JP2006325638A - Method of detecting abnormal shadow candidate and medical image processing system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a detecting method which facilitates collection of learning data used as the base for detecting abnormal shadow candidates and is highly versatile in detection. <P>SOLUTION: In a medical image processing system 100, an image processor 2 scans marked regions on the mammogram generated by an image generator 1, and various characteristic amounts are extracted. Then, every time the characteristic amount is extracted in each marked region, a Mahalanobis distance from an average in distributions of characteristic amounts of normal tissues to the characteristic amounts extracted in the marked regions is calculated. If the Mahalanobis distance is greater than a threshold value, the characteristic of the marked region is determined to be dissimilar to the characteristics of the normal tissues. The system detects that the marked region constitutes an abnormal shadow candidate region. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a method for detecting an abnormal shadow candidate and a medical image processing system for analyzing a medical image and detecting a candidate region for an abnormal shadow.

  In the medical field, a medical image obtained by imaging a patient with an imaging device such as CT (Computed Tomography) or MRI (Magnetic Resonance Imaging) is converted into digital data and displayed on a display when the doctor makes a diagnosis. I have come to do interpretation. In particular, in recent years, computer-aided diagnosis (hereinafter referred to as CAD) that detects candidates for abnormal shadows such as cancerous parts for the purpose of reducing the burden on interpreting physicians and reducing oversight of abnormal shadows (referred to as image parts of lesions). Has been developed.

In the above-mentioned CAD, abnormal shadow candidates are detected by converting the features of abnormal shadows into features using a technique such as image processing (see, for example, Patent Document 1). In other words, an area with features similar to an abnormal shadow was detected as an abnormal shadow candidate area using the feature value of the abnormal shadow as learning data, but the collection of abnormal shadow cases depends on the incidence of the lesion. It takes a long time and cooperation with many hospitals is required. For this reason, it takes a lot of time and cost to collect information on the characteristics of abnormal shadows, that is, learning data.
JP 2002-112985 A

  In addition, for the reasons described above, from the previously limited abnormal shadow information, pre-processing, abnormal shadow candidate detection processing, false positive candidates (candidates in which normal tissue or the like is erroneously detected among the detected candidates) And the like, and by performing a plurality of steps and classifying the feature quantity at each step, the detection accuracy of abnormal shadow candidates has been improved. However, it has been a problem to construct a highly versatile algorithm that can simplify the feature quantity that needs to be classified in each step and can deal with various lesion types as much as possible.

  The subject of this invention is providing the detection method with high versatility of a detection while making easy the collection of the learning data used as the foundation which detects an abnormal shadow candidate.

The invention according to claim 1 is a method of detecting an abnormal shadow candidate.
Storing the feature amount extracted from the medical image of the normal tissue in the storage means;
Extracting a feature amount from a medical image to be detected as an abnormal shadow candidate, and comparing the feature amount of the medical image to be detected with the feature amount of the stored normal tissue;
Detecting an image region having a feature dissimilar to the normal tissue feature based on the result of the comparison as a candidate region of an abnormal shadow;
It is characterized by including.

The invention according to claim 2 is the method of detecting an abnormal shadow candidate according to claim 1,
The method includes a step of displaying detection information indicating the detected abnormal shadow candidate region on a display means.

The invention according to claim 3 is the method of detecting an abnormal shadow candidate according to claim 1 or 2,
Including a step of extracting a feature amount from a medical image of the designated normal tissue when the medical image including only the normal tissue is designated by the doctor via the operation unit, and causing the storage unit to additionally update. To do.

Invention of Claim 4 is the detection method of the abnormal shadow candidate as described in any one of Claims 1-3,
Recognizing a subject area from the medical image, and classifying the subject area into a plurality of areas,
The extracted feature amount includes biometric information related to the subject of the medical image, photographing information related to photographing, image feature amount obtained by analyzing the medical image, position information on the image from which the image feature amount is extracted, or its It includes the classification region information at the extraction position.

The invention according to claim 5 is the method for detecting an abnormal shadow candidate according to claim 4,
The image feature amount is extracted for each frequency band obtained by frequency-resolving the medical image, and is stored in the storage unit.

The invention according to claim 6 is the method of detecting an abnormal shadow candidate according to claim 5,
The image feature amount extracted for each frequency band is texture information.

The invention according to claim 7 is the method for detecting an abnormal shadow candidate according to any one of claims 4 to 6,
The medical image is scanned with a region of interest, and the image feature amount is extracted for each region of interest.

