JP2008036068A - Osteoporosis diagnosis support apparatus, method and program, computer-readable recording medium recorded with osteoporosis diagnosis support program and lsi for supporting osteoporosis diagnosis - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve the determination accuracy of an osteoporosis screening using dental panoramic radiographs. <P>SOLUTION: A threshold for tentatively determining "a low bone density" or "a normal state" is determined for every characteristic value by a threshold processing based on the fuzzy logic. Three fuzzy sets are set based on the thresholds. The connection weight of a three-layer neural network where the product of membership values to the respective fuzzy sets is defined as an input and "the low bone density" or "the normal state" is defined as an output is determined. When the characteristic value of a certain undetermined image is given, the product of the membership values to the fuzzy sets is input in the neural network to acquire the output of "the low bone density" or "the normal state". <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an osteoporosis screening technique using a dental panoramic radiograph.

  Osteoporosis, in which bone density decreases and bones become brittle and easily fracture, has become a problem in recent years. Along with the further aging of society in the future, it is expected that the patient's quality of life will decrease due to fractures caused by osteoporosis, the mortality will increase, and the national medical expenses will increase. However, most postmenopausal women, who make up the majority of osteoporosis patients, live a high risk of fracture without subjective symptoms, and it is important to find this disease as early as possible to prevent the onset of fractures .

  Diagnosis of osteoporosis is generally based on bone density measurement of the femur. However, there are few examples of people who have no subjective symptoms bothering to measure the bone mineral density of the femur. Therefore, it is the actual situation that osteoporosis is not found until the occurrence of a fracture or the like. On the other hand, it is considered that there are many opportunities for potential osteoporosis patients without subjective symptoms to visit dentistry for treatment of dental diseases. Today, in dentistry, teeth and surrounding alveolar bone are diagnosed by panoramic X-rays that provide an overview of the entire jaw, and are used to treat dental diseases. Previous studies have shown that mandibular cortical bone thickness measurement and morphological analysis observed in dental panoramic radiographs is extremely useful for screening postmenopausal osteoporosis patients. These studies show that if cortical bone thickness is below a certain threshold, bone density is low and there is a high suspicion of osteoporosis, otherwise bone density is normal and suspicion of osteoporosis is low. Has been. Further, it has been shown that when the cortical bone is roughened, the bone density is lower than that of normal cortical bone and the suspicion of osteoporosis is high. Since panoramic X-ray photography equipment is available at any dental clinic, if it is used for screening osteoporosis patients, it will help early detection and treatment of osteoporosis through cooperation between dentists and orthopedists, and is also useful in terms of cost It is. In addition, if a dental panoramic X-ray photograph can be used, a large amount of X-ray photographs can be easily collected, so that an epidemiological investigation on osteoporosis can be performed.

