JP2008167950A - Radiographic image processing method and apparatus, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the quality of a radiation image indicating an object without increasing the dose of radiation to be emitted to the object in a radiographic image processing method. <P>SOLUTION: The radiation image 11 for input composed of an object image 11H indicating an object 1P and a part specifying image 11C indicating the boundary Pc of a specified part Px in the object 1P and the other part Po is prepared for each of the plurality of objects 1P of the same kind, the radiation image 33 for a teacher which emphatically indicates the specified part Px of the object image 11H is prepared for each of the respective objects 1P by the radiography of the object 1P, and a teacher learning completed filter 40 for which learning is performed with the radiation image 11 for the input as an object and the radiation image 33 for the teacher as the teacher is obtained. Thereafter, the radiation image 21 of the same kind as the radiation image 11 for the input prepared for a given object 3P is inputted to the teacher learning completed filter 40, and the radiation image 60 whose image quality degradation is compensated and which emphasizes the specified part Px is formed. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a radiographic image processing method, apparatus, and program for obtaining a radiographic image that emphasizes and expresses a specific part in a subject.

  Conventionally, in medical radiography and the like, a high-pressure image and a low-pressure image are obtained by radiography of a subject using radiation having different energy distributions, and the subject is identified by weighted subtraction processing between the high-pressure image and the low-pressure image. There is known a technique for obtaining an energy subtraction image that emphasizes a part exhibiting the radiation absorption characteristics of, for example, a bone part or a soft part of a living tissue (see Patent Document 1). This energy subtraction image is an image formed based on the difference between the high pressure image and the low pressure image.

  In radiography for obtaining the high-voltage image and the low-voltage image, for example, two types of radiations having different energy distributions generated with different tube voltages of the radiation source are applied to the subject twice in total at different timings. Each of the two stimulable phosphor sheets disposed with a copper plate interposed between them by two-shot radiography to irradiate the above-described high-pressure image and low-pressure image, or by one-time irradiation of the subject. A one-shot radiography method for simultaneously recording a high-pressure image and a low-pressure image of a subject is known.

  The energy subtraction image formed using the high-pressure image and the low-pressure image is more specific than the radiographic image (hereinafter referred to as simple radiographic image) obtained by a normal radiographic method (hereinafter referred to as simple radiographic method). Although it is excellent in that the part can be emphasized, the image includes more noise. In the simple radiography method, a simple radiographic image of the subject is acquired by radiography in which one type of radiation is irradiated to the subject once without using a plurality of types of radiation having different energy distributions. .

  Noise generated in the energy subtraction image is mainly caused by a shortage of radiation dose to be irradiated when the high pressure image or the low pressure image is acquired.

  That is, in medical radiography and the like, it is desired to reduce the burden on patients by reducing the dose of radiation used for radiography. For example, in radiography (two-shot radiography) in which normal radiography is required twice, radiography (one-shot radiography) in which a sufficient dose is not used twice or the dose is attenuated by the absorption of radiation by a copper plate The energy subtraction image created using the radiographic image (high-pressure image or low-pressure image) obtained by the radiography method has a lower image quality than the simple radiographic image obtained by the simple radiography method.

  In both the one-shot radiography method and the two-shot radiography method, it is required to reduce the dose of radiation applied to the subject in the radiography. If an attempt is made to reduce the dose, the amount of noise generated in the radiographic image obtained by this radiography increases and the image quality deteriorates.

  On the other hand, a radiological image that estimates a component representing a bone portion of a subject from one radiographic image obtained by simple radiography and emphasizes the bone portion in the subject without performing weighted subtraction processing of the plurality of radiographic images. There are known methods for forming. This technique is intended to obtain an image similar to the bone image that is the energy subtraction image without performing radiation imaging using radiation having different energy distributions.

  More specifically, this method is intended to obtain an image similar to the bone part image by the following procedure.

That is, a radiographic image for a teacher is formed in advance by emphasizing a bone portion obtained by radiography of a human chest that is a subject. When a learning simple radiographic image obtained by radiography of a human chest that is the same type of subject as the subject is input, a radiographic image using the teacher radiographic image as a teacher, that is, a bone portion is emphasized. Learning is repeated so that a radiographic image is output, and a teacher-learned filter (a filter using an artificial neural network (ANN)) is created. Thereafter, a simple radiographic image obtained by radiography of the human chest that is the same type of diagnosis subject as the subject is input to the teacher-learned filter, and a radiographic image for diagnosis of the human chest with the bones emphasized is input. It is to be obtained (see Patent Document 2).
JP-A-3-285475 US Patent Publication No. 2005 / 0100208A1

  However, the method using the teacher-learned filter is not reliable enough to estimate a bone part that is a specific part in the subject, and the distinction between the bone part and a part other than the bone part is unclear. Thus, the component representing the soft part in the subject may enter the image area representing the bone part expressed in the emphasized manner. That is, the soft part component may appear in the bone region as a false image.

  Therefore, there is a demand for obtaining a radiographic image that emphasizes and expresses a specific part in a subject without generating the false image.

  The present invention has been made in view of the above circumstances, and provides a radiographic image processing method, apparatus, and program capable of improving the quality of a radiographic image representing the subject without increasing the dose of radiation applied to the subject. It is intended to provide.

  In the radiographic image processing method of the present invention, a part representing a boundary between a specific part in the subject obtained by radiography of the subject and another part different from the specific part for each of a plurality of same-type subjects A radiographic image for input composed of a specific image and a subject image representing the subject is prepared, and a radiographic image for teacher representing the specific portion in the subject obtained by radiography of the subject is displayed for each subject. And for the teacher corresponding to the subject so that a radiographic image of the subject in which the specific part in the subject is emphasized is output in response to an input radiation image representing each subject. Obtain a teacher-learned filter that trains radiation images as a teacher, and then create a radiation image of the same type as the input radiation image for a given subject of the same type as the subject The radiographic image is input into the teacher-learned filter, and a radiographic image of the subject is formed in which the same part as the specific part in the given subject is emphasized. .

  It is desirable that the radiation dose used for radiography for creating the teacher radiographic image be larger than the radiation dose used for radiography for creating the subject image.

  The teacher radiation image may be formed by a weighted subtraction process using a high-pressure image and a low-pressure image obtained by radiation imaging using radiation having different energy distributions, so-called energy subtraction process.

  The input radiation image may be formed by a weighted subtraction process using a high-pressure image and a low-pressure image obtained by radiation imaging with radiation having different energy distributions, so-called energy subtraction process. The input radiation image may be a simple radiation image obtained by simple radiation imaging.

  The subject image can be a simple radiation image obtained by simple radiation imaging.

  The specific part may be a part having specific radiation absorption characteristics different from other parts.

  The subject may be a living tissue, and the specific part may include at least one of a bone part, a rib, a posterior rib, an anterior rib, a clavicle, and a spine.

  The subject may be a living tissue, and another part different from the specific part may include at least one of the lung field, the mediastinum, the diaphragm, and the ribs.

  The subject can be a living tissue, and the specific part can be a bone or soft part of the living tissue.

