JP2005278690A - Method for detecting nodular shadow from three-dimensional chest ct image using computer, apparatus therefor and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method for detecting nodular shadow from three-dimensional chest CT images using a computer, for highly accurately detecting the nodular shadow in a simple process. <P>SOLUTION: In the method for detecting the nodular shadow from the three-dimensional chest CT images using the computer, three-dimensional volume data are prepared 11 from chest CT slice image data, a lung field region is selected 13 from the three-dimensional volume data, a background trend components are removed 12 by a morphological filtering processing in chest CT slice images, and false-positive shadows are removed from the images from which the background trend components are removed. For the removal of the false-positive shadows, there are blood vessel extraction 17 by a nonlinear line emphasis filtering processing, the removal 18 of false positive utilizing edge direction images using a Sobel filter and the removal 19 of the false positive by feature quantities analysis, etc. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a method and apparatus for detecting a nodular shadow from a three-dimensional chest CT image using a computer, and a computer program.

  In recent years, CT images have been widely used for chest image diagnosis, and the transition from simple X-ray photography to diagnosis by CT has progressed (for example, see Patent Document 1). In addition, CT apparatuses equipped with a helical method or a multi-detector have become widespread, and the number of slice images per examination has increased from several tens to several hundreds, increasing the burden on the interpreting doctor. Thus, in order to support the detection of nodular shadows by a doctor using a computer, a technique for automatically detecting nodular shadows from a three-dimensional chest CT image has been developed.

  Currently, as a technique for automatically detecting a nodular shadow from a chest CT image, for example, the following method is known. First, a nodule-like shadow that is a candidate for cancer is used as a spherical model, and blood vessels are connected to a cylindrical model, and multiple templates with different sizes are created, and the shadows included in those template images and CT images There is a method of detecting nodular shadows by distinguishing between nodular shadows and blood vessel shadows by matching with (see Non-Patent Document 1). In addition, there is a method of calculating a cross-correlation value between various sizes of three-dimensional Gaussian distributions and nodular shadows, selecting an optimal template by a genetic model, and detecting a nodular shadow (see Non-Patent Document 2). There is a method (see Non-Patent Document 3) that uses a method of detecting a nodular shadow on the chest wall called a rolling ball method after threshold processing of the lung field region. In addition to these, a technique of selectively enhancing blood vessel shadows, nodular shadows, and airway shadows by using a Hessian matrix (see Non-Patent Document 4) is known.

However, these conventional techniques need to use a special method for detecting the nodular shadow of the chest wall portion, and the detection of the nodular shadow may be affected by the change in the CT value of the background. There was a problem of complexity of processing steps and accuracy.
JP 2001-78996 A Kanae Shigemoto, Hotaka Takizawa, Koji Yamamoto, Toru Nakagawa, Toru Matsumoto, Norio Kanno, Takeshi Iinuma, Mitsumi Matsumoto, High-speed recognition ", MEDICAL IMAGING TECHNOLOGY, 2003, Vol. 21, No. 2, p. 147-155 Li Xun, Hara Takeshi, Fujita Hiroshi, "Evaluation of GA template matching method by simulation using chest helical CT image", Journal of Medical Image Information and Communication Society, 2000, Vol. 17, No. 3, p. 118-129 Yongbum Lee, Takeshi Hara, Hiroshi Fujita, Shigeki Itoh, and Takeo Ishigaki, "Automated Detection of Pulmonary Nodules in Helical CT Images Based on an Improved Template-Matching Technique", IEEE TRANSACTIONS ON MEDICAL IMAGING, 2001, 20 (7), p .595-604 Qiang Li, Shusuke Sone, Kunio Doi, "Selective enhancement filters for nodules, vessels, and air walls in two- and three-dimensional CT scans", Medical Physics, 2003, 30 (8), p.2040-2051

  The present invention has been made in view of such circumstances, and is a method for detecting a nodular shadow from a three-dimensional chest CT image using a computer, and a method for detecting a nodular shadow with high accuracy by a simple process. The purpose is to provide

To achieve the above object, the method of the present invention is a method for detecting a nodular shadow from a three-dimensional chest CT image using a computer,
Creating three-dimensional volume data from chest CT slice image data;
Selecting a lung field region from the three-dimensional volume data;
In the chest CT slice image, removing a background trend component by morphological filter processing;
Removing false positive shadows from the background trend component removed image.

The apparatus of the present invention is an apparatus for detecting a nodular shadow from a three-dimensional chest CT image,
Means for creating three-dimensional volume data from chest CT slice image data;
Means for selecting a lung field region from the three-dimensional volume data;
In the chest CT slice image, means for removing the background trend component by morphological filtering,
And a means for removing false positive shadows from the background trend component removed image.

  The program of the present invention is a program capable of executing the method of the present invention on a computer.

  When the background trend component is removed by the morphological filter processing, the nodular shadow and the blood vessel can be clarified to a considerable degree as compared with the conventional technique. If the false positive shadow is removed from this processed image, high accuracy can be obtained. Can detect nodular shadows. Moreover, the morphological filter process is a simple processing method. Therefore, according to the present invention, the nodular shadow can be detected with high accuracy by a simple process.

  In the method of the present invention, the step of removing the background trend component by the morphological filter processing includes performing an open process using a circular kernel on the original image of the chest CT slice, and obtaining the obtained open process image from the original image. It is preferable that it is a process of deducting.

  In the method of the present invention, it is preferable that the step of removing the false positive shadow includes a step of removing the false positive shadow by feature amount analysis. The feature amount analysis includes shadow volume, shadow sphericity, average value of shadow voxel value, standard deviation of shadow voxel value, coefficient of variation of shadow voxel value, second central moment, skewness, kurtosis, An analysis that calculates at least one feature quantity selected from the group consisting of the maximum value of the shadow pixel value, the minimum value of the shadow pixel value, and the contrast of the shadow, and determines whether the shadow is a false positive. Is preferred.

  In the method of the present invention, it is preferable that the false positive shadow removing step includes a false positive removing step based on an edge direction image by a Soble filter process. In the false positive removal process by the edge direction image by the Sobel filter process, the Sobel filter process is performed on the chest CT slice image to create an edge direction image, and the two-dimensional Gauss that is a simulated nodular shadow on the edge direction image A two-dimensional cross-correlation value is calculated using an edge direction image obtained by performing Sobel filter processing on the distribution image as a template, the obtained two-dimensional cross-correlation value is represented as an image, this image is subjected to threshold processing, and this threshold processing image and Preferably, this is a step of examining the overlap with the initial candidate for the nodular shadow and removing the false positive shadow caused by the blood vessel. It is preferable that the false positive removal step by the edge direction image by the Sobel filter processing is performed on the cross-sectional chest CT slice image and the sagittal chest CT slice image.

