JP4823204B2 - Medical image processing device - Google Patents

Description

The present invention relates to a medical image processing apparatus that performs image analysis and image processing of a head image obtained by photographing a patient's head.

  In recent years, with the spread and high performance of magnetic resonance imaging apparatuses (hereinafter referred to as MRI; Magnetic Resonance Imaging), the number of brain dock examinations has increased rapidly. One of the purposes of brain dock testing is to detect unruptured cerebral aneurysms that occur in blood vessels at an early stage, and to take appropriate measures and treatments to treat severe diseases such as subarachnoid hemorrhage due to rupture of cerebral aneurysms. It is to prevent disease.

  Detection of an unruptured cerebral aneurysm by a doctor is performed using an MRA image (Magnetic Resonance Angiography) obtained by imaging the blood flow in the blood vessel by MRI. Usually, when interpreting by a doctor, an image obtained by two-dimensionalizing three-dimensional image data from various angles by MIP processing (Maximum Intensity Projection) is used, but an unruptured cerebral aneurysm that occurs in a blood vessel is small. It is necessary to distinguish from the surrounding blood vessel images that are displayed in a superimposed manner, and the doctor's fatigue is severe. There is also the possibility of oversight due to this fatigue.

  Conventionally, in order to support such a doctor's diagnosis, an apparatus for detecting an image area of a lesion portion by image processing has been developed (see, for example, Patent Document 1, Non-Patent Documents 1 and 2). This apparatus is generally called CAD (Computer-Aided Diagnosis).

  On the other hand, when an image based on MIP is generated and displayed in the medical image processing apparatus at the time of interpretation, a plurality of blood vessels may be displayed in an overlapping manner depending on the viewing direction. In this case, in order to detect a small unruptured aneurysm, it is necessary to distinguish between a blood vessel image that the doctor wants to pay attention to and a surrounding blood vessel image. Fatigue may overlook unruptured aneurysms that should be detected originally.

  It is known that blood vessel sites where aneurysms frequently occur are the middle cerebral artery bifurcation, the anterior traffic artery, the internal carotid artery posterior traffic artery bifurcation, and the like. Therefore, in order to observe these multiple sites in detail, it is only necessary to generate a target MIP image that focuses only on the blood vessel image of the blood vessel site in which the doctor is interested.

In a region where a plurality of blood vessel parts intersect, the depth information along the line-of-sight direction is used to determine the blood vessel image with the larger depth information, that is, the luminance of the intersection of the blood vessel parts on the rear side in the line-of-sight direction. There is also disclosed a technique for lowering the display (see, for example, Patent Document 2). According to this method, a sense of perspective can be given to the blood vessel image, and it becomes easy to observe the blood vessel portion on the near side.
JP 2002-112986 A Japanese Patent Laid-Open No. 5-277091 Norio Hayashi et al., Automatic extraction method of cerebellum and brain affected area in head MRI image using morphological processing, Journal of Medical Image Information Society, vol21.no1.pp109-115,2004 Yokoyama Ryujiro et al., Automatic detection of lacunar infarct area in brain MR images, Journal of Japanese Society of Radiological Technology, 58 (3), 399-405, 2002

  However, there has not yet been proposed a method for accurately detecting an unruptured cerebral aneurysm by distinguishing it from a vascular tissue.

  In addition, since the conventional medical image processing apparatus cannot automatically determine each blood vessel part in the blood vessel image, in order to automatically create the target MIP image, a blood vessel of interest from three-dimensional image data is used. An operation for manually specifying only the blood vessel image of the site is required. Such an operation is not easy and requires time.

  In addition, the method described in Patent Document 2 makes it easy to observe a blood vessel site closer to the observer. Therefore, in the MIP image created according to the line-of-sight direction that the doctor wants to observe, if another blood vessel part intersects the front side of the blood vessel part that the doctor is interested in, it cannot be removed, and the desired It is not always possible to observe a desired blood vessel site in detail from the viewing direction.

  An object of the present invention is to accurately detect a lesion from a head image. Also, it is possible to observe while paying attention to a specific blood vessel site.

According to the invention of claim 1,
Analyzing means for filtering the head MRA image, which is a three-dimensional image, performing multi-resolution analysis, and calculating a vector concentration level for each resolution level using partial images decomposed for each resolution level;
In reconstructing the original head MRA image from the partial image, the weighted image of the partial image is created by weighting the region where the calculated vector concentration degree is a predetermined value, and the partial image is multiplied by the weighted image Reconstruction means for reproducing the original image from the image;
Using the reconstructed head MRA image, a deletion unit that deletes a false positive candidate region that is a normal blood vessel from candidate regions of a lesion portion reproduced in the head MRA image;
A medical image processing apparatus is provided.

According to invention of Claim 2,
The analysis unit performs multi-resolution analysis by performing filter processing using filters H (z) and G (z) represented by the following formulas 11 and 12 on the head MRA image. A medical image processing apparatus according to item 1 is provided.
Note that z in the above formula represents z conversion.

According to invention of Claim 3,
The analysis unit performs filtering on the x, y, and z directions constituting the three-dimensional shape of the head MRA image to obtain partial images in the x, y, and z directions, respectively. A medical image processing apparatus according to 2 is provided.

According to invention of Claim 4,
The reconstruction means filters the partial images obtained by multiplying the weighted image by filters L (z), K (z), H (z) L (z) represented by the following expressions 11, 13, and 14: 4. The medical image processing apparatus according to claim 2, wherein an original image is reproduced by performing processing.
Note that z in the above formula represents z conversion.

According to the invention of claim 5,
The deletion unit calculates a feature amount for a candidate region of a lesion portion reproduced in the reconstructed head MRA image, and deletes a false positive candidate region from the candidate region based on the feature amount. A medical image processing apparatus according to any one of claims 1 to 4 is provided.

According to the invention of claim 6,
The feature amount is at least one of a size of a candidate area of a lesion part reproduced in the reconstructed head MRA image, a sphericity, an average of vector concentration in the candidate area, or a maximum value of vector concentration. The medical image processing apparatus according to claim 5, wherein the medical image processing apparatus is one or more.

According to the invention of claim 7,
The lesion is an aneurysm that occurs in a blood vessel,
An extraction means for extracting a blood vessel image from the head MRA image;
The medical image processing according to any one of claims 1 to 6, wherein the reconstruction unit performs multi-resolution analysis and reconstruction of an original image using the image from which the blood vessel image has been extracted. An apparatus is provided.

According to the invention described in claim 8,
Image control means for discriminating one or a plurality of blood vessel sites included in the blood vessel image of the head MRA image;
Display means for MIP display of the head MRA image;
Operation means for performing display operation;
When a blood vessel part to be displayed in the head MRA image is selected and operated via the operation means, only a blood vessel image corresponding to the selected blood vessel part is extracted from the head MRA image, and the extracted Display control means for displaying only a blood vessel image on the display means;
A medical image processing apparatus according to any one of claims 1 to 7 is provided.

According to the first to sixth aspects of the present invention, the vector concentration degree is calculated for each resolution level based on the partial image obtained by performing the multi-resolution analysis on the head image. By calculating the vector concentration for each resolution level, it is possible to cope with detection of aneurysms of various sizes, and detection processing with higher accuracy is possible. In addition, a lesion such as an aneurysm exhibits a certain high value vector concentration, but according to the present invention, reconstruction is performed so that only a region that is likely to be such a lesion is reproduced. It is possible to obtain a reconstructed image in which only some candidate regions are imaged. By calculating the feature amount using the reconstructed image, image elements other than the candidate region can be excluded from the calculated feature amount, and the feature amount for the candidate region can be accurately calculated. Therefore, it is possible to further improve the accuracy of the detection process itself.

