CN114266788A - Anti-artifact characteristic extraction method based on ultrasonic tomography reflected image convex target - Google Patents

Abstract

The invention belongs to the fields of computer-aided diagnosis, medical image processing, ultrasonic tomography, and medical image analysis, and specifically discloses an anti-artifact feature extraction method based on a convex target of a reflection image of ultrasonic tomography, comprising the steps of: (1) selecting a node by (2) Calculate the geometric center of gravity O ₁ of the pixel represented by the target; (3) Node update iteration; (4) Node fitting and shaping, and the contour lines obtained by connecting the new nodes after shaping correspond to The enclosed area is the corresponding image of the extracted convex target. The invention improves the overall process design of the method and the details such as iterative conditions, so that the anti-artifact feature extraction method based on the convex target of the ultrasonic tomography reflection image in the invention can The extraction is accurate and effective, while being robust to artifacts and noise.

Description

Anti-artifact characteristic extraction method based on ultrasonic tomography reflected image convex target

Technical Field

The invention relates to the fields of computer-aided diagnosis, medical image processing, ultrasonic tomography and medical image analysis, in particular to an anti-artifact characteristic extraction method based on a convex target of a reflected image of ultrasonic tomography, and the method and a corresponding system can be used in cooperation with medical ultrasonic tomography equipment.

Background

The medical ultrasonic tomography imaging equipment is novel high-end medical imaging equipment, has the advantages of non-invasion, no radiation, high resolution and high sensitivity, can obtain information of different tissues by step scanning and reconstruct a three-dimensional image of the tissues, has important clinical significance and huge application prospect in the fields of breast cancer screening and diagnosis, newborn malformation screening and diagnosis, orthopedic diagnosis and the like, and gradually becomes one of the hot spots in the ultrasonic application field. In the image mode, a reflection map displays structural information of a target, and compared with a B-ultrasonic image, the imaging range is greatly improved; in addition, the reflected image has noise and image artifacts similar to those of B-mode ultrasound and new artifacts due to its imaging method itself, which may adversely affect image processing and image analysis.

On the other hand, computer-aided medical diagnosis is widely used in actual clinics, wherein medical image analysis is a key technology for assisting doctors to interpret pathological information in medical images. In medical image analysis, feature extraction (such as geometric features including a two-dimensional area, a perimeter, a three-dimensional volume, a surface area and the like, and gray scale features including a gray scale mean, a gray scale variance and the like) of a target, especially a three-dimensional target is a clinically significant technology, and can help doctors to know the size of a focus of a patient, evaluate disease progress, treatment effect and the like.

Disclosure of Invention

Aiming at the medical image analysis requirement of medical ultrasonic tomography equipment, the invention aims to provide an anti-artifact characteristic extraction method based on an ultrasonic tomography reflected image convex target, wherein the anti-artifact characteristic extraction method based on the ultrasonic tomography reflected image convex target can accurately and effectively extract the characteristics of the reflected image of the medical ultrasonic tomography equipment by improving the overall process design, iteration conditions and other details of the method, and has good robustness on artifacts and noise. In addition, the invention further provides an anti-artifact characteristic extraction method for the convex target of the ultrasonic tomography reflected image sequence, which can more quickly and efficiently extract the three-dimensional convex target in the three-dimensional ultrasonic tomography on the premise of ensuring the extraction effect.

In order to achieve the above object, according to one aspect of the present invention, there is provided an anti-artifact feature extraction method based on a convex target of a reflection image of an ultrasonic tomography, comprising the steps of:

(1) for a layer of ultrasonic tomography reflected images of the convex target to be extracted, drawing an initial convex shell aiming at the convex target in a node selection mode; the contour lines obtained by sequentially connecting the nodes are initial convex shells, and the initial convex shells surround corresponding images of convex targets in the ultrasonic tomography reflected image;

(2) calculating the geometric gravity center O of all the nodes according to each node₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; is connected withThen, filling the target representation pixel for the triangle block; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

(3) According to each node, recording the node and O₁On the connecting line of (A) by a distance O₁The farthest target represents that the pixel point of the pixel is a new node, so that all the nodes are updated;

then, carrying out iterative processing; in each iteration, in the anticlockwise direction or the clockwise direction, for each node, judging whether the node is towards O₁Point displacement; specifically, for each node, on a connection line between the node and a next node adjacent to the node in the clockwise direction, the ratio r of target representation pixels in all pixel points passing through the connection line is_localLess than a predetermined local threshold t_localThen the node is led to O₁The point is displaced according to a preset step length distance, so that the node is updated; when the target represents the proportion r of pixels_localGreater than or equal to a preset local threshold value t_localAnd then, the node is considered to reach the target surface, at the moment, the node is fixed, and meanwhile, in the subsequent iteration process, whether the node is towards O or not is judged₁Point displacement;

after each iteration is finished, whether the integral stopping condition is met is judged; specifically, if all the adjacent nodes are connected, the ratio r of target representation pixels in all the pixels passed by the connection line is_entiretyGreater than a predetermined overall threshold t_entiretyStopping iteration, otherwise, continuing to iterate until the node reaches O₁；

(4) Performing conic curve fitting on the nodes obtained by the processing in the step (3) by using a linear regression method; then, with the node and O₁The intersection point of the connecting line of the points and the fitting curve is a sampling point, and point sampling is carried out on the curve obtained by fitting to obtain a plurality of sampling points corresponding to each node; then, for each node, carrying out weighted fusion on the node and the corresponding sampling point pairwise to obtain a new node after shaping; the contour line obtained by connecting the shaped new nodes is positioned on the convex target to be extractedCorrespondingly surrounding areas on the ultrasonic tomography reflected images are corresponding images of the extracted convex targets; and the corresponding imaging of the convex target obtained by extraction is used for subsequent target feature extraction.

