Breast cancer detection system, breast cancer detection method, breast cancer detection program, and computer-readable recording medium having breast cancer detection program recorded thereon

Abstract

A model calculating device calculates a mammary gland normal architecture model that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to the X-ray image of the breast. An orientation extracting device extracts linear components orientations of a region image texture that form a shape of a shadow in the breast X-ray image using a Gabor filter. A lesion determining device compares the mammary gland orientation in the normal architecture model calculated by the model calculating device and the orientation extracted by the orientation extracting device with respect to a region of interest in the X-ray image of the breast including architectural distortion candidates of the mammary gland detected by a preprocessing device to calculate a feature quantity based on a difference between the orientations and determines whether the candidates are an architectural distortion of the mammary gland based on the feature quantity.

Claims

1. A breast cancer detection system for detecting an architectural distortion of the mammary gland included in an X-ray image of the breast, the breast cancer detection system comprising: a processor programmed to: calculate a normal architecture model of the mammary gland that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to the X-ray image of the breast; extract orientations of linear components of an image texture of a region that form a shape of a shadow in the X-ray image of the breast; and compare the orientation of the mammary gland in the calculated normal architecture model and the extracted orientation with respect to a region of interest in the X-ray image of the breast including architectural distortion candidates of the mammary gland detected by preprocessing to: (i) calculate a feature quantity based on a difference between the orientations, and (ii) determine whether the candidates are an architectural distortion of the mammary gland based on the feature quantity.

2. The breast cancer detection system according to claim 1, wherein the processor calculates the feature quantity R according to the following equation in which, for a plurality of coordinates (x,y) in the region of interest, Φ(x,y) is the orientation of the mammary gland in the calculated normal architecture model, Θ(x,y) is the extracted orientation, N.sub.m is a parameter for determining an allowable range of matching orientations, and N.sub.ROI is the number of coordinates $\begin{array}{cc}\left[\mathrm{Equation}1\right]& \\ R=\frac{\underset{x}{.Math.}\underset{y}{.Math.}h\left(x,y\right)}{N_{\mathrm{ROI}}}\mathrm{where}h\left(x,y\right)=\{\begin{array}{cc}0,& .Math.\Phi \left(x,y\right)-\Theta \left(x,y\right).Math.\leqq \frac{\pi }{N_m}\\ 1,& \mathrm{otherwise}\end{array}.& \end{array}$

3. The breast cancer detection system according to claim 2, wherein the processor extracts orientations of linear components of an image texture of a region in the X-ray image of the breast using a Gabor filter.

4. The breast cancer detection system according to claim 3, wherein the processor calculates the normal architecture model of the mammary gland using an exponential curve.

5. The breast cancer detection system according to claim 2, wherein the processor calculates the normal architecture model of the mammary gland using an exponential curve.

6. The breast cancer detection system according to claim 1, wherein the processor extracts orientations of linear components of an image texture of a region in the X-ray image of the breast using a Gabor filter.

7. The breast cancer detection system according to claim 6, wherein the processor calculates the normal architecture model of the mammary gland using an exponential curve.

8. The breast cancer detection system according to claim 1, wherein the processor calculates the normal architecture model of the mammary gland using an exponential curve.

9. A breast cancer detection method for detecting an architectural distortion of the mammary gland included in an X-ray image of the breast, comprising: a model calculating step of calculating, by a computer, a normal architecture model of the mammary gland that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to the X-ray image of the breast; an orientation extracting step of extracting, by the computer, an orientation of linear components of an image texture of a region that form a shape of a shadow in the X-ray image of the breast; and a lesion determining step of comparing, by the computer, the orientation of the mammary gland in the normal architecture model calculated by the model calculating step and the orientation extracted by the orientation extracting step with respect to a region of interest in the X-ray image of the breast including architectural distortion candidates of the mammary gland detected by preprocessing to: (i) calculate a feature quantity based on a difference between the orientations, and (ii) determine whether the candidates are an architectural distortion of the mammary gland based on the feature quantity.