As an apparatus for supporting the diagnosis of osteoporosis using the above-mentioned dental panoramic radiograph, the degree of the roughening of the lower mandibular lower margin cortical bone on the panoramic radiograph is a skeleton (the mathematical morphological image processing method). 1) to identify and screen osteoporosis patients [Patent Document 1, Non-Patent Document 9], a computer system that can accurately and semi-automatically measure the thickness of the cortical bone of the lower mandible on panoramic radiographs [ Patent Document 2 and Non-Patent Document 6] exist. If these are used, osteoporosis can be accurately diagnosed by a simple method without requiring special skills or experience of an operator involved in osteoporosis diagnosis.
JP 2004-209089 A International Publication No. 2006/043523 Pamphlet Bone health and osteoporosis: a surgeon general's report. (2004) .Available: http://www.surgeongeneral.gov/reportspublications.html JA Kanis, “Requirement for DXA for the management of osteoporosisin Europe”. Osteoporosis International, vol. 16, pp. 229-238, 2005. A. Taguchi, M. Tsuda, M. Ohtsuka, I. Kodama, M. Sanada, T. Nakamoto, K. Inagaki, T. Noguchi, Y. Kudo, Y. Suei, K. Tanimoto, and AM Bollen, “Useof dental panoramic radiographs in identifying younger postmenopausal women with osteoporosis ”, Osteoporosis International vol. 17, no. 3, pp. 387-394,2006 K. Lee, A. Taguchi, K. Ishii, Y. Suei, M. Fujita, T. Nakamoto, M. Ohtsuka, M. Sanada, M. Tsuda, K. Ohama, K. Tanimoto, and SC White, “Visual estimationof mandibular cortex on panoramic radiographs in identifying postmenopausalwomen with low bone mineral businesses ”, Oral Surg Oral Med Oral Pathol Oral RadiolEndod vol. 100, pp. 226-231, 2005. A. Taguchi, Y. Suei, M. Sanada, M. Ohtsuka, T. Nakamoto, H. Sumida, K. Ohama, and K. Tanimoto, “Validation of dental panoramic radiography measuresfor identifying postmenopausal women with spinal osteoporosis”, AmericanJournal of Roentgenology, vol. 183, pp. 1755-1760, 2004. AZ Arifin, A. Asano, A. Taguchi, T. Nakamoto, M. Ohtsuka, and K. Tanimoto, “Computer-aided system for measuring the mandibular cortical width on panoramicradiographs in osteoporosis diagnosis”, in Proc. SPIE Med Imaging 2005- ImageProcessing Conference, San Diego, 2005, pp. 813-821 AZ Arifin, A. Asano, A. Taguchi, T. Nakamoto, M. Ohtsuka, M. Tsuda, Y.Kudo, and K. Tanimoto, “Computer-aided system for measuring the mandibularcortical width on panoramic radiographs in identifying postmenopausal womenwith low bone mineral density ”. Osteoporosis International, vol. 17, pp.753-759, 2006. A. Taguchi, M. Sanada, E. Krall, T. Nakamoto, M. Ohtsuka, Y. Suei, K. Tanimoto, I. Kodama, M. Tsuda, and K. Ohama, “Relationship between dentalpanoramic radiographic findings and biochemical markers of bone turnover ”, J. Bone Miner Res., vol. 18, pp. 1689-169, 2003. T. Nakamoto, A. Taguchi, A. Asano, M. Ohtsuka, Y. Suei, M. Fujita, M. Sanada, K. Ohama, and K. Tanimoto, "Computer-aided diagnosis oflow skeletal bone mass on panoramic radiographs, "presented at the82nd General Session & Exhibition of the International Association forDental Research no. 1953, Hawaii, 2004. LX Wang, “The WM method completed: a flexible fuzzy system approach to data mining”, IEEE Transaction on fuzzy system, vol. 11, no. 6, pp.768-780, 2003. S. Horikawa, T. Furuhashi, and Y. Uchikawa, "On fuzzy modelingusing fuzzy neural network with the back-propagation algorithm". IEEETransaction on Neural Network, vol. 3, no. 5, pp. 801-806, 1992. Y. Lin, GA Cunningham III, SV Coggeshall, "Using fuzzypartitions to create fuzzy systems from input-output data and set the initialweights in a fuzzy neural network", IEEE Transaction on Fuzzy Systems, vol. 5, no. 4, pp. 614-621, 1997. M. Sezgin and B. Sankur, “Survey over image thresholding techniques and quantitative performance evaluation”, Journal of Electronic Imaging, vol.13, no.1, pp. 146-165, 2004. LK Huang and MJJ Wang, “Image thresholding by minimizing the measures of fuzziness”, Patter Recognition, vo. 28, no. 1, pp. 41-51, 1995. AZ Arifin and A. Asano, "Image thresholding by measuring the fuzzy sets similarity," Proc. Information and Communication Technology Seminar 2005, pp. 189-194, 2005. AZ Arifin and A. Asano, "Image Segmentation by Histogram Thresholding Using Hierarchical Cluster Analysis," Pattern Recognition Letters, inpress, available online 3 May 2006.

  An object of the present invention is to improve the determination accuracy of osteoporosis screening using a dental panoramic radiograph.

An osteoporosis diagnosis support apparatus according to the present invention is an apparatus for determining osteoporosis using a feature amount related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. A threshold value determination unit that is determined by threshold processing based on fuzzy logic;
A fuzzy set setting unit that sets a plurality of fuzzy sets for each of the three feature amounts, and sets a membership function of the fuzzy set based on the threshold;
A neural network unit that inputs a product of membership values to each fuzzy set set for each of the three feature values, and outputs “low bone density” or “normal”;
A connection weight determining unit for determining a connection weight of the neural network unit using the learning data;
A feature amount acquisition unit for acquiring “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic X-ray photograph including a cortical bone portion;
An input unit that inputs a product of membership values to each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” acquired by the feature amount acquisition unit to the neural network unit;
It is characterized by that.

In the osteoporosis diagnosis support apparatus,
The threshold determination unit
For each of the three feature values, a “low bone density” fuzzy set and a “normal” fuzzy set are set, and the membership function is set to a parameter (t) for moving the intersection of both membership functions. Defined using
The sum of fuzziness is obtained for the learning data, and (t) that minimizes the fuzziness is defined as the threshold (T).
It is preferable.