As a further preferable aspect of the present invention, the ultrasound tomography reflected image of the convex object to be extracted is a sequence of multilayer ultrasound tomography reflected images, the method further includes a step (5), and the step (5) includes the sub-steps of:

(5-1) respectively drawing corresponding initial convex shells and geometric gravity centers O on the upper adjacent layer image and the lower adjacent layer image of the ultrasonic tomography reflected image of which the target contour line is obtained on the basis of the contour line obtained in the step (4)₀', wherein, the point O₀Is a point O₁Projection points on the upper adjacent layer image and the lower adjacent layer image; the initial convex shells of the upper adjacent layer image and the lower adjacent layer image are similar to the shape of the obtained target contour line of the layer; for each initial convex hull, the node of the initial convex hull is obtained by amplifying the node on the obtained target contour line of the layer, specifically, the node distance point O of the initial convex hull₀A distance of' equal to α₀Distance point O of corresponding node on the obtained target contour line₁A distance of₀Is a preset constant with the value larger than 1;

then, repeating the steps (2) to (4), and respectively obtaining target contour lines of the upper adjacent layer image and the lower adjacent layer image, thereby obtaining three layers of ultrasonic tomography reflected images with the target contour lines;

(5-2) continuously obtaining adjacent to-be-processed adjacent images in the multilayer ultrasonic tomography reflected image sequence upwards or downwards, and recording three layers of ultrasonic tomography reflected images as a 0 th layer, a 1 st layer and a 2 nd layer according to the distance between the adjacent to-be-processed adjacent images and the to-be-processed adjacent images in the three layers of ultrasonic tomography reflected images with the target contour lines, wherein the 0 th layer is closest to the adjacent to-be-processed image, and the 1 st layer is the farthest from the adjacent to-be-processed image;

then theCalculating the mean value d of Euclidean distances between all nodes and the geometric center of gravity on each layer respectively aiming at the 0 th layer, the 1 st layer and the 2 nd layer_ave(ii) a Wherein d of the 0 th layer_aveIs marked as

Layer 1 d_aveIs marked as

D of layer 2_aveIs marked as

Next, the scaling coefficient α is calculated:

then, respectively drawing the initial convex hull and the geometric gravity center O corresponding to the to-be-processed adjacent layer image₀ ^’Wherein, point O₀ ^’Is a point O of the 0 th layer₁A projection point on the adjacent layer image to be processed; the shape of the initial convex hull of the to-be-processed adjacent layer image is similar to that of the target contour line of the 0 th layer, the node of the initial convex hull is obtained by amplifying the node on the target contour line of the 0 th layer, and specifically, the node distance point O of the initial convex hull of the to-be-processed adjacent layer image₀' is a distance equal to α × the distance of the corresponding node on the object outline of layer 0 from the point O of layer 0₁The distance of (d);

then, the target existence condition is judged, specifically:

calculating the geometric gravity center O of all the nodes according to the nodes of the initial convex shell₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

Then, for the trapezoidal partition, selecting a near O inside the trapezoidal partition₀Partial area of the point is used as a mark area, and the proportion r of target representation pixels in all pixel points in the mark area is calculated_obj；

When all trapezoids are blocked_objAre all equal to or larger than a preset threshold value t_objContinuing to execute the step (3), the step (4) and the step (5-2);

when there is a trapezoidal block r_obj<A predetermined threshold value t_objWhen the direction image is processed, the processing of the current direction image is terminated;

therefore, the target extraction of the multilayer ultrasonic tomography reflected image sequence can be completed.

As a further preferred aspect of the present invention, in the step (4), the weighted fusion satisfies the following formula:

wherein p is_bnewFor the shaped node, p_bFor nodes before shaping, p_cTo correspond to the sampling point, p_centerIs O₁Point, Δ is the laplacian operator; dist_α(p_b，p_c) Is the Euclidean distance between two points in the belt direction, when p_bDist when within the fitted conic section_α(p_b，p_c) Is composed ofNegative, otherwise positive.

As a further preferred aspect of the present invention, in the step (3), r is_entiretyIs calculated according to the following formula:

in the formula, N represents the total number of nodes, the node number is sequentially increased by 1 from 1 in the clockwise direction, and N_iRepresenting the number of sampled pixel points on the connection line of the ith node and the (i-1) th node, m_iRepresenting the number of pixel points sampled on the connection line of the ith node and the (i-1) th node as target representation pixel points; and, when the ith node remains stationary, let m_i＝n_i。

As a further preferred aspect of the present invention, the adaptive threshold segmentation is specifically performed by using an algorithm of madzu;

in the step (2), for each triangle, the triangle internal distance O is determined₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the step (5-2), for each triangle, the triangle internal distance O is determined₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the step (5-2), the marking region is specifically: and aiming at a certain trapezoid partition block, respectively quartering two waists of the trapezoid shape, respectively taking an equant point closest to the triangle partition block on the two waists as a mark point, and connecting the mark points on the two waists, wherein the area in the trapezoid partition block and between the mark point connecting line and the triangle partition block is a mark area.

According to another aspect of the present invention, the present invention provides an anti-artifact feature extraction system based on a convex target of a reflection image of an ultrasonic tomography, comprising:

an image pre-processing function module to: for a layer of ultrasonic tomography reflected images of the convex target to be extracted, drawing an initial convex shell aiming at the convex target in a node selection mode; the contour lines obtained by sequentially connecting the nodes are initial convex shells, and the initial convex shells surround corresponding images of convex targets in the ultrasonic tomography reflected image;

a geometric centroid calculation function for a target representation pixel, for: calculating the geometric gravity center O of all the nodes according to each node₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

A node retraction function module, configured to: according to each node, recording the node and O₁On the connecting line of (A) by a distance O₁The farthest target represents that the pixel point of the pixel is a new node, so that all the nodes are updated;

and for performing an iterative process: in each iteration, in the anticlockwise direction or the clockwise direction, for each node, judging whether the node is towards O₁Point displacement; specifically, for each node, the line connecting the node with the next node adjacent in the clockwise directionIn the above, the ratio r of the target representation pixels in all the pixel points through which the connection line passes_localLess than a predetermined local threshold t_localThen the node is led to O₁The point is displaced according to a preset step length distance, so that the node is updated; when the target represents the proportion r of pixels_localGreater than or equal to a preset local threshold value t_localAnd then, the node is considered to reach the target surface, at the moment, the node is fixed, and meanwhile, in the subsequent iteration process, whether the node is towards O or not is judged₁Point displacement;

after each iteration is finished, whether the integral stopping condition is met is judged; specifically, if all the adjacent nodes are connected, the ratio r of target representation pixels in all the pixels passed by the connection line is_entiretyGreater than a predetermined overall threshold t_entiretyStopping iteration, otherwise, continuing to iterate until the node reaches O₁；

A shaping function module to: performing conic curve fitting on the nodes obtained by processing the node retraction function module based on a linear regression method; then, with the node and O₁The intersection point of the connecting line of the points and the fitting curve is a sampling point, and point sampling is carried out on the curve obtained by fitting to obtain a plurality of sampling points corresponding to each node; then, for each node, carrying out weighted fusion on the node and the corresponding sampling point pairwise to obtain a new node after shaping; the area correspondingly surrounded by the contour line obtained by connecting the shaped new nodes on the ultrasonic tomography reflected image of the convex target to be extracted is the corresponding image of the extracted convex target; and the corresponding imaging of the convex target obtained by extraction is used for subsequent target feature extraction.

