KR102308059B1 - Level Set Segmentation method and apparatus for Medical Image - Google Patents

Abstract

The present invention relates to a 3D level set segmentation method for a medical image, comprising: allocating indexes for structure information recorded in level set data vector information to each of a plurality of core units that are multi-cores of a GPU; Each of the plurality of core units allocated with the index reads structure information on a corresponding voxel recorded in structure information corresponding to the allocated index and structure information on voxels surrounding the voxel, and calculates a level of the corresponding voxel based on the structure information and updating the level of the corresponding voxel to the calculated level; and deriving an object region using the updated voxel levels, wherein the multi-core plurality of core units performs parallel processing of calculating and updating the voxel level.

Description

3D level set segmentation method and apparatus for medical image {Level Set Segmentation method and apparatus for Medical Image}

The present invention relates to a 3D printing application technology, and more particularly, it minimizes the size of memory required for the execution of the 3D level set division algorithm so that the 3D level set division algorithm can be performed regardless of the image size by using the multi-core of the GPU. It relates to a method and apparatus for dividing a 3D level set for a medical image that can be implemented quickly.

The 3D printing application market is growing year by year, and its application in the medical field is expected to continue to increase and form the largest market in the future. Already, 3D printing is being used for models for surgical planning or educational models for medical staff or medical students for specific cases, and its utilization is increasing.

In order to produce such a model for surgical simulation or a model for education in the medical field, a process of creating a 3D model by segmenting a medical image is required. There are several methods for segmenting the MRI image among the medical images. FIG. 1 schematically illustrates a 3D level set segmentation algorithm, which is one of the methods. Referring to FIG. 1 , an object is divided by a level set division method by receiving medical image data such as MRI, and the divided object is used as 3D data for 3D model generation.

Although the above-described 3D level set division method is very useful, there is a problem that it takes a long time to implement and run in a CPU (Central Processing Unit) due to a large amount of calculation.

More specifically, the medical image is composed of 3D volume data, and the 3D Level Set Segmentation algorithm requires several 3D memories in the calculation process, and the total memory takes 10G to 20G Byte or more.

2 illustrates a segmentation process according to the 3D Level Set Segmentation algorithm. If the level of the target pixel is greater than 0, which is a reference value, it is determined as an object, and if it is less than or equal to 0, the background (background) ) is judged as Accordingly, when an initial seed point is set in a medical image, a high level is set at the seed point, and the level is extended to match the image to find and segment an object region.

When implementing the above 3D level set partitioning algorithm in GPU, all arrays of level, velocity function, and other variables necessary for calculation are allotted to the memory of GPU, and multi-cores of GPU each have one 3D coordinate is allocated and the level is calculated for the corresponding coordinates by referring to the arrays stored in the memory to update the level. In this process, if it is necessary to refer to adjacent voxel values such as gradients, the data stored in the memory is referred to.

This is a method of deriving a region to the final level generated by repeating this process. When this is implemented, the array of variables required for level, velocity function, and other calculations are all composed of a three-dimensional array of the size of the original image, so the required memory becomes very large. There was a problem in that the size of the original image that can be processed was limited.

Accordingly, in the prior art, it was urgently required to develop a technique capable of minimizing the size of a memory for performing a 3D level set division algorithm, thereby enabling the 3D level set division algorithm to be performed regardless of an image size.

Korean Patent Publication No. 10-2018-0080786 (2018.07.13.) Korean Patent Registration No. 10-0613106 (2006.08.09) Korean Patent Registration No. 10-1471646 (2014.12.04.)

The present invention provides a 3D level for medical imaging that minimizes the memory required for performing the 3D level set segmentation algorithm and enables the execution of the 3D level set segmentation algorithm to be performed quickly using the multi-core of the GPU regardless of the image size. An object of the present invention is to provide a set dividing method and apparatus.

To this end, according to one aspect of the present invention, in a 3D level set segmentation method for a medical image,

allocating indexes for structure information recorded in level set data vector information to each of a plurality of core units that are multi-cores of the GPU;

Each of the plurality of core units allocated with the index reads structure information on a corresponding voxel recorded in structure information corresponding to the allocated index and structure information on voxels surrounding the voxel, and calculates a level of the corresponding voxel based on the structure information and updating the level of the corresponding voxel to the calculated level; and

deriving the object region with the updated voxel levels;

The plurality of core units that are the multi-core units may parallelly process the level calculation and update of the voxels.

