JP2009211138A - Target area extraction method, device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To further improve extraction performance of a target area inside an input image. <P>SOLUTION: About each of pixels each expressing a reference point and pixels each expressing a point except the reference points inside a plurality of sample images wherein the reference points present each on an outline of the target area, and specifiable based on pixel value distributions of neighborhood areas are already known, the pixel value distributions of the neighborhood areas are previously performed with machine learning, and the reference points inside the input image are detected based on a result of the machine learning. An optional point C is set inside the target area inside the input image P, and a distinction area T thought that it includes the whole target area is set in the input image P. Outline likeness of each pixel inside the set distinction area T is calculated based on pixel value information of a neighborhood pixel of the pixel. The target area having the outline passing the reference points from the set distinction area T and including the optional point C is extracted based on the set optional point C, the detected reference points, and the calculated outline likeness of each pixel. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a method, apparatus, and program for extracting a target area from an input image.

  Conventionally, in the medical field, in order to provide an image having high diagnostic performance, processing for extracting and displaying a predetermined target region such as a lesion region or an organ region in a medical image has been performed.

  As a method for extracting a target area in an input image, for example, as disclosed in Patent Document 1, based on information on a point in a target area and a point in a background area specified by a user, The probability that each pixel of the pixel is a pixel indicating the target area is calculated, and the probability that the adjacent pixel is a pixel in the same area is calculated based on the local brightness difference (edge information) of the image. There is known a method of extracting a target area in an input image by distinguishing it from a background area using both the certainty factors.

Also, in Patent Document 2, a plurality of sample images including a target region are prepared, and pixel value information of a region in the vicinity of a pixel that is known whether it is a pixel representing an outline in the sample image is machine-learned in advance. Thus, the contour likeness of each pixel in the input image is calculated using the pixel value information of the neighboring region of the pixel, and the target region in the input image is extracted separately from the background region using the calculated contour likeness. Thus, a method for improving the extraction accuracy of the target region has been proposed.
US Pat. No. 6,973,212 JP 2007-307358 A

  However, in the above prior art, the target region is extracted based on the local pixel value distribution of the image, such as calculating the contour likeness of each pixel in the input image using the pixel value information of the neighboring region of the pixel. Therefore, when a pixel value distribution such as a contour exists inside or outside the target area, there is a problem that the target area may not be extracted correctly.

  For example, when extracting a specific organ region from a medical image, there is a tumor in the organ region, noise is present in the medical image, etc. When the pixel value distribution exists, the pixel value distribution due to such tumor or noise is erroneously recognized as the contour of the organ region, and the contour of the organ region may not be extracted correctly.

  In view of the above circumstances, an object of the present invention is to provide a target area extraction method, apparatus, and program capable of further improving the extraction performance of a target area.

  The target region extraction method of the present invention is a method for extracting a target region from an input image, and is present on the contour of the target region of the same type as the target region and can be specified based on the pixel value distribution of the neighboring region. Based on the result of the machine learning, the pixel value distribution of the neighboring region is pre-machine-learned for each of the pixels representing the reference point and the pixels representing the points other than the reference point in the plurality of sample images whose reference points are known. A detection step for detecting a reference point in the input image, a point setting step for setting an arbitrary point in the target region in the input image, and a discrimination region that is supposed to include the entire target region in the input image. A region setting step to be set, a calculation step of calculating the contour likeness of each pixel in the set discrimination region based on pixel value information of neighboring pixels of the pixel, an arbitrary set point, and a detected reference Points, and calculated Based on the contour likelihood of each pixel has a contour that passes through the reference point from the set determination area, and is characterized in that it comprises an extraction step of extracting a region of interest including the arbitrary point.

  In the above method, the calculation step calculates pixel value information of neighboring pixels for each of a pixel representing a point on the contour and a pixel representing a point other than the contour in a plurality of sample images whose contours of the target region are known. Machine learning may be performed in advance, and the contour-likeness of each pixel may be calculated based on the result of the machine learning.