As a further preferred aspect of the present invention, the system further comprises:

the upper and lower adjacent layer image preprocessing function module is used for: respectively drawing corresponding initial convex shells and geometric gravity centers O on the upper adjacent layer image and the lower adjacent layer image of the ultrasonic tomography reflected image with the obtained target contour line on the basis of the contour line obtained by the shaping function module₀ ^’Wherein, in the step (A),point O₀ ^’Is point O₁Projection points on the upper adjacent layer image and the lower adjacent layer image; the initial convex shells of the upper adjacent layer image and the lower adjacent layer image are similar to the shape of the obtained target contour line of the layer; for each initial convex hull, the node of the initial convex hull is obtained by amplifying the node on the obtained target contour line of the layer, specifically, the node distance point O of the initial convex hull₀ ^’Is equal to alpha₀Distance point O of corresponding node on the obtained target contour line₁A distance of₀Is a preset constant with the value larger than 1;

the upper and lower adjacent layer image preprocessing function module is also connected with the geometric gravity center calculation function module of the target representation pixel, the node retraction function module and the shaping function module, so that target contour lines of an upper adjacent layer image and a lower adjacent layer image can be obtained, and three layers of ultrasonic tomography reflected images with the target contour lines are obtained;

the function module for automatically extracting the three-dimensional object contour line is used for: continuously obtaining adjacent to-be-processed adjacent images in a multilayer ultrasonic tomography reflected image sequence upwards or downwards, and respectively recording three layers of ultrasonic tomography reflected images as a 0 th layer, a 1 st layer and a 2 nd layer according to the distance between the adjacent to-be-processed adjacent images and the to-be-processed adjacent images with the target contour line, wherein the 0 th layer is closest to the adjacent to-be-processed adjacent images, the 1 st layer is the same as the adjacent to-be-processed adjacent images, and the 2 nd layer is farthest from the adjacent to-be-processed images;

then, the mean value d of the Euclidean distances between all nodes and the geometric center of gravity on each layer is calculated respectively for the 0 th layer, the 1 st layer and the 2 nd layer_ave(ii) a Wherein d of the 0 th layer_aveIs marked as

Layer 1 d_aveIs marked as

D of layer 2_aveIs marked as

Next, the scaling coefficient α is calculated:

then, respectively drawing the initial convex hull and the geometric gravity center O corresponding to the to-be-processed adjacent layer image₀', wherein, the point O₀' Point O of layer 0₁A projection point on the adjacent layer image to be processed; the shape of the initial convex hull of the to-be-processed adjacent layer image is similar to that of the target contour line of the 0 th layer, the node of the initial convex hull is obtained by amplifying the node on the target contour line of the 0 th layer, and specifically, the node distance point O of the initial convex hull of the to-be-processed adjacent layer image₀' is a distance equal to α × the distance of the corresponding node on the object outline of layer 0 from the point O of layer 0₁The distance of (d);

then, the target existence condition is judged, specifically:

calculating the geometric gravity center O of all the nodes according to the nodes of the initial convex shell₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on all of the initial convex hullsThe coordinates of the pixel points of the target representation pixels are calculated, and the geometric gravity center O of the target representation pixels is calculated₁；

Then, for the trapezoidal partition, selecting a near O inside the trapezoidal partition₀Partial area of the point is used as a mark area, and the proportion r of target representation pixels in all pixel points in the mark area is calculated_obj；

The function module for automatically extracting the three-dimensional target contour line is connected with the node retraction function module and the shaping function module,

when all trapezoids are blocked_objAre all equal to or larger than a preset threshold value t_objThen, the target contour line of the to-be-processed adjacent layer image is obtained through the connected node retraction function module and the shaping function module, and meanwhile, the function module for automatically extracting the three-dimensional target contour line is continuously utilized for processing;

when there is a trapezoidal block r_obj<A predetermined threshold value t_objWhen the direction image is processed, the processing of the current direction image is terminated;

therefore, the target extraction of the multilayer ultrasonic tomography reflected image sequence can be completed.

As a further preferred aspect of the present invention, in the shaping function module, the weighted fusion satisfies the following formula:

wherein p is_bnewFor the shaped node, p_bFor nodes before shaping, p_cTo correspond to the sampling point, p_centerIs O₁Point, Δ is the laplacian operator; dist_α(p_b，p_c) Is the Euclidean distance between two points in the belt direction, when p_bDist when within the fitted conic section_α(p_b，p_c) Negative, otherwise positive.

As a further preferred aspect of the present invention, in the node retraction function module, r_entiretyIs calculated according to the following formula:

in the formula, N represents the total number of nodes, the node number is sequentially increased by 1 from 1 in the clockwise direction, and N_iRepresenting the number of sampled pixel points on the connection line of the ith node and the (i-1) th node, m_iRepresenting the number of pixel points sampled on the connection line of the ith node and the (i-1) th node as target representation pixel points; and, when the ith node remains stationary, let m_i＝n_i。

As a further preferred aspect of the present invention, the adaptive threshold segmentation is specifically performed by using an algorithm of madzu;

in the geometric gravity center calculation function module of the target representation pixel, for each triangle, the triangle internal distance O is calculated₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the functional module for automatically extracting the three-dimensional target contour line, for each triangle, the internal distance O of the triangle is used₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the functional module for automatically extracting the three-dimensional target contour line, the marking area is specifically as follows: and aiming at a certain trapezoid partition block, respectively quartering two waists of the trapezoid shape, respectively taking an equant point closest to the triangle partition block on the two waists as a mark point, and connecting the mark points on the two waists, wherein the area in the trapezoid partition block and between the mark point connecting line and the triangle partition block is a mark area.

Through the technical scheme, compared with the prior art, the anti-artifact characteristic extraction method and the system based on the ultrasonic tomography reflected image convex target have the advantages that firstly, the existence of artifacts is considered, the target is subjected to trapezoidal partitioning, and self-adaptive threshold partitioning is carried out on the partitions; secondly, considering the uniform characteristics of the convex shell nodes, such as the characteristic of smooth change of the surface of the convex target, the convex shell nodes are always associated with all the nodes in the process of carrying out iterative updating, and the influence caused by artifacts and noise is reduced to the maximum extent by using local-overall dual-threshold stopping conditions and a linear regression method to resist the artifact contour shaping, so that an accurate contour is extracted. For the ultrasonic tomography reflected image sequence, the pixel continuity and the target existence condition in the scanning direction of the image sequence are considered, the invention further provides a target extraction processing step for the images of the upper and lower adjacent layers of the processed image in the image sequence, thereby automatically realizing the segmentation of the whole three-dimensional target, saving the operation complexity and time cost and obtaining an accurate segmentation result. Through the method, the two-dimensional target or the three-dimensional target is subjected to accurate feature extraction finally. The method can be applied to the situation of extracting the geometric and gray features of the convex target in the ultrasonic fault reflection diagram (sequence) under the conditions of more serious artifacts and noises, has accurate result and good robustness to the artifacts and the noises, can help researchers, detectors and the like to accurately extract the contour and the features of the target, and is convenient for people to concentrate on subsequent analysis and processing.