The invention according to claim 8 is the method of detecting an abnormal shadow candidate according to claim 7,
A feature is that a subject region in a medical image is extracted, a region of interest is scanned with respect to the subject region, and an image feature amount is extracted.

The invention according to claim 9 is the method of detecting an abnormal shadow candidate according to any one of claims 6 to 8,
In the distribution of the feature amount of the normal tissue, the distance from the average feature amount of the normal tissue to the feature amount of the attention area is calculated, and the attention area where the distance is a predetermined value or more is not similar to the normal tissue. An abnormal shadow candidate region having a feature is detected.

The invention according to claim 10 is the method of detecting an abnormal shadow candidate according to claim 9,
The distance is a Mahalanobis distance,
An attention area in which the Mahalanobis distance is a predetermined value or more is detected as a candidate area for an abnormal shadow.

The invention according to claim 11 is the method of detecting an abnormal shadow candidate according to any one of claims 4 to 10,
Extracting the feature amount of the normal tissue for each classification region and storing it in the storage means,
The feature amount extracted from the detection target medical image is compared with the stored normal tissue feature amount for each classification region from which the feature amount is extracted.

Invention of Claim 12 in the detection method of the abnormal shadow candidate as described in any one of Claims 1-11,
The medical image is a breast image.

The invention according to claim 13 is the method of detecting an abnormal shadow candidate according to claim 12,
The subject area is a breast area in a breast image.

The invention according to claim 14 is the method of detecting an abnormal shadow candidate according to claim 13,
The classification regions are a breast region and a pectoral muscle region.

The invention according to claim 15 is the method of detecting an abnormal shadow candidate according to claim 13,
The position information is any one of a distance from a region of interest to a skin line of a breast region, a distance to a pectoral muscle region, and a distance to a papilla.

The invention according to claim 16 is the method of detecting an abnormal shadow candidate according to any one of claims 1 to 16,
When the feature amount extracted from the medical image to be detected has a difference greater than or equal to a predetermined value with respect to the average feature amount of the normal tissue, the abnormal shadow candidate detection process is stopped.

The invention according to claim 17 is the method of detecting an abnormal shadow candidate according to claim 16,
The average and variance of the feature amounts of the normal tissue are calculated, and the detection of the abnormal shadow candidate is stopped when the feature amount extracted from the medical image to be detected is a predetermined variance value or more.

The invention according to claim 18 is the medical image processing system,
Storage means for storing a feature amount extracted from a medical image of normal tissue;
Extracting a feature amount from a medical image to be detected as an abnormal shadow candidate, comparing the feature amount of the medical image to be detected with the feature amount of the stored normal tissue, and based on the result of the comparison, An abnormal shadow candidate detection means for detecting an image region having a feature dissimilar to the normal tissue feature as an abnormal shadow candidate region;
It is characterized by providing.

  According to the inventions of claims 1 and 18, since only the feature amount of normal tissue needs to be prepared as learning data for abnormal shadow candidate detection processing, learning data can be easily collected and various lesions can be collected. Species can also be handled, and versatility in detection processing is improved.

  According to invention of Claim 2, detection information can be provided as reference information at the time of a doctor's interpretation.

  According to the invention described in claim 3, since the doctor can additionally update the feature amount of the normal tissue, the learning ability can be enhanced by use and the detection accuracy of the abnormal shadow candidate can be improved.

  According to the fourth aspect of the present invention, it is possible to determine the similarity with the feature of the normal tissue based on various feature amounts, and to improve the determination accuracy.

  According to the fifth and sixth aspects of the present invention, the feature amount of the texture information can be compared for each frequency band having the same feature, and the accuracy is improved in determining the similarity with the feature of the normal tissue. Can be made.

  According to the eighth, ninth, and tenth aspects of the present invention, it is possible to determine the similarity to the characteristics of the normal tissue in consideration of the variance in the distribution of the characteristic amounts of the normal tissue.

  According to the eleventh aspect of the present invention, it is possible to improve the accuracy of determining similarity to the features of normal tissue by classifying the regions having the same characteristics and comparing the feature amounts for each of the classified regions. it can.

  According to the invention described in claims 12, 13, 14, and 15, it is possible to detect a candidate region for an abnormal shadow based on a feature amount in a breast region of a breast image.

  According to the inventions of claims 16 and 17, the detection is stopped when it has dissimilar characteristics apart from the characteristics of the normal tissue and it is impossible to judge whether the shadow of the normal tissue is abnormal or abnormal. Can do.

  In the present embodiment, an example using a breast image obtained by photographing a breast as a medical image will be described.