  In the method of the present invention, the false positive shadow removal step includes a step of removing a false positive shadow caused by a blood vessel by performing nonlinear line enhancement filter processing on an image subjected to MIP (maximum value projection method) processing. Is preferred. The MIP processing is preferably processing for creating partial volume data using three chest CT slice images and performing MIP processing on the partial volume data.

  The method of the present invention preferably further includes a step of performing nodule shadow enhancement processing using a three-dimensional zero-enclosed Gaussian template and performing threshold processing on the enhanced image to select a nodule shadow candidate.

  In the apparatus of the present invention, the means for removing the background trend component by the morphological filtering process performs an open process using a circular kernel on the original image of the chest CT slice, and the obtained open process image is obtained from the original image. A subtracting means is preferred.

  In the apparatus according to the present invention, it is preferable that the means for removing the false positive shadow includes a means for removing the false positive shadow by feature amount analysis. The feature amount analysis includes shadow volume, shadow sphericity, average value of shadow voxel value, standard deviation of shadow voxel value, coefficient of variation of shadow voxel value, second central moment, skewness, kurtosis, An analysis that calculates at least one feature quantity selected from the group consisting of the maximum value of the shadow pixel value, the minimum value of the shadow pixel value, and the contrast of the shadow, and determines whether the shadow is a false positive. Is preferred.

  In the apparatus of the present invention, it is preferable that the false positive shadow removing unit includes a false positive removing unit based on an edge direction image by a Soble filter process. The false positive removal means based on the edge direction image by the Sobel filter processing performs Sobel filter processing on the chest CT slice image to create an edge direction image, and a two-dimensional Gauss that is a simulated nodular shadow on the edge direction image. A two-dimensional cross-correlation value is calculated using an edge direction image obtained by performing Sobel filter processing on the distribution image as a template, the obtained two-dimensional cross-correlation value is represented as an image, this image is subjected to threshold processing, Preferably, this means is a means for examining the overlap with the initial candidate for the nodular shadow and removing the false positive shadow caused by the blood vessel. The false positive removal by the edge direction image by the Sobel filter processing is preferably performed on the horizontal chest CT slice image and the sagittal chest CT slice image.

  In the apparatus according to the present invention, it is preferable that the false positive shadow removing means includes means for removing a false positive shadow caused by a blood vessel by performing nonlinear line enhancement filter processing on an image subjected to MIP processing. The MIP processing is preferably processing for creating partial volume data using three chest CT slice images and performing MIP processing on the partial volume data.

  The apparatus of the present invention preferably further includes means for performing nodule shadow enhancement processing using a three-dimensional zero-enclosed Gaussian template and performing threshold processing on the enhanced image to select nodule shadow candidates.

  The program of the present invention may be stored in a computer-readable recording medium.

  Next, an example of the method of the present invention will be described with reference to the drawings. Note that the images, flowcharts, and functional block diagrams shown below are examples in the implementation of the present invention, and the present invention is not limited thereby.

  A flowchart of this example technique is shown in FIG. As shown in the figure, in this example, first, a step 11 for reading a chest CT image, and non-uniform CT value fluctuation (globally changing background trend component) in the lung field region are removed by a method based on morphological filter processing. Step 12, lung region extraction step 13, nodular shadow template creation step 14 using three-dimensional zero-surround Gaussian, nodular shadow enhancement step 15 by three-dimensional cross-correlation method, nodular shadow by threshold processing Candidate pick-up step 16, blood vessel extraction step 17 by nonlinear line enhancement filter processing, false-positive removal step 18 using edge direction image using Sobel filter processing, false-positive removal step 19 by feature amount analysis, Step 20 of determining a nodular shadow candidate is included. Each of these steps will be specifically described below.

(1) Reading of a chest CT slice image A CT image obtained by a chest CT examination is composed of a plurality of transverse slice images as shown in FIG. In this example, a nodule-like shadow exists at the position of the arrow of the slice image displayed enlarged on the right side of the figure. Such an original slice image of 512 × 512 pixels is preferably smoothed by a moving average process of 4 × 4 pixels, and three-dimensional volume data is preferably created using an image reduced to 256 × 256 pixels. The reason for the reduction to 256 × 256 pixels is that the matrix size is necessary and sufficient to detect the nodule shadow of 4.9 mm, which is the lower limit of the nodule shadow size in the image database, and the unnecessary high frequency components are attenuated. This is because the time required for processing is also shortened.

(2) Background trend component removal by morphological filter processing Next, background trend component removal by morphological filter processing (open processing) is performed to make the average density different depending on the position in the lung field. By this processing, the background CT value can be made uniform, and only blood vessels and nodular shadows in the lung field can be extracted. A flowchart of this process is shown in FIG. As illustrated, in this process, first, an open process 232 is performed on the chest CT slice original image 231. The open process 232 includes an image contraction process 232a by a minimum value filter process using a circular kernel and a background trend image generation process 232b by an image extension by a maximum value filter process using a circular kernel. Then, if the process 233 for subtracting the open process image from the original image is performed, the background trend component can be removed. Specifically, for example, a minimum value filtering process using a circular kernel having a diameter of 15 pixels is performed on the chest CT slice original image. By this process, blood vessels and nodular shadows smaller than the circular kernel used disappear, and shadows larger than the filter size shrink slightly. Next, in order to restore the size of the shrunken large shadow to the original size of the image subjected to the minimum value filter processing, the expansion processing is performed using a circular maximum value filter having the same kernel size. At this time, the shadow disappeared by the minimum value filter processing remains disappeared even when the maximum value filter processing is performed. Performing expansion processing on an image that has been subjected to contraction processing in this way is called open processing. FIG. 3B shows an image obtained by performing an open process on the original image in FIG. The vascular shadow and nodular shadow in the lung field almost disappeared, and the background background component, the heart, and the soft tissue outside the lung field remained. Finally, the background trend component in the lung field can be removed by subtracting the open processing image from the chest CT slice image that is the original image. FIG. 3C shows an image with the background component removed. It can be seen that the background in the lung field is uniform and blood vessels and nodular shadows are well depicted. By performing this process, it is possible to easily detect the nodular shadow without performing a special process particularly for the nodular shadow in contact with the chest wall.

