KR100545060B1 - Method and system for clustering bone images - Google Patents

Abstract

Osteotomy clustering method and system are provided. The bone skeletal image clustering method and system binarizes the bone skeletal image of interest, assigns it to a class based on the coincidence ratio of the binarized image and the binary image of the representative image, removes unnecessary classes among the classes, and determines the representative image of the class. Recluster by reviewing the number of classes. In addition, the number of representative images generated through the clustering method may be reduced by matching the representative images according to the user's selection.

Clustering, case, class, representative image, error value, match rate

Description

Method and system for clustering bone images

1 is a schematic block diagram of a spine image clustering system according to an embodiment of the present invention.

2 is a flowchart illustrating a process for clustering spine images according to an embodiment of the present invention.

Figure 3 is a detailed flowchart of the preparation step (S210) according to an embodiment of the present invention.

4 is a cross-sectional view showing the selected 256 contrast image of the spine and the user's region of interest.

5A through 5B are cross-sectional views showing areas of interest selected by a user in 256 shades and a binarized image through a binarization process.

6 is a detailed flowchart of the clustering step (S220) according to an embodiment of the present invention.

7 is a detailed flowchart of a representative image matching step S270 according to an embodiment of the present invention.

8 is an exemplary diagram of a representative image matching step according to an embodiment of the present invention.

9A to 9C are cross-sectional views of assigning contrast values to regions of interest of a case image and a matching template and matching each other according to an embodiment of the present invention.

FIG. 10 is a perspective view illustrating a region in which a matching template can scan a transition space in a case according to an embodiment of the present invention.

 (Explanation of symbols for the main parts of the drawing)

110: user interface 120: class assignment control module

130: control module 140: image source database

150: Representative video database by class

160: Binary image database 170: Image acquisition module

180: region of interest extraction module 190: binarization control module of region of interest

The present invention relates to a bone fracture image clustering method and system, and more particularly, to manage a more complex bone fracture image case by assigning it to a limited class based on a matching rate determined through self-learning, and to manage the bone fracture more systematically. An image clustering method and system are provided.

The human body can be divided into head, fuselage, upper limb and lower limb. The head is the skull, the body is the spine, sternum, ribs, pelvis, etc. In addition, the upper limbs include clavicle, scapula, humerus, radius, ulna, carpal bone, metacarpal and phalanges, and the lower extremities include femur, patella, tibia, fibula and femur. The osteolysis image clustering method forms a class of images according to a predetermined rule and uses a representative image of the class, and thus it can be applied to all general osteolysis images. However, as an embodiment, a clustering method and system for a spine image will be described below.

Spinal fixation device is a device used to depress the depressed spinal cord and nerve roots, correct the white spine and restore the support function of the spine. Implants, which are generally biocompatible metal products, are used here and are usually fixed in the motion segment disk space at the location of L3-L5. Implants reduce the pressure inside the disk and prevent disk space from narrowing.

As the material of the implant, soft interspinous implants, which can solidify the stability-related operation as well as rigid fixation and prosthetic replacement, are also used. The diameter of the silicon is determined by the size of the interspinous space, and it has been shown to contribute to the stability of the spine by being placed in the least surgical method. However, the dimensions of these implants are not determined to be of standard type. In other words, in the CT image, the slice is selected and observed in the z-axis direction, and based on information such as the length of the interspinous process, the maximum width of the spinous process, the height of the patient, and the weight of the spine. Several candidate implants in a custom fashion were made. In fact, custom implants are time consuming and expensive.

Therefore, it is necessary to manufacture implants of a standard type method which is classified into a finite number of forms. However, the huge spine image database consists of three-dimensional, various forms of spinous processes, including information on length and angle, including information on contrast values in two-dimensional images, and thus directly refers to meaningful criteria from these databases. It is difficult to elicit. In order to solve this problem, the present invention introduces Automatic Cluster Detection. Due to the lack of prior knowledge of subgroups, it is a good solution to divide them into clusters when it is not possible to subdivide them into a given number.

