JP2022010572A - Image processing apparatus, image processing method and program - Google Patents

Abstract

To provide an image processing apparatus which can easily create a correct answer image considering a slice interval and a slice thickness.SOLUTION: An image processing apparatus comprises: a first acquisition unit which acquires a plurality of first tomographic images that are obtained by photographing a subject and a plurality of second tomographic images that indicate a region of interest for each of the plurality of first tomographic images; a first generation unit which generates a third tomographic image expressing the subject by using the first tomographic image included in a range of a prescribed thickness among the plurality of first tomographic images; and a second generation unit which generates a fourth tomographic image that indicates a region of interest in the third tomographic image by using the second tomographic image included in the range of the prescribed thickness among the plurality of second tomographic images.SELECTED DRAWING: Figure 1

Description

The present invention relates to an image processing device, an image processing method, and a program for creating a correct image indicating a region of interest drawn in an image taken by an image pickup device.

In recent years, segmentation using machine learning has become important in the field of medical image analysis. Segmentation is a process of distinguishing a region of interest drawn in an image from an region other than the region of interest, and is also called region extraction, region division, or image division. In order to improve the accuracy of segmentation based on machine learning, it is necessary to prepare a large number of learning images and correct answer images. Non-Patent Document 1 discloses a technique of selecting two correct answer images from a plurality of already created correct answer images and artificially generating a new correct answer image from the selected correct answer images.

Zach Eaton-Rosen, "Improving Data Augmentation for Medical Image Segmentation", International Conference on Medical Imaging with Deep Learning 20

In a three-dimensional medical image composed of a plurality of slices such as a CT image, the appearance of the region of interest differs depending on the interval between slices or the thickness of the slices, even if the region of interest is the same. Therefore, it is important to create a correct image in consideration of the slice interval and slice thickness. However, in the technique disclosed in Non-Patent Document 1, a correct image is created without considering the slice interval and the slice thickness.

An object of the present invention is to provide an image processing apparatus capable of easily creating a correct image in consideration of a slice interval and a slice thickness.

The image processing apparatus according to the first aspect of the present invention is
A first acquisition unit that acquires a plurality of first tomographic images obtained by photographing a subject and a plurality of second tomographic images indicating a region of interest for each of the plurality of first tomographic images. When,
A first generation unit that generates a third tomographic image representing the subject by using the first tomographic image included in a predetermined thickness range among the plurality of first tomographic images.
Of the plurality of second tomographic images, the second tomographic image included within the predetermined thickness range is used to obtain a fourth tomographic image showing a region of interest in the third tomographic image. The second generator to generate and
To prepare for.

The image processing method according to the second aspect of the present invention is
It is an image processing method performed by a computer.
The first acquisition step of acquiring a plurality of first tomographic images obtained by photographing a subject and a plurality of second tomographic images indicating a region of interest for each of the plurality of first tomographic images. When,
A first generation step of generating a third tomographic image representing the subject by using the first tomographic image included in a predetermined thickness range among the plurality of first tomographic images.
Of the plurality of second tomographic images, the second tomographic image included within the predetermined thickness range is used to obtain a fourth tomographic image showing a region of interest in the third tomographic image. The second generation step to generate and
including.

According to the present invention, it is possible to easily create a correct image in consideration of the slice interval and the slice thickness.

The figure which shows the structure of the image processing apparatus which concerns on 1st Embodiment. The flowchart which shows the processing procedure of the teaching data generation part 101. The figure explaining the CT image and the correct answer image included in the teaching data Simput. The figure explaining the CT image generated by the 1st generation part 120. The figure explaining the correct answer image generated by the 2nd generation part 130. The figure explaining the correct answer image generated by the 2nd generation part 130. The flowchart which shows the processing procedure of the learning processing execution part 102.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. The same or equivalent components, members, and processes shown in the drawings shall be designated by the same reference numerals, and duplicate description thereof will be omitted as appropriate. In addition, some of the components, members, and processes are omitted in each drawing.