  FIGS. 4A and 4B show an original CT slice image having a matrix size of 512 × 512 and an image obtained by removing background trend components, respectively. A nodular shadow is present in the arrow part, but an image from which the trend component has been removed is observed as an isolated shadow without touching the chest wall. The results of simple threshold processing using these images are shown in FIGS. 5 (a) and 5 (b), respectively. In FIG. 5 (a), the nodular shadow of the lower lung field is detected in contact with the blood vessel and the chest wall due to the influence of the background component of the background, whereas in FIG. 5 (b), the nodular shadow and It can be seen that blood vessels are clearly extracted.

  Since only the shadow in the lung field can be extracted by this method, it can be used as a pre-process for performing MIP (maximum value projection method) that is frequently used when displaying a three-dimensional image. Three-dimensional display of blood vessel images is easier than performing MIP processing using original CT images. Further, this is a technique that can be applied not only to the CT image targeted this time but also to all images such as MRI and ultrasonic images.

(3) Extraction of lung field region A flowchart of lung field region extraction is shown in FIG. As shown in the figure, the lung region extraction process is performed by binarization using a threshold value determined by a step 222 of creating a CT value histogram from the three-dimensional chest CT image 221 created in (1) and a histogram discriminant analysis. A process 223 for performing processing, a process 224 for three-dimensional labeling, a process 225 for selecting a chest region using the volume of the label image, and a process 226 for filling the label image by the closing process and smoothing the edge. . Specifically, for example, using the three-dimensional CT volume data (matrix size: 256 × 256 × number of slice images) created in (1), a histogram of CT values as shown in FIG. 6 is obtained. Then, a threshold value is determined by a discriminant analysis method using the histogram, and the three-dimensional CT image is binarized. Further, after labeling the obtained three-dimensional binary image, only the lung field region is extracted by obtaining the volume of each shadow. FIG. 7A shows the extracted lung field region. As shown in this figure, a defect occurs in a relatively thick blood vessel region or the like simply by performing threshold processing. For this purpose, a closing process is performed on the three-dimensional binary image, and the defect portion is filled to obtain a final lung field region. The final lung field region thus obtained is shown in FIG. FIG. 8 is a diagram showing the extracted lung field region as a two-dimensional montage image. It can be seen that the lung field region is extracted almost accurately.

(4) Nodular Shadow Enhancement FIG. 24 shows a flowchart of a nodular shadow candidate enhancement process. As shown in the figure, this process includes a step 242 for creating a simulated nodular shadow template by a three-dimensional Gaussian function for the background trend component removal image 241 and a nodular shadow enhancement step 243 by a three-dimensional cross-correlation method. .

Creating a simulated nodular shadow template using a 3D Gaussian distribution For emphasis processing of nodular shadow candidates, it is necessary to create a simulated nodular shadow template using a 3D Gaussian function (distribution) as follows. is there.

  That is, first, since the nodular shadow to be detected is often spherical, it is detected using a spherical three-dimensional Gaussian distribution template. In addition, a three-dimensional zero-enclosed Gaussian template, including the surrounding zero region, is used as a simulated nodular shadow to more effectively enhance isolated nodular shadows. FIG. 9 shows a three-dimensional Gaussian template (upper left of the figure), and also shows cross sections of a transverse section (upper right of the figure), a coronal surface (lower left of the figure), and a sagittal plane (lower right of the figure). In the figure, the dotted line indicates the range of the zero region of the template.

  In a three-dimensional CT image, in many cases, the resolution in the cross section (xy plane) is about 0.5 mm to 0.6 mm, whereas the resolution in the body axis direction (z direction) is 5 mm to 10 mm. Taken as a degree. Therefore, when creating a spherical template, the length of each side of the voxel is taken into consideration so that the distribution is spherical. The formula of the three-dimensional Gaussian distribution is shown in the following formula (I).

上記式（Ｉ）において、Ｇ（ｘ，ｙ，ｚ）は座標（ｘ，ｙ，ｚ）における３次元ガウス分布、

In the above formula (I), G (x, y, z) is a three-dimensional Gaussian distribution at coordinates (x, y, z),

は３次元ガウス分布の中心座標、σ_ｘ、σ_ｙ、σ_ｚは各辺における標準偏差を、それぞれ表わす。

Is the center coordinate of the three-dimensional Gaussian distribution, and σ _x , σ _y , and σ _z are the standard deviations on each side.

(5) Enhancement of nodular shadow candidate by three-dimensional cross-correlation method Next, a cross-correlation value between a 256 × 256 × slice image with the background trend component removed and a three-dimensional zero-enclosed Gaussian template image is used as a template image. By shifting the position of, the pattern that closely resembles the nodular shadow is selectively emphasized. This calculation is performed only for the lung field region extracted by the lung field region extraction process (3), for example. The formula for obtaining the cross-correlation value is shown in the following formula (II).

上記式（ＩＩ）において、Ｃは相互相関値、Ｘ、Ｙ、Ｚは、それぞれ、ｘ、ｙ、ｚ方向のマトリクスサイズ、σ_ｆ、σ_ｇは、それぞれ、画像とテンプレートのボクセル値の標準偏差、

In the above formula (II), C is the cross-correlation value, X, Y, and Z are the matrix sizes in the x, y, and z directions, respectively, and σ _f and σ _g are the standard deviations of the voxel values of the image and the template, respectively. ,

は画像の平均ボクセル値、

Is the average voxel value of the image,

はテンプレート画像の平均ボクセル値を表わす。

Represents the average voxel value of the template image.

  The cross-correlation value is close to 1.0 when there is a shadow having exactly the same shape as the template, and close to 0.0 when the shape is greatly different from the template. Further, when the shape is the same and black and white are reversed, it is −1.0. As a result, a template image that has a shape that closely resembles a simulated nodular shadow has a correlation value close to 1.0, and a template image that does not resemble a template image, such as a blood vessel, is obtained. , A correlation value close to 0.0 is obtained. FIG. 10A shows the calculation result of the cross-correlation value calculated using a three-dimensional Gaussian template having a diameter of 6 mm. The nodular shadow at the arrow position is well emphasized, but other normal structures such as blood vessels are also emphasized. FIG. 10B shows a calculation result of a cross-correlation value calculated using a template in which a 2.5 mm area around a 3 mm Gaussian template having a diameter of 6 mm is wrapped with zeros. Compared with FIG. 10A, it can be seen that the enhancement of blood vessels is suppressed and the nodular shadow is more selectively enhanced. FIG. 10C shows a calculation result of a cross-correlation value calculated using a template in which a 5.0 mm area around a 3 mm Gaussian template having a diameter of 6 mm is wrapped with zeros. The nodular shadow is still emphasized, and the enhancement of blood vessels is further suppressed than in (b).

