CN104021539A - System used for automatically detecting tumour in ultrasonic image - Google Patents

Abstract

A system for automatic detection of tumors in ultrasound images is provided. The system includes: a candidate region detection device for detecting at least one candidate region including a tumor from an ultrasound image based on a deformable part model (DPM); a tumor region localization device for determining from the detected at least one candidate region Tumor area; tumor contour separation means for detecting a tumor by separating the contour of the tumor based on the determined tumor area. Through the system, the area including the tumor can be effectively detected from the ultrasound image without human participation, and on this basis, the outline of the tumor can be relatively accurately separated as one of the basis for the diagnosis of the tumor .

Description

System for automatic detection of tumors in ultrasound images

technical field

The invention relates to medical image processing technology, in particular to a system for detecting tumor parts from ultrasonic images.

Background technique

In modern medicine, ultrasound imaging is an important means for diagnosing various tumors (for example, thoracic tumors such as breast tumors), because ultrasound examination is relatively convenient, non-invasive and low-cost. However, doctors in the imaging department need to manually label each ultrasound image to reflect the imaging characteristics of the corresponding organ as the imaging basis for judging tumors. However, it takes a lot of time to manually label the ultrasound examination results of a large number of patients one by one, and the labor cost is high.

Therefore, attention has been paid to schemes for automatic detection of tumors from ultrasound images. For example, Drukker, K. et al. proposed in "Computerized lesion detection on breast ultrasound", Med. Phys., 29(7):1438-46 (2002) that the lesion area is almost darker than the background area to varying degrees, and based on this A strict assumption is used to generate tumor regions.

However, in the existing automatic tumor detection schemes such as those described above, due to the poor image quality of the ultrasonic detection results, and the very complex shape of the tumor, not only irregular patterns, but also often accompanied by calcification, it is difficult to effectively The area where the tumor is located can be accurately detected in the ultrasound image.

In addition, since whether the relevant area is a tumor is also closely related to the specific tissue where it is located, misjudgment is also likely to occur in the existing tumor detection technology.

Contents of the invention

It is an object of the present invention to provide a system capable of efficiently and automatically detecting tumors from ultrasound images.

According to an aspect of the present invention, there is provided a system for automatically detecting a tumor in an ultrasound image, comprising: candidate region detection means for detecting at least one candidate region including a tumor from an ultrasound image based on a deformable part model (DPM) tumor region localization means for determining a tumor region from the detected at least one candidate region; tumor contour separation means for detecting a tumor by separating the contour of the tumor based on the determined tumor region.

The candidate region detecting means may detect at least one candidate region including the tumor from the ultrasound image by changing an aspect ratio of the root template in the DMP.

The tumor region locating device may use a support vector machine (SVM)-based binary classifier to determine the tumor region from the at least one detected candidate region.

The feature vector of the binary classifier may be based on the context features of the candidate regions.

The feature vector may include at least one of the following items: the DPM detection score of the candidate area, the position and size of the candidate area, the offset of each component in the candidate area relative to the root, the distance between the foreground and the background in the candidate area The intensity difference, the coexistence part between the candidate region and the candidate region with the highest DPM detection score.

The tumor region localization device can use a multi-kernel learning (MKL) method to define the kernel function of the binary classifier as a linear combination of multiple basic kernel functions.

The tumor area localization device can linearly combine the RBF kernel function of three bandwidths and the polynomial kernel function of three dimensions, so as to train each feature component in the feature vector and the feature vector as a whole to obtain the kernel of the binary classifier function.

The tumor contour separation device may use a level set method to separate tumor contours, wherein the tumor contour curve is iterated with the boundary of the tumor region as an initial curve.

The tumor contour separation device can construct the energy function used in the level set method by maximizing the distance between the foreground and the background of the tumor image.

According to another aspect of the present invention, there is provided a system for automatically detecting an object in an ultrasound image, comprising: candidate region detection means for detecting at least one candidate including an object from an ultrasound image based on a deformable part model (DPM) an area; object area locating means for determining an object area from the detected at least one candidate area; object outline separating means for detecting an object by separating an outline of the object based on the determined object area.