In general, clustering analysis can be used to reduce the amount of data and categorize similar data. This clustering process is widespread in information processing and one of the objectives of using clustering algorithms is to provide an automated tool for defining or helping categories or classifications. Clustering tries to find similar groups in the hope that similar records will behave similarly, instead of requiring no distinction between independent and dependent variables or already categorized data sets.

Therefore, the first thing that should be applied to comparing or integrating data is to reduce the size of the data set without losing the core characteristics of the data. The most important way is to find a data set that is computationally efficient and can represent the original data. The basis for finding the data set is usually the average of the observations. The present invention proposes an algorithm that improves the existing k-means clustering method as a method for detecting the shape distribution of spinal processes.

Although the Euclidean distance between observations was used as a criterion for clustering analysis used in the existing k-means clustering, the Euclidean distance is limited to use as a criterion for defining similarity between images.

In addition, in the past, the number k of clusters was generally determined by trial and error and subjective viewpoint. The choice of k depending on the type of cluster can have a decisive effect on the results, which can lead to poor results if it does not fit the structure of the input data.

Traditional k-means clustering methods usually specify k centers by the user, either by reading k centers from a file or from another application. However, these arbitrarily determining centers may distort the result if the initial center is far from the data distribution.

Also, in general, a larger set of observations may result in insignificant vibrations at the Euclidean distance value calculated between the observations and the representative values of the clusters and thus may not converge. Therefore, if the iterative process fails to move to the complete cluster center, the convergence criterion is selected, in which the process is terminated when the threshold is reached from the center to the center. However, these special stop rules may hinder the overall optimal clustering process.

An object of the present invention is to provide a bone skeleton image clustering method.

Another technical problem to be achieved by the present invention is to provide a bone skeletal image clustering system.

The technical problems of the present invention are not limited to the above-mentioned technical problems, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In accordance with an aspect of the present invention, there is provided a bone fracture image clustering method according to an embodiment of the present invention. (A) Binarizing and providing a bone fracture image of interest, and (b) matching ratio of the binarized image and the binary image of the representative image. (C) removing unnecessary classes among the classes, (d) determining a representative image of the class, and (e) counting the number of classes for which the representative image is determined. It is characterized by performing steps (b), (c) and (d) repeatedly.

According to an aspect of the present invention, there is provided a bone fracture image clustering system comprising: a first storage unit for storing an input image, a second storage unit for storing a representative image for each class, and the input image. A third storage unit for storing the binarized image of the region of interest of the representative image of each class, if a new case is included in the class, the representative image of the corresponding class is revised, and if the number of allocated classes is less than the specified limit, the matching rate It includes a control module for changing, class allocation control module for allocating a case or creating a new class according to the matching rate determined by the learning module.

Other specific details of the invention are included in the detailed description and drawings.

Advantages and features of the present invention and methods for achieving them will be apparent with reference to the embodiments described below in detail with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various forms. It is provided to fully convey the scope of the invention to those skilled in the art, and the present invention is defined only by the scope of the claims. Like reference numerals refer to like elements throughout.

First, FIG. 1 is a schematic block diagram of a bone fracture image clustering system according to an embodiment of the present invention. The user interface 110, the class assignment control module 120, the control module 130, and the image source database 140 are illustrated in FIG. , A representative image database 150 for each class, a binary image database 160, an image acquisition module 170, an ROI extraction module 180, and a binarization control module 190 of an ROI.

The user interface 110 accepts a user's interest portion of the entire spine image or manages input / output for image collection.

The class assignment control module 120 assigns a case or creates a new class according to the matching rate determined by the learning module.

However, the term "module" used in the present embodiment refers to software or a hardware component such as an FPGA or an ASIC, and a module plays a role. However, modules are not meant to be limited to software or hardware. The module may be configured to be in an addressable storage medium and may be configured to play one or more processors. Thus, as an example, a module may include components such as software components, object-oriented software components, class components, and task components, as well as protrusions, functions, properties, procedures, and subroutines. , Segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functionality provided within the components and modules may be combined into a smaller number of components and modules or further separated into additional components and modules. In addition, the components and modules may be implemented to play one or more CPUs in a device or secure multimedia card.

