CN111402284A - Image threshold value determination method and device based on three-dimensional connectivity - Google Patents

Abstract

The present invention provides an image threshold method and device based on three-dimensional connectivity, including: importing all two-dimensional images collected at one time; setting a segmentation threshold interval o, a discriminant parameter threshold ε; selecting a threshold search initial value μ ₀ , the The threshold search initial value μ ₀ is less than or equal to the segmentation threshold; respectively calculate the total number N of voxels in the foreground region segmented under the threshold μ _0-2 *ο, the threshold μ ₀ -ο and μ ₀ , and/or the total number of segmented voxels The number M, and/or the total number L of the segmented non-foreground region voxels; set the discriminant parameter β; if the discriminant parameter β> the discriminant parameter threshold ε, take the threshold μ ₀ corresponding to the current β, and output μ ₀ ‑o, As the best threshold; otherwise, μ ₀ ←μ ₀ +ο, continue threshold detection. The method and device provided by the present invention utilize the significant changes of the voxel space feature to find the best segmentation pixel value, so that the segmentation of the foreground is more accurate.

Description

A method and device for image threshold determination based on three-dimensional connectivity

technical field

The invention belongs to the field of image processing, and relates to a method and device for determining an image threshold, in particular to a method and device for determining an image threshold based on three-dimensional connectivity.

Background technique

The threshold method is one of the commonly used methods for image segmentation. It uses the difference in gray value between the target and the background in the image to classify the pixels by setting a threshold, so as to achieve the separation of the target and the background. Commonly used threshold methods include:

(1) Manual experience selection method

According to prior knowledge, or by analyzing and summarizing the rules of the target and the background in the image, the pixel value interval of the target and the background is obtained, and on this basis, a better threshold is found. This method cannot achieve automatic threshold selection, therefore, it is inefficient and susceptible to image quality images, resulting in significant segmentation errors.

(2) Maximum inter-class variance method

Also known as the Otsu method, its basic idea is to divide the image into two parts, the foreground and the background, according to the grayscale characteristics of the image. When the difference between the two parts is the largest, the threshold is the best. between variance.

Let M be the number of gray levels of the image, N is the total number of pixels, N1 is the total number _of background pixels, _N2 is the total number of foreground pixels, and P _i is the total number of pixels with pixel value i, then

The proportion of background pixels: ω ₀ =N ₁ /N

Proportion of foreground pixels: ω ₁ =N ₂ /N

Gray mean of background pixels:

Grayscale mean of foreground pixels:

The grayscale mean of the image: μ=ω ₀ ×μ ₀ +ω ₁ ×μ ₁

Inter-class variance of the image: σ=ω ₀ ×(μ ₀ -μ) ² +ω ₁ ×(μ ₁ -μ) ²

Substitute the between-class mean formula into:

σ=ω ₀ ×ω ₁ ×(μ ₀ -μ ₁ ) ²

This method is currently the most commonly used threshold calculation method. Since an equal threshold is used for each pixel during segmentation, it is only suitable for images with continuous foreground grayscale features.

(3) Iterative method

Its basic idea is to divide the image into two parts, the foreground and the background. When the two parts are stable, the threshold is the best, and the standard for measuring the segmentation stability is the mean of the pixel centers of the two parts.

Let M be the number of gray levels of the image, and P _i represent the total number of pixels with pixel value i, then

Grayscale center value of background pixels:

The grayscale center value of the foreground pixel:

Mean of foreground and background centers:

Iteratively generates the T value as a new threshold, when T ^t =T ^(t-1) , the T value at this time is the optimal threshold. This method works for two parts of the image that are significantly different.

Both the maximum inter-class variance method and the iterative method require the pixel value distribution characteristics of the foreground and background regions to be considered at the same time. Therefore, the process of determining the threshold value will be disturbed by the background feature information. In addition, these two methods require the distribution of pixel values in each part of the foreground to be continuous. If there is a region with contrasting pixel values in the foreground or local information is missing, a large segmentation error will also occur.

SUMMARY OF THE INVENTION

In order to overcome the deficiencies in the prior art, the inventors of the present invention have conducted keen research to provide an image threshold determination method and device based on three-dimensional connectivity, which combines the spatial distribution characteristics of foreground voxels with the value distribution of pixels (voxels). The features are combined, and the best segmentation pixel value is found by using the significant change of the voxel space feature, thereby completing the present invention.

The object of the present invention is to provide the following technical solutions:

In a first aspect, an image thresholding method based on three-dimensional connectivity, comprising:

S100, import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

S200, set segmentation threshold interval ο, discriminate parameter threshold ε;

S300, select a threshold search initial value μ ₀ , and the threshold search initial value μ ₀ is less than or equal to the segmentation threshold;

S400, calculate the threshold value μ _0-2 *ο, the threshold value μ ₀ -ο and the total number N of foreground region voxels segmented under the threshold value μ ₀ , and/or the total number M of segmented voxels, and/or segmented out The total number of non-foreground region voxels L;

S500, a discriminant parameter β is set based on the parameters determined in S400, and the discriminant parameter β is used to measure whether the total number N of voxels in the segmented foreground region increases sharply;

S600, if the discriminant parameter β> the discriminant parameter threshold ε, skip to S800; otherwise, continue to S700;

S700, μ ₀ ←μ ₀ +ο, return to S400;

S800, take the threshold μ ₀ corresponding to the current β, and output μ ₀ -ο as the optimal threshold.

