KR20230172937A - Medical image segmentation method and apparatus - Google Patents

Abstract

A method and device for segmenting medical images into regions are disclosed. The region segmentation device inputs the medical image into the region segmentation model to obtain a predicted value in pixel or voxel units, applies a Gaussian mixture model to the distribution of the predicted value of the medical image, generates multiple Gaussian distributions, and divides the region between the multiple Gaussian distributions. The boundary is set as a threshold for region division, and the threshold and predicted values are compared to identify pixels or voxels corresponding to the segmented region in the medical image.

Description

Medical image segmentation method and apparatus {Medical image segmentation method and apparatus}

Embodiments of the present invention relate to a method and device for segmenting a region in a medical image, and more specifically, to a method and device for segmenting human tissue from a medical image using an artificial intelligence model.

There are various methods to image the internal tissues of the human body, such as X-ray imaging, CT (Computed Tomography), and MRI (Magnetic Resonance Imaging). X-ray imaging provides two-dimensional medical images, and CT and MRI provide three-dimensional medical images. Various human tissues such as bones, organs, and lesions are present in medical images. Segmentation of specific areas in medical images is necessary for various purposes, such as 3D printing of human tissue and lesion diagnosis. Artificial intelligence models are being used to segment human tissue. However, the existing artificial intelligence model has a fixed threshold for region segmentation, so there is a problem in that the quality of region segmentation of medical images varies for each patient.

The technical task to be achieved by embodiments of the present invention is to provide a method and device for segmenting regions from a medical image using an artificial intelligence model applying an optimal threshold for each medical image.

In order to achieve the above technical problem, an example of a method for segmenting medical images according to an embodiment of the present invention is a method for segmenting medical images performed by a region dividing device, where the medical image is input into a region partition model. obtaining a predicted value in pixel or voxel units; Generating a plurality of Gaussian distributions by applying a Gaussian mixture model to the distribution of predicted values of the medical image; Setting a boundary between the plurality of Gaussian distributions as a threshold for region division; and comparing the threshold and the predicted value to identify a pixel or voxel corresponding to the segmented area in the medical image.

In order to achieve the above technical problem, an example of a region segmentation device according to an embodiment of the present invention includes a prediction unit that obtains a prediction value in pixel or voxel units for a medical image through a region partition model; a distribution determination unit that generates a plurality of Gaussian distributions by applying a Gaussian mixture model to the distribution of predicted values of the medical image; a threshold setting unit that sets a threshold based on a point where the plurality of Gaussian distributions meet; and a region determination unit that compares the threshold value and the predicted value to determine a pixel or voxel corresponding to a segmented region in the medical image.

According to an embodiment of the present invention, the quality of region segmentation can be improved by applying the optimal threshold value for each medical image to an artificial intelligence model that divides the region of a medical image.

1 is a diagram showing an example of a medical image region segmentation device according to an embodiment of the present invention;
Figure 2 is a diagram showing an example of a region division model according to an embodiment of the present invention;
3 is a flowchart showing an example of a region division method according to an embodiment of the present invention;
Figure 4 is a diagram expressing the main body of predicted values of the region partition model according to an embodiment of the present invention as a heat map;
Figure 5 is a diagram expressing the distribution of predicted values of the region partition model according to an embodiment of the present invention as a histogram;
Figure 6 is a diagram showing the results of a region division method according to an embodiment of the present invention, and
Figure 7 is a diagram showing the configuration of an example of a region partition device according to an embodiment of the present invention.

Hereinafter, a method and device for segmenting areas of a medical image according to an embodiment of the present invention will be described in detail with reference to the attached drawings.

1 is a diagram illustrating an example of a medical image region dividing device according to an embodiment of the present invention.

Referring to FIG. 1 , the region division device 100 performs region division 130 on the medical image 120 through the region division model 110. The medical image 120 may be a two-dimensional or three-dimensional medical image, and for example, may be a CT or MRI image. The region division model 110 is an artificial intelligence model that receives the medical image 120 and outputs a predicted value in pixel or voxel units. The area division model 110 is reviewed again in FIG. 2.

Figure 2 is a diagram illustrating an example of a region division model according to an embodiment of the present invention.

Referring to FIG. 2, the region division model 110 is an artificial intelligence model that divides the region of human tissue (various organs, lesions, etc.) upon receiving the medical image 200. For example, the region division model 110 can be implemented through various artificial neural networks such as CNN (Convolutional Neural Network).

The region division model 110 outputs a predicted value in pixel or voxel units of the medical image 200. For example, if the region division model 110 is a model that classifies regions in a two-dimensional medical image composed of pixels, the region division model 110 outputs a predicted value in pixels, and the region division model 110 divides the region into voxels. If it is a model that divides a region in a constructed 3D medical image, the region division model 110 outputs a voxel-wise predicted value. The region segmentation model 110 may be any one of various existing medical image segmentation models. That is, various conventional region division models that output predicted values in pixel or voxel units can be applied to this embodiment. It is assumed that the region division model 110 has been trained in advance.