In medical imaging diagnosis, the method has great significance for extracting the features of an interested target, and can help researchers, detection personnel and doctors to carry out image quality evaluation, basic development condition evaluation, curative effect evaluation and the like. Feature extraction generally requires segmentation or contour extraction of the object of interest. The reflection image of the ultrasonic tomography equipment can better represent the internal structure information of the scanning part. As a new medical image modality, noise and image artifacts similar to B-mode ultrasound exist in a reflection image and new artifacts brought by an imaging method of the reflection image. The conventional contour extraction or segmentation methods such as snake model, watershed algorithm, region growing and the like are greatly influenced by noise and artifacts, and accurate results are difficult to obtain. In order to overcome the influence of the artifact, the invention corrects the contour abnormality generated by the local artifact through the integral unity, and realizes the artifact-resistant convex hull shrinkage based on the spiral iteration and the artifact-resistant contour shaping based on the linear regression method. Further, considering the processing of the image sequence, the invention finally obtains a method for acquiring the three-dimensional target contour on the basis of predicting the variation trend thereof by using the second order differential (because the target contour variation in the image sequence is not simply enlarged or reduced layer by layer).

In summary, the method for extracting the anti-artifact feature based on the ultrasound tomography reflected image convex target in the present invention is a target feature extraction method with strong robustness to the noise and artifact in the reflected image of the ultrasound tomography device (of course, is also suitable for performing anti-artifact feature extraction on the ultrasound tomography reflected image sequence, i.e. three-dimensional target feature extraction), and can suppress the influence of the noise and artifact, extract an accurate contour, and calculate an accurate feature.

Drawings

FIG. 1 is a three-dimensional display of an ultrasound tomography reflectometry; two objects are shown.

FIG. 2 is an ultrasound tomography reflectance map; in addition to the two objects, there are also multiple reflection artifacts and radial artifacts.

Fig. 3 (a) and 3 (b) are diagrams in which the operator manually draws the target convex hull for a single target, respectively.

Fig. 4 (a) and fig. 4 (b) are the result of contour extraction of two objects in the ultrasound tomography reflectogram by the method of the present invention, respectively. Taking fig. 4 (a) as an example, the outer contour line of the 2 contour lines shown in the figure is used for manually drawing the target convex hull (which is consistent with fig. 3 (a)) for the operator, and the inner contour line is used for connecting the nodes obtained after the nodes are retracted after the processing of the method of the present invention.

Fig. 5 is a contour extraction result of two three-dimensional objects in an ultrasonic tomography reflection map sequence.

FIG. 6 is a schematic flow chart of the method for extracting the anti-artifact characteristics of the convex target based on the ultrasonic tomography reflected image sequence.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

In general, the anti-artifact characteristic extraction method based on the ultrasonic tomography reflected image sequence convex target can be used for manually drawing an initial target convex shell by an operator, and then, the difficulty in segmentation caused by local artifacts and uneven gray scale is overcome by utilizing trapezoidal partitioning and self-adaptive threshold segmentation, so that an accurate target range is obtained; and then, performing spiral iteration retraction on all convex shells through a local-overall double-threshold stopping condition, positioning the target surface according to the relative position information of the nodes and the target range, and simultaneously considering the integrity of the convex shells to avoid node over-retraction and under-retraction caused by local artifacts and noise. After the local-overall dual-threshold stopping condition is met, the conic curve fitting is carried out on the sections by using a least square method (or other linear regression methods), and then fusion is carried out, so that the shape distortion caused by the fact that artifacts and noise on local contours are too heavy is further avoided, and a more accurate result is obtained. After the accurate target surface of the two-dimensional image is obtained, the convex hull of the adjacent image is generated in a self-adaptive mode according to the continuity of pixels in the sequence image and the target existence condition, the accurate target surface is extracted, meanwhile, the operation can be supervised by an operator, the operation complexity is reduced, meanwhile, the target is accurately segmented, and the features are extracted.

The method of the present invention is described in detail below by taking an example of an ultrasound tomography reflectometer convex profile extraction method based on an iterative profile and a least square method. The method can specifically comprise the following steps:

step one, acquiring a reflection image sequence:

firstly, a medical ultrasonic tomography device is used for data acquisition and image reconstruction, and a reflection image sequence and image parameters including voxel width (mm), height (mm) and thickness (mm) are obtained. A three-dimensional object to be feature extracted is then determined from the image. The three-dimensional display of the image sequence is shown in fig. 1, wherein the object of feature extraction is a sphere of an ellipsoid of low brightness in the figure. As shown in fig. 2, both have strong artifacts on the image, wherein the streak artifact on the ellipse is multiple reflection, and the radial artifact at the circle is the artifact caused by the reflection map imaging algorithm.

Step two, drawing a target convex shell:

on the single layer reflection image (fig. 2), an initial convex hull of the target object can be drawn by the operator by selecting a node, as shown in fig. 3.

Step three, self-adaptive threshold segmentation based on block trapezoids:

firstly, according to the geometric gravity center O of the node and the convex shell selected in the step two₀Trapezoidal partitioning is carried out on the inner part of the convex shell; then, each trapezoid block is subjected to adaptive threshold segmentation, so that the problem of unacceptable target segmentation deviation caused by uneven distribution of pixels on the surface of a target is avoided, and the accuracy and the robustness are improved; finally obtaining the geometric gravity center O of the segmentation result₁。

Step four, artifact-resistant convex hull retraction based on spiral iteration:

updating convex shell nodes according to the segmentation result of the step three; and then the convex shell retraction is controlled according to the local-overall double-threshold stopping condition, the shape of the convex shell is kept flexibly, and the adverse effect of local artifacts is inhibited.

Step five, artifact-resistant contour shaping based on a least square method:

and fitting the corresponding conical curve of the node by a least square method, and then fusing the result of the step four and the conical curve to shape the convex shell, so that the error caused by the local artifact in the step four is eliminated, the robustness is improved, and the outline of the single-layer reflection image is obtained, as shown in fig. 4.