  The specific part may be a part where the position in the subject has changed between the high-pressure image and the low-pressure image.

  The teacher-learned filter performs learning for obtaining the teacher-learned filter and formation of a radiographic image for a given subject for each of a plurality of different spatial frequency bands, and is formed for each spatial frequency band. Each of the radiographic images can be synthesized to obtain one radiographic image.

  The radiological image processing apparatus according to the present invention includes a part specifying image representing a boundary between a specific part in the subject obtained by radiography for each of the same kind of subjects and another part different from the specific part, and A radiographic image for input composed of a subject image representing a subject, and a radiographic image for teacher for each subject expressed by emphasizing the specific part in the subject obtained by radiography of the subject. The teacher radiation image corresponding to the subject is used as a teacher so that a radiation image of the subject in which the specific part in the subject is emphasized is output in response to an input radiation image representing the subject. A filter acquisition unit that obtains a learned teacher-learned filter and creates a radiation image of the same type as the input radiation image for a given subject of the same type as the subject. A part-enhanced image that forms a radiographic image of the subject in which the same part as the specific part in the given subject is enhanced by inputting the radiographic image into the teacher-learned filter And a forming means.

  The program of the present invention represents a part specifying image representing a boundary between a specific part in the subject obtained by radiation imaging for a plurality of subjects of the same type and another part different from the specific part, and the subject. An input radiation image composed of a subject image and a teacher radiation image for each subject that highlights and displays the specific part in the subject obtained by radiography of each subject represent each subject. The teacher radiograph corresponding to the subject was learned as a teacher so that the radiographic image of the subject in which the specific part in the subject is emphasized is output in response to the input of the input radiographic image A procedure for obtaining a teacher-learned filter, a procedure for creating a radiation image of the same type as the input radiation image for a given subject of the same type as the subject, and the radiation image A radiological image processing method for executing a procedure for forming a radiographic image of the subject in which the same part as the specific part in the given subject is emphasized It is for execution.

  The same kind of subject means, for example, a subject having substantially the same size, shape, structure, and radiation absorption characteristics of each part. For example, the human body is the same part, and the chests of different adult males are the same type of subject. The abdomen of different adult women or the heads of different children are also the same kind of subject. For example, in the case of an industrial product, it means a subject having substantially the same size, shape, structure, material, and the like. Further, for example, the same type of subject may be a part of the chest of an adult male different from each other (for example, 1/3 of the chest close to the neck). In addition, the same type of subjects can be different small areas in the same subject.

  “To create a radiation image of the same type as the input radiation image for the given subject” means that the given subject is subjected to the same processing as the processing for obtaining the input radiation image. This means that a radiographic image of a given subject is created. That is, for example, when performing radiography of the given subject under the same imaging conditions as when acquiring an input radiographic image, and acquiring the input radiographic image for the radiographic image obtained by this radiography The radiographic image of the given subject can be created by performing the same image processing.

  The emphasis on the specific part is not limited to the case where the specific part is expressed more conspicuously than the other part, and only the specific part may be expressed.

  The region specifying image is, for example, an image in which a predetermined tissue to which the local region mainly belongs is determined for each local region in a radiographic image, or an image in which a boundary between different tissues is determined. Moreover, the said site | part specific image can be obtained by the discrimination | determination process of a specific site | part and a site | part different from this specific site | part, for example.

  The radiographic image processing method, apparatus, and program of the present invention are directed to an input radiographic image, and the radiographic image for the teacher is output so that the radiographic image of the subject in which a specific part in the subject is emphasized is output. A learning filter that has been trained as follows, and for the given subject, a radiation image of the same type as the input radiation image is created, and the radiation image is input to the teacher-learned filter, and the given subject Since the radiographic image in which the same part as the specific part is emphasized is formed, the quality of the radiographic image representing the subject can be improved without increasing the dose of radiation applied to the subject. .

  That is, the image input to the teacher-learned filter as in the past is not limited to a simple radiation image, but a part specifying image that represents a boundary between a specific part in the subject and another part and a subject image that represents the subject. Since the image is input to the learned filter, the occurrence of the false image can be suppressed.

  More specifically, in the conventional method in which only a simple radiation image is input to the teacher learned filter, a false image is generated because the reliability of estimating a specific part in the subject is insufficient. On the other hand, in the present invention, since the part specifying image representing the boundary in addition to the subject image representing the subject is input to the teacher learned filter, the specific part in the subject is compared with the input of only the simple radiation image. More information can be used to distinguish between and other parts. Therefore, it is possible to improve the reliability of estimating the specific part by the supervised learned filter, thereby compensating for the false image generated in the radiation image of the given subject.

  By the above, it is possible to create a radiographic image in which a specific part in the subject is emphasized without increasing the dose of radiation irradiated to a given subject, and improve the quality of the radiographic image representing the subject. be able to.

  Further, if the teacher radiological image is less subject to image quality degradation due to noise or the like than the subject image and the part specifying image constituting the learning radiographic image corresponding to the teacher radiographic image, the teacher learned filter is It can be learned to compensate for image quality degradation. Then, by passing a radiation image of the same type as the input radiation image through a teacher-learned filter, radiation that compensates for image quality degradation that has occurred in the radiation image when a radiation image of the same type as the input radiation image is created. An image can be obtained.

  In addition, if the radiation dose used for radiography for creating the radiographic image for teacher is made larger than the dose of radiation used for radiography for creating the subject image, the above-mentioned radiation for teacher can be more reliably obtained. The image can be an image with less image quality deterioration than the subject image constituting the input radiation image, and thereby the quality of the radiation image representing the subject can be improved.

  In addition, if the teacher radiograph is an energy subtraction image formed by a weighted subtraction process using a high-pressure image and a low-pressure image obtained by radiation imaging with radiation having different energy distributions, so-called energy subtraction process, it is more reliable. In addition, the teacher radiation image can be an image that emphasizes the specific part.

  Here, if the specific part is a part having a specific radiation absorption characteristic different from the other part, the boundary between the specific part and the other part in the subject can be determined more reliably. A radiographic image formed by accurately emphasizing a specific part in a subject can be formed.

  The radiation image processing method and apparatus and program of the present invention will be described below. The radiological image processing method according to the first embodiment of the present invention includes, as an input radiographic image, a simple radiographic image representing a subject, and a specific part and other parts in the subject created from the simple radiographic image. Two types of images of the part specifying image representing the boundary are adopted.

  FIG. 1 is a diagram showing a procedure for acquiring a supervised learned filter used in the radiological image processing method according to the first embodiment, and FIG. 2 is a procedure for obtaining a diagnostic radiographic image using the supervised learned filter. FIG. Note that the hatched portion in the figure indicates an image or image data representing the image.

  The radiographic image processing method is obtained by performing simple radiography 10 on each chest 1P for each chest 1Pα, 1Pβ (hereinafter collectively referred to as chest 1P) of an adult male that is a plurality of similar subjects. A subject image 11H for learning, which is a simple radiation image representing the chest 1P, and a bone portion Px which is a specific part in the chest 1P obtained by subjecting the subject image 11H to boundary extraction processing 12 and the bone portion Px An input radiation image 11 including a learning part specifying image 11C representing a boundary Pc with another part Po different from the above is prepared.

