the biological brain according to the x / y / z axes to obtain 2D sequence images, i.e., original images, calculating Gaussian blur for the original images to obtain Gaussian blur images, subtracting the Gaussian blur images from the original images and then adding the original images to obtain sequence images I after image texture enhancement;

[0008] (2) marking the brain region contour line of the image and sampling the contour points through the contour line: selecting a texture-enhanced image I t with a target brain region, drawing the brain region contour line through manual marking, and then sampling all points on the brain region contour line to obtain a brain region contour point set P;

[0009] (3) calculating the normal slope of each brain region contour point through the coordinates of adjacent brain region contour points, and calculating the coordinates of a plurality of pixel points in the normal direction of the brain region contour point;

[0010] (4) calculating the DenseSIFT features of the texture-enhanced image I t and the next image I t+1 adjacent thereto through the DenseSIFT feature algorithm;

[0011] (5) extracting the feature vectors of the DenseSIFT features corresponding to the pixel point set P coordinates on the texture-enhanced image I t , and calculating the distance between the feature vectors of the DenseSIFT features of a plurality of adjacent pixel point coordinates in the normal direction of the point set P coordinates on the image I t+1 to obtain an initial contour recommendation point set S of the image I t+1 ;

[0012] (6) The initial contour recommendation point set S is filtered by calculating the offset value, and the constraint calculation of the initial contour recommendation point is offset corrected to obtain the corrected contour recommendation point;

[0013] (7) The corrected contour recommendation point is fitted by a Bezier curve to obtain the recommended contour line of the image I t+1 , and the recommended contour line is resampled as a contour point as the input of step (3), and the images I t+1 , I t+2 are input as step (4), and the operations of steps (3) to (7) are repeated to realize the brain region contour drawing of each image in the sequence image I.

[0014] The step (2) specifically comprises the following steps:

[0015] (2a) The image I t containing the target brain region on the texture-enhanced sequence image I is drawn by manual annotation to obtain one or more closed curves;

[0016] (2b) The closed curve is thinned, and each point on the closed curve is saved clockwise, and the points of different closed curves are marked;

[0017] (2c) The saved points are equidistantly sampled to obtain the brain region contour point set P.

[0018] The step (3) specifically comprises the following steps:

[0019] (3a) Assuming that the known brain region contour point P n is a point in the brain region contour point set P, the coordinates of the known brain region contour point P n are (X n , Y n ), and the previous brain region contour point P n-1 (X n-1 , Y n-1 ) and the next brain region contour point P n+1 (X n+1 , Y n+1 ), the brain region contour point P n is first subtracted from the previous brain region contour point P n-1 to obtain the vector P n-1 P n (X n -X n-1 , Y n -Y n-1 ), which is unitized to obtain its unit perpendicular segment L1, and the vector representation of the unit perpendicular segment L1 is:

[0020] (-(Y n -Yn-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 ,(X n -X n-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 )

[0021] Similarly, the vector P is calculated. n P n+1 The unit perpendicular segment L2 is used to obtain the brain region contour point P by summing the unit perpendicular segments L1 and L2. n The normal vector is finally used as the brain region contour point P. n Dividing the y-component of the normal vector by the x-component yields the brain region contour point P. n The slope of the normal line k n ;

[0022] (3b) Through brain region contour point P n and its normal slope k n The absolute value of the point is used to construct the normal equation, and the coordinates of the multiple pixels that the normal equation passes through in the first quadrant are obtained. If the slope of the normal is k... n If the value is negative, the obtained pixel coordinates will be reversed along the Y-axis;

[0023] (3c) The obtained multiple pixel coordinates are arranged along the brain region contour point P. n By performing central symmetry, the brain region contour point P is obtained. n Multiple pixel coordinates along the normal direction;

[0024] (3d) Record brain region contour points P n Multiple pixel coordinates along the normal direction and the pixel coordinates Q of these pixel coordinates. [n,1] Q [n,2] ...Q [n,m] .