Further, when the search range of the threshold μ can be predicted, the method can be implemented by the following steps:

S100, import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

S200, set segmentation threshold interval ο, discriminate parameter threshold ε;

S300, select a value smaller than the segmentation threshold as the threshold search lower bound μ ₀ ; select a value greater than the segmentation threshold as the threshold search upper bound μ ₁ ;

S400, with o as an interval, calculate the total number N of foreground region voxels segmented under each threshold μ between μ ₀ and μ ₁ , and/or the total number M of segmented voxels, and/or the segmented The total number L of non-foreground area voxels, and the above parameters are added to the queue list _N , and/or list _M , and/or list _L ;

S500, set a discriminant parameter β based on the parameters measured in S400, and this discriminant parameter β is used to measure whether the total number N of voxels in the segmented foreground region increases sharply; calculate the discriminant parameter β under each threshold and join the queue list _β ;

S600, according to the order of the threshold μ from small to large, the corresponding β values in the queue list _β are used for discrimination, until β>ε;

S700, take the threshold μ corresponding to the current β, and output μ-o as the optimal threshold.

In a second aspect, a three-dimensional connectivity-based image thresholding device for implementing the three-dimensional connectivity-based image thresholding method described in the first aspect, the device comprising:

The import module is used to import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

The parameter setting module is used for inputting the numerical value or the calculation method of the setting parameter, including the assignment of the initial value μ 0 for the segmentation threshold interval o, the discriminant parameter threshold ε, and the threshold search initial value μ ₀ , and the calculation method of the selection discriminant parameter β;

A voxel number determination module for determining the total number N of foreground region voxels segmented under thresholds _μ0-2 *ο, thresholds _μ0 -ο and _μ0 , and/or the total number of segmented voxels M, and / or the total number of segmented non-foreground region voxels L;

The discriminant parameter measurement module measures the discriminant parameter β in turn according to the order of the threshold μ from small to large;

The threshold determination module is used to determine the numerical relationship between the discriminant parameter β and the discriminant parameter threshold ε. If the discriminant parameter β> the discriminant parameter threshold ε, take the threshold μ ₀ corresponding to the current β, and output μ ₀ -o as the optimal threshold; If the discriminant parameter β≤the discriminant parameter threshold ε, the current threshold is increased by the segmentation threshold interval as a new search threshold, and the voxel quantity measurement module and the discriminant parameter measurement module are activated to perform the next threshold operation again.

Further, when the search range of the threshold μ can be predicted, the device includes:

The import module is used to import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

The parameter setting module is used to input the numerical value or calculation method of the setting parameter, including assigning values for the segmentation threshold interval o, the discriminant parameter threshold ε, the threshold search lower bound μ ₀ , and the threshold search upper bound μ ₁ , and the selection of the discriminant parameter β. Calculation;

The voxel number determination module is used to calculate the total number N of foreground region voxels segmented under each threshold μ between μ ₀ and μ ₁ , and/or the total number M of segmented voxels, with o as an interval, And/or the total number of non-foreground region voxels L, and adding the above parameters to the queue list _N , and/or list _M , and/or list _L ;

The discriminant parameter measurement module calculates the discriminant parameter β under each threshold and joins the queue list _β ;

The threshold judgment module, in the order of the threshold μ from small to large, takes the corresponding β value in the queue list _β for judgment, until the discriminant parameter β> the discriminant parameter threshold ε, takes the threshold value μ ₀ corresponding to the current β, and outputs μ ₀ -ο, as the optimal threshold.

According to a kind of image threshold determination method and device based on three-dimensional connectivity provided by the present invention, the following beneficial technical effects are brought:

Compared with the traditional method, the new method can automatically determine the threshold according to the spatial distribution characteristics of image voxels and pixel (voxel) value distribution characteristics; only the characteristics of the foreground area need to be considered, and the interference of the background can be avoided; through continuous threshold adjustment, The connectivity of 3D voxels is used to increase the effective pixels in the area where the foreground information is missing, reduce the contrast of pixel values within the foreground, and make the segmentation of the foreground more accurate.

The image threshold determination method and device based on three-dimensional connectivity provided by the present invention is suitable for any two-dimensional image in a set of two-dimensional images that can form a three-dimensional image through three-dimensional reconstruction, and is especially suitable for various images generated for target human body parts in medical treatment. Tomography-like images, such as CT (X-ray computed tomography) images, MRI (magnetic resonance imaging) images, PET (positron emission tomography) images, PET-CT images, PET-MRI images, digital breast tomography (3D mammography) images, etc. This is because in clinical observation, doctors pay more attention to the characteristics of the lesion area in the foreground in the relevant images, while the background area is often only used as a reference for comparison or even ignored directly. The threshold determination method based on three-dimensional connectivity provided by the present invention is centered on the foreground feature, which can avoid the interference of the threshold setting by the background feature, and retain as many pixels (voxels) of the foreground target as possible, which is more accurate for the later stage. Analysis offers the possibility. In contrast, the traditional maximum inter-class variance method and iterative method are more susceptible to the influence of background features and areas with large contrast in the foreground, which will cause defects in the foreground area or loss of more details, which is not conducive to later analysis and processing. .