(6) Picking up nodular shadow candidates by threshold processing FIG. 25 shows a flowchart of threshold processing. As shown in the figure, this process includes a step 252 for creating a binary image by threshold processing from a nodule shadow enhanced image 251 and a step 253 for determining a nodule shadow initial candidate. Specifically, threshold processing is performed on the nodular shadow enhanced image obtained in (5) above, and the nodular candidate shadow is detected as a binary image. Here, the threshold is set to 0.27, but this is determined empirically so that the nodule candidate shadow detected by binarization and the nodule shadow observed in the original image are approximately the same size. Value. Therefore, the threshold value is not limited to this, and other values may be used as the threshold value.

  FIG. 11A shows an image of cross-correlation values obtained with the three-dimensional zero-enclosed Gaussian template, and FIG. 11B shows nodule candidate shadows obtained as a result of the threshold processing. As shown in the figure, a small nodular shadow is detected at the position of the arrow. In addition, when there is a possibility that many false positive shadows may be included, it is preferable to remove false positives in the subsequent processing.

(7) Elimination of false positives using edge direction pattern obtained by Sobel filter Since the false positive shadows are often attributed to blood vessels, the processing here distinguishes the blood vessels from the linear pattern from the nodular shadows. This is a process for removing the result. This processing method is a technique originally developed by the present inventors. The Sobel filter is a filter that outputs edge strength and its direction. Here, the edge direction in a pattern extending linearly in a certain direction, such as a blood vessel, is constant, and in the case of a round pattern, such as a nodular shadow, the edge direction is at the center of the shadow. Pay attention to the change from 0 to 359 degrees toward the target, and remove false positives caused by blood vessels.

  FIG. 26 shows a flowchart of false positive removal processing using the edge direction pattern obtained by the Sobel filter. As shown in the figure, this process includes a step 261 of creating an edge direction image by Sobel filter processing, a step 262 of creating an edge direction template image of a simulated nodular shadow, and a two-dimensional view of the edge direction template image and the edge direction image. A step 263 for calculating a cross-correlation value and a step 264 for removing a false positive shadow having a direction pattern different from the edge direction template image.

  In this process, a two-dimensional slice CT image is used to simplify the process. FIGS. 12A and 12B show a two-dimensional Gaussian distribution that is a simulated nodular shadow and an edge direction image obtained by processing the distribution with a Sobel filter. In this figure, angles from 0 degrees to 359 degrees are expressed by shades of black to white. When the two-dimensional cross-correlation value is calculated by using the direction image of FIG. 12B as a template with respect to the edge direction image of FIG. 12C obtained by performing the Sobel filter processing on the original image, the edge direction of the simulated nodular shadow A high correlation value is shown at an edge direction distribution similar to the distribution diagram 12 (b), and a low correlation value is shown at a different edge direction pattern such as a blood vessel. In the enlarged view of FIG. 12C, a nodular shadow is present in the circled portion, and it can be seen that the pattern in that direction is very similar to the edge direction pattern of the template. FIG. 13A shows a cross-correlation value obtained by this method as an image. Further, FIG. 13B shows a result of performing threshold processing on this image. As described above, in the nodular shadow pattern, a high cross-correlation value is detected as a small dot-shaped shadow, and in the case of a blood vessel, it is not detected at all or is detected as a long pattern. Then, by examining the overlap between the binary image obtained by this method and the initial candidate for the nodular shadow, the false positive shadow caused by the blood vessel can be reduced. Further, the edge direction pattern by the Sobel filter may be similarly obtained for the two-dimensional cross-sectional image in the sagittal direction, and false positive shadows may be further reduced.

  Since this method uses the direction of the shadow edge, it does not depend on the contrast of the shadow. Therefore, even in the case of early light lung cancer shadows, if the CT value at the center is high and the area around it is low, it is a method that can be detected with high sensitivity.

  FIG. 14 shows the effect of removing false positives using the edge direction pattern obtained by the Sobel filter. FIG. 14A shows an initial nodule shadow candidate obtained in the process of emphasizing the nodule shadow candidate by the three-dimensional cross correlation method, and FIG. 14B shows the edge direction obtained by the Sobel filter. It shows after removal of false positives using a pattern. From this figure, it can be seen that a number of false positive shadows can be removed by the process of step 264 of removing false positive shadows having a direction pattern different from the edge direction template image.

(8) Extraction of blood vessels using nonlinear line enhancement filter To further reduce false positives caused by blood vessels, two-dimensional nonlinear line enhancement filters (Kanae Shigemoto, Hotaka Takizawa, Junji Yamamoto, Toru Nakagawa, Toru Matsumoto, Norio Kanno, Takeshi Iinuma and Mitsomi Matsumoto, “High-speed recognition of nodal shadows from chest X-ray CT images using 3D nodule / blood vessel model and template matching”, MEDICAL IMAGING TECHNOLOGY, 2003, Vol. 21, No. 2, p. 147-155) was used to extract blood vessel shadows.

  FIG. 27 shows a flowchart of blood vessel extraction processing using a nonlinear line enhancement filter. As shown in the figure, this process includes a process 271 for extracting shadows by MIP (maximum value projection) process for removing background trend components, a process 272 for binarizing the MIP-processed image, and a nonlinear line enhancement filter process. Step 273 of emphasizing the blood vessel shadow (line pattern) by the above, and step 274 of removing the false positive shadow caused by the blood vessel.

  A filter as shown in FIG. 15 was used for this processing. This filter is composed of a set of three linear templates and 12 groups of templates having different angles. In the drawing, for simplification, only two sets of angles (vertical direction and template rotated 45 degrees to the right) are shown. In each template, the sum of the pixel values overlapping the middle template (symbol in the figure: A), the sum of the left template (symbol in the figure: B), and the sum of the right template (symbol in the figure) : C), and the sum of AB and AC is divided by the length of the template. [Output value = (2A-B-C) / Template length (when A> B and A> C)] However, when the value of A is lower than B or C, the output value is 0 [ Output value = 0 (except when A> B and A> C)]. Finally, the maximum value of the output values from the templates in all 12 directions is set as the output value of the filter. This makes it possible to emphasize line patterns in any direction.

