KR20240044196A - Method and apparatus for multi-stage segmentation of organs based on semi-supervised learning - Google Patents

Abstract

A method and apparatus for identifying a target organ based on semi-supervised learning are disclosed. The method includes performing preprocessing on a first labeled dataset and a first unlabeled dataset for image data containing the target organ; Obtaining a second labeled dataset and a second unlabeled dataset for a region where the target organ is located among the image data, based on the preprocessed dataset; And inputting the second labeled data set and the second unlabeled data set into a first AI model and a second AI model, and obtaining an identification result of the target organization, wherein the first AI model And each of the second AI models includes a regression layer that performs a regression task and a split layer that performs a segmentation task, and the second AI model uses the EMA value of the weight used in the first AI model as a weight. A loss function for obtaining an identification result of the target organization can be constructed based on multi-scale data output through the regression layer and division layer of each of the first AI model and the second AI model.

Description

{METHOD AND APPARATUS FOR MULTI-STAGE SEGMENTATION OF ORGANS BASED ON SEMI-SUPERVISED LEARNING}

The present disclosure relates to a method and device for organ segmentation. More specifically, the present disclosure relates to a multi-level organ segmentation method and apparatus based on semi-supervised learning.

Accurate identification and segmentation of pancreatic segments through abdominal CT imaging can be used to understand the shape of the pancreas in the diagnosis, surgery, and treatment planning of pancreatic cancer.

However, the location and shape of the pancreas varies from patient to patient, and it is not easy to distinguish between the pancreas and surrounding organs. Accordingly, there is great technical difficulty in accurately identifying and segmenting the pancreas in abdominal CT images.

With the recent development of artificial intelligence-based image processing technology, a method of identifying the pancreas based on abdominal CT images is being introduced. Previously, the supervised learning method to identify the pancreas based on label data was widely used.

However, although label data plays an important role in the supervised learning method, there is a problem that there is currently no label data for analyzing the pancreas.

Registered Patent Publication No. 10-2202361, 2021.01.07

The purpose of an embodiment of the present disclosure is to provide a multi-level organ segmentation method and device based on semi-supervised learning.

Additionally, an embodiment of the present disclosure aims to provide a method and device for more accurately dividing and identifying organs that have large differences in shape or are located at different locations.

The problems to be solved by the present disclosure are not limited to the problems mentioned above, and other problems not mentioned can be clearly understood by those skilled in the art from the description below.

According to an embodiment of the present disclosure, a method of identifying a target organ based on semi-supervised learning, performed by a device, includes first information on image data including the target organ. performing preprocessing on a labeled dataset and a first unlabeled dataset; Obtaining a second labeled dataset and a second unlabeled dataset for a region where the target organ is located among the image data, based on the preprocessed dataset; And inputting the second labeled data set and the second unlabeled data set into a first AI model and a second AI model, and obtaining an identification result of the target organization, wherein the first AI model And each of the second AI models includes a regression layer that performs a regression task and a segmentation layer that performs a segmentation task, and the second AI model is used in the first AI model. Using the EMA (exponential moving average) value of the weight as a weight, based on multi-scale data output through the regression layer and division layer of each of the first AI model and the second AI model, A loss function may be configured to obtain an identification result of the target organization.

And, the step of performing the preprocessing includes performing intensity normalization, spacing normalization, and data augmentation on the first labeled dataset and the first labeled dataset. May include steps.

And, the intensity normalization is performed by cropping the intensity values of the first labeled dataset and the abdominal image included in the first labeled dataset by a predefined value and removing regions unrelated to the target organ. Afterwards, an operation to perform normalization may be included.

And, the step of acquiring the second labeled dataset and the second unlabeled dataset includes comparing the first labeled dataset on which the preprocessing was performed and ground truth data for the target organization. Based on the result, obtaining the second labeled dataset may be included.