According to the seventh aspect of the present invention, the area where the vector concentration degree is substantially calculated can be narrowed down to the blood vessel area where the lesion part of the aneurysm exists, and the calculation time can be shortened. .

According to the eighth aspect of the present invention, it is possible to extract and observe only a blood vessel image of a specific blood vessel region that a doctor wants to observe from the target image. Since a blood vessel image may be displayed by overlapping a plurality of blood vessel portions, it is difficult to observe the overlapping portion. Therefore, by displaying only the specific blood vessel site selected by the doctor, it is possible to eliminate the overlap of the blood vessel site and provide an environment in which the doctor can easily interpret.

<First Embodiment>
First, the configuration will be described.
FIG. 1 shows a configuration of a medical image processing apparatus 10 in the present embodiment.
The medical image processing apparatus 10 detects a candidate area of a lesion from the medical image by performing image analysis on the medical image obtained by examination imaging.
The medical image processing apparatus 10 is obtained by an image generation apparatus that generates a medical image, a server that stores and manages the medical image, and a medical image stored in the server in order to be read by a doctor. It may be provided in a medical image system in which various devices such as an image interpretation terminal are connected via a network. In the present embodiment, an example in which the present invention is realized by the medical image processing apparatus 10 alone will be described. However, the functions of the medical image processing apparatus 10 are distributed to each component of the medical image system, and the entire medical image system is used. The invention may be realized.

Hereinafter, each part of the medical image processing apparatus 10 will be described.
As illustrated in FIG. 1, the medical image processing apparatus 10 includes a control unit 11, an operation unit 12, a display unit 13, a communication unit 14, a storage unit 15, and a lesion candidate detection unit 16.

The control unit 11 includes a CPU (Central Processing Unit) and a RAM (Random Access Memory).
) And the like, and reads various control programs stored in the storage unit 15 and performs various calculations, and comprehensively controls processing operations in the units 12 to 16.

  The operation unit 12 includes a keyboard, a mouse, and the like. When these are operated by an operator, an operation signal corresponding to the operation is generated and output to the control unit 11. In addition, it is good also as providing the touchscreen comprised integrally with the display in the display part 13. FIG.

  The display unit 13 includes a display unit such as an LCD (Liquid Crystal Display). In response to an instruction from the control unit 11, the display unit 13 displays lesion candidates detected from various operation screens, medical images, and medical images. Various display information such as detection results and detection information is displayed.

  The communication unit 14 includes a communication interface and transmits / receives information to / from an external device on the network. For example, the communication unit 14 performs a communication operation such as receiving a medical image generated from the image generation apparatus and transmitting detection information of a lesion candidate in the medical image processing apparatus 10 to an interpretation terminal.

The storage unit 15 stores various processing programs such as a control program used in the control unit 11 and a detection process used in the lesion candidate detection unit 16 as well as data such as parameters necessary for execution of each program and processing results thereof. ing.
In addition, the storage unit 15 stores medical images that are detection targets of lesion candidates, information on the detection results, and the like.

  The lesion candidate detection unit 16 cooperates with the processing program stored in the storage unit 15 to perform various image processing (gradation conversion processing, sharpness adjustment processing, dynamic range compression processing, etc.) as necessary on the processing target image. ). In addition, the lesion candidate detection unit 16 executes detection processing and outputs the detection result. The contents of the detection process will be described later.

Next, lesion candidate detection processing by the medical image processing apparatus 10 will be described.
In this embodiment, an example will be described in which a lesion head of an unruptured cerebral aneurysm is detected from an MRA image (three-dimensional image) obtained by imaging a patient's head by MRI and imaging blood flow in the brain. . A cerebral aneurysm is a bulge (expansion) formed in the wall of an artery, and is caused by pressure applied by blood flow on the artery wall. Thrombus is likely to occur inside the cerebral aneurysm, and when this cerebral aneurysm ruptures, serious diseases such as subarachnoid hemorrhage develop.

FIG. 2 is a flowchart for explaining the flow of the detection process. This detection process is a process executed by reading the detection process program stored in the storage unit 15 by the lesion candidate detection unit 16 as described above.
As shown in FIG. 2, in the detection process, three-dimensional image data of MRA is first input (step S1). Specifically, the MRA image to be processed stored in the storage unit 15 is read by the lesion candidate detection unit 16.

Here, an image obtained by MRI will be described.
MRI is a method for obtaining an image using nuclear magnetic resonance (hereinafter referred to as NMR) in a magnetic field.
In NMR, a subject is placed in a static magnetic field, and then an RF pulse (radio wave) having a resonance frequency of a nucleus to be detected in the subject is irradiated. For medical purposes, the resonance frequency of hydrogen atoms constituting water, which is abundant in the human body, is usually used. When the object is irradiated with an RF pulse, an excitation phenomenon occurs, the atomic nuclear spins that resonate with the resonance frequency are in phase, and the nuclear spins absorb the energy of the RF pulse. When the irradiation of the RF pulse is stopped in this excited state, a relaxation phenomenon occurs, the phase of the nuclear spin becomes nonuniform, and the nuclear spin releases energy. The time constant for phase relaxation is T2, and the time constant for energy relaxation is T1.

  In MRI, it is possible to obtain various types of images according to the inspection purpose by changing the imaging method. For example, with respect to the repetition time TR and the echo time TE, TR = T1, TE << T2 and the shooting conditions adjusted are T1-weighted images, and TR >> T1, TE = T2 and the shooting conditions are adjusted are T2-weighted images. However, there is a difference in image contrast between them, and the T1-weighted image is mainly used for detecting the anatomical structure and the T2-weighted image is used for detecting the lesioned part. An image photographed by the FLAIR method is a T2-weighted image in which a signal from water is attenuated, and is particularly called a FLAIR image.

  MRA refers to a blood vessel imaging method in MRI. In MRI, it is possible to absorb energy only in a specific slice (tomographic section) by applying a gradient magnetic field in the direction from the foot of the subject to the head (this direction is called the body axis). At the time of imaging, blood is saturated with RF pulses in the blood vessels in the slice, but since blood flows constantly in the blood vessels, the signal intensity in that slice increases when non-saturated blood flows in over time. . MRA is a method of imaging blood vessels with blood flow by imaging this high signal.

FIG. 3A shows an example of an MRA image.
As shown in FIG. 3A, since a blood vessel region with a blood flow has a high signal, the blood vessel region appears white in the MRA image.

  When such 3D data of the MRA image is input, preprocessing is performed on the 3D image data as a preparation stage for detecting candidates (step S2). As preprocessing, normalization processing and gradation conversion processing of image data are performed. Normalization is performed by converting by linear interpolation so that all sides constituting the voxel become three-dimensional image data of the same size. Then, density gradation conversion processing is performed on the three-dimensional image data converted into voxels of the same size, and the signal value of each voxel is linearly converted into density gradations of 0 to 1024. At this time, the higher the signal value, the closer to the density value 1024, and the lower the signal value, the closer to the density value 0. The density gradation range is not limited to 0 to 1024 and can be set as appropriate.