First, the configuration will be described.
FIG. 1 shows a system configuration of a medical image processing system 100 in the present embodiment.
The medical image processing system 100 is a system that takes a medical image, performs image processing on the medical image, detects an abnormal shadow candidate, and provides the doctor with the detection information together with the medical image.

  As illustrated in FIG. 1, the medical image processing system 100 includes an image generation device 1, an image processing device 2, a printer 3, an image server 4, and a viewer 5. These devices 1 to 5 are connected to each other through a communication network N constructed in a medical institution such as a LAN (Local Area Network) so that data can be transmitted and received between them. The DICOM (Digital Imaging and Communication in Medicine) standard is applied to the communication network N.

Hereinafter, each component apparatus 1-5 is demonstrated.
The image generation apparatus 1 captures a human body and generates digital data of the captured image (medical image). For example, CR (Computed Radiography), FPD (Flat Panel Detector), CT, MRI, cassette-specific reading Modality such as an apparatus or a film digitizer can be applied. In the present embodiment, it is assumed that breast image data is generated by applying a breast-specific CR that performs X-ray imaging of the left and right breasts as the image generation apparatus 1.

  Note that the image generation device 1 is a device that complies with the DICOM standard described above, and includes various information attached to the generated medical image (for example, patient information regarding a patient whose medical image is captured, imaging information regarding imaging or examination, Inspection information, etc.) can be input from the outside and can be automatically generated. The image generation apparatus 1 adds the supplementary information as header information to the generated medical image and transmits it to the image processing apparatus 2 via the communication network N. In addition, when not conforming to the DICOM standard, incidental information can be input to the image generating apparatus 1 using a DICOM conversion apparatus (not shown).

  The image processing apparatus 2 performs various types of image processing on the medical image supplied from the image generation apparatus 1 and performs an image analysis of the medical image to detect an abnormal shadow candidate.

  The printer 3 outputs a medical image to a recording medium such as a film based on the medical image data received from the image processing apparatus 2 or the image server 4.

  The image server 4 includes an image DB (Data Base) 4a. In this image DB 4a, a medical image (original image) generated by the image generation apparatus 1 and an image-processed medical image received from the image processing apparatus 2 ( Process image) and manage its input and output.

  The viewer 5 is a terminal device used by a doctor for diagnosis, and includes a display unit such as an LCD (Liquid Crystal Display), an operation unit such as a keyboard and a mouse, and the like. The viewer 5 acquires detection information of a specified medical image or abnormal shadow candidate from the image server 4 and displays it in accordance with a doctor's operation instruction.

Next, the image processing apparatus 2 according to the present invention will be described in detail.
FIG. 2 shows an internal configuration of the image processing apparatus 2.
The image processing apparatus 2 includes a control unit 21, an operation unit 22, a display unit 23, a communication unit 24, a storage unit 25, an image processing unit 26, and an abnormal shadow candidate detection unit 27.

  The control unit 21 includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and the like. The CPU reads various control programs from the storage unit 25 and develops them in the RAM. Then, the execution of processing is comprehensively controlled according to the program, and the operation of each unit is centrally controlled. For example, registration processing and abnormal shadow candidate detection processing (each processing content will be described later) are executed in accordance with the registration processing program and abnormal shadow candidate detection processing program according to the present invention.

  The operation unit 22 includes a keyboard, a mouse, and the like, generates an operation signal corresponding to a key operation and a mouse operation, and outputs the operation signal to the control unit 21.

  The display unit 23 includes an LCD or the like, and performs various displays such as an operation screen at the time of image processing and a medical image after processing in accordance with instructions from the control unit 21.

  The communication unit 24 includes a communication interface such as a router or a modem, and communicates with an external device on the communication network N according to an instruction from the control unit 21. For example, a medical image to be processed is received from the image generation apparatus 1 or a processed medical image is transmitted to the image server 4 or the printer 3.

The storage unit 25 stores various control programs, parameters necessary for execution of the programs, or processing result data.
The storage unit 25 is a storage unit that stores a feature amount DB 251 and a learning DB 252.

As shown in FIG. 3, the feature amount DB 251 is a database constructed from feature amount data extracted from a breast image. The feature amount is extracted for each region of interest by the abnormal shadow candidate detection unit 27 scanning a region of interest of a predetermined size with respect to the medical image. Therefore, the feature amount DB 251 stores the position information of the attention area, the information of the area where the attention area is located, the biological information of the medical image, the photographing information, the image feature amount (texture information, contrast, etc.), and the like. Details of each feature amount will be described together with an extraction process in the abnormal shadow candidate detection unit 27 described later.
The feature amount DB 251 is created separately for feature amounts extracted from breast images including only normal tissues (hereinafter referred to as normal images) and feature amounts extracted from breast images to be detected as abnormal shadow candidates. , And stored in the storage unit 25.

