JP2010200925A - Image processing device, method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve performance in detecting object tissues comprising linear structure on a three-dimensional image. <P>SOLUTION: The main axis direction of a candidate object tissue comprising linear structure or the normal direction of a tissue comprising surface structure in a detection range is determined. Based on the main axis direction or the normal direction, a normalization process is performed to a candidate object range including the candidate object tissue. Feature quantities of the candidate object range after the normalization process is performed are computed. Whether or not the object tissue is included in the candidate object range is determined using the computed feature quantities. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to three-dimensional image detection processing, and more particularly to an image processing apparatus and method suitable for automatic detection of a target tissue having a linear structure or a planar structure in a three-dimensional image, and a program therefor.

  Conventionally, in the medical field, it is possible to construct a three-dimensional image by detecting a target tissue by applying a predetermined detection process (for example, a detection method using machine learning) to a plurality of two-dimensional images. A pseudo three-dimensional image representing a target tissue is generated by causing a computer to execute an image projection method (Intensity Projection method) or a volume rendering method. Thereby, the radiogram interpreter can easily check the status of the target organization.

  However, since a three-dimensional image has a large amount of information, a feature amount extracted from the three-dimensional image itself increases, and for example, calculation processing of a detection method by machine learning increases, which is a problem.

  Therefore, Patent Document 1 proposes a method of generating three cross-sectional images orthogonal to each other from a three-dimensional image and performing a target tissue detection process based on feature amounts extracted from the three cross-sectional images.

  Further, Patent Document 2 proposes a method of detecting a target tissue from a two-dimensional image by machine learning.

U.S. Pat. No. 7,346,209 JP 2007-307358 A

  The invention described in Patent Document 1 sets three cross-sectional images orthogonal to each other from a target tissue as detection target images. For example, when this method is applied to a linear structure, FIG. ) Three cross-sectional images as shown in (b) are obtained. According to the invention described in Patent Document 1, a cross-sectional image is set by randomly changing three orthogonal axes, and a plurality of detection processes are performed, thereby reducing the total amount of information used for calculation. Detect target tissue.

  However, in the case of a curved linear tissue as shown in FIG. 3B, there is a problem that information on the target tissue is largely lacking in the cross-sectional image.

  In addition, there is a method of using a Hessian matrix as a method for extracting a linear structure / planar structure, which is used for extraction of blood vessels and the like. Even if the blood vessel can be discriminated, it is difficult to correctly discriminate when the blood vessel is curved, branched or accompanied by a lesion. Similarly, since the planar structure in the human body does not always form an ideal plane, it has been difficult to determine correctly.

 Therefore, in view of the above circumstances, the present invention provides an image processing apparatus capable of improving detection performance even when a linear tissue or a planar tissue is curved, branched, or accompanied by a lesion, and It is an object to provide a method and a program therefor.

  An image processing apparatus according to the present invention includes a detection region setting unit that sets a detection region in a three-dimensional image obtained by capturing an image of a subject, and a main axis direction or planar structure of a target candidate tissue including a linear structure in the detection region. A main axis / normal direction calculating unit that calculates a normal direction of the tissue, and a normalization processing unit that performs normalization processing on the target candidate region including the target candidate tissue based on the main axis direction or the normal direction. And a determination unit that calculates a feature amount of the target candidate region that has been subjected to normalization processing, and determines whether the target candidate region includes the target tissue using the calculated feature amount. And The “image processing device” may include an imaging device that generates a three-dimensional image.

  The “main axis / normal direction calculation unit” calculates a main axis direction of a target candidate tissue having a linear structure or a normal direction of a tissue having a planar structure. For example, the main axis direction or the normal direction of the target candidate tissue may be calculated by calculating a Hessian matrix for the detection region and analyzing the eigenvalue of the calculated Hessian matrix.

  Furthermore, the “spindle / normal direction calculation unit” determines whether the target candidate tissue has a linear structure or a planar structure, depending on whether the calculated eigenvalue satisfies a predetermined threshold condition. May be determined.

  A “three-dimensional image” is an image composed of voxel data. For example, it is a three-dimensional image composed of a plurality of two-dimensional images. Examples of the two-dimensional image include a radiation image, a CT image, an MRI image, an RI image, and a PET image.