[0025] Step (4) specifically includes the following steps;

[0026] (4a) The DenseSIFT feature algorithm refers to: a local window is intercepted for each pixel point, and then the local window is divided into 4x4 units, the gradient direction of each unit is divided into 8 intervals, the gradient of each pixel is calculated, and the gradient histogram of the pixel points in each interval in eight directions is counted, thereby generating a 4x4x8=128-dimensional feature vector; the eight directions refer to: right, right upper 45°, upper, left upper 45°, left, left lower 45°, lower and right lower 45°;

[0027] (4b) Set the local window radius, and expand the edge of the texture-enhanced image according to the local window radius, the expansion size is consistent with the local window radius, and the pixel value of the expansion is 0;

[0028] (4c) The expanded image is equally divided into small pixel blocks, and the length and width of the pixel block are both twice the local window radius, and the local window is moved according to the step length, wherein the step length is set to 1, that is, each pixel is calculated;

[0029] (4d) Each pixel block is divided into 4x4=16 blocks, and the gradient histogram of each pixel block in eight directions is calculated;

[0030] (4e) After calculating the gradient of the center pixel of each small pixel block, the DenseSIFT feature of the entire image is obtained through Gaussian convolution.

[0031] The step (5) specifically comprises the following steps:

[0032] (5a) Extract the image I t of the drawn contour line n under the coordinates of the feature vector α n ;

[0033] (5b) Extract the brain region contour point P t under the coordinates of the feature vector β t+1 of the adjacent next image I n of the image I [n,1] , Q [n,2] ... Q [n,m] ; [n,1] [n,2] [n,m] ;

[0034] (5c) The Euclidean distances of the feature vectors α n and the feature vectors β [n,1] , β [n,2] ... β [n,m] are calculated in turn, the point with the smallest Euclidean distance is obtained, and the point is determined as the initial contour recommendation point set S corresponding to the image I t+1 . ​​

[0035] The step (6) specifically comprises the following steps:

[0036] (6a) calculating the displacement direction between each point of the initial contour recommended point set S of the image I t+1 and the brain region contour point set P of the image I t , specifically, making difference between the point P n and the point S n to obtain the difference value in the x and y directions, if the difference value is positive, it is recorded as 1, if the difference value is zero, it is recorded as 0, and if the difference value is negative, it is recorded as -1;

[0037] (6b) judging whether the displacement direction of the adjacent initial contour recommended point is reasonable according to the displacement direction of the adjacent initial contour recommended point, the judgment basis is that if the x and y displacement directions of the adjacent initial contour recommended points S n-1 and S n+1 are consistent, that is, 1 or 0 or -1, and one of the x and y displacement directions of the initial contour recommended point S n is consistent with S n-1 and S n+1 , it is judged to be reasonable, otherwise it is not reasonable; if only one of the x and y displacement directions of the adjacent initial contour recommended points S n-1 and S n+1 is consistent, and the displacement direction of the initial contour recommended point S n is consistent with the x and y displacement directions of the initial contour recommended points S n-1 and S n+1 , it is judged to be reasonable, otherwise it is not reasonable, and the unreasonable points are recorded;

[0038] (6c) calculating the deformation field matrix between the brain region contour point set P and the initial contour recommended point set S through the regularized thin plate spline algorithm, and performing product operation on the deformation field matrix and the unreasonable points to obtain new coordinates, replacing the coordinates of the unreasonable points, making them more smooth with other points of the initial contour recommended point set S, and obtaining the corrected contour recommended point.