The control module 130 mainly revises the representative image when a new case is included in the class. In addition, for convergence, the representative image of the previous round and the representative image of the current round may be compared. If the number of allocated classes is less than the specified number, the matching ratio is increased to recluster.

The image source database 140 stores the image collected by the image collection module 170 without modification, and provides an image source when making a binarized image or modifying a representative image for each class.

The representative image database 150 for each class stores a representative image reflecting the characteristics of each class, and the representative image provides an average of an image source belonging to a class.

The binary image database 160 extracts a region of interest of the original image of the case and stores the image obtained by transforming the region of interest into a binary format. In addition, it stores the binarized image of the representative image of each class. This is used to compare the binarized image of the case and the binarized image of the representative image when assigning each case to a class.

The image acquisition module 170 receives the original image from the outside and stores it in the image source database 140 or provides it to the ROI extraction module 150.

The ROI extraction module 180 receives the original image from the image acquisition module 170, and provides the binarization control module 190 of the ROI when the user extracts the ROI.

The region of interest binarization control module 190 receives the region of interest from the region of interest extraction module 170 and binarizes the image according to a predetermined criterion. This is provided to the binarized video database 160.

A detailed method of automatically classifying and managing images by the osteolysis image clustering system of the present invention will be described with reference to FIGS. 2 to 10.

FIG. 2 is a flowchart illustrating a process for clustering bone images according to an embodiment of the present invention. First, a region of interest is extracted before entering the clustering step, and a preparation step of binarizing the extracted region of interest is performed. . Then, the first case is assigned to the first class (hereinafter referred to as class A) (S210).

 The binarized ROI image of the second case is matched with the binarized representative image of Class A. If there is a match, the case is included in class A, and the representative image is revised. If it is not matched, class B is created and assigned. Repeat this process for all cases. That is, the binarized ROI image of the nth case is compared with the existing class A to class P, and assigned to the matching class. If there is no matching binarized representative image, a new class P + 1 is created and assigned. In addition, upon completion of the assignment for the last case, the classes of the clusters included will remove classes that are below a certain percentage of the total cases, for example less than 2%. This process can be selectively made according to the purpose of extracting the exceptional and different shape case separately (S230).

After removing the exceptional classes, check whether the representative image of each class has been revised. If there is a revision of the representative image, repeat the clustering step and class removal step again. If convergence, check the number of classes. This convergence process is repeated until there is no further revision of the representative image (S240).

The error degree of the class that has undergone the convergence step is examined (S247), and if the number of classes determined by the matching rate is smaller than the number of classes allowed by the user, the matching rate is increased to repeat the clustering step, class removing step, and convergence step. (S250)

Additionally, the matching template and the case of the representative image completed through the above steps may be matched to select only a more general representative image.

Hereinafter, each step will be described in detail with reference to the accompanying drawings.

Figure 3 is a flow chart illustrating in detail the preparation step (S210) according to an embodiment of the present invention.

This includes determining a region of interest (VOI) for the portion of the entire spine image where the implant is to be placed, and binarizing the image.

As shown in FIG. 4, when the entire spine image having the 256 contrast value is provided (S211), the user extracts the ROI in a rectangular shape as shown in the figure (S212). In the present embodiment, the three-dimensional point of interest (VOI) point is determined in the form of a three-dimensional hexahedron with x, y, z, and axes. First, a slice corresponding to a portion where an implant is to be located is selected, and a method of determining a region of interest on a plane within the slice is performed. The region of interest (VOI) adds a margin to the part where the implant is to be placed, and the number of slices in the z direction selects several sheets for the upper part and the lower part, and is located at a certain point in the plane. The pixels are the same size for all images. For example, you can select 10 z-direction slices and determine all images with 64 * 64 pixels in the plane. It also processes the region of interest to be located at the same location on all slices in one case. Each case handled in this way will be called {S ₁ , S ₂ , ..., S _N }.

