JP2987634B2 - Abnormal shadow judgment device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Industrial Application Field) The present invention relates to a method for determining whether an abnormal shadow candidate specified on a radiation image based on image data representing the density of each pixel of the radiation image of the subject is an abnormal shadow. The present invention relates to an abnormal shadow determination device that determines whether or not the determination is negative.

2. Description of the Related Art In various fields, a recorded radiation image is read, image data representing the density of each pixel is obtained, and after appropriate image processing is performed on the image data, the image is reproduced and recorded. ing. For example, an X-ray image is recorded using an X-ray film having a low gamma value designed to be compatible with the subsequent image processing, and the X-ray image is read from the film on which the X-ray image is recorded, and an electric signal ( After converting the image data into image data and performing image processing on the image data and reproducing the image data as a visible image on a copied photograph or the like, a reproduced image having good image quality performance such as contrast, sharpness, and graininess can be obtained. (See Japanese Patent Publication No. 61-5193).

In addition, the applicant (X-ray, α-ray, β-ray,
A part of the radiation energy at the time of irradiating gamma rays, electron beams, ultraviolet rays, etc., is accumulated, and then, when irradiating with excitation light such as visible light, the stimulable phosphor (stimulable luminescent material) that emits stimulable luminescence according to the accumulated energy ( Using a stimulable phosphor, radiation image information of a subject such as a human body is temporarily recorded on a sheet-like stimulable phosphor, and the stimulable phosphor sheet is scanned with excitation light such as laser light to shine. Generates stimulating light, photoelectrically reads the resulting stimulating light to obtain image data, and based on the image data, a radiation image of the subject is made a visible image on a recording material such as a photographic photosensitive material or a CRT. A radiation image recording / reproducing system for outputting has been already proposed (Japanese Patent Laid-Open No. 55-12 / 55).
Nos. 429, 56-11395, 55-163472, 56-104645
No. 55-116340, etc.).

This system has the practical advantage of being able to record images over a very large radiation exposure area compared to conventional radiographic systems using silver halide photography. That is, in the case of the stimulable phosphor, it has been recognized that the amount of emitted light that is stimulated by excitation after accumulation is proportional to the radiation exposure amount over an extremely wide range. Even if fluctuates considerably, the amount of the stimulating light emitted from the stimulable phosphor sheet is read by the photoelectric conversion means with the reading gain set to an appropriate value and converted into an electric signal. By outputting a radiation image as a visible image on a recording material such as a photographic light-sensitive material or a display device such as a CRT using, a radiation image which is not affected by a change in radiation exposure can be obtained.

In recent years, in a system using the X-ray film or the stimulable phosphor sheet, particularly a system configured for medical diagnosis of a human body, a reproduced image having good image quality suitable for observation (diagnosis) is simply obtained. In addition, automatic recognition of images has been performed (for example, see Japanese Patent Application Laid-Open No. Sho 62-1254).
No. 81).

Here, automatic image recognition refers to an operation of extracting a target pattern from a complex radiation image by performing various processes on image data. For example, various linear shapes such as a chest X-ray image of a human body are used. , For example, an operation of extracting a shadow corresponding to a tumor from a very complicated image mixed with a circular pattern.

Such a complicated radiographic image (for example, the chest X of the human body)
A target pattern (for example, tumor shadow) is extracted from the line image), and a visible image in which the extracted pattern is specified is reproduced and displayed, thereby assisting an observer in observation (for example, assisting a doctor in diagnosis). be able to.

(Problems to be Solved by the Invention) Japanese Patent Application Laid-Open No. 62-125481 discloses, for example, a method of scanning a chest X-ray image of a human body using a real space filter which does not change depending on the position on the image, and forming a circular pattern. And a linear pattern are extracted, and the circular pattern is displayed as a tumor candidate, and the linear pattern is displayed as a shadow of a blood vessel.

However, the radiographic image of the human body is very complicated, for example, in a chest X-ray image, the pattern is different between a case where a shadow of a tumor appears immediately beside a rib and a case where it appears in the middle of a rib. There is a problem in that a device having a simple configuration as described above often misses recognition or a pattern that is not a tumor is recognized as a tumor.
For example, when automatic recognition of images is performed in a medical system, when a doctor observes a reproduced image and makes a diagnosis,
The pattern that is automatically recognized tends to catch the eye,
If there is any omission in the automatic recognition, there is a high possibility that the doctor will miss it as it is, which is a serious problem.

To avoid this, it is conceivable to devise a filter or the like that scans the radiation image and extract a pattern that is considered as a candidate for a tumor or the like without omission. However, when a pattern that is considered as a candidate for a tumor is extracted as much as possible, many patterns (noise) that are not tumors are also extracted. This is better than leaking, but too much noise will reduce the reliability of the automatic recognition system and may even make the diagnosing physician tired.

In addition, a doctor who has not automatically recognized until now, for example, observes (diagnoses) a chest X-ray image recorded on an X-ray film, based on his / her knowledge and experience, determines where the tumor is located on the image. There is also the fact that even if the pattern is slightly deformed, the tumor is extracted fairly accurately.

In view of the above facts, the present invention accurately determines whether an abnormal shadow candidate, such as a tumor, specified on a radiographic image by, for example, scanning with a real space filter as described above is an abnormal shadow. It is an object of the present invention to provide an abnormal shadow judging device capable of performing the operation.

(Means for Solving the Problems) An abnormal shadow determination apparatus according to the present invention specifies an abnormal shadow on a radiation image by a method according to an abnormal shadow to be determined based on image data representing the density of each of the radiation pixels of a subject. An abnormal shadow determination device that determines whether the abnormal shadow candidate is an abnormal shadow, based on the image data of the abnormal shadow candidate or the abnormal shadow candidate and the image data in the vicinity of the candidate. Calculating means for obtaining a plurality of feature values in accordance with the abnormal shadow to be determined; and a quantity representing the level of the probability that the abnormal shadow candidate is an abnormal shadow based on the plurality of feature values. And a neural network that outputs

Here, it is preferable that at least one of the area, shape, and contrast of the abnormal shadow candidate is included in the plurality of feature amounts.

Note that the “density” in the “density of each pixel of the radiation image” refers to a signal value corresponding to the radiation dose, and specifically, for example, the density of an X-ray photographic film or the brightness of a stimulable phosphor. It refers to the intensity of the emitted light.