[0039] From the above technical solutions, the beneficial effects of the present application are: first, the present application is based on the traditional DenseSIFT feature, and the gradient feature can be used for feature operation on each pixel point in the image, so that the contour can be recommended without data set, thus solving the problem of lack of artificial annotation data set for new data or new brain region, and the method has strong universality; second, the points are recommended from the contour line, and the contour line is automatically drawn after the end, which greatly reduces the problems of uneven recommended contour and slow speed caused by dense points; third, the contour recommendation method proposed by the present application only needs to be drawn on one image, and can recommend subsequent images, has high accuracy and short time consumption. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 For the method flowchart of the present application;

[0041] Figure 2 For the monkey brain VISoR image after downsampling;

[0042] Figure 3 For the monkey brain VISoR image after texture enhancement;

[0043] Figure 4 For the partial brain region contour line image;

[0044] Figure 5 For the partial brain region contour point image;

[0045] Figure 6 For the adjacent brain image, which is also the target image to be recommended;

[0046] Figure 7 For the recommended result contour point image;

[0047] Figure 8 For the contour line image drawn by the recommended contour point; DETAILED DESCRIPTION

[0048] As shown in Figure 1 , a brain region contour fast recommendation method based on DenseSIFT features, the method comprising the following steps in sequence:

[0049] (1) Obtain 2D sequence biological brain images, and perform image texture enhancement: sort the 3D images of the biological brain according to the x / y / z axes to obtain 2D sequence images, i.e., original images, calculate the Gaussian blur of the original images to obtain Gaussian blur images, subtract the Gaussian blur images from the original images and then add the original images to obtain the sequence images I after image texture enhancement;

[0050] (2) Mark the brain region contour line of the image, and sample the contour point through the contour line: select a texture-enhanced image I t with a target brain region, t is the t-th in the sequence image I, draw the brain region contour line through manual annotation, and then sample all points on the brain region contour line to obtain the brain region contour point set P;

[0051] (3) Calculate the normal slope of each brain region contour point through the coordinates of the adjacent brain region contour points, and calculate the coordinates of multiple pixel points in the normal direction of the brain region contour point;

[0052] (4) Calculate the DenseSIFT features of the texture-enhanced image I t and its adjacent next image I t+1 through the DenseSIFT feature algorithm;

[0053] (5) extracting the texture enhanced image I t the feature vectors of the DenseSIFT features of the pixel points in the set P corresponding to the image I t+1 the feature vectors of the DenseSIFT features of the pixel points in the set P corresponding to the image I t+1 , and the initial contour recommended point set S is obtained.

[0054] (6) filtering the initial contour recommended points in the initial contour recommended point set S by calculating the offset value, and correcting the constraint calculation of the initial contour recommended points by offset, to obtain the corrected contour recommended points.

[0055] (7) fitting the corrected contour recommended points by a Bezier curve to obtain the recommended contour line of the image I t+1 , and resampling the recommended contour line into contour points as the input of step (3), and taking the images I t+1 , I t+2 as the input of step (4), repeating the operations of steps (3) to (7), and realizing the brain region contour drawing of each image in the sequence image I.

[0056] The step (2) specifically comprises the following steps:

[0057] (2a) drawing one or more closed curves on the image I t in the sequence image I n by manual annotation, and obtaining the closed curves.

[0058] (2b) thinning the closed curves, saving each point on the closed curves clockwise, and marking the points of different closed curves.

[0059] (2c) isometrically sampling the saved points to obtain the brain region contour point set P.

[0060] The step (3) specifically comprises the following steps.

[0061] (3a) assuming that the known brain region contour point P n is a point in the brain region contour point set P, and the coordinates of the known brain region contour point P n are (X n , Y n ), and the previous brain region contour point P n-1 (X n-1 , Y n-1 ) and the next brain region contour point P n+1 (X n+1 , Y n+1 ), the brain region contour point P n is first subtracted from the previous brain region contour point Pn-1 vector P n-1 P n (X n -X n-1 ,Y n -Y n-1 ), and after unitizing it, a unit perpendicular segment L1 is obtained, and the vector representation of the unit perpendicular segment L1 is:

[0062] (-(Y n -Y n-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 ,(X n -X n-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 )

[0063] Similarly, the unit perpendicular segment L2 of the vector P n P n+1 is calculated, and then the sum of the unit perpendicular segments L1 and L2 is obtained to get the normal vector of the brain region contour point P n , and finally the y component of the normal vector of the brain region contour point P n is divided by the x component to get the normal slope k n of the brain region contour point P n ;