  This nonlinear line enhancement filter is used for the binary image shown in FIG. 16A obtained by performing threshold processing on an image from which the background trend component has been removed. The reason is that a light line pattern can be easily detected as a line. The line pattern thus obtained is shown in FIG. It can be seen that a relatively thick blood vessel is detected. The blood vessel image detected by this processing is used to remove the nodular shadow and false positives caused by the blood vessel.

  In order to detect blood vessels more effectively, MIP (maximum intensity projection) processing is also used. In this process, the background background removed image created in (3) is used. It is preferable to create partial volume data using upper and lower slices in a relatively narrow range for all slice images including the lung field region as shown in FIG. This time, a set of three slice images is used, but the range can be determined according to the slice thickness. The range is, for example, a slice image of a range of 3 to 6 sheets. Next, MIP processing is performed using the partial volume data. The MIP processed image is shown in FIG. Of the upper three images shown in this figure, FIG. 17B is a central image for performing MIP processing, FIG. 17A is an upper slice image, and FIG. 17C is a lower slice image. It is. FIG. 17D shows an image obtained by performing MIP processing using the three slice images shown in FIGS. 17A to 17C. In this image, it can be seen that more blood vessels are included in one MIP image than the individual slice images due to the effect of the MIP processing. However, the image obtained by the MIP process may include not only a blood vessel shadow but also a nodular shadow as shown by a white arrow in FIG. Therefore, by performing binarization processing on the MIP image shown in FIG. 17D and performing nonlinear line enhancement filter processing, it becomes possible to selectively enhance only the linear blood vessels. By applying this method to a slice image including all lung fields, it is possible to selectively enhance the blood vessel shadows in all the lung fields. The MIP image obtained by this method is subjected to nonlinear line enhancement filter processing on FIG. 18A, and the detected blood vessel image is shown in FIG. As shown in the figure, it can be seen that more blood vessel shadows are selectively emphasized as compared with FIG.

(9) Elimination of false positives by feature value analysis Many false positive shadows are removed by the processing so far, but in order to reduce false positives to a practical level, feature value analysis further reduces false positive shadows. Is preferably removed. FIG. 28 shows a flowchart of the feature amount analysis process. As shown in the figure, this process includes a step 281 for obtaining a nodule candidate shadow after false positive shadow removal using the edge direction pattern obtained by the Sobel filter process, a step 282 for calculating a feature amount of the nodule candidate shadow, And step 233 for removing false positive shadows by threshold processing of the feature amount. In the feature amount analysis, it is preferable to appropriately select the type of image to be calculated according to the type of feature amount. Table 1 below shows the types of feature values.

(Table 1) Features used for analysis 1. Shadow volume 2. The sphericity of the shadow 3. Average value of voxels for each shadow 4. Voxel standard deviation for each shadow 6. Coefficient of variation for each shadow 6. Second-order central moment Skewness (third-order central moment)
8). Kurtosis (4th center moment)
9. 9. Maximum pixel value for each shadow 10. Minimum pixel value for each shadow 11. Contrast between maximum value of nodular shadow and average value of second pixel from surrounding area. 14. Contrast between average value of nodular shadow and average value of second pixel from surrounding area. Ratio of Lines Contained for Each Shadow In this example, the feature quantities 1 and 2 are calculated for individual shadows by labeling the binary image shown in FIG. The feature amounts 3 to 12 were calculated for the original image, background background removed image, line enhanced image, three-dimensional cross-correlation image, and MIP image, and false positives were removed. Only the feature amount 13 was calculated for the line-enhanced image. Details of the individual feature amounts are shown below.

1. The volume is measured every cubic millimeter (mm ³ ) for each shade with the label number created in volume (6). In this example, the aim is to detect a nodular shadow of 5 mm to 20 mm, and an extremely small shadow or a large shadow is removed as a false positive.

2. Sphericality Since the nodular shadow has a shape close to a sphere, we determined the sphericity and distinguished it from false positives. In this example, it is defined in the literature (Maryellen L. Giger, Kunio Doi and Heber MacMahon, "Automated detection of nodules in peripheral lung fields", Medical Physics, 1998, 15 (2), p.158-164) What extended circularity to three dimensions was used. For each shadow, a sphere with the same volume as the shadow was overlapped around the center of gravity of the shadow, and the ratio of the shadow and the sphere overlapped was determined. If it is a perfect sphere, the sphericity will be 1.0, and if the shadow is long, its value tends to be small. Therefore, a shadow with a low value was judged as a false positive shadow.

3.4.5. Average value, standard deviation, coefficient of variation Find the average value, standard deviation, and coefficient of variation of the voxel value for each shadow as the feature value. The average value is determined by the average density of the shadow, the standard deviation is calculated by the density variation of the shadow, and the coefficient of variation is obtained by dividing the standard deviation by the average value, which is useful for evaluating the magnitude of the substantial data variation.

6). Secondary central moment The secondary central moment indicates the dispersion of the distribution. By obtaining the variance, the distribution of each shadow can be found. A formula representing the distribution of the distribution is shown in Formula (III).

上記式（ＩＩＩ）において、σ^２は２次中心モーメント、

In the above formula (III), σ ² is the secondary central moment,

は、それぞれ、重心のｘ、ｙ、ｚ座標、Ｘ、Ｙ、Ｚは、テンプレート画像の縦・横方向のボクセル数、画像枚数、ｐ（ｘ，ｙ，ｚ）はボクセル値を表す。

Are the x, y, z coordinates of the center of gravity, X, Y, Z are the number of voxels in the vertical and horizontal directions of the template image, the number of images, and p (x, y, z) is the voxel value.

7). Skewness (third-order central moment)
The skewness represents the distortion of the distribution. Thereby, the object of distribution can be known. When the skewness is> 0, the distribution is distorted to the right. When the skewness is <0, the distribution is distorted to the left. When the skewness is 0, the distribution is the left and right. Indicates. The calculation is shown in Formula (IV).

8). Kurtosis (4th center moment)
The kurtosis is a value representing the flatness of the distribution and is used for comparison with the shape of the Gaussian distribution. Since the fourth-order moment is 3 when the distribution is a Gaussian distribution, the kurtosis K from the Gaussian distribution can be obtained by subtracting 3 from the fourth-order moment. Since the simulated nodular shadow is created using a Gaussian distribution, the kurtosis = 0 for a shape that is very similar to the nodular shadow. A flat shape such as a blood vessel has a kurtosis <0. This is shown in Formula (V).

上記式（Ｖ）において、Ｋはガウス分布からの尖度を表す。

In the above formula (V), K represents the kurtosis from the Gaussian distribution.

9.10. Maximum value, minimum value The maximum value and minimum value of the pixel value for each shadow are obtained as the feature amount.