The candidate region detecting means may detect at least one candidate region including the object from the ultrasound image by changing an aspect ratio of the root template in the DMP.

The object region locating device may determine the object region from the detected at least one candidate region using a support vector machine (SVM)-based binary classifier.

The feature vector of the binary classifier may be based on the context features of the candidate regions.

The object region localization device may use a multi-kernel learning (MKL) method to define the kernel function of the binary classifier as a linear combination of multiple basic kernel functions.

According to another aspect of the present invention, there is provided a method for automatically detecting an object in an ultrasound image, comprising: detecting at least one candidate region including an object from the ultrasound image based on a deformable part model (DPM); The at least one candidate area determines an object area; and the object is detected by separating a contour of the object based on the determined object area.

According to the above exemplary embodiments, the region including the tumor can be effectively detected from the ultrasound image without human participation, and on this basis, the outline of the tumor can be relatively accurately separated as the basis for the diagnosis of the tumor one.

Description of drawings

The above and other objects and features of the present invention will become more clear through the following detailed description of exemplary embodiments in conjunction with the accompanying drawings, wherein:

1 shows a block diagram of a tumor detection system according to an exemplary embodiment of the present invention;

Fig. 2 shows the processing flow of detecting a tumor by a tumor detection system according to an exemplary embodiment of the present invention;

Fig. 3 shows an example of a tumor candidate region detected according to an exemplary embodiment of the present invention;

Figure 4 shows an example of a tumor region determined according to an exemplary embodiment of the present invention;

Figure 5 shows an example of a tumor contour isolated according to an exemplary embodiment of the present invention;

Figure 6 shows a block diagram of an object detection system according to an exemplary embodiment of the present invention; and

FIG. 7 shows a process flow of detecting an object by an object detection system according to an exemplary embodiment of the present invention.

Detailed ways

Reference will now be made in detail to embodiments of the invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to like parts throughout. The embodiments are described below in order to explain the present invention by referring to the figures.

FIG. 1 shows a block diagram of a tumor detection system according to an exemplary embodiment of the present invention. As shown in FIG. 1 , a tumor detection system according to an exemplary embodiment of the present invention includes: a candidate region detection device 10 for detecting at least one candidate region including a tumor from an ultrasound image based on a deformable part model (DPM); tumor region localization A means 20 for determining a tumor area from the detected at least one candidate area; a tumor contour separating means 30 for detecting a tumor by separating the contour of the tumor based on the determined tumor area. Here, as an example, an ultrasound image may indicate a chest ultrasound examination result for a breast tumor, however, it should be noted that the present invention is not limited thereto, and the tumor detection system according to an exemplary embodiment of the present invention may be applied to tumors in other organs detection.

In the above tumor detection system, the candidate area detection device 10 can effectively detect the candidate area where tumors may appear by using the deformable part model (DPM) method, and on this basis, the tumor area positioning device 20 further determines the tumor area , and the tumor contour is extracted by the tumor contour separation device 30, so that the automatic detection of the tumor in the ultrasound image can be effectively completed.

Hereinafter, an example of tumor detection according to an exemplary embodiment of the present invention will be described with reference to FIG. 2 .

FIG. 2 shows a process flow of detecting a tumor by a tumor detection system according to an exemplary embodiment of the present invention.

Referring to FIG. 2 , at step S100 , at least one candidate region including a tumor is detected from an ultrasound image by the candidate region detecting device 10 based on a deformable part model (DPM).