The “abnormal shadow” refers to a shadow, such as a tumor, a calcification, a pleural thickening, or a pneumothorax, which is not found in a standard shadow, for example, in a chest X-ray image. The abnormal shadow determination device of the present invention may determine all of these abnormal shadows, but is not limited thereto. For example, only an abnormal shadow may be determined as an abnormal shadow.

The method for specifying the abnormal shadow candidate is not particularly limited. For example, the abnormal shadow candidate may be specified by scanning the radiation image using the real space filter as described above, and the observer of the radiation image may be used. You may specify it manually.

Further, the “feature amount” refers to a general term for various quantities that reflect the likelihood of a designated abnormal shadow candidate as an abnormal shadow. For example, the area, irregularity of the abnormal shadow candidate, and the abnormal shadow candidate Mean value of the image data, the variance value of the image data, the ratio (contrast) between the average value of the image data in the abnormal shadow candidate and the average value of the image data in the vicinity thereof. What kind of feature amount is calculated by the above calculation means,
The number of feature values to be determined is not particularly limited, and is determined in consideration of the type of abnormal shadow to be determined, determination accuracy, calculation time, and the like. However, as will be described later, the area, shape, and contrast of the abnormal shadow candidate all reflect the certainty of the abnormal shadow candidate as an abnormal shadow. It is preferable to use at least one.

Further, the above “neural network” refers to an algorithm that models the function of the human brain.

(Operation) As described above, the abnormal shadow determination device of the present invention obtains a plurality of feature amounts F ₁ , F ₂ ,..., Fn based on the image data in the vicinity of the specified abnormal shadow candidate. Is input to the neural network, and based on the plurality of characteristic amounts, an amount representing the level of probability that the abnormal shadow candidate is an abnormal shadow is output. Further, by providing the neural network with the error back propagation learning function, it is possible to "learn" past knowledge, experience, and the like, and to determine with high accuracy whether or not the abnormal shadow candidate is an abnormal shadow. Here, the “error back-propagation learning function” is a neural network in which the output of the neural network is compared with the correct answer (teacher signal), and the weight of the connection (the weight of the synaptic connection) is sequentially corrected from the output side to the input side. Refers to the “learning” algorithm in a network (for example, “DERumelhart, GEHinton and RJWilli
ams: Learning representations by back-propagating
errors, Nature, 323-9, 533-356, 986a "," Hideki Aso: Back Propagation Computrol No.24 53-6
0, "Kazuyuki Aihara, Neural Computer Tokyo Denki University Press."

In addition, by including at least one of the area, shape, and contrast of the abnormal shadow candidate in the feature amount, the area, shape, and contrast well reflect the likelihood of the abnormal shadow candidate as an abnormal shadow. Therefore, the determination accuracy can be sufficiently improved.

(Example) Hereinafter, an example of the present invention will be described with reference to the drawings. Here, an example will be described in which the above-described stimulable phosphor sheet is used to detect, as an abnormal shadow, a tumor shadow that typically occurs in the lungs of a human body as a substantially spherical shape. This tumor appears on the visible image based on the image data (image data representing the density of each pixel) S1 as a substantially circular pattern that is whitish (lower in density) than the surroundings.

 FIG. 1 is a schematic diagram of an example of an X-ray imaging apparatus.

The X-ray source 11 of the X-ray imaging apparatus 10 irradiates X-rays 12 toward the chest 13a of the human body 13 and irradiates the stimulable phosphor sheet 14 with the X-rays 12a transmitted through the human body 13 so that the human body 13 is exposed. The transmission X-ray image of the chest 13a is stored and recorded on the sheet 14.

FIG. 2 is a diagram schematically showing an example of a chest X-ray image stored and recorded on a stimulable phosphor sheet.

At the center, an X-ray image 15 of the lung (hereinafter, referred to as the lung portion 15; the same applies hereinafter), a skin portion 16 on both sides thereof, and X-rays 12 on both sides thereof.
A linear X-ray portion 17 is formed in which the sheet 14 (see FIG. 1) is directly irradiated onto the sheet 14 without passing through the subject 13.

FIG. 3 is a perspective view showing an example of an X-ray image reading apparatus and a computer system which is an embodiment of the abnormal shadow determining apparatus of the present invention.

The stimulable phosphor sheet 14 on which the X-ray image is recorded as shown in FIG. 2 is set at a predetermined position of the X-ray image reader 20. The stimulable phosphor sheet set at this predetermined position
14 is conveyed in the direction of arrow Y (sub-scan) by sheet conveying means 22 such as an endless belt driven by a motor 21
Is done. On the other hand, the light beam emitted from the laser light source 23
24 is reflected and deflected by a rotary polygon mirror 26 driven by a motor 25 and rotated at a high speed in the direction of the arrow, and after passing through a converging lens 27 such as an fθ lens, the light path is changed by a mirror 28 to be incident on the sheet 14 for sub-scanning. The main scanning is performed in an arrow X direction substantially perpendicular to the direction (arrow Y direction). From the portion of the sheet 14 irradiated with the excitation light 24, stimulated emission light 29 of an amount corresponding to the stored and recorded X-ray image information is diverged, and the stimulated emission light 29 is guided by the light guide 30. , Are photoelectrically detected by a photomultiplier (photomultiplier tube) 31. The light guide 30 is formed by molding a light-guiding material such as an acrylic plate, and is arranged such that a linear incident end face 30a extends along a main scanning line on the stimulable phosphor sheet 14. The light receiving surface of the photomultiplier 31 is joined to the emission end face 30b formed in an annular shape. The stimulated emission light 29 entering the light guide 30 from the incident end face 30a is
The light is emitted from the emission end face 30b, is received by the photomultiplier 31, and the photostimulated light 29 representing the X-ray image is converted into an electric signal by the photomultiplier 31.

The analog output signal S0 output from the photomultiplier 31 is logarithmically amplified by a logarithmic amplifier 32, and the A / D converter 3
Image data S1 digitized in 3 and as an electric signal
Is obtained.

The obtained image signal S1 is input to the computer system 40. The computer system 40 includes an example of the abnormal shadow determination device of the present invention, and includes a main unit 41 in which a CPU and an internal memory are built in, a drive unit 42 in which a floppy disk as an auxiliary memory is inserted and driven, The computer system 40 includes a keyboard 43 for inputting necessary instructions and the like to the computer system 40 and a CRT display 44 for displaying necessary information.