And, the step of obtaining the second labeled dataset and the second unlabeled dataset includes using a pseudo label based on the first unlabeled dataset, It may include a step of acquiring a data set that is not yet available.

And, the loss function is a confidence map based on the first multi-scale probability data output through the division layer of the first AI model and the second multi-scale probability data output through the division layer of the second AI model. -It may be based on uncertainty data about the boundary surface of the target organ output based on scale probability data.

And, the loss function may be based on supervised learning error data obtained through the difference between the first multi-scale probability data and ground truth data.

And, the loss function is obtained through the difference between the first multi-scale distance map output through the regression layer of the first AI model and the second multi-scale distance map output through the regression layer of the second AI model. It can be based on inter-model error data.

And, the loss function may be based on inter-layer error data obtained through the difference between the first multi-scale probability data and the second multi-scale probability data.

In another embodiment of the present disclosure, an apparatus for identifying a target organ based on semi-supervised learning includes one or more memories; and one or more processors, wherein the one or more processors pre-process a first labeled dataset and a first unlabeled dataset for image data containing the target organ. Do; Based on the preprocessed dataset, obtain a second labeled dataset and a second unlabeled dataset for a region where the target organ is located among the image data; and inputting the second labeled data set and the second unlabeled data set into a first AI model and a second AI model to obtain an identification result of the target organization, wherein the first AI model and the second AI model are set to obtain an identification result of the target organization. Each of the second AI models includes a regression layer that performs a regression task and a segmentation layer that performs a segmentation task, and the second AI model has weights used in the first AI model. Using the EMA (exponential moving average) value as a weight, based on multi-scale data output through the regression layer and division layer of each of the first AI model and the second AI model, the target A loss function may be constructed to obtain the identification result of the institution.

And, the one or more processors may perform intensity normalization, spacing normalization, and data augmentation on the first labeled dataset and the first labeled dataset. .

And, the one or more processors obtain the second labeled dataset based on a result of comparing the first labeled dataset on which the preprocessing was performed and ground truth data for the target institution. You can.

And, the one or more processors may obtain the second unlabeled data set using a pseudo label based on the first unlabeled data set.

In addition to this, a computer program stored in a computer-readable recording medium for execution to implement the present disclosure may be further provided.

In addition, a computer-readable recording medium recording a computer program for executing a method for implementing the present disclosure may be further provided.

According to the above-described problem-solving means of the present disclosure, a multi-level organ segmentation method and device based on semi-supervised learning can be provided.

In addition, according to the means for solving the above-described problems of the present disclosure, it is possible to more accurately analyze the structure of the pancreas that has a large difference in shape or exists in different positions.

The effects of the present disclosure are not limited to the effects mentioned above, and other effects not mentioned may be clearly understood by those skilled in the art from the description below.

1 is a schematic diagram of a system for implementing a multi-level organ segmentation method based on semi-supervised learning, according to an embodiment of the present disclosure.
Figure 2 is a block diagram for explaining the configuration of a multi-level organ segmentation device based on semi-supervised learning, according to an embodiment of the present disclosure.
Figure 3 is a flowchart for explaining a multi-level organization division method based on semi-supervised learning, according to an embodiment of the present disclosure.
FIG. 4 is a diagram illustrating the configuration and operation of each model for implementing a multi-level organization segmentation method based on semi-supervised learning, according to an embodiment of the present disclosure.
Figure 5 is a diagram for explaining a method by which a device performs preprocessing, according to an embodiment of the present disclosure.
Figures 6 and 7 are diagrams for explaining the configuration and operation of a model applied to a multi-level organ segmentation method based on semi-supervised learning, according to an embodiment of the present disclosure.
Figure 8 is a diagram showing the results of multi-level organization segmentation based on semi-supervised learning, according to an embodiment of the present disclosure.