Thereafter, the binarization operation is performed by Equation 1 (S213). In the following, the whole cases are divided into k clusters {P ^t _j , 1≤j≤k}, and the representative image vector { _c μ ^t _j (j = 1, ..., k) when the cases belong to cluster P _j } Is classified as t denotes a round and c denotes the number of updates in j. In addition, B ( _c μ ^t _j (x, y, z)) Denotes a function that binarizes the intensity of the voxel at the position (x, y, z) of the representative image vector.

As a threshold value for the binarization of the spine, a similar result can be obtained by selecting any value from 190 to 240, but 230 is preferable in consideration of noise reduction and spine density. In addition, in the case of low bone density person, 200 can be selected. The threshold is not limited to the above embodiment. FIG. 5A illustrates a region of interest (ROI) on a 64 * 64 plane, and FIG. 5B illustrates a result of binarizing a 64 * 64 region of interest image.

After that, by assigning the first case to class A (S214), the preparation step is completed.

6 is a flowchart illustrating a clustering step (S210) in detail according to an embodiment of the present invention.

First, when t = 0, the initial matching ratio Th _s = Th _start is determined for cluster allocation (S221). The method of determining the initial matching rate may be directly designated by the user or randomly determined within a predetermined range. For example, it is selectable within the range of 70 to 100%, and in the example, the initial matching ratio was selected as 76%. However, the initial matching rate does not significantly affect the clustering result, so it is acceptable to select an appropriate range of values.

Thereafter, a process of assigning the nth case to the Pth class is performed (S222). As shown in Equation 2, the binarized contrast value of the case to be processed and the binarized contrast value of the representative value image of the class are compared at the positions of the designated voxels to obtain a ratio of matching contrast values. That is, the cumulative value of the pixels in which the contrast value of the case and the contrast value of the class representative image match is converted into a percentage. In Equation 2, A _case means the number of pixels of the VOI, and is converted into the number of pixels * slice of the ROI. In this embodiment, A _case is 64 * 64 * 10. In addition, Con [S _i , _c μ ^t _j ] is a function that repeatedly compares the binarized image of the _case and the binarized image of the representative image of the class by A _case , and represents the ratio having the same contrast value.

The method of comparing the matching rates as described above is a method of comparing the contrast value and the position information together, and has a characteristic of obtaining a ratio in the total volume by comparing the contrast values of the corresponding voxels.

After the matching, the value of Con [S _i , _c μ ^t _j ] is compared with the initial matching rate determined previously, and if the value is less than the initial matching rate, a new class is created and the case used in the matching process is allocated. S223) When the initial matching ratio is over, the representative video is revised. (S224)

Explaining the class assignment process in detail, assign the first case S1 to class 1. In the case of the second case S2, the representative image of the cluster 1 allocated to the cluster 1 is compared with the contrast value of the corresponding voxel of the case. Here, if the match rate is smaller than the initial match rate, case S2 is assigned to cluster 2, and if the match rate is greater than the initial match rate, case S2 is assigned to cluster 1 and the representative value image of cluster 1 is revised. From the second case S2 and above, as shown in Equation 3, the match rate is compared with the binary representative image of all the made classes P _j (j = 1, ..., k), and the largest match rate is compared. Assign case S _i to the visible P _j . Representative image revision is required in P _j to which case S _i is assigned.

In the method of correcting the representative image (S224), the contrast value of the pixel is corrected using 256 contrast values instead of the binary image. When the number of cases belonging to the current class P _j is c, the new representative image is shown in Equation 4. Hereinafter, the process of changing the representative image by using Equation 4 is a learning process, and the ratio of the newly revised _{c + 1} μ ^t _j and the previous _c μ ^t _j is defined as the learning rate.

Equation 4 is an average of existing cases and newly allocated cases. The method of determining the representative image in the above method can obtain accurate clustering and the representative image by creating a flexible representative value by assigning a new case rather than determining a simple mean value of cases belonging to the same class as a representative value vector. have.

The learning rate is inversely proportional to the number of cases assigned to that cluster. When processing one learning case, since the value determined by the existing c cases is previously assigned to the representative value, the learning rate by the single case decreases as the number of existing cases forming the representative image increases.