10. The breast cancer detection method according to claim 9, wherein the lesion determining step calculates the feature quantity R according to the following equation in which, for a plurality of coordinates (x,y) in the region of interest, Φ(x,y) is the orientation of the mammary gland in the normal architecture model calculated by the model calculating step, Θ(x,y) is the orientation extracted by the orientation extracting step, N.sub.m is a parameter for determining an allowable range of matching orientations, and N.sub.ROI is the number of coordinates $\begin{array}{cc}\left[\mathrm{Equation}2\right]& \\ R=\frac{\underset{x}{.Math.}\underset{y}{.Math.}h\left(x,y\right)}{N_{\mathrm{ROI}}}\mathrm{where}h\left(x,y\right)=\{\begin{array}{cc}0,& .Math.\Phi \left(x,y\right)-\Theta \left(x,y\right).Math.\leqq \frac{\pi }{N_m}\\ 1,& \mathrm{otherwise}\end{array}.& \end{array}$

11. The breast cancer detection method according to claim 10, wherein the orientation extracting step extracts orientations of linear components of an image texture of a region in the X-ray image of the breast using a Gabor filter.

12. The breast cancer detection method according to claim 11, wherein the model calculating step calculates the normal architecture model of the mammary gland using an exponential curve.

13. The breast cancer detection method according to claim 10, wherein the model calculating step calculates the normal architecture model of the mammary gland using an exponential curve.

14. The breast cancer detection method according to claim 9, wherein the orientation extracting step extracts orientations of linear components of an image texture of a region in the X-ray image of the breast using a Gabor filter.

15. The breast cancer detection method according to claim 14, wherein the model calculating step calculates the normal architecture model of the mammary gland using an exponential curve.

16. The breast cancer detection method according to claim 9, wherein the model calculating step calculates the normal architecture model of the mammary gland using an exponential curve.

17. A non-transitory computer-readable medium storing a breast cancer detection program thereon, for detecting an architectural distortion of the mammary gland included in an X-ray image of the breast, the program causing a computer to perform steps comprising: a model calculating step that calculates a normal architecture model of the mammary gland that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to the X-ray image of the breast; an orientation extracting step that extracts orientations of linear components of an image texture of a region that form a shape of a shadow in the X-ray image of the breast; and a lesion determining step that compares the orientation of the mammary gland in the normal architecture model calculated by the model calculating step and the orientation extracted by the orientation extracting step with respect to a region of interest in the X-ray image of the breast including architectural distortion candidates of the mammary gland detected by preprocessing to: (i) calculate a feature quantity based on a difference between the orientations, and (ii) determine whether the candidates are an architectural distortion of the mammary gland based on the feature quantity.

Description

BRIEF DESCRIPTION OF EMBODIMENTS

(1) FIG. 1 is a block diagram illustrating a configuration of a breast cancer detection system according to an embodiment of the present invention.

(2) FIG. 2 is an explanatory diagram illustrating a method of calculating a normal architecture model in the breast cancer detection system illustrated in FIG. 1.

(3) FIG. 3 is a flowchart illustrating the flow of processes of the breast cancer detection system illustrated in FIG. 1.

(4) FIG. 4(a) illustrates an X-ray image of the breast received by the breast cancer detection system illustrated in FIG. 1, FIG. 4(b) illustrates an orientation Φ of the mammary gland in a normal architecture model obtained for the X-ray image of the breast in FIG. 4(a), and FIG. 4(c) illustrates an image illustrating an orientation Θ of linear components extracted from the X-ray image of the breast in FIG. 4(a).

(5) FIG. 5(a) illustrates an X-ray image of the breast received by the breast cancer detection system illustrated in FIG. 1, FIG. 5(b) illustrates an image illustrating an orientation Θ of linear components of region A in FIG. 5(a), FIG. 5(c) illustrates an image illustrating an orientation Θ of linear components of region B in FIG. 5(a), FIG. 5(d) illustrates an image illustrating an orientation Φ of the mammary gland in a normal architecture model of region A in FIG. 5(a), FIG. 5(e) illustrates an image illustrating an orientation Φ of the mammary gland in a normal architecture model of region B in FIG. 5(a), FIG. 5(f) illustrates an image illustrating h(x,y) of region A in FIG. 5(a), and FIG. 5(g) illustrates an image illustrating h(x,y) of region B in FIG. 5(a).

(6) FIG. 6 is a graph illustrating FROC curves obtained by the breast cancer detection system illustrated in FIG. 1.