When the image data S1 is input to the computer system 40, a tumor shadow candidate (abnormal shadow candidate) is first specified based on the input image data S1, and then the tumor shadow candidate is identified as a tumor shadow (abnormal shadow). Probabilities are required.

As described above, the designation of the tumor shadow candidate may be manually performed by the operator while observing the X-ray image, but hereinafter, the tumor shadow is manipulated on the X-ray image using the real space filter. (Hereinafter, this may be referred to as a tumor shadow), and the extracted tumor shadow is used as the designated tumor shadow candidate to determine whether the tumor shadow candidate is a true tumor shadow. An example will be described.

Figure 4 Figure extraction of the tumor shadow candidate, in order to explain an example of a real space filter for extracting the tumor shadow, drawn virtually about a predetermined image P ₀ on the X-ray image on the image It is. The predetermined pixel P ₀ is whether a pixel of said tumor Kagenai is determined. By scanning on the X-ray image using the filter as shown here, a tumor shadow on the X-ray image is extracted.

Figure 5 is centered on the given pixel P _0, which is a diagram showing an example of a profile of the X-ray image of FIG. 4 line segments L ₁ and L ₅ of extending direction (x-direction). Here, the predetermined pixel P ₀
Is located approximately at the center of the tumor shadow 7 that is very close to the rib shadow 6. The tumor shadow 7 typically appears as a substantially symmetrical profile, but as in this example, the tumor shadow 7
May be not bilaterally symmetric when is located very close to the rib shadow 6. Even in such a case, the tumor shadow 7
It is important to be able to extract Note that the broken line 8 in FIG. 5 is an example of the profile when there is no tumor.

As shown in FIG. 4, a plurality (eight in this case) of line segments extending from a predetermined pixel P _{0 in the} X-ray image to the periphery of the X-ray image
Assuming L _i (i = 1, 2,..., 8), circles R _j (j = 1, 2, and ₃ ) having radii r ₁ , r ₂ , and r ₃ centered on a predetermined pixel P ₀ , respectively.
Assume 3). The image data of the predetermined pixel P ₀ is f ₀ ,
Each pixel P _ij located at each intersection of each line segment _Li and each circle R _j (in FIG. 4, symbols are shown for P ₁₁ , P ₁₂ , P ₁₃ , P ₅₁ , P ₅₂ and P ₅₃ . It is assumed that the image data of ()) is f _ij .

Here, the difference delta _ij between the image data f ₀ of a given pixel P ₀ and the image data f _ij of each pixel P _ij is calculated according to the following equation (1).

_{Δ ij = f ij -f 0 ...} (1) (i = 1,2, ......, 8; j = 1,2,3) then for each line segment L _i, the difference obtained in (1) The maximum value of _Δij is obtained. That is, when an example line segments L _1, L _5, for the line segment L _1, pixel P _11, P _12, P each difference corresponding to _{13 Δ 11 = f 11 -f 0} Δ 12 = f 12 - The maximum value of f ₀ Δ ₁₃ = f ₁₃ −f ₀ is obtained. In this example, the _{Δ 13 <Δ 12 <Δ 11} <0 as shown in FIG. 5, therefore delta ₁₁
Is the maximum value. Pixel P ₅₁ for the line segment L ₅ Further, P _52, P
Maximum delta ₅₃ out of the difference _{Δ 51 = f 51 -f 0 Δ} 52 = f 52 -f 0 Δ 53 = f 53 -f 0 is obtained which corresponds to _53.

Thus the predetermined pixels P _0, by to obtain the maximum value of the difference between the plurality of pixels for each line segment L _i, it is possible to cope with the tumor shadow of various sizes.

Next, the person sets the two line segments extending in opposite directions from a given pixel P _0, i.e. the line segment L ₁ and the line segment L _5, a line segment L ₂ and the line segment L _6, a line segment L ₃ With the line segment L ₇ and each of the line segment L ₄ and the line segment L ₈ as one set, the average value of the two maximum values (M ₁₅ , M ₂₆ , M ₃₇ , and M ₄₈ , respectively) is obtained for each set. . The set of the segment L ₁ and the line segment L _5, the average value M ₁₅ is Is required.

The image data this by treating the two line segments extending in opposite directions from a given pixel P ₀ as human group as, in the vicinity of the tumor shadow 7 as shown in FIG. 5, for example, ribs Pictures 6 Can be reliably detected even if the distribution is asymmetric.

If the average value _{M 15, M 26, M 37} , M 48 as described above is determined, on the basis of these average values _{M 15, M 26, M 37} , M 48 , as described below, predetermined characteristic value C ₁ of the pixel P ₀ is used for determining whether or not a pixel of tumor Kagenai is required.

6 is a diagram for illustrating how to obtain the characteristic value C _1. The horizontal axis represents the average values M ₁₅ , M ₂₆ ,
M ₃₇ , M ₄₈ , the vertical axis is each evaluation value corresponding to these average values
C ₁₅ , C ₂₆ , C ₃₇ , and C ₄₈ .

If the average value M ₁₅ , M ₂₆ , M ₃₇ , M ₄₈ is smaller than a certain value M _{1, the} evaluation value is zero, and if it is larger than a certain value M _{2, the} evaluation value is 1.0, M _{1 to} M _1
In the middle of _2, the evaluation value is a value between 0.0 and 1.0 depending on the magnitude of the value. In this way, the average values M ₁₅ , M ₂₆ ,
Evaluation values C ₁₅ , C ₂₆ , C ₃₇ , C ₄₈ corresponding to M ₃₇ , M ₄₈ are obtained, and the sum of these evaluation values C ₁₅ , C ₂₆ , C ₃₇ , C ₄₈ C ₁ = C ₁₅ + C ₂₆ + C ₃₇ + C ₄₈ (3) is the characteristic value C ₁ . That is, this characteristic value C ₁ is the minimum value 0.0
And any value between 4.0 and the maximum value.

This characteristic value C ₁ is compared with a predetermined threshold value Th ₁ , and C ₁ ≧ T
whether it is h1, by either a C ₁ <Th1, a given pixel P ₀ is whether the pixel of each tumor Kagenai is determined.