The reason why the learning rate decreases is because the representative image is determined by the cumulative calculation value of all the cases assigned to the class, and the representative images of each class are made the representative images of all the cases assigned to the class. Therefore, each class is determined only by such a representative image, and all cases belong to the class of the representative image closest to itself.

After the representative image is revised, the above process is repeated until all the cases entering the current clustering stage are allocated, that is, S _N (S225).

After the allocation of the last case is completed, using Equation 5, the class of the included class is less than 2% of the total case is removed from the class list (S230). By eliminating exceptional classes, k classes can be composed of only meaningful classes, and the disadvantages of algorithms sensitive to k values and exception values are compensated for. Thus, this process can be understood as extracting cases of exceptional and distinctive shapes separately.

Table 1 shows the result of class removal using 100 cases. Looking at the first row, analyzing L34_upper based on 76% match rate, four classes were created, and the class consisting of two or less cases was removed, indicating that four cases were dropped.

division L34_upper L34_lower Match rate k number Number of cases dropped K count Number of cases dropped 76 4 4 2 2 77 4 3 2 3 78 4 4 2 6 79 4 4 2 6 80 5 5 2 6 81 6 3 3 6

When the class removal process is finished, the process proceeds to the convergence (Convergence) step for stabilization of each class (S240).

The convergence step checks whether there is a revision on the representative image of each class and repeats the round until the representative image converges. Rounding then means going through a clustering step and a class removal step. Thus, repeating three rounds means repeating the clustering step and the class removal step three times.

As shown in Equation 6, the round is repeated until the matching ratio between the representative image of each class of the previous round and the representative image of the newly formed class becomes a certain percentage or more. The value used as the round repetition criterion may be specified by the user. For example, 100% may be specified to require a perfect match.

If it does not reach a certain percentage or more, it is determined that convergence has not been made. In this case, since the case must be reassigned from the beginning, the case number is initialized to 1 (S245), and the round is repeated. If convergence is made, the process proceeds to the error value calculation step S247.

Error degree calculation step S247 is a step of calculating the error degree of the converged classes. The error degree is calculated as the sum of squares of distances between the cases and the representative image to which each case belongs, as shown in Equation (7). The distance is a conceptual expression, and in the case of the error observation value in the above equation, it is calculated by considering the degree of inconsistency with the representative image as 100%. After the distance from the representative image is also calculated, the process goes to the reclustering request step (S250).

The reclustering request step (S250) means that if there is no further revision of the representative image and the number of the corresponding classes is smaller than the class limit number (), the matching rate is increased by using Equation 8 (S255). If reclustering is required, the matching rate is raised (S255) to return to the clustering step, otherwise, the matching rate and k value determination step (S260) are passed. This is because the initially determined coincidence rate Th _start is an approximate value, and thus the k value determined accordingly is not completely suitable for a given case assignment. Therefore, through this process, it is possible to create the maximum match rate (Th _s ) within the limited class count () category. For example, if the number of the corresponding class is smaller than the class limit, increase the matching rate by 1% and then return to the clustering step (S220) to repeat class distribution for all cases.

In the determination of the matching rate and the k value (S260), the matching rate and the k value corresponding to the smallest error degree among the calculated error degrees according to the increase of the matching rate are determined as the most suitable. From Table 2, it can be seen that the distance squared value according to the coincidence rate is calculated and the calculated error degree is excluded except for the case removed in the class elimination step. Here, it can be seen that the error value is the smallest as 67414 when the agreement rate is 82%. Therefore, 82% is calculated to be the best match for the initial match of the cases.

After the matching rate and the k value determination step, the representative image matching step S270 may be selectively performed. By scanning all cases using an implant suitable for the representative image of the class, the number of classes can be reduced and a general representative image can be made. In the process of scanning, if the implant does not fit into the protrusion at a certain point, it stops and the position is moved to another point at that point and repeated comparison. Continue the comparison and find and remember the point with the least error, that is, the point with the least error space. This matching process is achieved by a per voxel comparison of the template 273 'with the case 271' and the implant geometry, for example, the dimensions of the case are horizontal * vertical * slice (64 * 64 *). 10) and the size of the template (48 * 48 * 4), the template is moved by one voxel with respect to the three-dimensional of the case to compare the entire section. In this case, the variation interval is (x, y, z) = (17, 17, 5).

