JPH1196339A - Device and method for detecting mass shadow in image, and recording medium recorded with mass shadow detection program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce areas from being missed or picked up too much by using or combining the method for discriminating similarity between a link area formed from picked-up mass shadow candidate points and a three-dimensional(3 D) shape formed from a mass shadow area and a pixel strength curved surface while extracting and identifying a non-directional similar circular area from a linear area or background. SOLUTION: The non-directional similar circular area can be identified from the directional linear area or background by a mass shadow candidate point extracting part 1. Besides, normal areas picked up too much are reduced by performing direction balance analysis or 2D shape analysis in the link area of direction differentiation to the link area formed from the picked-up mass candidate points by a part 2 for reducing candidate areas picked up too much.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a tumor shadow detecting apparatus and method for recognizing and detecting a tumor shadow from an image taken by an X-ray photograph or the like.

[0002]

2. Description of the Related Art Generally, a technique for recognizing a tumor shadow from an image taken by an X-ray photograph or the like is performed by the following two-stage processing. In the first stage, tumor shadow candidate points are extracted by filter processing, and in the second stage, after the connection process is performed, normal regions that are excessively picked up while leaving the tumor shadow are reduced.

[0003] Conventional techniques in the first and second stages will be described. As a filter for extracting the tumor shadow in the first stage, a Min-DD filter that outputs the minimum value in all directions of a second-order difference filter output is known.

The Min-DD filter suppresses the image portion of a streak region such as a blood vessel, and relatively removes a tumor shadow of a type constituted only by a pixel intensity curved surface having a relatively large change in inclination in all directions. It has an excellent effect to emphasize, and is described in detail in IEICE Transactions, Vol. J67-D-II, No.2,
pp. 241-249, 1993.

Several methods of using a neural network for classifying malignant or benign to a tumor shadow candidate area or an image have been proposed. For example, US Pat. No. 5,463,548 (1995), N. Asada and K. Doi.

As a method for reducing the excessively picked-up normal region in the second stage, a two-dimensional shape analysis such as the area and circularity of the region obtained as a binary image is performed. Also, returning to the original image, narrowing down is performed using an evaluation scale such as the standard deviation of the pixel intensity value in the region or the contrast between the region and the surroundings. For example, US Pat.
360, 1996, Mohamed. Abdel-Mottaleb.

[0007]

The above-mentioned Min-DD filter maintains a tumor shadow of a type consisting only of a pixel intensity curved surface in which the change in inclination is relatively large in any direction, while keeping the breast shadow intact. Although it has the effect of suppressing lines and blood vessels, it also suppresses the pixel intensity curved surface in the tumor shadow where the change in inclination is small in any direction. In fact, some tumor shadows have a pixel intensity curved surface in which the change in inclination is small in any direction.

[0008] However, since the Min-DD filter also suppresses such a portion, a missing portion may occur when extracting a tumor shadow candidate point. Moreover, there is no device that uses a neural network not for determining a candidate area but for determining a candidate point. Furthermore, when using a neural network for the determination of candidate points, the same teacher signal is given to the peripheral and internal points of the tumor shadow, and for the constituent points of the streaked region such as blood vessels, If another teacher signal is given to perform the learning, the discrimination rate between the constituent points of the tumor shadow and the constituent points of the streak region such as blood vessels deteriorates.

Further, the tumor shadow candidate points extracted by the first stage of the filtering process are picked up because the feature amounts are similar to the points constituting the tumor shadow, and these points are formed. Whether or not the pixel intensity curved surface of the connected region is similar to the three-dimensional shape of the tumor shadow region must be determined by the processing of the second stage.

[0010] However, in the prior art, the analysis of the two-dimensional shape of the connected region obtained as a binary image and the statistical analysis of the average pixel intensity value returned to the original image are:
However, there is no document that uses the differential information of the pixel intensity curved surface to examine the three-dimensional shape and examines the balance in that area in detail.

Further, when the picked-up points form a small area, the difference between a circle and a rectangle peculiar to the two-dimensional digital image processing becomes delicate, and the evaluation scale of the two-dimensional shape is not very effective. .