  The target region extraction device of the present invention is a device that extracts a target region from an input image, and is present on the contour of the same target region as the target region and can be specified based on the pixel value distribution of the neighboring region. Based on the result of the machine learning, the pixel value distribution of the neighboring region is pre-machine-learned for each of the pixels representing the reference point and the pixels representing the points other than the reference point in the plurality of sample images with known reference points Detection means for detecting a reference point in the input image, point setting means for setting an arbitrary point in the target area in the input image, and a discrimination area that is supposed to include the entire target area in the input image. Area setting means for setting, calculation means for calculating the contour likeness of each pixel in the set discrimination area based on pixel value information of neighboring pixels of the pixel, an arbitrary set point, and a detected reference Points, and calculated And an area extracting unit that extracts a target area having an outline passing through a reference point from a set discrimination area and including an arbitrary point, based on the contour likeness of each pixel. .

  In the above apparatus, the calculation means obtains pixel value information of neighboring pixels for each of a pixel representing a point on the contour and a pixel representing a point other than the contour in a plurality of sample images whose contours of the target region are known. Machine learning may be performed in advance, and the contour-likeness of each pixel may be calculated based on the result of the machine learning.

  The target area extraction program of the present invention is a program for extracting a target area from an input image, and is present on the contour of a target area of the same type as the target area in the computer and has a pixel value distribution in a neighboring area. Machine learning is performed in advance on the pixel value distribution of the neighboring region for each of the pixel representing the reference point and the pixel representing the point other than the reference point in a plurality of sample images whose reference points that can be specified based on the machine are known. A detection unit that detects a reference point in the input image based on a learning result, a point setting unit that sets an arbitrary point in the target region in the input image, and the entire target region in the input image An area setting means for setting a possible discrimination area; a calculation means for calculating the contour likeness of each pixel in the set discrimination area based on pixel value information of neighboring pixels of the pixel; and an arbitrary set point Based on the detected reference point and the calculated contour outline of each pixel, it functions as an area extracting means for extracting a target area having an outline passing through the reference point from a set discrimination area and including an arbitrary point It is characterized by making it.

  According to the target area extraction method, apparatus, and program of the present invention, an arbitrary point is set in the target area in the input image, and a determination area that is supposed to include the entire target area is set in the input image. The contour likeness of each pixel in the set discrimination region is calculated based on the pixel value information of the neighboring pixels of the pixel, and the input image is calculated based on the set arbitrary point and the calculated contour likeness of each pixel. When extracting a target region from a plurality of sample images that exist on the same type of target region contour as the target region and have known reference points that can be identified based on the pixel value distribution of the neighboring region, For each of the pixel representing the reference point and the pixel representing the point other than the reference point, the pixel value distribution in the vicinity region is previously machine-learned, and the reference point in the input image is detected based on the result of the machine learning, and the detection is performed. In the result Since the target area is extracted based on the target area, even if there is a pixel value distribution such as a contour inside or outside the target area, it is detected as a point that exists on the correct contour of the target area. The contour of the target area can be determined so as to pass through the reference point reliably, and the extraction performance of the target area can be further improved.

  Hereinafter, an embodiment in which a target region extraction apparatus of the present invention is applied to one that extracts a liver region from a medical image will be described with reference to the drawings. The configuration of the target area extracting apparatus 1 as shown in FIG. 1 is realized by executing a target area extracting apparatus program read into the auxiliary storage device on a computer (for example, a personal computer). At this time, the target area extraction 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.

  As shown in FIG. 1, the target region extraction device 1 extracts a liver region from a medical image P acquired by a CT device or the like, and includes a reference point detection unit 10, a point setting unit 20, and a region setting unit 30. Further, a contour-likeness calculating unit 40, a region extracting unit 50, and the like are provided.

  The reference point detection unit 10 is a pixel and a reference point that represent a reference point in a plurality of sample images that exist on the outline of the liver region and whose reference points that can be specified based on the pixel value distribution of the neighboring region are known. For each of the pixels representing points other than the above, the pixel value distribution in the neighboring region is machine-learned in advance, and the reference point in the medical image P is detected based on the machine learning result. And a detector 12.

