JP2006325937A - Image determination device, image determination method, and program therefor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve a precision when determining the abnormal shadow of lung cancer or the like. <P>SOLUTION: The candidate area of the abnormal shade is detected from a medical image. A small area containing at least a part of the candidate area is set as a remarkable area M and moreover, small areas existing near the remarkable area M are set as nearby areas R1, R2, ..., R8. The remarkable area M and the nearby areas R1, R2, ..., R8 generate normal artificial images, and a difference between the artificial image and the original image of the remarkable area M (minimum square distance) is compared with a difference between the artificial images and the original images of the nearby areas R1, R2, ..., R8 (minimum square distance). According to the presence/absence of the difference between both of the differences, it is determined whether a candidate area included in the remarkable area M is the abnormal shade. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to an image determination apparatus, an image determination apparatus method, and a program for supporting detection of abnormal shadow candidates by interpretation of medical images.

  Conventionally, in the medical field, two or more medical images with different shooting times for a patient are comparatively read, the difference between the images is examined, an abnormal shadow is detected based on the difference, and the progression of the disease The treatment policy is examined by grasping the situation and the healing situation.

  In particular, because small shadows such as lung cancer that have abnormal shadows in the early stages are easy to overlook, when detecting such abnormal shadows, subtraction that associates pixels between the images for two medical images with different shooting times A method for obtaining a difference by performing processing or the like, and detecting a region having a pixel value equal to or greater than a predetermined value or a characteristic shape, for example, a substantially circular shape as an abnormal shadow candidate in a difference image (temporal subtraction image) representing the difference Has been proposed (Patent Document 1, etc.).

  However, two time-series images with different imaging timings to be compared are subject to positional deviation due to changes in the posture of the patient at the time of imaging. Alignment is performed. As a method of alignment, global alignment such as rotation, translation, enlargement / reduction, etc., for at least one of the two images of the same subject to be compared (linear transformation such as affine transformation) ), Non-linear distortion transformation based on the corresponding positional relationship obtained by template matching for each local region (for example, non-linear distortion transformation using curve fitting by a two-dimensional polynomial) processing (warping), or these global positioning There have been proposed a method of combining positioning and local positioning, and a method of performing positioning again in a local region having a high degree of positional deviation and its vicinity (Patent Document 2, Patent Document 3, etc.).

  In the above method, since alignment is performed using images after being projected in two dimensions, large three-dimensional fluctuations (forward tilt, backward tilt, When the lateral rotation) occurs, the alignment cannot be completed, and an artifact (false image) is generated on the difference image, and the abnormal shadow is not accurately detected. In addition, in the case of patients who do not have medical images taken in the past, there is no image to be compared, and the technique based on difference images that is effective when detecting abnormal shadows that are easily overlooked, such as lung cancer, can be used well. I couldn't.

Therefore, each image in the database in which the normal structure image representing the normal structure of the subject is stored and stored, and the shape of the anatomical feature extracted from each image are used as the teacher data, and the normal structure statistics of the subject are used. Model (shape change model, texture change model, shape and texture correlation model), and using this model, a normal structure image corresponding to the input image was artificially generated and generated An apparatus that detects a candidate area for an abnormal shadow based on a normal structure image and an input medical image has been proposed (Patent Document 4 and the like).
JP 2002-158923 A JP-A-7-37074 JP 2002-32735 A JP 2004-41694 A

  However, the above-described method for artificially generating the normal structure image can be used even when a large three-dimensional variation occurs due to a change in the posture of the patient between images, or even in the case of a patient for which no medical image has been taken in the past. It is effective for detecting small abnormal shadows that are easily overlooked. Usually, chest CAD (computer aided diagnose) cuts out the detected lung cancer candidates into regions of interest (ROI) for each ROI. In other words, it is inefficient to generate an artificial image of the entire chest because it is determined whether or not the patient has lung cancer.

  In addition, since the features that appear in the part where the structural part such as the rib is present and the part where the structural object such as the rib does not exist are different, when an artificial image of the entire chest is generated, it is partially different from the original image. There may be cases where a characteristic is present.

  Therefore, it is desired to improve the detection of an abnormal shadow that is easily overlooked, such as lung cancer, by accurately generating an artificial image corresponding to a region of interest in which an abnormal shadow is detected instead of the entire chest.

  The present invention has been made in view of the above circumstances, and an image determination apparatus, an image determination method, and an image determination method capable of determining an abnormal shadow more efficiently and accurately than a conventional method using comparative interpretation. The purpose is to provide the program.

The image discriminating apparatus of the present invention is a medical image including a structure having anatomical features obtained by photographing a predetermined part of a subject, and a plurality of small regions obtained by subdividing the medical image are extracted from the structure. A classification means for classifying the obtained anatomical features into small regions similar to each other;
Artificial image generation means for generating an artificial image of a normal small area of the structure according to a classification result of the classification means for at least one of the classified small areas;
Candidate area detecting means for detecting a candidate area of an abnormal shadow from the medical image;
A region setting unit that sets a small region including at least a part of the candidate region detected by the candidate region detection unit as a region of interest, and further sets a small region existing in the vicinity of the region of interest as a neighborhood region;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. And determining means for determining whether or not.

Further, the image discrimination method of the present invention is a medical image including a structure having an anatomical feature obtained by photographing a predetermined part of a subject, and a plurality of small regions obtained by subdividing the medical image are represented by the structure. A classification means for classifying each anatomical feature obtained from an object into small regions similar to each other;
In an image determination apparatus comprising: an artificial image generation unit that generates an artificial image of a normal small region of the structure according to a classification result of the classification unit for at least one small region of the classified small regions ,
A candidate area detecting step of detecting a candidate area of an abnormal shadow from the medical image;
An area setting step for setting a small area including at least a part of the candidate area detected by the candidate area detection step as an attention area, and further setting a small area in the vicinity of the attention area as a vicinity area;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. And a determination step for determining whether or not.