The learning DB 252 is a database that stores feature amounts of normal tissue groups, and stores feature amounts extracted from normal images as learning data, as shown in FIG. Specifically, the mean μ, the variance σ, and the variance-covariance matrix Σ− ¹ , which are the feature amounts of the normal tissue group, calculated by the abnormal shadow candidate detection unit 27 are stored. These pieces of information are calculated and stored for each of the areas (Da to Dc) obtained by further classifying the breast area where the abnormal shadow exists in the breast area into a plurality of areas based on the difference in density. The average μ, variance σ, and variance covariance matrix Σ ⁻¹ will be described later together with the calculation method.

  The image processing unit 26 performs gradation conversion processing, sharpness adjustment processing, and the like on the medical image according to the image processing program. In the case of a breast image, the image processing unit 26 performs alignment processing for combining the left and right breast images side by side. Execute.

  The abnormal shadow candidate detection unit 27 includes a CPU and the like, and in cooperation with a processing program stored in the storage unit 25, a feature amount is obtained from a medical image designated and operated by a doctor as a medical image including only normal tissue. The learning DB 252 is constructed by extracting and updating the storage unit 25 as learning data. Also, an abnormal shadow candidate that extracts a feature amount from a medical image to be detected, compares the feature amount with a feature amount of a normal tissue stored in the learning DB 252, and detects an abnormal shadow candidate based on the comparison result. It is a detection means. At this time, as a pre-process for detecting an abnormal shadow candidate, image analysis of a medical image is performed to recognize a subject area, and the subject area is classified into a plurality of areas.

Next, the operation of the medical image processing system 100 will be described.
First, the overall flow from the generation of a medical image to the interpretation of the medical image by a doctor will be described.
In the medical image processing system 100, first, when an image is taken by the image generation apparatus 1, a medical image (breast image) is generated. As related information of the generated breast image, patient information (patient name, patient ID, age, weight, height, etc.), imaging information (imaging direction, imaging site, imaging conditions, imaging method, etc.), examination information ( (Examination date, name of doctor in charge, modality used, etc.) are attached to the breast image. The breast image (original image) with the accompanying information is output from the image generation apparatus 1 to the image server 4 and is output to the image processing apparatus 2 as a breast image to be detected as an abnormal shadow candidate.

  In the image processing apparatus 2, image processing for interpretation is performed on the breast image, and abnormal shadow candidate detection processing is performed (the detection processing will be described later), and the processed image is an image together with detection information of the abnormal shadow candidate. Output to the server 4. In the image server 4, the processed image and the detection information are stored in the image DB 4a in association with the original image stored previously. Then, in response to a request from the viewer 5, the processed image and detection information stored in the image DB 4 a are output to the viewer 5. In the viewer 5, as shown in FIG. 5, the processed image is displayed on the display. At this time, if an abnormal shadow candidate is detected, detection information indicating the detection position (an arrow marker shown in FIG. 5) is displayed. With this display, the doctor interprets the processed image and diagnoses an abnormal shadow using the detection information as reference information.

  At this time, as a result of the interpretation by the doctor, the feature amount of the normal image diagnosed as having only the normal tissue and not including the abnormal shadow is registered in the learning DB 252 of the image processing apparatus 2, thereby performing subsequent image processing. This can be reflected in the abnormal shadow candidate detection process in the apparatus 2. When performing registration in the learning DB 252 as described above, the doctor presses the registration button d11 displayed on the interpretation screen d1. By pressing the registration d11 button, the viewer 5 transmits registration instruction information of the breast image displayed on the interpretation screen d1 to the image server 4, and the image server 4 responds to the registration instruction information with a normal image and Instruction information for registering the breast image is transmitted to the image processing apparatus 2 together with the designated breast image (original image). In the image processing apparatus 2, registration processing in the learning DB 252 is performed using the breast image received together with the instruction information.

Hereinafter, the registration process in the image processing apparatus 2 will be described with reference to FIG.
In the registration process shown in FIG. 6, when the control unit 21 first determines that the input breast image is an image designated as a normal image (step S <b>1; Y), the abnormal shadow candidate detection unit 27 performs the normal image processing. Biological information and imaging information are extracted as feature quantities of normal images from the incidental information (patient information, imaging information, examination information, etc.) of the image (step S2). The biological information is information related to the body of the subject (patient) of the normal image, and for example, the age, sex, height, weight, blood pressure, obesity, etc. of the patient are extracted from the accompanying information. The shooting information is information related to shooting, and information such as shooting direction, shooting conditions (tube voltage, mAs value, tube, additional filter, etc.), shooting method, and the like are extracted from the supplementary information.