Hereinafter, the present invention will be described by taking as an example the liver depicted in a CT image taken by an X-ray computed tomography (X-ray CT) apparatus as an area of interest. However, the present invention is applicable not only to the liver but also to all structures of the human body such as other organs, bones and muscles. Further, the present invention applies not only to CT images but also to tomographic images captured by modalities other than X-ray CT devices such as magnetic resonance imaging (MRI) devices, positron emission tomography (PET) devices, and ultrasonic image pickup devices. It is also applicable to this. The embodiment of the present invention is not limited to the following embodiments.

<First embodiment>
The configuration of the image processing apparatus 100 according to the first embodiment will be described with reference to FIG. 1. The image processing device 100 is a device that learns a classifier that extracts a liver region from a CT image. The image processing device 100 includes a teaching data generation unit 101 and a learning processing execution unit 102. The teaching data generation unit 101 generates new teaching data from the teaching data created in advance based on a predetermined thickness consisting of the slice thickness and the slice interval specified by the operator or the like. The learning process execution unit 102 learns the classifier 103 using the teaching data generated by the teaching data generation unit 101. The classifier 103 is a classifier that receives a CT image as an input and extracts a region of interest drawn in the CT image. The classifier 103 learned by the learning process execution unit 102 can extract a region of interest from a CT image having a slice thickness and a slice interval specified by an operator or the like with high accuracy.

The image processing device 100 includes a processor and a memory, and each of the above-mentioned functional units is realized by the processor executing a program stored in the memory. A part or all of each functional unit may be realized by a dedicated hardware device.

In addition to the image processing device 100, there is a data server 200 for storing data to be input to the image processing device 100 and data output by the image processing device 100. The data server 200 is an example of a computer storage medium, and is a large-capacity information storage device typified by a hard disk drive (HDD) or a solid state drive (SSD). The data server 200 may be held inside the image processing device 100, or may be separately provided outside the image processing device 100 and configured to be communicable via a network.

The teaching data includes input data equivalent to the data to be segmented, and correct answer data which is an output expected when this input data is input to the classifier 103. In this embodiment, the input data is a CT image, and the correct answer data is an image showing a region of interest in the CT image. Teaching data is also called teacher data or training data. The input data is also called an input image, a learning image, and learning data. The correct answer data is also called a correct answer image.

[Teaching data generator]
The configuration of the teaching data generation unit 101 will be described in more detail with reference to FIG.

The first acquisition unit 110 acquires a plurality of first tomographic images obtained by photographing a subject and a plurality of second tomographic images indicating a region of interest for each of the plurality of first tomographic images. do. The first acquisition unit 110 also has a first slice interval and a first slice thickness of the first tomographic image and a second tomographic image, and a second of the third tomographic image and the fourth tomographic image. Get the slice spacing and the second slice thickness.

The first acquisition unit 110 acquires a set of CT images (input images) and correct answer images from the teaching data Sinput stored in the data server 200. Further, the first acquisition unit 110 acquires the slice thickness and the slice interval value of the CT image to be processed by the classifier 103 from the data server 200, and newly generates the slice thickness of the CT image and the correct answer image. And generate information about slice spacing. The first acquisition unit 110 transmits information regarding the acquired CT image and the slice thickness and slice interval of the newly generated CT image to the first generation unit 120. Further, the first acquisition unit 110 transmits the acquired correct image and the information regarding the slice thickness and the slice interval of the newly generated correct image to the second generation unit 130.

The CT image and the correct answer image acquired by the first acquisition unit 110 will be described with reference to FIG. The CT image (input image) acquired by the first acquisition unit 110 is a three-dimensional image (a plurality of first tomographic images) composed of a plurality of tomographic images. The tomographic image 310, tomographic image 320, tomographic image 330, and tomographic image 340 in FIG. 3 are examples of tomographic images constituting one CT image. The tomographic image 310, tomographic image 320, tomographic image 330, and tomographic image 340 are consecutive slices of Nimput sheets, which are aligned along the patient's body axis. The region of interest 311 and the region of interest 321 and the region of interest 331 and the region of interest 341 are the livers visualized on the CT image, and are regions (regions of interest) to be extracted by the classifier 103. Further, the region 312, the region 313, the region 324, the region 325, and the region 326 are the torso, stomach, spine, right kidney, and left kidney depicted in the CT image, respectively. These areas are not the areas set as the areas of interest in the correct image. All tomographic images show areas that are not set as such areas of interest, but in FIG. 3, some non-areas of interest are not designated for ease of viewing.