Step six, obtaining the outline of the three-dimensional target:

due to the fact that the adjacent layer images have pixel continuity, the contour of the whole three-dimensional target is extracted in three steps. And step one, automatically drawing a target convex hull for the images of the upper and lower adjacent 2 layers, and extracting the contour through the steps three, four and five. And step two, automatically drawing image target convex hulls of other layers by using the change condition of the adjacent layer image convex hull, repeatedly executing the step three, the step four, the step five and the step two, and taking the target existence condition as a termination condition. And thirdly, checking whether the contour extraction of the three-dimensional target is finished or not to obtain the final contour of the three-dimensional target, as shown in fig. 5 (if the automatic processing of the computer does not finish the extraction, the operator finishes the accurate contour extraction of the three-dimensional target).

Seventhly, extracting the target features based on the accurate contour:

according to the result, the geometric features, the gray features and other features of the target are extracted.

Example 1

The method comprises the following steps:

the first step is that the image characteristics of the reflection image sequence are obtained as follows:

the image sequence is obtained by scanning a medical ultrasonic tomography system based on a ring probe. The obtained reflection image has a certain degree of noise and artifacts due to the equipment, the imaged object, and the like, and the main artifacts include multiple reflections and radial artifacts in the middle of the image, as shown in fig. 2.

Step two, drawing an initial target convex shell, for example:

and the operator selects the nodes of the convex shells in a one-way mode outside the surface of the feature extraction target according to the clockwise direction or the anticlockwise direction. The nodes are selected as much as possible according to the curvature condition of the target surface, and the nodes are selected outside the surface with larger curvature.

After the nodes are selected, the connecting lines of the adjacent nodes form a convex shell, and the shape of the convex shell is similar to that of the target surface as much as possible.

Step three, the detail steps of the self-adaptive threshold segmentation based on the block trapezoid are as follows:

first, the geometric center of gravity O of the convex hull is calculated₀The calculation method is to calculate the average value of all node coordinates (thus ensuring O₀Inside the target).

Then, for each node-O in turn₀Trapezoidal partitioning is carried out in a triangular area surrounded by the connecting line and the convex shellAnd adaptive threshold segmentation. Specifically, the node is formed by 2 adjacent nodes and O₀In any triangular region formed, with node-O₀The midpoints of the two edges (i.e., node-O)₀The middle point of the connecting line) and 4 points of 2 nodes are taken as the trapezoidal vertexes together, the trapezoidal blocks are generated, and the image in each trapezoidal block is subjected to adaptive threshold segmentation by using an Otsu algorithm. For example, the value above the threshold is set to 1, and the value below the threshold is set to 0. After division, will be close to O₀One of the binary pixels with the largest number of distributions is a representation pixel of the target, and the other binary pixel with the smaller number is O₀Distributed is a background representation pixel.

Finally, filling target representation pixels in a central polygonal area surrounded by all the trapezoid blocks;

for the central polygonal area and the surrounding trapezoidal blocks, calculating O based on the representing pixel points of all targets₁The coordinates are the average of the coordinates of all targets representing the pixel points.

Step four, the detail steps of the artifact-resistant convex hull retraction based on the spiral iteration are as follows:

firstly, according to the division result of the third step, making node-O₁Distance on connecting line O₁The farthest object represents the pixel as the new convex hull node.

Then, all the nodes are orderly led to be towards O according to the node selection direction₁And conditionally performing a section of displacement on the points, regarding all node displacements as one iteration, and fixing all nodes to form a new convex shell when a local-overall dual-threshold stopping condition is met. Specifically, at each node to O₁Before the point displacement, the proportion r of the target representation pixel on the connecting line of the node and the next node is calculated_localWhen r is_localLess than a local threshold t_localTime (local threshold t)_localThe value of (c) can be preset, in this embodiment to 0.9), and the node is made to go to O₁The distance of the dot moving by d pixels (the value of d can be preset as well, in this embodiment, is preset as a unit pixel), if r_local≥t_localThe node is considered to arrive at the target surface and is fixed for subsequent iterationsTo the middle stop to O₁The point is shifted and the next node is processed. After each iteration is finished, whether the integral stopping condition is met is judged, namely the proportion r of target representation pixels on the connecting lines of all adjacent nodes on the convex shell_entiretyWhether it is greater than the overall threshold t_entirety(Overall threshold value t)_entiretyThe value of (c) can also be preset, in this embodiment to 0.95), when the condition is satisfied, the iteration is stopped, otherwise, the iteration is continued until the node reaches O₁At this time, it is considered that there is no object or the contour extraction of the object fails.

In a concrete operation, the proportion in which the target represents pixels can be calculated as r by uniformly sampling 20 points on the line connecting adjacent contour points_local(ii) a In calculating r_entiretyWhen the target representation pixel is a point which is locally stopped to retract, all 20 points on a connecting line between the target representation pixel and the last adjacent node are regarded as the target representation pixel, specifically:

in the formula, N represents the total number of contour points, N_iRepresenting the number m of sampled pixel points on the connecting line of the ith contour point and the (i-1) th contour point_iAnd the number of pixel points sampled on the connecting line of the ith contour point and the (i-1) th contour point is the target representation pixel point. R when the ith contour point_local>t_localWhen the turn is round (i.e., when the ith node reaches the target surface and is fixed), let m_i＝n_i。

Step five, the detail steps of the anti-artifact contour shaping based on the least square method are as follows:

firstly, a least square method is adopted to carry out conic curve fitting on the nodes in the step four.

Then, point sampling is carried out on the fitted curve, and the sampling point is the node-O₁The intersection of the line and the curve.

And finally, fusing the sampling points and the convex hull nodes to generate new convex hull nodes, and finishing the shaping of the convex hull obtained in the step four. Specifically, the weighted coordinates of the calculation node and the corresponding sampling point are used as new nodes, and the calculation formula of the weighted coordinates is as follows:

p_b，p_c，p_bnew∈R⁺²

wherein p is_bIs a node, p_cTo correspond to the sampling point, p_centerIs O₁Point, Δ is Laplace operator, Dist_α(p_b，p_c) Is the Euclidean distance between two points in the belt direction, when p_bDist when within the fitted conic section_α(p_b，p_c) Is negative, otherwise positive, p_bnewIs the shaped node. R⁺²The conventional meaning, i.e., a two-dimensional set of positive real numbers, is satisfied.

Thus, the target extraction of the single-layer ultrasonic image can be completed.

For the multi-layer ultrasound image corresponding to the three-dimensional object, in addition to processing each layer separately according to the above steps, in order to improve the efficiency and save the processing time, according to experience, the processing can also be performed according to the following steps.