  In the above simple radiography method, a radiographic image (simple radiographic image) of the subject is obtained by radiography in which one type of radiation is irradiated to the subject once without using a plurality of types of radiation having different energy distributions. To do.

  Further, a radiological image 33 for teacher is obtained for each chest 1P, which is obtained by emphasizing the bone Px, which is a specific part in each of the chests 1Pα, 1Pβ,. To do.

  Then, the teacher learned filter 40 is obtained in which the input radiation image 11 is the target and the teacher radiation image 33 is learned as a teacher.

  That is, the teacher-learned filter 40 uses a set of the input radiation image 11 and the teacher radiation image 33 prepared for each chest 1Pα, 1Pβ, and so on, and for each chest 1Pα, 1Pβ,. Corresponding to each chest 1Pα, 1Pβ,... So that when the created input radiation image 11 is inputted, the radiation image 50 of the chest 1P representing the bone Px is emphasized. This is obtained by learning the teacher radiation image 33 as a teacher. More specifically, the teacher-learned filter 40 uses a set of the input radiation image 11 and the teacher radiation image 33 prepared for the chest 1Pα, and the input radiation image 11 corresponding to the chest 1Pα is input. The training radiation image 33 corresponding to the chest 1Pα is learned as a teacher so that the radiation image 50 corresponding to the chest 1Pα representing the bone portion Px in an emphasized manner is output. is there.

  The teacher radiographic image 33 represents a bone image which is an energy subtraction image formed by weighted subtraction processing 32 of the high pressure image 31H and the low pressure image 31L obtained by the radiography 30 for each chest 1P, that is, energy subtraction processing. Is.

  As shown in FIG. 2, after obtaining the teacher-learned filter 40, simple radiography 20 is performed on the chest 3 </ b> P to be diagnosed which is one given subject of the same type as the chest 1 </ b> P, and the input radiation is obtained. A radiation image 21 of the same type as the image 11 is created.

  That is, a subject image 21H to be diagnosed representing the chest 3P, which is a simple radiation image obtained by simple radiation imaging 20 of the chest 3P, and a chest 3P obtained by subjecting the subject image 21H to boundary extraction processing 22 A radiographic image 21 is created that includes a bone-specific part Px, which is a specific part of the first part, and a part-specific image 21C to be diagnosed representing a boundary Pc between another part Po different from the bone part Px.

  The subject image 21H and the part specifying image 21C to be diagnosed are input to the teacher learned filter 40 obtained as described above, and the bone portion Px in the chest 3P that is a given subject is emphasized and represented. A diagnostic radiographic image is formed. This diagnostic radiographic image is a radiographic image in which the fake image of a part other than the bone portion in the image representing the bone portion is suppressed.

  The radiation image 21 of the same type as the input radiation image 11 is obtained on the basis of the simple radiography 20 under the same imaging conditions as the simple radiography 10 for the given chest 3P. . That is, the input radiation image 11 and the radiation image 21 are obtained by radiation imaging in which a subject is irradiated with substantially the same dose of radiation having substantially the same radiation energy distribution. The process performed in the boundary extraction process 22 is also equivalent to the process performed in the boundary extraction process 12.

  The input radiation image 11, the chests 1Pα, 1Pβ,... Used for creating the teacher radiation image, and the one chest 3P given when creating the diagnostic radiation image 60 are both the same type of subjects. It is. That is, each of the breasts 1Pα, 1Pβ,..., 3P is a living tissue having a shape, a structure, a size, and radiation absorption characteristics of each part that are substantially equal. Further, the bone part Px which is the specific part is a part having specific radiation absorption characteristics different from the other part Po in the chest which is the subject.

  Further, the boundary extraction process 12 (boundary extraction process 22) is performed for each small area portion in the simple radiation image 11H (simple radiation image 21H), where the tissue to which the small area portion mainly belongs is a bone portion or other than a bone portion. Are determined, and the determination results in each small region portion are integrated to acquire the region specifying image 11C (region specifying image 21C).

  As described above, according to the radiographic image processing method of the first embodiment, the specific part in the subject and the specific part are not increased without increasing the dose of radiation applied to the subject to be diagnosed. It is possible to generate a bone part image that more clearly represents a boundary with another different part.

  It should be noted that by using the teacher radiographic image 33 having less image quality degradation than the subject image 11H and performing learning by the teacher learned filter 40, the image quality degradation caused in the subject image 21H representing the given subject is compensated. In addition, it is possible to form a diagnostic radiographic image in which the bone portion Px in the given chest 3P is emphasized.

  However, the radiographic image processing method of the present invention can be applied regardless of the degree of image quality degradation of the teacher radiographic image 33 and the subject image 11H. That is, for example, the radiographic image processing method of the present invention can be applied even if the teacher radiographic image 33 includes image quality degradation similar to that of the subject image 11H.

  The radiation image processing method according to the second embodiment of the present invention will be described below. In this radiographic image processing method, as an input radiographic image, a high-pressure image representing a subject, a bone image with deteriorated image quality formed by weighted subtraction processing of the high-pressure image and the low-pressure image of the subject, and deterioration of the image quality with the high-pressure image Three types of images of a part specifying image representing a boundary between a specific part and another part in the subject formed using the bone part image are used.

  FIG. 3 is a diagram showing a procedure for acquiring a supervised learned filter used in the radiographic image processing method according to the second embodiment, and FIG. 4 is a radiographic image for acquiring a diagnostic radiographic image using the supervised learned filter. It is a figure which shows the procedure of a processing method.

  In the radiation image processing method of the second embodiment, radiations having different energy distributions are obtained for each of the breasts 1Qα, 1Qβ (hereinafter collectively referred to as the chest 1Q) of adult women, which are a plurality of similar subjects. An input radiation image 15 created using the high-pressure image 15H and the low-pressure image 15L representing the chest 1Q obtained by the radiation imaging 14 is prepared.

  The input radiation image 15 includes the high-pressure image 15H that is a subject image, the bone image 15K that is a subject image with degraded image quality formed by the weighted subtraction process 16 of the high-pressure image 15H and the low-pressure image 15L, and the high-pressure image. From the part specifying image 15C representing the boundary Qc between the bone part Qx in the chest 1Q and the other part Qo different from the bone part Qx formed by the boundary extraction process 17 using 15H and the bone part image 15K These three types of learning radiation images are included.

  The radiography 14 is radiography in which the dose of radiation applied when obtaining the high-pressure image 15H is larger than the dose of radiation applied when obtaining the low-pressure image 15L. Therefore, the high pressure image 15H is an image with less noise, and the low pressure image 15L is an image with more noise than the high pressure image. Further, the image quality of the bone portion image 15K created using the low-noise image 15L with much noise is deteriorated.

  For the boundary extraction process 17, various image processing techniques and the like that determine the boundary between a specific part and another part that are conventionally known can be applied.