Further, the program of the present invention provides a computer with a large number of small regions obtained by subdividing the medical image in a medical image including a structure having an anatomical feature obtained by photographing a predetermined part of the subject. A classifying means for classifying the anatomical features obtained from the structure into small areas similar to each other;
Artificial image generation means for generating an artificial image of a normal small area of the structure according to a classification result of the classification means for at least one of the classified small areas;
Candidate area detecting means for detecting a candidate area of an abnormal shadow from the medical image;
A region setting unit that sets a small region including at least a part of the candidate region detected by the candidate region detection unit as a region of interest, and further sets a small region existing in the vicinity of the region of interest as a neighborhood region;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. It is made to function as a determination means for determining whether or not.

  The “predetermined portion of the subject” refers to a portion of the human body or animal body such as the chest, head, or upper arm viewed from the viewpoint of which side it is located.

  “Structure” refers to a tissue that constitutes a predetermined part of a subject.

  “Similar anatomical features obtained from a structure” means that the features obtained from the location of each tissue and the configuration of each tissue are similar. For example, the chest is composed of tissues such as the ribs, soft parts, and heart, but the pixel value density differs between the location where only one rib is imaged and the location where two ribs are imaged. Although the features are not similar, the locations where only one rib is imaged are close in pixel value density, resulting in images with similar anatomical features. Also, depending on the position of the chest lung field area, such as whether it is the marginal part or the mediastinum part, the resulting density and texture will differ and the image will be similar, but if within the same marginal part, the ribs A similar image appears depending on whether one is overlapping or two overlapping.

  Further, “the structure is normal” means that the tissue constituting each part is normal.

  “Set a small region including at least a part of a candidate region as a region of interest” means setting a candidate region detected by the candidate region detection means within the small region to be set as a region of interest. The entire candidate area may not be included in the small area but may include a part of the candidate area.

  “Set a small area near the attention area as a neighboring area” sets the small area so that it overlaps the attention area even if the small area is set so that the neighboring area is adjacent to the attention area. You may do.

  “Original image” refers to a medical image obtained by photographing.

In addition, the determination unit is configured to determine a difference feature amount according to a difference between the artificial image of the attention area and the original image of the attention area, and a difference between the population image of the vicinity area and the original image of the vicinity area. Having a difference feature amount calculating means for calculating a difference feature amount,
The difference feature quantity of the attention area calculated by the difference feature quantity calculation means is compared with the difference feature quantity of the neighboring area, and the candidate area included in the attention area is an abnormal shadow depending on whether there is a difference between the two feature quantities. It is possible to use what determines whether or not there is.

  The difference feature quantity may be a least square distance between the artificial image and the original image or a normalized cross-correlation value.

Further, the artificial image generating means has a unique image storage means for storing a plurality of unique images having unique features that appear for each small area when the classified structure of each small area is normal. And
Weighting is performed according to the degree of similarity between the small area for generating the artificial image and each of the plurality of unique images, and the artificial image of the small area is generated by adding the plurality of unique images. Can do.

  As the “unique image”, for example, by using a plurality of images of different normal structures as samples and using a statistical method, an image having a characteristic that appears in each small region when the structure is normal is acquired. It can be a unique image.

  Further, the unique image stored in the unique image storage means is an image obtained by performing principal component analysis based on an image of each small region where the structure of the medical image obtained by photographing a subject is normal. desirable.

When the medical image is a chest image obtained by photographing the chest of the subject,
The classification means is
Lung field recognition means for recognizing a lung field region from the chest image;
Rib recognition means for recognizing the rib from the chest image,
It is desirable to classify the large number of small regions according to the position in the lung field region and the overlap of the ribs.

  According to the present invention, from a medical image obtained by photographing a predetermined part of a subject such as a human body, a large number of subdivided small regions are classified into small regions having similar anatomical features obtained from the structure. Since an artificial image corresponding to the characteristics of each area can be generated by generating an artificial image corresponding to the classified small area, when a candidate area for an abnormal shadow is detected, the candidate area and its neighboring area By comparing the artificial image and the original image in, it is possible to accurately determine whether or not the candidate area is an abnormal shadow rather than an artifact or the like. Furthermore, it is efficient because an artificial image of only a specified small area is generated instead of the entire chest.

  The difference between the artificial image and the original image can be quantitatively determined by using a difference feature amount corresponding to the difference between the artificial image and the original image such as a least square distance or a normalized cross-correlation value. It becomes possible.

  A plurality of unique images having unique features that appear in each small area when the structure of each classified small area is normal are stored in advance, and the similarity between each small area and each of the plurality of unique images If an artificial image is generated by weighting according to the above, an image of a normal structure corresponding to a specific small region can be generated artificially.

  If the unique image is an image obtained by principal component analysis based on the normal image of each region, it is possible to generate artificial images corresponding to various images by combining a small number of images.

  If the medical image is a chest image obtained by photographing the chest, the small area is classified according to the overlap of the position in the lung field and the rib from the recognition of the lung field and the recognition of the rib. It is possible to improve the determination accuracy when determining abnormal shadows such as lung cancer.

  The image determination apparatus of the present invention will be described with reference to the drawings. As shown in FIG. 1, the image determination apparatus 1 of the present invention includes a classification unit 10 that classifies a large number of small areas obtained by subdividing a part imaged in a medical image P into small areas having similar anatomical features. A candidate area detecting means 30 for detecting a candidate area for an abnormal shadow from a medical image, an artificial image generating means 20 for generating a normal artificial image for a structure of a small area including the detected candidate area for the abnormal shadow, and detection. A small area including the candidate area is set as the attention area, and further, an area setting means 40 for setting the small area existing in the vicinity of the attention area as the vicinity area, and an artificial image of the attention area and the vicinity area are generated, A determination unit 50 is provided that compares the difference between the artificial image of the attention area and the original image and the difference between the artificial image of the neighboring area and the original image to determine whether the candidate area is an abnormal shadow.