  Here, the reference points are determined according to the type of the target area to be extracted from the input image, and the number thereof is not limited. In the present embodiment, for example, as shown in FIG. 2, either or both of point A and point B existing at an angular position in the outline of the liver region consisting of a generally smooth curve is used as a reference point. .

  In addition, when extracting the heart region where the subject's heart is imaged from the medical image, among the points existing on the outline of the heart region, for example, the point existing at the boundary where the heart branches to the aorta, or The sharpest point at the bottom of the heart can be used as the reference point.

  The classifier acquisition unit 11 prepares a plurality of sample images including a liver region, as described in, for example, Japanese Patent Application Laid-Open No. 2006-139369, and includes pixels other than the pixels representing the reference points and the reference points in the sample images. For each pixel representing a point, the pixel value distribution in the neighborhood area is pre-machine-learned to determine whether each pixel in the sample image is a pixel indicating a reference point in the pixel value distribution in the neighborhood area of the pixel. The discriminator D for identifying based on this is acquired. The discriminator D acquired in this way can be applied when discriminating whether each pixel in an arbitrary medical image is a pixel indicating a reference point. When a liver region is extracted using two or more reference points (for example, reference points A and B), the classifier D is acquired for each reference point.

  Here, the neighboring area of each pixel is the reference point based on the pixel value distribution of the neighboring area such as the direction in which the pixel value changes in each pixel in the neighboring area and the magnitude of the change. It is desirable that the area has a size that can be identified. In addition, this neighborhood region may be a region centered on the pixel, or a region having the pixel at a position off the center.

FIG. 2 shows an example of the vicinity regions R _A and R _B of the reference points A and B. Here, the case where the vicinity area is a rectangular area is illustrated, but the vicinity area may be an area having various shapes such as a circle and an ellipse. Further, only the pixel value distribution of some pixels in the vicinity region may be used for the machine learning.

  In addition, for the process of obtaining the discriminator D, an adaboosting algorithm, a neural network, a support vector machine (SVM), or the like can be used.

  The detection unit 12 detects the reference points A and B in the medical image P by causing the classifier D acquired by the classifier acquisition unit 11 to scan the medical image P. In addition, before the detection of the reference point by the detection unit 21, when a determination region T that is supposed to include the entire target region is set in the region setting unit 30 described later, of the entire medical image P, The reference point in the medical image P may be detected by scanning the discriminator D only over the set discrimination region T or a part of the region including the discrimination region T.

  The point setting unit 20 sets an arbitrary point C (seed point) in the liver region in the medical image P. For example, the point setting unit 20 uses a keyboard or a pointing device provided in the target region extraction device 1 to The position on the medical image P designated based on the input may be set as an arbitrary point, and a constant mass is set at each point in the liver region automatically detected by the conventional target region detection method. May be set as the arbitrary point.

Further, when the reference points A and B are detected by the reference point detection unit 10 prior to the setting of the arbitrary point C by the point setting unit 20, the anatomical relationship between the liver region and the reference points A and B is detected. With an arbitrary positional relationship, an arbitrary point C (x _C , y _C ) is set using the reference point A (x _A , y _A ) and the reference point B (x _B , y _B ) according to the following equation (1). You may make it do.

  The arbitrary point C may be set at the approximate center of the liver region or may be set at a position deviated from the center of the liver region.

  The region determination unit 30 sets a determination region T that is supposed to include the entire liver region in the medical image P. For example, an operator using a keyboard or a pointing device provided in the target region extraction apparatus 1 The region on the medical image P designated based on the input of the above may be set as the discrimination region T, or there may be a liver region centered on the point C set in the point setting unit 20 An area having a size larger than the size may be automatically set as the discrimination area T. As a result, the range of the region of interest from the entire medical image P can be limited, and subsequent processing can be speeded up.

  The discrimination region T can adopt various shapes such as a rectangle, a circle, and an ellipse as the peripheral shape.