Specifically, as known to those skilled in the art, in the DPM method, there is a root filter with relatively coarse precision and multiple component filters with relatively fine precision, correspondingly, there is a root and multiple components, wherein, p ₀ indicates the root intended to approximately cover the entire object to be detected (for example, the entire region where the tumor is located), while p ₁ , p ₂ , . . . , p _n indicate individual different Components, correspondingly, the candidate region detection device 10 can calculate the DPM detection score of each candidate region through equation (1):

$\mathrm{<span\; class="notranslate">\; score</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; score</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; dx</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; dy</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

From equation (1), it can be seen that the DPM detection score score(p ₀ ,p ₁ ,...,p _n ) of the candidate region can be mainly expressed as the appearance score of the root filter/component filter at their respective positions (where F _i (p _i ) is the filter response of the root and each component, G _i (p _i ) is the characteristic of the root and each component) and the deformation term (used to represent the position of each component away from its anchor position deviation cost) (where x _i and y _i indicate the pixel positions of the respective components), on the basis of which a real-valued bias term b (where b can be determined experimentally) is added to the difference to obtain the final DPM detection score.

As a preferred mode, the above-mentioned DPM method can be improved so as to be more in line with the growth characteristics of the tumor itself. That is to say, since the tumor itself has various shapes and wide variations in aspect ratio, and the number of root templates corresponding to the root filter in DPM is small (usually 2 or 3), it is easy to fail to detect different width-to-height ratios. High ratio of tumors. In view of the above, the candidate region detection apparatus 10 according to an exemplary embodiment of the present invention may detect at least one candidate region including a tumor from an ultrasound image by changing an aspect ratio of a root template in a DMP.

Specifically, when calculating the dot product of F ₀ ( _{p 0 )} and G ₀ (p ₀ ) for the root p 0 , the dot product result can be determined as the maximum value:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; -</span>{{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; \&prime;</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; \&tau;</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

It can be seen from equation (2) that by determining the dot product of F ₀ (p ₀ ) and G ₀ (p ₀ ) to be within a predetermined deflection range (where τ indicates the deflection range, different The application determines the maximum dot product value of τ), so that the root template in DMP is no longer limited to a fixed value, but can be adjusted to a certain extent within a predetermined range, which helps to target various widths and heights Effective detection of specific tumors.

As a preferred mode, after obtaining each candidate region with a higher DMP detection score as described above, the obtained candidate region can be further screened to obtain a more reliable detection result, for example, the DPM can be removed A specific candidate region with a higher detection score overlaps less than 50% of the candidate regions with a lower detection score by DPM, thereby achieving effective screening of the candidate region.

FIG. 3 shows an example of tumor candidate regions detected according to an exemplary embodiment of the present invention. As shown in FIG. 3 , relative to the actual tumor and its location, several candidate tumor regions can be detected, wherein each rectangular region corresponds to a candidate region detected in step S100 .

Then, in step S200, the tumor region is determined by the tumor region locating device 20 from at least one detected candidate region.

Specifically, the tumor region locating device 20 may use a support vector machine (SVM)-based binary classifier to determine the tumor region from at least one detected candidate region.

Various suitable ways can be used to obtain the feature vector for the binary classifier, for example, the feature can be extracted from the position and size of the root template/part template output corresponding to the root filter/part filter vector.

In addition, in order to further improve the accuracy of tumor region localization, the feature vector of the support vector machine (SVM)-based binary classifier can be based on the context features of the candidate region. For example, the following equation (3) can represent the feature vector f of the binary classifier:

f=(s,r,offset _i ,I,S _MAX ,r _MAX ) (3)

The above feature vector f involves the following items: the DPM detection score s of the candidate area, the vector r representing the position and size of the candidate area, the offset _i of each component in the candidate area relative to the root, the distance between the foreground and the background in the candidate area The intensity difference I between them, the coexistence part (S _MAX , r _MAX ) between the candidate region and the candidate region with the highest DPM detection score. Through the above feature vector f according to the exemplary embodiment of the present invention, the context features of the candidate region can be used for classification, so that not only the position and size of the candidate region, but also the surrounding tissues where the candidate region is located are considered. In addition, since there are usually no more than three tumors in the same ultrasound image, it is helpful to further strengthen the accuracy of classification by considering the coexistence between the current candidate region and the candidate region with the highest DPM detection score. sex.