  The artificial image generating means 20 stores a plurality of unique images having unique features that appear when the structure of each small region classified according to anatomical features is a normal structure for each classified small region. A unique image storage means 21, and an artificial image of a designated small region among a number of small regions obtained by subdividing a part imaged in the medical image P is a unique image corresponding to the small region into which the small region is classified. An artificial image is generated by weighted addition of images.

  For example, in the case where the medical image P is a chest image obtained by photographing the chest, the position of the lung field and the overlap of the ribs (specifically, whether the overlap of the ribs is one or two) are anatomically determined. Different features appear. Also, the appearance of abnormal shadows such as lung cancer varies depending on whether they are on the ribs or other than the ribs. Therefore, in the present embodiment, a case where an artificial image of a small region in which an abnormal shadow such as lung cancer appears in a chest image will be specifically described below.

  The classification means 10 includes a lung field recognition means 11 for recognizing a lung field and a rib recognition means 12 for recognizing a rib, and according to anatomical features such as the position of a cardiothoracic cage and the degree of rib overlap. Classify into small areas with similar scientific features.

  The candidate area detection means 30 detects an abnormal shadow candidate area such as lung cancer, and the artificial image generation means 20 obtains an artificial image of a normal structure where the abnormal shadow candidate area is detected by the candidate area detection means 30. Generate.

  Further, the determination unit 50 includes a difference feature amount calculation unit 51 that calculates a difference feature amount according to the difference between the population image and the original image, and the artificial image generation unit 20 determines that the attention area and the neighboring area are normal structures. Whether or not the candidate area included in the attention area is an abnormal shadow based on the difference feature quantity between the artificial image of the attention area and the original image and the difference feature quantity of the artificial image and the original image of the neighboring area. Judgment is made.

  Hereinafter, the process flow of the image determination apparatus 1 of the present embodiment will be described based on the flowchart of FIG.

  First, when the subject's chest image P is input to the image determination apparatus 1 (step 1), the classification means 10 detects the cardiothoracic region from the chest image P by the lung field recognition means 11 to recognize the lung field region. (Step 2) Further, the rib recognition unit 12 performs rib recognition from the lung field region (Step 3). Next, the candidate area detection means 30 performs an enhancement process that emphasizes abnormal shadows such as nodules and tumors of the cancerous part to detect candidate areas of abnormal shadows such as lung cancer (step 4). Therefore, the region setting means 40 sets a small region including the candidate region as the attention region, and further sets eight neighboring small regions adjacent to the attention region as the adjacent region (step 5). The classifying means 10 classifies the attention area and the adjacent area into a predetermined image type according to the position of the attention area and the adjacent area in the lung field area and the degree of overlap of the ribs at the position of each area (step 6). ). In order to generate an artificial image corresponding to the classified small area, the artificial image generation means 20 calculates the similarity between the plurality of unique images stored in the unique image storage means 21 and the candidate area, and this similarity Each unique image is added according to the degree to generate an artificial image of each region (step 7). Further, the determination unit 50 compares the difference between the artificial image of the attention area and the original image of the medical image and the difference between the artificial image of the neighboring area and the original image of the medical image, and determines whether there is a difference between the two differences. Then, it is determined whether the candidate area included in the attention area is an abnormal shadow (step 8).

  Next, the above steps will be described in detail below.

(1) Lung field recognition (step 2)
The lung field region captured in the chest image is recognized by detecting the cardiothoracic cage. First, an approximate contour of the cardiothoracic cage is extracted from the chest image using an edge detection mask such as a Gabor function, a position that is the approximate center of the extracted cardiothoracic cage is obtained, and polar coordinates are converted based on the center position. On the polar coordinate plane, the contour of the cardiothoracic is automatically detected by performing a template matching process using a template that is substantially similar to the standard contour of the cardiothoracic. The region surrounded by the detected contour of the cardiothoracic region is recognized as a lung field region, and the lung apex (portions 2 and 7) and the edge as shown in FIG. 3 with reference to the obtained lung field region (Parts 3 and 8), mediastinal part (parts 4 and 9), and lower part of diaphragm (parts 5 and 10) are extracted (for example, Japanese Patent Application Laid-Open No. 2003-6661 filed by the present applicant). reference).

  Alternatively, using the method proposed in US Pat. No. 6,549,646, the apex of the lungs (portions 2 and 7), the edge (portions 3 and 8), the mediastinum (portions 4 and 9), It may be divided into each region of the lower part of the diaphragm (5 and 10 parts) and the edge part may be extracted.

(2) Rib recognition (step 3)
For rib recognition, for example, a statistical model of a structure having a normal structure that does not include an abnormal shadow is created in advance using a sample image of the chest imaged as teacher data, and the chest image P input from the model is created. A rib shape corresponding to is artificially generated.

  First, N images are selected as sample images in which the ribs are clearly captured from a number of chest images, the chest images are displayed, and a point on the ribs of each image is displayed using a pointing device such as a mouse. A model is created in advance by designating n points (for example, 400 points) of landmarks and using them as teacher data.

First, for each of the N sample images, the rib shape X = (x 1, y 1,..., Xi, y i,..., X n, yn) with n landmarks designated and N rib images The average shape Xave = (x _ave 1, y _ave 1,..., X _ave i, y _ave i,..., X _ave n, y _ave n) is obtained from the average shape (FIG. 4 ( See a), ○ indicates the anterior rib, and Δ indicates the posterior rib). A difference vector ΔXj = Xj−Xave (j = 1,..., N) between the rib shape X and the average shape Xave of N sample images is obtained, and N difference vectors ΔXj (j = 1,. .., N) to perform a principal component analysis to obtain eigenvectors (hereinafter referred to as principal component shapes) Ps (s = 1,..., M) from the first principal component to the m-th principal component.