  The contour likelihood calculation unit 40 calculates the contour likelihood of each pixel in the discrimination region T set by the region setting unit 30 based on the pixel value information of the neighboring pixels of the pixel, and the evaluation function acquisition unit 4 1 and a calculation unit 42.

  The evaluation function acquisition unit 41 prepares a plurality of sample images including a liver region, and pixel value information of neighboring pixels for each of a pixel representing a point on the contour and a pixel representing a point other than the contour in the sample image Is previously learned as a feature amount in the pixel, thereby obtaining an evaluation function F for evaluating whether each pixel in the sample image is a pixel indicating an outline based on the feature amount.

  Specifically, as described in, for example, Japanese Patent Laid-Open No. 2007-307358, pixel value information of neighboring pixels of each pixel, for example, an area of 5 pixels in the horizontal direction × 5 pixels in the vertical direction centering on the pixel An evaluation function combining a plurality of weak discriminators for determining whether or not the pixel is a pixel indicating an outline using a combination of pixel value values of a plurality of different pixels in the combination of all the weak discriminators F is sequentially generated until it can be evaluated with a desired performance whether each pixel in the sample image is a pixel indicating an outline.

  The evaluation function F acquired in this way can be applied when evaluating whether each pixel in an arbitrary medical image is a pixel indicating the outline of the liver region.

  Also, a machine learning technique such as an adaboosting algorithm, a neural network, or a support vector machine (SVM) can be used for the process of obtaining the evaluation function F.

  Based on the feature amount of each pixel in the discrimination region T, the calculation unit 42 calculates an evaluation value of whether each pixel is like a contour, that is, whether the pixel is a pixel indicating a contour, using the evaluation function F. Is.

  The region extraction unit 50 extracts a liver region from the discrimination region T using an arbitrary point C, a reference point S, and the contour likeness of each pixel. For example, Yuri Y. Boykov, Marie-Pierre Jolly, “Interactive Graph Cuts for Optimal Boundary and Region Segmentation of Objects in ND images”, Proceedings of “International Conference on Computer Vision”, Vancouver, Canada, July 2001 vol.I, p.105-112., US Pat. No. 6973212 When the discrimination region T is divided into a liver region and a background region by a graph cut method (Graph Cuts) described in the specification etc., the boundary between the liver region and the background region is present on the outline of the liver region. A liver region is extracted from the discrimination region T so as to pass through the reference points A and B without fail.

Specifically, first, as shown in FIG. 3, a node N _ij (i = 1, 2,..., J = 1, 2,...) Representing each pixel in the discrimination region T and labels that each pixel can take. (In this embodiment, nodes S and T representing liver regions or background regions), a link n-link connecting nodes of adjacent pixels, a node N _ij representing each pixel, and a node S representing a liver region A graph composed of a link s-link that connects, a node N _ij that represents each pixel, and a link t-link that connects a node T that represents the background region is created.

Here, in the n-link, there are four links from each node to four adjacent nodes for each node N _ij representing each pixel in the determination region, and each adjacent node has a mutual node. There are two links going to. Here, the four links from each node N _ij to the four adjacent nodes represent the probability that the pixel indicated by the node is a pixel in the same region as the four adjacent pixels. The likelihood is obtained based on the contour likeness of the pixel. Specifically, when the contour likelihood of the pixel indicated by the node N _ij is less than or equal to the set threshold value, a predetermined maximum value of the probability is set for each of those links, and the contour likelihood is set to the threshold value. In the case described above, the probability of a smaller value is set for each link as the contour likelihood increases. For example, when the maximum value of the probability is set to 1000, when the likelihood of the contour of the pixel indicated by the node N _ij is equal to or less than the set threshold value (zero), four lines from the node to four adjacent nodes are displayed. If a value of 1000 is set for the link and the contour likelihood is greater than or equal to the set threshold (zero), the value calculated by the following formula (1000-(the contour likelihood / maximum value of the contour likelihood) x 1000) is used. Can be set for each link. Here, the maximum value of the contour likelihood means the maximum value of all the contour likelihoods calculated for each pixel in the discrimination region T by the calculation unit 42.