Description of drawings

Fig. 1 is a flow chart of a method in a preferred embodiment of the present invention;

Fig. 2 is a flow chart of a method in another preferred embodiment of the present invention;

3 is a flowchart of the method described in Embodiment 1 of the present invention;

Figure 4 is a CT image of the lung with a large area of high-density shadow;

Fig. 5 is the segmentation result graph of threshold method based on manual experience selection method (μ=-350);

Fig. 6 is the segmentation result diagram of the threshold method based on the maximum between-class variance method;

Fig. 7 is the segmentation result diagram of threshold method based on iterative method;

FIG. 8 is a graph showing the segmentation result of the threshold method based on the three-dimensional connectivity threshold measurement method (μ=-280).

Detailed ways

The present invention will be further described in detail below through the accompanying drawings and embodiments. The features and advantages of the present invention will become more apparent from these descriptions.

The current image threshold calculation methods, the maximum inter-class variance method and the iterative method, all rely on the distribution characteristics of pixel values. This idea has the following problems:

(1) The threshold calculation is based on the pixel values of the foreground and background areas, which causes the threshold to be interfered by background information, which affects the accuracy of foreground area segmentation. In fact, we usually only care about the foreground segmentation results;

(2) They are very sensitive to the distribution of pixel values in the foreground and background regions. When there is a significant pixel value contrast region in the foreground, or local information is missing, serious segmentation errors will be caused.

In fact, in addition to the distribution characteristics of pixel values, the 3D image obtained from 2D image reconstruction also has voxel spatial distribution characteristics, such as voxel connectivity, which can be used to make up for the shortcomings of image segmentation based on pixel values. .

Aiming at the problems existing in the traditional threshold method, the present invention proposes an image threshold determination method and device based on three-dimensional connectivity, which combines the spatial distribution characteristics of foreground voxels with the distribution characteristics of pixel (voxel) values. Adjust the statistic of the observed voxel space feature, and use the significant change of the statistic to find the best segmentation threshold, which effectively improves the accuracy of foreground segmentation.

In the present invention, a pixel (voxel) value refers to a value of a pixel of a two-dimensional image or a voxel in a three-dimensional image. For example, in an ordinary two-dimensional grayscale image, its value range is 0 to 255; while in a CT image, the pixel value (ie CT value) can range from -1000 to 2000. In the three-dimensional image generated by the group CT cross-sectional images, the value range of each voxel may also be -1000 to 2000. Traditional thresholding methods, such as maximum inter-class variance, iterative methods, etc., all support image segmentation with a single value in pixels or voxels. In the present invention, unless otherwise specified, a pixel (voxel) value refers to a value taken by a pixel of a two-dimensional image or a voxel of a three-dimensional image.

The image threshold determination method and device based on three-dimensional connectivity in the present invention is suitable for any two-dimensional image in a set of two-dimensional images that can form a three-dimensional image through three-dimensional reconstruction, including CT image, MRI image, PET image, PET-CT image Images, PET-MRI images, digital breast tomography (3D mammography) images and other tomographic images generated for the target body.

In the medical process, CT, MRI and other methods can obtain the imaging of related parts through continuous cross-sectional scanning of a certain part of the human body, which can be used for the inspection of various diseases. These continuous sections are superimposed on each other to form a three-dimensional image of the relevant parts after technical processing. Therefore, there is a correlation between multiple sections, which reflects the three-dimensional spatial distribution characteristics of the relevant parts. It is the main idea of the present invention to combine this three-dimensional spatial distribution feature with its pixel (voxel) value distribution feature to perform threshold calculation. It can increase the meaningful pixel (voxel) points as much as possible under the premise of controlling the signal-to-noise ratio, thereby improving the accuracy of image segmentation. Although this method will also lead to an increase in part of the noise and cannot obtain completely accurate results, it loses less detailed information than the traditional method, and the results can be used as the basis for subsequent image processing methods for more accurate segmentation. The design of the algorithm provides the possibility.

The flow chart of the threshold value determination method of the present invention is shown in FIG. 1 .

S100, import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

S200, set segmentation threshold interval ο, discriminate parameter threshold ε;

The smaller the ο value, the more fine-grained statistical changes can be observed. However, this will lead to a greater amount of calculation, and the statistical value used for discrimination is also more susceptible to noise interference and fluctuations, affecting the accuracy of the judgment results. , therefore, the o value selection should consider the finer granularity under the control of noise interference, according to experience, 10 can be selected in lung CT images;

S300, select a threshold search initial value μ ₀ , and the threshold search initial value μ ₀ is less than or equal to the segmentation threshold;

The initial value μ ₀ of the threshold search can be selected from the range of the empirical value of the foreground. For example, in the lung CT image, the range of the empirical value of the window level of the lung tissue area is -450 to -600. Therefore, a lower value can be selected from this interval. value as the search initial value (for example, -600);

S400: Calculate the threshold value (μ ₀ -2*ο), the threshold value (μ ₀ -ο ₎ , and the total number of voxels N( abbreviated as the total number of voxels in the segmented foreground area), and/or the total number M of segmented voxels, and/or the total number L of segmented non-foreground area voxels; wherein, the threshold υ is set as the threshold (μ _0-2 *ο) the size of the identifiable three-dimensional connected domain in the image group or a value smaller than it, the value of the threshold υ is set to be able to distinguish and exclude the connected domain generated by other noises as the goal;

S500, a discriminant parameter β is set based on the parameters determined in S400, and the discriminant parameter β is used to measure whether the total number N of voxels in the segmented foreground region increases sharply;

S600, if the discriminant parameter β> the discriminant parameter threshold ε, skip to S800; otherwise, continue to S700;

S700, μ ₀ ←μ ₀ +ο, return to S400;

S800, take the threshold μ ₀ corresponding to the current β, and output μ ₀ −ο (that is, the threshold μ ₀ corresponding to the current β minus the segmentation threshold interval ο), as the optimal threshold.