When the principal component analysis is performed, as shown in FIGS. 4B and 4C, the first principal component shape P1 appears as a component that expands the rib in the direction of the arrow in FIG. 4B, and the second principal component shape P2 Appears as a component that expands the ribs in the direction of the arrow in FIG. The model of the arbitrary shape of the rib can be approximated by the linear sum of the average shape Xave and each principal component shape Ps (s = 1,..., M) as follows, and by changing the shape factor bs, Various rib shapes can be generated by warping from the average shape.

  Therefore, in order to artificially generate a rib shape that matches the input chest image P, the shape factor bs is obtained. Several points on the ribs imaged from the chest image P are acquired from the chest image P, and the coordinates of the points on the ribs are substituted into the above equation to find the solution of the shape factor bs (specifically, Can be obtained as a solution of simultaneous equations by substituting the same number of points as the shape factor bs, for example). Even for a chest image P in which the shape of the rib is not clearly photographed, the entire shape of the rib can be generated by substituting the obtained shape factor bs into the above equation. Specifically, since it is easy to extract the posterior rib etc. by edge detection in the chest image, it is possible to extract the point on the edge of the posterior rib and obtain the shape factor bs (the applicant of the present invention). (See Japanese Patent Application Laid-Open No. 2004-41694, etc.).

  Alternatively, as another method, a method may be used in which the edge of the rib is extracted from the chest image, and the extracted shape on the rib is interpolated by a B-spline or the like.

Specifically, as shown in FIG. 5, a plurality of points P1, P2, P3,... On the rib edge curve detected from the chest image are extracted, and a B-spline curve for interpolating these points is extracted. P (t) is obtained (see FIGS. 5 and 6). The nth-order B-spline curve P (t) is defined by the control point Qi (i = 1, 2,..., n) and the parameter t, and is expressed by the following equation.

The cubic B-spline curve P (t) when n = 3 can also be expressed as follows.

Therefore, when t = 0 in the formula (2), the following formula is derived.

Control points are given as shown in FIG. 6, the second control point Q ₂ and the third control point Q ₃ of the present on tangents t _1, t ₂ of the start point and the end point of the curve representing the rib edge . Therefore, a control point Q _i (i = 1, 2, 3,...) Is obtained so as to satisfy this relationship, the position of the point P _i on the radius curve and the relationship of the above equation (3) ( For details, refer to “MEDICAL IMAGING TECHNOLOGY Vol.20 No.6 November 2002, Page 694-Page 701”), and the point on the extracted edge may be interpolated with a B-spline curve to obtain the rib shape.

(3) Detection of lung cancer candidate regions (step 4)
First, the candidate area detection means 30 creates an enhancement-processed image 110 by performing enhancement processing that emphasizes abnormal shadow candidates such as nodules and tumors that appear on the chest image P. Abnormal shadows such as nodules and tumors of cancerous parts taken on the chest image P have roughly rounded outlines, and on the image, the pixel values are larger than the surrounding areas and the gradient vectors are concentrated As observed. Such an abnormal shadow is observed as a circular convex region having a hemispherical shape in which pixels having the same density spread concentrically, and the circular convex region has a distribution of pixel values (density values) from the peripheral part toward the central part. Therefore, the gradient of the pixel value that the pixel value is high (the density value is low) is recognized. The gradient lines are concentrated toward the center direction of the abnormal shadow, and the gradient of the pixel value can be calculated as a gradient vector, and the abnormal shadow can be detected from the degree of concentration of the gradient vector.

  Specifically, the concentration degree of the gradient vector is obtained as follows.

First, as for the gradient vector, the orientation φ of the gradient vector is obtained based on the calculation formula shown in the following formula (4) for all pixels constituting the image to be calculated.

  Here, as shown in FIG. 7, f11 to f55 are pixel values corresponding to pixels on the outer periphery of a mask of 5 vertical pixels × 5 horizontal pixels centering on the pixel j.

Therefore, the gradient degree C of the gradient vector is calculated according to the equation (5) for all the pixels P constituting the target image.

  Here, N is the number of pixels existing in a circle with a radius l around the pixel of interest, θj is a straight line connecting the pixel i and each pixel j in the circle, and the above equation (4) for each pixel j The angle formed with the gradient vector calculated in () (see FIG. 8).

  The reason why the degree of concentration C expressed by the above equation (2) is large is a large value in pixels where the gradient vectors of the respective pixels j are concentrated. The gradient vector of each pixel j in the vicinity of the shadow of the nodule or tumor is directed to the approximate center of the tumor shadow regardless of the contrast of the shadow, so that the pixel having the large concentration C is an abnormal shadow. This is the pixel at the center of.

  In addition, among those that evaluate the degree of concentration, there is one that devised the size and shape of the filter in order to achieve detection power that is not affected by the size or shape of abnormal shadows. An example is a ring filter (Adaptive ring filter).

  Abnormal shadows are not only those in which the pixel value gradually increases from the center to the periphery, but the change in the pixel value in the central part is not monotonous, such as nodules appearing in lung cancer, and the vector field is disturbed and the degree of concentration decreases. There is also. The adaptive ring filter can be applied not only to the case where the pixel value changes monotonously but also to the case where the pixel value in the central portion is not monotonous and the vector field is disturbed to reduce the degree of concentration.

  The mask area of the adaptive ring filter is a ring-shaped area shown in FIG. 9, and the inner radius is determined adaptively. When the radius of the inner circle and the outer circle is r and R, respectively, the relationship is expressed as in Equation (6).

R = r + d (6)
However, d is a constant.