The s-link that connects the node N _ij representing each pixel and the node S representing the liver region represents the probability that each pixel is a pixel included in the liver region, and the node N _ij representing each pixel The t-link connecting the nodes T representing the background region represents the likelihood that each pixel is a pixel included in the background region, and the certainty indicates whether the pixel is either the liver region or the background region. When information indicating whether the pixel is a pixel to be shown has already been given, the information is set according to the given information.

Specifically, since an arbitrary point C is a pixel set in the liver region, as shown in FIG. 10, an s-link that connects a node N ₃₃ indicating the point C and a node S indicating the liver region. Set the probability of a large value to. In addition, the discrimination area T set to include the liver area on the basis of an arbitrary point set in the liver area usually includes the liver area and the background area existing around the liver area. Therefore, it is assumed that each pixel on the periphery of the discrimination region TT2 is a pixel indicating the background region, and nodes N ₁₁ , N ₁₂ ,..., N ₁₅ , N ₂₁ , N ₂₅ indicating those pixels are indicated. , N ₃₁ , and a node T representing the background area are set to a probability of a large value.

  Further, as shown in FIG. 4, the reference point between the reference point A and the point C among the entire line segments extending in the direction passing through the reference points A and B from the point C set in the point setting unit 20. Since each pixel located in the portion between B and point C can be determined to be a pixel existing inside the liver region, the node indicating each pixel and the node S representing the liver region are connected. The probability of a large value is set for s-link, and each pixel located in a portion extending from the reference point A to the opposite side of the point C and a portion extending from the reference point B to the opposite side of the point C is the liver. Since it can be determined that the pixel exists outside the region, the probability of a large value is set in the t-link that connects the node indicating each pixel and the node T indicating the liver region.

  Since the liver region and the background region are mutually exclusive regions, for example, as shown by dotted lines in FIG. 5, all the n-links, s-links, and t-links are cut off. By separating the node S from the node T, the discrimination region T is divided into a liver region and a background region, and the liver region is extracted. Here, optimal region division can be performed by performing cutting so that the total probability of all n-links, s-links, and t-links to be cut is minimized.

  In this example, the case of extracting the liver region using the graph cut method (Graph Cuts) is illustrated, but instead, for example, dynamic planning as described in JP-A-2007-307358 is used. The liver region may be extracted using other methods such as determining the contour of the liver region using a method.

  Next, an example of processing performed when a liver region is extracted from the medical image P with the above configuration will be described.

First, the discriminators D _A and B that can identify whether or not each pixel in an arbitrary medical image acquired in advance by the discriminator acquisition unit 11 is a pixel that indicates one of the reference points _A and B. _{Using B} , reference points A and B in the medical image P are detected. Next, the point setting unit 20 uses the reference point A (x _A , y _A ) and the reference point B (x _B , y _B ) in the target region in the medical image P according to the above equation (1). An arbitrary point C (x _C , y _C ) (seed point) is set in. Next, the region setting unit 30 sets a discrimination region T that seems to include the entire target region in the medical image P. Next, the determination unit 42 uses the evaluation function F that is acquired in advance by the evaluation function acquisition unit 41 to evaluate whether each pixel in an arbitrary medical image is a pixel indicating the outline of the liver region. The contour likeness of each pixel in T is calculated. Finally, the region extraction unit 50 determines whether the contour of the liver region is the reference point A, based on the set arbitrary point C and the calculated contour profile of each pixel by, for example, the graph cut method (Graph Cuts). The target region is extracted from the discrimination region T so as to pass through B, and the process is terminated.

  According to the above embodiment, an arbitrary point is set in the target area in the input image, and a determination area that is supposed to include the entire target area is set in the input image. When calculating the contour likeness of each pixel based on pixel value information of neighboring pixels of that pixel and extracting the target area from the input image based on the set arbitrary point and the calculated contour likeness of each pixel In addition, a pixel that represents a reference point and a reference in a plurality of sample images that exist on the contour of the target region of the same type as the target region and have known reference points that can be identified based on the pixel value distribution of the neighboring region For each pixel representing a point other than a point, the pixel value distribution in the neighboring area is machine-learned in advance, the reference point in the input image is detected based on the machine learning result, and the target area is further based on the detection result To do the extraction Therefore, even if there is a pixel value distribution such as a contour inside or outside the target area, the target must be surely passed through the reference point detected as a point existing on the correct contour of the target area. The contour of the region can be determined, and the extraction performance of the target region can be further improved.