  The input radiation image 15 is prepared, the image quality is less deteriorated than the learning high-pressure image 15H obtained by radiography of the chest 1Q, and the bone portion Qx in the chest 1Q is highlighted. A radiographic image 36 for teacher is prepared for each chest 1Qα, 1Qβ,.

  The teacher subject image 36 representing the bone part can be created by using a conventionally known method. For example, the chest 1Q obtained by the radiation imaging 35 of the chest 1Q using a dose larger than the radiation dose used in the individual radiography for each chest Q1 when the input radiation image 15 is created The bone part image formed by the weighted subtraction process between the high-pressure image to be represented and the low-pressure image can be obtained.

  Next, a teacher-learned filter 41 is obtained by learning the input radiation image 15 and learning the teacher radiation image 36 as a teacher.

  That is, the teacher-learned filter 41 compensates for image quality degradation when the high-pressure image 15H for learning for each chest 1Q that is each subject, the bone image 15K, and the part specifying image 15C are input. The teacher radiographic image 36 is trained as a teacher so that a radiographic image 51 representing the bone portion Qx in the chest 1Q, which is the specific part, is output. More specifically, the teacher learned filter 41 is obtained by learning from a set of the input radiation image 15 and the teacher radiation image 36 corresponding to each of the prepared several types of subjects 1Qα, 1Qβ,. Can do.

  After obtaining the teacher-learned filter 41, a radiation image 25 of the same type of diagnosis object as the input radiation image 15 is obtained for the breast 3Q of a given adult woman to be diagnosed, which is the same kind of subject as the chest 1Q. The radiographic image 25 is created and input to the teacher-learned filter 41 obtained as described above, the image quality deterioration is compensated, and a bone part Qx which is a specific part in the given chest 3Q is obtained. An enhanced radiographic image is formed. This radiographic image is a radiographic image in which mixing of false images of parts other than the bone portion into the image representing the bone portion is suppressed.

  The radiation image 25 is created using the high-pressure image 25H and the low-pressure image 25L representing the chest 3Q obtained by radiation imaging 24 using radiation having different energy distributions with respect to the chest 3Q.

  That is, the radiation image 25 is a high-pressure image 25H that is a subject image to be diagnosed, and a bone image that is a subject image to be diagnosed that has deteriorated in image quality, formed by weighted subtraction processing 26 of the high-pressure image 25H and the low-pressure image 25L. Diagnosis representing the boundary Qc between the bone part Qx in the chest 3Q formed by the boundary extraction processing 27 using the 25K and the high-pressure image 25H and the bone part image 25K and another part Qo different from the bone part Qx It consists of three types of images of the target part specifying image 25C.

  As described above, according to the radiographic image processing method of the second embodiment, the quality of the radiographic image representing the subject can be improved without increasing the dose of radiation applied to the subject to be diagnosed. .

  Here, an example of a technique for performing the boundary extraction processes 17 and 27 will be described. As described above, conventionally known methods can be applied to this boundary extraction processing. FIG. 5 is a diagram showing a state of the boundary extraction process.

  In the boundary extraction process, two classes (for example, an image region having a value of −1 and an image region having a value of +1) are determined as a class to be discriminated.

  A bone image E representing the radiation image of the chest D1 obtained by weighted subtraction between the high-pressure image obtained by radiography of the chest D1 as the subject and the low-pressure image, and the high-pressure image F are used as input radiation images. .

  Further, a part specifying image G obtained by manually labeling the two classes for distinguishing a part other than the bone part and the bone part in the chest part from the radiographic image of the chest part D1 is used as a teacher radiation image. And when the bone part image E and the high voltage | pressure image F are input into the discrimination filter N1, the site | part specific image J showing the boundary of the bone part in the said chest D1 and parts other than this bone part is formed. The discriminating filter N1 is obtained by learning the teacher site identification image G as a teacher.

  The part specifying image J is a similar image to the teacher part specifying image G.

  The learning of the discrimination filter N1 is performed by, for example, defining a subwindow Sw in a small region corresponding to each other in each of the bone part image E, the high-pressure image F, and the part specifying image G, and pixels in the subwindow Sw A feature amount that is a value of and a class corresponding to the feature amount are determined and executed.

  The feature amount is the sub-window Sw in each of the different spatial frequency bands (bone image EH, EM, EL and high voltage image FH, FM, FL) obtained by multiresolution conversion of the bone image E and the high voltage image F. It is a pixel value of a rectangular area of 5 × 5 pixels, and if the spatial frequency band is 3 bands, 3 × 5 × 5 = 75 pixel values, and if it is 8 bands, 8 × 5 × 5 = 200 pixel values It will be represented by Then, using a support vector machine (SVM) to be described later, learning is performed for two classes, that is, two classes for discriminating between a bone part and a part other than the bone part.

  The boundary extraction process 17 and the boundary extraction process 27 having the boundary extraction process discriminating filter N1 learned as described above, when a bone image and a high-pressure image of the same subject are input, A part specifying image representing a boundary with a part other than the bone part is formed.

  Hereinafter, the discrimination of two classes (a bone part and a part other than the bone part) by the support vector machine (SVM) will be described. FIG. 6 is a diagram showing discrimination of two classes by the support vector machine.

  For this support vector machine (SVM), the following documents can be referred to.

  Nello Cristianini (Author), John Shawe-Taylor (Author), Takeshi Ohkita (Translation), titled “Introduction to Support Vector Machine”, Kyoritsu Shuppan, published on March 25, 2005, p. 149 to P.I. 156.

Regarding the problem of learning the following function that discriminates two classes y = {-1,1} corresponding to an n-dimensional feature vector x, first consider the case where the discriminant function is linear.

At this time, the geometric distance (margin) between the discriminant plane and the learning sample is

  It becomes.

The support vector machine learns the discriminant plane that maximizes this margin under the constraint that all learning samples are correctly separated by the discriminant function.

  ξ is a relaxation variable that allows learning samples that are not correctly identified. C is a parameter that sets a trade-off between model complexity and constraint relaxation.

The above problem is equivalent to solving the following dual problem, and a global solution can always be obtained from the characteristics of the convex quadratic programming problem.

The discriminant function obtained by solving this dual problem is expressed by the following equation.

  This function is a linear function, but in order to extend it nonlinearly, if the input X is mapped to a higher-order feature space Φ (X) and the vector Φ (X) in that feature space is regarded as the previous input X, Good (X → Φ (X)). In general, mapping to a high-dimensional space is accompanied by a significant increase in computational complexity, but the inner product terms appearing in the formula to be optimized are expressed as K (x, y) = <Φ (x), Φ (y)>. If it is replaced with a kernel function that satisfies, the same result as that calculated after mapping to a higher dimension in the calculation of the input dimension can be obtained. As the kernel function, an RBF kernel, a polynomial kernel, a sigmoid kernel, or the like can be used.

  Next, the discrimination filter N1 for boundary extraction processing will be supplemented. FIG. 7 is a diagram illustrating a state in which sub-windows are set in a radiographic image for determining the boundary to be subjected to boundary extraction processing and a teacher image of a class corresponding to the radiographic image.