  When the preprocessing is completed, a blood vessel image region is extracted from the three-dimensional MRA image (step S3). First, threshold processing is performed to binarize the MRA image. In general, in the MRA image, as shown in FIG. 3A, the blood vessel region appears white and the other regions appear black. Therefore, in the binarized image, the blood vessel region has a different value from the other regions. Therefore, the blood vessel region is extracted by the region expansion method. First, using the binarized image, the voxel that is the starting point (whitest and high density value voxel) is determined, and in the 3D MRA image before binarization processing, 26 voxels near the voxel determined as the starting point are examined. A neighboring voxel that satisfies a certain determination condition (for example, a density value of 500 or more) is determined as a blood vessel region. Then, the same processing as described above is repeated for the neighboring voxels determined to be the blood vessel region. In this manner, the blood vessel region can be extracted by sequentially extracting the voxels that satisfy the determination conditions while expanding the region. FIG. 3B shows a diagram in which the blood vessel region extracted from the MRA image of FIG. 3A is binarized with white (density value 1024) and the other regions with black (density value 0).

  Next, filter processing using a vector concentration filter as shown in FIG. 4 is performed on the extracted three-dimensional MRA image of the blood vessel region, and a primary candidate for a cerebral aneurysm is obtained from the processed image output by the filter processing. An area is detected (step S4). The vector concentration filter calculates the vector concentration in each voxel, and images and outputs the calculated vector concentration as a voxel value of the voxel. The vector concentration degree focuses on the direction of the gradient vector of the density change and evaluates how much the gradient vector in the neighboring region is concentrated with respect to a certain point of interest.

FIG. 5 shows a cerebral aneurysm model and a blood vessel model.
As shown in FIG. 5, in the case of a cerebral aneurysm, since a spherical aneurysm exists on a linear blood vessel, the gradient vector (the arrow in the figure indicates the direction of the gradient vector) goes to the center of the aneurysm. Tend. On the other hand, since the blood vessel has a linear shape, such a tendency does not occur. Therefore, the region close to the shape of the cerebral aneurysm model has a higher vector concentration value than other blood vessel regions. Therefore, the primary candidate region of the cerebral aneurysm can be detected by outputting only the region having a high vector concentration level by the vector concentration filter.

Specifically, when a focused voxel P as shown in FIG. 4 is scanned on the extracted image of the vascular region, and the extracted vascular region exists within a range of a sphere having a radius R centered on the focused voxel P. Calculate the vector concentration.
The vector concentration is calculated by the following equation 1.
Here, the angle θ represents an angle between the direction vector from the target voxel P to the peripheral voxel Q _j and the direction of the vector in the peripheral voxel Q _j , and M is the number of the peripheral voxels Q _j to be calculated. (See FIG. 4).

  When the vector concentration degree is obtained for each voxel constituting the blood vessel region, a filtered image as shown in FIG. 6 is output using the vector concentration degree as a voxel value. Since the vector concentration degree is output in the range of 0-1, in FIG. 6, the image is imaged so as to appear white and high in density as the vector concentration degree is larger (closer to 1).

  Next, a threshold process for binarizing the filtered image using, for example, a threshold value of 0.5 is performed, and a primary candidate area is detected. That is, an area composed of voxels having a vector concentration degree greater than the threshold value 0.5 by binarization is extracted as a primary candidate area. FIG. 7A shows a part of the filtered image shown in FIG. By binarizing this filtered image, a binarized image as shown in FIG. 7B is obtained. As shown in FIG. 7B, an image area that appears white in the binarized image, that is, has a large vector concentration degree, is output as a primary candidate area.

Note that the threshold for binarization is determined in advance using a teacher image in which the presence of a cerebral aneurysm is known. That is, by thresholding the teacher image, a threshold value that can extract only the image region of the cerebral aneurysm whose presence has already been found is obtained.
The threshold value may be obtained by statistical analysis by the p-tile method or the like. The p-tile method is a method of obtaining a density histogram and obtaining a density value at a certain area ratio p% in the density histogram as a threshold value. In the present embodiment, the density histogram of the filtered image is obtained, and the density value that occupies the area ratio p% from the highest density value side is determined as the threshold value.

  Next, in order to perform secondary detection, a feature amount indicating the feature of the primary candidate region is calculated (step S5). Since the cerebral aneurysm has a certain size and a spherical shape, in this embodiment, the size of the candidate region, the sphericity, and the average value of the vector concentration in each voxel in the region are calculated as feature quantities. I decided to. However, there is no particular limitation on what kind of feature value is used as long as it can characterize a cerebral aneurysm. For example, the maximum value of vector concentration of each voxel may be calculated, and the concentration of each voxel may be calculated. The standard deviation of the values may be calculated.

  As the size feature amount, the volume of each voxel constituting the candidate region is calculated. Here, in order to shorten the processing time, the number of voxels is calculated instead of the actual volume, and this is used as an index value indicating the volume in subsequent calculations. Further, as shown in FIG. 8, the feature amount of the sphericity is obtained when a sphere having the same volume as that of the primary candidate region is arranged so that the centroid of the primary candidate region and the centroid of the sphere coincide with each other. Is obtained from the ratio of the volume of the primary candidate region portion that coincides with the total volume of the primary candidate region.

  When the feature amount is calculated, secondary detection is performed based on the feature amount (step S6). In the secondary detection, a cerebral aneurysm is distinguished from a blood vessel that is a normal tissue. Identification is an example of a classifier using a rule-based method, but is not limited to this, and any method may be used as long as it can be identified, such as an artificial neural network, a support vector machine, or discriminant analysis.

  In the rule-based method classifier, as shown in FIGS. 9A and 9B, first, a profile indicating the relationship between the sphericity with respect to the size and the average of the vector concentration with respect to the size is created. In the discriminator, a range to be detected as a secondary candidate (a range surrounded by a solid line in FIGS. 9A and 9B) is determined in advance, and the profile is created using each feature quantity of the primary candidate to be identified as variable data. At this time, if the variable data of each feature quantity of the primary candidate is within the detection range, it is identified as a true positive candidate, and if not, it is identified as a false positive candidate. In other words, only the primary candidates in which the feature quantity variable data is distributed within the detection range are secondarily detected.

  The detection range is determined by using a teacher image that has been previously determined to be a cerebral aneurysm or a normal blood vessel. First, the average of size, sphericity, and vector concentration is obtained from cerebral aneurysm and normal blood vessel teacher images, respectively, and the relationship between the sphericity with respect to size and the average value of vector concentration with respect to size is obtained from these features. Profile is created. In FIG. 9A and FIG. 9B, the variable data indicated by the ◯ marker is true positive, that is, teacher data that has been determined to be a cerebral aneurysm, and the variable data indicated by the ● marker is false positive, that is, has been determined to be a blood vessel. Teacher data. At this time, since the image features are different between the cerebral aneurysm and the normal blood vessel, the distribution of the feature values of the average values of the size, the sphericity, and the vector concentration degree is biased. Therefore, two threshold values are determined for the average values of the size, the sphericity, and the vector concentration so that the range in which teacher data of all cerebral aneurysms is included in the profile is minimized. A region range surrounded by the threshold is a detection range.

  In the example of FIGS. 9A and 9B, a range surrounded by four dotted lines indicating the threshold (only a portion surrounding the range is indicated by a solid line) is a detection range, and “size-sphericity”, “size- A detection range is determined for each of “average vector concentration”. In any profile, a primary candidate (its variable data is indicated by a marker Δ) located within the detection range is secondarily detected as a cerebral aneurysm candidate.

  When secondary detection is performed in this way, tertiary detection is further performed on the secondary detection candidates (step S7). In tertiary detection, discriminant analysis is performed using three feature amounts. As a discriminant analysis method, any of Mahalanobis distance, principal component analysis, linear discriminant function, etc. can be applied, but a method different from the method at the time of secondary detection is applied.