The predicted value 210 output by the region segmentation model 110 is a value indicating the probability of whether a pixel or voxel corresponds to the segmentation region. For example, the region division model 110 may output a predicted value between 0 and 1 for each pixel or voxel. The region division device may identify pixels or voxels with a predicted value above a certain level as a division region based on the predicted value in pixel or voxel units output by the region division model 110. The value for dividing the partition area (i.e., the threshold 220) can be set in advance by the user or defined during the learning process of the region partition model 110. However, since the characteristics of medical images for each patient are different, the accuracy of the region segmentation results for each medical image may vary when the threshold value predetermined in the region segmentation model 110 is applied to medical images of various patients. Therefore, it is necessary to improve the results of region segmentation (230) by applying the optimal threshold (220) for each patient, considering the characteristics of medical images for each patient. The optimal threshold 220 for each medical image is not generated through a separate learning process, but is determined based on the predicted value output by the region partition model, as will be discussed below.

Figure 3 is a flowchart showing an example of a region division method according to an embodiment of the present invention.

Referring to FIG. 3, the region segmentation device 100 obtains a pixel or voxel-level predicted value for a medical image through a region partition model (S300). An example of a region division model is shown in Figure 2.

The region division device 100 generates a plurality of Gaussian eye distributions from the distribution of predicted values for medical images (S310). The distribution of predicted values indicates the frequency of pixels corresponding to each predicted value. The region dividing device 100 can identify a plurality of Gaussian distributions by applying a Gaussian Mixture Model to the distribution of predicted values. The Gaussian mixture model is an algorithm that expresses the probability distribution as the sum of several Gaussian distributions. Since the Gaussin-an mixture model itself is already widely known, further explanation about it will be omitted.

Since the predicted value is a value that represents the probability that each pixel or voxel corresponds to the segmentation area, a Gaussian mixture model can be applied to the distribution of the predicted value. For example, the region dividing device 100 generates a histogram for the predicted value of the medical image as shown in FIG. 5, generates a distribution curve of the predicted value through this, and then applies a Gaussian mixture model to obtain a plurality of Gaussian distributions. can do.

The region division device 100 sets a threshold based on the boundaries of a plurality of Gaussian distributions (S320). For example, the region dividing device 100 may set one of the values of a plurality of boundary points where a plurality of Gaussian models meet as a threshold. The region dividing device 100 may obtain the average of each Gaussian model, obtain the value of the boundary point using the re-averaged value of each Gaussian model, and set this as a threshold. In addition, various methods of calculating the threshold value by considering various statistical values (variance, average, weight, etc.) of each Gaussian model can be applied to this embodiment.

The types of predicted values output by the region partition model are TP (True Positive), where the actual value is true and the predicted value is also true, FP (False Positive), where the actual value is false but the model predicts to be true, and the actual value is false and the model is also false. It can be classified as predicted FN (False Negative) (if predicted value = 0), etc. In the case of TP and FP, the predicted value is not 0, and in the case of FN, the predicted value is 0, so the boundary between TP and FP can be set as the optimal threshold of the region partition model. To this end, the region division device 100 may express the distribution of the predicted value as two Gaussian distributions and set the boundary point between the two Gaussian distributions as a threshold. An example of this is shown in Figure 5.

When the threshold is set, the region segmentation device 100 compares the predicted value output by the region segmentation model with the threshold and identifies an area composed of pixels or voxels with a predicted value greater than the threshold as a segmented area (S330). The quality of region division can be improved by setting the optimal threshold value according to the predicted value output by the region division model and using it for region division.

Figure 4 is a diagram expressing the distribution of predicted values of the region partition model according to an embodiment of the present invention as a heat map, and Figure 5 is a diagram expressing the distribution of predicted values of the region partition model according to an embodiment of the present invention as a histogram.

Referring to FIGS. 4 and 5 together, when the region segmentation device 100 receives the medical image 400, it can input the medical image 400 into a region segmentation model to obtain a predicted value for each pixel (or voxel). In Figure 4, the distribution of predicted values for each pixel (or voxel) is expressed as a heatmap 410, and in Figure 5, the distribution of predicted values for each pixel is expressed as a histogram 500.

The region dividing device 100 may obtain a plurality of Gaussian distributions by applying a Gaussian distribution model (GMM) to the distribution of predicted values. As an example, the region segmentation device 100 may determine the distribution curve of the predicted value through a histogram representing the number of pixels corresponding to the predicted value between 0 and 1, and obtain a plurality of Gaussian distributions from the distribution curve of the predicted value. The size of each section of the histogram may vary depending on the embodiment. The two Gaussian distributions may correspond to the distribution of predicted values corresponding to FP and the distribution of predicted values corresponding to TP, respectively. That is, the distribution of predicted values can be divided into a TP area 510 and an FP area 520, respectively, based on the areas of the two Gaussian distributions. In the example of Figure 5, each Gaussian distribution is displayed as a dotted line and the curve representing the sum of two Gaussian distributions is displayed as a solid line. In sections where the dotted and solid lines are the same, only the solid lines are indicated.