By scanning the X-ray image using the real space filter, that is, by replacing each pixel on the X-ray image with the predetermined pixel P ₀
By determining whether each pixel is a pixel within the tumor shadow, the tumor shadow on the X-ray image is extracted. still,
In the scanning of the filter, there is a possibility that a pattern other than the tumor shadow similar to the tumor shadow may be extracted. Therefore, the pattern extracted by the scanning is referred to as a tumor shadow candidate. Further, the extraction of the tumor shadow is performed, for example, using a method similar to the method described in Japanese Patent Application Laid-Open No. Sho 62-125481, or the like, by using a predetermined area including a pixel determined to be a pixel in the tumor shadow. Can be performed by partitioning a region having a circle and defining the region as a tumor shadow. The extraction of the tumor shadow is the same in the case of using another real space filter described below.

The real space filter for extracting the tumor shadow candidate is not limited to the above filter. Hereinafter, another example of the real space filter that can be used for extracting a tumor shadow candidate will be described.

The gradient of the image data f _ij of each pixel P _ij (i = 1, 2,..., 8; j = 1, 2, 3) in FIG. Is required.

Here, the gradient refers to a certain pixel P on the X-ray image.
The xy coordinate is (m, n), and the coordinates of the pixel P ′, P ″ adjacent to the pixel P in the x direction and the y direction are (m + 1, n), (m, n +
1), and the image data of the pixels P, P ′, P ″ are f (m, n), f (m + 1, n), f (m, n + 1), respectively. Means a vector represented by

FIG. 7 is a diagram showing the gradient and the calculation method described below.

Gradient These gradients are sought after The length of the vector is aligned to 1.0. That is, the gradient The size of And the standardized gradient Is required.

Next, this standardized gradient Of the direction component of the line segment L _i is calculated. That is, each pixel P
_A unit vector from _ij to a predetermined pixel P ₀ is And when (Where * represents the inner product).

Thereafter, the components a inward (direction of the predetermined pixels P ₀₎ positive for, when the outward and negative, each line segment L _i (i = 1, 2,
……, 8) Each maximum value Are obtained, and each of these maximum values Value obtained by adding Is required. This sum As the characteristic value C _2, thresholds Th2 this characteristic value C ₂ is a predetermined
It is compared with either a C ₂ ≧ Th2, depending on whether it is C ₂ <Th2, a predetermined pixel P ₀ is whether the pixel of each tumor Kagenai is determined.

This filter uses the gradient Size of The normalized, that direction (degree direction difference between the line segment L _i)
By paying attention to only the characteristic value that has a large value due to the circular shape regardless of the contrast with the surroundings
C ₂ is determined, thereby the tumor shadow is extracted with great accuracy.

Next, another example of a real space filter that can be used for extracting a tumor shadow candidate will be described.

Q ₀ and Q _ij (i = 1, 2,..., 8; j = 1, 2,
3) (However, in FIG. 4, Q ₀ and Q ₁₁ , Q ₁₂ , Q
₁₃ , Q ₅₁ , Q ₅₂ , Q ₅₃ are shown only in the central region including pixel P ₀ and each pixel P _ij (i = 1, 2,..., 8; j = 1,
Each peripheral area including (2) and (3) is shown.

Each of these areas Q ₀ and Q _ij (i = 1, 2,..., 8; j = 1, 2, 3)
In each case, the average values Q ₀ , Q _ij (i = 1, 2,..., 8; j = 1, of a large number of image data corresponding to a large number of pixels in the regions Q ₀ , Q _ij
2,3) is required. For simplicity, each area Q
₀ , Q _ij (i = 1, 2,..., 8; j = 1, 2, 3) and the same symbol used to indicate the average value of the image data in each area. .

Next, each difference Δ _ij between the average value Q _{0 of the} central region and the average value Q _ij of each peripheral region (i = 1, 2,..., 8; j = 1, 2, 3)
There is obtained as _{Δ ij = Q ij -Q 0 ...} (5), further each line segment L _i, the maximum value delta _i of the difference delta _ij is calculated. That is, when an example line segments L _1, L _{5, Δ 11} for the line segment L _1, delta _12, the maximum value delta ₁ of the delta _13, the line segment _{L 5 Δ 51, Δ 52,} Δ maximum delta ₅ of ₅₃ is required.

Next, a second characteristic value V representing the variation of the maximum value Δ _{i (i} = 1~8) first characteristic value representative of the U and the maximum value Δ _{i (i} = 1~8) is obtained. For this purpose, first, the characteristic values U _{1 to} U ₄ and V _{1 to} V ₄ are obtained according to the following arithmetic expressions.

_{U 1 = (Δ 1 + Δ} 2 + Δ 5 + Δ 6) / 4 ... (6) U 2 = (Δ 2 + Δ 3 + Δ 6 + Δ 7) / 4 ... (7) U 3 = (Δ 3 + Δ 4 + Δ 7 + Δ 8 ) / 4 ... (8) U 4 = (Δ 4 + Δ 5 + Δ 8 + Δ 1) / 4 ... (9) V 1 = U 1 / U 3 ... (10) V 2 = U 2 / U 4 ... (11) V ₃ = U ₃ / U ₁ (12) V ₄ = U ₄ / U ₂ (13) Here, for example, a case where the characteristic value U ₁ is obtained according to the equation (6) will be described. delta ₁ and delta ₂ or delta ₅ and delta ₆₎ means that the smoothed summation over, the opposite sides sandwiching the pixel P ₀ region (delta ₁ +
As with the first filter that is described above for adding the delta ₂ and Δ _{5 +} Δ _6), because the image data as shown in FIG. 5 to be able to detect the tumor shadow be asymmetrical is there.

Further, for example, (10) will be described for obtaining the characteristic value V ₁ in accordance with formula, is a characteristic value determined for directions perpendicular to each other and the characteristic value U ₁ and the characteristic value U _3, thus the tumor shadow as shown in FIG. 6 If 7 is circular, V ₁ ≒ 1.0,
When deviating from a circular, i.e. pixel P ₀ is the from V ₁ deviates from 1.0 when within shadow straight as ribs shadows.