In addition, the structure information is characterized in that it is a variable for calculating the coordinates and level of a voxel belonging to a predetermined region of interest.

In addition, receiving medical image information, a velocity function, and a seed point, deriving an object region at a level from the seed point, and setting a region of interest extended by a predetermined distance from the object region ; may be further included.

In addition, whenever the level is updated, the object region and the region of interest are reset.

In addition, when the ROI is set, generating structure information on voxels in the ROI, and recording the generated structure information in level set data vector information;

The method may further include; recording index information that is an access address of a corresponding structure at a location of index array information corresponding to coordinates of voxels corresponding to the structure information recorded in the level set data vector information.

In addition, the structure is composed of a vector, and the index arrangement information is characterized in that the index information is recorded to correspond to the coordinates of the voxel of the 3D image information.

According to another aspect of the present invention, in an apparatus for dividing a 3D level set for a medical image,

a memory unit for storing level set data vector information for recording structure information; and

A plurality of core units that are internal multi-cores are assigned indexes for structure information recorded in the level set data vector information,

When the index is assigned, structure information on the corresponding voxel recorded in the structure information corresponding to the index and the voxels surrounding the voxel is read from the level set data vector information, and the level of the corresponding voxel is determined based on the structure information calculated, and updating the level of the voxel with the calculated level;

GPU for deriving the object region with the updated voxel levels;

The plurality of core units that are the multi-core units may parallelly process the level calculation and update of the voxels.

In addition, the structure information is characterized in that it is a variable for calculating the coordinates and level of a voxel belonging to a predetermined region of interest.

In addition, the GPU receives medical image information, a velocity function, and a seed point, derives an object region at a level from the seed point, and extends the region of interest by a predetermined distance from the object region. It is characterized by setting.

Also, the GPU resets the object region and the ROI whenever the level is updated.

In addition, when the region of interest is set, the GPU generates structure information for voxels of the region of interest, records the generated structure information in the level set data vector information, and records the generated structure information in the level set data vector information. It is characterized in that index information, which is an access address of the corresponding structure, is recorded at a position of the index array information corresponding to the coordinates of voxels corresponding to the structure information.

In addition, the structure is composed of a vector, and the index arrangement information is characterized in that the index information is recorded to correspond to the coordinates of the voxel of the 3D image information.

The present invention minimizes the size of memory required for the execution of the 3D level set division algorithm, thereby enabling the execution of the 3D level set division algorithm to be performed quickly using the multi-core of the GPU regardless of the image size.

1 is a diagram schematically illustrating a general 3D level set division process.
2 is a diagram illustrating a general 3D level set segmentation algorithm.
3 is a diagram showing the configuration of a 3D level set dividing apparatus according to a preferred embodiment of the present invention.
4 is a view showing a detailed configuration of a 3D level set dividing apparatus according to a preferred embodiment of the present invention.
5 is a diagram illustrating the structure of structure information according to a preferred embodiment of the present invention.
6 is a flowchart of a 3D level set segmentation method according to a preferred embodiment of the present invention.

The present invention minimizes the size of memory required for performing the 3D level set division algorithm, so that the 3D level set division algorithm can be performed using multiple cores of the GPU regardless of the image size.

The present invention configures and stores the values of the original medical image and the variables that require calculation only for the region of interest used when dividing the object among the 3D data of the original medical image as members of the structure information, and stores each of these structure information. The index information, which is the access address for the medical image, is recorded in the index arrangement information composed of 3D data corresponding to the original medical image and used when dividing the 3D level set, thereby minimizing the size of memory required to perform the 3D level set division algorithm. In particular, the structure information has a structure of a vector (VECTOR).

An apparatus and method for dividing a 3D level set according to the present invention will be described in detail with reference to the drawings.

<Configuration of 3D level set dividing device>

3 and 4 show the configuration of a 3D level set dividing apparatus according to a preferred embodiment of the present invention.

The 3D level set division apparatus is composed of the GPU 100 and the memory unit 200, finds the boundary between the object and the background in the current level, expands the region by a specific voxel in the boundary region, sets the region of interest, Add the voxels of the region to the struct vector. Next, the previously created structure vector is equally set in the memory unit 200 , data is copied, and calculation is performed in the GPU 100 . Also, the GPU 100 includes a plurality of core units 1041 to 104N that are multi-core. Accordingly, the plurality of core units 1041 to 104N of the GPU 100 are assigned to respective structure vector items to perform calculations and store the changed levels in x, y, and z positions defined in the structure. Therefore, as a result, GPU parallel processing is performed on the regions of interest and the resulting level is saved, and when one stage is finished, it goes back to the first stage, finds the region of interest at the changed level, and performs calculations with GPU parallel processing for the region of interest. The process is repeated to obtain the final result.

