JP2008161693A - Image processor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image processor which can effectively contribute to image processing technique to sufficiently suppress noise without blurring of image, for instance, pattern recognizing technique etc. <P>SOLUTION: The image processor is to judge similarity between first pixels and second pixels constituting an image by a test. The image processor is equipped with a calculation part 112 to carry out the judgement from a plurality of images different in time direction including the image and a calculation part 113 to calculate the weighted average of the first pixels and the second pixels on the basis of the judgement result. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image processing apparatus, and more particularly, to the same method and apparatus capable of reducing noise existing on an image.

  Currently, image processing techniques are used in various fields.

  Image processing is performed to deal with degradation or modification of images acquired by video tape recorders, digital cameras, etc., or to check whether structures are manufactured as designed. This is done for the purpose of clearly grasping the pattern and structure itself.

  Various image processing is also performed in various medical image diagnostic apparatuses such as an X-ray CT apparatus, a SPECT apparatus, and an MRI apparatus. The utility of drawing blood flow or contrast medium flow, or extracting lesions or contours of organs is widely recognized.

  The image processing technology includes various elemental technologies such as a noise suppression technology, a feature extraction technology, and a pattern recognition technology, and each technology is used alone or in appropriate combination. Of these elemental technologies, in particular, a technology for reducing random noise included in an image is an indispensable technology for more clearly reproducing an imaged or reconstructed object.

However, further improvements are required for conventional image processing techniques, particularly noise reduction techniques. For example, so-called “smoothing” is widely known as a noise reduction technique. In this smoothing, when there is an input value f (i, j) for a certain pixel (i, j), an average density in the vicinity of the pixel (i, j) is output for the pixel (i, j). The value is g (i, j). Specifically, if n × n pixels in the vicinity of the pixel (i, j) are used, the output value g (i, j) is

As required. However, a, b, c, and d in the above formula (1) are integers. Further, 1 / (b−a + 1) (d−c + 1) in the above equation (1) is a so-called weight. FIG. 13 shows a case where a, b, c, d = −1, 1, −1, 1.

Meanwhile Generally, the distribution of the population, such as variance is sigma ^2, when calculating the average value of n samples taken independently, the variance of the average value is known to be a sigma ^{2 /} n Therefore, according to the above equation (1), the “population” and the “distribution σ ² ” referred to above are components of noise values included in the values of the respective pixels (i, j). Therefore, the contribution of noise in the value f (i, j) of each pixel can be reduced.

However, by simply applying this, so-called “edge blur” occurs, the spatial resolution of the image is impaired, and the entire image becomes blurred. Taking the above-described medical image as an example, even when it is desired to depict a fine vascular structure with as little noise as possible, the vascular structure is not originally depicted according to the noise suppression processing according to the above equation (1). Since averaging (smoothing) is performed including the pixels, even if noise is suppressed, the contrast representing the blood vessel structure is also reduced by the smoothing, making it difficult to draw a detailed blood vessel structure There is.

  The present invention has been made in view of the above circumstances, and its object is to sufficiently suppress noise without causing blurring of the image, and other image processing techniques such as patterns. An object of the present invention is to provide an image processing apparatus that can contribute effectively to recognition technology and the like.

  The present invention is to determine the similarity between the first pixel and the second pixel constituting the image by a test, and a determination unit that determines from a plurality of images different in the time direction including the image; An image processing apparatus is provided, comprising: an averaging unit that weights and averages the first pixel and the second pixel based on a determination result by the determination unit.

  According to the present invention, it is possible to effectively contribute to other image processing techniques such as a pattern recognition technique as well as noise can be sufficiently suppressed without causing image blurring.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
In the present embodiment, first, a pixel value that is a vector value or a scalar value is configured for a pixel constituting an image, and the degree of fitness between the pixel value and another pixel value that is configured separately from the pixel value is quantified. When the new pixel value for the pixel is configured using the other pixel value, if the fitness is large, the contribution of the other pixel is increased and the fitness is small. Is an image processing method characterized in that the new pixel value is configured by reducing the contribution of the other pixels. In the present invention, the “other pixel values configured separately” may be configured based on other pixels configuring the image. That is, in this case, the pixel values are configured as “one pixel value” and “other pixel value” which are vector values or scalar values, respectively, for one pixel and another pixel. Furthermore, the present invention particularly determines a weight of each of the other pixel values by applying a weight function that is a function of the fitness to the fitness obtained for each of the other pixel values. By calculating a weighted average of the other pixel values using this weight, the new pixel value is configured by increasing the weight when the fitness is large, The contribution of the pixel value in the weighted average is increased, and when the fitness is small, the weight is decreased, thereby reducing the contribution of the other pixel value in the weighted average, and the new pixel It is also part of the value. According to such processing, since other pixels determined to be “similar” to one pixel are emphasized and a new pixel value is formed, the spatial resolution is reduced as in the conventional case. There is no loss of time resolution when the image is a moving image. (Weight Function and Weight) In the present invention, the weight function is preferably a non-negative monotonically increasing function related to the fitness. Further, the one pixel and the other pixel are x and y, respectively, and the one pixel value and the other pixel value are respectively v (x) = (v ₁ (x), v ₂ (x),. v _K (x)) and v (y) = (v ₁ (y), v ₂ (y),..., v _K (y)), and the new pixel value to be constructed is v ′ (x ) = (V ′ ₁ (x), v ′ ₂ (x),..., V ′ _K (x))

(However, N (x) represents a range including the other pixels.)
(Hereinafter, a form such as the right side of the above equation for obtaining a new pixel value v ′ (x) will be referred to as a “coherent filter according to the present embodiment”).

(Example of image processing technology applicable to this embodiment)
The present invention reduces the noise of the image by configuring the above-described processing, that is, the new pixel value, for example, with respect to the entire surface of the image, or also performs pattern recognition in the image. The present invention exhibits excellent effects in each element of these image processing techniques as will be described later.

(Fitness)
In the present embodiment, the fitness can be quantified based on a risk factor obtained as a result of applying a statistical test method to scalar values constituting the one pixel value and the other pixel values, respectively. It may be. As used herein, the “statistical test method” may be, for example, a “chi-square test method” or the like as will be described later. In general, the relationship between the risk factor introduced here and the degree of fitness particularly includes the case where one increases and the other decreases. That is, when the new pixel value is configured, if the risk factor is large (the fitness level is small), the weight for the other pixel is decreased (the contribution to the new pixel value to be configured is small), This is a case in which the weight for the other pixels is increased (the contribution to the new pixel value to be configured is increased) if the risk factor is small (the fitness level is large).

(Selection range of “Other pixels”)
In the present embodiment, the other pixel is selected from a predetermined region around the one pixel, and in particular, the predetermined region is a neighborhood centered on the one pixel. Good. When configured in this manner, the other pixels that are determined to be similar to the one pixel usually have a high possibility of being present around the one pixel. When the new pixel value is obtained effectively, useless calculation can be omitted.

(Pixel value configuration method)
In the present embodiment, it is possible to select various configuration methods of the pixel values (referring to both one pixel value and another pixel value). More specifically, first, a plurality of images are included. In the case of a single image, the pixel value can be a vector value in which scalar values of pixels at the same point through each of the plurality of images are arranged.

  At this time, the plurality of images may be a plurality of still images constituting a moving image, and in such a case, the present invention does not impair spatial resolution when obtaining the new pixel value. In addition, the time resolution is not impaired. A suitable application example in such a case is, for example, a case where the moving image is a dynamic CT image acquired by a medical image diagnostic apparatus.

  In addition, as a second pixel value configuration method, the image is a plurality of types of images related to the same subject, and the scalar value of pixels at the same point through each of the plurality of types of images is the same as described above. Can be used as a vector value. In addition, a suitable application example in such a case is, for example, an image obtained by a multi-window imaging method in a SPECT apparatus or a PET apparatus, and the plurality of types of images are different from radioisotopes, respectively. This is the case when the images are different from each other based on the emitted gamma rays. Another suitable application example is, for example, a case where the image is a color image, and the plurality of types of images are different images obtained by being separated into the three primary colors of light.

  Furthermore, as a third pixel value configuration method, the image is a single image, and the pixel value can be configured from the single image. More specifically, for example, the pixel value can be configured as a vector value in which a scalar value possessed by a pixel on the one image and a scalar value possessed by a pixel in the vicinity of the pixel are arranged. is there. Note that according to such a pixel value configuration method, the present invention can be applied to any digital image.

(Suitable apparatus configuration for realizing the image processing of the present embodiment)
Finally, as an apparatus configuration for realizing the above-described image processing method, a degree of fitness between a pixel value configured as a vector value or a scalar value and another pixel value configured separately for each pixel configuring an image. When the adaptability quantifying means for determining the new pixel value for the pixel is configured using the other pixel value, if the adaptability is large, the contribution of the other pixel is increased. However, when the degree of adaptation is small, it is preferable to include pixel value calculation means for reducing the contribution of the other pixels and forming the new pixel value.

  This will be described in detail below. In the following, first, a more general description (item number “I”; the same applies hereinafter) of the outline of the “coherent filter” according to the present invention will be given, and then the coherent filter is applied to various image processing. (Item numbers “II” and “III” to “VIII”) will be described sequentially.

Further, following the description of such various application examples, the most general form of the coherent filter in the present embodiment and its application example (item number “IX”) will be described. Supplementary items (item number “X”) will be described together.

(I General description of coherent filter according to the present invention)
First, a preliminary explanation is given for explaining the “coherent filter” according to the present invention.

(I-1 Pixel value v (x))
In general, a digital image acquired via an imaging means such as a camera is composed of a plurality of pixels (or the image can be considered as a set of such pixels). In the following description, the position of the pixel is represented as a vector x (that is, a coordinate value vector), and a value (for example, a numerical value representing light and shade) of the pixel x is represented as a K-dimensional vector. In the case of a two-dimensional image, the pixel x is a two-dimensional vector indicating coordinate values (x, y) representing a position on the image. The “pixel value v (x)” defined for a pixel x is

Is written. In the following, each of v ₁ (x), v ₂ (x),..., V _K (x) on the right side of the equation (3) will be referred to as a “scalar value” for the pixel x.

For example, when the image is a “color image”, each pixel has brightness (scalar value) of three primary colors (red, green, and blue), and therefore the pixel value v (x) of each pixel is It can be considered that the dimension is a vector of K = 3 (assuming that the subscripts are “red”, “green”, and “blue” in each term on the right side of the above equation (3), for example. (See equation (3.1) below.) For example, if the image is a moving image composed of K still images and each pixel of the nth image has a scalar value v _n (x), K-dimensional vector values v _n (x) = (v ₁ (x), v ₂ (x),..., Arranged by arranging pixel values (scalar values) of pixels x at the same common point (same coordinates). v _K (x)) is a pixel value as a vector value described below.

(I-2 Conformity or risk factor p (x, y) and weight w (p (x, y)))
Consider an appropriate pixel set N (x) for the pixel x (this set N (x) may include the pixel x). Next, a weight w (p (x, y)) is considered between each pixel y which is an element of N (x) and the pixel x. The weight w (p (x, y)) has the following properties.

(I-2-1 Conformity or risk factor p (x, y))
First, the meaning of the function p (x, y) that affects the value of w (p (x, y)) will be described. This p (x, y) is a means for quantifying the “fitness” referred to in the present invention. Generally speaking, how similar the pixel x and the pixel y∈N (x) are in some sense. A specific numerical value indicating whether or not (for example, the degree of statistical difference recognized between the pixel values v (x) and v (y) of both pixels x and y) is given.

  More specifically, for example, when p (x, y) gives a small value, the pixel x and the pixel y have a “statistically significant difference between the pixel values v (x) and v (y)”. Is not (= high fitness) ”, and when p (x, y) gives a large value when it is determined that there is a high possibility of similarity,“ there is a statistically significant difference (= small fitness) ” "Is judged as follows.

By the way, the pixel values v (x) and v (y) (or scalar values v ₁ (x),..., V _K (x) and v ₁ (y),..., V _K (y)) are always noise. Must be considered to be included. For example, when considering a case where an image is acquired by a CCD image pickup device, each pixel constituting the image includes noise or the like due to dark current in the device or irregular fluctuation of the amount of light incident from the outside.

  Since such noise generally takes various values for all pixels, even if the pixel x and the pixel y reflect the same object (in the outside world), they are actually observed. On the image, it may not have the same value. In other words, if it is assumed that the respective noises are removed from the pixel x and the pixel y reflecting the same object, they are displayed on the image as representing the same object. (= Recognized as such) and both have essentially the same (or very close) pixel values.

  Therefore, based on the above-mentioned noise characteristics, regarding the above p (x, y), using the concept of the “null hypothesis” that is well known in the statistical test method, this p (x, y) Specifically, we can say as follows. In other words, the null hypothesis H “pixel x and pixel y have the same pixel value when each noise is removed”, in other words, “v (x) = v (y). Excluding the difference caused by (that is, when such a proposition holds, it is considered that “both pixels x and y are similar (= high adaptability)”), and the function p (x, y ) Is a risk factor (or significance level) when rejecting this hypothesis H (in this case, p (x, y) is a function whose value range is [0, 1]). (P (x, y) ε [0,1])).