Therefore, the criterion for the small area must also be determined based on the three-dimensional shape of the pixel intensity curved surface. Therefore, the present invention is capable of extracting and identifying a similar circular area having no direction from a linear area or a background, and
According to a similarity determination method and a combination thereof between the tumor shadow region and the three-dimensional shape formed by the pixel intensity curved surface, the connected region formed by the picked-up tumor shadow candidate points has a small number of overlooked or over-picked regions in the image. It is an object of the present invention to provide a detecting device and method of the present invention, and a recording medium recording a tumor shadow detection program.

[0013]

In order to achieve the above object, the present invention provides a filter operation means for performing a composite product operation on input image data with a plurality of filters having different directions, and the filter operation means. Direction balance calculating means for finding an angle between a vector having a plurality of outputs in each direction as components and a vector (virtual reference vector) having the same component, and detecting a tumor shadow based on the output of the direction balance calculating means The present invention provides a tumor shadow detection device including a tumor shadow candidate point determining unit that performs the above operation.

A filter operation step of performing a composite product operation on the input image data with a plurality of filters having different directions; a vector having a plurality of outputs for each direction of the filter as a component; A tumor shadow detection method comprising: a direction balance calculation step for obtaining direction balance information based on an angle between an equal vector (a virtual reference vector); and a tumor shadow candidate point determination step for detecting a tumor shadow based on the obtained direction balance information. I will provide a.

Further, the present invention is a recording medium storing a tumor shadow detection program for detecting a tumor shadow from input image data by a computer, wherein a composite product operation of the input image data with a plurality of filters having different directions is performed. Is performed, and directional balance information is obtained based on an angle between a vector having a plurality of outputs in each direction of the filter as components and a vector (virtual reference vector) having the same component, and based on the obtained directional balance information. And a recording medium storing a tumor shadow detection program characterized by having recorded thereon a program for causing a computer to execute tumor shadow detection characterized by determining a tumor shadow candidate point.

In the tumor shadow detection apparatus and the tumor shadow detection method having the above-described configurations, the tumor shadow candidate point extracting unit identifies and picks up a circular area having no directivity from a linear area or a background having directivity. The excessively picked-up normal region is reduced by performing a directional balance analysis, a two-dimensional shape analysis, and the like in the connected region of the directional differentiation on the connected region formed by the picked-up tumor shadow candidate points by the excessively candidate region reducing unit.

The recording medium storing the tumor shadow detection program for implementing the above-described tumor shadow detection method can realize the detection of the tumor shadow even with a general-purpose computer or the like.

[0018]

Embodiments of the present invention will be described below in detail with reference to the drawings. The tumor shadow detection device of the present invention is roughly divided into a processing method different from that of the conventional method. The tumor shadow candidate point extraction unit 1 that extracts a tumor shadow candidate point by filter processing, performs a connection process, and then detects the tumor shadow. It is composed of an over-picked candidate area reduction unit 2 for reducing the remaining and picked-up normal areas. In the present embodiment, for example, for a breast X-ray image as a target, a process of extracting a tumor shadow candidate point having similar characteristics and a point forming a tumor shadow from the breast X-ray image is described. An example will be described. Of course, the target image to be processed is not limited to the breast X-ray image as long as the image is the same as the image described in the embodiment.

FIG. 1 shows a tumor shadow candidate point extracting unit for extracting a tumor shadow candidate point from an image as a first embodiment of a tumor shadow detection device according to the present invention. The tumor shadow candidate point extraction unit 1 includes an image input data unit 11 that optically captures a breast X-ray image on an X-ray film using a CCD camera or a scanner, etc., a pixel intensity in each captured image, and various sizes. A filter operation unit 12 for calculating a composite product of a function having a direction differential action consisting of a direction and a direction balance calculation unit 13 for calculating a declination, a standard deviation, and the like of a direction-dependent output value of a Gabor cosine filter. A feature vector calculation unit 14 that generates a feature vector using the declination, standard deviation, and the like, and a feature similar to the feature vector of the tumor shadow candidate point using a neural network such as a Kohonen neural network, which will be described later. And a tumor shadow candidate point determination unit 15 that picks up a point having a vector.

The tumor shadow candidate points extracted by the tumor shadow candidate point extraction unit 1 are sent to an excessive picking candidate region reduction unit 2 for reducing a normal region, which will be described later. First, an image is taken in from an image input data section 11 composed of an image input device such as a CCD camera or a scanner.