In the osteoporosis diagnosis support apparatus,
The fuzzy set setting unit includes:
Three fuzzy sets (L, M, H) are set for each feature quantity based on the threshold value (T),
Wherein for each feature quantity, a minimum value T _l of those discriminated erroneously when determining _the maximum value is taken as T _h in the learning data the threshold (T),
The membership function (μL) of the fuzzy set (L) is maximum at T _l , minimum at T, and monotonically decreases in the interval from T _{l to} T.
The membership functions of the fuzzy set (M) (μM) is at most T, becomes minimum at _{T l} and _{T h, T} l through T of the interval monotonically increasing, the section of T～T _h is made monotonically decreased And
The membership functions of the fuzzy set (H) (μH) is a minimum of T, becomes maximum at _{T h,} section T～T _h is made of a monotonous increase,
It is preferable.

In the osteoporosis diagnosis support apparatus,
The fuzzy set setting unit includes:
Each feature is divided into three sections (l, m, h) at the intersection of the membership functions (μL, μM, μH),
Each of the learning data is classified into 27 types according to which of the three sections (l, m, h) the three feature quantities belong to,
If each of the classifications includes at least “low bone density” learning data, label the classification with “low bone density”, label the other classifications with “normal”,
27 kinds of fuzzy rules are set by associating the section (l, m, h) with the fuzzy set (L, M, H),
A product of membership values for each fuzzy set in each of the 27 rules is input to the neural network unit.
It is preferable.

An osteoporosis diagnosis support method according to the present invention is a method for determining osteoporosis using a feature amount related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Determining by threshold processing based on fuzzy logic;
Setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
The product of membership values to each fuzzy set set for each of the three feature values is input, and the connection weight of the neural network means that outputs “low bone density” or “normal” is used for the learning. Determining using data;
Obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Inputting a product of membership values of each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph into the neural network means.
It is characterized by that.

An osteoporosis diagnosis support program according to the present invention is a computer program for determining osteoporosis using a feature amount related to cortical bone,
Computer
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Means for determining by threshold processing based on fuzzy logic,
Means for setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network means for inputting a product of membership values to each fuzzy set set for each of the three feature values and outputting “low bone density” or “normal”;
Means for determining the connection weight of the neural network means using the learning data;
Means for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Means for inputting the product of membership values to each fuzzy set of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network means;
It is made to function as.

A recording medium according to the present invention is a computer-readable recording medium that records an osteoporosis diagnosis support program,
The program is a computer,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Means for determining by threshold processing based on fuzzy logic,
Means for setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network means for inputting a product of membership values to each fuzzy set set for each of the three feature values and outputting “low bone density” or “normal”;
Means for determining the connection weight of the neural network means using the learning data;
Means for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Means for inputting the product of membership values to each fuzzy set of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network means;
Is a program to function as
It is characterized by that.

An osteoporosis diagnosis support LSI according to the present invention is an LSI that determines osteoporosis using a feature amount related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. A circuit that is determined by threshold processing based on fuzzy logic,
A circuit for setting a plurality of fuzzy sets for each of the three feature values, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network circuit that inputs a product of membership values to each fuzzy set set for each of the three feature values and outputs “low bone density” or “normal”;
A circuit for determining a connection weight of the neural network circuit using the learning data;
A circuit for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone portion;
A circuit that inputs a product of membership values to each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network circuit.
It is characterized by that.

  The present invention uses the three feature values “cortical bone thickness”, “minimum segment size”, and “average segment size” extracted from a dental panoramic radiograph to determine whether the bone density is low or normal. This is done by a neural network. Thus, combining the three feature amounts has the following advantages.

  Cortical bone morphology and cortical bone thickness, respectively, “medical bone morphology indicates the rate of post-menopausal bone resorption in women” and “cortical bone thickness mainly represents the maximum amount of bone acquired at a young age” There is a background that these combinations are useful for discrimination of osteoporosis in that they represent the result of bone mass absorbed from the maximum bone mass. For the above analysis, it should be sufficient if there is a feature quantity of either “minimum segment length” or “average segment length” in addition to “cortical bone thickness”. However, since the image calculation for obtaining a skeleton is easily affected by subtle changes in cortical bone morphology, and errors often occur in these feature amounts, the present invention suppresses this by using both feature amounts, and the determination accuracy Has improved.

  In the present invention, threshold values are determined for each of the three feature quantities “cortical bone thickness”, “minimum segment size”, and “average segment size” by threshold processing based on fuzzy logic, and set for each feature quantity. By defining the membership functions of a plurality of fuzzy sets based on the above thresholds, it is possible to obtain an appropriate membership function according to the available learning data, which also improves the system's judgment accuracy. It is improving.