7 is a flowchart illustrating a specific representative image matching process according to an embodiment of the present invention, and FIG. 8 is an exemplary diagram of the representative image matching process. The representative image matching process may be usefully used to make a general implant suitable for the most cases when making an implant.

In addition, the purpose of the matching process can be divided into two. First, since the implant is composed of geometric curved surfaces, it may be different from the three-dimensional shape of the representative image. The matching process is a process of reducing the number of implants and improving the shape more appropriately through quantitative measurement of such error space between the implant and the case. Secondly, since the representative image has errors as much as the coincidence rate with a single case, when the implant is manufactured, it is produced with a slightly larger dimension in consideration of such an error, and a matching process is required as a verification process for such an extended dimension.

After determining the matching rate and the value of k, a case is provided (S271). The case is provided in a comparable form, which in this example is 64 * 64 * 10.

Thereafter, a step S272 of providing a 2D image of an implant suitable for the representative image is performed. First, in the representative image, measure the dimensions of the width of the spine, the interspinous space, the angle of the bottom of the spine, and the depth of the spine. Create an implant with extra margin. Hereinafter, such an implant is referred to as a representative pattern. In addition, a number of representative patterns can be made according to how much margin is taken into account, and their suitability is determined through a matching process.

The two-dimensional image of the provided implant is provided in a form that can extract the matching template and the region of interest, it is provided in the form of a case, or preferably provided as a stereoscopic image of 64 * 64 * 10.

Subsequently, in operation S273, a matching template suitable for the 2D image is provided. The matching template is a process for allowing the presence of a spine process to be detected in the outer region of the template like the region of the template itself over 180 degrees based on the lower center point of the representative pattern. The area has a size of 48 * 48 * 4. As shown in FIG. 8, the matching template 273 ′ is more expanded than the 2D image of the representative pattern for matching with the case image.

Thereafter, a region of interest (ROI) is extracted from the matching template (S274). Referring to FIG. 8, a portion of the matching template is selected as the ROI 275 ′. For example, if the matching template is 64 * 64, the ROI is determined to be 48 * 48. This is to scan the case region 64 * 64 along the region of interest 48 * 48 along the transition space 17 * 17. When the ROI is determined, a contrast value is allocated to the ROI of the case image and the matching template in order to distinguish it from the case image to be compared (S275). 9A and 9B, in the case image, the projection A is represented by 0 and the background B is represented by 255, whereas the ROI of the matching template is the background C of 100 and the matching template D of 255. Treated separately. However, the contrast value allocation is not limited to the embodiment because it is possible to use other contrasts that can be distinguished, and is used to distinguish four areas in the process.

An error value is calculated by matching a region of interest and a case of the matching template given a contrast value (S276). Referring to FIG. 9C, the result of the matching is analyzed as follows.

The area was largely divided into four areas. In area 1, only the template exists. The contrast value is represented as (255 + 255) / 2 and is skipped in the process. Area 2 is where the template overlaps the case and the contrast value is (0 + 255) / 2. This is because the protrusion of the case is larger than the implant and protrudes out of the template, so if the voxel is detected during the process, the stereo is considered inappropriate. Area 3 is the area where only the protruding part of the case exists, and the contrast value is expressed as (0 + 100) / 2. Therefore, this part does not affect the processing, so it is skipped like the first area. Area 4 is an area where only the backgrounds of the two images exist, which means the error space between the template and the case. The intensity value is represented as (255 + 100) / 2 and is considered as an error space so that the number of voxels is accumulated. The cumulative value in area 4 is calculated as an error value and used to calculate the coincidence rate. Table 3 summarizes these results.

division domain case template Contrast distribution process One template - o (255 + 255) / 2 skip 2 Protrusion + Template o o (0 + 255) / 2 Nonconforming 3 spin o - (0 + 100) / 2 skip 4 background - - (255 + 100) / 2 Error value calculation

The error value in the fourth region is a region where the protrusion and the implant do not fit. The larger the error value, the larger the gap between the implant and the protrusion. Therefore, it is important to reduce the error value, and the cumulative error value is calculated by calculating all the error values while scanning the matching template with respect to the variation space for the whole case. That is, one matching template is 48 * 48 * 4, and when four slices are matched, an error value of each slice is calculated and accumulated to calculate a cumulative error value.