The correct image is also a three-dimensional image (a plurality of second tomographic images) composed of a plurality of tomographic images. The tomographic image 350, tomographic image 360, tomographic image 370, and tomographic image 380 in FIG. 3 are examples of tomographic images constituting one correct image. Each tomographic image of the correct image corresponds to the tomographic image of the CT image. In FIG. 3, the tomographic image 350, the tomographic image 360, the tomographic image 370, and the tomographic image 380 correspond to the tomographic image 310, the tomographic image 320, the tomographic image 330, and the tomographic image 340, respectively. Each tomographic image of the correct image is a binary image showing the region of the liver depicted in the tomographic image of the corresponding CT image. Region 351 and region 361, region 371, and region 381 show the region of interest 311 and region 321 of interest, the region of interest 331, and the region of interest 341, respectively. It is not essential that the area of interest is drawn in the tomographic image, and the correct image showing that the area of interest is not drawn is included in the teaching data as a set for the tomographic image in which the area of interest is not drawn. It may be included.

The CT image and the correct answer image acquired by the first acquisition unit 110 hold information about the slice thickness and the slice interval. The slice thickness is the thickness of the surface (slice) that constitutes each tomographic image. In general, all tomographic images that make up a CT image have the same slice thickness. The slice interval is the distance in the imaging space between two consecutive tomographic images. Generally, all tomographic images constituting one CT image are arranged at equal intervals along the patient's body axis. Therefore, one CT image (tomographic image 310, tomographic image 320, ..., tomographic image 330, tomographic image 340) has one (same) slice interval. Further, in a general CT image, the slice thickness and the slice interval have the same value. Here, since each tomographic image of the correct image corresponds to the tomographic image of the CT image, the slice thickness and the slice interval of the correct image are equal to the slice thickness and the slice interval of the CT image. Hereinafter, the slice thickness of the CT image and the correct answer image acquired by the first acquisition unit 110 will be referred to as Tipput, and the slice interval will be referred to as Dinput. Tinput corresponds to the first slice thickness and Dinput corresponds to the first slice interval.

Further, the first acquisition unit 110 acquires the slice thickness and slice interval values of the CT image to be processed by the classifier 103 from the data server 200. Then, the slice intervals and slice thickness values of the newly generated CT image and the correct answer image are set. Hereinafter, the slice thickness and the slice interval set by the first acquisition unit 110 will be described as Target and Dtaget, respectively. Ttarget corresponds to the second slice thickness and Dtarget corresponds to the second slice interval.

The first acquisition unit 110 sets the Target to a value equal to the slice thickness of the CT image to be processed by the classifier 103. For example, suppose you want to extract the liver depicted in a 5 mm thick CT image with the classifier 103. At this time, the Target is set to 5. Then, the image processing apparatus 100 outputs the parameters of the discriminator 103, which is the most effective for extracting the liver depicted in the CT image having a thickness of 5 mm, as the final result (output of the second output unit 170). ..

When the classifier 103 is a classifier that inputs and discriminates a CT image as a three-dimensional image, the first acquisition unit 110 sets the Dtaget to a value equal to the Ttaget. According to this, the slice thickness and the slice interval of the generated teaching data can be matched with the CT image of the object to be processed by the classifier 103. On the other hand, when the classifier 103 is a classifier that discriminates by inputting a CT image as a two-dimensional image for each tomographic image, the first acquisition unit 110 sets the Digital to a value equal to that of Dinput. According to this, it is possible to generate a number of teaching data close to the teaching data to be input. When the classifier 103 is two-dimensional, the value of Dtaget may be the same as that of Ttaget as in the case of three-dimensionality, or may be set to a value different from that of Dinput and Ttaget.

In the above, the first acquisition unit 110 acquires the slice thickness and slice interval values of the CT image to be processed by the classifier 103 from the data server 200, and these values are input by the user. May be obtained as. The first acquisition unit 110 may acquire either or both of the slice thickness and the slice interval as numerical values from the user. Alternatively, the first acquisition unit 110 may acquire the type of image input to the classifier 103 from the user and set either or both of the slice thickness and the slice interval according to the type of the image.