The output of the adaptive ring filter is

Thus, the output of the adaptive ring filter takes a local maximum value near the center of the circular convex region. For example, a circular convex region as shown in FIG. 10 (A) appearing on the original image has a pixel value as shown in FIG. 10 (B) on a white line, and when an adaptive ring filter is applied, FIG. 10 (C). As shown, a pixel value higher than that of the original image appears in the central portion. (For more information, see, for example, Sakaki, Yoshihiro Sugawara, Hidefumi Obata, “Gradient Vector Concentration Filter for Extracting Cancer Shadow Candidates”, IEICE Transactions (D-II) Vol.J83-D-II No. 1, pp. 118-125, Jan. 2000.).

  Therefore, this adaptive ring filter processing is performed on the chest image P to create an enhanced processed image 110 in which nodules and tumors are enhanced. FIG. 11 shows an example in which a nodule is emphasized using an adaptive ring filter with l = 20 mm and d = 4 mm (hereinafter, nodule detection will be described). When the nodule (white arrow) portion appearing on the original image in FIG. 11A is subjected to the adaptive ring filter processing, it is emphasized as shown in FIG.

  However, at the edge of the cardiothoracic region, the ribs and the like are photographed in an overlapping manner, so that the concentration gradient density is disturbed, and even if the enhancement process is performed using the chest image P of the original image, the circular convex region is not emphasized well. Therefore, it is better to remove the influence of the background image and perform the enhancement process at the edge.

  Specifically, a difference image 120 is created by subtracting the background image from the original image for the edge portions (portions 3 and 8) obtained using the above-described rib cage extraction method. By performing enhancement processing on the image, it is possible to remove the influence of the background image and enhance the nodule. Specifically, for example, the background image component can be removed by subtracting a smoothed image obtained by blurring the original image with a Gaussian filter from the original image.

  FIG. 12 shows a state when the enhanced image 110 is created using an adaptive ring filter. FIG. 12A shows a state where the enhancement processing image 110 is created by performing the enhancement processing using the original image, but the nodule is not emphasized well due to the influence of the background image. On the other hand, FIG. 12B shows an enhanced image 110 obtained by performing an enhancement process using an adaptive ring filter on the difference image 120 obtained by subtracting the smoothed image blurred by the Gaussian filter from the original image 100 at the edge. The appearance when created is shown, but the nodule is emphasized without being affected by the background image.

  Next, a binarization process is performed on the enhancement-processed image 110 by gradually changing the threshold value from a low value to a high value to create a plurality of binarized images 130. In the binarization processing, a pixel having a pixel value equal to or greater than a given threshold value is replaced with a first pixel value (for example, 255 (white) when the pixel value is represented by 8 bits), and the pixel value of the pixel equal to or smaller than the threshold value Is replaced with a second pixel value (for example, 0 (black) when the pixel value is expressed in 8 bits) to create a binary image 130. When binarization processing is performed, regions having high pixel values such as structures, nodules, and tumors on the binarized image 130 are replaced with the first pixel values, and the other regions are replaced with the second pixel values. The region where the pixels having the first pixel value are connected appears as an island-shaped isolated region. When the given threshold value is low, an isolated region appearing on the binarized image 130 is extracted including a white cloud-like portion appearing in the background image, but as the threshold value increases, nodules that do not include the background image are extracted. Only parts such as tumors and structures are extracted as isolated regions. In particular, a nodule or a tumor enhanced using an adaptive ring filter has a higher pixel value than other structures, and appears as an isolated region on the binarized image 130 binarized with a high threshold. From this, it can be said that the isolated region appearing on the binarized image 130 binarized with a high threshold has a high possibility of an abnormal shadow.

  FIG. 13 shows an example in which binarization processing is performed by changing the threshold value. FIG. 13A shows an enhancement processing image 110 in which an adaptive ring filter process is performed on the original image of the chest image P to emphasize a circular convex region. This enhanced image 110 is quantized with 8 bits and has a gradation of 0-255. When binarization processing is performed on the enhancement processing image 110 using 100 as a threshold value, a binarized image 130 as shown in FIG. ) Appears. Further, 39 levels of binarized images 130 are created by changing the threshold value in increments of 4, and an isolated region is extracted from each binarized image 130. FIGS. 13C and 13D are examples of a binarized image 130 binarized with threshold values 176 and 252.

  A nodule has a feature that the above-mentioned isolated region appears in a shape close to a circle and has a small area. On the other hand, an isolated region extracted by including a background image or an isolated region from which a structure is extracted often has a low circularity and a large area. Therefore, an isolated region extracted from each of a plurality of binarized images 130 obtained by binarization processing is extracted as a nodule candidate having a circularity of a predetermined value or less and a size of a predetermined value or less. .

  The circularity c is obtained, for example, from the extracted area S of the isolated region and its circumference L as follows.

The radius R of a perfect circle with area S is

The circumferential length L1 of a perfect circle having an area S is

It becomes. Therefore, the circularity c is determined to be 1 when it is a perfect circle and to approach 0 as it deviates from the perfect circle. Therefore, from the peripheral length L of the isolated region and the area S of the isolated region,

Asking.

Or you may obtain | require circularity as follows instead of calculating | requiring the circularity calculated | required by said Formula (10). As shown in FIG. 21, when an approximate circle Cir having a radius R is drawn around the center of gravity of the isolated region, the following equation (11) is obtained from the area S of the isolated region and the area A of the overlapping portion of the isolated region and the approximate circle. ), The circularity c may be obtained.

  Specifically, when determining the shadow of the nodule in the chest image, the circularity in the expression (11) is 0.7 or more and 2.26 mm < It is appropriate to extract an isolated region having a radius <4.94 mm as a nodule candidate region.

  In addition, as shown in FIG. 10, the nodule emphasized by the adaptive ring filter has a higher pixel value in the central part than the nodule appearing on the original image, so that it may appear as an isolated region at the same position even if the threshold value is changed. Many. Therefore, when the binarized image 130 is created by changing the threshold value step by step, a nodule candidate area that is frequently picked up at the same position is detected as a nodule candidate region.