  “Target tissue” refers to a tissue composed of a linear structure or a planar structure composed of a predetermined part of a subject represented in a three-dimensional image. For example, coronary artery, cerebral blood vessel, liver blood vessel, bronchus, pulmonary blood vessel, etc. The structure | tissue which consists of these linear structures may be sufficient. It may also be a planar structure such as skin, lumbar pleura, cortical bone or the like.

  “Detection region” refers to a predetermined region including a target region in a three-dimensional image obtained by imaging a subject. For example, when a coronary artery is extracted as a target tissue, a region including a heart region or a region of the region is included. It may be a partial area or the like.

  In the image processing apparatus of the present invention, the determination unit performs a normalization process on a three-dimensional image prepared in advance including a target tissue of the same type as the target tissue based on the main axis direction or the normal direction of the same type of target tissue. Whether or not the target tissue is included in the target candidate region by analyzing the feature amount calculated using the machine learning method using the teacher data. It may be a thing which discriminates.

  In learning of a discriminator for discriminating a coronary artery, positive teacher data can include data representing a lesioned part such as a curved part, a branched part, a stenosis, a calcification, or a stent placement part in addition to a straight part of the coronary artery. By learning by including these in the teacher data, it is possible to cope with variations of the coronary arteries and to determine a lesioned part or the like as a blood vessel with high accuracy. A portion other than the coronary artery can be randomly prepared for the negative teacher data.

  The image processing method of the present invention sets a detection region in a three-dimensional image obtained by capturing an image of a subject, and is a method of a tissue having a main axis direction or a planar structure of a target candidate tissue having a linear structure in the detection region. Calculate the line direction, perform a normalization process on the target candidate area including the target candidate tissue based on the main axis direction or the normal direction, calculate the feature amount of the target candidate area subjected to the normalization process, It is characterized by determining whether a target tissue is included in a target candidate area | region using the calculated feature-value.

  The program of the present invention has a function of setting a detection region in a three-dimensional image obtained by imaging a subject, and a target candidate tissue composed of a normal direction of a tissue composed of a linear structure or a planar structure in the detection region. A function for calculating the main axis direction, a function for performing normalization processing on the target candidate area including the target candidate organization based on the main axis direction or the normal direction, and a feature amount of the target candidate area subjected to the normalization process And the computer is caused to execute a function of determining whether or not the target tissue is included in the target candidate region using the calculated feature amount.

  According to the image processing apparatus, method, and program of the present invention, the main axis direction of the target candidate tissue composed of a linear structure or a planar structure in the detection region or the normal direction of the tissue composed of the planar structure is calculated, and the main axis direction or Based on the normal direction, normalization processing is performed on the target candidate region including the target candidate organization, the feature amount of the target candidate region subjected to the normalization processing is calculated, and the target is calculated using the calculated feature amount. By determining whether or not the target tissue is included in the candidate region, even a target tissue having a linear structure accompanied by a change in appearance due to curvature, branching, or lesion can be detected stably.

Functional block diagram of image processing apparatus A series of flowcharts in the embodiment of the present invention Conceptual diagram for explaining the prior art Conceptual diagram for explaining the Gaussian pyramid structure Conceptual diagram for explaining the calculation of the principal axis direction of a target candidate tissue having a linear structure Conceptual diagram showing an example before and after normalization

  Hereinafter, an embodiment of an image processing apparatus 1 according to the present invention will be described with reference to the drawings.

  FIG. 1 is a block diagram showing a preferred embodiment of an image processing apparatus 1 according to the present invention.

  The configuration of the image processing apparatus 1 as shown in FIG. 1 is realized by executing a program read into an auxiliary storage device (not shown) on a computer (for example, a personal computer). In addition, this program is stored in an information storage medium such as a CD-ROM or distributed via a network such as the Internet and installed in a computer.

  The image processing apparatus 1 automatically calculates a target tissue represented in a three-dimensional image using a plurality of two-dimensional images captured by, for example, the X-ray CT apparatus 10, and includes a detection region setting unit 30. , A main axis / normal direction calculation unit 40, a normalization processing unit 50, a discrimination unit 60, and the like.