However, those skilled in the art should understand that: the feature vector f is not limited to the items listed above, for example, the feature vector f may only include at least one or more of the items listed above, and does not necessarily include All terms listed in equation (3). In addition, any correlation vector that embodies the contextual features of the candidate region can be applied to the feature vector f.

Based on the feature vector f determined above, the tumor region localization device 20 may use a multi-kernel learning (MKL) method to define the kernel function of the binary classifier as a linear combination of multiple basic kernel functions. Specifically, in the multi-kernel learning MKL method, the parameters of the multiple basic kernel functions and the weights of each basic kernel function can be learned through training data. For example, the tumor region locating device 20 can linearly combine the RBF (radial basis) kernel function of three bandwidths and the polynomial kernel function of three dimensions, so as to perform Train to obtain the kernel function of the binary classifier. Here, the specific bandwidth of the RBF kernel function can be determined according to experiments.

After determining the feature vector and kernel function of the binary classifier, the tumor region localization device 20 can use the corresponding binary classifier to determine the tumor region from at least one detected candidate region, that is, the probability of tumor occurrence in the candidate region the largest candidate region.

Fig. 4 shows an example of a tumor region determined according to an exemplary embodiment of the present invention. As shown in FIG. 4 , the irregular figure included in the upper left rectangular area indicates the actual tumor, while the lower right rectangular area is the tumor area determined by the tumor area locating device 20 in step S200 .

Then, at step S300, a tumor is detected by the tumor contour separating device 30 by separating the contour of the tumor based on the determined tumor region.

Specifically, by employing the level set method, the contour curve of a tumor can be expressed as a zero level set of a higher dimensional function called a level set function. Here, the level set function can be made to evolve or iterate according to the development equation it satisfies (wherein, the boundary of the tumor region determined in step S200 is used as the initial curve for iteration). Since the level set function is constantly evolving, the corresponding zero The level set is also constantly changing, and when the evolution of the level set tends to be stable, the evolution stops, and the final contour curve of the tumor is obtained.

According to an exemplary embodiment of the present invention, the contour curve of the tumor can be obtained by minimizing the energy function represented by equation (4):

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{\mathrm{<span\; class="notranslate">\; inside</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&Integral;</span>}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; wxya</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; outside</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; )</span>}{<span\; class="notranslate">\; \&Integral;</span>}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; wxya</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&mu;</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; Length</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; area</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; inside</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; C</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

In Equation (4), C indicates the contour curve of the tumor, where the contour curve C of the tumor is iterated with the boundary of the tumor area as the initial curve, c ₁ indicates the inner area of the contour curve C of the tumor in the t-th iteration The average intensity of the pixels, _c2 indicates the average intensity of the pixels in the outer region of the tumor's contour curve C in the t-th iteration, inside(C) indicates the inner region of the tumor's contour curve C, and outside(C) indicates the tumor's contour curve The outer area of C, u ₀ (X,t) indicates the intensity of the pixel at the X position in the t-th iteration, Length(C) indicates the length of the contour curve C of the tumor, and Area(inside(C)) indicates the contour of the tumor The area of the inner region of the curve C, in addition, λ ₁ , λ ₂ , μ, v are fixed parameters greater than or equal to zero. From equation (4), it can be seen that the energy function adopted in the level set method can be expressed as the sum of the following items: the pixel intensity difference in the inner region of the contour curve C of the tumor Pixel intensity differences in the outer regions of the tumor's contour curve C The length factor μ·Length(C) of the contour curve C of the tumor, and the area factor v·Area(inside(C)) of the inner area of the contour curve C of the tumor.

In the above equation (4), the energy function F(c ₁ ,c ₂ ,C) mainly reflects the internal and external global image statistics and characteristics of the iterative contour curve. As an example, the parameters λ ₁ , λ ₂ can be set to 1 respectively, and the parameter v can be set to 0.