[0064] (3b) A point equation is made by the brain region contour point P n and the absolute value of the normal slope k n , and the pixel coordinate points crossed by the point equation in the first quadrant are obtained, and if the normal slope k n is negative, the obtained pixel coordinate points are inverted along the Y axis;

[0065] (3c) The obtained multiple pixel coordinate points are centrally symmetric along the brain region contour point P n , and multiple pixel coordinate points in the normal direction of the brain region contour point P n are obtained;

[0066] (3d) The multiple pixel coordinate points in the normal direction of the brain region contour point P n and the pixel coordinate points Q [n,1] , Q [n,2] ...Q[n,m] .

[0067] The step (4) specifically comprises the following steps:

[0068] (4a) The DenseSIFT feature algorithm refers to: a local window is intercepted for each pixel point, and then the local window is divided into 4x4 units, the gradient direction of each unit is divided into 8 intervals, the gradient of each pixel is calculated, and the gradient histogram of the pixel points in each interval in eight directions is counted, thereby generating a 4x4x8=128-dimensional feature vector; the eight directions refer to: right, right upper 45°, upper, left upper 45°, left, left lower 45°, lower, and right lower 45°;

[0069] (4b) Set the local window radius, and expand the edge of the texture-enhanced image according to the local window radius, the expansion size is consistent with the local window radius, and the pixel value of the expansion is 0;

[0070] (4c) The expanded image is equally divided into small pixel blocks, and the length and width of the pixel block are both twice the local window radius, and the local window is moved according to the step length, which is set to 1, that is, each pixel is calculated;

[0071] (4d) Each pixel block is further divided into 4x4=16 blocks, and the gradient histogram of each pixel block in eight directions is calculated;

[0072] (4e) After calculating the gradient of the center pixel of each small pixel block, the DenseSIFT feature of the entire image is obtained through Gaussian convolution.

[0073] The step (5) specifically comprises the following steps:

[0074] (5a) Extract the image I t of the contour line drawn n under the coordinates of the feature vector α n ;

[0075] (5b) Extract the brain region contour point P t of the next image I t+1 adjacent to the image I n under the coordinates of the normal direction Q [n,1] , Q [n,2] ... Q [n,m] , the feature vector β [n,1] , β [n,2] ... β [n,m] ;

[0076] (5c) Calculate the feature vector α n and the feature vector β [n,1] , β [n,2] in turn... beta [n,m] , get the point with the minimum Euclidean distance, and determine the image I t+1 corresponding to the initial contour recommended point set S.

[0077] The step (6) specifically comprises the following steps:

[0078] (6a) calculating the displacement direction between each point of the initial contour recommended point set S of the image I t+1 and the brain region contour point set P of the image I t , specifically, making difference between the point P n and the point S n to obtain the difference value in the x and y directions, if the difference value is positive, it is recorded as 1; if the difference value is zero, it is recorded as 0; if the difference value is negative, it is recorded as -1;

[0079] (6b) judging whether the displacement direction of the point is reasonable according to the displacement direction of the adjacent initial contour recommended point, the judgment basis is: if the x and y displacement directions of the adjacent initial contour recommended points S n-1 and S n+1 are consistent, i.e. all are 1 or 0 or -1, and one of the x and y displacement directions of the initial contour recommended point S n is consistent with S n-1 and S n+1 , it is judged as reasonable, otherwise it is judged as unreasonable; if only one of the x and y displacement directions of the adjacent initial contour recommended points S n-1 and S n+1 is consistent, and the displacement direction of the initial contour recommended point S n is consistent with the x and y displacement directions of the initial contour recommended points S n-1 and S n+1 , it is judged as reasonable, otherwise it is judged as unreasonable, and the unreasonable points are recorded.

[0080] (6c) calculating the deformation field matrix between the brain region contour point set P and the initial contour recommended point set S through the regularized thin plate spline algorithm, and performing product operation between the deformation field matrix and the unreasonable points to obtain new coordinates, replacing the coordinates of the unreasonable points, making them more smooth with other points of the initial contour recommended point set S, and obtaining the corrected contour recommended points.