  A subwindow Sa is set in the radiation image Za to be discriminated, and the value of the pixel Ga in the subwindow Sa is used as a feature amount. In the teacher image Zb of the class corresponding to the radiation image Ga, the class label at the position of the center pixel Gb of the subwindow Sb set at the same position is the teacher data.

  One-dimensional output values corresponding to N-dimensional inputs (features) form one learning sample in pairs. Discrimination filter learning is performed using a set of many learning samples.

  The discrimination result by the learned discrimination filter is a result for one pixel. Therefore, a discrimination image is obtained by causing the discrimination filter to scan all pixels. This is the same in the case of support vector regression described later.To generate a bone image as described later, at all pixel positions in a high spatial frequency band image, a middle spatial frequency band image, and a low spatial frequency band image, Nonlinear filtering is applied to determine the corresponding bone value.

  Next, the creation of the teacher learned filter 41 will be described in detail.

  FIG. 8 is a diagram showing a state in which a radiation image for each spatial frequency band for a given subject is input to a teacher-learned filter to create a diagnostic radiation image, and FIG. 9 is a teacher learning for each spatial frequency band It is a figure which shows a mode that a completed filter is created.

  Here, a plurality of part specifying images having different resolutions created from the part specifying image obtained by radiography and boundary extraction processing of the subject, and the different spatial frequencies representing the subject It is assumed that it consists of subject images for each band. Further, the teacher radiographic image is obtained by radiographing a subject of the same type as the subject, the image quality degradation is less than that of the subject image, and the same portion as the specific portion in the subject is emphasized. It is assumed that it consists of a plurality of radiographic images for teachers for different spatial frequency bands.

  That is, in order to obtain a plurality of part specifying images having lower resolutions different from each other from a part specifying image having one type of resolution, a reduction process for reducing the number of pixels with respect to the one type of part specifying image is reduced. To obtain a low-resolution part specifying image composed of a small number of pixels. As a result, the resolution of each part specifying image can be adjusted to different spatial frequency bands in the subject image. Note that a multi-resolution conversion technique for obtaining a plurality of subject images having lower spatial frequency bands that are different from each other from subject images having one type of spatial frequency band will be described later.

  The teacher-learned filter is intended for the input radiological image including the plurality of region specifying images having different resolutions and the subject images for the different spatial frequency bands, and the spatial frequency bands different from each other. It is assumed that a plurality of radiographic images for teachers are learned as teachers.

  Then, for a given subject to be diagnosed of the same type as the subject, a plurality of radiation images for the different spatial frequency bands of the same type as the input radiation image are created, and for each of the different spatial frequency bands. A plurality of radiation images are input to a supervised learned filter, and the supervised learned filter compensates for image quality degradation and emphasizes a specific part in the given subject. A plurality of radiation images for each band are formed, and the plurality of radiation images are combined to create one radiation image.

  That is, as shown in FIG. 8, the teacher-learned filter 41 multi-resolution converts each of the high-pressure image 25H that is a radiographic image of the chest 3Q that is a given subject to be diagnosed and the bone image 25K. A plurality of radiographic images 61H for each spatial frequency band to be diagnosed based on the input of a plurality of radiographic images for each spatial frequency band and the part specifying image 25C for each spatial frequency band different from each other. , 61M, 61L and a plurality of the created radiographic images 61H, 61M, 61L can be combined to obtain a diagnostic radiographic image 61.

  Here, the teacher learned filter 41 includes a high frequency band teacher learned filter 41H, a medium frequency band teacher learned filter 41M, a low frequency band teacher learned filter 41L, an image synthesis filter 41T, and the like. .

  As shown in FIG. 9, the radiographic images for training 36H, 36M, and 36L for each spatial frequency band representing the chest 1Q prepared for creating the teacher learned filter 41 are compensated for image quality degradation. In addition, the radiographic image 36 (bone high-resolution image) mainly representing the bone that is the specific part is obtained by multi-resolution conversion.

  Further, the bone image 15KH, 15KM, 15KL for each spatial frequency band, which is the input radiation image 15 for each spatial frequency band prepared for creating the teacher learned filter 41, for each spatial frequency band. The high-pressure images 15HH, 15HM, and 15HL are obtained by performing multi-resolution conversion on the bone image 15K representing the chest Q1 and the high-pressure image 15H, respectively, as in the case of the teacher radiation image 36.

  The part specifying images 15CH, 15CM, and 15CL for each spatial frequency band are obtained by reducing the part specifying image C.

  That is, the radiographic image 36 for teacher is subjected to multi-resolution conversion, and a radiographic image representing a high frequency band, which is a radiographic image for each spatial frequency band (hereinafter referred to as a teacher high frequency band image 36H), representing a medium frequency band. A radiographic image (hereinafter referred to as a teacher medium frequency band image 36M) and a radiographic image representing a low frequency band (hereinafter referred to as a teacher low frequency band image 36L) are formed.

  The learning bone image 15K is subjected to multi-resolution conversion, and a radiographic image representing a high-frequency band (hereinafter referred to as a bone high-frequency band image 15KH), a medium frequency band, which is a radiographic image for each spatial frequency band. (Hereinafter referred to as bone intermediate frequency band image 15KM) and a radiation image representing a low frequency band (hereinafter referred to as bone low frequency band image 15KL).

  Further, the high-voltage image 15H for learning is subjected to multi-resolution conversion, and a radiographic image representing a high-frequency band (hereinafter referred to as a high-frequency high-frequency band image 15HH) and a medium-frequency band representing the radiographic image for each spatial frequency band. A radiographic image (hereinafter referred to as a high-voltage medium frequency band image 15HM) and a radiographic image representing a low-frequency band (hereinafter referred to as a high-voltage low-frequency band image 15HL) are formed.

  FIG. 10 is a diagram illustrating a state in which an image is subjected to multiresolution conversion.

  As shown in FIG. 10, for example, the high-pressure and high-frequency band image 15HH is obtained by upsampling the high-pressure image 15H (high-pressure high-resolution image) and the high-pressure medium-resolution image H1 obtained by down-sampling the high-pressure image 15H. It is what was done.

  In the downsampling, a Gaussian low-pass filter with σ = 1 and 1/2 thinning of the high-pressure image 15H (high-pressure high-resolution image) are performed. The upsampling is performed using cubic B-spline interpolation.

  As in the case of the high-voltage high-frequency band image 15HH, the high-voltage medium-frequency band image 15HM is an increase of the high-pressure medium-resolution image H1 and the high-pressure low-resolution image H2 obtained by down-sampling the high-pressure medium-resolution image H1. It was obtained by sampling.

  The high-voltage low-frequency band image 15HL is obtained by down-sampling the high-voltage low-resolution image H2 and the high-pressure low-resolution image H2 in the same manner as when the high-voltage high-frequency band image 15HH or the high-pressure medium-frequency band image 15HM is acquired. It was obtained by upsampling with the high-pressure ultra-low resolution image H3.