DESCRIPTION OF EMBODIMENTS

(7) Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

(8) FIGS. 1 to 6 illustrate a breast cancer detection system, a breast cancer detection method, a breast cancer detection program, and a computer-readable recording medium having the breast cancer detection program recorded thereon according to an embodiment of the present invention. The breast cancer detection method according to the embodiment of the present invention is a method which is suitably implemented by the breast cancer detection system according to the embodiment of the present invention and can be executed by a computer with the aid of a computer-readable recording medium such as a hard disk, a DVD-ROM, a CD-ROM, and a flexible disk, having the breast cancer detection program recorded thereon.

(9) As illustrated in FIG. 1, the breast cancer detection system according to the embodiment of the present invention is a breast cancer detection system for detecting architectural distortions of the mammary gland included in X-ray images of the breast. The breast cancer detection system is implemented as a computer which includes a receiving means 11, a storage means 12 that stores various types of data, a main controller 13 having arithmetic and control functions, and an output means 14.

(10) The receiving means 11 is connected to a mammographic X-ray imaging device (mammography) so as to be able to receive X-ray images of the breast captured by the mammographic X-ray imaging device. The receiving means 11 may be connected to a medical X-ray image server rather than the mammographic X-ray imaging device so as to be able to receive X-ray images of the breast. Moreover, the receiving means 11 may be configured to be able to receive X-ray images of the breast stored in a storage medium such as a CD-R or a DVD-R from the corresponding readers. The storage means 12 is configured with a memory and is configured to store the X-ray images of the breast received from the receiving means 11.

(11) The main controller 13 is implemented as a CPU and is connected to the receiving means 11, the storage means 12, and the output means 14 so as to be able to control the respective components. The main controller 13 includes a preprocessing means 21, a model calculating means 22, an orientation extracting means 23, and a lesion determining means 24.

(12) [Preprocessing Means]

(13) The preprocessing means 21 is configured with the CAD system disclosed in Non-Patent Literature 3, capable of detecting as many lesion candidates as possible with a high true-positive fraction. The preprocessing means 21 includes a breast region extracting means 25, a DoG filtering means 26, and a threshold processing means 27. The breast region extracting means 25 determines the boundary between a pectoral region and a breast region using an image processing method of tracking high edge intensity points and removes a pectoral region in which the mammary gland is not present, from the X-ray image of the breast. Further, the breast region extracting means 25 reduces the image excluding pectoral muscles to the ¼ scale of the original image using a bicubic method in order to reduce the computation time.

(14) The central portion of a shadow of an architectural distortion tends to have a certain extent of contrast in such a way that the luminance intensity is generally higher or lower than that of the peripheral portion. The DoG filtering means 26 performs DoG (Difference of Gaussians) filtering on the X-ray image of the breast to detect a local contrast in order to extract brightness/darkness information associated with such an architectural distortion.

(15) The DoG filtering means 26 detects the local contrast in the following manner. That is, first, the convolution of a 2-dimensional Gaussian kernel G(xG,yG,σ) having a standard deviation σ and an input image I(x,y) is taken to obtain a smoothed image L(x,y,σ) using the following equation.
L(x,y,σ)=G(xG,yG,σ)*I(x,y)  (2)

(16) Here, xG and yG indicate the distance of the kernel from a target position (x,y), respectively. Moreover, the Gaussian kernel is defined by the following equation.

(17) $\begin{array}{cc}\left[\mathrm{Equation}4\right]& \\ G\left(x_G,y_G,\sigma \right)=\frac{1}{2{\mathrm{\pi \sigma }}^2}\exp \left(-\frac{x_G^2+y_G^2}{2{\sigma }^2}\right)& \left(3\right)\end{array}$

(18) Subsequently, a difference between two images having different smoothing levels is calculated according to Equation (4) to obtain a DoG image D.sub.k(x,y,σ).
D.sub.k(x,y,σ)=L(x,y,kσ)−L(x,y,σ)  (4)

(19) k is a parameter that determines the standard deviation (that is, the ratio between smoothing levels) of the 2-dimensional Gaussian kernel. Since the size (area) of a detection target lesion can be controlled by the parameter σ, the DoG filtering means 26 preferably set the parameter σ appropriate for the size of the detection target lesion.