  Therefore, when the risk factor p (x, y) is large, that is, when there is a high risk of rejection being incorrect, it can be said that the above hypothesis H is likely to be satisfied, and conversely, when it is small, that is, rejection is incorrect. If the risk is small, it can be said that there is a high possibility that the hypothesis H is not satisfied (note that although it is a well-known matter in the statistical test, the hypothesis H is not “rejected”, it is “true”. (In this case, it simply means that the proposition indicated by hypothesis H cannot be denied.)

(I-2-2 Weight w (p (x, y)))
The weight w (p (x, y)) is a function of the risk factor p (x, y) as described above (more generally, a function of fitness (the fitness is If ρ (x, y), it can be configured to be w (ρ (x, y)), and in order to obtain this weight w (p (x, y)), x and y Generally speaking, the weighting function w to be applied to the risk factor p (x, y) obtained for each combination has the effect of embodying the “rejection.” Specifically, the risk factor. When p (x, y) is large, the value of the weighting function w, that is, the weight w (p (x, y)) takes a large positive value, and vice versa, a small positive value (or “0”). ")", Etc. (The specific form of the weight function w will be described later.) That is, the weight w (p (X, y)) takes a large value when the pixel x and the pixel y seem to satisfy the proposition shown in the above hypothesis H, and takes a small value in the opposite case. It may be configured so that there are only two possible values of w, “0” or a fixed value other than “0”.

  The relationship among the hypothesis H, the risk factor p (x, y), and the weight w (p (x, y)) described above is collectively shown in FIG. Further, the weight function w (t) can be more generally referred to as “a non-negative monotonically increasing function defined by t∈ [0, 1]”, and the property to be satisfied by w (t) is at least That is all that is necessary.

(I-3 Coherent filter)
By the preliminary explanation so far, the “coherent filter” according to the present invention is derived as follows. That is, first, the weight w (p (x, y)) described above is calculated for all the pixels y that are elements of the set N (x) for a certain pixel x constituting the image. Next, a new scalar value v ′ _k (x) constituting the pixel x is calculated by the following expression using the plurality of weights w (p (x, y)). That is,

However, k = 1, 2,..., K. Then, using v ′ _k (x) obtained by this equation, the pixel value (new pixel value) v ′ (x) after the conversion of the pixel x is

Configure as.

Here, the case where the pixel value v (y) = (v ₁ (y), v ₂ (y),..., V _K (y)) (y = x) expressed by the above equation (4) is included. )) To v ′ (x) = (v ′ ₁ (x), v ′ ₂ (x),..., V ′ _K (x)) is a filter of the “coherent filter” according to the present invention. This is the format in this embodiment. As is apparent from the expression, this represents a weighted average value of the scalar values v _k (y) constituting the pixel values.

Such processing yields the following results. That is, the pixel value v ′ (x) is likely to have the same pixel value as the pixel x except for noise (= the weighted average value that emphasizes the pixel y that is likely to satisfy the proposition of the above hypothesis H). It represents a vector constructed from v ′ _k (x). Further, if there are a sufficient number of such pixels y, the pixel value v ′ (x) is not deviated from its true value that the pixel x should originally have, and noise is generated by the above-described averaging operation. It has the value which suppressed only.

  Even if the risk factor p (x, y) is small and the null hypothesis H is “rejected” and the weight w (p (x, y)) is small, the above description or ( As can be seen from the expression of 3), this is not necessarily completely “rejected”. Although this depends on the specific form of the weight function w described later, even when the risk factor p (x, y) is close to “0” (= 0%), w (p ( x, y)) ≠ 0 (where p (x, y) is a smaller positive value than when p (x, y) is close to “1”) (p (x, y) = 1) (Some cases are when v (x) = v (y), as will be described later.) In other words, it is not a complete rejection but a small contribution may be accepted (in this case, if w (p (x, y)) = 0, a complete rejection is performed. (See formula (14) below).

  Such processing can be generally described as follows. That is, when there are (a plurality of) pixels x constituting an image, the degree of matching between this pixel x and an arbitrary pixel y (in the above, yεN (x)) is quantified (above Is based on p (x, y).) When the degree of matching is large, a large contribution is recognized for the pixel y in the weighted averaging process using the pixel value v (y). It can be said that the image processing method effectively suppresses noise of the pixel x by allowing only a small contribution when the degree is small. In other words, when the pixel x and the pixel y are “similar”, the pixel y contributes more to the averaging process. When the pixel x and the pixel y are “similar”, the pixel y is almost or completely ignored. In other words,

  By applying such processing to the entire image, an extremely high noise suppression effect can be exhibited with almost no blurring of the image. Further, the present invention is not limited to the use of noise suppression. For example, also in the field of pattern recognition, an excellent effect can be exhibited by making a weighting function or a coherent filter into a suitable specific form. The specific effects of these effects will be described below in the description of various image processing.

(II Various application examples: Classification based on differences in construction method of pixel value v (x))
The present invention can be applied to various fields based on the general form of the coherent filter described above. And this can be summarized, for example, as shown in FIG. FIG. 2 shows various application examples in various fields classified according to the difference in the construction method of the pixel value v (x). That is, according to FIG. 2, the pixel value v (x) as the vector value described above is “from a plurality of images”, “from a plurality of types of images” (for example, obtained by one shooting or the like), or “one image It can be seen that it can be configured by each method of “from the image of” and various application examples relating to the present embodiment are classified into these respectively.

  Hereinafter, according to the classification table shown in FIG. 2, various application examples in which the coherent filter having the general form as described above is applied in a more specific situation, and three kinds shown in FIG. A specific description will be given of how the pixel value v (x) construction method is specifically realized (however, the order is different from that shown in FIG. 2).

(III When applying the present invention to X-ray CT imaging 1; Dynamic CT imaging (removing noise from a plurality of images))
First, the case where the coherent filter is applied to so-called “dynamic CT” imaging in an X-ray CT apparatus will be described. Here, as shown in FIG. 3, the X-ray CT apparatus 100 includes an X-ray tube 101 and an X-ray detector 102, a data acquisition unit 103 (DAS; Data Acquisition System), a preprocessing unit 104, a memory unit 105, It comprises a reconstruction unit 106, an image display unit 107, a control unit 108 that controls these units, an input unit 109 that can perform various settings and inputs such as X-ray irradiation conditions and imaging modes.

  According to the X-ray CT apparatus 100, the X-ray tube 101 and the X-ray detector 102 are rotated around the subject P, and the X-ray is exposed to the subject P, whereby the state inside the subject P is obtained. Can be observed as other CT images such as a tomographic image.

  At this time, each of the above-described components generally operates as follows. That is, the X-rays emitted from the X-ray tube 101 by the voltage application of the high voltage apparatus 101a and transmitted through the subject P are converted into analog electric signals by the X-ray detector 102, and hereinafter, digital conversion processing by the data collection unit 103 is performed. In response to various correction processes by the pre-processing unit 104, the data is stored in the memory unit 105 as projection data. Then, based on the projection data obtained by the X-ray tube 101 and the X-ray detector 102, for example, once around the subject P (= 360 °), other CTs such as tomograms by the reconstruction unit 106 are obtained. The image is reconstructed, and the image display unit 107 displays the CT image. It should be noted that the reconstructed CT image can be stored in the storage device 10M.

  Here, the “dynamic CT” imaging means that the X-ray tube 101 and the X-ray detector 102 repeatedly image the same part of the subject P (repetitive scanning. In a continuous rotation CT apparatus, repeated imaging by continuous rotation is performed. This is an imaging method in which projection data is acquired one after another, and reconstruction processing is performed one after another based on the projection data to obtain a series of time-series images (in this case, images The image display on the display unit 107 is controlled so as to be performed after a predetermined time from the scan start point or end point related to the collection of projection data that is the source of the image, for example, by a counter (not shown).

  Therefore, the image acquired / displayed in this way is a so-called moving image composed of a plurality of time-series still images as in a movie or the like. Note that such an imaging method typically injects a contrast medium into the subject P, observes and analyzes changes over time, and analyzes other pathological conditions such as stenosis and occlusion in blood vessels, for example. Used for. A method of performing CT imaging of the same part only twice before and after contrast medium administration can also be considered as dynamic CT imaging in a broad sense.

  Conventionally, at the time of “dynamic CT” imaging as described above, for example, some change in the subject P (for example, a change in the concentration of a contrast medium, respiratory movement, etc.) is generally considered while K imaging is performed. In order to suppress the image noise without impairing the spatial resolution, there is no choice but to smooth in the time direction. As a result, the adverse effect that the time resolution is impaired cannot be avoided.

However, as described above, an image acquired by dynamic CT imaging is a moving image and is performed for the purpose of observing temporal changes in detail, so that the temporal resolution is originally lost. This is not a favorable situation.

  If the coherent filter according to the present invention is used, the following processing (hereinafter referred to as “a”) that can suppress noise for all (a plurality of images) of K still images without impairing the time resolution. Dynamic coherent filtering ").

First, for the pixel x defined for the K still images, which are the moving images obtained as described above, as described above, as the pixel value v (x),

Can be configured. Here, the subscripts 1, 2,..., K in each term on the right side are numbers assigned through each of the K still images (see the description of the above-described equation (3)).

Next, a specific form of the weight function w1 in this case is given by the following equation, for example.

  However, yεN (x) and the set N (x) may be arbitrarily set for each pixel x (= may be set according to any criterion). However, in practice, the pixel x and the pixel y far away from the pixel x can satisfy the hypothesis “v (x) = v (y). However, excluding differences caused by noise between the two pixels”. Therefore, limiting the set N (x) on the basis of a set of pixels close to x has practical significance such as an increase in calculation speed.

  Therefore, here, as an example, the set N (x) is a set of pixels included in the surrounding rectangular area centered on the pixel x. More specifically, as the set N (x), for example, when all pixels constituting one still image of interest are 128 × 128 pixels, 3 × It may be an area for 3 pixels, or in the case of 512 × 512 pixels, it may be an area for 13 × 13 pixels centered on the pixel x.

In addition, σ _k in the above equation (6) is a noise standard deviation estimated on the assumption that each pixel of the kth still image has a certain degree common to all the pixels. C is a parameter that can determine and adjust the degree of action when the weight w1 (p (x, y)) is substituted into the above equation (4).

Hereinafter, these σ _k and C will be described in order.

First, σ _{k in the} equation (6) will be described (hereinafter, described as variance σ _k ² ). As described above, σ _k ² is the variance of the noise component of the scalar value of each pixel on the kth still image. The variance σ _k ² in the above equation (6) is estimated on the assumption that the scalar value of each pixel of the k-th image includes noise having a variance σ _k ² that is a constant value. is there. In general, such assumptions are sufficiently valid against the background described below.

First, when the size of the subject P, the structure of the X-ray tube 101 and the X-ray detector 102, the reconstruction unit 106, etc. are constant and the energy of the irradiated X-ray is constant, the noise of the CT image is The irradiation X-ray dose, that is, the product of the tube current and the irradiation time in the X-ray tube 101 that is proportional to this (the so-called tube current time product (mA · s)). On the other hand, it is also known that CT image noise is additive and generally follows a Gaussian distribution. That is, for an arbitrary scalar value v _n (x) (n = 1, 2,..., K) constituting the pixel value v (x) of a certain pixel x, the true value (value obtained by removing the noise contribution). V _n ⁰ (x), these difference values v _n (x) −v _n ⁰ (x) generally follow a Gaussian distribution with an average of 0 and a variance σ _k ² (note that the irradiation X-ray dose or the tube current The time product m · As and the noise variance σ _k ² are generally in inverse proportion.)

The variance σ _k ² also depends on the position of the pixel x itself (as described above, for example, each coordinate value x = (x, y)), but in the normal X-ray CT apparatus 100, Between the X-ray tube 101 and the X-ray detector 102, a physical X-ray filter (for example, a so-called “wedge” or “X-ray filter” composed of a copper foil, a metal lump or the like) that adjusts the X-ray irradiation amount. (Not shown) so that it can be ignored. This is because the wedge is irradiated so that the same amount of X-rays can be detected by any X-ray detector 102 by using the fact that the subject P is made of a substance having the same density as water. This has the effect of adjusting (absorbing or shielding) a part or all of the X-ray dose. Therefore, according to such a wedge, as a result, the noise variance σ _k ² hardly depends on the position of the pixel x. This is because an effect of making the value constant is produced (by the way, this wedge is generally installed for the purpose of effectively using the dynamic range of the X-ray detector 102).

From the above, on the K still images acquired by dynamic CT imaging, it is reasonable to estimate the variance σ _k ² that is a constant value for all the pixels on the kth still image. Of course, it may be possible to easily consider extending the present embodiment when the variance is different for each pixel (details will be described later in section X-3).

Now, in order to specifically calculate the above equation (5), what value is assigned as the variance σ _k ² becomes a problem. This is a problem because usually the shape of the noise distribution can be assumed (Gaussian distribution in the above), but the specific value of the variance σ _k ² is often unknown.

  Furthermore, in general, imaging may be performed by changing the irradiation dose (X-ray tube current × irradiation time (mAs)) for each imaging.

In the k-th image (k = 1, 2,..., K), the noise dispersion of the scalar value of each pixel is σ _k ^2, and the irradiation dose used for photographing the k-th image is When R _k , σ _k ² is proportional to R _k . Therefore, if σk ₀ ² can be specified for at least one k = k ₀ , the other k can also be specified.