  Although the detection of the nodule has been described above, it is possible to detect not only the nodule but also the mass by changing the size of the adaptive ring filter and the determination method based on the binarized image.

(4) Region setting (step 5)
Accordingly, as shown in FIG. 19, the area setting means 40 sets a small area so as to include the abnormal shadow candidate area detected by the candidate area detection means 30 described above, and uses this small area as the attention area. A small region in the vicinity of the region is set as the neighborhood region. Since nodules and the like of abnormal shadows are usually 30 mm or less, the small area is a rectangle having a side of about 40 mm, and the small area is set to include the detected candidate area as the attention area. Eight neighboring regions R1, R2, R3, R4, R5, R6, R7, and R8 (hereinafter referred to as adjacent ROIs) adjacent to the attention region M (hereinafter referred to as the central ROI) including the candidate region are set as neighboring regions. To do.

(5) Classification (step 6)
Next, a method for classifying lung field regions based on anatomical features will be described. First, the classification means 10 uses the thoracic apex (portions 2 and 7), the edge (portions 3 and 8), the vertical portion shown in FIG. The position of the partition (portions 4 and 9) or the lower part of the diaphragm (portions 5 and 10) is divided.

  First, as shown in FIG. 14, the coordinate values in the X direction are normalized so that the X coordinate of the outer position of the left lung (or right lung) cardiothoracic cage is 0 and the X coordinate of the inner position is 1. The coordinate value in the Y direction is normalized so that the Y coordinate of the upper position of the cardiothoracic cage is 0 and the Y coordinate of the lower position is 1.

For example, when the coordinates of the center position of the lung cancer candidate region detected by the candidate region detection means 30 are expressed with the upper left of the chest image P as the origin, and the coordinate value is (x, y), the lung cancer If the Y coordinate of the two points where the straight line in the Y direction passing through the center of the candidate area intersects the cardiothoracic cage is yup (top point) and ydown (bottom point), the normalized Y coordinate The value newy is
newy = (y −yup) / (ydown −yup)
It becomes. Similarly, if the Y coordinate of the two points where the straight line in the X direction passing through the center of the lung cancer candidate region intersects the cardiothoracic cage is xup (top point) and xdown (bottom point), the normal Y coordinate value newx when
newx = (xout −x) / (xout −xin)
It becomes. However, x = 1.0 is always set in the mediastinal portion (portions 4 and 9) of the cardiothoracic cage.

Based on the normalized coordinate values, the data is divided into 11 as shown in FIG.
(1) 0 ≦ newx <0.1, 0 ≦ newy <0.25
(2) 0.1 ≦ newx <0.8, 0 ≦ newy <0.25
(3) 0.8 ≦ newx <1.0, 0 ≦ newy <0.25
(4) 0 ≦ newx <0.25, 0.25 ≦ newy <0.75
(5) 0.25 ≦ newx <0.6, 0.25 ≦ newy <0.75
(6) 0.6 ≦ newx <1.0, 0.25 ≦ newy <0.75
(7) 0 ≦ newx <0.25, 0.75 ≦ newy <1.0
(8) 0.25 ≦ newx <0.6, 0.75 ≦ newy <1.0
(9) 0.6 ≦ newx <1.0, 0.75 ≦ newy <1.0
(10) newx = 1.0
(11) newy = 1.0

Next, from the position of the rib detected by rib recognition, the location where the rib is located or the number of overlapping ribs is calculated, and the rib features that represent this belong to the region (1) to (9) Small areas including candidate areas for lung cancer are further classified into three. Since the small area includes an image in which one rib overlaps in one part and includes a part in which there is no rib in some, the feature amount of the rib is the overlap of the rib included in the small area. The average is used to classify each image type.
(a) Rib feature = 0; No rib
(b) 0 <radius feature ≦ 1.0; one rib overlap
(c) 1.0 <rib feature amount; overlap of two ribs As described above, the lung field region is classified into 29 (= 9 × 3 + 2) image-type small regions.

(6) Artificial image generation (step 7)
Next, a method for generating an artificial image of each small area classified by the artificial image generating means 20 will be described. First, a method of generating a plurality of unique images having unique features that appear when the classified structures of each small region have a normal structure will be described. The unique image is generated for each image type of each small area classified by the classification means 10 described above. For example, the case where the feature amount of the rib is (b) (one rib overlap) (hereinafter referred to as category 4-1) in the small region belonging to the above (4) will be described as an example. As shown in FIG. 16, a rectangular small area of 80 × 80 pixels belonging to category 4-1 is extracted from a normal chest image in which no abnormal shadow has been imaged, and M images are obtained as sample images (FIG. 16 (a)). ) Select to use as teacher data. In FIG. 16, a portion where one rib overlaps is mainly extracted, but the principal component analysis is performed including the surrounding area.

First, an image vector Q = (q _1,1 ,..., Q _1,80 , q _2,1 ,, which includes 6400 pixel values included in M sample images (80 pixels × 80 pixels) as a component. _{···, q 2,80, ···, q} i, j, ···, q 80,1, ···, q 80,80) performed with and (see FIG. 17) principal component analysis, eigenvectors ( Eigenimage). FIG. 16 shows an example in which the principal images are analyzed for the sample images to obtain the top nine principal component vectors to obtain the unique images (80 pixels × 80 pixels) Q1, Q2,. The obtained unique images Q1, Q2,..., Q9 are linearly independent, and an artificial image can be generated by a linear sum of the unique images Q1, Q2,.

For example, when the candidate region for lung cancer detected by the candidate region detection means 30 belongs to category 4-b, an image R of a small region of 80 pixels × 80 pixels including the candidate region for lung cancer is used as the chest image P. To obtain correlation values a1, a2,..., A9 between the image R and the respective unique images Q1, Q2,. From this, the artificial image R-model corresponding to the image R of the small area is
R-model = a1 x Q1 + a2 x Q2 + ... + a9 x Q9
Is generated.