  The image processing apparatus 1 is connected to an image acquisition unit 20, an input unit 70, a display unit 80, and the like.

  The image acquisition unit 20 acquires a CT image (two-dimensional image) captured by, for example, an imaging apparatus as shown in FIG. 1, for example, the X-ray CT apparatus 10. Note that the image acquisition unit 1 may acquire not only a CT image but also a two-dimensional image such as a so-called MRI image, RI image, PET image, X-ray image, and the like. Moreover, the image acquisition part 20 acquires the three-dimensional image comprised by such a some two-dimensional image.

  The detection area setting unit 30 sets a detection area in the three-dimensional image acquired by the image acquisition unit 20. The detection area setting unit 30 sets a detection area by executing a detection algorithm, and executes a detection algorithm using, for example, threshold processing and segment division processing.

  The detection area setting unit 30 may set the area input by the input unit 80 as a detection area.

  The input unit 80 is a keyboard, a mouse, or the like.

  The main axis / normal direction calculation unit 40 calculates the main axis direction of the target candidate tissue having a linear structure in the detection region. For example, in the detection region set by the detection region setting unit 30, the main axis direction of the target candidate tissue having a linear structure or the normal direction of the target candidate tissue having a planar structure is calculated. Calculating the Hessian matrix and analyzing the eigenvalue of the calculated Hessian matrix to calculate the principal axis direction or the normal direction of the target candidate tissue, and the calculated eigenvalue further satisfies a predetermined threshold condition It is determined whether or not the target candidate tissue has a linear structure or a planar structure depending on whether or not the condition is satisfied. This makes it possible to roughly determine whether the structure is a linear structure or a planar structure. The determination unit 60 may be configured to more accurately determine whether the structure is a linear structure or a planar structure.

  The normalization processing unit 50 performs normalization processing on the target candidate region including the target candidate organization based on the main axis direction / normal direction calculated by the main axis / normal direction calculation unit 40.

  The determination unit 60 calculates the feature amount of the target candidate region that has been subjected to the normalization process by the normalization processing unit 50, and determines whether the target candidate region includes the target tissue using the calculated feature amount. To do.

  In addition, the determination unit 60 performs normalization processing on a three-dimensional image prepared in advance including a target tissue of the same type as the target tissue based on the main axis direction or normal direction of the same type of target tissue. It is provided as teacher data, and it is determined whether or not the target tissue is included in the target candidate region by analyzing the feature amount calculated using the machine learning method using the teacher data. It may be. In learning of a discriminator for discriminating a coronary artery, positive teacher data can include data representing a lesioned part such as a curved part, a branched part, a stenosis, a calcification, or a stent placement part in addition to a straight part of the coronary artery. By learning by including these in the teacher data, it is possible to cope with variations of the coronary arteries and to determine a lesioned part or the like as a blood vessel with high accuracy. A portion other than the coronary artery can be randomly prepared for the negative teacher data.

  Specifically, it is conceivable that a machine learning method based on Adaboost, which is a method of creating an integrated learning machine, is used as a target tissue detection method. The determination unit 60 can be used in the above-described target tissue detection method. The target organization sequentially updates the weights when re-sampling known methods such as feature point detection and learning data, and finally assigns the weights to the learning machines to the integrated learning machine. A machine learning method based on Adaboost is used. In the learning sample image, the center coordinate of the target tissue and its principal axis direction or the radius in the case of a linear structure are designated, and a cube rotated in the principal axis direction with respect to the center coordinate is set as a region of interest. The cube scale is normalized by the radius. The positive learning sample for discriminating the coronary artery includes data representing a curved portion of a straight blood vessel, a branch portion of the blood vessel, and a lesion such as stenosis, calcification, and stent placement portion. A portion other than the coronary artery is randomly prepared for negative teacher data.

  Next, for positive and negative learning sample images, a combination of the values of n pixel pairs selected at random is used as a feature value, and a positive and negative pattern is discriminated by a machine learning method based on Adaboost. Create a bowl. When a target tissue is detected, a three-dimensional image is scanned, and cubic regions of various sizes with the pixel of interest at the center are cut out to calculate feature amounts. This is input to the discriminator obtained in the learning stage, and a discriminant value is obtained.