This embodiment shows an example in which the distribution of the predicted values is expressed as two Gaussian distributions, assuming that the types of predicted values are largely TP and FP. However, this is just one example, and if the types of predicted values are divided into three or more types, the distribution of predicted values can be expressed as a plurality of Gaussian distributions corresponding to the number of types of predicted values. Hereinafter, for convenience of explanation, it is assumed that the distribution of predicted values is represented by two Gaussian distributions.

The region dividing device 100 sets the value of the boundary point where two Gaussian distributions meet as a threshold. For example, the region dividing device 100 may obtain the average value (512,514) of each Gaussian distribution and set the average value (520) of the average values (512,514) of the two Gaussian distributions as the threshold. The point where two Gaussian distributions meet can be obtained by various methods based on the statistical values of each Gaussian distribution, and is not limited to the average value of this embodiment. For example, the region dividing device 100 may set the lowest point in the center of a curve (solid line) representing the sum of two Gaussian distributions as the threshold.

Figure 6 is a diagram showing the results of a region division method according to an embodiment of the present invention.

Referring to FIG. 6, the result 610 of dividing a region in the medical image 600 using a predefined threshold in the region division model, and the medical image 600 by applying the optimal threshold according to this embodiment. ), the result of dividing the area (620) is shown. It can be seen that when the region is divided by applying the optimal threshold value of this embodiment, the region can be divided more precisely.

Figure 7 is a diagram showing the configuration of an example of a region partition device according to an embodiment of the present invention.

Referring to FIG. 7 , the region division device 100 includes a prediction unit 700, a distribution determination unit 710, a threshold value setting unit 720, and a region determination unit 730. The area division device 100 of this embodiment may be implemented with various types of computing devices including memory, processor, and input/output device. In this case, each configuration may be implemented in software, loaded into memory, and then performed by a processor.

The prediction unit 700 obtains a prediction value in pixel or voxel units for the medical image through a region partition model. The region segmentation model is an artificial intelligence model that divides a region in a medical image. Any model that outputs a predicted value in pixel or voxel units as shown in FIG. 2 can be applied to this embodiment. The area division model may be any one of various existing models.

The distribution determination unit 710 generates a plurality of Gaussian distributions by applying a Gaussian mixture model to the distribution of the predicted value of the medical image identified through the region partition model. For example, the distribution determination unit 710 may generate a plurality of Gaussian distributions from a histogram representing the distribution of predicted values, as shown in FIG. 5. In another embodiment, the distribution determination unit 710 may express the distribution of the predicted value as two Gaussian distributions. An example of this is shown in Figure 5.

The threshold setting unit 720 sets the threshold based on the point where a plurality of Gaussian distributions meet. For example, when two Gaussian distributions are obtained from the distribution of predicted values, the threshold setting unit 720 may set the threshold from the boundary area where the two Gaussian distributions meet, as shown in FIG. 5.

The region determination unit 730 compares the threshold value and the predicted value to identify pixels or voxels corresponding to the segmented region in the medical image. For example, the region determination unit 730 may identify an area consisting of pixels or voxels with a prediction value greater than the threshold set by the threshold setting unit 720 as a segmented region.

The present invention can also be implemented as computer-readable program code on a computer-readable recording medium. Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices. Additionally, computer-readable recording media can be distributed across networked computer systems so that computer-readable code can be stored and executed in a distributed manner.

So far, the present invention has been examined focusing on its preferred embodiments. A person skilled in the art to which the present invention pertains will understand that the present invention may be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered from an illustrative rather than a restrictive perspective. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the equivalent scope should be construed as being included in the present invention.

Claims (6)

In the area division method of medical images performed by the area division device,
Inputting a medical image into a region segmentation model to obtain a predicted value in pixel or voxel units;
Generating a plurality of Gaussian distributions by applying a Gaussian mixture model to the distribution of predicted values of the medical image;
Setting a boundary between the plurality of Gaussian distributions as a threshold for region division; and
Comparing the threshold and the predicted value to identify pixels or voxels corresponding to the segmented area in the medical image. The method of claim 1, wherein generating the Gaussian distribution comprises:
Representing the distribution of the predicted values as a histogram; and
A method for segmenting medical images comprising: generating at least two Gaussian distributions from the distribution of predicted values determined through the histogram. The method of claim 1, wherein setting the threshold value comprises:
calculating the average of each of the plurality of Gaussian distributions;
A method for segmenting medical images comprising: setting a re-averaged value of the plurality of Gaussian distributions as a threshold value. A prediction unit that obtains a pixel- or voxel-level prediction value for a medical image through a region division model;
a distribution determination unit that generates a plurality of Gaussian distributions by applying a Gaussian mixture model to the distribution of predicted values of the medical image;
a threshold setting unit that sets a threshold based on a point where the plurality of Gaussian distributions meet; and
A region segmentation device comprising: a region determination unit that compares the threshold value and the predicted value to identify pixels or voxels corresponding to the segmentation region in the medical image. The method of claim 4, wherein the distribution detection unit,
A region dividing device characterized by generating a plurality of Gaussian distributions from a histogram representing the distribution of the predicted values. A computer-readable recording medium recording a computer program for performing the method according to any one of claims 1 to 3.

Images