  Here, the case where an artificial image is generated by performing principal component analysis on a sample image has been described. However, like the method described in rib recognition, first, an average image of the sample image is obtained, and the average image and each sample image are obtained. The difference image may be subjected to principal component analysis, and the average image may be warped to generate an artificial image.

  FIG. 18A (category 4-1) shows an example in which an artificial image having a normal structure corresponding to an image in which cancer is present in category 4-1 is generated. If the artificial image is subtracted from the original image in which the cancer exists in category 4-1, it can be seen that the cancer appears black.

  On the other hand, FIG. 18B (category 4-1) shows an example in which an artificial image corresponding to an image of category 4-1 where no cancer exists is generated. Since there is no cancer in category 4-1, when the artificial image is subtracted from the original image, the entire image shows a uniform density.

(8) Determination method (step 8)
However, even when an artificial image that is as close as possible to the original image is created by the above-described method, it is difficult to reproduce the edges of the ribs on the artificial image. Therefore, as shown in FIG. If an attempt is made to emphasize a region where a cancer such as a nodule appears, an edge portion appears as an artifact in the difference image, and this may be determined as a cancer such as a nodule. In addition, when a characteristic texture partially appears on the background image depending on the photographed subject, the texture may be determined as a cancer such as a nodule.

  However, an abnormal shadow such as a nodule does not appear over a wide range, and a small shadow appears only in a place where a nodule is present. However, a similar shadow appears over a wide range in the texture having a feature of an artifact or a background image. That is, whether or not the candidate area is an abnormal shadow can be determined based on whether a similar shadow appears not only in the small area where the candidate area exists but also in the surrounding area.

  Therefore, the determination unit 50 determines whether the candidate region in the center ROI is an abnormal shadow such as a nodule by comparing the center ROI set by the region setting unit 40 and the shadow appearing in the adjacent ROI. .

  First, each of the central ROI and the adjacent ROI is classified into a corresponding image type by the classifying means 10, and the central ROI and the adjacent ROI are used by using the above-described artificial image generating means 20 according to the image type into which each ROI is classified. An artificial image corresponding to is generated.

  When the candidate shadow in the central ROI is a nodule, there is little possibility that a similar nodule appears in the neighboring ROIs in the vicinity of the surrounding eight. On the other hand, when the candidate shadow is an artifact, there is a high possibility that similar artifacts appear in the surrounding ROIs in the vicinity. That is, when the candidate shadow is a nodule, when the difference between the original image of the central ROI and its population image is taken, the nodule shadow as shown in FIG. 18 (a) appears on the difference image. The difference image obtained by subtracting the difference between the original image of the neighboring ROIs in the vicinity of 8 and the population image is a flat image. On the other hand, when the candidate shadow is an artifact, similar shadows appear continuously in the central ROI and the neighboring ROIs in the vicinity of eight.

  Therefore, if the candidate region of the central ROI is an abnormal shadow, the difference between the central ROI original image and its artificial image and the difference between the original image of the neighboring neighboring ROI and the artificial image are compared. Is considerably larger than the adjacent ROI.

  Therefore, the determination unit 50 uses the difference feature amount calculation unit 51 to calculate the difference feature amount between the generated artificial image of each ROI and the original image, and whether or not the candidate region of the central ROI is an abnormal shadow. Judgment is made. The difference feature amount only needs to be an index for determining whether the image on the difference image is due to an artifact or the like. Specifically, the least square distance, the normalized cross-correlation, or the like can be used. .

  For example, when comparing the minimum square distance of the original image of the central ROI and the artificial image, and the minimum square distance of the original image of the adjacent ROI and the artificial image, if there is an abnormal shadow due to cancer such as a nodule in the central ROI, As shown in FIG. 20, the least square distance of the center ROI is relatively larger than the least square distance of the adjacent ROI (the normalized cross-correlation between the original image and the artificial image is small). On the other hand, when there is no abnormal shadow due to cancer such as a nodule in the center ROI, the least square distance of the center ROI is approximately the same as other least square distances.

As a specific judgment method,
・ Difference between the difference feature amount of the neighboring ROIs in the vicinity of 8 and the difference feature amount of the central ROI ・ Comparison of the difference feature amount of the neighboring ROIs in the vicinity of 8 and the difference feature amount of the central ROI ・ 8 The sum total of the difference between the difference feature quantity of the neighboring ROI in the vicinity and the difference feature quantity of the center ROI can be used.

  Or, using the least square distance or normalized cross-correlation as one of the feature quantities, and combining with other feature quantities, discriminators such as SVM (support vector machine), Mahalanobis distance, and neural network It is also possible for a person to determine whether or not it is an abnormal shadow according to Japanese Patent Application Laid-Open Nos. 9-167238 and 2002-74325.

  In the above description, the neighboring area in the vicinity of the attention area has been described as the adjacent ROI set so as not to overlap the central ROI. However, the neighboring area may be set so as to partially overlap the attention area.

  As described above, the chest image has been described. The breast image is also divided into a pectoral muscle region, a mammary gland region, a fat region, and the like (see, for example, Japanese Patent Application Laid-Open No. 2003-334183 of the present applicant) to generate an artificial image. Alternatively, abnormal shadows may be determined.

  In the above description, the case of detecting a nodule of 30 mm or less has been described, but a shadow of 30 mm or more such as a tumor may be detected by appropriately changing the size of the small region.

  Further, an image determination apparatus can be obtained by installing a program for executing the above-described processing on a computer on a computer such as a personal computer via a CD-ROM or a network.