Next, the filter operation unit 12 calculates a composite product of the pixel intensity in each image and a function having a directional differentiation function having various sizes and directions. Here, a Gabor function having a directional differentiation action is used as such a function. The two-dimensional Gabor function is given by the following equation (1):

[0022]

(Equation 1) Given by Here, x and y are spatial variables, u and v are spatial frequencies, and σ is a scale variable.

[0023]

(Equation 2) Here, a Gabor cosine function having a second-order differential action and a Gabor sine function having a first-order differential action are used for various sizes and directions.
The calculation of the composite product in the filter operation unit 12 is obtained by performing an inverse Fourier transform on the product of the Fourier transform of the input image and each Gabor function. This is performed for a Gabor function consisting of s sizes and d directions.

Next, in the direction balance calculation section 13,
The bias of the output value for each direction of the Gabor cosine filter is calculated as follows. Considering that the size is arbitrarily fixed, attention is paid to a vector of the pixel (xi, yi) having the absolute value g, k (xi, yi) of the output for each direction after passing through the filter. Here, k is a subscript indicating a direction. The length of this vector (L2 norm) is represented by the following equation.

[0025]

(Equation 3) First, a d-dimensional vector (1,1,
.., 1,1) and the above-described vector are defined as θi (hereinafter, referred to as θi).
The cos value (the amount of deviation of the output value in each direction) is calculated with θi being referred to as the argument.

[0026]

(Equation 4)

For example, assuming a three-dimensional vector as shown in FIG. 2, a three-dimensional vector which is a virtual reference vector
With respect to (1,1,1), when the deflection angle θα is formed with the vector α of one candidate point (pixel), the deviation amount is cos θα.

The deviation amount is such that the deviation angle θβ formed with the vector of the point (pixel) on the mammary gland becomes larger with respect to the three-dimensional vector, while the deviation angle θα formed with the vector of the candidate point of the tumor shadow becomes smaller. Become.

Further, the directional balance calculator 13 calculates a standard deviation of the output direction of the Gabor cosine filter. Next, the feature vector calculation unit 14 generates a feature vector using the argument, the standard deviation, and the like calculated by the direction balance calculation unit 13. When the pixel intensity value is represented by f (xi, yi), the first feature amount feature1 relating to the argument and the pixel intensity and the second feature amount f relating to the standard deviation regarding the direction of the Gabor cosine filter output
eature2 is calculated for each size according to the following equation.

[0030]

(Equation 5) Here, the standard deviation is used as the second feature amount, but the variance may be used. At this time, for example, when four filter sizes are set, eight feature amounts are obtained.

At a point inside the tumor shadow, the deviation σi becomes small because the deviation of the filter output for each direction is small. On the other hand, the deviation angle σi of a breast line, a blood vessel, or the like increases because the filter output in a specific direction increases.

Therefore, it is effective to use the argument σi as a scale for separating a tumor shadow from a breast line, a blood vessel, or the like. However, also at the background point, the deviation angle σi is small because the deviation of the filter output for each direction is small.

In order to separate the tumor shadow from the background,
The argument σi is invalid and requires a measure of the pixel intensity value and the magnitude of the filter output. For example, when the above two feature amounts are used, the values of the background such as the tumor shadow and the breast line or blood vessel have the sizes as shown in the following table.

[0034]

[Table 1]

Although the separation can be performed to some extent by using only the first characteristic amount or the second characteristic amount, it is more effective to use a combination of two characteristic amounts. In the following tumor shadow candidate point determination unit 15, for example, when using a Kohonen neural network, a filter output value having directionality is used as a feature amount without using the above-described feature conversion process. In such a case, it is necessary to virtually construct a separation surface in a feature space having a dimension equal to the number of directions, and as many neuron elements are required.

Therefore, the feature conversion processing also has the effect of reducing the number of neuron elements when using a Kohonen neural network. Subsequently, the tumor shadow candidate point determination unit 15 picks up a point having a feature vector similar to the feature vector of the tumor shadow candidate point.

An example in which this procedure is performed using a Kohonen neural network will be described below. First, an input vector is randomly selected from a set of feature vectors of tumor shadow candidate points. Next, the similarity between the input vector and the vector of each neuron element is examined. The calculation of the similarity judgment is usually performed by Euclidean distance or inner product calculation.