The first generation unit 120 generates a new CT image (a plurality of third tomographic images) from the CT image (a plurality of first tomographic images) received from the first acquisition unit 110. The first generation unit 120 generates a third tomographic image representing the subject by using the first tomographic image included in the range of a predetermined thickness among the plurality of first tomographic images. do. Now, the CT image received by the first generation unit 120 from the first acquisition unit 110 is described as Cinput, and the CT image generated by the first generation unit 120 is described as Cartget. The first generation unit 120 averages a plurality of tomographic images included in a predetermined thickness range Rj among the tomographic images Cinput [i] (i = 1, ..., Ninput) constituting Cinput. As a result, a new tomographic image Category [j] (j = 1, ..., Average) is generated. Here, the thickness range Rj is calculated by the following equation for each j.

Since the tomographic image Cinput [i] has a thickness, only a part thereof may be included in the thickness range Rj. The first generation unit 120 may generate a new tomographic image Target [j] by averaging the tomographic images whose at least a part is included in the thickness range Rj. Alternatively, the first generation unit 120 may generate a new tomographic image Cartget [j] by averaging the tomographic images in which a portion having a predetermined ratio or more is included in the thickness range Rj. Further, the first generation unit 120 weighted and averages the tomographic images whose at least a part is included in the thickness range Rj by using the weight according to the ratio of the portions included in the thickness range Rj, and creates a new tomographic image. You may generate a Target [j].

FIG. 4 is a diagram illustrating an example of generating a new tomographic image. In this example, Dtarget = Ttarget = 2 × Dinput = 2 × Tinput. The first generation unit 120 averages the two tomographic images 310 and 320 to generate a new tomographic image 410, and also averages the two tomographic images 330 and 340 to generate a new tomographic image 420. Is being generated. The same processing is performed for a tomographic image (not shown) existing between the tomographic image 320 and the tomographic image 330. The first generation unit 120 regards the generated tomographic image 410, ..., The tomographic image 420 as one new CT image.

The second generation unit 130 generates a new correct image (a plurality of fourth tomographic images) from the correct image (a plurality of second tomographic images) received from the first acquisition unit 110. The second generation unit 130 uses the second tomographic image included in the range of a predetermined thickness among the plurality of second tomographic images to indicate the region of interest in the third tomographic image. Generate a tomographic image of. The pixel value of each coordinate of the fourth tomographic image may be determined as a representative value of the pixel value of the coordinates of the second tomographic image included in the range of a predetermined thickness. The correct answer image newly generated by the second generation unit 130 may be a binary image or a continuous value image. Which expression is to be generated is determined according to the expression of the correct image required for learning of the classifier 103 executed by the learning process execution unit 102 (so that the correct image of the required expression can be obtained). ..

Now, the correct answer image received by the second generation unit 130 from the first acquisition unit 110 is described as Minput, and the new correct answer image generated by the second generation unit 130 is described as Mtaget. When the correct image is a binary image, each pixel of the correct image takes one of the two pixel values. The first pixel value indicates that the pixel is a pixel included in the region of interest. Another pixel value indicates that the pixel is not included in the region of interest. On the other hand, when the correct image is a continuous value image, each pixel of the correct image takes two predetermined values, for example, any value between 0 and 1. At this time, the size of the pixel value represents the certainty that the pixel is included in the region of interest. That is, the larger the pixel value, the higher the possibility that the pixel is included in the region of interest.

First, a case where a new correct image Mtaget is generated as a binary image will be described. At this time, the second generation unit 130 displays a plurality of tomographic images included in the predetermined thickness range Rj among the tomographic images Minput [i] (i = 1, ..., Ninput) constituting the Minput. Average. Then, by applying the threshold value processing to the result, a new tomographic image Mtaget [j] (j = 1, ..., Ntaget) is generated. Rj is calculated by the above formula (Equation 1).