The maximum difference delta _i of _(i = 1 to 8) as the first characteristic value U representing the the maximum value U = MAX of _{U 1 ~U 4 (U 1,} U 2, U 3, U 4) ... (14) is employed, as the second characteristic value V representing the variation of the maximum value of the difference delta _{i (i} = 1 to 8), the maximum value V = MAX (V ₁ of V ₁ ~V _{4, V 2, V 3, V 4} ) ... (15) is employed. Thus, the first and second special values
U, the V is obtained, as a characteristic value C ₃ for a given pixel P ₀ to determine whether the pixel of tumor Kagenai, the ratio of these first and second characteristic values There is employed, the characteristic value C ₃ is compared with a predetermined threshold value Th3, or it is C ₃ ≧ Th3, by either a C ₃ <Th3, the pixel
It is determined whether each of P ₀ is a pixel in the tumor shadow.

As described above, in the computer system 40 shown in FIG. 3, a circular pattern considered as a tumor shadow is extracted by first scanning an X-ray image using a real space filter.

In each of the above filter examples, as shown in FIG. 4, pixels P _ij and average values Q _ij on _eight line segments L _{1 to} L ₈ are used, but the number of these line segments is eight. Needless to say, for example, 16 or the like may be used. In addition, the calculation was performed for three distances r ₁ , r ₂ , and r ₃ with respect to the distance from the predetermined pixel P ₀ , but the calculation is not limited to the three distances, and the size of the tumor shadow to be extracted is not limited to three. is often even distances one if you are substantially constant and, in order to extract more accurately the tumor shadow various sizes, distances performs the calculation continuously varied from r ₁ to r ₃ You may. Also,
The extraction of the tumor shadow candidate is not limited to the method using each of the above filters, but may use a different filter. However, here, not only patterns that are definitely considered to be tumor shadows are extracted, but also some noise (extracting non-tumor shadow patterns as tumor shadows or extracted non-tumor shadow patterns) is included. It is preferable to extract the tumor shadow without leaking,
Therefore, the filter used here is preferably suitable for the purpose.

When scanning on an X-ray image using each of the above filters, an intersection of linear shadows, such as a region where blood vessels are dense, may be extracted as a tumor shadow candidate. Therefore, here, after extracting the tumor shadow candidates, dense regions such as blood vessels that are not tumor shadows among the extracted tumor shadow candidates are excluded from the tumor shadow candidates as follows.

FIGS. 8A and 8B are diagrams showing X-ray images of dense regions such as genuine tumor shadows and blood vessels, which are once extracted as tumor shadow candidates. In each figure, a region surrounded by a broken line 9 is a region A once extracted as a tumor shadow candidate, and each graph is a profile (a plot of the image data S1) in the x direction and the y direction in each region A. .

The tumor shadow (FIG. 9A) has a relatively flat profile having a valley near the center in both the x and y directions, and in an area where blood vessels and the like are dense (FIG. 9B), in most cases, one direction ( In FIG. 9A, the profile has a small variation in the x direction) and a relatively flat profile in the other directions (y direction). Therefore, here, by utilizing the difference between the profiles, a region where blood vessels and the like once extracted as the tumor shadow candidates are densely excluded is excluded from the tumor shadow candidates. That is, the pixels arranged in the x direction are represented by m (m = 1, 2,...), The pixels arranged in the y direction are represented by n (n = 1, 2,...), And the pixels represented by (m, n) Let the image data be f (m, n). At this time, as shown in the following equation, the average value of the square of the primary difference value of the image data in the area A is calculated.

(However, Represents that the primary difference value is added in the area A,
N is then represent) the number of pixels in the area A, as the characteristic value C ₄ to determine whether or excluded tumor shadow candidate to survive as a tumor shadow candidate, and the Z _x
The smaller one value of Z _y min and (Z _x, Z _y), the larger the value
max (Z _x , Z _y ) There is calculated, compares the characteristic value C ₄ with a predetermined threshold value Th4, to continue as a tumor shadow candidate when the C ₄ ≧ Th4, C ₄ <Th4
In the case of, it is excluded from tumor shadow candidates.

As the above-mentioned characteristic value C ₄ is not limited to that calculated by the equation (19), for example, C ₂ = | Z _x −Z _y | (21) In the above example, the primary differences f (m + 1, n) -f (m, n) and f (m, n + 1) -f (m,
Although n) is obtained, for example, a difference in an oblique direction (a direction not orthogonal to both the x direction and the y direction) may be obtained.

Figure 3 in the computer system 40, the extraction of the tumor shadow candidate is performed by scanning of the X-ray image using the real spatial filter as described above, continue as a tumor shadow candidate by (19) characteristic value C ₄ of the A determination is made as to whether or not the binary image data is obtained, whereby 1 is assigned to each pixel in the area of the surviving tumor shadow candidate and 0 is assigned to each pixel in the other areas. Thereafter the tumor shadow candidate after the determination is performed using (19) the characteristic value C ₄ of the called tumor shadow candidate.

Arithmetic Means When the tumor shadow candidate is obtained as described above, a plurality of feature values F ₁ , F ₂ ,... For each candidate are next determined based on the obtained image data S1 near the tumor shadow candidate. , Fn, that is, in the present embodiment, three feature amounts F of an area F ₁ of each tumor shadow candidate, a shape F ₂ of each tumor shadow candidate, and a contrast F ₃ of each tumor shadow candidate obtained as follows. _{1 ~F 3} is required. It should be noted that the function of the computer system 40 (see FIG. 3) for obtaining this characteristic amount is considered as an example of the calculation means of the present invention.

FIG. 9 is a diagram showing an example of a tumor shadow candidate extracted as described above. The central region 7 is a region extracted as a tumor shadow candidate.

Among the extracted tumor shadow candidates, if the area is small (for example, 25 mm ² or less), the blood vessel shadow extending in the direction (X-ray irradiation direction) perpendicular to the blood vessel tangent (X-ray image (see FIG. 2)) ) is a more likely, especially when the contrast F ₃ high to be described later is strong suspicion. The size of a tumor shadow usually seen is about 10 mm to 40 mm in diameter, and an excessively large tumor shadow candidate exceeding this range is often not a tumor shadow.

Therefore the area F ₁ of the first region 7 in this embodiment is obtained,
The area F ₁ is employed as one of the plurality of feature quantity that reflects the likelihood of the tumor shadow the tumor shadow candidate.