Subsequently, the nearest neuron element having the maximum similarity calculated above, that is, the closest neuron element is selected. When the calculation of the similarity determination is the Euclidean distance, the nearest neuron element is selected based on c expressed by the following equation.

[0039]

(Equation 6)

Next, the weight vectors of the selected neuron element and the neuron elements located near the selected neuron element are brought closer to the input feature vector. (Update of neuron element weight vector)

[0041]

(Equation 7) Here, Nc: near the nearest neuron element, α (t):
Positive number, t: represents the number of updates. While repeating the update, the magnitude of Nc and the value of α (t) are gradually reduced.

First, weight vectors are arranged in a space in the self-organizing feature map. This is unsupervised learning that does not use information on the category to which the input vector belongs. Next, the neuron elements are labeled. Subsequently, learning vector quantization (LVQ: Learning Vector Quantization)
), Supervised learning for updating the weight vector in consideration of the information of the category to which the input vector belongs is performed.

In the self-organizing feature map, the neuron elements located in the vicinity of the nearest neuron element have also updated the weight vector, but the LVQ does not update the neuron elements in the vicinity.

In the LVQ learning, different teacher signals are given to points near the boundary of the tumor shadow and points inside the boundary except for the boundary. For example, if the tumor shadow is circular, a label of 1 is given only to points whose two-dimensional distance from the center of the circle is within 0.7 times the radius, and a label of 0 is given to other points. With respect to the unlearned data after the learning is completed, the data to which one label is added is extracted as a tumor shadow candidate point. The above 2
Although the dimensional distance was set only to a point within 0.7 times the radius, this 0.7 times is a numerical value that became suitable when the test was performed on a trial basis. Any numerical value exceeding the numerical value or a numerical value smaller than the numerical value may be used as long as it can process properly.

The above-described process may be executed on a computer, or a part of the process may be performed by an optical or dedicated digital signal processor (DS).
P: Digital Signal Processor) or an electric circuit. For example, the filter operation unit 1
In 2, the image may be held in transmissive spatial light modulators having circular openings of various sizes and positions different from each other, converted into an image by a lens, and transmitted through the transmissive spatial light modulator at a focal position. .

At this time, a circular open filter that approximately realizes the Fourier transform of the Gabor function is given in a form as shown in FIG. Here, a filter having four types of sizes and four types of directions has been described as an example.

Further, using the winner-take-all circuit,
The closest weight vector can be determined. The weight vector of the neuron element may be obtained by modulating the light intensity in accordance with the value, and the updating of the weight vector can be realized by employing a rewritable spatial light modulation element.

Next, as a second embodiment of the tumor shadow detecting apparatus according to the present invention, an over-picked candidate area reducing section 2 will be described in detail. Also in the present embodiment, the target image is
The detection of a tumor shadow in a breast X-ray image will be described as an example.

In this embodiment, as shown in FIG. 4, the connection processing unit 21 performs connection processing on candidate points extracted by the tumor shadow candidate point extraction unit 1 and labels the connection area to which each point belongs. And a region evaluation unit 22 that evaluates the directional balance calculated in each connected region labeled and calculates the degree of elongation for each connected region, and a numerical value obtained from the region evaluation unit 22 An area deletion unit 23 that deletes an area based on the determined threshold value.

The area evaluation section 22 calculates an in-area direction balance calculation and evaluation section 22a for calculating and evaluating the directional balance in each connected area to which a label is attached, and calculates an elongation value for each connected area. And a two-dimensional shape evaluation unit 22b.

The operation of the excessively picked-up candidate area reducing section 2 configured as described above will be described. First, the connection processing unit 21
In, a connection process is performed on the extracted candidate points, and a label indicating which connected region each point belongs to is attached. Subsequently, in-region direction balance calculation and evaluation unit 22a of region evaluation unit 22 calculates and evaluates the direction balance in each region for each connected region. Hereinafter, this evaluation scale will be described.

First, the in-region direction balance calculation and evaluation unit 22
As a first evaluation scale in a, the directional balance in the region of the inclination of the pixel intensity curved surface is calculated and evaluated. For this purpose, a gradient vector may be used. Here, the output value of the Gabor sine filter of the tumor shadow candidate point extraction unit 1 is used as the first order differential value.