In the present invention, the above-mentioned two-dimensional image is a grayscale image. If the two-dimensional image is an RGB color image, the RGB color image needs to be converted into a grayscale image. Conversion methods include, but are not limited to, an average method, a weighted average method, or a maximum-minimum average method. The above method is suitable for the case where the pixel value of the foreground area in the image is lower than the pixel value of the background area. If the pixel value of the foreground area in the image is higher than the pixel value of the background area, the maximum value of the current image pixel value is taken, and the maximum value is subtracted. The current value of each pixel is used as the new value of the pixel, thereby constructing a new image. Using this method, the situation where the pixel value of the foreground area is higher than the pixel value of the background area can be restored.

It can be seen from the above method that the basic idea of the method of the present invention is divided into three steps, which are: determination of pixel (voxel) segmentation threshold search range; foreground area recognition based on voxel connectivity; threshold judgment based on voxel statistical features .

(1) Determination of pixel (voxel) segmentation threshold search range

A pixel (voxel) value smaller than the segmentation threshold is selected as the initial threshold μ ₀ , as the lower bound of the threshold search range; the segmentation threshold interval o is increased step by step. Obviously, with the increase of the threshold, the pixels (voxels) points of non-zero value in the binarized image also increase continuously (set the background area to zero value, and the non-background area to non-zero value), the increased pixels In addition to some foreground pixels hidden in the foreground area, the (voxel) points also include some noise pixels (voxels) points located in the foreground area or the background area.

(2) Foreground region recognition based on voxel connectivity

Since the foreground is a meaningful whole, its voxel distribution is connected, and, due to its subjectivity in a 3D image, the larger the interconnected regions, the higher the probability of being a foreground region. Based on this assumption, the three-dimensional connected domain with the number of voxels greater than or equal to υ under different threshold μ can be marked as the foreground region corresponding to the current threshold μ.

(3) Threshold judgment based on voxel statistical features

As the threshold increases, the number of voxels in the foreground region that can be segmented will continue to increase, and synchronously, the noise data will also increase. In the early stage of threshold iteration, the noise data is mainly random noise in the image, and only part of the noise close to the foreground area may be connected to the foreground area, so the impact on the foreground area is small. Until the pixel (voxel) value interval of the background area is reached, a large number of pixels in the background area will suddenly appear, and they will form a continuous pixel (voxel) area together with some of the noise points in the early stage, which will be integrated with the foreground area, resulting in The number of voxels in the foreground area increases abruptly and substantially. Obviously, the previous threshold before such a sharp change is the best threshold that can be obtained based on the segmentation threshold interval o. At this time, as many foreground area pixels as possible are identified, and random noise is within a relatively controllable range , the interference of the background area is as small as possible.

We can design a statistic, a discriminant parameter, to detect such abrupt changes to determine the optimal threshold.

In a preferred embodiment, the discriminant parameter β may be the total number of voxels N (referred to as the total number of segmented foreground region voxels) of the three-dimensional connected domain segmented under the threshold μ and greater than or equal to the set threshold υ. Differential Adjacent Ratio A. When A exceeds the threshold ε, it is considered that the current threshold μ has exceeded the optimal segmentation threshold, and (μ-ο) is selected as the optimal segmentation threshold.

已经研究发现，在匀质区域如脏器对应的图像区域中，该统计量在未达到边界条件时具有稳定性，当阈值μ达到背景区域的像素(体素)值区间时，该值会发生陡增。Assuming that the total number of voxels in the three-dimensional connected domain with the number of voxels greater than υ under the t-th threshold μ ^t is N ₀ , the volume of the three-dimensional connected domain with the number of voxels greater than υ at the (t-1)th threshold μ ^(t-1) The total number of voxels is N _-1 , and the (t-2)th threshold μ ^(t-2) has a three-dimensional connected domain whose number of voxels is greater than υ. The total number of voxels is N _-2 , then from (t-2) to The difference value of the total number of voxels in the foreground area in the iteration of step (t-1) is (N _-1 -N _-2 ), and the difference between the total number of voxels in the foreground area in the iterations of (t-1) and t steps is (N -1 -N -2 ) The value is (N ₀ -N _-1 ), so the difference neighbor ratio of the total number of foreground voxels in the (t-2), (t-1) and t-step iterations

It has been found that in a homogeneous area such as an image area corresponding to an organ, this statistic is stable when the boundary conditions are not met. When the threshold μ reaches the pixel (voxel) value range of the background area, the value will occur. sharp increase.

In another preferred embodiment, the discriminant parameter β may also be the difference adjacent ratio B of the total number M of voxels segmented under different thresholds μ. The total number of voxels M includes voxels that are foreground regions and other non-zero valued voxels (ie, non-foreground region voxels) that are not included in them.