Step six, the detailed steps of obtaining the outline of the three-dimensional target are as follows:

step five, drawing corresponding convex shells and geometric gravity centers O on the upper and lower adjacent layer images according to the result of the step five₀', point O₀' coordinate is point O₁And (4) obtaining coordinates of convex shell nodes by amplifying the nodes obtained in the step five, wherein the formula is as follows:

p′_b＝α*p_bnew

where α is an empirical value, the value is greater than 1, and the specific value may be preset, and in this embodiment, α is set to 1.3. And then, executing the third step to the fifth step to obtain the target outlines of the upper layer image and the lower layer image.

And secondly, drawing the adjacent layer target convex hull up and down in two directions according to the three-layer target contour change condition obtained in the first step. In particular, for an upwardly expanded image to be processed, it will beThe processed three layers of images are respectively marked as a 0 th layer, a 1 st layer and a 2 nd layer according to the distances from the processed three layers of images to the image to be processed, wherein the 0 th layer is closest to the current image to be processed, the 1 st layer is the closest to the current image to be processed, and the 2 nd layer is the farthest from the current image to be processed; respectively calculating the average value of Euclidean distances between all target profile nodes of the 0 th layer, the 1 st layer and the 2 nd layer and the geometric center of gravity

Then, calculating the scaling coefficient through the following formula, and updating the value of alpha (in the formula, 1.3 is a preset value of alpha in the first step; if the value of alpha in the first step changes, the formula also needs to change synchronously):

and (4) after drawing the target convex hull according to the formula of the first step, judging the existence condition of the target. Specifically, step three is executed to judge the proportion r of the target representation pixels in the area between the connecting line of the upper 1/4 quantile points and the upper bottom of the two waists of all trapezoids_objWhen all r is_objWhen the threshold value is not less than the threshold value, continuously and circularly executing the second steps of the step four, the step five and the step six; when there is one r_obj<At threshold, processing of the current image is terminated (this layer may be subsequently processed separately by a human, e.g., a node may be manually selected from the beginning of step). The specific value of the threshold may be set as expected, and is 0.8 in this embodiment (of course, the value of the threshold may also be flexibly adjusted, for example, it is preset to be other real numbers within a range greater than 0 and less than 1).

Similar operations can also be applied to the downward-expanded image to be processed.

Thus, the target extraction of the multilayer ultrasonic image can be completed.

And thirdly, browsing all images containing the target and checking whether each layer of image accurately extracts the target contour. And if no deviation exists, the result obtained in the step six is feasible. If the deviation exists or no extraction contour exists, the result obtained in the step six is skipped, and the single feature extraction processing operation is carried out on the layer image (namely, the step two, the step three, the step four and the step five are executed) until the accurate three-dimensional target contour extraction is completed, so that the final three-dimensional target contour is obtained.

Seventhly, extracting target features based on the accurate contour; for example:

according to the voxel parameters obtained in the first step, calculating geometric characteristics such as the area and the perimeter of the target in each layer of image, the volume and the surface area of the whole three-dimensional target and the like; and calculating pixels surrounded by the contour, and extracting gray features such as a target gray mean value, a gray variance and the like. What characteristics are finally obtained can be adjusted according to actual requirements.

The above embodiments are only examples, for example, in addition to the least square method, other linear regression methods may be used to replace the least square method to achieve a comparable feature extraction effect based on the present invention. In addition, the convex target in the invention meets the conventional definition, namely, two connecting lines on any boundary are in the target; the convex hull surrounds the contour of the convex object both near the edge of the convex object.

It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (10)

1. An anti-artifact feature extraction method based on an ultrasonic tomography reflected image convex target is characterized by comprising the following steps:

(1) for a layer of ultrasonic tomography reflected images of the convex target to be extracted, drawing an initial convex shell aiming at the convex target in a node selection mode; the contour lines obtained by sequentially connecting the nodes are initial convex shells, and the initial convex shells surround corresponding images of convex targets in the ultrasonic tomography reflected image;

(2) calculating the geometric gravity center O of all the nodes according to each node₀(ii) a Then, based on O₀Point, with O₀The line connecting each node with the initial convexDividing the shell to obtain₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

(3) According to each node, recording the node and O₁On the connecting line of (A) by a distance O₁The farthest target represents that the pixel point of the pixel is a new node, so that all the nodes are updated;

then, carrying out iterative processing; in each iteration, in the anticlockwise direction or the clockwise direction, for each node, judging whether the node is towards O₁Point displacement; specifically, for each node, on a connection line between the node and a next node adjacent to the node in the clockwise direction, the ratio r of target representation pixels in all pixel points passing through the connection line is_localLess than a predetermined local threshold t_localThen the node is led to O₁The point is displaced according to a preset step length distance, so that the node is updated; when the target represents the proportion r of pixels_localGreater than or equal to a preset local threshold value t_localAnd then, the node is considered to reach the target surface, at the moment, the node is fixed, and meanwhile, in the subsequent iteration process, whether the node is towards O or not is judged₁Point displacement;

after each iteration is finished, whether the integral stopping condition is met is judged; specifically, if all the adjacent nodes are connected, the target table in all the pixels passed by the connection line isShowing the proportion r of pixels_entiretyGreater than a predetermined overall threshold t_entiretyStopping iteration, otherwise, continuing to iterate until the node reaches O₁；

(4) Performing conic curve fitting on the nodes obtained by the processing in the step (3) by using a linear regression method; then, with the node and O₁The intersection point of the connecting line of the points and the fitting curve is a sampling point, and point sampling is carried out on the curve obtained by fitting to obtain a plurality of sampling points corresponding to each node; then, for each node, carrying out weighted fusion on the node and the corresponding sampling point pairwise to obtain a new node after shaping; the area correspondingly surrounded by the contour line obtained by connecting the shaped new nodes on the ultrasonic tomography reflected image of the convex target to be extracted is the corresponding image of the extracted convex target; and the corresponding imaging of the convex target obtained by extraction is used for subsequent target feature extraction.

