KR102842416B1 - The method for constructing a modality-agnostic AI model using synthetic data and the brain image segmentation method using the constructed AI model - Google Patents

Abstract

The present invention relates to a method for segmenting brain images using virtual data, and more particularly, to a method for constructing a universal modality artificial intelligence model using virtual data and a method for segmenting brain images using the constructed artificial intelligence model. To this end, a method for constructing a universal modality artificial intelligence model using virtual data is provided, characterized in that it includes a step (S100) of inputting a label map (100) of a medical image including at least one region related to a disease of the human body; a step (S120) of generating a plurality of transformed images by transforming the label map (100) by an image generation model (120); a step (S140) of training a universal modality agnostic artificial intelligence model (200) based on at least one of the medical image, the label map (100), and the plurality of transformed images; and a step (S160) of determining that the universal modality artificial intelligence model (200) has been constructed when an error between the predicted label map (220) output by the universal modality artificial intelligence model (200) and the label map (100) is within a predetermined range.

Description

Method for constructing a general-purpose modality-agnostic AI model using synthetic data and a brain image segmentation method using the constructed AI model

The present invention relates to a method for segmenting brain images using virtual data, and more specifically, to a method for constructing a general-modality artificial intelligence model using virtual data and a method for segmenting brain images using the constructed artificial intelligence model.

In general, various forms of image data or image data are used in the fields of brain imaging or neuroimaging. For example, Figure 1a is an image of an amyloid scan, which is used to diagnose neurodegenerative diseases such as Alzheimer's disease. Figures 1b to 1d also show examples of various neuroimaging images used to study the brain. These brain images play a crucial role in volumetric measurements, morphological analysis, connectivity, physiology, and molecular studies.

Figure 2a is an example of a high-resolution T1-weighted image for medical research purposes, and Figure 2b is an example of a low-resolution MRI scan image that can be acquired in clinical settings. As illustrated in Figures 2a and 2b, high-resolution T1-weighted images for medical research purposes can be used to quantitatively measure the brain using various analysis software (e.g., FreeSurfer, SPM, FSL, Heuron AD, etc.). However, the number of images is limited to 4,000 to 5,000, which is a disadvantage, and the subjects are limited to Caucasian subjects.

On the other hand, low-resolution MRI scan images that can be acquired clinically, such as those in Figure 2b, are characterized by low resolution and different imaging settings at each hospital, although more than 10 million brain images are taken annually. These clinical brain images vary in imaging settings, contrast, resolution, and image size at each hospital. Even when taken with the same modality (e.g., X-ray, CT, MRI, etc.), there are cases where the results can differ when taken at different hospitals.

[Table 1] is a table comparing the characteristics of processing and analysis of conventional image data.

method speed Processing and analysis of image data (modality agnostic) manual --- +++ Multi-atlas segmentation
(Multi-Atlas segmentation) - + Bayesian partitioning
(Bayesian Segmentation) + ++ Supervised Convolutional Neural Network
(Supervised CNN) +++ ---

Therefore, when attempting to build an artificial intelligence model using these clinical brain images through supervised learning, there was a disadvantage in that the cost of data collection was high because each brain image had to be manually segmented.

Moreover, conventional label map segmentation methods, which mostly rely on convolutional neural networks (CNNs), suffer from domain gap issues and have limitations in generalizing to data with different resolutions or MR contrasts. Generalization is even poor when using data acquired with different parameters or hardware within the same modality. While attempts to address this issue using techniques like data argumentation require retraining for new MRs, this requires labeled data and incurs additional costs.

Furthermore, for registration, conventional methods estimated the deformation field between two images, but this method faced optimization issues and the hassle of re-estimating for each new image. Furthermore, not only did training take a long time, but deep learning-based registration also suffered from the impact of learned MR contrast and modality.

Republic of Korea Patent Registration No. 10-2466479 (Tau protein accumulation prediction device and tau protein accumulation prediction method using the same).

Accordingly, the present invention has been devised to solve the above problems, and the problem to be solved by the present invention is to increase the amount of data by modifying virtual data related to brain images, and to build an artificial intelligence model with excellent inference ability for segmentation and analysis by training artificial intelligence with the increased brain image data.

That is, the purpose of the present invention is to provide a method for constructing a general modality artificial intelligence model using virtual data and a method for segmenting brain images using the constructed artificial intelligence model.