Then, the detection result is output on the assumption that the third candidate area is the final cerebral aneurysm candidate (step S8).
FIG. 10 shows an example of a detection result displayed on the display unit 13 as a detection result.
As shown in FIG. 10, the display unit 13 displays marker information (markers indicated by arrows in FIG. 10) indicating the candidate regions of the third detected cerebral aneurysm in the MIP image created from the three-dimensional MRA image. Is done. By performing such display, the cerebral aneurysm candidate region and other regions can be identified. The MIP image is a two-dimensional image created by performing MIP processing on the three-dimensional MRA image data, and it is possible to display the structure in the image three-dimensionally. The MIP process is called a maximum brightness projection method, which is a two-dimensional image that is projected from a certain direction with parallel rays and enables the three-dimensional observation by reflecting the maximum brightness (signal value) in the voxel on the projection plane. Is the process of creating.

  Information regarding detection of cerebral aneurysm candidates, such as the average value of vector concentration calculated for the cerebral aneurysm candidate region, may be output and used as reference information at the time of diagnosis by a doctor. Further, the color of the marker information may be changed according to the degree of vector concentration. For example, change the color of the arrow marker to red if the vector concentration is 0.8 or higher, yellow if it is 0.7 to 0.8, blue if it is 0.7 to 0.5, etc. Thus, the doctor can easily visually recognize that the vector concentration, that is, the degree of the spherical shape of the aneurysm is strong.

  FIG. 10 shows an example in which the position of a cerebral aneurysm candidate can be identified by marker information. However, if the cerebral aneurysm candidate region is displayed so as to be distinguishable from other regions, such a display method is used. Not exclusively. For example, instead of MIP processing, the volume rendering method may be used for three-dimensional display. The volume rendering method is a method of performing three-dimensional display by giving color information and opacity to each voxel for each partial area. The opacity is set high for the attention area and low for the other areas. By doing so, the attention area can be raised. Therefore, at the time of display, opacity is set for each region, and color information setting processing corresponding to the opacity is performed.

  Moreover, it is good also as showing a detection result not using a MIP image etc. but using the filter process image (refer FIG. 6) obtained by the vector concentration degree filter. At this time, for example, a region with a low vector concentration is colored blue, and a region with a high value is red, so that the doctor can visually grasp the calculated vector concentration. Good. Further, the filtered image in which the color of the blood vessel region is changed according to the vector concentration degree may be displayed superimposed on the corresponding position of the MIP image. As described above, it is possible to provide vector concentration information as reference information for detecting a cerebral aneurysm by a doctor.

Next, the blood vessel part determination process will be described with reference to FIG.
This blood vessel part discrimination process is a software process realized by cooperation of the control unit 11 and the blood vessel part discrimination process processing program stored in the storage unit 15. In the blood vessel part discrimination processing, one or a plurality of blood vessel parts included in the blood vessel image are discriminated for the blood vessel image appearing on the three-dimensional MRA image obtained by photographing the head.

  MRA is one type of MRI blood vessel imaging method. In MRI, it is possible to absorb energy only in a specific slice (tomographic section) by applying a gradient magnetic field in the direction from the foot of the subject to the head (this direction is called the body axis). At the time of imaging, blood is saturated in the blood vessel in the slice by the RF pulse, but since the blood vessel always flows, the signal intensity in the slice increases when the unsaturated blood flow flows with time. MRA is a method of imaging blood vessels with blood flow by imaging this high signal.

First, a reference image necessary for determining a blood vessel site will be described.
As shown in FIG. 12A, the reference image is obtained by presetting positions and names of one or a plurality of blood vessel sites in a blood vessel image on the three-dimensional MRA image. Here, the blood vessel part refers to anatomical blood vessel classification, and the position of the blood vessel part refers to the position of the voxel belonging to the blood vessel part.

  In FIG. 12A, eight vascular sites (anterior cerebral artery, right middle cerebral artery, left middle cerebral artery, right internal carotid artery, left internal carotid artery, right posterior cerebral artery, left posterior cerebral artery, brain base included in the vascular image are shown. An example in which the position of the artery and the name of the blood vessel part is set is shown. In FIG. 12A, the names are shown only for three blood vessel parts (the right middle cerebral artery, the anterior cerebral artery, and the basilar artery) among the eight blood vessel parts, but the names are set for all eight blood vessel parts. The

  The reference image g2 is created from the three-dimensional data of the head MRA image g1 selected for the reference image as shown in FIG. 12B. First, an axial image (a two-dimensional tomographic image in which voxels are cut out in a plane perpendicular to the body axis) is created from this three-dimensional data at certain intervals. Based on the doctor's indication in one axial image, a voxel belonging to each blood vessel part is designated by manual operation, and further, the name of the blood vessel part is designated. By repeating this for each axial image sliced by changing the position in the body axis direction, it is possible to set the position of the voxel belonging to each blood vessel part and the name of the blood vessel part among all the voxels constituting the three-dimensional data. .

  In the reference image g2, landmark voxels are set at characteristic points such as a blood vessel bending point, an end point, and an intersection of blood vessel parts. The landmark is used for alignment between the target image and the reference image, and will be described in detail later. The landmark is also set according to a manual operation based on a doctor's indication.

The reference image g2 created as described above is stored in the storage unit 15.
The reference image g2 may be created by the control unit 11 of the medical image processing apparatus 10, or an externally created image may be stored in the storage unit 15. In addition, in order to show eight blood vessel parts included in the blood vessel image, each blood vessel part is identified and displayed in FIG. 12A, but the actual reference image g2 has a black background (low signal value) and a white blood vessel image (high signal value). ) And a binarized image. The position information of the voxels belonging to each blood vessel site, the name information, and the position information of the voxels that are landmarks are attached to the reference image or stored in the storage unit 15 as separate files in association with the reference image. ing.

Next, the blood vessel part discrimination process using the reference image will be specifically described.
In the medical image processing apparatus 10, a normalization process is first performed by the control unit 11 on a three-dimensional MRA image (hereinafter referred to as a target image) to be discriminated (step S11).
Since the MRA image used as the target image is an image of the blood flow, the voxel may be a cuboid that is not the same size or the maximum and minimum values of the voxel value may vary depending on the subject and imaging conditions. To do. Therefore, normalization processing is performed to unify the preconditions regarding the target image.

  In the normalization process, first, the target image is converted by linear interpolation so that all sides constituting the voxel have the same size. Next, a histogram is created for the voxel values of all the voxels of the target image. The voxel values of the upper 5% or more of the histogram are set to 1024, the minimum voxel value is set to 0, and all the voxel values of the target image are 0 to 1024. Linear conversion to gradation. At this time, the higher the voxel value, the closer to the density value 1024, and the lower the signal value, the closer to the density value 0. The density gradation range is not limited to 0 to 1024 and can be set as appropriate.

An example of the normalization process is shown in FIGS. 13A and 13B.
The target image g3 shown in FIG. 13A and the target image g4 shown in FIG. 13B are different patients as subjects. Therefore, the histogram h1 obtained from the target image g3 (see FIG. 13A) and the histogram h3 obtained from the target image g4 (see FIG. 13B) share the characteristic that there are two local maximum points. It can be seen that there is a considerable difference in the value range, and the histogram characteristics are different as a whole. When the normalization processing is performed on the target images g3 and g4 having such histogram characteristics and then the histogram is created again, the histogram h2 shown in FIG. 13A and the histogram h4 shown in FIG. 13B are obtained. As can be seen from the histograms h2 and h4, the histogram characteristics of the target images g3 and g4 are substantially the same by the normalization process.