(20) A region having a locally high contrast within the X-ray image of the breast has a high value on an image D.sub.k obtained after the DoG filtering. Thus, The threshold processing means 27 performs threshold processing in order to detect peak values in the image obtained after the DoG filtering. In this case, a threshold is set so as to detect a larger number of peak values while suppressing the area of each region. Moreover, threshold processing is performed on the area of the remaining area in a two-valued image obtained after the threshold processing in order to remove regions having an extremely large or small area. In this way, the preprocessing means 21 extracts the candidates for architectural distortions of the mammary gland.

(21) [Model Calculating Means]

(22) In the case of a normal mammary gland architecture, linear components of an image texture of a region present on a breast region scatter in an approximately fan shape from the nipple to the greater pectoral muscle. The model calculating means 22 calculates a normal architecture model of the mammary gland that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to a target X-ray image of the breast in order to model the normal mammary gland architecture. First, the model calculating means 22 detects the nipple. In a craniocaudal (CC) image, the greater pectoral muscle is assumed to be vertical and a point farthest from an image end is assumed to be the nipple position. In a mediolateral oblique (MLO) image, a point farthest from a straight line passing through the end points of the greater pectoral muscle is assumed to be the nipple position.

(23) Subsequently, in a coordinate (x,y) illustrated in FIG. 2 when the nipple is at the bottom and the pectoral muscle is on the top, the linear component of a normal mammary gland is approximated by such an exponential curve as the following equation based on the nipple position.
y=f(x)=Aexp(Bx)  (5)

(24) Here, A and B are parameters of an exponential function, and the coordinate of the nipple serving as the reference is (0, A). On the other hand, among the exponential curves of Equation (5) that passes through the nipple, the parameter B.sub.i of a curve that passes through a certain point (x.sub.i,y.sub.b) on the pectoral muscle side illustrated in FIG. 2 is expressed by the following equation.
B.sub.i=B=(1/x.sub.i)log(y.sub.b/A)  (6)

(25) Here, y.sub.b is the y-coordinate when the pectoral muscle is approximated to a straight line.

(26) If the length of the pectoral muscle (that is, the number of pixels present between points P.sub.1 and P.sub.2 in FIG. 2) is N.sub.f, an exponential curve that passes through the point (x.sub.i,y.sub.b) (i=1, 2, . . . , N.sub.f) is expressed by the following equation.
y=f.sub.i(x)=Aexp(B.sub.ix)  (7)

(27) The model calculating means 22 calculates the normal architecture model of the mammary gland by moving the point (x.sub.i,y.sub.b) from point P.sub.1 to point P.sub.2 by a step of one pixel to calculate B.sub.i according to Equation (6) and calculate the exponential function f.sub.i according to Equation (7).

(28) Moreover, the model calculating means 22 calculates an angle Φ between the tangent of the exponential function calculated by Equation (7) and the positive direction of the x-axis as the orientation of the mammary gland of the normal architecture model. Φ is given by the following equation.
Φ(x,.sup.iy)=arctan(f′.sub.i(x))=arctan(AB.sub.iexp(B.sub.ix))  (8)

(29) Here, .sup.iy is the y-coordinate that satisfies y=f.sub.i(x). Since Equation (8) cannot be applied to pixels through which the exponential curve does not pass, the pixels are processed by assuming that the pixels through which the exponential curve does not pass have the angle values of pixels which are located closest to the pixels.

(30) [Orientation Extracting Means]

(31) The orientation extracting means 23 extracts the orientation of the linear components that form the shape of the shadow in the X-ray image of the breast using Gabor filters. Gabor filters are used as orientation component feature extraction filters in many image recognition systems such as fingerprint authentication or iris recognition and are widely used in processing medical images such as X-ray images of the breast or computerized tomographic images.

(32) The output g of a Gabor filter is expressed by Equation (9).

(33) $\begin{array}{cc}\left[\mathrm{Equation}5\right]& \\ g\left(x,y\right)=\exp \left(-\frac{1}{2{\sigma }_G^2}\left(x^{\mathrm{\prime 2}}+\gamma y^{\mathrm{\prime 2}}\right)\right)\cos \left(2\pi \frac{x^{\prime }}{\lambda }\right)\mathrm{where}{x}^{\prime }=x\cos \theta +y\sin \theta y^{\prime }=-x\sin \theta +y\cos \theta & \left(9\right)\end{array}$

(34) Here, x and y are coordinates, θ is an angle, σ is a dispersion, γ is an aspect ratio, and λ is a wavelength.