  In the above-described embodiment, the case where the target region extraction apparatus 1 of the present invention is applied to the one that extracts the target region from the two-dimensional input image has been described. However, the target region is extracted from the three-dimensional input image. It can also be applied to things.

  For example, when determining the outline of a liver region in a three-dimensional medical image, the same as when extracting a liver region from a two-dimensional medical image, the corner of the outline of the liver region consisting of a smooth surface as a whole A point existing in is used as a reference point.

  The discriminator acquisition unit 11 prepares a plurality of three-dimensional sample images including the liver region, and for each of the voxels representing the reference point and the voxels representing points other than the reference point in the sample image, By performing machine learning in advance on the pixel value distribution, a discriminator is obtained that identifies whether each voxel in the sample image is a voxel indicating a reference point based on the pixel value distribution in the region near the voxel. Here, the neighborhood area of each voxel is the reference point based on the pixel value distribution in the neighborhood area, such as the direction in which the pixel value in each voxel within the neighborhood area changes, the magnitude of the change, and the like. It is desirable that the region is a three-dimensional region that is large enough to be identified. The detection unit 12 scans the classifier acquired by the classifier acquisition unit 11 on the three-dimensional medical image, thereby detecting a reference point in the medical image.

  Further, the point setting unit 20 sets an arbitrary point C in the three-dimensional coordinate system in the liver region of the three-dimensional medical image, and the region determination unit 30 includes the liver region in the three-dimensional medical image. A three-dimensional existence range that is supposed to include the whole of is set. Here, the existence range can adopt various shapes such as a hexahedron and a sphere as the peripheral shape.

  In addition, the evaluation function acquisition unit 41 prepares a plurality of three-dimensional sample images including the liver region, and for each of the voxels representing points on the contour and points other than the contour in those sample images, By performing machine learning in advance on the pixel value information of neighboring voxels, an evaluation function F that can evaluate whether each voxel in an arbitrary three-dimensional medical image is a voxel indicating the outline of the liver region is acquired. Here, as the pixel value information of neighboring voxels, for example, pixels in a plurality of different voxels in a cubic region of 5 voxels in the X-axis direction × 5 voxels in the Y-axis × 5 voxels in the Z-axis centering on the voxel. A combination of values can be used. Next, the calculation unit 42 uses the evaluation function F to calculate the likelihood of each voxel contour, that is, whether the voxel is a voxel indicating a contour, based on the feature amount of each voxel in the discrimination region T. To calculate.

  When the region extraction unit 50 divides the three-dimensional discrimination region T into a liver region and a background region by a three-dimensional graph cut method (Graph Cuts) described in, for example, US Pat. No. 6,732,212, The liver region is extracted from the discrimination region T so that the boundary between the liver region and the background region always passes through the reference point detected by the detection unit 12, and the process is terminated.

  In the above embodiment, the case where the target region extraction apparatus 1 of the present invention is applied to one that extracts a liver region from a medical image has been described. However, the present invention is not limited to this, and regions to be extracted from various input images, for example, The present invention can be applied when extracting target regions such as various organ regions or lesion regions in a medical image.

  As described above, since the number of reference points used for extracting the target area and the characteristics of each reference point differ depending on the type of target area to be extracted, two or more target areas are selectively extracted. In this case, the recording means included in the target area extraction apparatus 1 records in advance a reference point used for extraction of the target area for each type of target area, and the target area extraction apparatus 1 uses the target area. When extracting the target area, the user may select a target area to be extracted and automatically set a reference point suitable for the extraction of the target area.

  Further, the various correspondences described above when determining the contour of the liver region in the medical image can also be applied when extracting the region from various input images.