However, the technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

In order to achieve the above technical task, a method for constructing a universal modality artificial intelligence model using virtual data is provided, characterized by including a step (S100) of inputting a label map (100) of a medical image including at least one area related to a disease of a human body; a step (S120) of generating a plurality of transformed images by transforming the label map (100) by an image generation model (120); a step (S140) of training a universal modality agnostic artificial intelligence model (200) based on at least one of the medical image, the label map (100) and the plurality of transformed images; and a step (S160) of determining that the universal modality artificial intelligence model (200) has been constructed when an error between the expected label map (220) output by the universal modality artificial intelligence model (200) and the label map (100) is within a predetermined range.

In addition, the areas related to human diseases include at least one of the frontal lobe, parietal lobe, temporal lobe, occipital lobe, ventricle, areas related to Alzheimer's disease, ARIA-E (Amyloid-Related Imaging Abnormalities with Effusion or Edema) area, ARIA-H (Amyloid-Related Imaging Abnormalities with Hemosiderin) area, WMH (White Matter Hyperintensities) area, lacune area occurring in deep brain structures such as white matter or basal ganglia, cerebral microbleeds, gray matter area, thalamus area, caudate area, putamen area, skull area, and cerebrospinal fluid (CSF) area.

Additionally, the human body's disease can be one of dementia, Alzheimer's disease, vascular dementia, brain atrophy, Parkinson's disease, stroke, brain tumor, epilepsy, multiple sclerosis, amyloid vaccine side effect disease, and brain cancer.

Additionally, the image generation model (120) can generate multiple transformed images using a generative adversarial network model (GANM).

In addition, the image generation model (120) can generate multiple transformed images using a spatial transformation method that applies at least one of a rigid transformation method, an affine transformation method, and a non-linear transformation method.

Additionally, the image generation model (120) can generate multiple transformed images using a Gaussian Mixture Model (GMM) or a specific rule-based method in which data is generated according to a predetermined defined rule.

Additionally, the image generation model (120) can generate multiple transformed images using an artifact application method or a noise application method.

Additionally, the image generation model (120) can generate multiple transformed images by changing the resolution of the label map (100).

In addition, the plurality of transformed images may include a generated image (140) and a transformed label map (160), and the learning step (S140) may include a step (S142) in which a general modality artificial intelligence model (200) outputs an expected label map (220) based on the label map (100) and the generated image (140); a step (S144) in which an average dice loss unit (170) calculates an average from the similarity between the segmented regions based on the expected label map (220) and the transformed label map (160) and the dice loss; and a step (S146) in which a backpropagation unit (180) updates weights and biases to minimize an error between the expected label map (220) and the input label map (100).

Additionally, the general modality artificial intelligence model (200) may be a convolutional neural network (CNN) model or a Vision Transformer (ViT) or a mixed model thereof.

In addition, the label map (100) in the input step (S100) may be a plurality of label maps (100), and some of the label maps (100) among the plurality of label maps (100) may be a label map (100) in which the skull region is processed as the background of the region of interest (ROI) or a label map (100) in which the outer cerebrospinal fluid (Outer CSF) region is processed as the background of the region of interest (ROI).

Additionally, the label map (100) is a label map based on at least one of original medical images such as X-ray, MRI, and CT.

Additionally, the label map (100) is a label map of virtual data generated based on at least one of original medical images such as X-ray, MRI, and CT.

The above-described object of the present invention can also be achieved by a brain image segmentation method using a universal modality artificial intelligence model using virtual data, which is characterized by including, as another embodiment, a step (S200) of inputting a new medical image of the brain into a universal modality artificial intelligence model (200) constructed by the above-described method; and a step (S220) of the universal modality artificial intelligence model (200) inferring and outputting a label map based on the new medical image of the brain.

In addition, in the input stage (S200), the new medical image is an image that is different from the label map (100) input in the artificial intelligence model construction stage in at least one of resolution, size, shooting position, contrast, and format.

According to one embodiment of the present invention, a high-precision artificial intelligence model can be built by training artificial intelligence using a large amount of virtual data.

Furthermore, even when labeled brain images are input from different hospitals, with different imaging settings, resolution, size, format, or contrast, the inference results can be obtained with high segmentation robustness for the label map. Furthermore, the AI model constructed according to the present invention has the advantage of demonstrating accurate segmentation performance without error, even when modality changes.