When the normalization process is finished, the control unit 11 extracts a blood vessel image from the normalized target image (step S12).
First, threshold processing is performed on the target image, and binarization is performed. In general, in the MRA image, as shown in FIG. 14A, the blood vessel image appears white and the other tissue portions appear blackish. Therefore, in the binarized image, the blood vessel image has a value different from other regions. Therefore, the region having the same signal value as the blood vessel image is extracted by the region expansion method.

  In the region expansion method, a voxel that is a starting point (whitest and high density value voxel) is determined using a binarized image, and 26 voxels in the vicinity of the voxel determined as the starting point in the target image before binarization processing are determined. A nearby voxel satisfying a certain determination condition (for example, having a density value of 500 or more) is determined as a blood vessel image. Then, the same processing as described above is repeated for the neighboring voxels determined to be the blood vessel image. In this manner, the image region of the blood vessel image can be extracted by sequentially extracting the voxels that satisfy the determination condition while expanding the region.

FIG. 14B shows a blood vessel extraction image g6 obtained by extracting blood vessel images.
A blood vessel extraction image g6 is obtained by extracting a blood vessel image from the normalized target image g5 shown in FIG. 14A, and the blood vessel image area is white (density value 1024) and the other areas are black (density value 0) and binary. It has become.

  Next, in order to make the position of the blood vessel image of the blood vessel extraction image substantially coincide with the position of the blood vessel image of the reference image, the control unit 11 performs alignment based on the position of the center of gravity of each image (step S13). The center-of-gravity position is the position of the voxel that is the center of gravity of all the voxels belonging to the blood vessel image.

A specific description will be given with reference to FIGS. 15A and 15B.
FIG. 15A is a diagram in which the blood vessel extraction image and the reference image before alignment are superimposed. It can be seen from FIG. 15A that the positions of the respective blood vessel images do not coincide by simply combining the blood vessel extraction image and the reference image.
Therefore, the control unit 11 obtains the positions of the centroid P (x1, y1, z1) of the blood vessel extraction image and the centroid Q (x2, y2, y3) of the reference image as shown in FIG. 15A. Next, the blood vessel extraction image or the reference image is translated so that the barycentric positions P and Q coincide. FIG. 15B shows the result of matching the gravity center positions P and Q by translation. It can be seen from FIG. 15B that the positions of the blood vessel image of the blood vessel extraction image and the blood vessel image of the reference image roughly match.

Further, in order to perform alignment with high accuracy, the control unit 11 performs rigid body deformation on the blood vessel extraction image (step S14).
First, a corresponding point search using a cross-correlation coefficient is performed as preprocessing for rigid body deformation. This is because a plurality of corresponding points are set for each of the two images to be aligned, and one of the images is rigidly deformed so that the corresponding points set in the two images match each other. Here, a voxel of a landmark determined in advance in the reference image and a voxel of a blood vessel extraction image having locally similar image characteristics are set as corresponding points. The similarity of the image characteristics is determined based on the cross-correlation coefficient obtained for the blood vessel extraction image and the reference image.

  Specifically, as illustrated in FIG. 16A, corresponding points corresponding to 12 landmarks set in advance in the blood vessel image of the reference image g7 are searched from the blood vessel extraction image. When searching for corresponding points, as shown in FIG. 16B, in the blood vessel extraction image g8, a voxel at a position corresponding to each landmark of the reference image g7 is set as a start point, and the start point and the blood vessel extraction image g8 and the reference image g7 Voxels within a range of −10 to +10 voxels in the X-axis, Y-axis, and Z-axis directions (21 × 21 × 21 voxel cubic regions) are searched from the landmark voxels. A relation number C (hereinafter referred to as a correlation value C) is calculated.

In Equation 2, A (i, j, k) represents the voxel position of the reference image g7, and B (i, j, k) represents the voxel position of the blood vessel extraction image g8. IJK indicates the size of the search area, and IJK = 21 × 21 × 21.

Α and β are average values of the voxel values in the search region in the reference image g7 and the blood vessel extraction image g8, respectively, and are expressed by the following equations 3 and 4. σ _A and σ _B are standard deviations of the voxel values in the search region in the reference image g7 and the blood vessel extraction image g8, respectively, and are expressed by the following expressions 5 and 6.

The correlation value C has a range of −1.0 to 1.0, and the closer to the maximum value 1.0, the more similar the image characteristics of the reference image g7 and the blood vessel extraction image g8 are.
Therefore, the position of the voxel having the largest correlation value C is set as the corresponding point of the blood vessel extraction image g8 corresponding to the landmark of the reference image g7.

  When the corresponding point is set, the control unit 11 performs rigid body deformation on the blood vessel extraction image g8 based on the corresponding point, thereby aligning the blood vessel image of the blood vessel extraction image g8 and the blood vessel image of the reference image g7. Is done. Rigid body deformation is one of affine transformations, in which coordinate transformation is performed by rotation and translation. The alignment is performed by an ICP (Iterative Closest Point) algorithm that repeats rigid body deformation using the least square method a plurality of times so that the corresponding point of the blood vessel extraction image g8 matches the landmark of the reference image g7. According to this algorithm, every time a rigid body is deformed, the least square error of the distance at the corresponding point between the landmark of the reference image and the blood vessel extraction image is calculated, and the end condition such that the least square error exceeds a threshold value is satisfied. The rigid body deformation is repeated until. Note that rigid body deformation may be replaced with affine transformation (enlargement / reduction, rotation, coordinate transformation by parallel movement).

Next, based on the reference image, each blood vessel part included in the blood vessel image of the blood vessel extraction image aligned by rigid body deformation is determined by the control unit 11 (step S15).
First, the square of the Euclidean distance from a certain target voxel in the blood vessel extraction image is obtained for all voxels belonging to each blood vessel part of the reference image (this is referred to as a target voxel). Then, it is determined that the blood vessel part to which the target voxel having the shortest Euclidean distance belongs is the blood vessel part to which the target voxel belongs. At this time, the name of the blood vessel part of the target voxel is determined from the name of the blood vessel part set in the target voxel.

  For all the voxels constituting the blood vessel image of the blood vessel extraction image, the target voxel is set and the above processing is performed. When the corresponding blood vessel portion is determined for all the voxels, the voxel belonging to the determined blood vessel portion is determined. Blood vessel part information indicating the position and name of the blood vessel part is generated by the control unit 11 and attached to the target image (step S16). For example, when it is determined that the voxel at the position (x3, y3, z3) is a blood vessel part of the anterior cerebral artery, the voxel at the position “(x3, y3, z3)” has the blood vessel name “anterior cerebral artery”. The blood vessel part information indicating that it is a blood vessel part is appended to the header area of the target image.

FIG. 17A and FIG. 17B show the results of determining the blood vessel site.
The blood vessel extraction image g9 shown in FIG. 17A and the blood vessel extraction image g11 shown in FIG. 17B are images obtained from different subjects, respectively, and the image g10 shown in FIG. 17A and the image g12 shown in FIG. This is an image in which each blood vessel part is discriminated from g11 and is identified and displayed by changing the color for each blood vessel part. By performing alignment by rigid body deformation from the images g10 and g12, the blood vessel images are the same even though the blood vessel images of the images g10 and g12 have different forms (blood vessel position, size, extending direction, etc.). It can be seen that the part can be identified.

The above is a flow from determining a blood vessel part in the target image to attaching the blood vessel part information.
When a display instruction operation is performed on such a target image via the operation unit 12, the control unit 11 performs MIP processing on the target image to generate a MIP image and displays it on the display unit 13. Hereinafter, the display of the MIP image is referred to as MIP display.