<Osteoporosis diagnosis support system>
In recent years, screening for osteoporosis patients using dental panoramic radiographs has attracted attention. The osteoporosis diagnosis support system according to the present embodiment supports this. The overall configuration of this system is shown in FIG. This system includes an X-ray image input unit 100, a cortical bone thickness extraction unit 200, a segment extraction unit 300, a learning database 400, a fuzzy calculation unit 500, and a neural network 600. This system uses the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” extracted from dental panoramic radiographs to determine whether it is “low bone density (low BMD)” or “normal”. Is determined by a fuzzy neural network. The processing performed by this system is roughly divided into a “learning phase” and a “determination phase”. In the “learning phase”, the system is optimized using a panoramic X-ray photograph of a learning patient who has already been determined to be “low BMD” or “normal”. In the “determination phase”, it is determined whether the panorama X-ray photograph that has not been determined is “low BMD” or “normal” by using the system optimized in the learning phase.

<Learning phase>
The flow of processing performed in the learning phase is shown in FIG. Hereinafter, the processing in each step will be specifically described.

≪ST200≫
The learning data used in this system is the postmenopausal of 100 people who have obtained consent at the age of 50 years or older who visited the clinic in which the inventors of the present application were involved during bone density (BMD) evaluation between 1996 and 2001. Based on panoramic radiographs of mandibular cortical bone obtained from women. The 100 learning patients randomly collected clinics involving the present inventors from 531 postmenopausal women who visited during 1996-2001 for DXA measurements. That is, for these 100 people, the judgment result (“low BMD” or “normal”) of the clinic based on the bone density data of the lumbar spine using the DXA method and a panoramic radiograph of the mandibular cortical bone are obtained. ing. Note that the learning data is merely an example, and the learning data usable in the present system is not limited to the above. Learning if there is a panoramic radiograph of the mandibular cortical bone of the patient for learning and bone density determination results ("low BMD" or "normal") obtained by a diagnostic method other than this system for the patient for learning It can be used as business data.

  A panoramic X-ray photograph of 100 learning patients is converted into a digital image by the X-ray image input unit 100 and supplied to the cortical bone thickness extraction unit 200. The cortical bone thickness extracting unit 200 extracts the cortical bone thickness from the digitized panoramic X-ray image. The cortical bone thickness extracting unit 200 is realized by using the technique disclosed in [Patent Document 2, Non-Patent Document 6]. The cortical bone thickness extraction unit 200 cuts out a part of the lower jaw body including the left and right mental holes and the mandibular base on the panoramic radiograph as a target area image (see FIG. 3A), Cortical bone thickness is acquired by a predetermined calculation process based on the hole (see FIGS. 3B and 3C). Details of this processing are disclosed in [Patent Document 2, Non-Patent Document 6]. The cortical bone thickness extraction unit 200 calculates the average value of cortical bone thickness obtained for the left and right target region images, and accumulates this average value in the learning database 400 as the cortical bone thickness for the panoramic radiograph. .

  On the other hand, a panoramic X-ray photograph of 100 learning patients is converted into a digital image by the X-ray image input unit 100 and supplied to the segment extraction unit 300. The segment extraction unit 300 extracts the minimum segment size (minimum SEG) and the average segment size (average SEG) from the digitized panoramic X-ray image. The segment extraction unit 300 is realized using the technique disclosed in [Patent Document 1, Non-Patent Document 9]. The inner surface of the mandibular molar lower marginal cortical bone on the panoramic radiograph (dotted line part in Fig. 4 (a)) is smooth and has a certain thickness for those who have not experienced a decrease in bone density. It has (refer FIG.4 (b)). However, it is clear that when the bone density is reduced, the inner surface absorbs linearly, resulting in a crude structure (see FIG. 4C). Such a difference in form can be identified by applying a skeleton process (one of mathematical morphological image processing techniques) to the cortical bone portion of the panoramic X-ray image. When skeleton processing is applied to a panoramic X-ray image of cortical bone (see FIGS. 5 (a1) and (a2)), images in which fine lines run in various directions are obtained (FIGS. 5 (b1) and (b2)). reference). Only the components parallel to the lower edge of the mandible are extracted from this skeleton image, and the difference in cortical bone morphology is identified by analyzing the size (number of pixels) and number of line elements (segments) that appear in the binarized image. (See FIGS. 5 (c1) and (c2)). Details of the skeleton processing and the like are disclosed in [Patent Document 1, Non-Patent Document 9].