(35) A Gabor filter kernel having an inclination of an i-th angle θ.sub.i=πi/180 (i=1, 2, . . . , m) among m discretized angles (m is a natural number) of 180 degrees is defined as g.sub.i. In this case, a convolutional integration between g, and the original image I is expressed by I.sub.i(x,y)=(I*g.sub.i)(x,y), and an orientation map Θ(x,y) of the original image I(x,y) is defined by the following equation.

(36) $\begin{array}{cc}\left[\mathrm{Equation}6\right]& \\ \Theta \left(x,y\right)=\frac{\pi }{180}{i}_{\max }\left(x,y\right)\mathrm{where}{i}_{\max }\left(x,y\right)=\underset{i}{\arg \max }\left[.Math.I_i\left(x,y\right).Math.\right]& \left(10\right)\end{array}$

(37) That is, the orientation extracting means 23 calculates the orientation map Θ(x,y) by assuming that an angle πi.sub.max/180 corresponding to i that gives the largest value among m convolutional integration results |I.sub.i(x,y)| is an orientation component of a pixel at the position (x,y).

(38) [Lesion Determining Means]

(39) The lesion determining means 24 compares the orientation Φ(x,y) of the mammary gland in the normal architecture model calculated by the model calculating means 22 and the orientation Θ(x,y) extracted by the orientation extracting means 23 with respect to a region of interest in the X-ray image of the breast including the candidates for architectural distortions of the mammary gland detected by the preprocessing means 21, calculates a feature quantity R based on the difference between the two orientations according to Equation (1), and determines whether the architectural distortion candidates are the architectural distortions of the mammary gland based on the feature quantity R. In calculation of the feature quantity R, h(x,y) is calculated for all coordinates (x,y) in the region of interest. In this case, N.sub.ROI is the number of all pixels.

(40) The feature quantity R means that a strict match only is allowed when a parameter N.sub.m that determines an allowable range of the matching orientations has a large value whereas a rough match is also allowed when N.sub.m has a small value. When the orientation map Θ(x,y) matches the orientation Φ(x,y) of the mammary gland in the normal architecture model, the feature quantity R decreases. When architectural distortion shadows are included, since the proportion of the orientation map of which the orientations do not match the orientations of the normal mammary gland increases, the feature quantity R increases. The lesion determining means 24 sets a threshold to the feature quantity R and determines that the architectural distortion candidates are architectural distortions of the mammary gland when the feature quantity R is larger than the threshold.

(41) The main controller 13 transmits the image of the region of interest including the architectural distortion candidates extracted by the preprocessing means 21, the calculation result of the Gabor filter, the calculation result of the feature quantity R, the image of the region of interest determined to be architectural distortions by the lesion determining means 24, and the like to the storage means 12 so that the images and the calculation results are stored in the storage means 12.

(42) The output means 14 is configured with a monitor or a printer. The output means 14 is configured to be able to output the X-ray images of the breast stored in the storage means 12, the image of the region of interest including the architectural distortion candidates extracted by the preprocessing means 21, the calculation result of the Gabor filter, the calculation result of the feature quantity R, the image of the region of interest determined to be architectural distortions by the lesion determining means 24, and the like to a monitor or a printer.

(43) Next, the flow of processes of the breast cancer detection system according to the embodiment of the present invention will be described with reference to FIG. 3 to FIGS. 5(a) to 5(g). As illustrated in FIG. 3, first, the receiving means 11 receives X-ray images of the breast captured by a mammographic X-ray imaging device (step 31), and the preprocessing means 21 of the main controller 13 extracts architectural distortion candidates of the mammary gland. FIG. 4(a) illustrates an example of the X-ray image of the breast. In the preprocessing means 21, first, the breast region extracting means 25 removes the pectoral region from the X-ray image of the breast to extract the breast region (step 32), and the DoG filtering means 26 performs DoG filtering to detect a local contrast (step 33). Subsequently, the threshold processing means 27 performs threshold processing on the image obtained after the DoG filtering to extract the architectural distortion candidates of the mammary gland (step 34).