  The feature amount for discrimination is a first-order differential value of each CT value in the X, Y and Z directions in the CT image. In addition, an absolute value of CT value, a histogram of CT value, a secondary differential value, etc. may be used.

  Further, various other statistical analysis methods and machine learning methods such as a linear discrimination method, a neural network, and a support vector machine can be used as a method for detecting the target tissue.

  The discriminating unit 60 detects a plurality of positions of the target tissue from the three-dimensional image by the method described above. The display unit 70 is a monitor that displays a two-dimensional image or a three-dimensional image, a CRT screen, or the like. By displaying the area determined as the target tissue by volume rendering, it is possible to overview the entire linear structure or planar structure and visualize its continuity.

  Next, processing in the embodiment performed in the image processing apparatus 1 having the above configuration will be described.

  FIG. 2 is a series of processing flowcharts until the image processing apparatus 1 determines a target tissue having a linear structure or a planar structure in a three-dimensional image.

  First, as shown in FIG. 2, a three-dimensional image captured by the X-ray CT apparatus 10 is input to the image acquisition unit 20 (step S1).

  Next, the detection region setting unit 30 detects the heart region by executing the detection algorithm described above. The detection region setting unit 30 sets a detection region from the detected heart region (step S2). This detection area refers to a predetermined area including a target area in a three-dimensional image obtained by imaging a subject, and may be, for example, an area including a heart area or a partial area thereof. .

  The detection region setting unit 30 generates a Gaussian pyramid by performing multi-resolution conversion on a three-dimensional image in advance in order to detect a target candidate tissue having a linear structure in the detection region.

  Thereafter, as illustrated in FIG. 4, the detection region setting unit 30 scans a detection algorithm for each generated Gaussian pyramid to detect a target candidate tissue (for example, a coronary artery) having a linear structure with a different size.

  The detection area setting unit 30 scans the detection area of FIG. 4A having a Gaussian pyramid structure, and sequentially scans the detection area of FIG. 4B and the detection area of FIG. Set the coordinates to execute the detection process.

  By sequentially scanning the multi-resolution image, target candidate tissues (for example, coronary arteries) of different sizes can be detected.

  Next, the main axis / normal direction calculating unit 40 calculates the main axis direction of the target candidate tissue (for example, coronary artery) having a linear structure in the local region centered on the detected coordinates (step S3).

The main axis / normal direction calculation unit 40 performs eigenvalue analysis of the Hessian matrix in the region including the target candidate tissue, and calculates the main axis direction. The Hessian matrix is a matrix having a second-order partial differential coefficient as an element, and for a three-dimensional image, the Hessian matrix is a 3 × 3 matrix as shown in Expression (1).

  When the Gaussian kernel (f) function is used, the filter coefficient for obtaining the Hessian matrix is obtained by Expression (2). σ corresponds to a linear structure having a size to be detected.

  When the Hessian matrix is subjected to eigenvalue decomposition to obtain eigenvalues and eigenvectors, the eigenvector corresponding to the eigenvalue closest to 0 is the principal axis direction.

  As shown in FIG. 5, the linear structure is known to have a characteristic that two of the three eigenvalues are large and one is close to zero. Therefore, it is effective to determine the likelihood of a linear structure from the eigenvalues and make a more detailed determination on the remaining candidates. Further, in the case of a planar structure, one of the three eigenvalues has a large absolute value, and two having features close to 0 are determined to be likely to be a planar structure.

The eigenvalue of the formula (1) has a relationship like the formula (3) with respect to the target candidate tissue having a linear structure, for example.

  The normalization processing unit 50 performs normalization processing on the target candidate region including the target candidate tissue based on the main axis direction calculated by the main axis / normal direction calculation unit 40 (step S4). As shown in FIG. 6A, the normalization processing unit 50 cuts out a normalized image along the calculated principal axis direction. As shown in FIG. 6B, the three-dimensional image of the target candidate tissue after the normalization process has a rotation-invariant feature. However, the image processing apparatus 1 does not necessarily need to perform normalization processing, and the determination unit 60 may use a machine learning method by acquiring feature amounts for determination under the same conditions as when normalized. .