  The MIP method is a method of creating a two-dimensional image by performing projection with parallel rays from a certain direction and reflecting the maximum luminance (voxel value) in the voxels on the projection line on the projection plane. This projection direction is the line-of-sight direction that the doctor desires to observe. Here, the doctor can freely operate the observation direction, and the control unit 11 creates an MIP image from the target image according to the observation direction instructed via the operation unit 12, and the display unit 13. Shall be displayed.

  When an instruction operation for identifying and displaying a blood vessel part is performed in a state where the target image is displayed in MIP, the control unit 11 refers to blood vessel part information attached to the target image, and based on this blood vessel part information, the MIP image Display control is performed so that each blood vessel part can be identified in the blood vessel image in the middle (step S17).

  For example, when the voxel position and name of each blood vessel part are determined based on the blood vessel part information, the control unit 11 displays blue for the voxel belonging to the blood vessel part of the anterior cerebral artery in the MIP image, and the blood vessel part of the basilar artery. A color for each blood vessel part is set to a voxel at a position determined for each blood vessel part such as green for the voxel to which it belongs. Then, the set color is reflected in the MIP image of the target image. Furthermore, an annotation image indicating the name of the blood vessel part is created and synthesized with the corresponding blood vessel part of the MIP image.

FIG. 18 shows an example of identification display.
In FIG. 18, a MIP image g13 is a target image that is displayed in MIP from above the head. When the identification display of each blood vessel part is instructed with the MIP image g13 displayed, an identification display image g14 is displayed. The identification display image g14 is obtained by identifying and displaying each blood vessel part by giving different colors to the eight kinds of blood vessel parts in the blood vessel extraction image.
In the identification display image g14, when a doctor selects a blood vessel part, an annotation m indicating the name of the blood vessel part such as “basal artery” is associated with the selected blood vessel part by display control of the control unit 11. Is displayed.

  When MIP display from the side surface direction is instructed with respect to the MIP image g13 in the head upward direction, the control unit 11 creates and displays an MIP image g15 corresponding to the side surface direction. In the MIP image g15, the blood vessel image on the near side and the blood vessel image on the rear side overlap each other, and the display is difficult to observe. Even in such an MIP image g15, the blood vessel part can be identified and displayed. Since it is possible to determine which voxel corresponds to which blood vessel part by referring to the blood vessel part information, the voxel to be identified and displayed can be specified regardless of the change in the observation direction of the MIP display. is there.

The identification display image corresponding to the MIP image g15 is the image g16. In this identification display image g16, it is possible to extract and display only one of the blood vessel sites.
When any one blood vessel part is selected and operated via the operation unit 12, the control unit 11 projects from the MIP image in which only the luminance of the voxel of the selected blood vessel part is projected, that is, the MIP image g15 of the target image. A blood vessel selection image g17 in which only the selected blood vessel site is extracted is displayed. In the blood vessel selection image g17, only the selected blood vessel part is displayed in MIP and the other blood vessel parts are not displayed, so that the doctor can observe only the blood vessel part of interest.

  These display images g13 to g17 may be displayed side by side on the same screen, or may be switched and displayed as one image per screen. In the former case, the MIP display image g13, the identification display image g14, the blood vessel selection image g17, and the like can be compared and observed. In the latter case, the images g13 to g17 can be observed on a full screen display. , Making it easier to observe details.

Hereinafter, examples of the detection process will be described.
Three-dimensional MRA image data for 20 patients was obtained. These image data have a matrix size of 256 × 256, a spatial resolution of 0.625 to 0.78 mm, and a slice thickness of 0.5 to 1.2 mm. These image data are known to contain seven unruptured cerebral aneurysms. An unruptured cerebral aneurysm has been determined by an experienced neurosurgeon.

  First, these three-dimensional MRA image data were converted into image data of equal voxel size using a linear interpolation method, and normalized. By this normalization, all the image data became equal voxel image data having a matrix size of 400 × 400 × 200 and a spatial resolution of 0.5 mm. Next, after performing gradation conversion processing and linearly converting the density gradation from 0 to 1024, the size (volume), sphericity, and vector concentration of the cerebral aneurysm region (true positive) determined by the doctor Are calculated for the normal blood vessel region (false positive). These feature amounts were used as teacher data for determination of the detection range of the rule-based method, which is a discriminator for secondary detection, and discriminant analysis. Similarly, learning of the classifier was performed using it as a teacher image in the classifier at the time of tertiary detection.

  7 was performed on the three-dimensional MRA image data of 20 cases using the classifier constructed as described above. As a result, 1.85 false positive candidates were included in each case. The detection accuracy rate of the aneurysm was 100%, and it was confirmed that it could be detected with high accuracy.

  As described above, according to the first embodiment, using a vector concentration filter, a cerebral aneurysm candidate having a characteristic that a gradient vector concentrates in a central portion is accurately detected, and the detection information is provided to a doctor. Can do. Therefore, fatigue and oversight at the time of doctor's interpretation work can be prevented, and improvement in diagnostic accuracy is expected.

  In addition, since the cerebral aneurysm candidate region is displayed so as to be distinguishable from other regions, the doctor can easily grasp the detection result on the image, and the interpretation work can be facilitated.

  Further, since the vector concentration degree filter is applied not to the MRA image itself but to the extracted image obtained by extracting the blood vessel region, the processing time required for the filter processing can be shortened.

In addition, the said embodiment is a suitable example to which this invention is applied, and is not limited to this.
For example, in the above description, a cerebral aneurysm candidate is detected using a three-dimensional MRA image, but detection may be performed using a two-dimensional MRA image. In this case, a two-dimensional feature amount is calculated with the size of the cerebral aneurysm candidate as the number of pixels and the sphericity as the circularity.

  Moreover, although the example which detects an aneurysm candidate using an MRA image was demonstrated, you may use the MRI image by other imaging methods, such as a contrast MRA image which image | photographed the blood vessel area | region using the contrast agent, or CTA. An image obtained by imaging a blood vessel region by another imaging apparatus such as (Computed Tomography Angiography) or DSA (Digital Subtraction Angiography) may be used. The detection target is not limited to an aneurysm, but can be detected by applying the present invention as long as the lesion has a spherical aneurysm.

  In addition, according to the above-described embodiment, the reference image in which the positions and names of the eight blood vessel portions included in the blood vessel image on the image are predetermined and the target image are aligned with respect to the blood vessel image. The position and name of each blood vessel part included in the blood vessel image are determined. Since the position and name information is attached to the target image as blood vessel part information, each blood vessel part can be easily determined based on the blood vessel part information when performing MIP display of the target image. The blood vessel part can be identified and displayed. Therefore, the doctor can observe the target image while paying attention to a specific blood vessel site in the target image, and can improve the interpretation efficiency.

When performing MIP display, depending on the direction in which the MIP display is performed, a plurality of blood vessel portions may overlap, and it may be difficult to observe the blood vessel portion to be noticed.
However, according to the present embodiment, each blood vessel part can be identified and displayed. Therefore, the doctor can easily specify the position and name of each blood vessel part, and the interpretation work can be made more efficient.

  In addition, a target MIP image in which only the selected blood vessel part is extracted and displayed in response to an operation for selecting an arbitrary blood vessel part among the identified blood vessel parts is generated and displayed. It is possible to observe only the desired blood vessel site. Therefore, duplication of a plurality of blood vessel sites can be eliminated, and observation can be performed by specifying a specific blood vessel site such as a site where multiple aneurysms occur.