Can accurately estimate σ _k ² .

In the present embodiment (this situation applies), a specific numerical value of σ _k ² can be estimated for at least one k by the following method.

Of the K images, the expected value E [σ _k for the variance σ _k ^{2 is} measured by using N images (1 <N ≦ K) that can be assumed to have almost no change in the subject P. ² ] is effective. For simplicity of explanation below, it is assumed that the irradiation doses in these N images are the same, and therefore σ _k ² is constant (denoted as σ ² ) for k = 1, ² ,. Noise included in each of the scalar values v ₁ (x _f ), v ₂ (x _f ),..., V _K (x _f ) constituting the pixel value v (x _f ) of a certain pixel x _f in these N images. Is expected to follow a Gaussian distribution with mean 0 and variance σ ² as described above.

The expected value E [σ ² ] for the true variance σ ² ,

Can be obtained as The expected value E [σ ² ] of the variance can be considered to be appropriate for all pixels x on all K still images as described above, and can be used as a substitute for the true variance σ ^2. This is a value for which the certainty is guaranteed over a certain level. Therefore, in the actual calculation of the above equation (6), this E [σ ² ] may be substituted for σ ² of equation (6).

More specifically, such E [σ ² ] may be obtained from actually measured values based on, for example, the first and second still images in K still images (the above (7)). And, in terms of equation (8), this corresponds to N = 2.) Further, for the pixel x _f to be subjected to the actual operation of the (7) and (8), for example, air or bone are selected only appropriate pixel x _f excluding the portion being imaged (s If selected, an average of all the obtained E [σ ² ] may be taken. Furthermore, in general, it is better to devise measures to suppress the influence of the movement of the subject P (note that the evaluation of the noise variance σ ^{2 in} the equation (6) will be described last [in this embodiment (Please refer to “Additional Items” again.)

Even when the irradiation dose is not constant in taking these N images, it can be easily estimated that σ _k ² is correctly estimated by using the fact that σ _k ² is proportional to R _k .

Next, the parameter C in the above equation (6) will be described. First, the expression (6) includes the concept of the risk factor p (x, y) described in the above general form as follows. That is, the expression in the root sign in the right-hand side molecule of the equation (6) corresponds to the χ ² value which is supposed to follow the so-called χ square distribution, and is divided by (2σ) ² to Is the risk factor p1 (x, y) itself. That means

It is a thing.

  Eventually, in Equation (6), the risk factor p (x, y) described in the general form as described above is not explicitly displayed, but the actual weight w1 (p (x, y)). Can be seen as a function of the risk factor (= p1 (x, y)) as described above (equation (10)), that is, a “function of goodness of fit” (where the risk factor and As described above, the degree of fitness has a relationship in which when one increases, the other increases.

  As can be seen from the above equation (10), the parameter C has an effect of determining how sensitively the weight w1 (p (x, y)) reacts to the risk factor p1 (x, y). That is, when C is increased, p1 (x, y) is slightly decreased, and w1 (p (x, y)) approaches 0. Moreover, when C is made small, such a sensitive reaction can be suppressed. Specifically, C may be about 1 to 10 and preferably C = 3.

  In this embodiment, the similarity determination regarding both the pixels x and y, in other words, the determination of rejection of the above-described null hypothesis H regarding both the pixels x and y is, as is clear from the above, the risk factor p1 ( x, y) based on the so-called chi-square test method (statistical test method).

  Further, as can be seen from the expression of the above formula (6), in the present invention, after calculating the risk factor p (x, y) for each combination of x and y, the weight w (p (x, y)) It is not always necessary to go through the procedure of obtaining, and it may be configured to directly calculate (wop) as a composite function without specifically obtaining the risk factor p (x, y).

As described above, by estimating the variance σ ² (for example, E [σ ² ] in the equation (8)) and appropriately determining the parameter C (for example, C = 3), the equation (6) is obtained. For all pixels y included in the set N (x) defined for a certain pixel x (for example, an area corresponding to 3 × 3 pixels centered on the pixel x, for example) The weight w1 (p (x, y)) can be obtained. After that, by using this w1 (p (x, y)) instead of w (p (x, y)) in the above equation (4), a specific numerical calculation of the coherent filter is performed. Is possible. As a result, the pixel values v ′ (x) = (v ′ ₁ (x), v ′ ₂ (x),..., V, in which the noise is strongly suppressed without impairing the spatial resolution as well as the temporal resolution. ′ _K (x)) (= equation (5)), that is, such K still images or moving images can be obtained.

FIG. 4 illustrates such image processing so that it can be conceptually easily understood. That is, first, in FIG. 4A, in a still image of 1, 2,..., K, for a certain pixel x, a rectangular area N ^{3 × 3} (3 × 3 pixels centered on the pixel x). x) is assumed. The pixels in the left corner corner of the rectangular area N ^{3 × 3} (x), if _{y 1,} the pixel _{y 1,} as also shown in FIG. 4, a pixel value v _{(y 1)} ing.

Then, the scalar value _v 1 constituting the pixel value _{v (y 1) (y 1} ), v 2 (y 1), ..., v K (y 1) and the scalar value of the pixel values v (x) _v 1 ( x), v ₂ (x),..., v _K (x), the weight w1 (p (x, y ₁ )) is calculated by the above equation (6) (FIG. 4B). Also, the remaining pixels y ₂ ... In the rectangular area N ^{3 × 3} (x). .., Y ₈ , and as shown in FIG. 4B, w1 (p (x, y ₁ )),..., W1 (p (x, y ₈ )) and w1 (p (x , X)). (In this case, the risk factor p (x, x) is “1” from equation (9), and therefore the weight w1 (p (x, x)) is also “1” from equation (10) (= The maximum weight has been)).

Next, the weights w1 (p (x, y ₁ )),..., W1 (p (x, y ₈ )), w1 (p (x, x)) obtained in this way are used as the corresponding pixels. Are multiplied by scalar values v _k (y ₁ ), v _k (y ₂ ),..., V _k (y ₈ ), v _k (x) in the _{k-th image} , respectively, to obtain the sum ((4 ), And this is divided by the sum of the weights w1 related to the rectangular area N ^{3 × 3} (x) (also corresponds to the denominator of equation (4)), the k-th A scalar value v ′ _k (x) in which noise is suppressed can be obtained for the pixel x in the image (FIG. 4C). Further, for all images of k = 1, 2,..., K, the same weights w1 (p (x, y ₁ )),..., W1 (p (x, y ₈ )), w1 (p (x, x) )) To obtain a scalar value v ′ _k (x) in which noise is suppressed, thereby obtaining a pixel value v ′ _k (x) = (v ′ ₁ (x), v in which noise in the pixel x is suppressed. ′ ₂ (x),..., V ′ _K (x)) is obtained. If the above calculation is repeated for all the pixels x, K images with reduced noise can be obtained.

  An image composed of the pixel values v ′ (x) calculated in this way is as shown in FIG. 5, for example. FIG. 5A shows an original image (k = 11) before applying the coherent filter among K = 23 still images in total, and FIGS. 5B and 5C show k = 11. The images obtained when the coherent filter is applied are respectively shown in FIGS. It can be seen that the random noise seen in the original image of FIG. 5A is sufficiently suppressed in FIGS. 5B and 5C.

  Each processing described above may be performed according to a flowchart as shown in FIG. 6, for example, and calculation / image display and the like related to each processing are performed on the actual X-ray CT apparatus 100. In order to achieve this, for example, as shown in FIG. 3, an image processing unit 110 configured by a variance value estimation unit 111, a weight calculation unit 112, and a pixel value calculation unit 113 may be provided and implemented.

Among these, the weight calculation unit 112 is configured to obtain the weight w1 (p (x, y)) directly from the pixel values v (x) and v (y) as described above. Therefore, the calculation unit 112 does not specifically calculate the value of the risk factor p1 (x, y) (that is, the “risk rate calculation unit (incorporating the“ fitness quantification unit ”according to the present invention) and weights). It should be noted that the “risk rate calculation unit (fitness quantification unit)” that specifically calculates the value of the risk factor p1 (x, y) and the output thereof are not the above-described configuration. The weight calculation unit 112 may be configured to take a two-step procedure called “weight calculation unit” that calculates the weight w1 (p (x, y)) based on the above. The weight w1 (p (x, y)) is calculated using the variance σ ² and v (x) and v (y).

In addition, the pixel value calculation unit 113 uses the pixel values v (x) and v (y) and the weight w1 (p (x, y)) numerically calculated by the weight calculation unit 112 to use the pixel value v ′ ( x) is calculated. That is, the calculation unit 113 actually performs processing for suppressing noise of the original image, that is, application of a coherent filter (hereinafter, this is expressed as “applying a coherent filter”).

  When applying a coherent filter to a moving image composed of K still images in the dynamic coherent filter processing as described above, the processing in the image processing unit 110 once reconstructs all the still images. Then, these may be stored in the storage device 10M and subjected to a coherent filter later as post-processing, but the present invention is not limited to such a form, and the above-described continuous In the flow of scanning, continuous projection data collection, continuous reconstruction, and continuous display, a process for applying a coherent filter may be performed in real time (hereinafter, referred to as “real time coherent filtering”). .

In the preferred embodiment of real-time coherent filtering, each time a new image is taken and reconstructed, the following processing is performed. Among the first obtained image (image number 1) to the latest image (image number M), common identical points on K still images having image numbers M, M-1,..., M-K + 1. The pixel values (scalar values) of the pixel x at (same coordinates) are arranged and the K-dimensional vector value v (x) = (v _M (x), v _M−1 (x),..., V _{M−K + 1} (x) ). In this way, a coherent filter can be applied in exactly the same manner as the “dynamic coherent filtering” described above. However, the pixel value calculation unit 113 does not actually calculate all the elements of the pixel value v ′ (x), but only the scalar value v _M ′ (x) corresponding to the latest image (image number M). calculate. As a result, since the calculation speed is improved, the latest image in which noise is suppressed can be displayed in real time.

As another preferred embodiment of this “real time coherent filtering”, when the first K images are obtained, a coherent filter is applied in the same manner as described above, and v ₁ ′ (x),. v _K ′ (x) is obtained, and thereafter, v (x) = (v _M ) using K still images having K-dimensional vector values having image numbers M, M−1,..., M−K + 1. (X), v _M−1 ′ (x),..., V _{M−K + 1} ′ (x)), and the real time coherent filter processing described above may be applied to this. . It is convenient that the dimension K of the pixel value vector v (x) can be changed at any time by manual setting or automatic setting during these real-time coherent filtering processes.

(When the present invention is applied to IV X-ray CT imaging 2; (one image with reduced noise is obtained))
Next, a specific example of obtaining one image with reduced noise, unlike the dynamic CT, will be described. Conventionally, in general, when it is desired to obtain one CT image with less noise, in order to achieve this, time is required for K CT images repeatedly taken K times using a normal X-ray tube current. Simple average of directions (for example, if the concept of the pixel value v (x) in the present embodiment is used, the simple average is {v ₁ (x) +... + V _K (x)} / K). As a result, there are two methods: a method of obtaining one image and a method of obtaining one image by one imaging using a normal K-fold X-ray tube current. Whichever method is used, the variance of the noise is still 1 / K times that of a single image obtained by using a normal X-ray tube current. As a result of averaging, the latter method has an inversely proportional relationship between the irradiation X-ray dose and noise variance as described above). Therefore, in the past, due to the efficiency of operation, the latter method of taking images at a time within the range permitted by the performance of the X-ray tube, or when the performance of the X-ray tube cannot withstand a large amount of X-ray tube current, In general, a single CT image is obtained by the former method of repeatedly photographing K times.

  However, in these methods, the irradiation X-ray dose increases as the noise is suppressed. That is, the exposure dose of the subject P increases.

  Here, by using the coherent filter according to the present invention, it is possible to suppress such noise included in one CT image without increasing the exposure dose to the subject P.

First, in the present embodiment, K CT images are captured with an X-ray irradiation dose lower than usual during a time period during which it can be assumed that there is no change in the subject P. Therefore, in this case, for the pixel x defined for K CT images, as in the dynamic coherent filtering process,

Thus, the pixel value v (x) can be configured. In this case, since the amount of X-ray irradiation is small, the variance σ ² of noise included in each CT image is relatively large.

  Next, as the weight function in this case, a specific format similar to the above equation (6) is given.

However, in this case, since the image to be finally obtained is one CT image, it can be considered that the following method is adopted.

  For example, in the same manner as the dynamic CT imaging described above, all K images are subjected to a coherent filter, and the resulting K images are weighted averaged or simply averaged to obtain a target CT image. Or a method of applying a coherent filter to only one image arbitrarily selected from the K images (= the calculation of FIG. 4C is the selected one image) And the like.

By the way, according to the former method, the calculation for obtaining one image obtained by averaging these images from K images, which is the result of applying a coherent filter to all K images, is specifically, for example, ,

It is. Here, v ^* ′ (x) on the left side of the equation (11) is a pixel value (scalar) at the pixel x of one image obtained as a result of simple averaging. At this time, the noise variance (σ ² ) ^* included in v ^* ′ (x) is approximately

Here, since w (p (x, y)) ≧ 0, the summed term is always smaller than 1 on the right side of equation (12). Therefore, the noise suppression effect according to the present embodiment is superior to the noise suppression effect (σ ² / K) achieved by the conventional method. In other words, if the present embodiment is used to obtain a noise suppression effect equivalent to that of the prior art, the amount of X-ray irradiation required is smaller than that of the conventional method. The exposure dose can be reduced.