On this basis, in order to further improve the accuracy of separating tumors, according to the characteristics of tumor images, for example, the intensity of the image is often uneven due to calcification in the middle of the tumor, the tumor contour separation device 30 can preferably make the foreground of the tumor image and The distance between the backgrounds is maximized to construct the energy function used in the level set method.

As an example, the tumor contour separation device 30 can maximize the distance between the foreground and the background of the tumor image by adding a regular term according to the Bhattacharyya distance to the energy function of equation (4). , the corresponding energy function is optimized as shown in the following equation (5):

E(C)=βF(c ₁ ,c ₂ ,C)+(1-β)B(C) (5)

In equation (5), Represents the Bhattacharya distance between the probability density function p _in (X) of the contour curve C and the probability density function p _out (X), where p _in (X) means that the pixel at position X is inside the contour curve C (ie, located in the foreground of the tumor image), p _out (X) represents the probability density function of the pixel located at position X outside the contour curve C (ie, located in the background of the tumor image). Additionally, β∈[0,1] is used to control the participation of the Bhattacharya term B(C).

On this basis, in order to satisfy the properties of the signed distance function φ And to avoid reinitialization, we can consider adding an additional penalty function R _p (φ) to the energy function of Equation (5), so as to obtain the energy function shown in Equation (6):

E(C)=βF(c ₁ ,c ₂ ,C)+(1-β)B(C)+αR _p (φ) (6)

In equation (6), α>0, and the penalty function Among them, Ω represents the whole image.

As a preferred manner, the above-mentioned processing can be performed within a narrow band range, so as to increase the speed of iterative processing. In the manner described above, the tumor contour separating device 30 can separate the contour of the tumor as the detected tumor.

Fig. 5 shows an example of a tumor contour isolated according to an exemplary embodiment of the present invention. As shown in FIG. 5 , the tumor outline separated in step S300 has a higher degree of approximation than the actual tumor.

It should be noted that, in addition to the detection system shown in FIG. 1 , the method shown in FIG. 2 can also be implemented by computer programming, and can also be executed by a dedicated processor composed of specific hardware. Those skilled in the art may use any means known or available in the art to implement the method shown in FIG. 2 .

By designing a special experiment (for example, for 2758 ultrasound images from 1941 patients (among them, 913 patients with benign tumors and 1028 patients with malignant tumors), it can be seen that: Compared with other methods, the tumor region localization method according to the exemplary embodiment of the present invention not only significantly increases the hit rate of the tumor, but also significantly reduces the omission or positioning error, while the tumor separation method according to the exemplary embodiment of the present invention can obtain a higher than Existing methods have better accuracy.

It can be seen that according to the above-described automatic tumor detection system, the area including the tumor can be effectively detected from the ultrasound image without human participation, and the outline of the tumor can be relatively accurately separated on this basis , as one of the basis for tumor diagnosis.

The automatic detection system and method according to the exemplary embodiments of the present invention have been described above for tumor detection, however, it should be understood that the above exemplary embodiments can also be applied to other object detection in ultrasound images, specifically, ultrasound images as a This non-destructive testing method can be used in many fields, such as, it can be used to test castings, forgings, plates, composite materials, pipes, bars, profiles, Inspection of welded parts, machined parts and the above-mentioned workpieces in use. Therefore, the above-described automatic tumor detection system and method can be applied to the detection of other objects in the ultrasound image without limiting the detected object to a tumor.

An object detection system according to an exemplary embodiment of the present invention and a process flow of detecting an object by the object detection system will be described below with reference to FIGS. 6 and 7 .

FIG. 6 shows a block diagram of an object detection system according to an exemplary embodiment of the present invention. As shown in FIG. 6 , an object detection system according to an exemplary embodiment of the present invention includes: a candidate region detection device 101 for detecting at least one candidate region including an object from an ultrasound image based on a deformable part model (DPM); object region localization The means 201 is configured to determine an object area from the detected at least one candidate area; the object contour separating means 301 is configured to detect an object by separating the contour of the object based on the determined object area.