The generation of a new correct image is not limited to the above. If the pixel value of each coordinate in the new correct image Mtaget [j] is determined based on the representative value of the pixel value of the coordinate in each tomographic image Minput [i] included in the range Rj, a new method is used. A correct image may be generated. For example, the pixel value of each coordinate of the new correct image may be determined by a majority determination of the pixel value of the coordinate of each tomographic image Minput [i] included in the range Rj. Further, a new pixel value of the correct image may be determined by performing threshold processing on the average of the pixel values of the coordinates. In this case, the average of the pixel values may be a simple average or a weighted average in which the weight of the tomographic image near the center of the range Rj is set relatively large. Further, the pixel value of the new correct image may be determined by the product of the pixel values.

FIG. 5 is a diagram illustrating an example of generating a new correct image as a binary image. The setting of the slice thickness and the slice interval is the same as in the case of FIG. The second generation unit 130 generates a new tomographic image 510 from the two tomographic images 350 and 360, and also generates a new tomographic image 520 from the tomographic images 360 and 370. The same processing is performed for the tomographic image (not shown) existing between the tomographic image 360 and the tomographic image 370. The second generation unit 130 regards the generated tomographic image 510, ..., The tomographic image 520 as one new correct image.

Next, a case where a new correct image Mtaget is generated as a continuous value image will be described. At this time, the second generation unit 130 displays a plurality of tomographic images included in the predetermined thickness range Rj among the tomographic images Minput [i] (i = 1, ..., Ninput) constituting the Minput. By averaging, a new tomographic image Mtaget [j] is generated. Rj is calculated by the above formula (Equation 1). In addition, if the pixel value of each coordinate in the new correct image Mtaget [j] is determined based on the representative value of the pixel value of the coordinate in each tomographic image Minput [i] included in the range Rj, any method can be used. A new correct image may be generated. For example, instead of the simple average, the pixel value of each coordinate of the new correct image may be determined by the weighted average or the sum.

FIG. 6 is a diagram illustrating an example of generating a new correct image as a continuous image. The setting of the slice thickness and the slice interval is the same as in the case of FIG. The second generation unit 130 generates a new tomographic image 610 from the two tomographic images 350 and 360, and generates a new tomographic image 620 from the tomographic images 360 and 370. The same processing is performed for the tomographic image (not shown) existing between the tomographic image 360 and the tomographic image 370. The second generation unit 130 regards the generated tomographic image 610, ..., The tomographic image 620 as one new correct image.

The first output unit 140 associates the CT image received from the first generation unit 120 with the correct image received from the second generation unit 130 as a set. Then, the first output unit 140 adds the associated CT image and the correct answer image to the teaching data stored in the data server 200.

The first output unit 140 may output a correct image corresponding to the CT image to a display unit (not shown in FIG. 1). An example of a display device included in the display unit is a display. The display unit may display only the correct answer image corresponding to the CT image. Further, the display unit may simultaneously display the CT image and the corresponding correct answer image. Further, the display unit may simultaneously display the CT image and the correct answer image acquired from the data server 200 and the CT image and the correct answer image generated by the teaching data generation unit 101, or may display only one of them. ..

Next, with reference to FIG. 2, the processing procedure of the image processing method performed by the teaching data generation unit 101 will be described.

(S110)
In step S110, the first acquisition unit 110 acquires a set of CT images and a correct answer image from the teaching data stored in the data server 200. Further, the first acquisition unit 110 acquires the slice thickness and the slice interval value of the CT image to be processed by the classifier 103 from the data server 200, and newly generates a slice of the CT image and the correct answer image. Set information about thickness and slice spacing. The details of the processing are as described in the description of the first acquisition unit 110. Then, the first acquisition unit 110 transmits the acquired CT image and the information regarding the slice thickness and the slice interval of the newly generated CT image to the first generation unit 120. Further, the first acquisition unit 110 transmits the acquired correct image and the information regarding the slice thickness and the slice interval of the newly generated correct image to the second generation unit 130.

(S120)
In step S120, the first generation unit 120 acquires information regarding the CT image and the slice interval from the first acquisition unit 110. Then, the first generation unit 120 generates a new CT image from the CT image received from the first acquisition unit 110. The details of the processing are as described in the description of the first generation unit 120. Finally, the first generation unit 120 transmits the newly generated extracted CT image to the first output unit 140.