  When aligning the target image and the blood vessel image of the reference image, first the rough alignment is performed based on the position of the center of gravity, and then the target image is rigidly deformed by the cross-correlation coefficient to obtain the target image. Fine alignment is performed so that the feature points of the blood vessel image coincide with the corresponding feature points of the blood vessel image of the reference image.

Even in different subjects (patients), the shape of the main blood vessel site (the length, extension direction, thickness, etc.) of the blood vessel is generally the same. The shape of the blood vessel site varies depending on the subject.
However, as in the present embodiment, the position of the characteristic points such as the main inflection points of the blood vessel image is finally adjusted by rigid body deformation, so that each blood vessel portion in the target image can take any form. It is possible to accurately correspond to the blood vessel portion of the reference image. Therefore, it is possible to determine the blood vessel part uniformly regardless of the subject, and the versatility is high.

  Furthermore, since alignment is performed in two stages (based on the center of gravity position and based on rigid body deformation), it is possible to improve the determination accuracy for determining the blood vessel site. In addition, by performing alignment based on the position of the center of gravity before the rigid body deformation, it is possible to shorten the processing time related to the rigid body deformation, and the processing efficiency is good.

In addition, embodiment mentioned above is a suitable example to which this invention is applied, and is not limited to this.
For example, although an example using an MRA image has been described, an MRI image obtained by another imaging method such as a contrast MRA image obtained by imaging a blood vessel using a contrast agent may be used. Moreover, it is good also as using the image which image | photographed the blood vessel with other imaging devices, such as CTA (Computed Tomography Angiography) and DSA (Digital Subtraction Angiography).

Second Embodiment
In the second embodiment, an example of detection processing by the GC filter bank will be described.
The medical image processing apparatus according to the second embodiment has the same configuration as the medical image processing apparatus 10 according to the first embodiment, and only the operation is different. Therefore, the same components as those of the medical image processing apparatus 10 (see FIG. 1) according to the first embodiment are denoted by the same reference numerals, and the operation of the medical image processing apparatus 10 of the second embodiment will be described below.

FIG. 19 is a flowchart showing a detection process according to the second embodiment.
In the detection process shown in FIG. 19, first, three-dimensional image data of MRA is input (step S101), and pre-processing is performed on the three-dimensional image data (step S102). When the preprocessing is completed, a blood vessel image region is extracted from the three-dimensional image data (step S103). Note that steps S101 to S103 are the same processing as steps S1 to S3 described with reference to FIG. 2 in the first embodiment, and thus detailed description thereof is omitted here.

Next, the primary candidate region of the cerebral aneurysm is detected by the GC filter bank using the extracted three-dimensional MRA image of the blood vessel region (step S104).
Hereinafter, a detection method using the GC filter bank will be described.
The GC filter bank is a combination of various filter processes, and is divided into an analysis bank and a reconstruction bank. The analysis bank performs multi-resolution analysis of an original image (three-dimensional MRA image of a blood vessel region) to create images with different resolution levels (hereinafter referred to as partial images), and creates a weighted image from the partial images. On the other hand, the reconstruction bank weights each partial image with a weighted image, and then reconstructs the original image from the weighted partial image.

FIG. 20 shows an analysis bank.
As shown in FIG. 20, in the analysis bank, a filter process is performed in the filter bank A (z ^j ) using the three-dimensional MRA image of the blood vessel region as the original image S ₀ , and partial images of each resolution level j are sequentially created. Here, an example of creating partial images with resolution levels 1 to 3 will be described.
As shown in FIG. 21, the filter bank A (z ^j ) is subjected to filter processing using filters H (z ^j ) and G (z ^j ), and the image S _j−1 is converted into partial images S _j , Wz _j , Wy _j , It is decomposed into Wx _j . Here, S _j is obtained by applying a smoothing filter H (z ^j ) to S _j−1 in each of the x, y, and z directions. The smoothing filter H (z ^j ) is represented by the following formula 11.
In addition, z used in the said Formula 11 shows z conversion (Hereinafter, it is the same in Formula 12-14 which shows a filter.).

The partial images Wz _j , Wy _j , and Wx _j are first-order difference images obtained by applying the filter G (z ^j ) for each of the x, y, and z directions of the image S _j−1 . The filter G (z ^j ) is represented by the following expression 12.
The partial images Wz _j , Wy _j , and Wx _j have higher detectability with respect to fine fluctuations as the resolution level is smaller (j = 1). On the other hand, the greater the resolution level (j = 3), the higher the detectability for rough fluctuations. Therefore, the resolution level is lower partial image _{Wz j,} Wy _j, Wx j is suitable for the detection of small size aneurysm, the resolution level is high partial image _{Wz j,} Wy _j, Wx j is suitable for the detection of large aneurysms ing.

S _j is further subjected to filter processing by the filter bank A (z ^{j + 1} ) at the resolution level j + 1. In this way, the partial filter S _j , Wz _j , Wy _j , Wx _j is created for each resolution level j (j = 1 to 3) by performing the repeated filter processing.

Next, as shown in FIG. 20, the partial image Wz _j , Wy _j , Wx _j of each resolution level is subjected to filter processing by the vector concentration filter GC _j to calculate the vector concentration for each resolution level. The Since the calculation method of the vector concentration degree has been described above, the description thereof is omitted here.

The calculated vector concentration degree is input to the neural network NN.
In the neural network NN, an output value is output in a range of 0 to 1 so that a value of 1 is output as the possibility of a cerebral aneurysm is increased using teacher data in advance, and a value of 0 is output if it is lower. Designed to.

  When an output value is obtained from the neural network NN, the determination unit NM generates and outputs a weight image V based on the output value. In the weighted image V, a voxel value is set to 1 for a voxel having an output value larger than a certain threshold value (here, 0.8), and a voxel value is set to 0 for a voxel having an output value equal to or less than the threshold value 0.8. Is created. That is, there is a high possibility that a voxel for which the vector concentration degree exceeding the threshold value of 0.8 is calculated is a voxel constituting a region of a cerebral aneurysm. On the other hand, a voxel having a vector concentration degree equal to or less than the threshold value 0.8 is highly likely to be a voxel constituting a normal blood vessel region. Therefore, the weighted image is created by binarizing the voxel value with the threshold as the boundary. The weight image V is input to the reconstruction bank.

FIG. 22 is a diagram illustrating a reconfiguration bank.
The reconstruction bank performs weighting processing on each partial image S _j , Wz _j , Wy _j , Wx _j and then reconstructs the original original image S ₀ via the filter bank S (z ^j ).
In the weighting process, each partial image S _j , Wz _j , Wy _j , Wx _j is multiplied by the weight image V. That is, the value of 1 or 0 set for each voxel in the weight image V is used as a weighting coefficient in the weighting process. Each partial image multiplied by the weight image V is input to the filter bank S (z ^j ).

As shown in FIG. 23, the filter bank S (z ^j ) is subjected to filter processing using filters L (z ^j ), K (z ^j ), H (z ^j ) L (z ^j ), and then S _j , Wz _j , S _j-1 is reconstructed from Wy _j and Wx _j .
Here, the filters L (z ^j ) and K (z ^j ) are expressed by the following equations 13 and 14.

That is, S _j is subjected to filter processing by the filter L (z ^j ) in the x, y, and z directions. Further, Wz _j filter ^K for x-direction for ^(z j), y, for z-direction filter ^{H (z j) L (z} j) is applied. Wy _j For the y-direction filter K ^(z j), the z-direction filter ^{H (z j) L (z} j) is applied, for z-direction with respect to Wx _j filter K ^{(z j)} is Applied. Then, S _j−1 is obtained by adding S _j , Wz _j , Wy _j , and Wx _j after the respective filter processes.