[0081] The application will be further described below with reference to Figures 1 to 8 .

[0082] To understand the brain, it is necessary to determine the distribution of various cells in the brain, especially neurons, and to create a cell type map. Some neurons may span multiple regions, and to facilitate researchers in tracking and locating cells and to make it easier to draw the map, it is necessary to divide the brain into regions according to some distinctive features of the brain image or the unique cell groups in the region, which requires researchers to draw the brain region contours.

[0083] Current methods of drawing brain region contours and problems: For many biological brain images, the extraction method is usually through slice imaging. Because the original data volume is large, and many brain images do not have corresponding contour markers, only professional people can manually draw the brain region contours on 2D slices according to experience and observation. Moreover, due to the small amount of data with markers, the segmentation or tracking method of deep learning is difficult to achieve high accuracy. For some regions without markers, it is also impossible to automatically identify or generate new brain region contours. Manual contour marking is slow and inefficient. It took the Allen Institute for Brain Science in the United States 3 years to complete the manual annotation of the mouse whole brain reference map. Today, the macaque brain image data imaged by VISoR technology is thousands of times larger than the mouse brain. Even if the brain region drawing work is performed on the down-sampled image, it still takes months or even a year. Therefore, a method for assisting manual drawing is necessary.

[0084] Taking the VISoR macaque brain image as an example, the overall data is a 3D image, which can also be regarded as a sequence of 2D slices. According to observation, there is no big difference between adjacent 2D slices of the macaque brain image. The contour trajectory of most regions basically maintains the direction perpendicular to the contour line, that is, the normal direction. This method requires researchers to draw the brain region contour on a 2D slice, calculate the features of the points on the contour line, match the features of the pixel points in a certain range on the next image, find the most similar points, and then fit all the obtained points into a curve to automatically draw the recommended contour line on the next image. After manual correction, the next image can be drawn, and the contour drawing of the entire macaque brain can be completed. According to the above idea, the features of specific points need to be obtained, so DenseSIFT features are used here. It can extract the gradient features of each pixel point in the image to form a 128-dimensional vector for describing the pixel point. It does not need to rely on any dataset, which makes it suitable for drawing some currently determined brain region contours and new data or newly discovered brain regions, and solves the problem of no labeled dataset in the early stage of research. The traditional feature calculation combined with point matching method is fast and accurate, and the results can be adjusted in real time according to the modification of researchers, greatly saving time and manpower consumption, and ensuring the interactivity of the whole method.

[0085] As shown in Figure 2 , the image is a monkey brain VISoR image after downsampling, and the size is 1408*1203, and it can be seen from the figure that the outline of part of the brain area is not obvious; as shown in Figure 3 , the image is a monkey brain VISoR image after texture enhancement, and it can be seen from the figure that the gradient of the brain area edge is improved.

[0086] As shown in Figure 4 , the figure is the outline of part of the brain area drawn on the brain image with a brush; as shown in Figure 5 , the figure is the outline point of part of the brain area, and since the recommendation speed is slow by directly using the outline line, the outline point is used for recommendation by equidistant sampling of the outline line.

[0087] As shown in Figure 6 , the figure is adjacent to the brain image Figure 2 , which is also the target image recommended.

[0088] As shown in Figure 7 , the figure is the result of the outline point on the target image after recommendation, and it can be seen that it basically fits the brain area texture of the image; as shown in Figure 8 , the figure is the outline line fitted by the recommended outline point through the Bezier curve, and it can be seen that it basically fits the brain area texture of the image.