Assuming that the total number of voxels segmented under the t-th threshold μ ^t is M ₀ , the total number of segmented voxels under the (t-1)-th threshold μ ^(t-1) is M _-1 , the (t- 2) The total number of voxels segmented under the threshold μ ^(t-2) is M _-2 , then the difference value of the total number of segmented voxels in the iterations of (t-2) to (t-1) steps is (M _-1 -M _-2 ), the difference value of the total number of voxels segmented in the (t-1) and t step iterations is (M ₀ -M _-1 ), then, in (t-2 ), (t-1), and the difference neighbor ratio of the total number of voxels segmented in t-step iterations

In another preferred embodiment, the discriminant parameter β may also be the difference neighbor ratio C of the total number of voxels (L=M-N) in the non-foreground region segmented under the threshold μ.

Assuming that the total number of non-foreground region voxels segmented under the t-th threshold μ ^t is L ₀ , and the total number of non-foreground region voxels segmented under the (t-1)th threshold μ ^(t-1) is L _{− 1} , the total number of non-foreground region voxels segmented under the (t-2)th threshold μ ^(t-2) is L _-2 , then the segmentation is performed in the (t-2) to (t-1) step iterations The difference value of the total number of non-foreground area voxels is (L _-1 -L _-2 ), and the difference value of the total number of non-foreground area voxels segmented in the (t-1) and t step iterations is ( L ₀ -L _-1 ), then, the difference neighbor ratio of the total number of non-foreground region voxels segmented in (t-2), (t-1) and t-step iterations

In another preferred embodiment, the discriminant parameter β can also be the absolute value of the ratio of the total number N of voxels in the foreground region segmented under the threshold μ to the total number M of segmented voxels (R=|N/M| ) of the differential neighbor ratio D.

Assuming that the absolute value of the ratio of the total number N of voxels in the foreground region segmented to the total number M of segmented voxels under the t-th threshold μ ^t is R ₀ , under the (t-1)th threshold μ ^(t-1) The absolute value of the ratio of the total number N of voxels in the segmented foreground area to the total number of voxels M is R _-1 , and the voxels in the foreground area segmented under the (t-2)th threshold μ ^(t-2) The absolute value of the ratio of the total number N to the total number M of segmented voxels is R _-2 , then the total number N of foreground area voxels segmented in the iterations from (t-2) to (t-1) steps is the same as the segmented voxel. The difference value of the absolute value of the ratio of the total number of voxels M is (R _-1 -R _-2 ), and the total number N of voxels in the foreground area segmented in the (t-1) and t step iterations is the same as the segmented The difference value of the absolute value of the ratio of the total number of voxels M is (R ₀ -R _-1 ), so the total number of voxels in the foreground area segmented in (t-2), (t-1) and t-step iterations Difference adjacent ratio of the absolute value of the ratio of the number to the total number of voxels segmented

In another preferred embodiment, the discriminant parameter β may further include the absolute value of the ratio ( S=|N/L|), the difference neighbor ratio E; or the absolute value of the ratio of the total number L of non-foreground region voxels segmented to the total number M of segmented voxels under the threshold μ (T=|L/M |), the difference adjacent ratio F; or the difference of the absolute value (1/R=|M/N|) of the ratio of the total number of voxels M segmented to the total number N of segmented foreground area voxels under different threshold μ Adjacent ratio; or the absolute value of the ratio (1/S=|L/N|) of the total number of non-foreground region voxels segmented under the threshold μ (L=M-N) to the total number N of segmented foreground region voxels The difference neighbor ratio; or the difference neighbor ratio of the absolute value (1/T=|M/L|) of the ratio of the total number M of voxels segmented under the threshold μ to the total number L of voxels in the non-foreground area.

The sensitivities of these statistics are different, so there will be some differences in the results of the discrimination, but they are all close, reflecting the observation of this sharp increase phenomenon from different perspectives.

For the convenience of description, we collectively refer to such statistics used for threshold judgment as discriminant parameters.

In the present invention, when the upper bound of the segmentation threshold can be determined, the search range of the threshold μ can be predicted. At this time, the number of voxels under each threshold μ is greater than or equal to the total number of voxels N of the three-dimensional connected domain of υ (that is, the total number of voxels in the segmented foreground area), and/or the threshold (μ ₀ -2*ο) The total number M of voxels segmented down, and/or the total number L of non-foreground region voxels segmented under the threshold (μ ₀ -2*ο), and the discriminant parameter β value based on the above parameters can all be obtained by parallel computing .

In view of this, the flow chart of an accelerated threshold determination method is shown in FIG. 2 .

S100, import all the two-dimensional images collected at one time to obtain a two-dimensional image group list ₀ , all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

S200, set segmentation threshold interval ο, discriminate parameter threshold ε;

S300, select a value smaller than the segmentation threshold as the threshold search lower bound μ ₀ ; select a value greater than the segmentation threshold as the threshold search upper bound μ ₁ . A value can be selected from the range of empirical values of the foreground as the lower bound μ ₀ of the threshold search. For example, in the lung CT image, the empirical value range of the window level of the lung tissue area is -450 to -600. Therefore, a value can be selected from this interval. The lower value is used as the initial search value (for example, -600); similarly, according to experience, the CT value of the normal lung tissue area is generally less than 0. Therefore, 0 can be set as the upper bound μ ₁ of the threshold search accordingly, and the search limit can be limited. The upper and lower bounds range can effectively reduce the amount of calculation, and also avoid the interference of other segment values;

S400, with o as an interval, calculate the total number N of voxels in the three-dimensional connected domain where the number of voxels in the foreground region is greater than or equal to the set threshold υ under each threshold μ between μ ₀ and μ ₁ (referred to as segmented The total number of voxels in the foreground area), and/or the total number of segmented voxels M, and/or the total number of segmented non-foreground area voxels L, and the above parameters are added to the queue list _N , and/or list _M , and/or list _L ;

S500, set a discriminant parameter β based on the parameters measured in S400, and this discriminant parameter β is used to measure whether the total number N of voxels in the segmented foreground region increases sharply; calculate the discriminant parameter β under each threshold and join the queue list _β ;

S600, according to the order of the threshold μ from small to large, the corresponding β values in the queue list _β are used for discrimination, until β>ε;

S700, take the threshold μ corresponding to the current β, and output μ-o as the optimal threshold.