First, the direction in which the output value of the Gabor sine filter at the pixel (xi, yi) becomes maximum is found.
Here, it is assumed that a Gabor sine filter is prepared for 2d directions.

Subsequently, for each direction, the number of pixels in the connected area which is the output of the Gabor sine filter in that direction is counted. Based on this, a vector Pj having a length corresponding to the number of pixels in each direction is created.

At this time, the vectors are arranged radially as shown in FIG. 5, but as for the tumor shadow candidate area, a direction-balanced shape is obtained as shown in FIG. 5 (a). As a result, an area with poor directional balance as shown in FIG. 5B is deleted from the tumor shadow candidate area. The evaluation of the balance is made as follows.

The starting point of the radially arranged vector is defined as the origin, and a separation line in a certain direction k passing through the origin defines
In each of the divided two half planes, a vector sum of the above vectors is obtained, and a value of a ratio at which a ratio of two lengths of the vectors (Expression (10)) becomes 1 or less is adopted (Expression (11)). ).

This separation line is rotated by (360 degrees / 2d) to calculate a value, and the minimum value (unbalance) is obtained (Equation (12)). In general, since the contour line of the pixel intensity value of the tumor shadow has a shape close to a concentric circle, Gabor-si
A large value is obtained for the value of n-balance-measure.

On the other hand, since the contour line of the pixel intensity value in the pectoral muscle region is not circular but close to a group of parallel straight lines, Gabor-
The value of sin-balance-measure is small. Therefore, by removing the direction balance value that is equal to or less than a certain threshold value from the tumor shadow candidate area, it can be useful for deleting the pectoral muscle area.

[0059]

(Equation 8)

Next, the in-region direction balance calculation evaluation section 22
The second evaluation scale in a will be described. This is a point that was picked up because the feature amount is similar to the part of the medium shadow shadow that is slightly unbalanced in the direction, and if that point is unbalanced, deletes the area containing that point It is. This is based on the assumption that small tumor shadows are well balanced and have extremely good directional balance. In a connected region where the area of the largest inscribed circle is small, rotational-symmetry-m is defined as
Exclude those whose easure value is below the threshold.

[0061]

(Equation 9)

Note that the cos value here uses the cos value of the argument calculated by the direction balance calculation unit 13 described in the first embodiment. Now, the two-dimensional shape evaluation unit 2
In 2b, the value of the degree of elongation defined by the following equation is calculated for each connected region.

Elongation degree = (area of connected area) / (area of maximum inscribed circle) (14) Here, the area of the connected area is the total number of pixels constituting the area, and Area maximum-inscribed-aera
Is calculated by the following equation.

[0064]

(Equation 10) Here, I and B are suffix sets representing a set of inner points and a set of boundary points of the connected area, respectively.

The well-known calculation of the degree of elongation is to divide the area of a region by the square of the number of times of degeneration in morphology, and to calculate the elongation of the region. The elongation used here is only the elongation. Instead, it can also express the degree of circularity. That is, as shown in FIG. 6A, the degree of elongation is large in an elongated area, and as shown in FIG. 6B, the degree of elongation of a tumor shadow (similar circular area) is small (close to 1). .

In addition, if necessary, a well-known two-dimensional shape evaluation scale such as complexity may be adopted. In the region deletion unit 23, the in-region direction balance calculation evaluation unit 2
Based on the numerical values obtained from 2a and the two-dimensional shape evaluation unit 22b, the former deletes the area of the numerical value equal to or less than the predetermined threshold, and deletes the latter the area of the numerical value equal to or more than the predetermined threshold. An area having a numerical value equal to or smaller than a predetermined threshold is deleted.

The area remaining after such deletion is a tumor shadow candidate area. It is needless to say that further effects can be obtained by using the first and second embodiments in combination.

A series of these processes will be described with reference to the flowchart of FIG. First, an image is input (step S1), and after Fourier transform (step S2), the directional filter is integrated (step S1).
3) The output is subjected to inverse Fourier transform (step S)
4). These processes are equivalent to a composite product operation.

Next, the bias of the output for each direction is calculated (step S5), and a feature vector is generated from this (step S6). This feature vector is input to the neural network, and a tumor shadow candidate point is extracted (step S7), and connectivity processing is performed on this (step S8).