2. The method as claimed in claim 1, wherein the ultrasound tomography reflection images of the convex object to be extracted are a sequence of multi-layered ultrasound tomography reflection images, the method further comprising a step (5), wherein the step (5) comprises the sub-steps of:

(5-1) respectively drawing corresponding initial convex shells and geometric gravity centers O on the upper adjacent layer image and the lower adjacent layer image of the ultrasonic tomography reflected image of which the target contour line is obtained on the basis of the contour line obtained in the step (4)₀', wherein, the point O₀Is a point O₁Projection points on the upper adjacent layer image and the lower adjacent layer image; the initial convex shells of the upper adjacent layer image and the lower adjacent layer image are similar to the shape of the obtained target contour line of the layer; for each initial convex hull, the node of the initial convex hull is obtained by amplifying the node on the obtained target contour line of the layer, specifically, the node distance point O of the initial convex hull₀A distance of' equal to α₀Distance point O of corresponding node on the obtained target contour line₁A distance of₀Is a preset constant with the value larger than 1;

then, repeating the steps (2) to (4), and respectively obtaining target contour lines of the upper adjacent layer image and the lower adjacent layer image, thereby obtaining three layers of ultrasonic tomography reflected images with the target contour lines;

(5-2) continuously obtaining adjacent to-be-processed adjacent images in the multilayer ultrasonic tomography reflected image sequence upwards or downwards, and recording three layers of ultrasonic tomography reflected images as a 0 th layer, a 1 st layer and a 2 nd layer according to the distance between the adjacent to-be-processed adjacent images and the to-be-processed adjacent images in the three layers of ultrasonic tomography reflected images with the target contour lines, wherein the 0 th layer is closest to the adjacent to-be-processed image, and the 1 st layer is the farthest from the adjacent to-be-processed image;

then, the mean value d of the Euclidean distances between all nodes and the geometric center of gravity on each layer is calculated respectively for the 0 th layer, the 1 st layer and the 2 nd layer_ave(ii) a Wherein d of the 0 th layer_aveIs marked as

Layer 1 d_aveIs marked as

D of layer 2_aveIs marked as

Next, the scaling coefficient α is calculated:

then, respectively drawing the initial convex hull and the geometric gravity center O corresponding to the to-be-processed adjacent layer image₀', wherein, the point O₀' Point O of layer 0₁A projection point on the adjacent layer image to be processed; the shape of the initial convex hull of the to-be-processed adjacent layer image is similar to that of the target contour line of the 0 th layer, the node of the initial convex hull is obtained by amplifying the node on the target contour line of the 0 th layer,specifically, the node distance point O of the initial convex hull of the to-be-processed adjacent layer image₀' is a distance equal to α × the distance of the corresponding node on the object outline of layer 0 from the point O of layer 0₁The distance of (d);

then, the target existence condition is judged, specifically:

calculating the geometric gravity center O of all the nodes according to the nodes of the initial convex shell₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

Then, for the trapezoidal partition, selecting a near O inside the trapezoidal partition₀Partial area of the point is used as a mark area, and the proportion r of target representation pixels in all pixel points in the mark area is calculated_obj；

When all trapezoids are blocked_objAre all equal to or larger than a preset threshold value t_objContinuing to execute the step (3), the step (4) and the step (5-2);

when there is a trapezoidal block r_obj<A predetermined threshold value t_objWhen the direction image is processed, the processing of the current direction image is terminated;

therefore, the target extraction of the multilayer ultrasonic tomography reflected image sequence can be completed.

3. The method of claim 1, wherein in the step (4), the weighted fusion satisfies the following formula:

wherein p is_bnewFor the shaped node, p_bFor nodes before shaping, p_cTo correspond to the sampling point, p_centerIs O₁Point, Δ is the laplacian operator; dist_α(p_b，p_c) Is the Euclidean distance between two points in the belt direction, when p_bDist when within the fitted conic section_α(p_b，p_c) Negative, otherwise positive.

4. The method of claim 1, wherein in step (3), r is_entiretyIs calculated according to the following formula:

in the formula, N represents the total number of nodes, the node number is sequentially increased by 1 from 1 in the clockwise direction, and N_iRepresenting the number of sampled pixel points on the connection line of the ith node and the (i-1) th node, m_iRepresenting the number of pixel points sampled on the connection line of the ith node and the (i-1) th node as target representation pixel points; and, when the ith node remains stationary, let m_i＝n_i。

5. The method of claim 2, wherein the adaptive threshold splitting is performed using the Otsu algorithm;

in the step (2), for each triangle, the triangle internal distance O is determined₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangleFor O to₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the step (5-2), for each triangle, the triangle internal distance O is determined₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the step (5-2), the marking region is specifically: and aiming at a certain trapezoid partition block, respectively quartering two waists of the trapezoid shape, respectively taking an equant point closest to the triangle partition block on the two waists as a mark point, and connecting the mark points on the two waists, wherein the area in the trapezoid partition block and between the mark point connecting line and the triangle partition block is a mark area.

6. An anti-artifact feature extraction system based on a convex target of an ultrasonic tomography reflected image is characterized by comprising:

an image pre-processing function module to: for a layer of ultrasonic tomography reflected images of the convex target to be extracted, drawing an initial convex shell aiming at the convex target in a node selection mode; the contour lines obtained by sequentially connecting the nodes are initial convex shells, and the initial convex shells surround corresponding images of convex targets in the ultrasonic tomography reflected image;

a geometric centroid calculation function for a target representation pixel, for: calculating the geometric gravity center O of all the nodes according to each node₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, the image within each trapezoid partition is adaptively thresholded and binary values are set based on the thresholdPixel, and simultaneously filling binary pixels into all pixel points in the trapezoidal partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

A node retraction function module, configured to: according to each node, recording the node and O₁On the connecting line of (A) by a distance O₁The farthest target represents that the pixel point of the pixel is a new node, so that all the nodes are updated;

and for performing an iterative process: in each iteration, in the anticlockwise direction or the clockwise direction, for each node, judging whether the node is towards O₁Point displacement; specifically, for each node, on a connection line between the node and a next node adjacent to the node in the clockwise direction, the ratio r of target representation pixels in all pixel points passing through the connection line is_localLess than a predetermined local threshold t_localThen the node is led to O₁The point is displaced according to a preset step length distance, so that the node is updated; when the target represents the proportion r of pixels_localGreater than or equal to a preset local threshold value t_localAnd then, the node is considered to reach the target surface, at the moment, the node is fixed, and meanwhile, in the subsequent iteration process, whether the node is towards O or not is judged₁Point displacement;

after each iteration is finished, whether the integral stopping condition is met is judged; specifically, if all the adjacent nodes are connected, the ratio r of target representation pixels in all the pixels passed by the connection line is_entiretyGreater than a predetermined overall threshold t_entiretyStopping iteration, otherwise, continuing to iterate until the node reaches O₁；

A shaping function module to: performing conic curve fitting on the nodes obtained by processing the node retraction function module based on a linear regression method; then, with the node and O₁The connection line of the points andthe intersection points of the fitted curves are sampling points, point sampling is carried out on the fitted curves, and a plurality of sampling points corresponding to each node are obtained; then, for each node, carrying out weighted fusion on the node and the corresponding sampling point pairwise to obtain a new node after shaping; the area correspondingly surrounded by the contour line obtained by connecting the shaped new nodes on the ultrasonic tomography reflected image of the convex target to be extracted is the corresponding image of the extracted convex target; and the corresponding imaging of the convex target obtained by extraction is used for subsequent target feature extraction.