In this method, the setting of the discriminant parameter β is consistent with the setting of the discriminant parameter β in the method when the search range of the threshold μ cannot be predicted.

The method of the invention can well solve the image segmentation of the complex pixel value distribution in the background area, the lack of local information in the foreground or the existence of pixel value contrast areas. Compared with the traditional method, this method fully combines the spatial distribution characteristics of foreground voxels in the image with the value distribution characteristics of pixels (voxels) for image segmentation, and uses the changes of the foreground spatial distribution characteristics to find the best segmentation threshold. It can not only avoid the interference of the background, but also increase the effective pixels (voxels) in the foreground area under the premise of ensuring a suitable signal-to-noise ratio, reduce the contrast of pixel values in the foreground, and increase the available information in the information-missing area by adjusting the threshold. , to make the foreground segmentation result more accurate.

According to a second aspect of the present invention, an image thresholding device based on three-dimensional connectivity is provided, the device comprising:

The import module is used to import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

The parameter setting module is used for inputting the numerical value or the calculation method of the setting parameter, including the assignment of the initial value μ 0 for the segmentation threshold interval o, the discriminant parameter threshold ε, and the threshold search initial value μ ₀ , and the calculation method of the selection discriminant parameter β;

The voxel number determination module is used to determine the threshold value (μ ₀ -2*ο), the threshold value (μ ₀ -ο) and the volume of the three-dimensional connected domain whose voxel number is greater than or equal to the set threshold value υ in the foreground area under the threshold value μ ₀ the total number of voxels N (referred to as the total number of voxels in the segmented foreground area), and/or the total number of segmented voxels M, and/or the total number of segmented non-foreground area voxels L;

The discriminant parameter measurement module measures the discriminant parameter β in turn according to the order of the threshold μ from small to large;

Threshold judgment module, for judging the numerical relationship between the discriminant parameter β and the discriminant parameter threshold ε, if the discriminant parameter β > the discriminant parameter threshold ε, take the threshold μ ₀ corresponding to the current β, and output μ ₀ - ο, as the best threshold; If the discriminant parameter β≤the discriminant parameter threshold ε, the current threshold is increased by the segmentation threshold interval (μ ₀ ← μ ₀ +ο) as a new search threshold, and the voxel quantity measurement module and the discriminant parameter measurement module are activated to perform the next threshold operation again. .

Further, when the upper bound of the segmentation threshold can be determined, the search range of the threshold μ can be predicted. The three-dimensional connectivity-based image threshold device can be adjusted accordingly, and the device includes:

The import module is used to import all the two-dimensional images collected at one time to obtain a two-dimensional image group, and all the two-dimensional images can obtain a three-dimensional image for the target in the image through three-dimensional reconstruction;

The parameter setting module is used to input the numerical value or calculation method of the setting parameter, including assigning values for the segmentation threshold interval o, the discriminant parameter threshold ε, the threshold search lower bound μ ₀ , and the threshold search upper bound μ ₁ , and the selection of the discriminant parameter β. Calculation;

The voxel number determination module is used to calculate the total number N of voxels in the three-dimensional connected domain where the number of voxels in the foreground area is greater than or equal to the set threshold υ under each threshold μ between μ ₀ and μ ₁ (referred to as the total number of voxels in the foreground area), and/or the total number of voxels M, and/or the total number of voxels in the non-foreground area, L, and the above parameters are added to the queue list _N , and/or list _M , and/or list _L ;

The discriminant parameter measurement module calculates the discriminant parameter β under each threshold and joins the queue list _β ;

The threshold judgment module, in the order of the threshold μ from small to large, takes the corresponding β value in the queue list _β for judgment, until the discriminant parameter β> the discriminant parameter threshold ε, takes the threshold value μ ₀ corresponding to the current β, and outputs μ ₀ -ο, as the optimal threshold.

In the above device, the setting of the discriminant parameter β is consistent with the setting of the discriminant parameter β in the corresponding method.

Preferably, the device further includes a conversion module for converting the RGB color image into a grayscale image.

The above-mentioned apparatus in the present invention corresponds to a technical solution that can be used to execute the above-mentioned method, and the implementation principle and technical effect thereof are similar, and will not be repeated here.

Those skilled in the art can understand that all or part of the steps of implementing the above method can be completed by program instructions related to hardware. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps including the above method are executed; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other media that can store program codes.

Example

Test environment: use a desktop computer, including an Intel i9 chip, 64GB memory. Using the window10 operating system, based on the C++ environment, install the ITK package.

Example 1

Perform lung segmentation on lung CT images with large areas of high-density shadows, and evaluate the difference between the threshold determination method of the present invention and the results of the traditional threshold method; the method flow is shown in FIG. 3 .