  As described above, the bone high-frequency band image KH, the bone mid-frequency band image KM, and the bone low-frequency band image KL are also created for the bone image E.

  Further, in order to match the resolution of the part specifying image 15C for learning with the resolution of each image, a reduction process for reducing the number of pixels is performed on the part specifying image 15C to obtain a lower resolution image. As a result, the high-resolution part specifying image 15C (hereinafter also referred to as a boundary high-frequency band image 15CH), a medium-resolution radiation image (hereinafter referred to as a boundary high-frequency band image 15CM), and a low-resolution radiation image (hereinafter referred to as a boundary high-frequency band image 15CM). , A boundary low frequency band image 15CL) is formed.

  Each of the boundary high frequency band image 15CH, the boundary medium frequency band image 15CM, and the boundary low frequency band image 15CL is limited to the case where the high resolution image is subjected to the reduction process to obtain the low resolution image as described above. Absent. For example, a part specifying image corresponding to this spatial frequency band may be created from an image of a specific spatial frequency band for each different resolution.

  The teacher learned filter 41 is also created for each of the three spatial frequency bands. That is, the high frequency band teacher learned filter 41H, the medium frequency band teacher learned filter 41M, and the low frequency band teacher learned filter 41L are obtained by learning for each spatial frequency band.

  Hereinafter, a case where the high frequency band teacher learned filter 41H and the like are created by learning will be described.

  As shown in FIG. 9, the bone high-frequency band image 15KH for learning, the high-voltage high-frequency band image 15HH for learning, the boundary high-frequency band image 15CH that is a high-resolution part specifying image for learning, and the high-frequency band image 36H for teachers For each, a sub-window Sw consisting of a rectangular area of 5 pixels × 5 pixels (a total of 25 pixels), which is a small area corresponding to each other, is set.

  Then, with respect to the feature amount consisting of the value of 25 pixels constituting each sub-window Sw of the bone high-frequency band image 15KH, the high-voltage high-frequency band image 15HH, and the boundary high-frequency band image 15CH, the sub-window Sw of the teacher high-frequency band image 36H A learning sample whose target value is the value of the center pixel is taken out, and a plurality of learning samples are extracted while moving the subwindow. The high frequency band teacher learned filter 41H is obtained by learning, for example, 10,000 kinds of learning samples extracted in this manner.

  Note that the high frequency band image 51H, the medium frequency band image 51M, and the low frequency band image 51L described later are similar images of the teacher high frequency band image 36H, the teacher medium frequency band image 36M, and the teacher low frequency band image 36L, respectively. It is.

  The high frequency band teacher learned filter 41H and the like are obtained by learning a regression model using the following support vector regression. This regression model has a boundary between the feature amount of the input bone high-frequency band image 15KH (image represented by 25 pixels) and the feature amount of the high-pressure high-frequency band image 15HH (image represented by 25 pixels). High-frequency band that outputs a high-frequency band image 51H that mainly compensates for the bone portion that is the specific part, with image quality degradation compensated according to the feature amount of the high-frequency band image 15CH (the image represented by the 25 pixels). This is a nonlinear filter.

  Further, the above-mentioned medium frequency band teacher learning is performed by the same learning using the bone medium frequency band image 15KM, the high voltage medium frequency band image 15HM, the boundary medium frequency band image 15CM, and the teacher medium frequency band image 36M. A finished filter 41M is obtained.

  Further, the low frequency band teacher learning is performed by learning similar to the above using the bone low frequency band image 15KL, the high voltage low frequency band image 15HL, the boundary middle frequency band image 15CL, and the teacher low frequency band image 36L. A finished filter 41L is obtained.

  As described above, the learning of the regression model is performed for each spatial frequency band, and the teacher learned filter 41 including the teacher learned filter 41H, the teacher learned filter 41M, and the teacher learned filter 41L is obtained.

  As shown in FIG. 8, the teacher-learned filter 41 created as described above has the same kind of diagnosis as the input radiation image 15 created for the breast 3Q of a given adult woman to be diagnosed. An image for each spatial frequency band for each of the bone image 25K, the high-pressure image 25H, and the region specifying image 25C, which are the target radiographic images 25, is input.

  That is, a high-frequency high-frequency band obtained by multi-resolution conversion of the bone high-frequency band image 25KH, the bone intermediate-frequency band image 25KM, the bone low-frequency band image 25KL, and the high-pressure image 25H obtained by multi-resolution conversion of the bone image 25K. Supervised learning of boundary high-frequency band image 25CH, boundary high-frequency band image 25CM, and boundary low-frequency band image 25CL obtained by reducing image 25HH, high-voltage medium-frequency band image 25HM, high-voltage low-frequency band image 25HL, and part specifying image 25C To the completed filter 41.

  The teacher-learned filters 41H, 41M, and 41L to which the images for the respective spatial frequency bands of the bone part image 25K, the high-pressure image 25H, and the region specifying image 25C are input are diagnostic radiation for each spatial frequency band. The images 61H, 61M, and 61L are estimated, and the estimated radiation images 61H, 61M, and 61L are synthesized by the image synthesis filter 41T to obtain a diagnostic radiation image 61.

  That is, when the bone high-frequency band image 25KH, the high-pressure high-frequency band image 25HH, and the boundary high-frequency band image 25CH are input to the high-frequency band supervised learned filter 41H, the image quality deterioration is compensated and the bone that is the specific part A radiographic image 61H for diagnosis in a high frequency band mainly representing the portion is formed.

  In addition, when the bone intermediate frequency band image 25KM, the high-voltage intermediate frequency band image 25HM, and the boundary intermediate frequency band image 25CM are input to the medium frequency band teacher learned filter 41M, the image quality deterioration is compensated for, and the specific A radiographic image 61M for diagnosis in the medium frequency band mainly representing the bone part as a part is formed.

  Furthermore, when the bone low-frequency band image 25KL, the high-voltage low-frequency band image 25HL, and the boundary low-frequency band image 25CL are input to the low-frequency band teacher-learned filter 41L, the image quality degradation is compensated for and A diagnostic radiographic image 61L in the low frequency band mainly representing the bone portion, which is a part of the above, is formed.

  The formed high-frequency band diagnostic radiation image 61H, medium-frequency band diagnostic radiation image 61M, and low-frequency band diagnostic radiation image 61L are combined by the image combining filter 41T to compensate for image quality degradation. And the said diagnostic radiation image 61 which mainly represents the bone part which is a specific site | part is created.

  FIG. 11 is a diagram illustrating how upsampling and addition are performed in the image synthesis filter.

  As shown in FIG. 11, the image synthesis filter 41T repeats upsampling and addition in the order of the diagnostic radiographic image 61L in the low frequency band, the diagnostic radiographic image 61M in the medium frequency band, and the diagnostic radiographic image 61H in the high frequency band. The diagnostic radiation image 61 is obtained.

  That is, an image obtained by upsampling the diagnostic radiographic image 61L in the low frequency band and the diagnostic radiographic image 61M in the medium frequency band is obtained, and an image obtained by upsampling the image is obtained. The diagnostic radiographic image 61 is obtained by adding the diagnostic radiographic image 61H in the high frequency band.