  Similar to the cortical bone thickness extracting unit 200, the segment extracting unit 300 cuts out the left and right target area images (see FIG. 3A), and extracts segments from the cortical bone portion of each target area image by the above-described series of processes. The segment extraction unit 300 calculates a minimum value (minimum segment size) and an average value (average segment size) of the obtained segment sizes for each of the left and right target area images. The segment extraction unit 300 calculates the left and right average values of the minimum segment size and the average segment size, and uses these average values as the minimum segment size (minimum SEG) and average segment size (average SEG) for the panoramic radiograph, respectively. Are stored in the learning database 400.

  As a result of the above processing, various data about 100 learning patients are accumulated in the learning database 400 as shown in FIG. In the learning database 400, for each learning patient, the panoramic X-ray image ID, the determination result by the DXA method, the cortical bone thickness obtained by the cortical bone thickness extracting unit 200, and the minimum obtained by the segment extracting unit 300 The segment size (minimum SEG) and the average segment size (average SEG) are stored in association with each other.

≪ST210≫
The fuzzy operation unit 500 performs a threshold processing based on fuzzy inference to determine whether each of the three feature values “cortical bone thickness”, “minimum segment size”, and “average segment size” is “low BMD” or “normal”. Determine the threshold to be determined. This process is performed as follows.

  First, for each feature quantity, two fuzzy sets, a fuzzy set (L) corresponding to “low BMD” and a fuzzy set (N) corresponding to “normal”, are considered, and the membership function is represented by 2 Define with the following function.

In Equation 1, μ _L represents a fuzzy membership function corresponding to “low BMD”, and μ _N represents a membership function corresponding to “normal”. x is an element of each feature quantity, x _min is a minimum value, and x _max is a maximum value. t indicates a threshold value, and an arbitrary value between x _min and x _max is assigned. The parameter t moves the intersection of the two membership functions μ _L and μ _N to the left and right. Z represents a Z function, and S represents an S function.

Next, for each feature amount, fuzziness indexes γ _L (t) and γ _N (t) are obtained by [Equation 2] using data stored in the learning database 400.

In Equation 2, γ _L (t) represents a fuzziness index for the fuzzy set L, and γ _N (t) represents a fuzziness index for the fuzzy set N. X _L represents a set of each feature amount associated with the determination result of “low BMD” in the learning database 400, and X _N corresponds to the determination result of “normal” in the learning database 400. A set of each feature quantity is shown. | X _L | indicates the number of members of the set X _L , and | X _N | indicates the number of members of the set X _N. As the value of t increases, the fuzziness index γ _L (t) decreases and γ _N (t) tends to increase.

One of the fuzziness indices (here, γ _L (t)) is normalized by the normalization coefficient α shown in [Equation 3], and t (= T) that minimizes the reference function J (t) shown in [Equation 4] is obtained. This is used as a threshold value.

By the above thresholding method, the threshold value T _{cw for} the “cortical bone thickness”, the threshold value T _min for the “minimum segment size”, and the threshold value T _avg for the “average segment size” are obtained. These threshold values correspond to the optimum cutoff threshold values of “cortical bone thickness”, “minimum segment size”, and “average segment size” as shown in FIG.

≪ST220≫
Next, the fuzzy computing unit 500 uses three threshold values T _cw , T _min , and T _avg for three feature values of “cortical bone thickness”, “minimum segment size”, and “average segment size”. L, M, and H are set, and their membership functions μL, μM, and μH are defined. Details of this processing will be described by taking the case of “cortical bone thickness” as an example.

First, each learning data stored in the learning database 400 (FIG. 6) is determined by the threshold value T _cw . Specifically, each of the cortical bone thicknesses c _{1 to} c ₁₀₀ stored in the learning database 400 is compared with the threshold value T _cw . When the cortical bone thickness c _i is equal to or less than the threshold value T _cw , it is determined as “low BMD”, and otherwise it is determined as “normal”. Next, the determination result by the threshold value T _cw is compared with the determination result by the DXA method for the cortical bone thickness c _i . Minimum value T _l of the cortical bone thickness of both judgment results do not _match, the maximum value T _h. The fuzzy set _{L cw,} M _{cw, H} cw of membership function μL _cw, μM _cw, the .mu.H _cw, the _{T cw, T l,} based on the _{T h,} is set by a linear function as shown in FIG. 8 . Membership function μL _cw of fuzzy set _{L cw} is, up to _{T l (1),} the smallest in _{T cw} (0) _{next, T} l _{~T cw} of the section is a monotonically decreasing. Further, the interval between x _{min and} T _l takes the maximum value (1), and the interval below x _min and above T _cw takes the minimum value (0). Membership function μM _cw of fuzzy set _{M cw is,} up to _{T cw} (1), minimum (0) _{T l} and _{T h next, T} l _{~T cw} of the section is monotonically _{increasing, T} cw ~T _h section of Is monotonically decreasing. In addition, the interval of T ₁ or less and T _h or more takes the minimum value (0). Membership function μH _cw of fuzzy set _{H cw} is _a minimum of _{T cw} (0), up to _{T h} (1) _{next, T} cw ~T _h of the section will be monotonically increasing. _Further, the section of the T _{h ~x max} is a maximum value _{(1), T cw} below and _{x max} or more sections takes the minimum value (0).