The effect obtained by using this embodiment is as shown in FIG. 7, for example. FIG. 7A shows an image obtained by a conventional method, that is, an image obtained by simply averaging images taken repeatedly K times in the time direction, and FIG. 7B shows a coherent image for all K images according to the present embodiment. An image formed by simple averaging of K images obtained as a result of filtering. It can be seen that the random noise seen in FIG. 7A is sufficiently suppressed in FIG. 7B. (Note that FIG. 7 shows a case where a 0.5 second scan is taken three times (K = 3), and a rectangular area of 13 × 13 pixels is used as the set N (x), and C = 3 is an image under the condition of 3.
Further, FIG. 7C is configured in consideration of the fact that the noise suppression effect in FIG. 7B is so excellent that the image becomes unfamiliar compared to the conventional image. Fig. 7 is an image obtained by taking the average of Fig. 7 (a) and Fig. 7 (b). That is, in the image of FIG. 7C, the pixel value is v ^** ′ (x), and the pixel value in FIG. 7A is v (x) (= {v ₁ (x) +... + V _K (x )} / K)

In this case, u = 1/2. The present invention is not limited to this embodiment itself (obtains FIG. 7 (b)), and for example, the value of u in the above equation (13) (however, 0 ≦ u ≦ 1) can be used by the diagnostician. A configuration that can be adjusted accordingly is also preferable.

(When the present invention is applied to V X-ray CT imaging 3; multi-energy imaging (reducing noise of multiple types of images))
Next, the case where the present invention is applied to so-called “multi-energy imaging” will be described. Here, multi-energy imaging means changing the voltage applied to the X-ray tube 101 or inserting an X-ray filter made of a thin metal plate or the like between the X-ray tube 101 and the subject P. The X-ray energy distribution to be irradiated is varied in various ways, and the same subject P is imaged by the plurality of energy distributions thus obtained. In this case, although the subject P has not changed, the image changes depending on the energy distribution of the X-rays, and as a result, a plurality of types of images are obtained.

Here, assuming that K times of photographing are performed while changing the energy distribution, pixel values v (x) = (v) can be obtained by arranging the scalar values at the pixels x of K types of images obtained by the K times of photographing. ₁ (x), v ₂ (x),..., V _K (x)). Therefore, for this pixel value v (x), dynamic coherent filtering in (III, when applying the present invention to X-ray CT imaging 1; dynamic CT imaging (removing noise from a plurality of images)) The same means can be applied to apply a coherent filter, and as a result, K types of images with random noise suppressed can be obtained.

(When the present invention is applied to a medical image diagnostic apparatus other than a VI X-ray CT apparatus)
In the following, the coherent filter according to the present invention is applied to a medical image diagnostic apparatus other than an X-ray CT apparatus, such as a nuclear magnetic resonance imaging apparatus (so-called “MRI apparatus”), a nuclear medical diagnostic apparatus (“scintillation camera, SPECT apparatus, Alternatively, an embodiment applied to a PET apparatus "), an X-ray fluoroscopic apparatus, or the like will be described.

(When the present invention is applied to a VI-I MRI apparatus)
First, an application example of the present invention relating to an MRI apparatus will be described. The present invention relates to this MRI apparatus, and to an imaging method conventionally referred to as “averaging”, “high-speed imaging”, and “three-dimensional imaging”, and “imaging using a plurality of pulse sequences”. It is possible to apply to “method” and the like.

First, averaging is, as the name suggests, for the purpose of suppressing image noise by averaging, imaging is repeated while no change occurs in the subject P, and a plurality of images obtained (K images) ) Is a simple average to obtain one image. If the scalar values at the pixels x of the K images are arranged, the pixel values v (x) = (v ₁ (x), v ₂ (x),..., V _K (x)) can be configured. For the weight function w, for example, the above equation (5) may be used. Therefore, for this pixel value v (x), dynamic coherent filtering in (III, when applying the present invention to X-ray CT imaging 1; dynamic CT imaging (removing noise from a plurality of images)) A coherent filter can be applied by applying the same means as the above, and as a result, a plurality of images in which noise superior to that due to averaging is suppressed can be obtained. Alternatively, the same processing as in (IV when applying the present invention to X-ray CT imaging 2; (obtaining one image with reduced noise)) is used to obtain one image with suppressed noise. You can also

  In addition, high-speed imaging is a method of obtaining a moving image representing a change with time by repeatedly capturing images when a change occurs in the subject P, and is also called dynamic MRI. In this case as well, a plurality of (K) still images can still be obtained. Therefore, as in the above-described averaging (III, the present invention is applied to X-ray CT imaging 1; dynamic CT imaging). The noise can be suppressed by applying a coherent filter (similar to the dynamic coherent filter processing in removing noise from a plurality of images). Further, in high-speed imaging (III. When applying the present invention to X-ray CT imaging 1; dynamic CT imaging (removing noise from a plurality of images)), exactly the same processing as real-time coherent filter processing May be applied.

  Three-dimensional imaging is an imaging method for obtaining a three-dimensional image at a time. Even in this imaging method, averaging (that is, processing for repeatedly performing imaging and taking a simple average for the purpose of suppressing noise) or high-speed imaging (when the subject P has changed) is performed. In these cases, the pixel x is expressed as a three-dimensional vector x = (x, y, z), except that it is described above. This is exactly the same as the method of applying a coherent filter in averaging or high-speed imaging. Furthermore, in order to apply the coherent filter when the three-dimensional imaging is performed only once, the method described in later-described (VIII constructing the pixel value v (x) from one image) should be applied. Can do.

  On the other hand, in the MRI apparatus, different physical parameters can be converted into an image by changing the pulse sequence. The pulse sequence is a combination of the application sequence of the RF pulse and the gradient magnetic field, the applied intensity and time, and the pause time. By changing this, a very diverse image can be obtained. A few examples of pulse sequences that are typically used include the names of physical parameters that contribute the most to image contrast, such as T2 * -weighted image, T1-weighted image, proton-density image, etc. What is called is known. In any case, depending on the pulse sequence, different images are obtained even though the subject has not changed, resulting in substantially multiple types of images. Therefore, the coherent filter can be applied in exactly the same manner as (when the present invention is applied to V X-ray CT imaging 3; noise of a plurality of types of images is reduced).

More specifically, if K images are captured while changing the pulse sequence, the vector v (x) = (v ₁ ) can be obtained by arranging the scalar values at the pixel x of the image obtained by the K images. (X), v ₂ (x),..., V _K (x)) can be configured. Therefore, in this case as well, in the same way as in the above-described averaging, it is possible to apply the same means as the dynamic coherent filter process and apply the coherent filter to suppress noise.

(When the present invention is applied to a VI-2 scintillation camera, SPECT apparatus, or PET apparatus)
Secondly, an application example of the present invention relating to a scintillation camera, a SPECT apparatus, or a PET apparatus will be described. These devices administer a drug (radiopharmaceutical) containing a radioisotope (hereinafter referred to as “RI”) in the subject, and detect gamma rays generated by the spontaneous decay of RI to thereby detect the radiopharmaceutical in the subject. In particular, when a living body is used as a subject, this distribution is used for the purpose of obtaining a so-called functional map representing the degree of biochemical metabolism and blood circulation. The scintillation camera captures the radiopharmaceutical distribution as a two-dimensional image as if the subject was seen through. Further, the SPECT apparatus and the PET apparatus capture a three-dimensional image representing a three-dimensional distribution of the radiopharmaceutical using a computer tomography technique.

  In a scintillation camera, a SPECT apparatus, or a PET apparatus, since the detected gamma dose is usually extremely small, so-called “photon noise” due to the variation in the number of photons caused by the quantum mechanical properties of RI appears in the image.

  Conventionally, in order to suppress photon noise without reducing the resolution of an image, there is no method other than increasing the detected gamma dose. Specifically, a method of increasing the amount of RI to be administered and a longer imaging time are required. There was a way to do it.

  If the amount of RI that can be administered can be increased (for example, increased doses are allowed for laboratory animals for medical research), this can reduce photon noise, but such methods are Accompanied by an increase in radiation exposure. Further, when diagnosing the human body (= the human body is a subject), the increase in the amount of RI as described above is not allowed in the first place (legally regulated). Furthermore, it is impossible to increase the amount of imaging using radiation emitted by RI that is naturally contained in natural objects.

  Further, when using a method for extending the imaging time, it is necessary that the subject does not move for a very long time (for example, several hours) in order to obtain an effective noise reduction effect. This condition is difficult to realize when a living body (particularly a human body) is used as a subject, and there is a problem that the processing capability (or cost / performance) of the imaging apparatus decreases due to a long imaging time. Therefore, in reality, it is unavoidable that a strong noise appears in the image only by setting an appropriate photographing time which is limited to an allowable time (for example, within one hour). Therefore, in order to suppress this noise, it is only possible to perform smoothing of the image. Therefore, it has been considered that it is inevitable that the resolution of the image is impaired as an adverse effect of the smoothing.

  In such a case, according to the present invention, the following processing can be performed. That is, during a time equivalent to the appropriate shooting time, a shorter time of shooting is performed a plurality of times to obtain a plurality of images having a relatively large amount of noise. And, when applying the present invention to (IV X-ray CT imaging 2; (obtaining one image with reduced noise)), the same processing is applied, and the noise is further reduced without impairing the spatial resolution. A suppressed image can be obtained, and can be realized without increasing the amount of RI.

  Further, in the scintillation camera, the SPECT apparatus, or the PET apparatus, similarly to the X-ray CT, an imaging method that repeatedly images the same subject and tracks changes in the distribution of the radiopharmaceutical in the subject, that is, “ There are imaging methods called “dynamic scintigram”, “dynamic SPECT” or “dynamic PET”. Also in this case, dynamic coherent filter processing or real-time coherent filter processing in (III, when applying the present invention to X-ray CT imaging 1; dynamic CT imaging (removing noise from a plurality of images)) Exactly the same processing can be applied, and image noise can be effectively suppressed.

  Furthermore, in a PET apparatus or a SPECT apparatus, multiple types of RI can be administered to a subject at the same time by using the fact that RI emits gamma rays of energy specific to the type, thereby simultaneously administering multiple types of RI to a subject. In other words, a so-called “multi-window photographing method” for obtaining respective distribution images may be used. This is because the energy of the detected gamma ray is measured, and the gamma ray emitted by any of the plurality of types of RI depends on whether the value falls within an artificially set numerical range (energy window). It is based on the technology to determine whether there is.

  More specifically, by setting an appropriate number of energy windows (for example, windows W1 and W2), an image based on gamma rays corresponding to one of the windows (for example, window W1), another window (for example, By separately forming images such as images based on gamma rays corresponding to the window W2), the above-described plurality of RI distribution images are obtained by one imaging.

  That is, the same object is photographed using an imaging apparatus, and a plurality of types of images are obtained under different processing conditions. More generally speaking, it can be said that “this is an imaging method in which different types of K images are obtained by one imaging using K types of RI”. Therefore, the coherent filter can be applied in exactly the same manner as (when the present invention is applied to V X-ray CT imaging 3; noise of a plurality of types of images is reduced).

That is, also in this case, the pixel value v (x) = (v ₁ (x), v ₂ (x),..., V _K (x)) is obtained for each pixel x on such K types of images. Can be configured. In addition, the variance of noise included in each pixel value v (x) varies depending on the position of the pixel x and the type of image, but is almost proportional to the total amount of gamma rays detected for each type of image. From this property, it is possible to estimate the variance of noise for each pixel x. As described above, the present invention can be applied to “multi-window shooting” to apply a coherent filter, and as a result, K types of images in which random noise is suppressed can be obtained.

  In nuclear medicine diagnostic devices, there are reconstruction methods (fan beam MU-EN reconstruction method and fan beam, parallel hole collimator super-resolution reconstruction method, etc.) that are effective in improving resolution. Combining the coherent filter according to the present invention has no principle restriction, and as a result, it is possible to create an image with high resolution and sufficiently suppressed random noise.

(VI-3 When the present invention is applied to a fluoroscope)
Thirdly, an application example of the present invention relating to an X-ray fluoroscopic apparatus will be described. First, the “X-ray fluoroscopy device” mentioned here is a device that repeatedly performs imaging at regular intervals (for example, 1/24 seconds) while continuously exposing low-dose X-rays to a subject. One imaging is called a “frame”, and an apparatus capable of observing a so-called fluoroscopic image (moving image) by continuously displaying X-ray images obtained for each frame one after another is assumed. . Therefore, in general, an apparatus having a configuration called a “C arm” or “Ω arm” is applicable, and the X-ray CT apparatus 100 described above is also equipped with the functions just described. As long as it is a thing, the form described below is naturally applicable.

  In such an X-ray fluoroscopic apparatus, since the dose is low as described above, the fluoroscopic image contains a lot of random noise due to photon noise. If the amount of X-ray irradiation is sufficiently increased in order to solve such a problem, continuous X-ray irradiation is performed, so that not only the amount of exposure to the subject greatly increases, but also the exposure to the examiner is remarkably increased. It will increase.