  The feature amount input in the learning of the regression model will be specifically described below. FIG. 12 is a diagram illustrating an example of a region indicating the feature amount.

  The feature amount may be a pixel value itself in a radiation image for each spatial frequency band, or may be obtained by performing a special filter process. For example, as shown in FIG. 12, each pixel value of a region U1 consisting of three pixels adjacent to each other in the vertical direction or a region U2 consisting of three pixels adjacent to each other in the horizontal direction in a radiographic image in a specific spatial frequency band. The average value may be a new feature amount. In addition, wavelet transform may be performed and wavelet coefficients may be used as feature amounts. Further, pixels over a plurality of frequency bands may be used for the feature amount.

  The contrast normalization performed in the learning of the regression model will be described below.

  A standard deviation is calculated for the pixel value of each pixel included in the sub-window Sw (see FIG. 9) set in the image for each spatial frequency band. The pixel value of the band image is multiplied by a coefficient so that the standard deviation matches a predetermined target value.

I '= I × (C / SD)
Here, I is the pixel value of the original image, I ′ is the pixel value after contrast normalization, SD is the standard deviation for the pixel value in the sub-window Sw, C is the target value of the standard deviation (the above C is set in advance) Constant).

  The sub-window Sw is scanned so as to cover the entire region in each radiation image, and the pixel value in the sub-window Sw is set so that the standard deviation approaches the target value for all sub-windows Sw that can be set in each image. Normalize by multiplying by a predetermined coefficient.

  As a result of the normalization, the magnitudes (contrasts) of the amplitudes of the image components for each spatial frequency band are aligned. Thereby, since the variation of the image pattern in the radiographic image for every spatial frequency band input into the said teacher learning filter 41 reduces, the effect which improves the estimation precision of the said bone part is acquired.

  In the step of learning the supervised learned filter that is the non-linear filter, the high-pressure image is subjected to the above-described contrast normalization process, and the multiplied coefficient is also multiplied to the bone part image without deterioration. A learning sample is prepared from a pair of normalized high-pressure image and bone part image to learn a nonlinear filter.

  In the step of estimating the diagnostic radiographic image mainly representing the bone part of the diagnostic object, the input high-voltage image is subjected to contrast normalization, and the pixel value of each normalized spatial frequency band is set as a teacher-learned filter. To enter. The output value of the teacher-learned filter is multiplied by the reciprocal of the coefficient used when the normalization is performed to obtain an estimated value of the bone.

  In addition, as described above, one image is converted into a plurality of images for each different spatial frequency band, and image processing is performed for each converted image to create a plurality of processed images for each spatial frequency band. As a technique for combining the plurality of processed images to obtain one processed image, various conventionally known techniques can be employed.

  Next, support vector regression (regression by support vector machine (SVR)) will be described.

  FIG. 13 is a diagram showing how an approximate function is obtained by support vector regression.

Regarding the problem of learning a function that approximates a real value y corresponding to a d-dimensional input vector x, first consider the case where the approximation function is linear.

  The ε-SVR algorithm proposed by Vapnik finds f that minimizes the following loss function.

For the ε-SVR algorithm proposed by Vapnik, the above document “Introduction to Support Vector Machine” can be referred to.

<W · w> is a term representing the complexity of a model that approximates data, and Remp [f] is expressed as follows.

  Here, | y−f (x) | ε = max {0, | y−f (x) | −ε}, which means that errors smaller than ε are ignored. ξ and ξ * are relaxation variables that allow an error exceeding ε in the positive direction and the negative direction, respectively. C is a parameter that sets a trade-off between the complexity of the model and relaxation of constraints.

The above main problem is equivalent to solving the following dual problem, and a global solution can always be obtained from the characteristics of the convex quadratic programming problem.

The regression model obtained by solving this dual problem is expressed by the following equation.

This function is a linear function, but in order to extend it nonlinearly, if the input X is mapped to a higher-order feature space Φ (X) and the vector Φ (X) in that feature space is regarded as the previous input X, Good (X → Φ (X)). In general, mapping to a high-dimensional space is accompanied by a significant increase in computational complexity, but the inner product terms appearing in the formula to be optimized are expressed as K (x, y) = <Φ (x), Φ (y)>. If it is replaced with a kernel function that satisfies, the same result as that calculated after mapping to a higher dimension in the calculation of the input dimension can be obtained. As the kernel function, an RBF kernel, a polynomial kernel, a sigmoid kernel, or the like can be used.

  As a learning method in the discrimination, AdaBoost or the like may be used in addition to the support vector machine (SVM).

  The classes to be identified are not limited to two classes consisting of bones and parts other than bone parts, posterior ribs, and ribs, but three classes including posterior ribs, ribs, and clavicles, and more than three classes including clavicles. It may be.

  FIG. 14 is a diagram showing motion artifacts that occur in the bone image representing the chest.

  As shown in FIG. 14, in a bone part image FK representing an adult female breast that is an energy subtraction image formed by weighted subtraction using a high-pressure image and a low-pressure image obtained by radiation imaging with different energy distributions. In some cases, a motion artifact Ma generated in response to the heartbeat may occur. Since such motion artifacts are required to be removed from the radiographic image, the specific part is set as a motion artifact, and a radiographic image representing the motion artifact Ma is formed through the supervised learned filter. By subtracting the radiographic image thus formed from, for example, the bone image FK, a bone image from which the motion artifact Ma has been removed can be created.

  As described above, the specific part can be a part where the display position in the subject has changed in the high-pressure image and the low-pressure image obtained at different timings. In addition, as described above, the specific part that is emphasized may be an unnecessary part (defect part) in the radiographic image. In such a case, by subtracting the radiation image representing the unnecessary part from the radiation image including both the unnecessary part and the necessary part other than the unnecessary part, the unnecessary part is removed and the desired image representing only the necessary part is represented. A radiographic image can be obtained.

  In addition, when acquiring the radiographic image for teachers based on radiography using a dose larger than the dose of radiation used for one radiography when obtaining an input radiographic image, one radiography In some cases, the radiation dose to the subject may exceed the permissible value. However, by limiting the sum of the radiation doses to the subject for a predetermined period, etc. Radiography can be performed.

  As shown in FIGS. 1 and 2, the radiographic image processing apparatus 110 that performs the radiographic image processing method of the present invention obtains simple radiography of the chest 1P obtained for each chest 1P of an adult male that is a plurality of similar subjects. Subject image 11H for learning, which is a simple radiation image representing the chest 1P obtained in step 10, and a bone portion that is a specific part in the chest 1P obtained by subjecting the subject image 11H to boundary extraction processing 12 An input radiation image 11 composed of a learning part specifying image 11C representing a boundary Pc between Px and another part Po different from the bone part Px, and the subject image 11H obtained by radiography of the chest 1P. The teacher radiation image 33 is used for the input radiation image 11 and the teacher radiation image 33 is expressed by emphasizing the bone portion Px, which is a specific part in the chest 1P. A filter acquisition unit Mh1 (see FIG. 1) that obtains a teacher-learned filter 40 that has learned the radiographic image 33 as a teacher, and a chest 3P that is a subject to be diagnosed that is the same type of subject as the chest 1P that is a given subject. The same-type image creation unit Mh2 (see FIG. 2) that performs radiography 20 and creates a radiation image 21 of the same type as the input radiation image 11, and inputs the diagnosis-target radiation image 21 to the teacher-learned filter 40. And a region-enhanced image forming unit Mh3 (see FIG. 2) that forms a diagnostic radiographic image that emphasizes and represents the bone portion Px in the given chest 3P.