(S130)
In step S130, the second generation unit 130 acquires information on the correct image and the slice interval from the first acquisition unit 110. Then, the second generation unit 130 generates a new correct answer image from the correct answer image received from the first acquisition unit 110. The details of the processing are as described in the description of the second generation unit 130. Finally, the second generation unit 130 transmits the newly generated extracted correct image to the first output unit 140.

(S140)
In step S140, the first output unit 140 associates the CT image received from the first generation unit 120 with the correct image received from the second generation unit 130. Then, the first output unit 140 adds the associated CT image and the correct answer image to the teaching data Snew stored in the data server 200. If the Snew does not exist in the data server, the first output unit 140 creates the Snew in the data server 200. Then, the CT image and the correct answer image associated with the created Snew are added.

(S150)
In step S150, the control unit (not shown) of the image processing apparatus 100 determines whether or not the CT image and the correct answer image that have not been processed exist in the teaching data Sinput stored in the data server 200. If there is an unprocessed CT image and a correct answer image, the control unit (not shown) causes the teaching data generation unit 101 to execute the process of step S110. On the contrary, when the unprocessed CT image and the correct answer image do not exist, the control unit (not shown) ends the processing of the teaching data generation unit 101.

According to the above procedure, the teaching data generation unit 101 generates new teaching data Snew from the teaching data Sinput stored in the data server 200.

[Learning process execution unit]
Subsequently, the configuration of the learning process execution unit 102 will be described with reference to FIG. 1. The second acquisition unit 150 acquires the teaching data Snew generated by the teaching data generation unit 101 from the data server 200 and transmits it to the learning unit 160. The teaching data Snew consists of a plurality of CT images and correct answer images. Each correct answer image is associated with a corresponding CT image, and one CT image and one correct answer image form a set of data. These CT images and correct answer images are CT images and correct answer images generated by the first generation unit 120 and the second generation unit 130. In addition to the Snew, the second acquisition unit 150 may acquire the teaching data Sinput from the data server 200 and transmit the Sinput and the Snew to the learning unit 160.

The learning unit 160 acquires teaching data from the second acquisition unit 150. Then, the learning unit 160 learns the classifier 103 for extracting the region of interest using the acquired teaching data. As a result of learning, the learning unit 160 acquires the parameters of the classifier 103. Finally, the learning unit 160 transmits the acquired parameters of the classifier 103 to the second output unit 170.

The classifier 103 may be any classifier as long as it can be used for segmentation. The learning algorithm executed by the learning unit 160 is uniquely determined according to the type of the classifier 103. Therefore, the learning unit 160 executes the learning algorithm corresponding to the classifier 103. These learning algorithms are known techniques, and their processing procedures are disclosed in the literature and the like. Therefore, the details of the processing procedure will be omitted here. In the following, a case of configuring a two-dimensional classifier that processes a CT image for each two-dimensional cross-sectional image and outputs an image as a result of extracting a region of interest as a two-dimensional cross-sectional image will be described.

One of the processing steps of the learning algorithm is the calculation of loss. The loss function used to calculate the loss is whether the classifier 103 is a classifier that outputs, for each pixel in the CT image, whether or not the pixel is included in the region of interest in binary, or the pixel is the region of interest. It is selected according to whether it is a classifier that outputs the certainty (continuous value) contained in.

When the classifier 103 is a classifier that outputs a binary value, a binary image is used as the correct answer image. At this time, the learning unit 160 uses a binary cross entropy or the like as a loss function. Now, it is assumed that the CT image Ii and the correct answer image Ri of the T set are stored in the teaching data. However, i = 1, ..., T. The classifier 103 in the middle of learning is applied to the T CT images Ii, and the image as a result of extracting the region of interest is designated as Oi. The image sizes of the images Ri and Oi are M × N, the pixel value of the correct image is Ri (x, y), and the pixel value of the extracted image is Oi (x, y). However, 1 <= x <= W and 1 <= y <= H. Then, the binary cross entropy is calculated as follows.

On the other hand, when the classifier 103 is a classifier that outputs certainty, a continuous value image is used as the correct image. At this time, the learning unit 160 uses a square error or the like as a loss function. The squared error is calculated as follows.