Next, referring to FIG. 10, the matching template 920 scans the shift space 17 * 17 * 5 in the x, y, and z directions to obtain the cumulative error value at each shift space in the case 910. do. Even if the images of the matching template and the case are not exactly registered in the horizontal, vertical, and height, the matching template may be detected where the matching template is located and may be detected even when the angle is different.

If the matching process and the error calculation process are implemented in Visual C ++, for example, they are as follows.

For (Z_range) {

For (X_range) {

For (Y_range) {

if (region_①)

;

else if (region_②)

loop exit;

else if (region_③)

;

else

error_rate ++;

}

}

}

If (error_rate> threshold)

exit;

Repeat for x_range, y_range, z_range by the first for statement, skip processing for 1 and 3 areas for each area, exit the loop for 2 areas, and error only for 4 areas. Accumulate the values.

When 17 * 17 * 5 cumulative error values are calculated through this process, the cumulative error value which is the minimum among the cumulative error values is determined as the representative error value of the case with respect to the implant of the representative image.

This scanning process and the representative error value determination process should be repeated for each representative image and case, and the amount of repetitive scans can be calculated by Equation (9).

x_range, y_range, and z_range are x, y, and z values of the disparity space, respectively, k is the number of the determined class or the number of representative images, and case is the number of cases. According to the above formula, the number of repetitions for one representative image of one case having a variation space of 17 * 17 * 5 is 17 * 17 * 5 * 1 * 1 = 1445 times. In addition, in order to process 100 cases of 40 representative images, 17 * 17 * 5 * 40 * 100 = 5,780,000 repetitions are performed.

When the representative error value of the case for the implant corresponding to the representative image of each class is calculated through repetition, the case is assigned to the class having the smallest representative error value (S277). Table 4 shows that case 1 is assigned to class B because the representative error value in class B is the smallest. The method of allocating a class may use a representative error value as described above, or obtain a representative error rate through a ratio of each representative error value to a total representative error value and allocate the same to the smallest representative error rate. It is also possible to obtain (100-representative error rate)% and assign it to the largest value, and the assignment method is not limited by parameters.

case Implant that matches the representative image Representative Error Value Assignment Case 1 Class A 69300 Class B Class B 67800 Class C Does not fit ... ... Class J 76666

When making an implant, it is not made to fit exactly into the intervertebral to be inserted, but the allowance is taken into account. Therefore, the same implant can be used even for representative images of different classes. In addition, due to the high k value and the matching rate, there is a possibility to be allocated to another class due to the difference of the detailed image. Therefore, when more general implant production is required, it is necessary to perform representative image matching, and in order to produce detailed and various kinds of implants according to a user's selection, the representative image matching step may be omitted.

Table 5 shows the (100-representative error rate)% of each case and class. As described above, it can be expressed through the representative error value or the representative error rate. The results obtained by matching each case with the implant corresponding to the representative image of each class are described. The approximate frequency of matching can be found and the class that can hardly match can be filtered and used as the primary evaluation item.

division Class A Class B Class D Class F Class H Case 1 86.4 81.2 72.7 80.6 69.7 Case 2 0 81.9 79.4 0 82.1 Case 3 0 0 79.9 0 77.6 Case 4 0 0 79.5 0 77.2 Case 5 0 76.9 72.8 81.1 66.9 Case 6 0 88.1 76.7 0 78.8 Case 7 88.5 90 82.9 0 78.5 Case 8 0 90 82.9 0 78.5 Case 9 77.2 61.5 51.9 59.9 47.5 Case 10 0 92.8 85.9 0 82.5 Case 11 0 84 73.6 0 71.3 Case 12 0 89.3 83.1 0 77.9 Case 13 83 78.5 68.4 0 65.4 Case 14 0 81.5 66.2 0 68.7 Case 15 0 83.8 68.8 0 73.4 Case 16 0 0 77.7 0 71.7 Case 17 0 93.1 87.1 0 86.9 ... ... ... ... ... ... Case 95 0 88.2 76.5 0 80.6