11.12. As shown in FIG. 19, the average pixel value around 2 pixels away from the outer periphery of the shadow is obtained as a feature amount, and the contrast is obtained by subtracting from the maximum pixel value of the shadow. In addition, the contrast obtained by subtracting the average pixel value of the shadow from the average pixel value around the pixel 2 pixels away from the outer periphery of the shadow is also calculated.

13. Ratio of lines (lines) included in each shadow The ratio of lines included in each shadow is calculated by dividing the number of voxels emphasized as a line by the volume of the shadow in the line-enhanced image.

(10) Determination of Nodular Shadow Candidate The final nodular shadow candidate is determined from the initial nodular shadow detected in (6) and determined as a nodular shadow by the analysis of (7) to (9). To do. Then, a threshold is set for determining the nodular shadow or false positive for all of the above-mentioned features, and when all the features have nodular shadow features, the shadow is the final nodule. Judge as a shadow. Therefore, if any one of them is classified as a false positive, the shadow is removed as a false positive shadow.

  FIG. 20 shows a scatter diagram of the volume of all shadows and the standard deviation of voxel values. The small dots in the figure represent false positive shadows, and the large dots represent nodular shadows. From this graph, it can be seen that a shadow with an extremely large volume or a small shadow, and a shadow with a large standard deviation contain many false positive shadows. From this figure, an appropriate threshold for removing false positive shadows on volume and standard deviation can be determined. In this way, threshold values are set for all feature amounts.

  As examples of threshold setting, examples of thresholds 3 to 12 shown in (Table 1) in the original image are shown below.

3. Average value of voxels for each shadow 13.0 ≧ average value average value ≧ 114.0
Average value ≧ 0.12333 × volume + 60.0
4). Standard deviation of voxels for each shadow 1.2 ≧ standard deviation
Standard deviation ≧ 28.5
Standard deviation ≧ 0.0733 × volume + 3.0
5). Coefficient of variation for each shadow 0.49 ≥ Coefficient of variation
Coefficient of variation ≥ 7.0
Coefficient of variation ≧ 0.004333 × volume + 3.0
6. Secondary central moment −0.38 ≧ Secondary central moment Secondary central moment ≧ 1.12
Secondary center moment ≧ 0.0047 × volume + 0.12
7). Skewness (third-order central moment)
-1.0 ≧ distortion skewness ≧ 1.88
Skewness ≧ 0.0061 × Volume-4.7
8). Kurtosis (4th center moment)
14.0 ≧ kurtosis
Kurtosis ≧ 154.0
Kurtosis ≧ 0.3 × volume + 40.0
9. Maximum pixel value for each shadow Maximum value ≧ 78.0
Maximum value ≧ 0.16833 × volume + 14.0
10. Minimum pixel value for each shadow 0.102 ≧ minimum value minimum value ≧ 0.46
Minimum value ≧ 0.00125 × volume + 0.15
11. Contrast between maximum value of nodular shadow and average value of second pixel from surrounding area 8.0 ≧ contrast contrast ≧ 144.0
Contrast ≧ 0.46666 × volume 12. Contrast between average value of nodular shadow and average value of second pixel from surrounding area 6.8 ≧ contrast contrast ≧ 103.0
Contrast ≧ 0.2216666 × volume + 30.0

  The image shown in FIG. 21 is an example in which a nodular shadow is emphasized by a three-dimensional cross-correlation method using a 5 mm wide template wrapped in zero. FIG. 21A is a diagram illustrating an example of enhancement of nodular shadow candidates in the original image, and FIG. 21B is a diagram illustrating an example of enhancement of nodular shadow candidates in the background component removed image. It is. When the nodular shadow is emphasized using the original image as it is, an arcuate false positive as shown in FIG. 21A appears near the chest wall due to the partial volume effect, whereas the background When the image from which the trend has been removed is used, it can be seen that there is no false positive around the pleura as shown in FIG. From this, it can be understood that the process of the step 333 of removing the background trend component by the morphological filter process is very effective and important in detecting the nodular shadow.

  Next, FIG. 29 shows an example of functional blocks of an apparatus using a computer for detecting the nodular shadow of this example. As shown in the figure, this apparatus includes a three-dimensional chest CT image processing / control unit 295, an operation input unit 292, a data storage unit 297, a display processing unit 306, and a monitor (display) 307 as main components. To do.

The three-dimensional chest CT image processing / control unit 295 uses a three-dimensional CT image data input unit 296, a lung region selection processing unit 298, a background trend component removal processing unit 299, and a three-dimensional zero-enclosed Gaussian template. Nodular shadow enhancement processing unit 302, edge direction image creation unit 301 using three-dimensional Sobel filter processing, edge direction image false positive removal processing unit 304, MIP processing unit 300, and line enhancement filter processing It has a blood vessel shadow extraction unit 303 and a false positive shadow removal processing unit 305 by feature amount analysis. Note that the arrows in the figure indicate the flow of data or the flow of processing. The three-dimensional CT image data input unit 296 is connected to the CT image reading device 291 and the data storage unit 297. A CT slice image reading process is executed. That is, here, the image data from the CT image reading device 291 is received, and this image data is transmitted to the lung field region selection processing unit 298 and the false positive shadow removal processing unit 305 by feature amount analysis, and necessary. In response, the data is transmitted to the data storage unit 297, where the data is stored. The lung field region selection processing unit 298 executes the above-described (3) lung field region extraction processing. The background trend component removal processing unit executes background trend component removal by the above-described (2) morphological filter processing. The nodular shadow enhancement processing unit 302 using the three-dimensional zero-enclosed Gaussian template includes the above-mentioned (4) nodular shadow enhancement processing, and (5) nodular shadow candidate enhancement processing by the three-dimensional cross-correlation method, (6) A process for picking up nodular shadow candidates by threshold processing is executed. The edge direction image creating unit 301 using the three-dimensional Sobel filter processing and the false positive removal processing unit 304 using the edge direction image are false positives using the edge direction pattern obtained by the above-described (7) Sobel filter. Perform removal. The MIP processing unit 300 and the blood vessel shadow extraction unit 303 by the line enhancement filter process execute blood vessel extraction by the above-described (8) nonlinear line enhancement filter. The false positive shadow removal processing unit 305 by feature amount analysis executes (9) false positive removal by feature amount analysis and (10) determination of nodular shadow candidates. Then, the determined nodule-like shadow candidate is displayed on the monitor 307 by the display processing unit 306. These controls are performed by operations using a keyboard 293 and a mouse 294 connected to the operation input unit 292.