  When the coherent filter according to the present invention is applied, it is possible to solve these problems and provide a fluoroscopic image in which noise is sufficiently suppressed.

  Specifically, for the following explanation, if image numbers 1, 2,..., M are sequentially added to the images obtained for each frame, the newest K images, that is, image numbers M, M−1. ,..., M−K + 1 images are stored in a memory, and for these K images, (III When applying the present invention to X-ray CT imaging 1; Dynamic CT imaging (removing noise from a plurality of images) The same process as the dynamic coherent filter process or the real-time coherent filter process in)) can be applied.

  Although the medical image diagnostic apparatus is left, the above method can be applied to an apparatus used for so-called nondestructive inspection as an “X-ray fluoroscopic apparatus”.

  The present invention can be similarly applied to a tomographic image obtained by an ultrasonic diagnostic apparatus. In this case as well, there is no change in the generation of images one after another while continuously photographing the subject. Therefore, it is possible to suppress noise by applying a coherent filter with the same configuration as in the X-ray fluoroscopic apparatus. .

(When applying the present invention more generally than VII; color image etc.)
The present invention can also be applied to “general devices” or “general images” that handle images other than the medical image diagnostic apparatus described above. In the following, an application example of the present invention to such a general apparatus or image will be described. It should be noted that here, a description will be mainly given of a case where a general “color image” which is a particularly important field is targeted.

In the color image, as described above, for each pixel x constituting the image, a three-dimensional vector having the brightness of the three primary colors of light as an element can be configured. That is,

(Here, V _R , V _G , and V _B mean the light of the three primary colors, R (red), G (green), and B (blue)). As shown in FIG. 2, this corresponds to the case where “the pixel value v (x) is configured from a plurality of types of images”.

In addition, for example, when the color image is captured by a digital camera configured with a CCD image sensor, a CMOS image sensor, or the like, the dispersion of noise included in the pixel value of each pixel is determined by the amount of light incident on the pixel x. . Specifically, scalar values v _R (x), v _G (x), v _B (x) and noise variances σ ² _R (x), σ ² _G (x), σ ² _B included in these scalar values. (X) is approximately proportional to each other. Using this property, an estimated value of noise variance can be obtained. Accordingly, a specific form of the weighting function w similar to that in the above equation (6) is given, and thereby the coherent filter of the above equation (4) is configured, thereby suppressing the pixel value v ′ (x) = (V ′ _R (x), v ′ _G (x), v ′ _B (x)) can be obtained.

Taking “color” in the broadest sense, in addition to the above-mentioned red, green and blue, including, for example, infrared rays and ultraviolet rays, the above equation (3.1) can be expressed as a higher order vector or RGB. It can also be configured as a vector that does not matter. For example, it is known that a method of expressing a color image using pixel values of vector values is not necessarily limited to the three primary colors, and the present invention is applied to the case where such a vector display that does not use the three primary colors is used. You may apply. Also, four primary colors, including ultraviolet light V _U, purple V _P The separated 5 primaries (incidentally which corresponds to visual Chorui) case, for example, 5 if if if pixel x primary colors V (x, such as ) = (V _R (x), V _G (x), V _B (x), V _U (x), V _P (x) A coherent filter can be applied.

  Furthermore, for example, when trying to measure the light reflection spectrum of a subject, the wavelength of the illumination light that illuminates the subject is switched stepwise or continuously, and shooting is performed repeatedly with a monochrome camera for each illumination wavelength, or a color filter In an imaging apparatus that performs multi-wavelength imaging by switching in stages and repeatedly imaging, a pixel value v (x) is configured based on these multiple (single wavelength or single color) images. Can do. That is, a coherent filter can be applied by applying the same method as in (when applying the present invention to V X-ray CT imaging 3; (reducing noise of plural types of images)), and as a result. Multiple types of images with reduced random noise can be obtained.

  Furthermore, many of these imaging devices usually perform one shooting to obtain one image. However, if the present invention is used, as described in the above "X-ray CT imaging" (IV when the present invention is applied to X-ray CT imaging 2; (obtaining one image with reduced noise)), Rather, it is better to obtain a plurality of (K) images by repeatedly performing short-time shooting during the same shooting time. This is because the K images are randomized by applying a coherent filter (although these K images contain a lot of random noise because the amount of light decreases in short-time shooting). This is because it is possible to construct one image in which noise is more strongly suppressed. Further, when each of these K images is a color image, for example, it may be handled as “a plurality of types (three types) of images obtained by the three primary colors of light are a plurality of sets (K sets)”. That is, for each pixel x constituting each image, a three-dimensional vector having the brightness of the three primary colors of light as an element can be configured, and therefore the pixel value v (x) of the 3K-dimensional vector can be configured for the entire K images. . A coherent filter may be applied to the pixel value of the 3K-dimensional vector.

  In an imaging apparatus that captures a moving image, such as a video camera, (III When applying the present invention to X-ray CT imaging 1; dynamic CT imaging (obtaining a plurality of images with reduced noise)) The described dynamic coherent filtering or real-time coherent filtering can be applied. Further, in the same manner as described above, a coherent filter can be applied by treating as “a plurality of types (three types) of images obtained by the three primary colors of light are a plurality of sets (K sets)”.

  In addition to the digital camera described above, the present invention is also applied to an image shot by a video camera, a night vision camera, a high-speed camera, etc. (hereinafter generally referred to as “imaging device”) in the same manner as described above. Is possible. In particular, the night vision camera and the high-speed camera are originally used in a situation where the amount of light per frame is low, and there is a lot of random noise included in the image, but not to mention the spatial resolution. In particular, it is not preferable to impair temporal resolution in a high-speed camera. Therefore, it is one of the most suitable devices for applying the present invention.

Therefore, also in the imaging apparatus, an image with further reduced noise can be obtained by such an imaging method.

  On the other hand, the present invention is applied to an image reproduced in an apparatus capable of reproducing an image based on a signal recorded or transmitted on some recording medium, for example, a video tape recorder or a DVD player, instead of “imaging”. When the coherent filter according to is applied, a reproduced image in which random noise is suppressed can be obtained.

  Furthermore, in these devices, even when still images (so-called “pause images” or images for printing) are generated and displayed, still images with suppressed random noise can be obtained by applying a coherent filter. Suitable for printing. To this end, (III When applying the present invention to X-ray CT imaging 1; Dynamic CT imaging (removing noise from a plurality of images)), (V Applying the present invention to X-ray CT imaging) 3) (reducing the noise of multiple types of images)), the method described later in (VIII Constructing the pixel value v (x) from one image), or a combination thereof be able to.

(VIII Configure pixel value v (x) from one image)
The application examples described above correspond to those in which the pixel value v (x) is configured from “from a plurality of images” and “from a plurality of types of images” in the classification shown in FIG. In the present invention, in addition to the construction method of these two pixel values v (x), as shown in FIG. 2, it is also conceivable to “construct a pixel value v (x) from one image”. Can be done as described below.

In a single two-dimensional image, for example, a single CT image acquired by the X-ray CT apparatus 100 or MRI apparatus, or a single image obtained by monochrome imaging by the imaging apparatus or the like, the value of the pixel x is Consists of a single scalar value. In other words, the pixel value v (x) as a vector value is a one-dimensional vector.

In such a case, one method for applying the coherent filter according to the present invention considers the pixel value v (x) as the “one-dimensional vector” itself. That is, in the above equation (3), K = 1,

And Thereafter, the configuration of the coherent filter or the suppression of the image noise can be executed in the same manner by the above-described procedure. The present invention includes a case where the “pixel value” v (x) as the “vector value” is a one-dimensional vector, that is, a scalar value.

For the weight w3 in this case, for example, the following equation can be used.

  Here, D has the same properties as the parameter C in the equation (6). That is, when D is small, the weight w3 (p (x, y)) is smaller, and when D is larger, it is larger.

  An image based on the weight w3 (p (x, y)) and the pixel value v ′ (x) obtained by the above equation (4) is as shown in FIG. 8, for example. 8 (a), (b), (c), and (d) are contrast-enhanced images showing an example of 3D extraction of blood vessels, and FIG. 8 (e) is an image in which brain gray matter is extracted as another extraction example. is there. FIG. 8A shows a CT image (original image) to be processed by the coherent filter. FIGS. 8B and 8D show the result of applying a coherent filter using a rectangular area of 3 × 3 pixels as the set N (x) used in the equation (6). (B) shows the result when D = 9 in the above equation (14), and FIG. 8 (d) shows the result of further thresholding the image subjected to the coherent filter of (c). Yes. FIG. 8C shows an image obtained by performing threshold processing on the original image in FIG.

  When comparing the images of FIG. 8A and FIG. 8B, the image of FIG. 8B has noise compared to the image of FIG. 8A due to the coherent filter. It can be seen that it is moderately suppressed and sufficient quality is guaranteed for diagnosis.

Comparing the images of FIG. 8C and FIG. 8D, it is difficult to extract blood vessels based on the threshold value of the image of FIG. 8C in which only threshold processing is performed on the original image. On the other hand, it can be seen that the image of FIG. 8 (d) subjected to threshold processing by applying a coherent filter can extract blood vessels based on the threshold, and thin blood vessels are not lost.

  Moreover, according to FIG.8 (e), it turns out that the area | region of gray matter has been extracted.

By the way, according to such a method, although the effect as shown in FIG. 8 is obtained, the one-dimensional vector value v (x) = v (x) is used to measure the degree of matching between the pixel x and the pixel y. Therefore, only the risk factor p (x, y) that is statistically uncertain can be obtained. For this reason, the effect of suppressing random noise that can be performed without losing resolution is not very high. Therefore, in such a case, it is more preferable to adopt another method as follows. That is, it is effective to construct a multidimensional vector using a set Z (x) composed of several pixels in the vicinity of the pixel x and use this as the pixel value v (x) of the pixel x. For example, as shown in FIG. 9, in addition to the scalar value of the pixel x = (x, y) itself, that is, v (x) = v (x, y), 4 adjacent to the pixel x vertically and horizontally 4 Using a set Z (x) composed of (x + 1, y), (x-1, y), (x, y + 1), (x, y-1) pixels (hereinafter, a set Z around these pixels x) Pixels that are elements of (x) are expressed as “z ₁ (x), z ₂ (x), z ₃ (x), z ₄ (x)”), and scalar values of these four pixels are added. Consider a five-dimensional vector. That is,

It is. Such a vector can be configured in the same manner for all the pixels constituting the image G shown in FIG. 9 (FIG. 9 shows only a part of the entire image in an enlarged manner). .

In this way, when the pixel value of the one-dimensional vector is used by configuring the coherent filter according to the present invention using the pixel value v (x) of the multi-dimensional vector composed of one image. Compared with this, it becomes possible to determine the fitness with high accuracy. As a result, random noise can be more strongly suppressed without degrading the resolution of the image.

When the pixel value is configured in this way, the above-described null hypothesis H “pixel x and pixel y” used for quantifying the degree of fitness of the pixel is the same when each noise is removed. The term “has pixel values” means that the null hypothesis H ′ “a total of five pixels of the pixel x and surrounding pixels z ₁ (x) to z ₄ (x)”, the pixel y and surrounding pixels z ₁ ( The total of five pixels y) to z ₄ (y) has the same pixel value as the corresponding pixel except for noise ”.

More specifically, as shown in FIG. 9, a certain pixel x and a certain pixel y existing in a set Z (x) (this is an area around the pixel x as described above, for example) In both cases, for example, when existing inside the region E1 of some object E1 appearing on the image G as shown in the figure, the risk factor p (x, y) calculated by, for example, the equation (9) For example, the value of the weight w (p (x, y)) calculated by the equation (6) or the like is large. Because in this case, the pixels z ₂ (x) and z ₄ (x) and the pixels z ₂ (y) and z ₄ (y) are in the region of the image E1 of the object, and the pixels z ₁ (x) and z _{Since 3} (x) and the pixels z ₁ (y) and z ₃ (y) are in the image of the object E2, the vector values V (x) and V (y) represented by the expression (3.3) are similar. It is because it is doing.

Now, in the above embodiment, focusing on the pixel x at the edge of the object image E1, the pixel y having a small risk factor p (x, y) when rejecting the null hypothesis H ′ is: in general, too much in (for example, FIG. 9 to be presumed that would not be found, either even pixels x and y _p existing inside edge E1E the image of the object E1, to reject the null hypothesis H ' In this case, the risk factor p (x, y) is relatively large compared to the case of the pixels x and y shown in FIG.

  More generally speaking, with respect to a pixel x existing at the edge of an image of an object, the value of the risk factor p (x, y) calculated by, for example, equation (9) for most y∈N (x). Is high, for example, the weight w (p (x, y)) calculated by the equation (6) or the like is small. Then, since the smoothing effect is weak even if weighted averaging is performed, for example, the noise suppression effect by the coherent filter calculated by the equation (4) is poor.

Therefore, another method obtained by further improving the above method will be described below. That is, in this method, the pixel value v (x) is

It is composed. However, v _{1 to} v _{4 in the} equation (3.4) are v (x + 1, y), v (x−1, y), v (x, y + 1) and v (x) in the equation (3.3). x, y-1) are rearranged according to a certain rule (for example, in descending order of value).