The block diagram which shows embodiment of the target area extraction apparatus of this invention The figure which shows an example of the reference point of a liver area | region, and its vicinity area | region The figure for demonstrating one method of extracting an object area | region by the area | region extraction part of FIG. Diagram for explaining one method of setting s-link and t-link values based on the position of seed point and reference point The figure for demonstrating one method of extracting an object area | region by the area | region extraction part of FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Target area | region extraction apparatus 10 Reference point detection part 20 Point setting part 30 Area setting part 40 Contourness calculation part 50 Area extraction part A, B Reference point P Medical image C Arbitrary point T in the liver area | region T Discrimination area | region

Claims (5)

A method of extracting a target area from an input image,
A pixel representing the reference point and a reference in a plurality of sample images that exist on the outline of the target region of the same type as the target region and whose reference points that can be specified based on the pixel value distribution in the neighboring region are known For each pixel representing a point other than a point, a detection step of previously learning a pixel value distribution of a neighboring region and detecting a reference point in the input image based on the result of the machine learning;
A point setting step for setting an arbitrary point in the target area in the input image;
In the input image, an area setting step for setting a discrimination area that is supposed to include the entire target area;
A calculation step of calculating the contour likeness of each pixel in the set discrimination region based on pixel value information of neighboring pixels of the pixel;
Based on the set arbitrary point, the detected reference point, and the calculated contour likeness of each pixel, the contour has a contour that passes through the reference point from the set discrimination region, and the arbitrary An extraction process for extracting the target area including a point.   In the calculation step, pixel value information of neighboring pixels is pre-machined for each of a pixel representing a point on the contour and a pixel representing a point other than the contour in a plurality of sample images whose contours of the target region are known. The target region extraction method according to claim 1, wherein learning is performed and the contour-likeness of each pixel is calculated based on the result of the machine learning. An apparatus for extracting a target area from an input image,
A pixel representing the reference point and a reference in a plurality of sample images that exist on the outline of the target region of the same type as the target region and whose reference points that can be specified based on the pixel value distribution in the neighboring region are known For each pixel representing a point other than a point, a detection unit that performs machine learning in advance on a pixel value distribution in a neighboring region and detects a reference point in the input image based on the result of the machine learning;
Point setting means for setting an arbitrary point in a target area in the input image;
In the input image, an area setting means for setting a discrimination area that is supposed to include the entire target area;
Calculating means for calculating the contour likeness of each pixel in the set discrimination region based on pixel value information of neighboring pixels of the pixel;
Based on the set arbitrary point, the detected reference point, and the calculated contour likeness of each pixel, the contour has a contour that passes through the reference point from the set discrimination region, and the arbitrary A target region extracting apparatus comprising: a region extracting unit that extracts the target region including a point.   The calculation means previously calculates pixel value information of neighboring pixels for each of a pixel representing a point on the contour and a pixel representing a point other than the contour in a plurality of sample images whose contours of a target region are known. 4. The target area extracting apparatus according to claim 3, wherein learning is performed and the likelihood of the contour of each pixel is calculated based on a result of the machine learning. A program for extracting a target area from an input image,
On the computer,
A pixel representing the reference point and a reference in a plurality of sample images that exist on the outline of the target region of the same type as the target region and whose reference points that can be specified based on the pixel value distribution in the neighboring region are known For each pixel representing a point other than a point, a detection unit that performs machine learning in advance on a pixel value distribution in a neighboring region and detects a reference point in the input image based on the result of the machine learning;
Point setting means for setting an arbitrary point in a target area in the input image;
In the input image, an area setting means for setting a discrimination area that is supposed to include the entire target area;
Calculating means for calculating the contour likeness of each pixel in the set discrimination region based on pixel value information of neighboring pixels of the pixel;
Based on the set arbitrary point, the detected reference point, and the calculated contour likeness of each pixel, the contour has a contour that passes through the reference point from the set discrimination region, and the arbitrary A target region extraction program that functions as region extraction means for extracting the target region including a point.

Images