  Next, the determination unit 60 calculates the feature amount of the target candidate region that has been subjected to the normalization process by the normalization processing unit 50, and uses the calculated feature amount to determine whether the target candidate region includes the target tissue. Is determined (step S5). The determination unit 60 extracts the feature amount from the target candidate region that has been subjected to the normalization process by the normalization processing unit 50 by the above-described machine learning method and performs the determination.

  In the embodiment of the present invention, the coronary artery is identified as an example of the target tissue. However, it can be used to extract other linear structures such as cerebral blood vessels, liver blood vessels, pulmonary blood vessels, and bronchi. It can also be used to extract planar structures such as skin, interlobar pleura, and cortical bone.

  As described above, the image processing apparatus 1 calculates the principal axis direction of the target candidate tissue composed of a linear structure in the detection region or the normal direction of the tissue composed of a planar structure, and based on the principal axis direction or the normal direction, The target candidate region including the candidate organization is normalized, the feature amount of the target candidate region subjected to the normalization processing is calculated, and the target tissue is included in the target candidate region using the calculated feature amount It is possible to stably detect even a target tissue having a linear structure accompanied by a change in appearance due to curvature, branching, or lesion.

1 Image processing device
DESCRIPTION OF SYMBOLS 10 X-ray CT apparatus 20 Image acquisition part 30 Detection area setting part 40 Main axis and normal direction calculation part 50 Normalization process part 60 Discriminating part 70 Display part

Claims (8)

A detection area setting unit for setting a detection area in a three-dimensional image obtained by imaging a subject;
A main axis / normal direction calculation unit for calculating a main axis direction of a target candidate tissue composed of a linear structure in the detection region or a normal direction of a tissue composed of a planar structure;
A normalization processing unit that performs a normalization process on the target candidate region including the target candidate tissue based on the main axis direction or the normal direction;
A determination unit that calculates a feature amount of the target candidate region that has been subjected to the normalization process, and determines whether the target candidate region includes a target tissue using the calculated feature amount; An image processing apparatus.   The principal axis / normal direction calculation unit calculates a Hessian matrix for the detection region, and calculates a principal axis direction or a normal direction of the target candidate tissue by analyzing an eigenvalue of the calculated Hessian matrix. The image processing apparatus according to claim 1, wherein the image processing apparatus is one.   The main axis / normal direction calculation unit determines whether the target candidate tissue has a linear structure or a planar structure depending on whether the calculated eigenvalue satisfies a predetermined threshold condition. The image processing apparatus according to claim 2, wherein: The discriminating unit performs normalization processing on a three-dimensional image prepared in advance including a target tissue of the same type as the target tissue based on a main axis direction or a normal direction of the target tissue of the same type. It is prepared as teacher data,
Using the teacher data to determine whether the target candidate region includes a target tissue by analyzing the calculated feature value using a machine learning method. The image processing apparatus according to any one of claims 1 to 3.   The image processing apparatus according to claim 1, wherein the target tissue is a coronary artery.   The image processing apparatus according to claim 1, wherein the detection area is an area including a heart area of the subject. Set the detection area in the 3D image obtained by imaging the subject,
Calculating the principal axis direction of the target candidate tissue consisting of a linear structure in the detection region or the normal direction of the tissue consisting of a planar structure;
Based on the main axis direction or the normal direction, the target candidate region including the target candidate tissue is subjected to normalization processing,
Image processing characterized in that a feature amount of the target candidate region subjected to the normalization process is calculated, and whether or not a target tissue is included in the target candidate region using the calculated feature amount apparatus. A function for setting a detection region in a three-dimensional image obtained by imaging a subject;
A function of calculating a principal axis direction of a target candidate tissue composed of a linear structure in the detection region or a normal direction of a tissue composed of a planar structure;
Based on the main axis direction or the normal direction, a function of performing a normalization process on the target candidate region including the target candidate tissue;
A feature amount of the target candidate region subjected to the normalization process is calculated, and a function of determining whether the target candidate region includes a target tissue using the calculated feature amount is realized by a computer. A program characterized by that.

Images