  Since the operation of the radiographic image processing apparatus 110 is the same as that of the radiographic image processing method already described, a description thereof will be omitted. Note that each image handled by the filter acquisition unit Mh1, the same kind of image creation unit Mh2, and the part-enhanced image formation unit Mh3 can be an image itself or image data representing an image.

  Note that the teacher-learned filter is not learned for each small region, but only one type is prepared for each frequency, and all small regions are processed by one filter. In the learning of the filter, learning samples are taken out from various small regions of a single (or a small number) of radiographic images, and these many samples are treated as a set at the same time. That is, for example, learning samples consisting of Mr. A's collarbone, Mr. A's collarbone, Mr. A's rib contour, Mr. A's rib center, etc. are collectively learned. In addition, although the feature quantity to be input to the filter is 25 pixels, the teacher who is the output corresponding to the 25 pixels is not 25 pixels but one pixel at the center of the small area.

  Further, a program for executing the function of the radiation image processing apparatus of the present invention can be installed in a personal computer, and the same operation as in the above embodiment can be executed in the personal computer. That is, a program for causing a computer to execute the radiographic image processing method of the above embodiment corresponds to the program of the present invention.

The figure which shows the procedure which acquires the teacher learning completed filter used for the radiographic image processing method of 1st Embodiment. The figure which shows the procedure of the radiographic image processing method of 1st Embodiment The figure which shows the procedure which acquires the teacher learning completed filter used for the radiographic image processing method of 2nd Embodiment. The figure which shows the procedure of the radiographic image processing method of 2nd Embodiment. Diagram showing the boundary extraction process Diagram showing discrimination of two classes by support vector machine The figure which shows a mode that a subwindow is set to the radiographic image and teacher image which should be distinguished The figure which shows a mode that the radiographic image for every spatial frequency band about the given subject is inputted into a teacher learned filter, and the radiographic image for diagnosis is created. The figure which shows a mode that a teacher learned filter is produced for every spatial frequency band The figure which shows a mode that multi-resolution conversion of the image The figure which shows the state of upsampling and addition in the image composition filter The figure which shows the field which constitutes the feature quantity Diagram showing how approximate function is calculated by support vector regression It is a figure which shows the motion artifact which arose in the bone part image showing a chest part.

Explanation of symbols

1P Subject 3P Subject 10 Radiography 11H Subject image 11C Site specific image 11 Radiation image for input 20 Radiography 21 Radiation image 33 Radiation image for teacher 40 Teacher-trained filter 60 Radiation image

Claims (13)

For each of a plurality of the same type of subjects, a part specifying image representing a boundary between a specific part in the subject obtained by radiography of the subject and another part different from the specific part, and a subject image representing the subject Prepare an input radiation image consisting of
Prepared for each subject is a teacher radiographic image obtained by radiographing each subject, representing the specific portion of the subject in an emphasized manner,
Obtaining a teacher-learned filter that trains the teacher radiation image corresponding to the subject as a teacher with respect to the input radiation image representing each subject,
Then, for a given subject of the same type as the subject, create a radiation image of the same type as the input radiation image,
The radiographic image processing method, wherein the radiographic image is input to the teacher-learned filter to form a radiographic image of the subject in which the same part as the specific part in the given subject is emphasized .   2. The radiation according to claim 1, wherein a radiation dose used for radiography for creating the teacher radiographic image is larger than a radiation dose used for radiography for creating the subject image. Image processing method.   3. The teacher radiation image according to claim 1, wherein the teacher radiation image is formed by a weighted subtraction process using a high-pressure image and a low-pressure image obtained by radiation imaging with radiation having different energy distributions. Radiation image processing method.   4. The input radiation image is formed by weighted subtraction processing using a high-pressure image and a low-pressure image obtained by radiation imaging with radiation having different energy distributions from each other. 5. A radiation image processing method according to claim 1.   The radiographic image processing method according to claim 1, wherein the subject image is a simple radiographic image obtained by simple radiography.   The radiographic image processing method according to claim 1, wherein the specific part is a part having specific radiation absorption characteristics different from other parts.   7. The object according to claim 1, wherein the subject is a living tissue, and the specific part includes at least one of a bone part, a rib, a posterior rib, an anterior rib, a clavicle, and a spine. The radiation image processing method according to claim 1.   8. The object according to claim 1, wherein the subject is a living tissue, and another part different from the specific part includes at least one of a lung field, a mediastinum, a diaphragm, and a rib. The radiographic image processing method of any one of Claims 1.   The radiographic image processing method according to claim 1, wherein the subject is a biological tissue, and the specific part is a bone part or a soft part of the biological tissue.   The radiographic image processing method according to claim 4, wherein the specific part is a part whose position in the subject has changed between the high-pressure image and the low-pressure image.   The teacher-learned filter performs learning for obtaining the teacher-learned filter and formation of a radiation image for the given subject for each of a plurality of different spatial frequency bands, and is formed for each spatial frequency band. The radiographic image processing method according to claim 1, wherein each radiographic image is synthesized to obtain one radiographic image. An input comprising a part specifying image representing a boundary between a specific part in the subject and another part different from the specific part, and a subject image representing the subject, obtained by radiography for a plurality of subjects of the same type An input radiographic image representing each subject using the radiographic image for each subject and the radiographic image for each subject representing the particular portion in the subject emphasized by radiography of the subject. Filter acquisition means for obtaining a teacher-learned filter in which the teacher radiograph corresponding to the subject is learned as a teacher in response to the input;
For a given subject of the same type as the subject, the same kind of image creating means for creating a radiation image of the same kind as the input radiation image;
A site-enhanced image forming unit configured to input the radiographic image to the teacher-learned filter and form a radiographic image of the subject in which the same site as the specific site in the given subject is enhanced A radiographic image processing apparatus. An input comprising a part specifying image representing a boundary between a specific part in the subject and another part different from the specific part, and a subject image representing the subject, obtained by radiography for a plurality of subjects of the same type An input radiographic image representing each subject using the radiographic image for each subject and the radiographic image for each subject representing the particular portion in the subject emphasized by radiography of the subject. A procedure for obtaining a teacher-learned filter for learning the teacher radiation image corresponding to the subject as a teacher in response to the input;
Creating a radiation image of the same type as the input radiation image for a given subject of the same type as the subject;
A radiographic image processing method for inputting the radiographic image to the teacher-learned filter and executing a procedure for forming a radiographic image of the subject in which the same portion as the specific portion in the given subject is emphasized A program that causes a computer to execute.