Next, the abnormal shadow candidate detection unit 27 performs image analysis of the normal image, extracts the subject area, and classifies the subject area into a plurality of areas (step S2).
Hereinafter, a procedure for extracting each region from the medical image will be described with reference to FIG.
FIG. 7 is a diagram showing a breast image S taken in an oblique direction (hereinafter referred to as MLO). A subject area Sa is extracted from the breast image S, and a breast area Sa1 and a pectoral muscle area Sa2 are extracted from the subject area Sa. And the breast region Sa1 is classified into three regions Da, Db, and Dc.

<1> First, a variance histogram of pixel values in the breast image S is obtained, and a discriminant analysis method (when the variance histogram is divided into two classes, the discrimination ratio (dispersion ratio) between intra-class variance and inter-class variance in two classes is obtained). The threshold is determined using a method for determining the threshold so as to be maximized, and the breast image S is binarized using the determined threshold.

  At this time, in the breast image S, an unexposed region (a region that does not correspond to the subject portion and is a region directly irradiated with X-rays) exhibits a high density and thus becomes “1” by binarization. The subject area is expected to be “0”. Therefore, the binarization can divide the breast image S into the subject area Sa and the other missing areas Sb. Also, the boundary between the two regions (the subject region and the missing region) by binarization is recognized as the skin line SL.

<2> On the other hand, when the shooting direction is MLO, the pectoralis muscles lie in the subject portion, so that the pectoral muscle region Sa2 is recognized next from the subject region Sa. When the photographing direction is the front direction (hereinafter referred to as CC), the pectoral muscle is not reflected in the subject portion, and the subject region Sa becomes the breast region Sa1. For the pectoral muscle region Sa2, for example, a density gradient in the subject region Sa is calculated, and the subject region Sa is classified into a pectoral muscle region Sa2 and a breast region Sa1 based on the density gradient. As described in Japanese Patent Application Laid-Open No. 2001-238868, a local region is set, a threshold is set based on the pixel value in the local region, and the subject region Sa is binarized, whereby the pectoral muscles The region Sa2 and the breast region Sa1 may be recognized.

<3> Next, the breast region Sa1 is classified into three regions Da, Db, and Dc.
The breast region is a mixture of mammary gland and fat, and the concentration varies depending on the density. Therefore, when a doctor makes a diagnosis, the breast region is classified based on the concentration. Classification is high-concentration region with a high fat content and mammary gland content of less than 10%, medium-concentration region with mammary gland content of 10% or more and less than 50%, and mammary gland content of 50% or more In many cases, it is classified into three regions, which are considered to be low concentration regions. Abnormal shadows appear white and at low density on the image, so it is difficult to visually detect abnormal shadows when there is a lesion in the low-density area corresponding to the whitest density of the pectoralis major muscle or in a slightly whitish medium-density area. Become.

Therefore, the breast region Sa1 is classified into the high concentration region Da, the intermediate concentration region Db, and the low concentration region Dc with reference to the density of the pectoral muscle region Sa2. Specifically, a dispersion histogram of the pectoral muscle region Sa2 is created, a region having a relatively uniform density is extracted from the shape, and the average density is used as a threshold value and classified into each region Da to Dc. For example, the remaining area from which the area having a higher density than the threshold is deleted is denoted by Dc, and the area having a density of 30% or higher of the threshold is Db, and the area having a high density by 60% or higher is Da. By doing, it can classify | categorize into each area | region Da-Dc.

When each region is extracted from the normal image in this way, the attention region F is sequentially set in the extracted breast region Sa1, and the image feature amount is calculated by the abnormal shadow candidate detection unit 27 for each attention region F. (Step S4).
More specifically, with reference to FIG. 8, the quadrangular attention area F is set to scan on the breast area Sa1. The size of the region of interest F may be set according to the abnormal shadow to be detected, for example, in the case of detecting a microcalcification cluster shadow that is a kind of breast cancer, for example, a square shape of 5 mm square. A plurality of sizes may be set. For example, when detecting a shadow of a tumor that is a kind of breast cancer, an image feature amount is calculated using, for example, a 1 cm square area and a 3 cm square area. May be.