FIG. 7 shows a process flow of detecting an object by an object detection system according to an exemplary embodiment of the present invention.

Referring to FIG. 7 , at step S101 , at least one candidate region including an object is detected from an ultrasound image by the candidate region detecting device 101 based on a deformable part model (DPM).

Then, in operation S201 , an object area is determined from at least one detected candidate area by the object area locating device 201 .

Finally, in operation S301, an object is detected by the object contour separating means 301 by separating the contour of the object based on the determined object region.

It should be noted that each device in the detection system shown in Figure 6 adopts a method similar to that shown in Figures 1 and 2 when performing the operation process shown in Figure 7, the only difference is that the various parameters and processes related to the tumor are changed for arguments and handling for arbitrary objects.

While the invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those skilled in the art that changes may be made in form and detail without departing from the spirit and scope of the invention as defined by the claims. various changes.

Claims (15)

1. A system for automatically detecting tumors in ultrasound images, comprising: Candidate region detection means for detecting at least one candidate region including a tumor from an ultrasound image based on a deformable part model (DPM); a tumor region locating device, configured to determine a tumor region from the detected at least one candidate region; A tumor contour separating device for detecting a tumor by separating the contour of the tumor based on the determined tumor area. 2. The system of claim 1, wherein the candidate region detecting means detects at least one candidate region including the tumor from the ultrasound image by changing an aspect ratio of the root template in the DMP. 3. The system according to claim 1, wherein the tumor region locating device determines the tumor region from the at least one detected candidate region using a support vector machine (SVM)-based binary classifier. 4. The system of claim 3, wherein the feature vectors of the binary classifier are based on contextual features of candidate regions. 5. The system according to claim 4, wherein the feature vector comprises at least one of the following items: the DPM detection score of the candidate area, the position and size of the candidate area, the deviation of each component in the candidate area relative to the root The amount of displacement, the intensity difference between the foreground and background in the candidate area, the coexistence part between the candidate area and the candidate area with the highest DPM detection score. 6. The system according to claim 5, wherein the tumor region locating device uses a multi-kernel learning (MKL) method to define the kernel function of the binary classifier as a linear combination of multiple basic kernel functions. 7. The system according to claim 6, wherein the tumor region localization device linearly combines the RBF kernel function of three bandwidths and the polynomial kernel function of three dimensions, so that for each eigencomponent and eigenvector in the eigenvector The whole is trained to obtain the kernel function of the binary classifier. 8. The system according to claim 1, wherein the tumor contour separation device uses a level set method to separate the contour of the tumor, wherein the contour curve of the tumor is iterated with the boundary of the tumor area as an initial curve. 9. The system according to claim 8, wherein the tumor contour separating means constructs the energy function used in the level set method by maximizing the distance between the foreground and the background of the tumor image. 10. A system for automatically detecting objects in ultrasound images, comprising: Candidate region detection means for detecting at least one candidate region including an object from an ultrasound image based on a deformable part model (DPM); Object area locating means for determining an object area from the detected at least one candidate area; Object outline separating means for detecting an object by separating an outline of the object based on the determined object area. 11. The system of claim 10, wherein the candidate region detecting means detects at least one candidate region including the object from the ultrasound image by changing an aspect ratio of the root template in the DMP. 12. The system of claim 10, wherein the object region locating means determines the object region from the detected at least one candidate region using a support vector machine (SVM) based binary classifier. 13. The system of claim 12, wherein the feature vectors of the binary classifier are based on contextual features of candidate regions. 14. The system of claim 13, wherein the object region locating device uses a multi-kernel learning (MKL) method to define the kernel function of the binary classifier as a linear combination of a plurality of basic kernel functions. 15. A method for automatically detecting an object in an ultrasound image, comprising: detecting at least one candidate region comprising an object from the ultrasound image based on a deformable part model (DPM); determining an object region from the detected at least one candidate region; Objects are detected by separating contours of the objects based on the determined object regions.