As described _{above, T cw, T l, T} h is because it is calculated by using the learning data stored in the learning database 400, also changes these values in response to the training data available come. As a result, the shapes of the membership functions μL _cw , μM _cw , μH _cw also change as shown in FIG. 9, for example, according to the available learning data. As described above, in this system, the optimum membership functions μL _cw , μM _cw , and μH _cw are set according to the available learning data.

A process of setting fuzzy sets L _min , M _min , and H _min for the “minimum segment size” based on the threshold T _min and defining their membership functions μL _min , μM _min , μH _min , and The process for setting the fuzzy sets L _avg , M _avg , H _avg for the “average segment size” based on the threshold value T _avg and defining their membership functions μL _avg , μM _avg , μH _avg is the same as above. To be done.

≪ST230≫
Next, as shown in FIG. 10A, the fuzzy computing unit 500 divides each feature amount into three sections l, m, and h at the intersections of the membership functions μL, μM, and μH. As a result, the three-dimensional feature space composed of cortical bone thickness, minimum segment size, and average segment size is classified into 27 (= 3 × 3 × 3) subspaces [[cortical bone thickness, minimum SEG, Average SEG) is 27 (l, l, l), (l, l, m), ..., (h, h, m), (h, h, h)]. As the membership functions μL, μM, and μH change according to available learning data, the three sections l, m, and h also change as shown in FIG. 10B, for example. Further, the three sections l, m, and h vary depending on the feature amount.

  Next, the fuzzy computing unit 500 converts the learning data stored in the learning database 400 (FIG. 6) according to whether the three feature values belong to the sections of l, m, and h, respectively. Classify into one of the subspaces. Based on this classification result, each subspace is labeled “low BMD” or “normal”. Specifically, if at least one learning data whose determination result by the DXA method is “low BMD” is included, the subspace is labeled “low BMD”. ".

Next, the fuzzy calculation unit 500 has three inputs corresponding to the three feature quantities “cortical bone thickness”, “minimum segment size”, and “average segment size”, and one output corresponding to “low BMD” or “normal”. Is set based on the above labeling. Specifically, the sections l, m, and h are associated with the fuzzy sets L, M, and H. For example, “cortical bone thickness” is L _cw , “minimum segment size” is M _min, and “average segment size” is In the case of H _avg , 27 types of fuzzy rules such as “normal” are set for the label.

  Next, the fuzzy computing unit 500 obtains the product of the membership values for each fuzzy set in the rule for each of the 27 rules for each learning data in the learning database 400. Then, as shown in FIG. 11, these 27 products are used as inputs to 27 neurons in the input layer of the neural network 600. The neural network 600 has two neurons in the intermediate layer and one neuron in the output layer. In the output layer, “normal” is 0 and “low BMD” is 1. The weight of the neural network 600 is learned by the BP algorithm. Note that the initial value of the connection weight from the input layer to the intermediate layer is determined using the fuzzy rule. That is, if the label in the rule corresponding to a certain neuron is “normal”, the connection weight between the neuron in the input layer and the neuron in the intermediate layer is 0, and the weight is 1 if it is “low BMD”. The connection weights to both neurons in the intermediate layer have the same initial value.

<Judgment phase>
In the “determination phase”, whether the panorama X-ray photograph that has not been determined is “low BMD” or “normal” is determined as follows using the system optimized in the learning phase described above.

  The undetermined panoramic radiograph is converted into a digital image by the X-ray image input unit 100 and supplied to the cortical bone thickness extraction unit 200 and the segment extraction unit 300. The cortical bone thickness extracting unit 200 extracts the cortical bone thickness from the digitized panoramic X-ray image, and the segment extracting unit 300 extracts the minimum segment size and the average segment size. The extraction processing by the cortical bone thickness extraction unit 200 and the segment extraction unit 300 is the same as the processing in the learning phase. With respect to the extracted “cortical bone thickness”, “minimum segment size”, and “average segment size”, the product of membership values for each fuzzy set in the rule is calculated by the fuzzy computing unit 500 for each of the 27 rules described above. These 27 products are input to 27 neurons in the input layer of the neural network 600, and a “low BMD” or “normal” output is obtained from the output layer.