Next, in order to perform different processing on the small area and the other area, the diameter of the largest inscribed circle is compared with a certain threshold value ε and determined (step S9). For this determination, a small area may be selected according to the number of pixels, but here, the diameter of the largest inscribed circle is used in relation to the subsequent processing. For small regions where the diameter of the largest inscribed circle is less than a certain threshold ε, (N
O) The deviation (declination) at the point of the second order differentiation is checked (step S10). Then, by comparing the most skewed value in the area with a predetermined threshold, the area other than the well-balanced small area is deleted (step S11).

On the other hand, for a region where the diameter of the largest inscribed circle is equal to or larger than a certain threshold (YES), the directional balance in the region of the first order differential is checked (step S12). Further, two-dimensional shape processing is performed (step S13), and the candidate area that is too picked up is deleted (step S11).

As a result of these processes, a tumor shadow candidate area is extracted (step S14). As described above, according to the tumor shadow candidate point extracting unit of the present embodiment, it is possible to identify a circular area having no direction from a linear area having directionality or a background. Also, according to the excessively picked candidate area reduction unit, the connected area formed by the picked-up tumor shadow candidate points is subjected to directional balance analysis, two-dimensional shape analysis, and the like in the connected area of the directional differentiation, so that the too picked The area can be reduced.

Further, by combining the tumor shadow candidate point extraction unit and the excessive picking candidate area reducing unit, it is possible to suppress the overlooking rate of the tumor shadow and the excessive picking-up rate of the normal area to very small values.

In the first and second embodiments described above,
A computer may be constructed as a tumor shadow detection program for detecting a tumor shadow from input image data, and the tumor shadow detection program may be stored in a recording medium such as a magneto-optical disk or a semiconductor storage device and executed.

As the tumor shadow detection program, for example, a composite product operation is performed on input image data with a plurality of filters having different directions, and a vector having a plurality of outputs for each direction of the filter as a component. , Software for realizing tumor shadow detection by determining direction balance information based on an angle between each component and a vector having the same component (virtual reference vector), and determining tumor shadow candidate points based on the obtained direction balance information Is fine.

Although the above embodiments have been described, the present invention includes the following inventions. (1). A filter operation unit for performing a composite product operation on the input image data with a plurality of filters having different directions, a vector having a plurality of outputs for each direction of the filter operation unit as a component, and a vector having the same components ( A directional balance calculating means for obtaining an angle between the imaginary reference vector and a virtual reference vector; and a tumor shadow candidate point determining means for detecting a tumor shadow based on an output of the directional balance calculating means. .

(2). Standard deviation calculating means for calculating a standard deviation of a plurality of outputs for each direction of the filter calculating means, wherein the tumor shadow candidate point determining means is based on outputs from the standard deviation calculating means and the direction balance calculating means. The mass shadow detection device according to the above (1), wherein the mass shadow is detected.

(3). The image processing apparatus further includes a feature vector calculation unit that calculates a feature amount based on the pixel intensity, and the tumor shadow candidate point determination unit detects a tumor shadow based on an output of the direction balance calculation unit and the feature vector calculation unit. The tumor shadow detection device according to any one of the above (1) or (2), wherein

(4). The tumor shadow candidate point determination means includes a neural network having a learning function, and the learning of the tumor shadow candidate point includes providing different teacher signals to points near the tumor shadow boundary and inside points excluding the boundary vicinity. The tumor shadow detection device according to any one of the above (1) to (3), characterized in that:

(5). In a plurality of directions in the tumor shadow area detected by the tumor shadow candidate point determining means, a pixel number calculating means for obtaining the number of pixels at which the filter output is maximum in each direction, and a pixel number calculating means for calculating the number of pixels. When a vector is formed by setting the number of pixels as a length and the corresponding direction as a direction, the plurality of directions are divided into two half planes, and in each half plane, a vector sum of the vectors is obtained. Item (1) to Item (1), further comprising: a vector ratio calculating means for calculating a ratio of the sum lengths; and a tumor shadow determining means for detecting a tumor shadow based on an output of the vector ratio calculating means. 4) The tumor shadow detection device according to any one of the above items.