  As the image feature amount, the average pixel value (density value) of each pixel value in the attention area F, contrast (difference between the average pixel value in the attention area F and the pixel value in the peripheral area), the standard deviation of the pixel value, and the fractal In addition to various feature quantities such as dimensions and curvature, texture information after frequency decomposition and the like are calculated.

The texture information is an image feature amount indicating a feature of the texture (structure) such as a straight line or a dot constituting the image. The texture information is calculated for each decomposed frequency band after frequency decomposition of the breast image by wavelet transformation defined by the following equation (1), and stored in the feature amount DB 251.

  The texture information calculation method includes a method using a density co-occurrence matrix and a method using a density difference matrix. Here, an example of calculating a feature amount obtained from the density co-occurrence matrix will be described. The density co-occurrence matrix is a pixel pair (x1, x2), (x2, y2) having a specific relative positional relationship in the image f (x, y), and the density pair is (i, j). I.e., f (x1, x2) = i and f (x1, x2) = j.

Using the density co-occurrence matrix, each feature quantity of energy, entropy, correlation, local uniformity, and inertia is calculated from the following equations (2) to (6).

  Further, when calculating each image feature amount, position information of the attention area F from which the image feature amount is calculated is calculated. The position information may be any of the distance from the center point of the attention area F to the skin line SL (indicated by the number of pixels; the same applies hereinafter), the distance to the pectoral muscle area Sa2, and the distance to the nipple. Furthermore, it is determined on which area of the areas Da to Dc the attention area F is classified in the breast area Sa1 from the calculated position information.

When the calculation of all the feature values is completed, the information on the feature values is stored in the normal image feature value DB 251 (step S5). Then, for each of the classification regions Da to Dc in which the attention region F exists, the average μ, variance σ, and variance covariance matrix Σ ^{−1 of} the normal shadow group using these feature values as variables are calculated, and each of the classification regions Da to Dc is calculated. Is stored in the learning DB 252 as learning data of the normal shadow group (step S6).

This will be specifically described. Assuming that i feature values are calculated in step S4, a population (normal shadow group) is formed by elements having the feature values extracted from the normal image as variables, and their average vertical vector (hereinafter simply referred to as average). Μ is expressed by the following equation (7). Further, a sample value vertical vector x of an element (hereinafter referred to as a sample) consisting of a feature quantity to be detected is expressed by the following equation (8).

The variance-covariance matrix Σ ⁻¹ is a parameter necessary for calculating the Mahalanobis distance between the average (center of gravity) of the normal shadow group and the sample to be detected in the subsequent abnormal shadow candidate detection process. This is the inverse of the group variance / covariance matrix. From this variance-covariance matrix Σ ⁻¹ , the Mahalanobis distance can be calculated by the following equation (9).

If the correlation matrix is R, it can also be calculated by the following equation (10).

  As described above, every time a normal image is registered, extraction of the feature amount, calculation of the normal shadow group average μ and the like are executed, and the normal image learning DB 252 is updated.

Next, detection processing for detecting a candidate region for an abnormal shadow using the learning DB 252 will be described.
FIG. 9 is a flowchart for explaining the flow of the abnormal shadow candidate detection process executed in the image processing apparatus 2.
In the detection process shown in FIG. 9, first, when a breast image to be detected (hereinafter referred to as a target image) is input, the abnormal shadow candidate detection unit 27 uses the incidental information of the target image to capture the subject's subject, Biometric information and imaging information are extracted as feature quantities related to (step T1). The extracted feature quantity is stored in the feature quantity DB 251 for the target image.

  Next, as shown in FIG. 7, the subject region is extracted from the target image, and when the pectoral muscle region and the breast region are further extracted from the subject region, the breast region is classified into a plurality of regions Da to Dc. (Step T2). Next, an attention area F is set for the breast area, and an image feature amount in the attention area F and position information of the attention area F are calculated. Further, from the position of the attention area F, it is determined which of the classification areas Da to Dc is the area where the attention area F exists (step T3). Since the contents of the processing steps so far are the same as those in steps S2 to S4 in the registration processing described above, description thereof is omitted.

  Next, the average μ and the variance σ of the normal shadow group corresponding to the determined classification areas Da to Dc are read from the learning DB 252. Then, it is determined whether or not the distribution position of the sample of the feature amount extracted from the attention area F is equal to or larger than a predetermined dispersion value such as 3σ or more with respect to the average μ of the normal shadow group (step T4). . If the distribution position is 3σ or more (step T4; Y), it is determined that neither normal shadow nor abnormal shadow can be determined because it is dissimilar from the characteristics of the normal shadow group (step T5). Thereafter, the process proceeds to step T10.