S _j−1 is further subjected to filter processing by the filter bank S (z ^j−1 ), and S _j−2 is reconfigured. It is possible to reconstruct the original image S ₀ by repeating such filtering.

The reconstruction of the original image S ₀ is performed by the filters H (z ^j ), G (z ^j ), L (z ^j ), and K (z ^j ) in the filter banks A (z ^j ) and S (z ^j ) Only in the case of the combination of filters shown in FIGS. 21 and 23, the original image can be completely reproduced. Therefore, as a result of the reconstruction, the voxel value is set to ₀ in the weight image V while the original original image S ₀ is completely reproduced for the voxel in which the voxel value is set to 1 in the weight image V. not reproduced the original image S ₀ for voxels, the voxel value becomes zero. Thus, GC filter bank output image output by the S ₀ is an image in which only the image area is likely a cerebral aneurysm has been reproduced. Images appearing in the output image S ₀ is the primary candidate for cerebral aneurysm.

When the output image _{S 0} obtained feature quantity calculation is performed by using the output image _{S 0} (step S105). As in the first embodiment, the feature amount is calculated by calculating the average value of the size of the candidate region, the sphericity, and the vector concentration in each voxel in the region. After secondary detection is performed using the calculated feature amount (step S106), further tertiary detection is performed (step S107), and the final detection result is output (step S108). Since the processing of steps S105 to S108 is the same as steps S5 to S8 described with reference to FIG. 2, detailed description thereof is omitted.

  In addition, the blood vessel part determination process is also performed in the second embodiment, but since the process contents are the same as those in the first embodiment, the description of the process contents and effects will be omitted.

  As described above, according to the second embodiment, the vector concentration degree is calculated for each resolution level based on the partial image obtained by performing the multiresolution analysis on the head image by the GC filter bank. By calculating the vector concentration for each resolution level, it is possible to cope with detection of aneurysms of various sizes, and detection processing with higher accuracy is possible.

  In addition, a voxel value is set to 1 so that the original image is reproduced only for voxels having a vector concentration level equal to or greater than a threshold value, and a weighted image with the voxel value set to 0 is generated for the other voxels. The partial image is multiplied and weighted, and the original image is reconstructed from the weighted partial images. Therefore, only the region having a high vector concentration degree, that is, the region having a high possibility of being an aneurysm is reconstructed, and a reconstructed image obtained by imaging only the candidate region of the aneurysm can be obtained. By calculating the feature amount using such a reconstructed image, image elements other than the candidate region can be excluded from the calculated feature amount, and the feature amount for the candidate region can be accurately calculated. It becomes. Therefore, it is possible to further improve the accuracy of the detection process itself.

It is a figure which shows the internal structure of the medical image processing apparatus which concerns on embodiment. It is a flowchart explaining the flow of the detection process performed by a medical image processing apparatus. It is a figure which shows the example of MRA image. It is a figure which shows the example of an extraction image which extracted the blood vessel area | region. It is a figure which shows a vector concentration degree filter. It is a figure which shows a cerebral aneurysm model and a blood vessel region model. It is a figure which shows the example output image of a vector concentration degree filter. It is a figure which shows the filter process image before a threshold value process. It is a figure which shows the filter process image after a threshold value process. It is a figure explaining the calculation method of sphericity. It is a figure explaining the identification method by a rule base method. It is a figure explaining the identification method by a rule base method. It is a figure which shows the example of an output of the detection result of a cerebral aneurysm candidate. It is a flowchart explaining the vascular region discrimination | determination process performed by the medical image processing apparatus. It is a figure which shows the example of a reference image. It is a figure which shows the original image used for preparation of a reference image. It is a figure which shows the histogram of a target image before and after performing a target image and a normalization process. It is a figure which shows the histogram of the target image before and behind performing the target image different from FIG. 13A, and a normalization process. It is a figure which shows a target image. It is a figure which shows the blood-vessel extraction image which extracted the blood vessel from the target image of FIG. 14A. It is a figure which shows the blood vessel extraction image and a reference image. It is the figure which superimposed the blood vessel extraction image and reference image which are shown to FIG. 15A. It is a figure which shows the landmark in a reference image. It is a figure which shows the corresponding point in a blood vessel extraction image. It is a figure which shows the discrimination result of the blood vessel extraction image and its blood vessel part. It is a figure which shows the discrimination result of the blood vessel extraction image and its blood vessel part. It is a figure which shows the example of an identification display of each blood-vessel site | part discriminate | determined in the target image. It is a flowchart which shows the detection process which concerns on 2nd Embodiment. It is a figure which shows the analysis bank of GC filter bank. It is a figure which shows filter bank A ( ^zj ). It is a figure which shows the reconstruction bank of GC filter bank. It is a figure which shows filter bank S ( ^zj ).

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Medical image processing apparatus 11 Control part 12 Operation part 13 Display part 14 Communication part 15 Storage part 16 Lesion candidate detection part

Claims (8)

Analyzing means for filtering the head MRA image, which is a three-dimensional image, performing multi-resolution analysis, and calculating a vector concentration level for each resolution level using partial images decomposed for each resolution level;
In reconstructing the original head MRA image from the partial image, the weighted image of the partial image is created by weighting the region where the calculated vector concentration degree is a predetermined value, and the partial image is multiplied by the weighted image Reconstruction means for reproducing the original image from the image;
Using the reconstructed head MRA image, a deletion unit that deletes a false positive candidate region that is a normal blood vessel from candidate regions of a lesion portion reproduced in the head MRA image;
A medical image processing apparatus comprising: The analysis unit performs multi-resolution analysis by performing filter processing using filters H (z) and G (z) represented by the following formulas 11 and 12 on the head MRA image. Item 2. The medical image processing apparatus according to Item 1.
Note that z in the above formula represents z conversion.   The analysis unit performs filtering on the x, y, and z directions constituting the three-dimensional shape of the head MRA image to obtain partial images in the x, y, and z directions, respectively. The medical image processing apparatus according to 2. The reconstruction means filters the partial images obtained by multiplying the weighted image by filters L (z), K (z), H (z) L (z) represented by the following expressions 11, 13, and 14: The medical image processing apparatus according to claim 2, wherein an original image is reproduced by performing processing.
Note that z in the above formula represents z conversion.   The deletion unit calculates a feature amount for a candidate region of a lesion portion reproduced in the reconstructed head MRA image, and deletes a false positive candidate region from the candidate region based on the feature amount. The medical image processing apparatus according to any one of claims 1 to 4.   The feature amount is at least one of a size of a candidate area of a lesion part reproduced in the reconstructed head MRA image, a sphericity, an average of vector concentration in the candidate area, or a maximum value of vector concentration. The medical image processing apparatus according to claim 5, wherein the medical image processing apparatus is one or more. The lesion is an aneurysm that occurs in a blood vessel,
An extraction means for extracting a blood vessel image from the head MRA image;
The medical image processing according to any one of claims 1 to 6, wherein the reconstruction unit performs multi-resolution analysis and reconstruction of an original image using the image from which the blood vessel image has been extracted. apparatus. Image control means for discriminating one or a plurality of blood vessel sites included in the blood vessel image of the head MRA image;
Display means for MIP display of the head MRA image;
Operation means for performing display operation;
When a blood vessel part to be displayed in the head MRA image is selected and operated via the operation means, only a blood vessel image corresponding to the selected blood vessel part is extracted from the head MRA image, and the extracted Display control means for displaying only a blood vessel image on the display means;
The medical image processing apparatus according to any one of claims 1 to 7, characterized in that it comprises a.