(6). For the area of the tumor shadow detected by the tumor shadow candidate determination means, further comprising an area calculation means for calculating the area of the largest inscribed circle inscribed in the area,
When the output of the area calculation means is smaller than a predetermined value, a tumor shadow is detected based on the output of the direction balance calculation means. 2. The tumor shadow detection device according to claim 1.

(7). For the area of the tumor shadow detected by the tumor shadow candidate determining means, the area ratio calculating means for calculating a ratio of the area of the area to the area of the largest inscribed circle inscribed in the area is further provided. The mass shadow saving device according to any one of (1) to (6), wherein the mass shadow is detected based on an output of the means.

(8). A filter operation step of performing a composite product operation on the input image data with a plurality of filters having different directions, a vector having a plurality of outputs for each direction of the filter as a component, and a vector having the same component (a virtual reference Vector), a direction balance calculation step of obtaining direction balance information based on an angle formed with the vector, a tumor shadow candidate point determination step of detecting a tumor shadow based on the obtained direction balance information,
A method for detecting a tumor shadow, comprising:

(9). A standard deviation calculating step of calculating a standard deviation of a plurality of outputs for each direction of the filter, wherein the tumor shadow candidate point determining step detects a tumor shadow based on the calculated standard deviation and the direction balance information. The method of detecting a tumor shadow according to the above mode (8), wherein:

(10). The method has a feature vector calculation step of calculating a feature amount based on pixel intensity, and the tumor shadow candidate point determination step detects a tumor shadow based on the direction balance information and the feature amount. (8) The method for detecting a tumor shadow according to any one of the above (9) and (9).

(11). The tumor shadow candidate point determination step includes:
Tumor shadow candidate points are determined by using a neural network with a learning function that learns tumor shadow candidate points by giving different teacher signals to points near the tumor shadow boundary and points inside the boundary excluding the boundary. The method for detecting a tumor shadow according to any one of the above (8) to (10), which is a feature.

(12). In a plurality of directions within the area of the tumor shadow detected by the tumor shadow candidate point determination step, a pixel number calculation step for obtaining the number of pixels at which the filter output is maximum in the direction is obtained by the pixel number calculation step. When a vector is formed with the number of pixels as the length and the corresponding direction as the direction, the plurality of directions are divided into two half planes, and the vector sum of the vectors is obtained in each half plane. (8) to (11), characterized by having a vector ratio calculation step for obtaining a length ratio of the image and a tumor shadow determination step for detecting a tumor shadow based on the obtained vector ratio. The method for detecting a tumor shadow according to any one of the above items.

(13). For the area of the tumor shadow detected by the tumor shadow candidate point determination step, the area has an area calculation step of calculating the area of the largest inscribed circle inscribed in the area, and when the obtained area is smaller than a predetermined value. The method for detecting a tumor shadow according to any one of the above items (8) to (12), wherein a tumor shadow is detected based on the direction balance information.

(14). The area of the tumor shadow detected by the tumor shadow candidate determination step further includes an area ratio calculation step of calculating a ratio between the area of the area and the area of the largest inscribed circle inscribed in the area. The method for detecting a tumor shadow according to any one of the above items (8) to (13), wherein the method detects a tumor shadow based on the ratio.

(15). A recording medium storing a tumor shadow detection program for detecting a tumor shadow from input image data by a computer, wherein the filter performs a composite product operation with a plurality of filters having different directions on the input image data, Directional balance information is obtained based on an angle between a vector having a plurality of outputs for each direction as a component and a vector (virtual reference vector) having the same component, and a tumor shadow candidate point is obtained based on the obtained directional balance information. A recording medium on which is recorded a program for causing a computer to execute mass shadow detection, characterized by determining a mass shadow detection program.

(16). The tumor shadow according to (15), wherein a standard deviation of a plurality of outputs for each direction of the filter is obtained, and a tumor shadow candidate point is determined based on the obtained standard deviation and the direction balance information. A recording medium on which a detection program is recorded.

(17). (15) The feature amount is calculated based on the pixel intensity, and a tumor shadow candidate point is determined based on the direction balance information and the feature amount.
A recording medium recording the tumor shadow detection program according to any one of the above item or the item (16).