However, the effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects that are not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present invention belongs from the description below.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the invention described below, serve to further understand the technical idea of the present invention, and therefore, the present invention should not be interpreted as being limited to matters described in such drawings.
Figures 1a to 1d are examples of various neuroimaging images used to study the brain.
Figure 2a is an example of a high-resolution T1-weighted image for medical research purposes.
Figure 2b is an example of a low-resolution MRI scan image that can be acquired in a clinical setting.
Figures 3a to 3d are examples of virtual synthetic data of the brain used in a method for constructing a universal modality artificial intelligence model using virtual data according to the present invention.
Figure 4 is a block diagram schematically showing a method for building a general modality artificial intelligence model using virtual data according to the present invention.
Figure 5 is a block diagram showing the process of outputting virtual synthetic data of the brain from an arbitrary label image in Figure 4.
Figure 6 is a block diagram of another representation of the block diagram shown in Figure 4.
Figure 7a is an example of a label map for the frontal lobe, parietal lobe, temporal lobe, occipital lobe, and ventricular region of the brain.
Figure 7b is an example of a label map for gray matter, white matter, thalamus, caudate nucleus, putamen, skull, and cerebrospinal fluid regions of the brain.
Figure 8a is an example of a label image input without modification in the present invention.
Figure 8b is an example of an incomplete label image in which the skull region is processed as the background of the region of interest (ROI) in the present invention.
Figure 8c is an example of a fully labeled image in which the outer CSF region is processed as the background of the region of interest (ROI) in the image shown in Figure 8b.
Figure 9 is a flowchart schematically showing a method for constructing a universal modality artificial intelligence model using virtual data according to the present invention and a method for segmenting brain images using the constructed artificial intelligence model.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings so that those skilled in the art can easily practice the present invention. However, the description of the present invention is merely an embodiment for structural and functional explanation, and therefore the scope of the present invention should not be construed as being limited by the embodiments described in the text. In other words, since the embodiments can be modified in various ways and can have various forms, the scope of the present invention should be understood to include equivalents that can realize the technical idea. In addition, the purposes or effects presented in the present invention do not mean that a specific embodiment must include all of them or only such effects, and therefore the scope of the present invention should not be construed as being limited thereby.

The meanings of terms described in the present invention should be understood as follows.

Terms such as "first" and "second" are intended to distinguish one component from another, and the scope of the rights should not be limited by these terms. For example, a first component could be referred to as a second component, and similarly, a second component could also be referred to as a first component. When a component is referred to as being "connected" to another component, it should be understood that it may be directly connected to that other component, but there may also be other components in between. Conversely, when a component is referred to as being "directly connected" to another component, it should be understood that there are no other components in between. Meanwhile, other expressions describing the relationship between components, such as "between" and "immediately between" or "adjacent to" and "directly adjacent to", should be interpreted similarly.

Singular expressions should be understood to include plural expressions unless the context clearly indicates otherwise, and terms such as "comprises" or "has" should be understood to specify the presence of stated features, numbers, steps, operations, components, parts, or combinations thereof, but not to exclude the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this invention pertains. Terms defined in commonly used dictionaries should be interpreted to be consistent with their meaning within the context of the relevant technology, and should not be interpreted as having ideal or overly formal meanings unless explicitly defined herein.

How to build a general-purpose AI model

Hereinafter, the configuration of a preferred embodiment will be described in detail with reference to the attached drawings. FIGS. 3A to 3D are examples of virtual synthetic data of the brain used in a method for constructing a universal modality artificial intelligence model using virtual data according to the present invention, FIG. 4 is a block diagram schematically showing a method for constructing a universal modality artificial intelligence model using virtual data according to the present invention, and FIG. 9 is a flowchart schematically showing a method for constructing a universal modality artificial intelligence model using virtual data and a method for segmenting brain images using the constructed artificial intelligence model according to the present invention.

As illustrated in FIGS. 4 and 9, first, a label map (100) including one or more regions related to a disease of the human body is input (S100). Examples of the input label map (100) are as shown in FIGS. 3A to 3D. The label map (100) is a label data structure mainly used in tasks such as image segmentation or object detection, and is a map that expresses which class (or category) each pixel or object belongs to. In other words, it is a map containing class information for each part of a brain image or 3D data. The label map (100) is a map that assigns a class label to each pixel, voxel, or object of a brain image or 3D data, and is mainly used in tasks such as image segmentation, object detection, and medical image analysis. This clearly defines which class each part belongs to, and plays an important role in learning and performance evaluation. In particular, in medical imaging, when pixels of a brain MRI image need to be divided into classes such as gray matter, white matter, and cerebrospinal fluid, the label map (100) is an array that displays the label of each class (tissue) as a number such as 0, 1, 2, etc. for each pixel of the image.