In the present embodiment, a tumor shadow that typically appears as a circle on an X-ray image is extracted. Therefore, an amount representing the degree (shape) of the circle is employed as another feature amount. Here, the center of gravity 0 of the area 7 is determined based on the binary image data in which a value of 1 is assigned to the area 7 and a value of 0 is assigned to the area 7 and a number of line segments extending from the center of gravity 0 to the periphery (FIG. 12). in l ₁ to l 8 pieces of ₈₎ is a virtual line segment _{l i (i = 1,2, ......} , each distance from the center of gravity 0 along 8) to the peripheral edge 7a of the area 7 d _{i (i} = 1,2, ……, 8), and the following equation Dispersion F ₂ is obtained, the dispersion F ₂ is employed as the feature quantity that represents the degree (shape) of the circular tumor shadow candidate according to.

Furthermore, the average A _V2 of the image data S1 of tumor average A _V1 and zone area of constant width around the region 7 7 of the image data S1 in the shadow region 7 of the candidate 'is obtained, the following equation F ₃ = contrast F ₃ determined according to _{A V1 -A V2 ... (23)} , the contrast
F ₃ is employed as a further different feature amounts. The tumor shadow is
Since those appearing on the generally spherical shape of the tumor X-ray image as a projection image of the lung, the contrast F ₃ may close to a predetermined value in many tumor shadow candidate or too low a contrast deviates from that value since in many cases the tumor shadow candidate contrast too high is not a tumor shadow, it is capable of adopting the contrast F ₃ as a feature quantity.

In the present embodiment, the above three feature values F _{1 to} F ₃ are obtained. However, depending on the type of the abnormal shadow to be extracted, the extraction accuracy, etc., together with the feature values F _{1 to} F ₃ or the above feature values Quantity F ₁
It further variety of different feature quantities instead to F ₃ may be adopted as a matter of course.

Neural Network The three feature values F _{1 to} F ₃ obtained as described above are
Input to the neural network.

FIG. 10 is a diagram showing an example of a neural network having an error back propagation learning (back propagation) function.

As shown in the figure, the first of this neural network
The layer (input layer), the second layer (hidden layer), and the third layer (output layer) are respectively composed of three, three, and one units. Kth
The i-th unit of layer ▲ u ^k _i ▼, the unit ▲ u ^k _i ▼
All inputs the ▲ x ^k _i ▼ to, the total output _{^{▲ y k i ▼, ▲ u}} k i ▼ from ▲ u ^{k + 1} _j ▼ weights binding to _{^{▲ W k i ▼ ▲ k +}} 1 j ▼ age,
Each unit ▲ u ^k _j ▼ identical characteristic function Shall be provided. In this case, each unit ▲ u ^k _j ▼ input ▲ x ^k _j ▼, output ▲ y ^k _j ▼ is ▲ y ^k _j ▼ = f (▲ x ^k _j ▼) (26) However, each unit ▲ u ¹ _i ▼ constituting the input layer
Each input F ₁ to _{(i = 1,2,3), F 2} , F 3 is directly input to the respective units _{^{▲ u 1 i ▼ (i =}} 1,2,3) without being weighted. The input three feature values F ₁ , F ₂ , and F ₃ are transmitted to the final output ｙy ³ ₁な が ら while being weighted by the weights kW ^k _i ▲ ^{k + 1} _j ▼ of each combination. .

Here, a method of determining the weights ＷW ^k _i ▲ ^{k + 1} _j ▼ of each connection will be described. First, the weight ▲ W of each connection by random numbers
The initial value of ^k _i ▼ ▲ ^{k + 1} _j ▼ is given. At this time, input F ₁
Be varied to F ₃ is the maximum, output ▲ y ³ ₁ ▼ so that a value or a value close thereto within the range of 0 to 1, it is preferable to limit the scope of the random number.

Next, a large number of X-ray images are prepared for which the presence or absence of the tumor shadow and the position of the tumor shadow when known are prepared, and a candidate for the tumor shadow C is extracted as described above. The above three feature values F ₁ , F ₂ and F ₃ are obtained. The three feature amounts F _1, F _2, F ₃ are input to the neural network shown in FIG. 10, each unit ▲ u ^k _i ▼ output ▲ y ^k _i ▼ is monitored.

When each output ▲ y ^k _i ▼ is obtained, the final output ▲ y ³ ₁ ▼ and the tumor shadow candidate corresponding to the input feature values F ₁ , F ₂ , F ₃ (as described above, It is known whether the tumor shadow candidate is a tumor shadow.) The square error with the teacher signal d having a value of 1 when the tumor shadow is a tumor shadow and 0 when the tumor shadow is not a tumor shadow Is required. In order to minimize the square error E, the weights ▲ W ^k _i ▲ ^{kk + 1} _j ▼ of each connection are corrected as follows.

In order to minimize the squared error E, E should be ▲ W ^k _i ▼ ▲
^Because it is a function of ^{k + 1} _j ▼ The weights ＷW ^k _i ▲ ^{k k + 1} _j ▼ of each connection are corrected as in
Here, η is a coefficient called a learning coefficient.

here, And from equation (25) Therefore, equation (29) is Becomes

Here, from equation (27), By transforming equation (31) using equation (26), Here, 'because it is (x) = f (x) (1-f (x)) ... (33), f' from _{^{(24), f (▲ x 3 1 ▼}} ) = ▲ y 3 1 ▼ _{^{· (1- ▲ y 3 1 ▼}} ) ... is (34).

In equation (30), we set k = 2 and substitute equations (32) and (34) into equation (30). The (35) by substituting expression (28) _{^{below, ▲ W 2 i ▼ ▲ 3}} 1 ▼ = ▲ W 2 i ▼ ▲ 3 1 ▼ -η · (▲ y 3 1 ▼) ·
_{^{▲ y 3 1 ▼ · (1-}} ▲ y 3 1 ▼) · ▲ y 2 i ▼ ... is (36). According to the equation (36), ▲ W ² _i ▼ ▲ ³ ₁ ▼ (i =
The weight of each connection of (1, 2, 3) is modified.

next, Therefore, substituting equations (25) and (26) into equation (37), Here, from equation (33), f ′ (▲ x ² _j ▼) = ▲ y ² _j ▼ · (1- ▲ y ² _j ▼) (39) Therefore, this equation (39) and (32) ) And (34) to (3
8) Substituting into the formula, In Equation (30), if k = 1, and Equation (40) is substituted into Equation (30), Substituting this equation (41) to (28), at the _{^{k = 1, ▲ W 1 i}} ▼ ▲ 2 j ▼ = ▲ W 1 i ▼ ▲ 2 j ▼ -η · (▲ y 3 1 ▼ -
d) ・ ▲ y ³ ₁ ▼ ・ (1- ▲ y ³ ₁ ▼) ・ ▲ y ² _j ▼ ・ (1- ▲ y ² _j ▼) ・ ▲
y ¹ _i ▼ ▲ w ² _j ▼ ▲ ³ ₁ ▼ ... (42) and ▲ W ² _i ▼ ▲ ³ ₁ ▼ (i = 1,
(2,3) is substituted into this equation (42), and ▲ W ¹ _i ▼ ▲ ² _j ▼ (i, j
= 1,2,3) is modified.