In more detail, when a seed (SEED) point is set, the GPU 100 derives an object region based on the level of the seed point, and sets a region of interest that is extended to a certain distance from the derived object region. and generates structure information composed of variables that require calculating the values and levels of the original medical image for the voxels of the region of interest. In this case, each of the plurality of core units 1041 to 104N calculates a level for the voxel of the ROI and re-derives the object region based on the calculated level by repeating in parallel and repeating the process. When the final level is determined through this process, the GPU 100 derives an object region based on the final level. As described above, in the present invention, since only information for level update and derivation of the object region is recorded in the memory unit 200 for the region of interest, the memory requirement can be significantly reduced. In addition, since the plurality of core units 1041 to 104N parallelly process the level calculation process for voxels, it is possible to quickly process 3D medical image information.

The memory unit 200 stores various information including the processing program of the GPU 100 , and in particular stores 3D medical image information, level set data vector information, and index arrangement information. The 3D medical image information is information about the original medical image. In addition, the level set data vector information is a structure that is vector information composed of variables requiring level calculation and original medical image values for voxels included in the region of interest extended by a predetermined distance from the object region derived based on the level of the seed point. information is recorded. In this way, variables in the original image value and calculation are composed of structure information rather than each array type, and the structure is set as a vector.

For example, as shown in FIG. 5 , the structure information includes voxel position information (int x; int y; int z;) and variables necessary for level calculation (float phi; vec3f dphi; float g; vec3 dg; float). dphi_norm; vec3f N;). The index arrangement information corresponds to 3D medical image information, and index information that is access address information of level set data vector information in which structure information for each of the coordinates (x, y, z) of voxels of the 3D medical image information is recorded is recorded

Now, a 3D level set dividing method applicable to the 3D level set dividing apparatus of the present invention will be described with reference to the drawings.

<Procedure of 3D level set division method>

6 is a flowchart of a 3D level set segmentation method according to a preferred embodiment of the present invention. Referring to FIG. 6, when the medical image information, the velocity function, and the seed point are set (steps 300, 302, and 304), the GPU 100 derives the object area at the current level, which is the level at the seed point, and A region of interest extended by a predetermined distance from the region is set (step 306).

When the region of interest is set, the GPU 100 generates structure information on voxels of the region of interest and records the generated structure information in level set data vector information (step 308). The structure information is composed of voxel position information (int x; int y; int z;) and variables necessary for level calculation (float phi; vec3f dphi; float g; vec3 dg; float dphi_norm; vec3f N;). Here, the variables required for level calculation are already known, and detailed descriptions thereof will be omitted.

Thereafter, the GPU 100 records index information, which is an access address of the corresponding structure information, at a position of the index arrangement information corresponding to the coordinates of the voxels corresponding to the structure information recorded in the level set data vector information (step 310).

Thereafter, the GPU 100 allocates an index for the structure information recorded in the level set data vector information to each of the plurality of core units 1041 to 104N that are multi-core (step 312). Each of the plurality of core units 1041 to 104N, which is a multi-core to which the index is assigned, reads structure information about a voxel recorded in structure information corresponding to the corresponding index from a memory and structure information on voxels surrounding the voxel, and the structure information The level of the corresponding voxel is newly calculated based on the values (step 314). Here, since the process of calculating the level of the voxel is already known, a detailed description thereof will be omitted.

As described above, when the level calculation of the voxel is completed, the levels of the voxels corresponding to each multi-core are updated to the calculated level (step 316).

If the level calculation and update of the voxels have not been performed a predetermined number of times, that is, the level of all voxels set as the region of interest has not been updated, the process returns to step 306 and repeats the level update for the remaining voxels. .

Contrary to this, if the level of the voxel is calculated and updated a predetermined number of times, that is, the levels of all voxels set as the ROI are updated, an object region is derived based on the finally updated level value (step 320).