In this way, the null hypothesis H ′ is the null hypothesis H ″ “the pixel x and the surrounding pixels z ₁ (x) to z ₄ (x), and the corresponding pixel y. and a total of five pixels around the pixel _{z 1 (y) ~z 4 (} y) thereof, respectively, [pixel _{z 1 (x) ~z 4 (} x) within and _z 1 (y) _{~z 4} (Although it is allowed to change the order in (y), it has the same value except for noise) ”(the inside of [] is added).

In this case, for example, also in the relationship between the pixel x and the pixel y existing inside the edge E2E of the object image E2 shown in FIG. 10, the risk factor p (x in the case where the null hypothesis H ″ is rejected. , Y) becomes smaller (because the pixels z ₂ (x), z ₃ (x), and z ₄ (x) and the pixels z ₁ (y), z ₂ (y), and z ₄ (y) are the above (3. 4) as v _{1 to} v ₃ in the equation, and the pixel z ₁ (x) and the pixel z ₃ (y) also coincide as v ₄ ). As a result, the pixel value v ( The noise suppression effect is greater than that of the configuration of x), and the spatial resolution is not impaired.

As described above, in the present invention, the pixel value v (x), which is a vector value, can be constructed from one image, and therefore, noise is suppressed using a coherent filter without impairing the spatial resolution. it can. Since such a method needs only one image, it can be generally applied to basically any image. That is, the present method can be applied to any image as well as various application examples (X-ray CT apparatus, medical image diagnostic apparatus, color image, etc.) described above. Also, even if the image is a moving image (since it can be viewed as a large number of still images as mentioned several times above), this method (" Of course, the pixel value v (x) may be applied from one image ”. Furthermore, when using this method to the three-dimensional or multi-dimensional image Z (x) etc. can be composed of six pixels and chi ₁ itself contacting the example chi _1, how to extend it inferred also very easily So there is no need to explain.

  In the present invention, the pixel value v (x) configured by a single method of “from a plurality of images”, “from a plurality of types of images”, or “from a single image” shown in FIG. Of course, a coherent filter may be configured (as already described), and further, pixel values v (x) represented as higher-dimensional vector values by appropriately combining these methods. It is also possible to configure a coherent filter based on this.

As already described in VII, for example, if there are two color images related to the same subject, the scalar value v _R (x) shown on the right side of the above equation (2.2) for a certain pixel x. , V _G (x) and v _B (x) are obtained from each of the two images as a pixel value v (x) of the pixel x,

A 6-dimensional vector can be obtained. In this way, when the pixel value v (x) represented by a vector having a higher dimension is used, for example, as a risk factor p (x, y) calculated by Equation (9) and Equation (6), A more reliable value can be calculated. As a result, random noise can be more strongly suppressed without impairing the resolution of the image.

(IX Most common form of coherent filter in this embodiment)
Now, a description will be given below of a more general form regarding the coherent filter introduced as shown in the above equation (4), apart from the description along FIG.

  In the general form of the coherent filter (equation (4)) or various application examples, the null hypothesis H “pixel x and pixel y described in the beginning is the same value when the respective noises are removed. Or “v (x) = v (y), except for the difference caused by noise between both pixels” is used. This is a more general expression as described below. can do.

That is, by introducing a model function f (a, v (y)) including M-dimensional unknown parameters a = (a ₁ , a ₂ ,..., A _M ) (where M <K), The null hypothesis H ₀ “v (x) = f (a, v (y)) + ξ (where ξ follows a (known) probability distribution)” is established in two vectors x and y representing positions. . At this time, the estimated values a 1 ^to a of the parameter a are determined from v (x) and v (y) by using an appropriate fitting method (for example, a least square method, which will be described later), thereby obtaining the above null. The hypothesis can be paraphrased as H ₀ “v (x) = f (a ^to , v (y)) + ξ, where ξ follows a (known) probability distribution.” ”Which is an equivalent proposition. Here, the above f (a ^to , v (y)) means that the degree of freedom (that is, the parameter a) allowed by the model described by the function f with respect to the pixel value v (y) of the pixel y is optimal. The pixel value v (x) of the pixel x is adjusted and converted so as to have the highest matching degree. And such optimal parameters are a ^~ .

As a result, the risk factor p (x, y) can be specifically calculated using a known probability distribution (where ξ is assumed to follow if the null hypothesis is correct), and thus the weight w ( p (x, y)) can be calculated, so the weighted average v ′ _k (x) is

Can be calculated by: This is the most common type of coherent filter in this embodiment. If the above equation (4) is compared with this equation (15), the former is actually the latter,

(IX-1 Application Example Based on Formula (15) above: Configuration of Time Density Curve Based on Dynamic Image)
Another application example of the present invention using the form shown in the equation (15) of the coherent filter will be described. Here, as the application example, when performing quantitative measurement of a time-density curve for a dynamic CT image obtained by the above-described medical image diagnostic apparatus, the present invention is applied to perform random measurement. An embodiment aiming at obtaining the time density curve with sufficiently suppressed noise will be described.

First, the “dynamic image” here refers to a series of images (image numbers 1, 2,..., K) obtained as a result of repeatedly shooting the same subject, and respective shooting times (t _1). , T ₂ ,..., T _K ), for example, a dynamic CT image obtained by the medical image diagnostic apparatus, that is, the X-ray CT apparatus 100, an X-ray fluoroscopic apparatus, a scintillation camera, an MRI, or the like. Images are taken using a device, a SPECT device, a PET device, an ultrasonic diagnostic device, or the like. In addition, the dynamic image may be configured using a series of images obtained as a result of repeated shooting by the above-described general shooting apparatus, that is, a digital camera, a video camera, a night vision camera, a high-speed shooting camera, or the like.

Further, the “time density curve” referred to here is a curve representing a change with time of the density value of an image at a specific portion in the dynamic image. In particular, in the above-mentioned medical image diagnostic apparatus, the temporal change of the contrast medium concentration in a specific tissue of a human body is used as a time concentration curve for the purpose of examining details such as blood flow dynamics and metabolic functions in the human tissue. Measuring is done. In astronomical observation or the like, a time concentration curve is used for the purpose of analyzing a change in luminous intensity of a specific celestial object. And more formally specified, i.e., the time-density curve, when the density value of a site at time _{t k} and _{d k,} the pair row _{{<t k, d k>} (k = 1,2, · .., K)}. Also, for many applications of time density curves, it is not always necessary to have an absolute value of d _k , but rather it is sufficient if only the increment (d _k −d ₁ ) relative to the first image 1 is obtained. is there. Moreover many of such applications, (here A unknown proportional coefficient) simply _(d k -d ₁₎ proportional to the data _A (d k _-d 1) is sufficient as long obtained only. In this case, therefore, the pair of columns {<t _k , A (d _k −d ₁ )> (k = 1, 2,..., K)} is the time density curve to be obtained.

In order to obtain such a time density curve, in principle, a part of the image k (k = 1, 2,..., K) constituting the dynamic image is to be measured. Or {<t _k , A (v _k (x) −v _1] using the scalar value v _k (x) of the pixel x included in the pair of columns {<t _k , v _k , (x)>}. (X))>} may be configured.

  However, in practice, there is a problem that the time density curve to be originally measured cannot be accurately obtained because random noise is included in the dynamic image taken by the medical image diagnostic apparatus or the like.

Furthermore, in practical use, so-called “partial volume effect” occurs in these dynamic images. The partial volume effect means that an image of a minute object in a subject is represented by a small number of pixels on the image, but these few pixels have an image of an adjacent object in the subject. This is a phenomenon in which the pixel values of these small number of pixels show only relatively small fluctuations (although they are proportional to fluctuations in the density value to be measured originally). In other words, the pixel values of these few pixels contain only a few signals. Therefore, when the partial volume effect is generated, the pair of columns {<t _k , v _k (x)> (k = 1, 2,..., K)} is obtained regardless of which pixel x is taken. Since the signal level is very low and it is influenced by the change of the density value in the tissue that is not originally measured, and there is random noise, the time density curve to be measured {<t _k , d There is a problem that _k >} cannot be obtained accurately.

Therefore, conventionally, in order to suppress random noise, first, a pair of columns {<t _k , v _k (x)> (k = 1, 2,...) Relating to the pixel x included in the region to be measured. ., K)} and then the time average value v _k # (x)

And the time average of the corresponding time

(Where m <n is an appropriate integer, G _j is an appropriate weighting factor), and a time density curve {<t _k #, v _k # (x> (k = 1, 2,..., K)} is used, and this is used as a time concentration curve.

Alternatively, conventionally, in order to suppress random noise, a set of pixels R (that is, a region of interest (ROI) on the image) that roughly corresponds to a region to be measured is set, and an image number k is set. , The spatial average value v _k (R) of the scalar values v _k (x) of all the pixels x∈R included in this set is obtained by the following equation (19), and a pair of columns {<t _k , v _k ( R)> (k = 1, 2,..., K)} is used, and a spatial averaging method is used in which this is used as a time density curve.

A control method is also used.

  However, according to these conventional methods, time resolution is lost when time averaging is performed, and when time averaging is performed, a change with time in the concentration of a portion other than the original portion to be measured becomes a measured value. There was a problem of mixing.

  In order to solve such problems and obtain a more accurate time density curve, the coherent filter according to the present invention is applied based on the above (IX The most general form of the coherent filter in the present embodiment). can do.

First, the null hypothesis to be used in the coherent filter of this embodiment will be described. When it is assumed that the true time concentration curve at the site to be measured is {<t _k , d _k > (k = 1, 2,..., K)}, {<t _k , A (d _k −d ₁ )> (k = 1, 2,..., K)} (where A is an unknown coefficient), the part to be measured A set R of pixels substantially corresponding to is set. For any pixel x∈R that is an element of the set R, the condition Q: “If this pixel x well reflects the true time density curve, it is almost affected by changes in density over time in other parts. If not, the partial volume effect and the pixel value (as vector values) v (x) = (v ₁ (x), v ₂ (x),..., V _k (x)) and By considering the effects of random noise,

Can be assumed to hold. Here, p (x) and q (x) are unknown coefficients that differ for each pixel x but do not change depending on the image number k (that is, the photographing time t _k ), and model the partial volume effect. is there. Γ _k (x) is a model of random noise, and has a value different for each pixel x and for each image number k, but its expected value is 0, and its statistical distribution is the pixel x. And the image number k (in the following, for the sake of explanation, an example in which the statistical distribution is a normal distribution having an average of 0 and a variance σ ² will be used. It is only necessary to be approximately approximated by the distribution of, and in that case, a method of modifying the following embodiment to be suitable is obvious.)

According to the above assumption, for any two pixels x and y that are elements of the set R, if the proposition “both pixels x and y satisfy the condition Q (above)” holds. It can be proved that the relationship of the following equation holds.

Here, a ₁ and a ₂ are unknown coefficients that differ for each pixel set x and y but do not change depending on the image number k (ie, the photographing time t _k ). Ξ _k is random noise, and the value is different for each pixel set x, y and for each image number k, but the expected value is zero.

Equation (21) is derived as follows. That is, an expression obtained by substituting y for x in the expression (19).

Get. Therefore,

By doing so, the equation (21) is derived. Here, a ₁ and a ₂ in the equation (24) are parameters representing the partial volume effect, and ξ _k in the equation (24) represents random noise.

From the above, the proposition that “pixels x and y both satisfy the condition Q” is the null hypothesis H ₀ ′ “v _k (x) = a ₁ v _k (y) + a ₂ + ξ _k (k = 1,... , K) ”.

Next, the null hypothesis H ₀ ′ “v _k (x) = a ₁ v _k (y) + a ₂ + ξ _k (k = 1,..., K)” is substantially equivalent and actually. Describes how to convert the proposition to a form that can be tested. This null hypothesis can be described in mathematically exact terms. The null hypothesis H ₀ ′ “There are certain constants a ₁ and a ₂ and ξ _k = v _k (x) −a ₁ v _k , ( y) −a ₂ (k = 1,..., K) follows a normal distribution with an average of 0 and variance (σ ² h (a ₁ )). Here, the coefficient h (a ₁ ) is

It is. (Equation (25) is immediately derived from equation (24), which is the definition of a ₁ and ξ _k , and the general nature of the variance relating to random variables.) The value of the variance σ ² is as described above. (III When the present invention is applied to X-ray CT imaging 1; dynamic CT imaging (obtaining a plurality of images with reduced noise)) and the like can be estimated simply and practically sufficiently accurately.

From the above, if the above constants a ₁ and a ₂ can be determined, it is possible to test the null hypothesis H ₀ ′. In practice, it is sufficient ^to obtain optimal estimates a ₁ ^to a ₂ ^to these constants.

For calculating the optimum estimated values of the constants a ₁ and a ₂ , a well-known fitting method can be used as it is. Therefore, an outline in the case of using the linear least square method will be described below as a typical example of such a fitting method. In order to apply the linear least square method to this embodiment, the sum of squares of ξ _k of the above null hypothesis is simply set as S (a), that is,

Define The value of S (a) depends on the constant vector a = (a ₁ , a ₂ ), that is, the values of the above constants a ₁ and a ₂ . Be calculated constant vector a as the S (a) takes a minimum value, to the constant a ₁ and a _2, the optimum estimated values a ₁ ^~ and a ₂ ^~ of the sense of unbiased estimate can be obtained. In addition, as a specific calculation method of the linear least square method, various known methods can be used, and these known calculation methods are all very simple and require a very short calculation time. It is.