Theoretically, by using the equations (36) and (42), the learning coefficient η is made sufficiently small and the number of times of learning is made sufficiently large, so that the weight ▲ W ^k _i ▼ ▲ ^{k + 1} _{j of} each connection is ^obtained. Can be converged to a predetermined value, but making the learning coefficient η too small is not realistic because the progress of learning is slowed down. On the other hand, the learning coefficient η
When the value is set to be large, the learning may oscillate (the weight of the connection does not converge to a predetermined value). Therefore, in practice, vibration is suppressed by adding an inertia term such as the following equation to the correction amount of the weight of the connection, and the learning coefficient η is set to a relatively large value.

(E.g., DERumelhart, GEHinton and RJWilliam
s: Learning internal representations by error propa
gation In Parallel Distributed Processing, Volume
1, jLMcClelland, DERumelhart and The PDP Researc
h Group, MIT Press, 1986b ”) However _{^{Δ ▲ W k i ▼ ▲ k}} + 1 j ▼ (t) is, in the t-th learning, the connection weight of the modified _{^{▲ W k i ▼ ▲ k +}} 1 j ▼ from the previous modification of said binding weight ▲ W Represents the correction amount obtained by subtracting ^k _i ▼ ▲ ^{k + 1} _j ▼. Α is a coefficient called an inertia term.

As the inertia term α and the learning coefficient η, for example, α = 0.9η0.25
Of the weight ▲ W ^k _i ▼ ▲ ^{k + 1} _j ▼ of each connection by using (learning)
Is performed, for example, 200,000 times, and thereafter, the weight of each combination ▲ W
^k _i ▼ ▲ ^{k + 1} _j ▼ is fixed to the final value. At the end of this training, a value close to 1 for tumor shadow of a tumor shadow candidate, so close to zero for those that are not tumor shadow of tumor shadow candidate is output outputted ▲ y ³ ₁ as ▼.

Therefore, after the learning is completed, a tumor shadow candidate whose tumor shadow is unknown is extracted this time, and characteristic values F ₁ , F ₂ , and F ₃ are obtained and input to the neural network shown in FIG. is, the thus obtained output ▲ y ³ ₁ ▼ becomes a signal whose tumors shadow candidate represents the level of probability is a tumor shadow. Since this signal is obtained after learning as described above, the signal accurately represents the probability that the tumor shadow candidate is a tumor shadow.

The neural network is not limited to a three-layer structure, but may be a multilayer structure. In addition, the number of units in each layer is not limited to the number in the above embodiment, but may be any number according to the number of input characteristic values, the required accuracy of a signal representing the probability of being a tumor shadow, and the like. It goes without saying that each layer can be constituted by a unit.

When the signal as described above (the output ▲ y ³ ₁ ▼) is obtained, the signal (output ▲ y ³ ₁ ▼) a predetermined threshold value Th5
(E.g. 0.5) is compared with, ▲ y ³ ₁ ▼ ≧ Th5 a either,
▲ y ³ ₁ ▼ <depending on whether the Th5, tumor shadow candidate corresponding to the feature quantity that is input into the neural network is a tumor shadow, respectively, or a determination not a tumor shadow is performed.

Here, the threshold value Th5 does not need to be fixed, and may be changed based on, for example, the anatomical information of the subject in the X-ray image.

Here, “anatomical information” refers to information on the structure of the subject, which is expressed on the radiographic image, and specifically,
For example, it refers to positional information of a lung field, a hilum, a rib, a heart, a diaphragm, and the like in a chest X-ray image.

Lung tumor shadows are more likely to appear on the shaded lung fields 15b on both sides than on the hilum 15a near the center shown in FIG. Therefore, here, the lung portion 15 is divided into a hilum portion 15a and a lung field portion 15b as described below, and the information is compared with the output ｙy ³ ₁と and the threshold value Th5 in comparison with the threshold value Th5. Is used to determine the value of

FIG. 11 is a diagram showing a histogram of image data S1 obtained from the X-ray image shown in FIG. The horizontal axis represents the value of the image data S1, and the vertical axis represents the appearance frequency of the image data S1 having that value.

The peak 17 on the right side of this histogram corresponds to the straight X-ray part 17 shown in FIG. The central mountain 15 is the lung 15 shown in FIG.
And the mountain 16 on the left side corresponds to the skin portion 16 shown in FIG. For the sake of simplicity, the same numbers are used for the corresponding parts in FIG. 2 and the respective peaks in FIG.

Here, a predetermined ratio (for example, 50%) of the entire area of the mountain 15 is obtained from the high concentration side based on the rule of thumb that the area of the lung field 15b with respect to the area of the lung 15 is substantially constant (the hatched area in FIG. 11). ), A set of pixels corresponding to image data in this area is defined as a lung field part 15b.

In this way, the hilum 15a and the lung field 15b are divided,
When determining whether or not the tumor shadow candidate in the lung field 15b is a tumor shadow, the threshold value Th5 is set to a low value (for example,
0.4), and is set higher (for example, 0.6) when determining a tumor shadow candidate in the hilar portion 15a.

 Here, a rib shadow is also extracted.

FIG. 12 is an enlarged view of a part of a rib shadow (not shown in FIG. 2) in the chest X-ray image shown in FIG. FIG. 12 also shows a diagram of the filter shown in FIG. 4, which will be described later.