[0089] In summary, based on the traditional DenseSIFT feature, the gradient feature can be used for feature operation on each pixel point in the image, and the contour can be recommended without data set, so the problem of lack of artificial annotation data set for new data or new brain area is solved, and the method is very strong in universality; the points are recommended from the contour line, and the contour line is automatically drawn after the end, which greatly reduces the problems of uneven recommended contour and slow speed caused by dense points, and only needs to be drawn on one image, so that the subsequent images can be recommended, the accuracy is high, and the time consumption is short; compared with the popular machine learning method, the contour can be recommended without training data set, so the problem of lack of artificial annotation data set for new data or new brain area is solved, the method is very strong in universality, the problems of uneven recommended contour and slow speed caused by dense points are greatly reduced, the accuracy is high, and the time consumption is short.

Claims

1. A method for fast recommendation of brain region contours based on DenseSIFT features, characterized in that: The method comprises the following steps in sequence: (1) obtaining a 2D sequence biological brain image, and performing image texture enhancement: arranging a 3D image of a biological brain according to x / y / z axes to obtain a 2D sequence image, i.e., an original image, calculating Gaussian blur of the original image to obtain a Gaussian blur image, subtracting the Gaussian blur image from the original image and then adding the original image to obtain a sequence image I after image texture enhancement; (2) marking the brain region contour line of the image, sampling the contour points through the contour line: selecting a texture-enhanced image I t having a target brain region, drawing the brain region contour line through manual marking, and then sampling all points on the brain region contour line to obtain a brain region contour point set P; (3) calculating a normal slope of each brain region contour point through adjacent brain region contour point coordinates, and calculating a plurality of pixel point coordinates in the normal direction of the brain region contour point; (4) calculating DenseSIFT features of the texture-enhanced image I t and its adjacent next image I t+1 . (5) extracting the texture-enhanced image I t the feature vectors of the DenseSIFT features of the pixel points in the set P corresponding to the image I t+1 the feature vectors of the DenseSIFT features of the pixel points in the set P corresponding to the image I t+1 the initial contour recommendation point set S of the image I (6) performing initial contour recommendation point screening in the initial contour recommendation point set S through offset value calculation, performing offset correction on the initial contour recommendation point calculation constraint to obtain a corrected contour recommendation point; (7) fitting the recommended contour points after correction by a Bezier curve to obtain a recommended contour line of the image I t+1 , resampling the recommended contour line as contour points as the input of step (3), and taking the images I t+1 , I t+2 as the input of step (4), repeating the operations of steps (3) to (7) to realize the brain region contouring of each image in the sequence images I; The step (6) specifically comprises the following steps: (6a) calculating the initial contour recommendation point set S of the image I t+1 and the displacement direction between each point of the brain region contour point set P of the image I t , specifically, the difference between point P n and point S n is obtained to obtain the difference value in the x and y directions, if the difference value is positive, it is recorded as 1; if the difference value is zero, it is recorded as 0; if the difference value is negative, it is recorded as -1; (6b) judging whether the displacement direction of the point is reasonable according to the displacement direction of the adjacent initial contour recommended point, the judgment basis is: if the displacement directions of x and y of the adjacent initial contour recommended points S n-1 and S n+1 are consistent, that is, both are 1 or 0 or -1, and one of the displacement directions of x and y of the initial contour recommended point S n is consistent with S n-1 and S n+1 , then it is judged to be reasonable, otherwise it is not reasonable; if only one of the displacement directions of x and y of the adjacent initial contour recommended points S n-1 and S n+1 is consistent, and the displacement direction of the initial contour recommended point S n is consistent with the displacement directions of x and y of the initial contour recommended points S n-1 and S n+1 , then it is judged to be reasonable, otherwise it is not reasonable, and the unreasonable points are recorded; (6c) calculating a deformation field matrix between the brain region contour point set P and the initial contour recommendation point set S through a regularized thin-plate spline algorithm, performing product operation on the deformation field matrix and unreasonable points to obtain new coordinates, replacing the coordinates of the unreasonable points to make them more smooth with other points in the initial contour recommendation point set S, and obtaining the corrected contour recommendation point.