7. The system of claim 6, further comprising:

the upper and lower adjacent layer image preprocessing function module is used for: respectively drawing corresponding initial convex shells and geometric gravity centers O on the upper adjacent layer image and the lower adjacent layer image of the ultrasonic tomography reflected image with the obtained target contour line on the basis of the contour line obtained by the shaping function module₀', wherein, the point O₀Is a point O₁Projection points on the upper adjacent layer image and the lower adjacent layer image; the initial convex shells of the upper adjacent layer image and the lower adjacent layer image are similar to the shape of the obtained target contour line of the layer; for each initial convex hull, the node of the initial convex hull is obtained by amplifying the node on the obtained target contour line of the layer, specifically, the node distance point O of the initial convex hull₀A distance of' equal to α₀Distance point O of corresponding node on the obtained target contour line₁A distance of₀Is a preset constant with the value larger than 1;

the upper and lower adjacent layer image preprocessing function module is also connected with the geometric gravity center calculation function module of the target representation pixel, the node retraction function module and the shaping function module, so that target contour lines of an upper adjacent layer image and a lower adjacent layer image can be obtained, and three layers of ultrasonic tomography reflected images with the target contour lines are obtained;

the function module for automatically extracting the three-dimensional object contour line is used for: continuously obtaining adjacent to-be-processed adjacent images in a multilayer ultrasonic tomography reflected image sequence upwards or downwards, and respectively recording three layers of ultrasonic tomography reflected images as a 0 th layer, a 1 st layer and a 2 nd layer according to the distance between the adjacent to-be-processed adjacent images and the to-be-processed adjacent images with the target contour line, wherein the 0 th layer is closest to the adjacent to-be-processed adjacent images, the 1 st layer is the same as the adjacent to-be-processed adjacent images, and the 2 nd layer is farthest from the adjacent to-be-processed images;

then, the mean value d of the Euclidean distances between all nodes and the geometric center of gravity on each layer is calculated respectively for the 0 th layer, the 1 st layer and the 2 nd layer_ave(ii) a Wherein d of the 0 th layer_aveIs marked as

Layer 1 d_aveIs marked as

D of layer 2_aveIs marked as

Next, the scaling coefficient α is calculated:

then, respectively drawing the initial convex hull and the geometric gravity center O corresponding to the to-be-processed adjacent layer image₀', wherein, the point O₀' Point O of layer 0₁A projection point on the adjacent layer image to be processed; the shape of the initial convex hull of the to-be-processed adjacent layer image is similar to that of the target contour line of the 0 th layer, the node of the initial convex hull is obtained by amplifying the node on the target contour line of the 0 th layer, and specifically, the node distance point O of the initial convex hull of the to-be-processed adjacent layer image₀' is a distance equal to α × the distance of the corresponding node on the object outline of layer 0 from the point O of layer 0₁The distance of (d);

then, the target existence condition is judged, specifically:

calculating the geometric gravity center O of all the nodes according to the nodes of the initial convex shell₀(ii) a Then, based on O₀Point, with O₀Dividing the initial convex hull by the connecting line of the point and each node to obtain a node O₀The points and the adjacent 2 nodes are a plurality of triangles with vertexes; then, for each triangle, the triangle internal distance O₀The distance between the points is divided into trapezoidal blocks and triangular blocks, wherein the distance between the triangular blocks is O₀The point distance is closer, the trapezoidal block distance O₀The points are far apart; then, performing self-adaptive threshold segmentation on the image in each trapezoid partition, setting binary pixels based on the threshold, and simultaneously performing binary pixel filling on all pixel points in each trapezoid partition; after the division is finished, the part is close to O₀One of the binary pixels with the largest number distributed in the region of the point is marked as a target representation pixel, and the other binary pixel is marked as a background representation pixel; then, filling target representation pixels into the triangle blocks; then, based on the pixel point coordinates of all target representation pixels in the initial convex hull, the geometric gravity center O of all the target representation pixels is calculated₁；

Then, for the trapezoidal partition, selecting a near O inside the trapezoidal partition₀Partial area of the point is used as a mark area, and the proportion r of target representation pixels in all pixel points in the mark area is calculated_obj；

The function module for automatically extracting the three-dimensional target contour line is connected with the node retraction function module and the shaping function module, and when r of all trapezoid blocks is divided into blocks_objAre all equal to or larger than a preset threshold value t_objThen, the target contour line of the to-be-processed adjacent layer image is obtained through the connected node retraction function module and the shaping function module, and meanwhile, the function module for automatically extracting the three-dimensional target contour line is continuously utilized for processing;

when there is a trapezoidal block r_obj<A predetermined threshold value t_objWhen the direction image is processed, the processing of the current direction image is terminated;

therefore, the target extraction of the multilayer ultrasonic tomography reflected image sequence can be completed.

8. The system of claim 6, wherein in the shaping function module, the weighted fusion satisfies the following formula:

wherein p is_bnewFor the shaped node, p_bFor nodes before shaping, p_cTo correspond to the sampling point, p_centerIs O₁Point, Δ is the laplacian operator; dist_α(p_b，p_c) Is the Euclidean distance between two points in the belt direction, when p_bDist when within the fitted conic section_α(p_b，p_c) Negative, otherwise positive.

9. The system of claim 6, wherein in the node retraction function, r is_entiretyIs calculated according to the following formula:

in the formula, N represents the total number of nodes, the node number is sequentially increased by 1 from 1 in the clockwise direction, and N_iRepresenting the number of sampled pixel points on the connection line of the ith node and the (i-1) th node, m_iRepresenting the number of pixel points sampled on the connection line of the ith node and the (i-1) th node as target representation pixel points; and, when the ith node remains stationary, let m_i＝n_i。

10. The system of claim 6, wherein the adaptive threshold splitting implementation is using the Otsu algorithm;

in the geometric gravity center calculation function module of the target representation pixel, for each triangle, the triangle internal distance O is calculated₀Distance of pointsThe method comprises the following steps of dividing into trapezoidal blocks and triangular blocks: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the functional module for automatically extracting the three-dimensional target contour line, for each triangle, the internal distance O of the triangle is used₀The distance of the points is divided into trapezoid blocks and triangle blocks, and the method specifically comprises the following steps: for each triangle, pass O₀Connecting the middle points of the two triangular sides of the point to obtain a trapezoidal block and a triangular block;

in the functional module for automatically extracting the three-dimensional target contour line, the marking area is specifically as follows: and aiming at a certain trapezoid partition block, respectively quartering two waists of the trapezoid shape, respectively taking an equant point closest to the triangle partition block on the two waists as a mark point, and connecting the mark points on the two waists, wherein the area in the trapezoid partition block and between the mark point connecting line and the triangle partition block is a mark area.

Images