The chest X-ray image has two rib shadows 15 as shown in FIG.
A crossed region 15d of c appears, and this crossed region 15d may be extracted as a tumor shadow candidate. Moreover, the image data in the intersection area 15d is almost flat, and unlike the case where the area where the blood vessel shadows are collected is extracted as a tumor shadow candidate,
Aforementioned (19) are not excluded from the tumor shadow candidate by calculation and the determination of the characteristic quantity C ₄ by an equation. Therefore, when the rib shadow is extracted and the tumor shadow is obtained from the tumor shadow candidate, when the tumor shadow candidate matches the cross region 15d of the rib shadow, the candidate is excluded from the tumor shadow.

As an example of the method of extracting the rib shadow 15c, “Discrimination of a rib image in an indirect chest X-ray photograph” (The Institute of Electronics, Communication and Communication Engineers, October 26, 1972, Document No. IT72-24 (1972- The method described in 10)) is adopted. In this method, a chest X-ray image is scanned using a line-sensitive filter to extract a line figure, and a line corresponding to a rib shadow is determined from the position of the line figure on the X-ray image, the direction in which the line extends, and the like. Are extracted, and a rib shadow is extracted by approximating the rib boundary line by a quadratic function.

Incidentally, the threshold value Th5 is kept fixed, lung portion 15b in which whether hilar 15a in which one of the information feature amounts F ₁ a, F _2, ..., the input to the neural network described above as one of the Fn You may make it.

Further, the anatomical information may be used when extracting abnormal shadow candidates.

For example, when determining whether the predetermined pixel P ₀ described above with reference to FIG. 4 is a pixel in the tumor shadow, the pixel P ₀ shown in FIG. In determining whether or not the pixel is within the tumor shadow, L ₁ , L ₃ , L ₅ , L ₇
Is used for r ₁ and r ₂ , and for L ₂ and L ₆ , r ₁
using the information of r _3, the rib shadow can be avoided to influence this determination by using only the information of r ₁ for L ₄ and L _8.

Although the above embodiment is an example of obtaining a tumor shadow that typically appears as a circle in a chest X-ray image of a human body obtained using a stimulable phosphor, the present invention is applicable to a case where a circular tumor shadow is obtained. The present invention is not limited to a chest X-ray image and is not limited to a system using a stimulable phosphor, and is specified on a radiographic image based on image data representing a radiographic image of a subject. This can be widely used when determining whether or not the detected abnormal shadow candidate is an abnormal shadow.

(Effect of the Invention) As described in detail above, the abnormal shadow determination device of the present invention provides a plurality of abnormal shadow candidates based on the image data of the designated abnormal shadow candidate or the abnormal shadow candidate and the image data near the candidate. There is an arithmetic means for calculating a feature amount, and a neural network which inputs the plurality of feature amounts and outputs an amount representing the probability of the abnormal shadow candidate being an abnormal shadow based on the plurality of feature amounts. Therefore, it is possible to determine with high accuracy whether or not the designated abnormal shadow candidate is an abnormal shadow. In addition, in order to perform highly accurate determination, it is preferable that at least one of the area, shape, and contrast of the abnormal shadow candidate is input to the neural network as one of the feature amounts.

[Brief description of the drawings]

FIG. 1 is a schematic diagram of an example of an X-ray imaging apparatus, FIG. 2 is a diagram schematically showing an example of a chest X-ray image stored and recorded on a stimulable phosphor sheet, and FIG. FIG. 4 is a perspective view showing an example of an X-ray image reading device and a computer system which is an embodiment of the abnormal shadow detection device of the present invention. FIG. 4 illustrates an example of a real space filter for extracting a tumor shadow. FIG. 5 is a diagram drawn virtually on a predetermined pixel P ₀ on the X-ray image on the image, and FIG. 5 is a line segment L of FIG. 4 centered on the predetermined pixel P ₀ . diagram showing an example of a profile of the X-ray image ₁ and L ₅ of extending direction (x-direction), FIG. 6, the predetermined pixel P ₀ is used for determining whether or not a pixel of tumor Kagenai characteristics FIG. 7 is a diagram for explaining how to obtain a value. FIG. 7 is a diagram _showing a gradient of pixel data f _ij . FIGS. 8A and 8B are X-ray images of dense regions such as genuine tumor shadows and blood vessels, which are once extracted as tumor shadow candidates, and their x- and y-directions. FIG. 9 is a diagram showing a profile, FIG. 9 is a diagram showing an example of a tumor shadow candidate extracted by the candidate extracting means, and FIG. 10 is provided with an error back propagation learning function adopted by the abnormal shadow extracting means. FIG. 11 is a diagram showing an example of a neural network, FIG. 11 is a diagram showing a histogram of image data obtained from the X-ray image shown in FIG. 2, and FIG. 12 is a diagram of a chest X-ray image shown in FIG. It is the figure which expanded and showed some rib shadows (not shown in FIG. 2). 6 ... rib shadow, 7 ... tumor shadow 10 ... X-ray imaging device, 14 ... stimulable phosphor sheet 15 ... lung, 15a ... hilum 15b ... lung field, 16 ... skin 17: Direct X-ray part, 20: X-ray image reading device 23: Laser light source, 26: Rotating polygon mirror 29: Stimulated emission light, 30, Light guide 31: Photomultiplier 40 Computer system

 ────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Shoji Hara 798 Miyadai, Kaisei-cho, Ashigara-gun, Kanagawa Prefecture Fuji Photo Film Co., Ltd. (56) References JP-A-62-125481 JP, A) JP-A-1-114899 ( JP, A)

Claims (2)

(57) [Claims] 1. An abnormal shadow candidate designated on a radiation image by a method according to an abnormal shadow to be determined based on image data representing the density of each of the radiation pixels of the subject as to whether the abnormal shadow candidate is an abnormal shadow or not. Based on the image data of the abnormal shadow candidate or the abnormal shadow candidate and the image data in the vicinity of the candidate, a plurality of feature amounts according to the abnormal shadow to be determined Calculating means for obtaining, and a neural network for inputting the plurality of feature amounts and outputting a quantity representing a level of probability that the abnormal shadow candidate is an abnormal shadow based on the plurality of feature amounts. Abnormal shadow judgment device. 2. The abnormal shadow determination apparatus according to claim 1, wherein at least one of the area, shape, and contrast of the abnormal shadow candidate is included in the plurality of feature amounts.