  Next, a configuration example of a computer system of this apparatus is shown in a block diagram of FIG. As shown in the figure, this computer system includes a CPU 500 for executing processing, a memory 501 for storing programs and data, a memory 501 or a hard disk 502 for storing programs and data used by the CPU 500, and a monitor (display for displaying data or images). 503, a keyboard 504 for inputting data or commands, an external recording medium receiving unit 505 such as a CD-ROM, and a communication processing unit 506 for communicating with a CT apparatus or other computer system via a network . In order to execute the detection method of the present invention on a computer system, the program of the present invention is installed on the hard disk 502 by meeting the CD-ROM drive 505 or the communication processing unit 506, and the program is started by the CPU 500. Executed.

  Next, an example in which the detection of nodular shadow according to this example is applied to an actual clinical example will be shown. The clinical cases used were 40 cases, and these cases were analyzed by the method of the present invention described above. The three-dimensional chest CT image including the nodular shadow used was taken with a HI MEDICAL SYSTEM Hi Speed CT / i and HISPEED-RP CT apparatus. The photographing conditions are: scan mode: HELICAL MODE, tube voltage: 140 kv and 120 kv (only one of 40 examples), reconstruction function: STANDARD, slice image matrix size: 512 × 512, pixel size: 0.585 mm to 0.703 mm, slice thickness: 5 mm to 10 mm, and reconstruction interval: 5 mm to 14 mm. The nodular shadows in the image database are all cases in which a definitive diagnosis has been made, and the diameter is 4.9 mm to 13.2 mm. Since clinical image data is used, the CT apparatus and imaging conditions vary depending on the case. As a result of this analysis, nodular shadows could be detected with a probability of 92.5%, and the average number of false positives at that time was 1.45 / slice.

  As described above, the present invention is a method for detecting a nodular shadow from a three-dimensional chest CT image using a computer, which can detect the nodular shadow with high accuracy by a simple process. Therefore, by using the result detected by the present invention when the doctor performs image diagnosis, it is possible to reduce the time required for diagnosis. In addition, an effect of reducing oversight of a light shadow that is difficult to detect can be expected.

It is a flowchart figure which shows an example of the method of this invention. It is the figure which showed an example of the CT image obtained by a chest CT examination. It is the figure which showed an example of the background trend component removal by morphological processing. (A) It is the figure which showed an example of the original image of CT obtained by chest CT examination. (B) It is the figure which showed an example of the background background component obtained by the open filter process of the image of Fig.3 (a). (C) It is an example of the image which removed the background component of FIG.3 (b) from the image of Fig.3 (a). (A) It is the figure which showed an example of the original image of CT obtained by chest CT examination. (B) It is an example of the image which removed the background component from the image of Fig.4 (a). (A) It is the figure which showed the image after the threshold value process of the image of Fig.4 (a). (B) It is the figure which showed the image after the threshold value process of the image of FIG.4 (b). It is the figure which showed an example of the histogram of a three-dimensional CT image. (A) It is the figure which showed the image which extracted only the lung field area | region by calculating | requiring the volume of each shadow, after labeling a three-dimensional binary image. (B) It is the figure which showed the image of the final lung field area | region obtained by performing the closing process with respect to a three-dimensional binary image, and filling a defect | deletion part. It is the figure which showed the extracted lung field area | region as a two-dimensional montage image. It is the figure which showed the example of the three-dimensional zero siege Gaussian template (simulated nodular shadow). (A) A cross section of a three-dimensional Gaussian template is shown. (B) The cross section of a cross section is shown. (C) A cross section of the coronal surface is shown. (D) shows a cross section of the sagittal plane. (A) It is the figure which showed the example of the image of the three-dimensional cross-correlation value using the three-dimensional Gaussian template (no surrounding zero area | region) of diameter 6mm. (B) It is the figure which showed the example of the image of the three-dimensional cross-correlation value using the three-dimensional Gaussian template (diameter of surrounding zero area | region = 2.5mm) of diameter 6mm. (C) It is the figure which showed the example of the image of the three-dimensional cross-correlation value using the three-dimensional Gaussian template (diameter of surrounding zero area | region = 5.0 mm) of diameter 6mm. (A) It is the figure which showed the example of the image of the cross correlation value obtained by the three-dimensional zero surrounding Gaussian template. (B) It is the figure which showed the example of the image after carrying out the threshold value process of the image of Fig.11 (a). (A) It is a figure which shows the example of two-dimensional Gaussian distribution. (B) It is a figure which shows the example of the image of an edge direction obtained by processing a 2-dimensional Gaussian distribution with a Sobel filter. (C) It is a figure which shows the example of the edge direction image obtained by processing an original image with a Sobel filter. (D) It is the elements on larger scale of FIG.12 (c). (A) It is a figure which shows the example of the image of the two-dimensional cross correlation value using a Sobel image. (B) It is a figure which shows the image which threshold-processed the image of Fig.13 (a) by cross-correlation value 0.75 or more. (A) It is the figure which showed the example of the nodule initial candidate shadow. (B) It is the figure which showed the example which reduced the false positive using the edge direction pattern obtained by the Sobel filter. It is the figure which showed the example of the nonlinear line emphasis filter. (A) It is the figure which showed the example of the binary image obtained by carrying out the threshold value process of the image from which the background trend component was removed. (B) It is the figure which showed the example of the nonlinear line emphasis filter process image. (A) It is the figure which showed the example of the upper slice image. (B) It is the figure which showed the example of the slice image before performing MIP processing. (C) It is the figure which showed the example of the lower slice image. (D) It is the figure which showed the example of the image by which the MIP process was carried out. (A) It is the figure which showed the example of the image by which the MIP process was carried out. (B) It is the figure which showed the example of the image by which the nonlinear line emphasis filter process was carried out. It is the figure which showed the example of the method of calculating | requiring the contrast with the maximum value of a nodular shadow, and the average value of the 2nd pixel from the circumference | surroundings. It is the figure which showed the example of the scatter diagram of the standard deviation of the volume of all the shadows, and a voxel value. It is the figure which showed the example of the emphasis of the nodular shadow by the cross-correlation method (width 5mm wrapped in 0). (A) It is the figure which showed the example of emphasis of the nodular shadow candidate in an original image. (B) It is the figure which showed the example of emphasis of the nodular shadow candidate in a background component removal image. It is a flowchart figure which shows an example of the process of extraction of a lung field area | region. It is a flowchart figure which shows an example of the process of the background trend component removal by a morphological filter process. It is a flowchart figure which shows an example of the process of emphasizing a nodular shadow candidate. It is a flowchart figure which shows an example of the process of picking up a nodular shadow candidate. It is a flowchart figure which shows an example of the process of a false positive removal using the edge direction pattern obtained with the Soble filter. It is a flowchart figure which shows an example of the process of the false positive shadow removal by a nonlinear emphasis filter. It is a flowchart figure which shows an example of the process of the false positive removal by feature-value analysis. It is a figure which shows an example of the functional block of the apparatus using the computer for detecting the nodular shadow in one Example of this invention. It is a block diagram which shows the structural example of the computer system of the apparatus using the computer for detecting the nodular shadow in one Example of this invention.