(18). Tumor shadow candidate points are determined by a neural network with a learning function in which different teacher signals are applied to the points near the tumor shadow boundary and points inside the boundary except for the boundary to learn the tumor shadow candidate points. A recording medium on which the tumor shadow detection program according to any one of the above items (15) to (17) is recorded.

(19). In a plurality of directions in the region of the tumor shadow, the number of pixels at which the filter output is maximum in the direction is obtained, and the obtained number of pixels is set to the length, and the vector is configured with the corresponding direction as the direction. Is divided into two half planes, a vector sum of the vectors is obtained in each half plane, a length ratio of the vector sum is obtained, and a tumor shadow is detected based on the obtained vector ratio. A recording medium recording the tumor shadow detection program according to any one of the above (15) to (18).

(20). For the area of the detected tumor shadow, determine the area of the largest inscribed circle inscribed in the area, and when the calculated area is smaller than a predetermined value, detect the tumor shadow based on the directional balance information. A recording medium on which the tumor shadow detection program according to any one of the above items (15) to (19) is recorded.

(21). (15) The method according to (15), wherein the ratio of the area of the tumor shadow region to the area of the largest inscribed circle inscribed in the region is determined, and the tumor shadow is detected based on the determined area ratio. A recording medium on which the tumor shadow detection program according to any one of the items (19) to (19) is recorded.

[0097]

As described above in detail, according to the present invention, a circular area having no directivity can be extracted from a linear area or a background and identified, and a connection can be made in which picked-up tumor shadow candidate points are formed. The region consists of the tumor shadow region and the pixel intensity surface 3
By the similarity determination method and the combination with the dimensional shape,
It is possible to provide a device and a method for detecting a tumor shadow in an image and a recording medium in which a tumor shadow detection program is recorded, which has few overlooked or over-picked areas.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating a configuration of a tumor shadow candidate point extraction unit that extracts a tumor shadow candidate point from an image as a first embodiment of a tumor shadow detection device according to the present invention.

FIG. 2 is a diagram for explaining a Gabor function used in a tumor shadow detection device according to the present invention.

FIG. 3 is a diagram showing an example of a circular open filter used in the first embodiment.

FIG. 4 is a diagram showing a configuration of an excessively picked-up candidate area reducing unit as a second embodiment of the tumor shadow detection apparatus according to the present invention.

FIG. 5 is a diagram for describing directional balance in an evaluation scale tumor shadow candidate area by an intra-area directional balance calculation evaluation unit according to the second embodiment.

FIG. 6 is a diagram for describing an extension degree by an area deletion unit according to the second embodiment.

FIG. 7 is a flowchart illustrating a process of detecting a tumor shadow according to the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Tumor shadow candidate point extraction part 2 ... Too picked up candidate area reduction part 11 ... Image input data part 12 ... Filter calculation part 13 ... Direction balance calculation part 14 ... Feature vector calculation part 15 ... Tumor shadow candidate point determination part 21 ... Connection Processing unit 22: Area evaluation unit 22a: In-region direction balance calculation evaluation unit 22b: Two-dimensional shape evaluation unit 23: Area deletion unit

Claims (3)

[Claims] 1. A filter operation means for performing a composite product operation of a plurality of filters having different directions on input image data, a vector having a plurality of outputs for each direction of the filter operation means as components, Tumor shadow detection, comprising: directional balance calculating means for calculating an angle between the component and the vector having the same component; and tumor shadow candidate point determining means for detecting a tumor shadow based on an output of the directional balance calculating means. apparatus. 2. A filter operation step of performing a composite product operation on input image data with a plurality of filters having different directions, a vector having a plurality of outputs for each direction of the filter as a component, A tumor shadow, comprising: a direction balance calculation step for obtaining direction balance information based on an angle between the vector and an equal vector; and a tumor shadow candidate point determination step for detecting a tumor shadow based on the obtained direction balance information. Detection method. 3. A recording medium for recording a tumor shadow detection program for detecting a tumor shadow from input image data by a computer, wherein a composite product operation of the input image data with a plurality of filters having different directions is performed. Based on the angle between a vector having a plurality of outputs for each direction of the filter as components and a vector having the same component, direction balance information is obtained, and based on the obtained direction balance information, a tumor shadow candidate point A recording medium on which a program for causing a computer to execute mass shadow detection, characterized by determining a mass shadow, is recorded.