When the learning unit 160 learns the classifier 103 using the loss function (Equation 3), the resulting classifier 103 determines the certainty that the pixel is included in the region of interest for each pixel in the CT image. It will be output.

The second output unit 170 stores the parameters of the classifier 103 received from the learning unit 160 in the data server 200.

This is the end of the explanation of the learning process execution unit 102.

Next, the processing procedure of the learning processing execution unit 102 will be described with reference to FIG. 7. As described above, the learning process execution unit 102 executes the same process as the known learning algorithm except that the calculation method of the loss function is different. Therefore, in the following, the processing procedure of the learning processing execution unit 102 will be briefly described.

(S210)
In step S210, the second acquisition unit 150 acquires the teaching data from the data server 200.

(S220)
In step S220, the learning unit 160 executes a learning algorithm determined according to the type of the classifier 103. The details of the processing executed in this step are as described in the explanation of the learning unit 160.

(S230)
In step S230, the second output unit 170 stores the parameters of the classifier 103 obtained in step S220 in the data server 200.

According to the above procedure, the learning process execution unit 102 learns the classifier 103.

The image processing apparatus 100 according to the first embodiment can efficiently generate a correct image in consideration of the slice thickness and the slice interval of the CT image to be processed by the classifier 103. Then, the classifier 103 is trained based on the acquisition result. The resulting classifier 103 can extract the region of interest with high accuracy from the CT image having the slice thickness and the slice interval.

In the above description, the case where the slice thickness of the CT image to be processed by the classifier 103 is a fixed value is taken as an example, and the case where Snew is generated as the teaching data used for the learning is described as an example. Other embodiments may be used.

For example, when there are a plurality of slice thicknesses (for example, 1 mm, 2 mm, 3 mm, 5 mm) of the CT image for which region extraction is desired, the above processing may be individually performed for each slice thickness. In other words, a new correct image may be generated and the discriminator may be learned for each slice thickness. In this case, the teaching data generation unit 101 generates new teaching data Snew for each assumed slice thickness. Further, the learning processing execution unit 102 learns and generates the classifier 103 corresponding to each slice thickness by using the teaching data Snew of the corresponding slice thickness. At this time, when the region of the unknown CT image is extracted using the learned discriminator 103, the discriminator 103 having the slice thickness is selected based on the slice thickness of the CT image to be processed, and the discriminator is used. The area extraction process may be performed using this. Further, not only the slice thickness but also a plurality of slice intervals may be set. For example, for each of the plurality of sets of assumed slice thickness and slice interval, teaching data of the corresponding slice thickness and slice interval may be generated and used for learning of the classifier 103.

Further, the assumed slice thickness T is divided into several groups (for example, T ≦ 1 mm, 1 mm <T ≦ 2 mm, 2 mm <T ≦ 3 mm, 3 mm <T ≦ 5 mm, 5 mm <T), and each of them is divided into several groups. The classifier 103 corresponding to the group may be generated. In this case, the teaching data generation unit 101 generates new teaching data Snew for each group of assumed slice thickness while setting various slice thicknesses included in the range of the group as Target. Further, the learning processing execution unit 102 learns and generates the classifier 103 corresponding to each group by using the teaching data Snew of each corresponding group. At this time, when the region of the unknown CT image is extracted using the learned classifier 103, the classifier 103 of the group including the slice thickness is selected based on the slice thickness of the CT image to be processed. The area extraction process may be performed using the classifier. Further, grouping is not always necessary, and the same processing may be performed with all slice thicknesses as one group. By any of the above methods, the region of interest can be extracted with high accuracy for CT images having various slice thicknesses.

In the above description, the image to be processed is described by taking a CT image captured by an X-ray CT apparatus as an example, but the image to be processed can be any arbitrary image such as an MRI tomographic image, a PET tomographic image, and an ultrasonic tomographic image. It may be a tomographic image.

(Other examples)
The present invention is also realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiment is supplied to the system or device via a network or various storage media, and the computer (or CPU, MPU, etc.) of the system or device reads the program. This is the process to be executed.