<Advantages of this system>
As described above, the osteoporosis diagnosis support system according to the present embodiment uses the three feature amounts “cortical bone thickness”, “minimum segment size”, and “average segment size” extracted from the dental panoramic radiograph. The determination of “low bone density” or “normal” is performed by a fuzzy neural network. Thus, combining the three feature amounts has the following advantages.

  Cortical bone morphology and cortical bone thickness, respectively, “medical bone morphology indicates the rate of post-menopausal bone resorption in women” and “cortical bone thickness mainly represents the maximum amount of bone acquired at a young age” There is a background that these combinations are useful for discrimination of osteoporosis in that they represent the result of bone mass absorbed from the maximum bone mass. For the above analysis, it should be sufficient if there is a feature quantity of either “minimum segment length” or “average segment length” in addition to “cortical bone thickness”. However, since the image calculation for obtaining the skeleton is easily affected by subtle changes in cortical bone morphology, and errors often occur in these feature quantities, this system suppresses this by using both feature quantities, and the judgment accuracy Has improved.

  Further, in this system, threshold values are determined for each of the three feature quantities “cortical bone thickness”, “minimum segment size”, and “average segment size” by threshold processing based on fuzzy logic (ST210), and each feature quantity is determined. By defining the membership functions of the three fuzzy sets set for each based on the threshold value (ST220), it is possible to obtain an appropriate membership function according to the available learning data. Has also improved the system's judgment accuracy.

<Experimental result>
An experimental result using this system is shown in FIG. Here, performance is evaluated in terms of sensitivity and specificity. The system was optimized using learning data for 50 people, and “low BMD” or “normal” was determined for panoramic radiographs of 50 test subjects. The learning data and the experiment subjects were randomly selected and repeated 10 times with different combinations. The sensitivity (Sensitivity) and specificity (Specificity) shown in FIG. 12 are average values of 10 times. As a comparison, the results of a fuzzy neural network method that does not use the thresholding algorithm (ST210) and a fuzzy inference system method that does not use the neural network (ST230) are shown. The sensitivity (Sensitivity) and specificity (Specificity) of this system incorporating a thresholding algorithm, fuzzy reasoning system, and neural network are 90.92% and 63.69%, respectively. This performance advantage is thought to indicate the importance of designing good algorithms to optimize membership functions and the importance of learning steps using neural networks.

<Other embodiments>
The osteoporosis diagnosis support system according to the present embodiment can be realized as a computer system, as a computer program, or as a computer-readable recording medium on which the program is recorded. Furthermore, it can also be realized as a dedicated LSI that executes the functions described in this embodiment.

  By using the present invention, a dentist and an orthopedic surgeon can collaborate to discover a patient with osteoporosis using a panoramic X-ray photograph of a jaw bone taken at a dental clinic and recommend an orthopedic surgeon to receive an appropriate treatment for osteoporosis. Can be done before it takes care.

The block diagram which shows the whole structure of the osteoporosis diagnosis assistance system by embodiment of this invention. Flow chart showing the flow of processing performed in the learning phase (A) An example of a dental panoramic radiograph, (b) (c) An example of cortical bone thickness Example of cortical bone morphology change on panoramic radiograph Diagram showing the segment extraction process Example of data stored in the learning database Fuzzy index for "cortical bone thickness", "minimum segment size" and "average segment size" Example of membership function of fuzzy set L _cw , M _cw , H _cw for “cortical bone thickness” Example of membership function of fuzzy set L _cw , M _cw , H _cw for “cortical bone thickness” An example of dividing each feature into three sections l, m, h at the intersection of membership functions Example of neural network configuration Experimental results using the system of this embodiment

Explanation of symbols

100 X-ray image input unit 200 Cortical bone thickness extraction unit 300 Segment extraction unit 400 Learning database 500 Fuzzy calculation unit 600 Neural network

Claims (8)