In addition, the label map (100) in the input step (S100) is a plurality of label maps (100). For example, FIG. 7a is an example of a label map for the frontal lobe, parietal lobe, temporal lobe, occipital lobe, and ventricular region of the brain, and FIG. 7b is an example of a label map for the gray matter, white matter, thalamus, caudate nucleus, putamen, skull, and cerebrospinal fluid regions of the brain.

And, FIG. 8a is an example of a label image input without modification in the present invention, FIG. 8b is an example of an incomplete label image in which a skull region is processed as a background of a region of interest (ROI) in the present invention, and FIG. 8c is an example of a complete label image in which an outer CSF region is processed as a background of a region of interest (ROI) in the image shown in FIG. 8b.

As shown in FIGS. 8A to 8C, some of the label maps (100) among the plurality of label maps (100) input in this manner are label maps (100) in which the skull region is processed as the background of the region of interest (ROI) or label maps (100) in which the outer cerebrospinal fluid (Outer CSF) region is processed as the background of the region of interest (ROI).

In addition, the label map (100) is a label map of virtual data generated based on at least one of original medical images such as X-ray, MRI, and CT, and is preferably based on a 3D T1 image among MRI images. Specifically, the label map (100) refers to a 2D, 3D, or 4D image, and is a label map generated from MRI, X-ray, CT, T1-weighted image of MRI, T2-weighted image of MRI, etc.

And, the region related to the human body's disease includes at least one of the frontal lobe, parietal lobe, temporal lobe, occipital lobe, ventricle, region related to Alzheimer's disease, ARIA-E (Amyloid-Related Imaging Abnormalities with Effusion or Edema) region, ARIA-H (Amyloid-Related Imaging Abnormalities with Hemosiderin) region, WMH (White Matter Hyperintensities) region, lacune region occurring in deep brain structures such as white matter or basal ganglia, cerebral microbleeds, gray matter region, thalamus region, caudate region, putamen region, skull region, and cerebrospinal fluid (CSF) region.

And, the human body's disease can be one of dementia, Alzheimer's disease, vascular dementia, brain atrophy, Parkinson's disease, stroke, brain tumor, epilepsy, multiple sclerosis, amyloid plaque side effect disease, and brain cancer.

Next, the image generation model (120) transforms the label map (100) to generate multiple transformed images (S120). There are several ways to transform the label map (100), as follows.

For example, the image generation model (120) can generate multiple transformed images using a generative adversarial network model (GANM). The generative adversarial network model (GANM) is a generative model designed with a structure in which two neural networks compete with each other to learn.

In addition, the image generation model (120) can generate multiple transformed images by using a spatial transformation method that applies at least one of a rigid transformation method, an affine transformation method, and a non-linear transformation method. The rigid transformation method is a method that simply transforms the position and direction without translation, rotation, or scale transformation while maintaining the shape of an object or image, and is used when rotating an image or moving it to a specific position.

Affine transformations are transformations that can encompass translation, rotation, scaling, and twisting of images or objects. Affine transformations preserve the linear nature of lines, but can alter angles and length ratios. In other words, parallel lines remain, but they can be transformed so that they are no longer perpendicular. For example, an image of the brain can be scaled up or down by a certain ratio, or a tilt applied.

Nonlinear transformation (NMT) is a method of transforming coordinates in a nonlinear manner, rather than a linear one. It can be used to apply curved distortions or complex nonlinear transformations to brain images. For example, it is possible to distort parts of a brain image or elongate only certain areas.

Additionally, the image generation model (120) can generate multiple transformed images using a Gaussian Mixture Model (GMM) or a specific rule-based method in which data is generated according to a predetermined defined rule.

A Gaussian mixture model (GMM) is a data modeling technique that combines multiple Gaussian distributions. It is a sampling technique used to describe or generate data by representing complex distributions as mixtures of multiple, simpler Gaussian distributions.