Thus, as a result of calculating the optimal estimated values a ₁ ^to a ₂ ^to the constants a ₁ and a ₂ , residuals defined by the following equations are calculated.

Can be calculated specifically. Therefore, using this residual r _k ^˜ , the above-described null hypothesis H ₀ ′ is substantially equivalent to the null hypothesis H ₀ ″ “r _k ^˜ (x, y) (k = 1,..., K ) mean 0, variance ^{_{(1+ (a 1 ~) 2}} ) σ 2 of normally distributed. "and can be translated. This is a concrete proposition that can actually perform the calculation of the test.

In addition, expression by vector

(However, the vectors a and ξ depend on the pixel set x, y.)

In other words, the null hypothesis H ₀ ″ is “v (x) = f (a ^to , v (y)) + ξ, where ξ is an average of 0 and variance (1+ (a ₁ ^to ) ² ) σ ² It follows the normal distribution.) ”, Which is exactly the same form as the null hypothesis H ₀ described in (IX Most common form of coherent filter in this embodiment). That is, it is clear that this embodiment is a modification of the coherent filter according to the present invention. Here, the above f (a ^to , v (y)) means that the pixel a is adjusted by optimally adjusting the parameter a representing the partial volume effect with respect to the pixel value v (y) of the pixel y. It means a pixel value v (x) of x that has been converted so as to have the highest fitness.

Next, in the present embodiment, a method for obtaining a time density curve by a coherent filter using the above-described null hypothesis H ₀ ″ will be described. For a set R of pixels roughly corresponding to a region to be measured, For one pixel x∈R included in the set R, the following calculation is performed for all pixels y∈R that are elements of the set R. That is, the residual r _k is actually calculated using the above method. ^{~ (x, y) (k} = 1, ..., K) is calculated, and then, above the null hypothesis _{H 0} "" ^{_{r k ~ (x, y)}} (k = 1, ..., K) is average The risk factor p (x, y) or weight w (p (x, y)) for rejecting “0, variance (1+ (a ₁ ^to ) ² ) follows a normal distribution of σ ² ” according to equation (6) Calculate specifically. Then, the weighted average v ′ _k (x) is calculated by the following equation (15 ′), and the time density curve {<t _k , v ′ _k (x) −v ′ ₁ (x)> (k = 1, 2,..., K)}.

The time density curve thus obtained is {<t _k , A (d _k −d ₁ )>} (which is a linear transformation of the true time density curve {<t _k , d _k >} at the pixel x. A is a measurement value approximating an unknown coefficient), and random noise is suppressed by the weighted average effect of equation (15). In addition, as is apparent from the equation, a pixel value vector corrected for the effect of the partial volume effect is used for the pixel value vector of the other pixel y used in the calculation by the equation (15). Furthermore, the present embodiment has a common feature of the coherent filter, that is, “a temporal average is not used at all, and a spatial average is calculated using a weight based on the degree of fitness with the pixel x”. Therefore, according to the present embodiment, it is possible to obtain a time density curve in which the influence of the partial volume effect is suppressed and random noise is suppressed without impairing the time resolution. The method for obtaining the time density curve in this way is particularly referred to as a “coherent regression method”.

  Next, an example of clinical use of a time density curve in a dynamic image obtained by dynamic CT imaging or the like in medical X-ray CT will be specifically described. In this application example, imaging such as dynamic CT is performed while rapidly injecting a contrast medium into a blood vessel, and a change in the density of an image of an artery present in a human tissue is measured as a time density curve, whereby blood in the tissue is measured. It is intended to diagnose the flow dynamics.

In this application example, in many cases, arteries in human tissue are generally very thin, and thus an image of an artery appearing on a tomographic image by CT produces a partial volume effect. Furthermore, it goes without saying that the image contains random noise. For this reason, in the conventional method, it is difficult to obtain a sufficiently accurate time concentration curve related to the artery, and if measurement is performed forcibly, it is a linear transformation of the true time concentration curve <t _k , D _k > related to the artery. <T _k , A (D _k −D ₁ )> (where D _k represents a pixel value (which is a scalar value) of a group of pixels corresponding to an artery image at time t _k , and k = 1. , 2,..., K) to some extent, only measured values <t _k , (v _k (x) −v ₁ (x))> were obtained. This measurement includes random noise. Also, the coefficient A remains unknown due to the effect of the partial volume effect.

Therefore, when the above-described method according to the present invention is applied, the measured value <t _k , (v ′ _k −3 (x) −v ′) that sufficiently approximates <t _k , A (D _k −D ₁ )>. ₁ (x))> (k = 1, 2,..., K). On the other hand, some veins that can be observed on the same tomographic image are considerably thick. Therefore, for those veins, a sufficiently good approximation value <t _k , (J _k− J ₁ )> (k = 1, 2,..., K) can be obtained. Here, J _k represents a pixel value at a time t _k of a group of pixels corresponding to a vein image.

By the way, in the time concentration curve relating to blood circulation, the proposition S: “If the contrast agent concentration in the blood at time t ₁ is 0, the time concentration curve relating to which blood vessel d <t _k , (d _k −d ₁ )> is also known to have the property that the area under the curve (AUC: Area Under Curve) matches. The area under the curve here means the integration of the time density curve <t _k , (d _k −d ₁ )> with respect to time t.

Therefore, the area under the curve AUC (d) of the time concentration curve <t _k , (d _k −d ₁ )> relating to a certain blood vessel d can be approximately calculated by the following equation, for example.

Therefore, the area under the curve AUC (J) for the time concentration curve {<t _k , (J _k −J ₁ )>} obtained by the conventional method for veins can be calculated using equation (30). (J can be substituted for d.) Also, regarding the artery, if the time concentration curve {<t _k , (D _k −D ₁ )>} is known, the area under the curve AUC (D) is calculated. (26) can be calculated in the same way, and according to the proposition S above

Therefore, AUC (D) cannot be calculated.

On the other hand, the time concentration curve <t _k , (v ′ _k (x) −v ′ ₁ (x))> obtained by the method according to the present invention is expressed as <t _k , A (D _k −D ₁ )>. The latter contains an unknown coefficient A. Therefore, the area under the curve AUC (v ′) that can be specifically calculated from {<t _k , (v ′ _k (x) −v ′ ₁ (x))>} using the equation (30) is AUC ( It must be exactly A times D). That is,

The value of the unknown coefficient A can be specifically determined. Therefore, if the value of the coefficient A is used to construct a time concentration curve <t _k , (v ′ _k (x) −v ′ ₁ (x)) / A>, this represents the arterial time concentration curve <t _{It is} nothing but an approximation of _k , (D _k −D ₁ )>. In this way, a method of constructing a time concentration curve in which the value of the proportionality coefficient A that has been unknown is determined using the area under the curve is referred to as an “AUC method”.

  From the above, in the clinical use of the time density curve in dynamic images obtained by dynamic CT imaging etc., it is difficult or impossible to measure with the conventional method by combining the AUC method with the coherent regression method. Even for the fine arterial time-concentration curve, which was possible, the measurement of the partial volume effect and random noise is eliminated, and the unknown proportional coefficient A is not obtained.

Of course, the AUC method can also be applied to a time-concentration curve <t _k , (v ′ _k (x) −v ′ ₁ (x))> relating to an artery measured by a conventional method alone (randomly). However, it is possible to construct a time-concentration curve in which the value of the proportionality coefficient A, which has been unknown), is determined, although the effects of noise and the partial volume effect cannot be excluded.

  A time density curve obtained by further combining the AUC method with the coherent regression method is, for example, as shown by a symbol (A) in FIG. 11 for a certain pixel x. In this figure, the time concentration curve constructed by the conventional method is also shown by applying the AUC method alone (B). In (B), the influence of random noise is clearly seen, whereas in (A), the noise is sufficiently suppressed and the time resolution is not impaired at all.

In the above description, the scalar value d _k is used as the density value of the time density curve. However, the present embodiment is not limited to this case. For example, individual images constituting a dynamic image are color images. Needless to say, the present invention can be easily extended and applied even when the density value of the time density curve is expressed as a vector value in a set of many types of images.

(X Supplementary matter of this embodiment)
Hereinafter, supplementary items of the embodiment described above will be described.
(X-1 General supplementary matters)
First, in the above embodiment, “risk rate” p (x, y) when the null hypothesis H is rejected is assumed as means for quantifying the “fitness” according to the present invention. The invention is not limited to such a form. As described above, the “goodness” may be a numerical “index” indicating whether or not the pixels x and y are similar in some sense.

  In relation to this, in the above, the chi-square test method is used to obtain the risk factor when the various null hypotheses are rejected (see equations (6) and (9)). ), The present invention is not limited to this. In other words, the present invention is not limited to only a form using a “statistical test method” or a “χ square test method” which is a specific form thereof.

  Furthermore, in the above embodiment, as a specific form of the weight w (p (x, y)), two kinds of convenience weight functions w1 (Equation (6) or (9) are substituted (Equation 10)). And w3 (formula (14)) are mentioned, but the present invention is not limited to this. As already stated, the property to be satisfied by the weight function w as a function of the risk factor p (x, y) ∈ [0, 1] is that w (t) is defined in the domain t ∈ [0, 1]. It is just a non-negative monotonically increasing function. And most generally speaking, the weighting function w may be “a non-negative monotonically increasing function related to goodness of fit”.

(X-2 more general form of conformity)
The “fitness” referred to in the present invention is not limited to the above-described method of quantifying the risk hypothesis p (x, y) when rejecting the null hypothesis. Accordingly, the “weight” is not limited to the method of calculating by applying the weight function w to the risk factor p (x, y). In the following, a more general form regarding such a degree of conformity will be described along with a specific example.

  First, generally speaking, according to the present invention, the present invention relates to a certain proposition appropriately set according to the purpose of image processing, and the probability that the proposition holds in the image to be processed depends on the purpose. In other words, the image is digitized by an appropriate method, and the numeric value is applied to the image processing. For example, in order to quantify the fitness of a pair of pixels x and y, a more gradual scale may be configured in the present invention. As such a specific example, a configuration in which the fitness is numerically expressed using a characteristic function in a fuzzy logic or the like can be considered.

(Example of applying the present invention to X-2-1 pattern recognition)
Here, as a specific example of a more general form regarding the degree of conformity, an example in which the present invention is applied to so-called “pattern recognition” that identifies and extracts a characteristic pattern on an image will be described. .

In this case, based on the characteristics of the pattern to be identified, a function m (x) representing the probability of the proposition “pixel x is a pixel constituting the pattern” is defined in advance. However, m (x) ε [0, 1] is assumed. More specifically, m (x) is a pixel value v (x) that is a pixel value of the pixel x, and a pixel value v that each pixel y included in a set N (x) of pixels in the vicinity of the pixel x has. This is a function that calculates a numerical value representing the probability of the proposition from (y) and the vector x itself representing the position of the pixel. Therefore, m (x) needs to be designed appropriately according to the difference between the pattern to be extracted and the pattern that should not be extracted. Also, the weight function

Define in. If a set X of pixels whose m (x) is equal to or greater than a certain “threshold value” is formed, a region of the pixel x corresponding to the pattern is obtained.

  Now, when an image P to be processed is given, w (T, m (x)) is calculated by substituting a specific threshold T and the m (x) into the weight function for all pixels x of the image. (Thus, a set X having w (T, m (x)) as a characteristic function (that is, a set of pixels x such that w (T, m (x)) = 1) is defined. This X is nothing but the area occupied by the pattern on the image P). Further, a new image P ′ is generated as follows. That is, the pixel x of P ′ has v (x) as a pixel value when w (T, m (x)) = 1, and 0 (zero vector) as a pixel value otherwise. The image P ′ thus created is obtained by deleting from the image P pixel values of pixels other than those corresponding to the pattern to be identified by 0.

  Of course, the calculation of m (x), the calculation of w (T, m (x)), and each process of the configuration of P ′ may be sequentially performed, or calculation is performed by software integrating these processes. It is also possible to perform.

  In this example, the function m (x) collates the feature of the pattern with the vector value v (x) that is the pixel value of the pixel x and the value of x itself, and the “fitness” between the pattern and the pixel x. Is digitized. Then, a weight w (T, m (x)) is calculated by the weight function defined by the equation (34), and a new image P ′ as a processing result is constructed using the weight and the image P.

  As a more specific example regarding this, a configuration of image processing for extracting thin lines appearing in an aerial photograph or a micrograph of a metal surface will be described.

  First, a method for constructing an index that numerically represents the probability of the proposition “pixel x is a pixel constituting a thin line” will be described.

  A set of pixels N (x) in the vicinity of the pixel x is defined in advance. Typically, N (x) may be defined as a rectangular area centered on the pixel x. Then, a multidimensional vector value v (x) in which pixel values (scalar values or vector values) in N (x) pixels are arranged is defined. Let K be the dimension of this vector. Typically, K is preferably about 25 to 169.

Next, a normalization function nrm, which is a K-dimensional vector function that is used supplementarily for later calculations, is defined as follows. That is, for any K-dimensional vector v where v ≠ 0, the normalization function nrm is

However, k = 1, 2,..., K. With this normalization function, for any K-dimensional vector v (unless v is a zero vector)

Is guaranteed.