  As described above in detail, a medical image including a structure having anatomical features is subdivided, each small region is classified and a unique image serving as a base is generated for each classified small region, and the unique image is generated. Since an artificial image is generated from the image, an image in which features corresponding to the small area appear can be generated, and the accuracy of the artificial image can be improved. Further, since the partial image is generated, it is possible to generate the image more efficiently than the generation of the entire image.

  Further, when a candidate area for an abnormal shadow is detected, it is determined whether the candidate area detected from the difference between the population image and the original image in the small area including the detected candidate area and the neighboring area in the vicinity is an abnormal shadow. By doing so, it is possible to reduce erroneous detection due to artifacts and the like.

The figure which shows schematic structure of the image determination apparatus of this invention The flowchart for demonstrating the flow of a process of the image determination apparatus of this invention. An example of lung field segmentation results Diagram for explaining the method of rib recognition by principal component analysis The figure for demonstrating the method of extracting the shape of a rib using a B-spline curve Diagram for explaining control points of B-spline curve Illustration for explaining the gradient Diagram for explaining the method for explaining concentration calculation The figure for demonstrating the method of calculating calculation of an adaptive ring filter The figure for demonstrating the pixel value output with the adaptive ring filter Example of output image of adaptive ring filter The figure for demonstrating the output result of the adaptive ring filter using a difference image The figure showing the mode of change of the binarized image when changing the threshold Diagram for explaining how to normalize lung field Diagram for explaining classification based on normalized position Diagram showing an example of principal component analysis The figure for demonstrating the component of the pixel contained in a sample image The figure which shows the example which generated the artificial image An example of a region of interest and a neighboring region nearby An example of the least square distance between an artificial image and an original image Diagram for explaining how to determine circularity

Explanation of symbols

1 Image determination device
10 Classification means
11 Lung field recognition means
12 rib recognition means
20 Artificial image generation means
21 Unique image storage means
30 Candidate area detection means
40 Area setting method
50 Judgment means
51 Difference feature quantity calculation means P Medical image

Claims (8)

In a medical image including a structure having an anatomical feature obtained by imaging a predetermined part of a subject, an anatomical feature obtained from a plurality of small regions obtained by subdividing the medical image is obtained from the structure. Classification means for classifying each similar small area,
Artificial image generation means for generating an artificial image of a normal small area of the structure according to a classification result of the classification means for at least one of the classified small areas;
Candidate area detecting means for detecting a candidate area of an abnormal shadow from the medical image;
A region setting unit that sets a small region including at least a part of the candidate region detected by the candidate region detection unit as a region of interest, and further sets a small region existing in the vicinity of the region of interest as a neighborhood region;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. An image determination apparatus comprising: determination means for determining whether or not The determination means includes a difference feature amount according to a difference between the artificial image of the attention area and the original image of the attention area, and a difference feature according to a difference between the population image of the vicinity area and the original image of the vicinity area. A difference feature amount calculating means for calculating the amount,
The difference feature quantity of the attention area calculated by the difference feature quantity calculation means is compared with the difference feature quantity of the neighboring area, and the candidate area included in the attention area is an abnormal shadow depending on whether there is a difference between the two feature quantities. The image determination apparatus according to claim 1, wherein it is determined whether or not there is an image.   The image determination apparatus according to claim 2, wherein the difference feature amount is a least square distance or a normalized cross-correlation value between the artificial image and the original image. The artificial image generating means is
A unique image storage means for storing a plurality of unique images having unique features that appear in each small region when the structure of each classified small region is normal;
Weighting is performed according to the degree of similarity between the small area for generating the artificial image and each of the plurality of unique images, and the artificial image of the small area is generated by adding the plurality of unique images. The image determination apparatus according to any one of claims 1 to 3.   The unique image stored in the unique image storage means is an image obtained by principal component analysis of the structure of the medical image obtained by photographing the subject based on the normal image of each small region. The image determination apparatus according to claim 4. The medical image is a chest image obtained by photographing the chest of the subject,
The classification means is
Lung field recognition means for recognizing a lung field region from the chest image;
Rib recognition means for recognizing the rib from the chest image,
6. The image determination apparatus according to claim 1, wherein the plurality of small regions are classified according to an overlap between a position in the lung field region and a rib. In a medical image including a structure having an anatomical feature obtained by imaging a predetermined part of a subject, an anatomical feature obtained from a plurality of small regions obtained by subdividing the medical image is obtained from the structure. Classification means for classifying each similar small area,
In an image determination apparatus comprising: an artificial image generation unit that generates an artificial image of a normal small region of the structure according to a classification result of the classification unit for at least one small region of the classified small regions ,
A candidate area detecting step of detecting a candidate area of an abnormal shadow from the medical image;
An area setting step for setting a small area including at least a part of the candidate area detected by the candidate area detection step as an attention area, and further setting a small area in the vicinity of the attention area as a vicinity area;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. And a determination step for determining whether or not the image is determined. Computer
In a medical image including a structure having an anatomical feature obtained by imaging a predetermined part of a subject, an anatomical feature obtained from a plurality of small regions obtained by subdividing the medical image is obtained from the structure. Classification means for classifying each similar small area,
Artificial image generation means for generating an artificial image of a normal small area of the structure according to a classification result of the classification means for at least one of the classified small areas;
Candidate area detecting means for detecting a candidate area of an abnormal shadow from the medical image;
A region setting unit that sets a small region including at least a part of the candidate region detected by the candidate region detection unit as a region of interest, and further sets a small region existing in the vicinity of the region of interest as a neighborhood region;
The attention unit and the neighboring region are classified by the classification unit, and the artificial image generation unit generates a normal artificial image for each of the classified regions, and the artificial image of the attention region and the attention region And the difference between the artificial image of the neighboring area and the original image of the neighboring area is compared, and the candidate area included in the attention area is abnormally shaded depending on the difference between the two differences. A program that functions as a determination unit that determines whether or not