On the other hand, if the distribution position is within 3σ (step T4; N), the Mahanobis distance between the sample x having the feature quantity extracted from the target image as a variable and the average μ of the normal shadow group is the above-described formula (9). (Or Equation (10)) (Step T6). When calculating the Mahalanobis distance, the mean μ and the variance-covariance matrix Σ ⁻¹ corresponding to the areas Da to Dc of the attention area F are read from the learning DB 252 and used for the calculation. Then, the Mahalanobis distance is compared with a threshold value prepared in advance, and it is determined whether the sample is similar or dissimilar to the normal shadow feature (step T7).

  If the computed Mahalanobis distance is greater than or equal to the threshold, that is, the sample is outside the normal shadow group distribution, and it is determined that it has features that are not similar to normal tissue features, that is, abnormal shadow features (Step T7; Y), it is determined that the attention area F for which the Mahalanobis distance is calculated includes an abnormal shadow (Step T8). On the other hand, when the Mahalanobis distance is less than the threshold value, that is, when it is determined that the sample is in the distribution of the normal shadow group and has the same or similar feature as that of the normal tissue (step T7; N) It is determined that the attention area F does not include an abnormal shadow (step T9).

  When the determination of whether or not an abnormal shadow is included for one region of interest F is completed, it is determined whether or not the region of interest F has been scanned for all breast regions, that is, whether the entire region of the breast region is abnormal or normal. It is determined whether or not the operation has been performed (step T10). If not yet completed (step T10; N), the attention area F is set at the next adjacent position, and the image feature amount and position information in the attention area F are calculated in the same manner as the processing of step T3 ( The process proceeds to step T11) and step T4.

  When the determination is completed for the entire breast region (step T10; Y), detection information is generated with the attention region F determined to contain an abnormal shadow as a candidate region for the abnormal shadow (step T12). This detection information is output to the image server 4 together with the target image, and is used for interpretation by a doctor in the viewer 5 as shown in FIG.

  As described above, according to the present embodiment, the learning data used for the abnormal shadow candidate detection process is only normal tissue, and the feature is located outside the population distribution of the normal tissue learning data population. An image region having a quantity, that is, an image region having a feature dissimilar to that of a normal tissue is detected as an abnormal shadow candidate region. As a result, learning data can be easily collected and various types of lesions can be dealt with, and versatility in detection processing is improved.

  In addition, it has a configuration in which normal tissue learning data can be additionally updated from a normal image that the doctor has determined that there is no abnormal shadow during interpretation, thus improving the learning ability and improving the detection accuracy of abnormal shadow candidates. Can be made. Furthermore, since a large amount of learning data can be obtained from a single normal image, learning efficiency can be improved and the versatility of detection processing can be further improved.

  Further, the breast region Sa1 in which the abnormal shadow exists is classified into regions Da to Dc having different image features, and the image feature amount is calculated when the learning data is stored or the abnormal shadow candidate is detected. Since the image is stored for each of the classification areas Da to Dc, the image feature amounts of the normal tissue and the attention area F can be compared for each of the classification areas Da to Dc having similar features, and the detection accuracy can be improved.

  Further, at the time of comparison, the average (centroid) of the learning data group of the normal tissue and the Mahalanobis distance of the image feature amount of the attention area F are calculated, and whether or not the attention area F is dissimilar from the normal tissue feature based on this Mahalanobis distance. That is, it is determined whether or not it is an abnormal shadow candidate. Therefore, it is possible to determine the difference in characteristics from the attention area F while considering the dispersion of the learning data group of the normal tissue, and the detection accuracy can be improved.

  In the above description, the image processing apparatus 2 includes the storage unit and the abnormal shadow candidate detection unit. However, the present invention is not limited to this, and any unit of the medical image processing system 100 can realize each unit. Also good.