(44) After the processing of the preprocessing means 21 ends, the model calculating means 22 calculates a normal architecture model of the mammary gland that scatters in a fan shape from the nipple toward the greater pectoral muscle with respect to the received X-ray image of the breast to calculate the orientation Φ(x,y) of the mammary gland (step 35). Moreover, the orientation extracting means 23 extracts the orientation Θ(x,y) of the linear components of an image texture of a region that form the shape of the shadows in the received X-ray image of the breast using Gabor filters (step 36). FIGS. 4(b) and 4(c) illustrate an image illustrating the orientation Φ(x,y) of the mammary gland in the normal architecture model calculated for the X-ray image of the breast of FIG. 4(a) and an image illustrating the extracted orientation Θ(x,y) of the linear components. In FIG. 4(c), m in Equation (10) is 12.

(45) Subsequently, the lesion determining means 24 calculates a feature quantity R using Equation (1) from the orientation Φ(x,y) of the mammary gland in the normal architecture model calculated by the model calculating means 22 and the orientation Θ(x,y) extracted by the orientation extracting means 23 with respect to the region of interest in the X-ray image of the breast including the architectural distortion candidates of the mammary gland detected by the preprocessing means 21 (step 37). Further, the lesion determining means 24 determines whether the architectural distortion candidates are architectural distortions of the mammary gland based on the feature quantity R (step 38).

(46) FIGS. 5(a) to 5(g) illustrate examples of the feature quantity R calculated for a region (region A in FIG. 5(a)) including architectural distortions of the X-ray image of the breast of FIG. 4(a) and a region (region B in FIG. 5(a)) of the normal mammary gland architecture. FIG. 5(f) illustrates h(x,y) of Equation (1), calculated from the orientation Θ(x,y) illustrated in FIG. 5(b) and the orientation Φ(x,y) of the mammary gland in the normal architecture model illustrated in FIG. 5(d) with respect to region A including architectural distortions. The feature quantity R calculated from h(x,y) was 0.99. Moreover, FIG. 5(g) illustrates h(x,y) of Equation (1), calculated from the orientation Θ(x,y) illustrated in FIG. 5(c) and the orientation Φ(x,y) of the mammary gland in the normal architecture model illustrated in FIG. 5(e) with respect to region B of the normal mammary gland architecture. The feature quantity R calculated from h(x,y) was 0.59. When a value such as 0.8, for example is set as the threshold of the feature quantity R, it is possible to accurately determine architectural distortions of the mammary gland illustrated in FIGS. 5(a) to 5(g).

(47) The breast cancer detection system according to the embodiment of the present invention calculates a normal architecture model of the mammary gland, calculates a feature quantity of architectural distortion candidates in the X-ray image of the breast based on the normal architecture model, and determines whether the architectural distortion candidates are architectural distortions based on the feature quantity. Thus, it is possible to accurately quantify the features of lesions by taking the positions of architectural distortion candidates on the breast region into consideration and to make determination. When the feature quantity is the difference between the orientation of the mammary gland in the normal architecture model and the orientation of the linear components that form the shape of the shadows in the X-ray image of the breast, the feature quantity can be easily calculated by a difference calculation. Thus, the feature quantity of architectural distortions such as speculations wherein the mammary gland scatters in a radial form from one point can be quantified with high accuracy. In this manner, the breast cancer detection system according to the embodiment of the present invention can reduce a false-positive fraction and accurately detect architectural distortions of the mammary gland as compared to a case where the normal mammary gland architecture is not taken into consideration.

(48) By using R in Equation (1) as the feature quantity, it is possible to accurately identify the features of architectural distortions. By adjusting the value of the parameter N.sub.m in Equation (1) according to the size, the contrast, and the like of noise in the X-ray image of the breast, it is possible to secure optimal detection performance. Moreover, it is possible to adjust a true-positive fraction and a false-positive fraction according to the value of the threshold of the feature quantity R used when determining architectural distortions.

(49) In the breast cancer detection system according to the embodiment of the present invention, although the CAD system disclosed in Non-Patent Literature 3 is used as the preprocessing means 21, other conventional CAD systems disclosed in Patent Literature 2, Non-Patent Literature 1, or Non-Patent Literature 2 may be used. Moreover, the normal architecture model of the mammary gland used in the breast cancer detection system according to the embodiment of the present invention is not limited to the model which uses an exponential curve but may be an optional model as long as the mammary gland scatters in a fan shape from the nipple toward the greater pectoral muscle.