(1) Input a group of lung CT images with large areas of high-density shadows, one of which is shown in Figure 4, with a significant high-density shadow area, all CT images in the image group are scanned in CT according to their corresponding original images (2) Set the segmentation threshold interval ο (take 10 by default in this example), the discriminant parameter threshold ε (take 100 in this example); (3) select a value μ ₀ smaller than the segmentation threshold as Threshold search initial value. According to clinical experience, the window level of human lung tissue is between -450 and -600. Therefore, -600 can be selected as the initial search value; (4) Calculate the threshold (μ ₀ - 2*ο) the number of voxels is greater than or equal to the three-dimensional connected domain of the set threshold υ and calculate their total number of voxels N ₋₂ ; (5) calculate the number of voxels under the threshold (μ ₀ −ο) in all CT images is greater than or equal to the set Set the three-dimensional connected domain of the threshold υ and calculate their total number of voxels N ₋₁ ; (5) Calculate the three-dimensional connected domain of all CT images with the number of voxels greater than or equal to the set threshold υ under the threshold μ ₀ and calculate their total voxel number. Number N ₀ ; (6) Calculate the discriminant parameter β=(N ₀ -N _-1 )/(N _-1 -N _-2 ); (7) If β>ε, skip to (9), otherwise continue ( 8); (8) N _-2 ←N _-1 , N _-1 ←N ₀ , μ ₀ ←μ ₀ +ο, return to (5); (9) Take the threshold μ ₀ corresponding to β, and output μ ₀ -o . In this group of lung CT images, if μ ₀ =-270, input -280. The binarization result when μ ₀ =-280 corresponding to FIG. 4 is shown in FIG. 8 . Obviously, in Figure 8, a lot of pixels have been added to the high-density shadow area, including the previously hidden lung tissue pixels and some noise data added due to the increased threshold. Due to the appearance of these pixels, the outline of the high-density shadow area is clearer.

Using the manual experience selection method, according to clinical experience, the CT value window width of lung tissue is about 1500 to 2000, and the window level is about -450 to -600. We choose -350, which is higher than the maximum window level, as the threshold. The result of the transformation is shown in Figure 5. Using the maximum inter-class variance method, the calculated threshold value is μ ₀ =-400, and the binarization result is shown in FIG. 6 . Using the iterative method, the calculated threshold value is μ ₀ =-400, and the binarization result is shown in FIG. 7 . Obviously, due to the low value, there are fewer pixels in the high-density shadow area, and the contour is not complete.

From the lung segmentation results obtained by the three-dimensional connectivity-based threshold determination method, manual empirical selection method, maximum inter-class variance method, and iterative method in the present invention, it can be known that the three-dimensional connectivity-based threshold determination method can be used in lung CT in areas with high density shadows. The results obtained in image segmentation are obviously better than those obtained by other methods. The manual experience selection method relies on prior knowledge and cannot automatically adjust the threshold according to the actual situation of the image; the maximum inter-class variance method and the iterative method both rely on The pixel characteristics of the image, when there is a large contrast area of pixel value in the image, such as a large area of high-density shadow area in the lung CT image, the CT value of this part is higher than that of normal lung tissue. This method will divide the contrast area, that is, the high-density shadow area, and cause a large segmentation error. Compared with these two methods, the threshold determination method based on 3D connectivity can retain meaningful pixels in the high-density shadow area, making the contour more complete, which is conducive to later image analysis and processing, although it will also add a small amount of noise to the edge of the contour. , but its interference can be reduced by post-processing such as denoising, which is worthwhile compared to the local loss caused by missing pixels.

The present invention has been described above with reference to the preferred embodiments, but these embodiments are merely exemplary and serve only for illustrative purposes. On this basis, various substitutions and improvements can be made to the present invention, which all fall within the protection scope of the present invention.

Claims (10)

1. An image thresholding method based on three-dimensional connectivity, comprising:

s100, importing all the two-dimensional images acquired at one time to obtain a two-dimensional image group, and obtaining a three-dimensional image aiming at a target in the image through three-dimensional reconstruction of all the two-dimensional images;

s200, setting a partition threshold interval omicron, and judging a parameter threshold;

s300, selecting a threshold value to search an initial value mu₀The threshold value is searched for an initial value mu₀Less than or equal to a segmentation threshold;

s400, respectively calculating threshold values mu₀-2 o, threshold μ₀-and a threshold μ -₀The total number of voxels N in the next segmented foreground region, and/or the total number of voxels M in the segmented foreground region, and/or the total number of voxels L in the segmented non-foreground region;

s500, setting a judgment parameter β based on the parameters measured in S400, wherein the judgment parameter β is used for measuring whether the total number N of the voxels in the segmented foreground region is increased steeply;

s600, if the discrimination parameter β is larger than the discrimination parameter threshold value, jumping to S800, otherwise, continuing to S700;

S700，μ₀←μ₀+ o, return to S400;

s800, taking the threshold value mu corresponding to the current β₀And output mu₀-as optimal threshold.

2. The method of claim 1, wherein the two-dimensional image is a grayscale image, and the foreground region pixel values are lower than the background region pixel values in the image;

if the two-dimensional image is an RGB color image, converting the RGB color image into a gray image;

if the pixel value of the foreground area in the image is higher than the pixel value of the background area, the maximum value of the pixel value of the current image is taken, the current value of each pixel is subtracted from the maximum value to serve as the new value of the pixel, and therefore a new image is constructed.