As described above, according to the present invention, data of all coordinates ((float phi; vec3f dphi; float g; vec3 dg; float dphi_norm; vec3f N;) is not recorded in the memory in a three-dimensional array form, but only the data of the coordinates of the region of interest are in a vector form. Also, for level calculation, the vector index of each coordinate is separately stored in the index array information to refer to the surrounding voxel values, and when referring to the data of the desired coordinate, the corresponding coordinate After referring to the index of , it is possible to refer to the desired data by referring to the corresponding index in the level set data vector.

The GPU level set segmentation algorithm according to the present invention is not limited by the size of the medical image and can be driven at a very high speed using the parallel processing of the GPU. It is expected to help in creation and contribute to various uses such as surgery planning and education.

As described above, the technical ideas described in the embodiments of the present invention may be implemented independently or in combination with each other. In addition, although the present invention has been described through the embodiments described in the drawings and detailed description of the invention, these are merely exemplary, and those of ordinary skill in the art to which the present invention pertains may make various modifications and equivalent other embodiments therefrom. possible. Accordingly, the technical protection scope of the present invention should be defined by the appended claims.

100 : GPU
1041 ~ 104N: Multiple core parts
200: memory unit

Claims (12)

In the 3D level set segmentation method for a medical image,
(a) Medical image information, velocity function, and seed point are set, an object region is derived at a level from the seed point, and a region of interest extended by a predetermined distance from the derived object region setting;
(b) allocating indexes for structure information recorded in level set data vector information to each of a plurality of core units that are multi-cores of the GPU;
(c) each of the plurality of core units allocated with the index reads structure information on a corresponding voxel recorded in structure information corresponding to the allocated index and structure information on voxels surrounding the voxel, and, based on the structure information, of the corresponding voxel calculating a level and updating a level of a corresponding voxel with the calculated level; and
(d) deriving an object region with the updated voxel levels;
The method of dividing a 3D level set for a medical image, wherein the plurality of core units that are multi-core perform parallel processing of calculating and updating the level of the voxel. According to claim 1,
The 3D level set segmentation method for a medical image, wherein the structure information is a variable for calculating coordinates and a level of a voxel belonging to a predetermined region of interest. delete According to claim 1,
Each time the level is updated, the object region and the region of interest are reset. According to claim 1,
(a-1) when the region of interest is set in step (a), generating structure information for voxels of the region of interest, and recording the generated structure information in level set data vector information; and
(a-2) recording the index information that is the access address of the structure at the location of the index array information corresponding to the coordinates of the voxels corresponding to the structure information recorded in the level set data vector information; A 3D level set segmentation method for a characterized medical image. 6. The method of claim 5,
The structure consists of a vector,
The method of dividing a 3D level set for a medical image, wherein the index arrangement information records index information to correspond to the coordinates of voxels of the 3D image information. A 3D level set dividing apparatus for a medical image, comprising:
a memory unit for storing level set data vector information for recording structure information; and
A plurality of core units that are internal multi-cores are assigned indexes for structure information recorded in the level set data vector information,
When the index is assigned, structure information on a corresponding voxel recorded in structure information corresponding to the index and surrounding voxels of the voxel is read from the level set data vector information, and the level of the corresponding voxel is determined based on the structure information. calculated, and updating the level of the voxel with the calculated level;
GPU for deriving the object region with the updated voxel levels;
The GPU receives medical image information, a velocity function, and a seed point, derives an object region at a level from the seed point, and extends the region of interest by a predetermined distance from the derived object region. set,
The apparatus for dividing a 3D level set for a medical image, wherein the plurality of core units, which are multi-core, parallelly process the level calculation and update of the voxels. 8. The method of claim 7,
The apparatus for dividing a 3D level set for a medical image, wherein the structure information is a variable for calculating coordinates and a level of a voxel belonging to a predetermined region of interest. delete 8. The method of claim 7,
3D level set dividing apparatus for a medical image, characterized in that the GPU resets the object region and the region of interest whenever the level is updated. 8. The method of claim 7,
the GPU generates structure information on voxels of the region of interest when the region of interest is set, and records the generated structure information in level set data vector information;
3D level set division for a medical image, characterized in that index information that is an access address of a corresponding structure is recorded at a position of index array information corresponding to coordinates of voxels corresponding to structure information recorded in the level set data vector information Device. 12. The method of claim 11,
The structure consists of a vector,
The apparatus for dividing a 3D level set for a medical image, wherein the index arrangement information records index information corresponding to the coordinates of voxels of the 3D image information.

Images