An apparatus for determining osteoporosis using features related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. A threshold value determination unit that is determined by threshold processing based on fuzzy logic;
A fuzzy set setting unit that sets a plurality of fuzzy sets for each of the three feature amounts, and sets a membership function of the fuzzy set based on the threshold;
A neural network unit that inputs a product of membership values to each fuzzy set set for each of the three feature values, and outputs “low bone density” or “normal”;
A connection weight determining unit for determining a connection weight of the neural network unit using the learning data;
A feature amount acquisition unit for acquiring “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic X-ray photograph including a cortical bone portion;
An input unit that inputs a product of membership values to each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” acquired by the feature amount acquisition unit to the neural network unit;
An osteoporosis diagnosis support apparatus characterized by the above. In claim 1,
The threshold determination unit
For each of the three feature values, a “low bone density” fuzzy set and a “normal” fuzzy set are set, and the membership function is set to a parameter (t) for moving the intersection of both membership functions. Defined using
The sum of fuzziness is obtained for the learning data, and (t) that minimizes the fuzziness is defined as the threshold (T).
An osteoporosis diagnosis support apparatus characterized by the above. In claim 2,
The fuzzy set setting unit includes:
Three fuzzy sets (L, M, H) are set for each feature quantity based on the threshold value (T),
Wherein for each feature quantity, a minimum value T _l of those discriminated erroneously when determining _the maximum value is taken as T _h in the learning data the threshold (T),
The membership function (μL) of the fuzzy set (L) is maximum at T _l , minimum at T, and monotonically decreases in the interval from T _{l to} T.
The membership functions of the fuzzy set (M) (μM) is at most T, becomes minimum at _{T l} and _{T h, T} l through T of the interval monotonically increasing, the section of T～T _h is made monotonically decreased And
The membership functions of the fuzzy set (H) (μH) is a minimum of T, becomes maximum at _{T h,} section T～T _h is made of a monotonous increase,
An osteoporosis diagnosis support apparatus characterized by the above. In claim 3,
The fuzzy set setting unit includes:
Each feature is divided into three sections (l, m, h) at the intersection of the membership functions (μL, μM, μH),
Each of the learning data is classified into 27 types according to which of the three sections (l, m, h) the three feature quantities belong to,
If each of the classifications includes at least “low bone density” learning data, label the classification with “low bone density”, label the other classifications with “normal”,
27 kinds of fuzzy rules are set by associating the section (l, m, h) with the fuzzy set (L, M, H),
A product of membership values for each fuzzy set in each of the 27 rules is input to the neural network unit.
An osteoporosis diagnosis support apparatus characterized by the above. A method for determining osteoporosis using features related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Determining by threshold processing based on fuzzy logic;
Setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
The product of membership values to each fuzzy set set for each of the three feature values is input, and the connection weight of the neural network means that outputs “low bone density” or “normal” is used for the learning. Determining using data;
Obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Inputting a product of membership values of each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph into the neural network means.
An osteoporosis diagnosis support method characterized by the above. A computer program for determining osteoporosis using features related to cortical bone,
Computer
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Means for determining by threshold processing based on fuzzy logic,
Means for setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network means for inputting a product of membership values to each fuzzy set set for each of the three feature values and outputting “low bone density” or “normal”;
Means for determining the connection weight of the neural network means using the learning data;
Means for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Means for inputting the product of membership values to each fuzzy set of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network means;
Program to function as. A computer-readable recording medium recording an osteoporosis diagnosis support program,
The program is a computer,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. Means for determining by threshold processing based on fuzzy logic,
Means for setting a plurality of fuzzy sets for each of the three feature quantities, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network means for inputting a product of membership values to each fuzzy set set for each of the three feature values and outputting “low bone density” or “normal”;
Means for determining the connection weight of the neural network means using the learning data;
Means for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone part;
Means for inputting the product of membership values to each fuzzy set of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network means;
Is a program to function as
A recording medium characterized by the above. An LSI for determining osteoporosis using feature quantities related to cortical bone,
The threshold value for discriminating between “low bone density” and “normal” for each of the three features “cortical bone thickness”, “minimum segment size”, and “average segment size” related to cortical bone is used using learning data. A circuit that is determined by threshold processing based on fuzzy logic,
A circuit for setting a plurality of fuzzy sets for each of the three feature values, and setting a membership function of the fuzzy sets based on the threshold value;
A neural network circuit that inputs a product of membership values to each fuzzy set set for each of the three feature values and outputs “low bone density” or “normal”;
A circuit for determining a connection weight of the neural network circuit using the learning data;
A circuit for obtaining “cortical bone thickness”, “minimum segment size”, and “average segment size” from a panoramic radiograph including a cortical bone portion;
A circuit that inputs a product of membership values to each of the fuzzy sets of “cortical bone thickness”, “minimum segment size”, and “average segment size” obtained from the panoramic radiograph to the neural network circuit.
An LSI for osteoporosis diagnosis support characterized by the above.