1 is a diagram showing a system configuration of a medical image processing system 100 in the present embodiment. 2 is a diagram illustrating an internal configuration of an image processing apparatus 2. FIG. It is a figure which shows the data structural example of feature-value DB. It is a figure which shows the data structural example of learning DB. It is a figure which shows the example of an interpretation screen which can update and register the learning data of a normal organization. It is a flowchart explaining the flow of the registration process performed by an image processing apparatus. It is a figure showing what classified a subject field in a breast image into a plurality of fields. It is a figure explaining the setting method of the attention area which is a processing unit which calculates an image feature-value. It is a flowchart explaining the detection process of the abnormal shadow candidate performed with an image processing apparatus.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Medical image processing system 1 Image generation apparatus 2 Image processing apparatus 21 Control part 25 Storage part 251 Feature-value DB
252 Learning DB
27 Abnormal Shadow Candidate Detection Unit 3 Printer 4 Image Server 4a Image DB
5 Viewer

Claims (18)

Storing the feature amount extracted from the medical image of the normal tissue in the storage means;
Extracting a feature amount from a medical image to be detected as an abnormal shadow candidate, and comparing the feature amount of the medical image to be detected with the feature amount of the stored normal tissue;
Detecting an image region having a feature dissimilar to the normal tissue feature based on the result of the comparison as a candidate region of an abnormal shadow;
A method of detecting an abnormal shadow candidate characterized by comprising:   The method for detecting an abnormal shadow candidate according to claim 1, further comprising a step of displaying detection information indicating the detected abnormal shadow candidate area on a display means.   Including a step of extracting a feature amount from a medical image of the designated normal tissue when the medical image including only the normal tissue is designated by the doctor via the operation unit, and causing the storage unit to additionally update. The method for detecting an abnormal shadow candidate according to claim 1 or 2. Recognizing a subject area from the medical image, and classifying the subject area into a plurality of areas,
The extracted feature amount includes biometric information related to the subject of the medical image, photographing information related to photographing, image feature amount obtained by analyzing the medical image, position information on the image from which the image feature amount is extracted, or its The method for detecting an abnormal shadow candidate according to any one of claims 1 to 3, further comprising classification region information at an extraction position.   The abnormal shadow candidate detection method according to claim 4, wherein the image feature amount is extracted for each frequency band obtained by frequency-resolving the medical image and stored in the storage unit.   The method of detecting an abnormal shadow candidate according to claim 5, wherein the image feature amount extracted for each frequency band is texture information.   The abnormal shadow candidate detection method according to any one of claims 4 to 6, wherein a region of interest is scanned with respect to a medical image, and the image feature amount is extracted for each region of interest.   8. The method for detecting an abnormal shadow candidate according to claim 7, wherein a subject area in a medical image is extracted, a region of interest is scanned with respect to the subject area, and an image feature amount is extracted.   In the distribution of the feature amount of the normal tissue, the distance from the average feature amount of the normal tissue to the feature amount of the attention area is calculated, and the attention area where the distance is a predetermined value or more is not similar to the normal tissue. The method of detecting an abnormal shadow candidate according to any one of claims 6 to 8, wherein the abnormal shadow candidate area is detected as a characteristic abnormal shadow candidate region. The distance is a Mahalanobis distance,
The abnormal shadow candidate detection method according to claim 9, wherein an attention area in which the Mahalanobis distance is equal to or greater than a predetermined value is detected as an abnormal shadow candidate area. Extracting the feature amount of the normal tissue for each classification region and storing it in the storage means,
11. The feature amount extracted from the medical image to be detected is compared with the stored feature amount of normal tissue for each classification region from which the feature amount is extracted. The method for detecting an abnormal shadow candidate according to one item.   The method of detecting an abnormal shadow candidate according to any one of claims 1 to 11, wherein the medical image is a breast image.   13. The abnormal shadow candidate detection method according to claim 12, wherein the subject area is a breast area in a breast image.   14. The abnormal shadow candidate detection method according to claim 13, wherein the classification regions are a breast region and a pectoral muscle region.   14. The abnormal shadow candidate according to claim 13, wherein the position information is any one of a distance from a region of interest to a skin line of a breast region, a distance to a pectoral muscle region, and a distance to a nipple. Detection method.   The abnormal shadow candidate detection process is stopped when a feature amount extracted from a medical image to be detected has a difference of a predetermined value or more with respect to an average of the feature amounts of the normal tissue. The method for detecting an abnormal shadow candidate according to any one of the above.   The average and variance of feature amounts of the normal tissue are calculated, and detection of the abnormal shadow candidate is stopped when a feature amount extracted from the detection target medical image is equal to or greater than a predetermined variance value. Item 18. A method for detecting an abnormal shadow candidate according to Item 16. Storage means for storing a feature amount extracted from a medical image of normal tissue;
Extracting a feature amount from a medical image to be detected as an abnormal shadow candidate, comparing the feature amount of the medical image to be detected with the feature amount of the stored normal tissue, and based on the result of the comparison, An abnormal shadow candidate detection means for detecting an image region having a feature dissimilar to the normal tissue feature as an abnormal shadow candidate region;
A medical image processing system comprising:

Images