Furthermore, specific rule-based sampling is a method of selecting or generating data based on specific rules or criteria. This specific rule-based method contrasts with random sampling and is used to select data that satisfies specific conditions. Rule-based sampling is a method in which data is generally selected or generated based on defined rules or algorithms.

In addition, the image generation model (120) can generate multiple transformed images by using an artifact application method or a noise application method. Artifacts mainly refer to abnormal or unintended errors, distortions, or abnormal phenomena that occur in medical imaging, data processing, image processing, or signal processing. Artifacts can interfere with analysis and can mainly occur due to equipment, environmental factors, or data processing. An example of such artifacts is a phenomenon in which images are distorted or have errors due to various factors such as limitations of the equipment or movement of the patient in medical imaging such as MRI, CT, and X-ray.

Additionally, the image generation model (120) can generate multiple transformed images by changing the resolution of the label map (100).

The generated multiple transformed images include a generated image (140) and a transformed label map (160), as illustrated in FIG. 4. The generated image (140) is input to a general modality artificial intelligence model (200), and the transformed label map (160) is input to an average die loss unit (170).

Next, a general modality agnostic artificial intelligence model (200) is trained using the generated image (S140). The specific detailed process is as follows. First, the general modality agnostic artificial intelligence model (200) outputs an expected label map (220) based on the label map (100) and the generated image (140) (S142). This general modality agnostic artificial intelligence model (200) can be a convolutional neural network (CNN) model, a Vision Transformer (ViT), or a hybrid model thereof.

Next, the average dice loss (170) calculates the average from the dice loss and the similarity between the segmented areas based on the expected label map (220) and the transformed label map (160) (S144).

Next, the backpropagation unit (180) updates the weights and biases to minimize the error between the expected label map (220) and the input label map (100) (S146).

During the learning process, if the error between the expected label map (220) output by the general modality artificial intelligence model (200) and the input label map (100) is within a predetermined range, the general modality artificial intelligence model (200) is determined to have been constructed (S160), and learning is terminated.

Fig. 5 is a block diagram showing a process of outputting virtual synthetic data of a brain from an arbitrary label image in Fig. 4, and Fig. 6 is a block diagram of another representation of the block diagram shown in Fig. 4. As shown in Figs. 5 and 6, virtual synthetic data is generated using an arbitrary label, and the generation module can generate synthetic brain image data using one or more of rule-based, machine learning, or artificial intelligence-based methods.

Brain image segmentation method using an established artificial intelligence model

Hereinafter, the operation of a preferred embodiment will be described in detail with reference to FIG. 9. First, a new medical image of the brain is input into the general-purpose modality artificial intelligence model (200) constructed using the aforementioned method (S200). At this time, the new medical image in the input step (S200) is an image that differs from the label map (100) input in the artificial intelligence model construction step in at least one of the following: resolution, size, shooting position, contrast, and format.

Next, the general-purpose modality artificial intelligence model (200) infers and outputs a label map based on new medical images of the brain (S220). Through this process, a label map with robust segmentation can be obtained from the input label image.

The detailed description of the preferred embodiments of the present invention disclosed above has been provided to enable those skilled in the art to implement and practice the present invention. While the above description has been made with reference to preferred embodiments of the present invention, those skilled in the art will appreciate that various modifications and variations can be made to the present invention without departing from the scope of the present invention. For example, those skilled in the art can utilize the individual components described in the above-described embodiments in combination with each other. Accordingly, the present invention is not intended to be limited to the embodiments described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

The present invention may be embodied in other specific forms without departing from the spirit and essential characteristics thereof. Therefore, the above detailed description should not be construed in any way as limiting but rather as illustrative. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all changes coming within the equivalent scope of the claims are intended to be included therein. The present invention is not intended to be limited to the embodiments set forth herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein. Furthermore, claims that are not explicitly cited in the claims may be combined to form an embodiment or incorporated into a new claim by post-application amendment.

100: Label Map,
120: Image generation model,
140: Generated image,
160: Deformed image,
170: Average Dice Loss,
180: Back Propagation,
200: General-purpose modality artificial intelligence model,
220: Expected label map.
230: Synthetic brain imaging data output unit.