2. The method of claim 1, wherein the method is a DenseSIFT feature based fast recommendation of brain region contours. The step (2) specifically comprises the following steps: (2a) one image I of the texture-enhanced sequence of images I exists a target brain region by manual annotation t rendering is performed to obtain one or more closed curves; (2b) thinning the closed curve, saving each point on the closed curve clockwise, and marking the points of different closed curves; (2c) equally sampling the saved points to obtain the brain region contour point set P. 3.The method of claim 1, wherein: The step (3) specifically comprises the following steps; (3a) Let P be a known brain region contour point n P is a point in the set of brain region contour points P n with coordinates (X n , Y n ), and let P n-1 (X n-1 , Y n-1 ) be the previous brain region contour point and P n+1 (X n+1 , Y n+1 ) be the next brain region contour point. Subtract P n from P n-1 to get the vector P n-1 P n (X n - X n-1 , Y n - Y n-1 ), and let L1 be the unit normal to this vector. The vector representation of L1 is given by (-(Y n -Y n-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 ,(X n -X n-1 ) / [(X n -X n-1 ) 2 +(Y n -Y n-1 ) 2 ] 1 / 2 ) Similarly, the vector P is calculated. n P n+1 The unit perpendicular segment L2 is used to obtain the brain region contour point P by summing the unit perpendicular segments L1 and L2. n The normal vector is finally used as the brain region contour point P. n Dividing the y-component of the normal vector by the x-component yields the brain region contour point P. n The slope of the normal line k n ; (3b) Through brain region contour point P n and its normal slope k n The absolute value of the point is used to construct the normal form equation, and the coordinates of the multiple pixels that the normal form equation passes through in the first quadrant are obtained. If the slope of the normal is k... n If the value is negative, the obtained pixel coordinates will be reversed along the Y-axis; (3c) obtaining a plurality of pixel coordinate points along the normal direction of the brain region contour point P n performing center symmetry to obtain a plurality of pixel coordinate points along the normal direction of the brain region contour point P n ​ (3d) recording the brain region contour point P n a plurality of pixel coordinate points in the normal direction of the contour point P and the pixel coordinate points Q in the neighborhood of these pixel coordinate points 4 [n,1] , Q [n,2] ...Q [n,m] .

4. The method of claim 1, wherein the method is a DenseSIFT feature based fast recommendation of brain region contours. The step (4) specifically comprises the following steps: (4a) the DenseSIFT feature algorithm refers to: taking a local window for each pixel point, then dividing the local window into 4×4 units, dividing the gradient direction of each unit into 8 intervals, calculating the gradient of each pixel, and counting the gradient histogram of the pixel points in each interval in eight directions, thereby generating a 4×4×8=128-dimensional feature vector; the eight directions refer to right, right upper 45°, upper, left upper 45°, left, left lower 45°, lower, and right lower 45°; (4b) setting a local window radius, expanding the edges of the image after texture enhancement according to the local window radius, and the expansion size is consistent with the local window radius; the pixel value of the expansion is 0; (4c) equally dividing the expanded image into small pixel blocks, and the length and width of the pixel block are both twice the local window radius; meanwhile, moving the local window according to a step length, and the step length is set to 1, i.e., calculating each pixel; (4d) dividing each pixel block into 4×4=16 blocks, and calculating the gradient histogram of each pixel block in eight directions; (4e) after calculating the gradient of the center pixel of each small pixel block, the DenseSIFT feature of the entire image is obtained through Gaussian convolution.

5. The method of claim 1, wherein the method is a DenseSIFT feature based fast recommendation of brain region contours. The step (5) specifically comprises the following steps: (5a) extracting the image I having the drawn contour line t upper brain region contour point P n feature vector a in the coordinates n ; (5b) extracting the image I t the next image I t+1 the brain region contour point P n Q in the normal direction at the coordinates [n,1] , Q [n,2] ... Q [n,m] the feature vector β at the coordinates [n,1] , β [n,2] ... β [n,m] ; (5c) sequentially calculate the eigenvector α n with the eigenvector β [n,1] , β [n,2] ... β [n,m] of the Euclidean distance, get the minimum point of the Euclidean distance, determine as the image I t+1 corresponding to the initial contour recommended point set S.

Images