Next, an arbitrary number of pixels x is arbitrarily determined, and an appropriate number J of typical images are collected such that there is only one thin straight line image passing through the pixels x with a background of 0. Image numbers 1, 2,. . . , J. The value of v (x) in the image with the image number j is calculated, and the vector r (j)

It is specifically configured by. This calculation is expressed as j = 1, 2,. . . , J. (The J vectors r (j) constructed in this way are hereinafter referred to as “pattern vectors”. Note that this pattern vector corresponds to “a separately configured pixel value” in the present invention.)
The function m (x) is defined by the following equation.

This function retrieves a value. The function m (x) thus configured takes a large value when the pixel x matches or resembles any of the pattern vectors r (j) (j = 1, 2,..., J), Otherwise, since it takes a small value, it is an index that numerically represents the probability of the proposition that “the pixel x is a pixel constituting a thin line”, that is, a function that represents the “fitness”. In addition, the weight function

Define in. Here, T is a constant serving as a threshold, and an appropriate value is set.

Now, when an image P to be processed is given, w (T, m (x)) is calculated by substituting a specific threshold T and the m (x) into the weight function for all pixels x of the image. . (Thus, a set X having w (T, m (x)) as a characteristic function (that is, a set of pixels x such that w (T, m (x)) = 1) is defined. This X is nothing but the area occupied by the pattern on the image P.)
Further, a new image P ′ is generated as follows. That is, the pixel x of P ′ has v (x) as a pixel value when w (T, m (x)) = 1, and 0 (zero vector) otherwise as a pixel value. The image P ′ thus created is obtained by deleting from the image P pixel values of pixels other than those corresponding to the pattern to be identified by 0.

Also, w (T, t) is an ambiguous discriminant function, for example

(Where C is a positive constant)
The pixel x of P ′ may be configured such that w (T, m (x) V (x) is a pixel value. This is a “thin line” by ambiguous logic. By reducing the contrast of pixels that are judged to be less relevant, the effect of configuring the image P ′ so that the “thin line” portion is raised is produced.

(Supplementary items regarding the X-3 variance σ ² estimation method)
Next, a supplementary explanation will be given regarding a method for estimating the variance σ ² (or standard deviation σ) used in the above equation (6) or (14).

In the above “when applied to X-ray CT imaging”, the expected value E [σ ² ] of variance obtained by the above equations (6) and (7) is valid for all pixels x on all K images. In view of the background, the E [σ ² ] is uniformly used for all the pixels x. However, if such a premise is not satisfied, that is, each pixel x (here, x ₁ ,..., X _F ) is unique (even if the distribution shape is the same (for example, Gaussian distribution)). The variance σ ² (x ₁ ),..., Σ ² (x _F ) may generally be assumed.

In such a case, according to the above formulas (6) and (7), the expected dispersion value E [σ ² (x ₁ )],..., E [σ ² (x _F ) is individually obtained for each pixel x. ] May be requested. Thereafter, the expected values E [σ ² (x ₁ )],..., E [σ ² (x _F )] individually obtained in this way are set according to the respective cases (= the pixel x of interest) ₁ ,..., X _F ) and (3), and v ′ (x) = (v ′ ₁ (x), v ′ ₂ (x),..., V ′ _K (x), respectively. ).

In addition, when it is assumed that the noise included in the pixel values of all the pixels has a common variance σ ² , the following method can be used to estimate the variance σ ² . First, for the set Ω (= {x ₁ ,..., X _F }) of all the pixels x, the average of the expected value E [σ ² (x)] of the variance obtained by the above formulas (7) and (8) (Σ ² ) ⁺ is obtained, and this is used as the first approximation of the estimated value of σ ² in equation (6). That is,

It is. However, in the above equation, | Ω | represents the number of elements in the set Ω (that is, | Ω | = F in this case). Next, let M be a set of pixels x such that E [σ ² (x)] is within, for example, several times the average value (σ ² ) ⁺ . If (σ ² ) ⁺⁺ is the average of only the pixels x that are elements of M, and the average is re-established,

It becomes. This can then be used as a more plausible estimate of σ.

  For example, this is the case as specifically described below, that is, as shown in FIG. 12A, there are two still images G1 and G2 constituting the moving image. As shown, the difference between the images G1 and G2 is effective when the images Z1 and Z2 are slightly shifted from each other (that is, the image Z1 moves like the image Z2).

The variance (σ ² ) ^{+ in the} equation (35) corresponds to the variance obtained based on all the pixels x (= Ω) constituting the image G3 as shown in FIG. . However, in this case, 12 as shown in (c), the probability distribution diagram for the image G3, due to the image G1 and G2 is a partial zeta ₁ and zeta ₂ which do not overlap each other, for example two mountains Since ζ ₁ ′ and ζ ₂ ′ are generated, the variance (σ ² ) ⁺ is overestimated. Therefore, it is inappropriate to use it for all still images as they are. On the other hand, the portion ζ _M shown in FIG. 12B is considered to correspond to the central mountain ζ _M ′ in FIG. 12C, and therefore the two peaks ζ ₁ are left with the mountain ζ _M ′ remaining. It is preferable to find the variance for the probability distribution with the ′ and g ₂ ′ removed. Such dispersion corresponds to the dispersion (σ ² ) ⁺⁺ in the equation (32).

That is, the average of only the set M “of pixels x such that E [σ ² (x)] is within several times the (σ ² ) ⁺ ” in the above is shown in FIG. This is nothing but the process of extracting only the peaks ζ _M ′ and taking the average, and as a result, (σ ² ) ⁺⁺ is the probability distribution of the state in which the peaks ζ ₁ ′ and ζ ₂ ′ are removed in FIG. We are looking for the variance for the figure.

Therefore, in this case, this (sigma ^{2) ++,} the by substituting the expression (5) or the like, each pixel _x 1, ..., for _{x F,} the pixel value v after conversion '(x) = A more plausible value can be obtained for (v ′ ₁ (x), v ′ ₂ (x),..., V ′ _K (x)).

  In addition, with regard to noise estimation, in general, an object with uniform brightness is placed at a position that appears in the corner of the image, and the noise distribution is measured by measuring the distribution of the value of the set of pixels corresponding to this part. It is also effective to estimate the distribution of.

  In the present invention, the noise distribution or variance estimation method may basically be any method, and is not limited to the several methods described herein. In practice, a method that is most suitably obtained should be adopted as appropriate using a priori information, imaging process theory, and measured or actually measured values that can be used in various embodiments. At this time, the estimation method is configured as simply as possible in accordance with the practical accuracy required in each embodiment, thereby simplifying the configuration of the apparatus and improving the processing speed. Preferred above.

(X-4 Still another embodiment of the present invention)
Finally, a supplementary description will be given of still another embodiment of the present invention that has not been mentioned above.
(Example of X-4-1 data compression pre-processing)
In general, when an image Q ′ is created by adding random noise to an image Q, the image Q ′ has a much larger amount of information than the image Q (the exception is that the original image Q has very large noise). This is the case. Therefore, if such an image Q ′ is to be recorded on a communication or recording medium, the amount of information for describing noise having no substantial meaning will waste the number of bits in the communication or recording. Become. In other words, if an image Q ′ with reduced noise is created from the image Q ′ and this is the target of communication and recording, an image having the same meaning with a small amount of information (and it is clear). It becomes possible to express.

  In general, when an image is an object of communication or recording, the image is often subjected to so-called “compression processing”. The purpose of this is to reduce the amount of information to be transmitted or recorded as much as possible, and to enable high-speed communication or mass recording (on one recording medium). For practical purposes, it is particularly the case of “moving images” that wants to reduce the information amount. These are facts well known in the lossy image compression technology.

  In consideration of these matters, the number of bits (that is, the amount of information) required for communication and recording is smaller when the image Q ″ that suppresses the noise is compressed than when the image Q ′ that includes the noise is compressed. can do.

  Therefore, preprocessing (specifically, a “preprocessing device” is configured by hardware on the image Q ′ that is to be communicated or recorded to suppress noise by applying the coherent filter according to the present invention, Or, it can be configured by software. The same shall apply hereinafter), and an image Q ″ with reduced noise can be obtained without impairing temporal resolution and spatial resolution, and the image Q ″ can be compressed at a high compression rate. It is possible to save time and memory required for communication and recording.

(X-4-2 Example as pre-processing of 2D or 3D region extraction processing)
A process of extracting a specific region from a two-dimensional image such as an aerial photograph or a three-dimensional distribution image by the MRI apparatus or the like is often performed. For example, a road-only region is extracted from the former aerial photograph, or a blood-vessel-only region is extracted from a three-dimensional distribution image of the human head by the latter MRI apparatus or the like, and is rendered and displayed as computer graphics. It is a process like this. In particular, the latter image of blood vessels can be rotated or observed from various directions, and is recognized as an important technique in the diagnosis of cerebrovascular disorders.

  When performing such region extraction, the simplest and most basic processing is so-called “threshold processing”. That is, the threshold processing is processing for classifying each pixel x depending on whether the scalar value of each pixel x constituting the two-dimensional image or the three-dimensional distribution image is equal to or greater than a certain threshold value. . Then, by rendering a set of pixels consisting only of pixels x that fall into a specific classification, an image that has undergone the region extraction, that is, for example, a “three-dimensional image only of blood vessels” can be represented (however, as a conventional technique, Various methods other than "threshold processing" are used.)

  By the way, in such a region extraction process, it is desirable that the original image (such as the above-described two-dimensional image or three-dimensional distribution image) has less noise. This is because if there is noise in the image, for example, a pixel that is originally a blood vessel has a pixel value as if it is a blood vessel, and conversely, a pixel that is originally a blood vessel is a blood vessel. This is because a state of having a pixel value as if it occurs. Therefore, in this case, it is necessary to remove such misclassified pixels by complicated post-processing. In addition, in the present situation, in order to remove the noise on the original image, the “smoothing process” as described in the section of the prior art is generally performed. There is a problem that the spatial resolution is lost and the fine structure (for example, thin blood vessels) of the region to be extracted is lost.

  Therefore, if preprocessing for suppressing the noise by applying the coherent filter according to the present invention to the original image, that is, the secondary reduced image or the three-dimensional distribution image, is performed, the image is not impaired. As a result, accurate region extraction processing can be performed.

  An embodiment similar to this has already been described in the case of “constructing the pixel value v (x) from one image” (see FIG. 8C).

FIG. 1 is a table showing the relationship between the risk factor p (x, y), the weight w (p (x, y)), and the null hypothesis. FIG. 2 is a table showing various application examples of the coherent filter according to the present embodiment according to the present invention, classified according to the difference in the configuration method of the pixel value v (x). FIG. 3 is a diagram showing a schematic configuration example of an X-ray CT apparatus incorporating a coherent filter. FIG. 4 is an explanatory diagram conceptually illustrating image processing by the coherent filter according to the present invention. FIG. 5A shows an example in which noise suppression processing using a coherent filter according to the present invention is performed on a plurality of images constituting a dynamic CT image, and particularly shows an unprocessed original image. FIG. 5B shows an example in which noise suppression processing is performed on a plurality of images constituting a dynamic CT image by the coherent filter according to the present invention, and particularly shows processed images. FIG. 5C shows an example in which noise suppression processing using a coherent filter according to the present invention is performed on a plurality of images constituting a dynamic CT image, and particularly shows processed images. FIG. 6 is a flowchart showing the flow of noise suppression processing by the coherent filter in the present embodiment. FIG. 7A shows an example in which noise suppression processing by the coherent filter according to the present invention is performed on one CT image, and an unprocessed original image is shown. FIG. 7B shows an example in which noise suppression processing by the coherent filter according to the present invention is performed on one CT image, and particularly shows a processed image. FIG.7 (c) shows the example which implemented the noise suppression process by the coherent filter based on this invention about one CT image, The image shown to Fig.5 (a), the image shown to FIG.5 (b), The image which took the average of is shown. FIG. 8A is a contrasted image showing an example of 3D extraction of blood vessels. FIG. 8B is a contrasted image showing an example of 3D extraction of blood vessels. FIG. 8C is a contrasted image showing an example of 3D extraction of blood vessels. FIG. 8D is a contrasted image showing an example of 3D extraction of blood vessels. FIG. 8E shows an image obtained by extracting the gray matter of the brain as another extraction example. FIG. 9 is an explanatory diagram showing an example of a method for constructing the pixel value v (x) from one image. FIG. 10 is an explanatory diagram showing another example of a method for constructing the pixel value v (x) from one image. FIG. 11 is a diagram showing a time density curve when the coherent regression method and the AUC method are combined, and the same curve when the conventional method and the AUC method are combined. FIG. 12 is an explanatory diagram for explaining a method of calculating the variance σ2 + according to the other estimation method with respect to another estimation method of the variance σ2 in equation (4) in the specification text. FIG. 13 is an explanatory diagram for explaining “smoothing processing” which is a conventional noise suppression method.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 101 ... X-ray tube, 102 ... X-ray detector, 103 ... Data acquisition part, 104 ... Pre-processing part, 105 ... Memory part, 106 ... Reconstruction part, 107 ... Image display part, 108 ... Control part, 109 ... Input Department.

Claims (1)

Determination means for determining the similarity between the first pixel and the second pixel constituting the image by a test, and determining from a plurality of different images in the time direction including the image;
Based on the determination result by the determination means, an averaging means for weighted averaging the first pixel and the second pixel;
An image processing apparatus comprising:

Images