Referring to Table 6, the average for each class, the number of allocated cases, and the optimal decision rate are shown. The average match rate is the average of the (100-representative error rate)% of Table 5, the number of cases represents the number of cases allocated to each class, and the optimal decision rate is the ratio of the number of cases to the total cases in%. It is shown. For example, for class A, this is calculated as (19/95) * 100 = 20%.

division Class A Class B Class D Class F Class H Average by class (%) 13.94 75.99 72.79 12.51 74.84 Number of cases 19 60 4 One 11 Optimal Decision Rate (%) 20 63 4 One 12

Although embodiments of the present invention have been described above with reference to the accompanying drawings, those skilled in the art to which the present invention pertains may implement the present invention in other specific forms without changing the technical spirit or essential features thereof. I can understand that. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive.

According to the osteoblast image clustering method and system as described above, there are one or more of the following effects.

First, various three-dimensional bone fracture image cases can be classified into a limited number of classes.

Second, the class can be revised by clustering new bone image cases based on the determined class.

Third, it is possible to easily create an implant for general bone damage correction using the representative image of the classified class.

Claims (14)

(a) binarizing and providing an osteoblast of interest; (b) assigning a class to a class based on a matching ratio between the binarized image and the binarized image of the representative image; (c) removing unnecessary classes of the classes; And (d) determining a representative image of the class; The bone fracture image clustering method of repeating the steps (b), (c) and (d) by examining the number of classes for which the representative image is determined. The osteotomy image clustering method of claim 1, wherein the osteotomy image is a spine image. The method of claim 1, wherein step (a) (a1) providing a full osteolysis image; (a2) extracting a region of interest from the whole bone marrow image; and (a3) binarizing the region of interest Osteotomy clustering method comprising a The method of claim 3, wherein the step (a2) is performed by the user. The method of claim 3, wherein the region of interest to be binarized is

_C ^{o t} _j (x, y, z) represents a representative image of a class, and (x, y, z,) represents a bone fracture image clustering method representing a voxel position of a representative image The method of claim 1, wherein step (b) (b1) determining an initial match rate; (b2) calculating a cumulative value of pixels in which the binarized image of the spine of interest and the representative image of each class match; And (b3) A bone fracture image clustering method comprising the step of revising a representative image of the class when a new observation is assigned. The boneless image clustering method of claim 1, wherein the unnecessary class is determined according to a ratio of the number of cases included in the class to the total number of cases. The bone marrow image clustering method of claim 1, wherein steps (b) and (c) are repeated until the representative image matches a predetermined ratio or more compared with the previous round. 10. The method according to claim 8, wherein the ratio is 100%. The bone bone image clustering method of claim 1, wherein the repetitive performance is performed when the number of classes for which the representative image is determined is less than a number designated by a user. The method of claim 1, further comprising: matching the representative image with a matching template. The method of claim 11, wherein the matching step Making a two-dimensional image of an implant suitable for the representative image; Creating a matching template based on the 2D image; Extracting a region of interest from the matching template; Assigning a contrast value to a region of interest of the matching template; And Comparing bone region image case with the ROI of the matching template 13. The osteotomy image clustering method according to claim 12, wherein the comparison is performed using at least one of a representative error value, a representative error rate, and (100-representative error rate)%. A first storage unit which stores an input image; A second storage unit which stores a representative image for each class; A third storage unit storing the binarized image of the ROI of the input image and the representative image of each class; A control module for revising the representative image of the corresponding class when the new case is included in the class, and changing the matching rate when the number of allocated classes is less than the specified limit number; And And a class assignment control module for allocating a case or generating a new class according to the matching rate determined by the learning module.

Images