3. The method of claim 1, wherein the criterion β is a differential neighbor ratio of the total number N of voxels in the segmented foreground region, or

The criterion β is the difference neighbor ratio of the total number M of divided voxels, or

The discrimination parameter β is a differential neighborhood ratio of the total number L of voxels in the non-foreground region that have been segmented.

4. The method of claim 1, wherein the criterion β is a difference neighbor ratio between the absolute value of the ratio of the total number of segmented foreground region voxels N to the total number of segmented voxels M, or

The discrimination parameter β is a difference neighborhood ratio of absolute values of ratios of the total number M of the divided voxels to the total number N of the divided voxels in the foreground region.

5. The method according to claim 1, wherein the discriminant parameter β is a difference neighborhood ratio of the absolute value of the ratio of the total number of voxels N in the segmented foreground region to the total number of voxels L in the segmented non-foreground region, or

The discrimination parameter β is a difference neighbor ratio of absolute values of ratios of the total number of non-foreground region voxels L and the total number N of foreground region voxels.

6. The method of claim 1, wherein the discriminant parameter β is a difference neighbor ratio of the absolute value of the ratio of the total number of voxels L in the segmented non-foreground region to the total number of voxels M, or

The discrimination parameter β is a difference neighborhood ratio of the absolute value of the ratio of the total number M of the divided voxels to the total number L of the divided voxels in the non-foreground region.

7. Method according to one of claims 1 to 6, characterized in that, when the search range of the threshold μ can be predicted, the method can be implemented by:

s100, importing all the two-dimensional images acquired at one time to obtain a two-dimensional image group, and obtaining a three-dimensional image aiming at a target in the image through three-dimensional reconstruction of all the two-dimensional images;

s200, setting a partition threshold interval omicron, and judging a parameter threshold;

s300, selecting a value smaller than the segmentation threshold as a threshold to search a lower bound mu₀(ii) a Selecting a value greater than the segmentation threshold as the upper threshold search μ₁；

S400, calculating mu at intervals of o₀To mu₁The total number of voxels N in the segmented foreground region and/or the total number of voxels M in the segmented foreground region and/or the total number of voxels L in the segmented non-foreground region under each threshold value mu are added into the queue list_NAnd/or list_MAnd/or list_L；

S500, setting a discrimination parameter β based on the parameters measured in S400, wherein the discrimination parameter β is used for measuring whether the total number N of the voxels in the divided foreground region is increased steeply, calculating the discrimination parameters β under each threshold value and adding the discrimination parameters into a queue list_β；

S600, according to the threshold value mu, from small to smallBig order, get it in queue list in turn_βThe β value corresponding to (1) is judged until reaching β>；

S700, taking the threshold value mu corresponding to the current β and outputting mu-o as the optimal threshold value.

8. An image thresholding device based on three-dimensional connectivity for implementing the method of one of the preceding claims 1 to 6, characterized in that it comprises:

the import module is used for importing all the two-dimensional images acquired at one time to obtain a two-dimensional image group, and three-dimensional images aiming at targets in the images can be obtained through three-dimensional reconstruction of all the two-dimensional images;

a parameter setting module for inputting the value or calculation mode of the set parameter, including dividing threshold value interval omicron, judging parameter threshold value and threshold value search initial value mu₀Assigning values and selecting a calculation mode of the discrimination parameter β;

a voxel quantity determination module for determining the threshold value mu₀-2 o, threshold μ₀-and μ -₀The total number of voxels N in the next segmented foreground region, and/or the total number of voxels M in the segmented foreground region, and/or the total number of voxels L in the segmented non-foreground region;

a discrimination parameter measuring module for sequentially measuring discrimination parameters β according to the order of the threshold value mu from small to large;

a threshold judging module for judging the value relationship between the discrimination parameter β and the discrimination parameter threshold, if the discrimination parameter β>Judging parameter threshold value, and taking current threshold value mu corresponding to β₀And output mu₀And if the discrimination parameter β is less than or equal to the discrimination parameter threshold, the current threshold is increased by a segmentation threshold interval and then is used as a new search threshold, and the voxel number determination module and the discrimination parameter determination module are started to perform next threshold operation again.

9. An image thresholding device based on three-dimensional connectivity for implementing the method of claim 7, wherein the device comprises:

the import module is used for importing all the two-dimensional images acquired at one time to obtain a two-dimensional image group, and three-dimensional images aiming at targets in the images can be obtained through three-dimensional reconstruction of all the two-dimensional images;

a parameter setting module for inputting the value or calculation mode of the set parameter, including dividing threshold interval omicron, judging parameter threshold, and searching lower limit mu of threshold₀And threshold search upper bound mu₁Assigning values and selecting a calculation mode of the discrimination parameter β;

a voxel number determination module for calculating mu at intervals of omicron₀To mu₁The total number of voxels N in the segmented foreground region and/or the total number of voxels M in the segmented foreground region and/or the total number of voxels L in the segmented non-foreground region under each threshold value mu are added into the queue list_NAnd/or list_MAnd/or list_L；

A discrimination parameter measuring module for calculating discrimination parameters β under each threshold value and adding the discrimination parameters into the queue list_β；

A threshold value judging module, which takes the threshold values mu in the queue list in turn according to the sequence from small to large of the threshold values mu_βUntil the corresponding β value reaches the judgment parameter β>Judging parameter threshold value, and taking current threshold value mu corresponding to β₀And output mu₀-as optimal threshold.

10. The apparatus of claim 9, further comprising a conversion module for converting the RGB color image into a grayscale image.

Images