100: Image processing device 110: First acquisition unit 120: First generation unit 130: Second generation unit

Claims (16)

A first acquisition unit that acquires a plurality of first tomographic images obtained by photographing a subject and a plurality of second tomographic images indicating a region of interest for each of the plurality of first tomographic images. When,
A first generation unit that generates a third tomographic image representing the subject by using the first tomographic image included in a predetermined thickness range among the plurality of first tomographic images.
Of the plurality of second tomographic images, the second tomographic image included within the predetermined thickness range is used to obtain a fourth tomographic image showing a region of interest in the third tomographic image. The second generator to generate and
An image processing device. The first generation unit averages the first tomographic image included in the predetermined thickness range to generate the third tomographic image.
The image processing apparatus according to claim 1. The second generation unit determines the pixel value of each coordinate of the fourth tomographic image as a representative value of the pixel value of the coordinate of the second tomographic image included in the predetermined thickness range. ,
The image processing apparatus according to claim 1 or 2. The second generation unit performs the fourth tomographic image by averaging the second tomographic image included in the predetermined thickness range, or by performing threshold processing after averaging. To generate,
The image processing apparatus according to claim 3. The averaging performed by the second generation unit is a weighted average in which the weight of the second tomographic image near the center of the predetermined thickness range is set relatively large.
The image processing apparatus according to claim 4. Further, a learning unit for learning a classifier for extracting the region of interest by using the third tomographic image and the fourth tomographic image is provided.
The image processing apparatus according to any one of claims 1 to 5. The fourth generation unit takes either of two pixel values indicating whether each pixel in the third tomographic image is included or not included in the region of interest. Generate a tomographic image and
The learning unit causes the discriminator to output one of two pixel values indicating whether or not the pixel is included in the region of interest for each pixel in the third tomographic image. To learn,
The image processing apparatus according to claim 6. The second generation unit generates the fourth tomographic image in which each pixel in the third tomographic image takes a value representing the certainty that the pixel is included in the region of interest.
The learning unit learns so that the discriminator outputs a value indicating the certainty that the pixel is included in the region of interest for each pixel in the third tomographic image.
The image processing apparatus according to claim 6. The first acquisition unit includes the first slice interval and the first slice thickness of the first tomographic image and the second tomographic image, and the third tomographic image and the fourth tomographic image. Get the second slice interval and the second slice thickness,
The predetermined thickness range is determined based on the second slice interval and the second slice thickness.
The first tomographic image and the second tomographic image included within the predetermined thickness range are obtained based on the first slice interval and the first slice thickness.
The image processing apparatus according to any one of claims 6 to 8. The first acquisition unit acquires the type of image input to the classifier and determines the second slice interval and the second slice thickness based on the type of image input to the classifier. decide,
The image processing apparatus according to claim 9. The first acquisition unit acquires a plurality of sets of the second slice interval and the second slice thickness.
The first generation unit generates the third tomographic image for each set of the second slice interval and the second slice thickness.
The second generation unit generates the fourth tomographic image for each set of the second slice interval and the second slice thickness.
The image processing apparatus according to claim 9 or 10. The learning unit learns the classifier for each set of the second slice interval and the second slice thickness.
The image processing apparatus according to claim 11. The first tomographic image is a tomographic image constituting any one of a CT image, an MRI tomographic image, a PET tomographic image, and an ultrasonic tomographic image.
The image processing apparatus according to any one of claims 1 to 12. The first acquisition unit acquires the first tomographic image and the second tomographic image from the data server, and obtains the first tomographic image and the second tomographic image.
The first generation unit and the second generation unit output the third tomographic image and the fourth tomographic image to the data server.
The image processing apparatus according to any one of claims 1 to 13. It is an image processing method performed by a computer.
The first acquisition step of acquiring a plurality of first tomographic images obtained by photographing a subject and a plurality of second tomographic images indicating a region of interest for each of the plurality of first tomographic images. When,
A first generation step of generating a third tomographic image representing the subject by using the first tomographic image included in a predetermined thickness range among the plurality of first tomographic images.
Of the plurality of second tomographic images, the second tomographic image included within the predetermined thickness range is used to obtain a fourth tomographic image showing a region of interest in the third tomographic image. The second generation step to generate and
Image processing methods, including. A program for causing a computer to execute each step of the image processing method according to claim 15.

Images