(50) The breast cancer detection program according to the embodiment of the present invention is provided in a form of being recorded on a computer-readable recording medium such as CD (CD-ROM, CD-R, CD-RW, and the like), and DVD (DVD-ROM, DVD-RAM, DVD-R, DVD-RW, DVD+R, DVD+RW, and the like), for example. In this case, a computer can read the breast cancer detection program from the recording medium, transmit and store the program to and in an internal or external storage means of the computer, and use the program. Moreover, the breast cancer detection program according to the embodiment of the present invention may be recorded on a storage means (recording medium) such as a magnetic disk, an optical disc, or an magneto-optical disc and be provided from the storage means to a computer via a communication line.

(51) Here, the computer is a concept that includes hardware and an operating system (OS) and means hardware operating under the control of the OS. Moreover, when an OS is not required and hardware is operated by an application program only, the hardware itself corresponds to the computer. Hardware includes at least a microprocessor such as a CPU and means for reading a computer program recorded on a recording medium.

(52) An application program as the breast cancer detection program according to the embodiment of the present invention includes program codes for causing a computer to realize the above-described functions. Moreover, some of the functions may be realized by an OS rather than the application program. Various compute-readable media such as an internal storage means (memory such as RAM or ROM), an external storage means, and the like of the computer or a print on which symbols such as barcodes are printed may be used as the computer-readable recording medium according to the embodiment of the present invention in addition to the flexible disk, CD, DVD, magnetic disk, optical disc, and magneto-optical disc.

PRACTICAL EXAMPLE 1

(53) The breast cancer detection system according to the embodiment of the present invention was applied to clinical data and the architectural distortion detection performance was examined. DDSM (digital database for screening mammography) which is a worldwide standard database was used as performance evaluation data, and 100 X-ray images of the breast including 50 cases including architectural distortions and 50 normal case examples were selected from the DDSM and were used. The spatial resolution of these images was 0.05 mm/pixel and the density resolution thereof was 12 bits.

(54) FROC (free-response receiver operating characteristic) curves were used in evaluation of the detection performance. The FROC curve is a graph of which the horizontal axis represents the number of false positives per image (FPI) and the vertical axis represents true-positive fraction (TPF). The FROC curve is frequently used for performance evaluation when a plurality of false positives occur as in mammography CAD. In the FROC curve, it is determined that the corner is closer to the top-left corner of the graph, the higher the detection performance of the curve is.

(55) K=1.6, σ=37.5, 50, 62.5, 75, and 87.5 were used as the parameters of the DoG filtering. λ=4, γ=1/256, σ=λ/{2(2 ln 2).sup.1/2}, and m=12 were used as the parameters of the Gabor filter, the line width was set to 0.8 mm, and the length was set to 12.8 mm.

(56) The calculated FROC curve is illustrated in FIG. 6. Moreover, comparison FROC curves calculated by the CAD systems of Non-Patent Literatures 1 to 3 are also illustrated in FIG. 6. The result obtained by Non-Patent Literature 3 is the result obtained by the preprocessing means 21 of the breast cancer detection system according to the embodiment of the present invention, and the same parameters as those of the breast cancer detection system according to the embodiment of the present invention are used as the parameters of the DoG filtering.

(57) As illustrated in FIG. 6, it was confirmed that the FROC curve of the breast cancer detection system according to the embodiment of the present invention was positioned closer to the top-left corner than the curves of the other methods and had a performance superior to that of the conventional systems. Moreover, in the CAD system of Non-Patent Literature 3, the number of false positives per image was 7.2 at a true-positive fraction of 80%. In the breast cancer detection system according to the embodiment of the present invention, since the normal architecture model of the mammary gland was introduced and the feature quantity R was used, the number of false positives per image was improved remarkably to 3.1 at the true-positive fraction of 80%. Therefore, it can be said that the breast cancer detection system according to the embodiment of the present invention can reduce the false-positive fraction remarkably while maintaining the true-positive fraction and improve the overall performance.

REFERENCE SIGNS LIST

(58) 11: Receiving means

(59) 12: Storage means

(60) 13: Main controller

(61) 21: Preprocessing means

(62) 25: Breast region extracting means

(63) 26: DoG filtering means

(64) 27: Threshold processing means

(65) 22: Model calculating means

(66) 23: Orientation extracting means

(67) 24: Lesion determining means

(68) 14: Output means