Claims (15)

It is performed by a computer,
A step (S100) of receiving a label map (100) of a medical image including one or more areas related to a disease of the human body;
A step (S120) in which the image generation model (120) transforms the label map (100) to generate multiple transformed images;
A step (S140) of training a universal modality agnostic artificial intelligence model (200) based on at least one of the above medical image, the label map (100) and the plurality of transformed images;
A step (S160) of determining that the universal modality artificial intelligence model (200) has been constructed when the error between the predicted label map (220) output by the universal modality artificial intelligence model (200) and the label map (100) is within a predetermined range;
The above image generation model (120) generates the plurality of transformed images using a spatial transformation method that applies at least one of a rigid transformation method and a non-linear transformation method, and
The above multiple transformed images include a generated image (140) and a transformed label map (160),
The above learning step (S140) is
A step (S142) in which the general modality artificial intelligence model (200) outputs the expected label map (220) based on the label map (100) and the generated image (140);
A step (S144) in which the average dice loss unit (170) calculates an average from the similarity between the segmented areas based on the predicted label map (220) and the modified label map (160) and the dice loss; and
The backpropagation unit (180) includes a step (S146) of updating weights and biases to minimize the error between the expected label map (220) and the input label map (100);
In the above input step (S100), the label map (100) is a plurality of label maps (100), and
A method for constructing a general modality artificial intelligence model using virtual data, characterized in that some of the label maps (100) among the above plurality of label maps (100) are label maps (100) in which the skull region is treated as the background of the region of interest (ROI) or label maps (100) in which the external cerebrospinal fluid (Outer CSF) region is treated as the background of the region of interest (ROI). In the first paragraph,
A method for building a general modality artificial intelligence model using virtual data, characterized in that the region related to the above human disease includes at least one of the frontal lobe, parietal lobe, temporal lobe, occipital lobe, ventricle, region related to Alzheimer's disease, ARIA-E (Amyloid-Related Imaging Abnormalities with Effusion or Edema) region, ARIA-H (Amyloid-Related Imaging Abnormalities with Hemosiderin) region, WMH (White Matter Hyperintensities) region, lacune region occurring in deep brain structures such as white matter or basal ganglia, cerebral microbleeds, gray matter region, thalamus region, caudate region, putamen region, skull region, and cerebrospinal fluid (CSF) region. In the first paragraph,
A method for constructing a universal modality artificial intelligence model using virtual data, wherein the above human disease is one of dementia, Alzheimer's disease, vascular dementia, brain atrophy, Parkinson's disease, stroke, brain tumor, epilepsy, multiple sclerosis, amyloid plaque side effect disease, and brain cancer. In the first paragraph,
The above image generation model (120) is a method for constructing a general modality artificial intelligence model using virtual data, characterized in that it generates the plurality of transformed images using a generative adversarial network model (GANM). delete In the first paragraph,
The above image generation model (120) is a method for constructing a general modality artificial intelligence model using virtual data, characterized in that it generates the plurality of transformed images using a Gaussian Mixture Model (GMM) or a specific rule-based method in which data is generated according to a predetermined defined rule. In the first paragraph,
The above image generation model (120) is a method for constructing a universal modality artificial intelligence model using virtual data, characterized in that it generates the plurality of transformed images using an artifact application method or a noise application method. In the first paragraph,
A method for constructing a universal modality artificial intelligence model using virtual data, characterized in that the image generation model (120) generates the plurality of transformed images by changing the resolution of the label map (100). delete In the first paragraph,
A method for constructing a general modality artificial intelligence model using virtual data, characterized in that the general modality artificial intelligence model (200) is a convolutional neural network (CNN) model or a Vision Transformer (ViT) or a mixed model thereof. delete In the first paragraph,
A method for constructing a universal modality artificial intelligence model using virtual data, characterized in that the above label map (100) is a label map based on X-ray, MRI, and CT images. In paragraph 12,
A method for constructing a universal modality artificial intelligence model using virtual data, characterized in that the above label map (100) is a label map of virtual data generated based on X-ray, MRI, and CT images. It is performed by a computer,
A step (S200) of inputting a new medical image of the brain into a general modality artificial intelligence model (200) constructed by any one of claims 1 to 4, 6 to 8, 10, 12 and 13; and
The above general modality artificial intelligence model (200) includes a step (S220) of inferring and outputting a label map based on a new medical image of the brain; and
A brain image segmentation method using a universal modality artificial intelligence model using virtual data, characterized in that in the above input step (S200), the new medical image is an image having at least one different resolution, size, shooting position, contrast, and format from the label map (100) input in the above artificial intelligence model construction step. delete