Explanation of symbols

11 CT image reading step 12 Lung field region non-uniform CT value fluctuation (globally changing background trend component) is removed by a method based on morphological filter processing 13 Lung field region extracting step 14 3D zero-surround Gaussian 15 Nodular shadow template creation process using three-dimensional cross-correlation method 16 Nodular shadow candidate picking process using threshold processing 17 Blood vessel extraction process using nonlinear line enhancement filter process 18 Sobel filter process was used Step of removing false positive using edge direction image 19 Step of removing false positive by feature analysis 20 Step of determining nodular shadow candidate

Claims (22)

A method of detecting a nodular shadow from a three-dimensional chest CT image using a computer,
Creating three-dimensional volume data from chest CT slice image data;
Selecting a lung field region from the three-dimensional volume data;
In the chest CT slice image, removing a background trend component by morphological filter processing;
Removing false positive shadows from the background trend component removed image. The step of removing a background trend component by the morphological filter processing is a step of performing an open process using a circular kernel on an original image of a chest CT slice and subtracting the obtained open process image from the original image. The method according to 1. The method according to claim 1, wherein the step of removing the false positive shadow includes a step of removing the false positive shadow by feature amount analysis. The feature amount analysis includes shadow volume, shadow sphericity, average value of shadow voxel value, standard deviation of shadow voxel value, coefficient of variation of shadow voxel value, second central moment, skewness, kurtosis, This is an analysis for calculating at least one feature amount selected from the group consisting of the maximum value of the shadow pixel value, the minimum value of the shadow pixel value, and the contrast of the shadow, and determining whether or not the shadow is a false positive. Item 4. The method according to Item 3. 5. The method according to claim 1, wherein the false positive shadow removal step includes a false positive removal step based on an edge direction image by a Soble filter process. In the false positive removal process by the edge direction image by the Sobel filter process, the Sobel filter process is performed on the chest CT slice image to create an edge direction image, and the two-dimensional Gauss that is a simulated nodular shadow on the edge direction image A two-dimensional cross-correlation value is calculated using an edge direction image obtained by performing Sobel filter processing on the distribution image as a template, the obtained two-dimensional cross-correlation value is represented as an image, this image is subjected to threshold processing, and this threshold processing image and The method according to claim 5, wherein the method is a step of removing a false positive shadow caused by a blood vessel by examining an overlap with an initial candidate for a nodular shadow. The method according to claim 5 or 6, wherein the false positive removal step based on the edge direction image by the Sobel filter processing is performed on a cross-sectional chest CT slice image and a sagittal chest CT slice image. The method according to claim 1, wherein the false positive shadow removing step includes a step of removing a false positive shadow caused by a blood vessel by performing nonlinear line enhancement filter processing on an image subjected to MIP processing. 9. The method according to claim 8, wherein the MIP processing is processing for generating partial volume data using three chest CT slice images and performing MIP processing on the partial volume data. 10. The method according to claim 1, further comprising a step of performing nodule shadow enhancement processing using a three-dimensional zero-enclosed Gaussian template and performing threshold processing on the enhanced image to select a nodule shadow candidate. Method. An apparatus for detecting a nodular shadow from a three-dimensional chest CT image,
Means for creating three-dimensional volume data from chest CT slice image data;
Means for selecting a lung field region from the three-dimensional volume data;
In the chest CT slice image, means for removing the background trend component by morphological filtering,
And means for removing false positive shadows from the background trend component removed image. The means for removing the background trend component by the morphological filter processing is means for performing an open process using a circular kernel on an original image of a chest CT slice and subtracting the obtained open process image from the original image. 11. Apparatus according to 11. The apparatus according to claim 11 or 12, wherein the means for removing the false positive shadow includes a means for removing a false positive shadow by feature amount analysis. The feature amount analysis includes shadow volume, shadow sphericity, average value of shadow voxel value, standard deviation of shadow voxel value, coefficient of variation of shadow voxel value, second central moment, skewness, kurtosis, This is an analysis for calculating at least one feature amount selected from the group consisting of the maximum value of the shadow pixel value, the minimum value of the shadow pixel value, and the contrast of the shadow, and determining whether or not the shadow is a false positive. Item 14. The device according to Item 13. 15. The apparatus according to claim 11, wherein the false positive shadow removing unit includes a false positive removing unit based on an edge direction image by a Soble filter process. The false positive removal means based on the edge direction image by the Sobel filter processing performs Sobel filter processing on the chest CT slice image to create an edge direction image, and a two-dimensional Gauss that is a simulated nodular shadow on the edge direction image. A two-dimensional cross-correlation value is calculated using an edge direction image obtained by performing Sobel filter processing on the distribution image as a template, the obtained two-dimensional cross-correlation value is represented as an image, this image is subjected to threshold processing, and this threshold processing image and 16. The apparatus according to claim 15, which is means for removing false positive shadows caused by blood vessels by examining an overlap with an initial candidate for a nodular shadow. The apparatus according to claim 15 or 16, wherein false positive removal based on an edge direction image by the Sobel filter processing is performed on a cross-sectional chest CT slice image and a sagittal chest CT slice image. The apparatus according to any one of claims 11 to 17, wherein the false positive shadow removing means includes means for removing a false positive shadow caused by a blood vessel by performing nonlinear line emphasis filter processing on an image subjected to MIP processing. The apparatus according to claim 18, wherein the MIP processing is processing for generating partial volume data using three chest CT slice images and performing MIP processing on the partial volume data. 20. The method according to any one of claims 11 to 19, further comprising means for performing nodule shadow enhancement processing using a three-dimensional zero-enclosed Gaussian template and performing threshold processing on the enhanced image to select a nodule shadow candidate. apparatus. A program capable of executing the method according to any one of claims 1 to 10 on a computer. A computer-readable recording medium on which the program according to claim 21 is recorded.

Images