Like reference numerals refer to like elements throughout this disclosure. The present disclosure does not describe all elements of the embodiments, and general content or overlapping content between the embodiments in the technical field to which the present disclosure pertains is omitted. The term 'unit, module, member, block' used in the specification may be implemented as software or hardware, and depending on the embodiment, a plurality of 'unit, module, member, block' may be implemented as a single component, or It is also possible for one 'part, module, member, or block' to include multiple components.

Throughout the specification, when a part is said to be “connected” to another part, this includes not only direct connection but also indirect connection, and indirect connection includes connection through a wireless communication network. do.

Additionally, when a part "includes" a certain component, this means that it may further include other components rather than excluding other components, unless specifically stated to the contrary.

Throughout the specification, when a member is said to be located “on” another member, this includes not only cases where a member is in contact with another member, but also cases where another member exists between the two members.

Terms such as first and second are used to distinguish one component from another component, and the components are not limited by the above-mentioned terms.

Singular expressions include plural expressions unless the context clearly makes an exception.

The identification code for each step is used for convenience of explanation. The identification code does not explain the order of each step, and each step may be performed differently from the specified order unless a specific order is clearly stated in the context. there is.

Hereinafter, the operating principle and embodiments of the present disclosure will be described with reference to the attached drawings.

In this specification, 'multi-level organ segmentation device based on semi-supervised learning' includes various devices that can perform computational processing and provide results to the user. For example, the multi-level organ segmentation device based on semi-supervised learning according to the present disclosure may include all of a computer, a server device, and a portable terminal, or may take the form of any one.

Here, the computer may include, for example, a laptop, desktop, laptop, tablet PC, slate PC, etc. equipped with a web browser.

The server device is a server that processes information by communicating with external devices, and may include an application server, computing server, database server, file server, game server, mail server, proxy server, and web server.

The portable terminal is, for example, a wireless communication device that guarantees portability and mobility, such as PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), and PDA. (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), WiBro (Wireless Broadband Internet) terminal, smart phone ), all types of handheld wireless communication devices, and wearable devices such as watches, rings, bracelets, anklets, necklaces, glasses, contact lenses, or head-mounted-device (HMD). may include.

In explaining the present disclosure, “image data” may include, but is not limited to, a CT (computer tomography) image of a part of the patient's body.

In describing the present disclosure, “organ” may be a human (eg, a patient) or an animal, or a part or all of a human or animal. For example, the organ may include at least one of organs such as the liver, heart, uterus, brain, breast, abdomen, etc., blood vessels (eg, arteries or veins, etc.), fatty tissue, etc. In addition, “target organ” may refer to a part of the body that is the target of actual surgery or a part of the body to be identified/segmented through an AI model (e.g., pancreas, etc.).

1 is a diagram for implementing a multi-stage organ segmentation method based on semi-supervised learning (or a method of identifying a target organ based on semi-supervised learning) according to an embodiment of the present disclosure. This is a schematic diagram of system 1000.

As shown in Figure 1, the system 1000 for implementing a multi-level organ segmentation method based on semi-supervised learning includes a device 100, a hospital server 200, a database 300, and an AI model 400. It can be included.

Here, FIG. 1 illustrates the case where the device 100 is implemented in the form of a single desktop, but is not limited thereto. As described above, device 100 may refer to various types of devices or a group of devices in which one or more types of devices are connected.

The device 100, hospital server 200, database 300, and artificial intelligence (AI) model 400 included in the system 1000 can communicate through the network (W). . Here, the network W may include a wired network and a wireless network. For example, the network may include various networks such as a local area network (LAN), a metropolitan area network (MAN), and a wide area network (WAN).

Additionally, the network W may include the known World Wide Web (WWW). However, the network (W) according to an embodiment of the present disclosure is not limited to the networks listed above, and may include at least some of a known wireless data network, a known telephone network, and a known wired and wireless television network.

The device 100 may provide a multi-stage organ segmentation method based on semi-supervised learning (or a method of identifying a target organ based on semi-supervised learning).

For example, the device 100 may perform preprocessing on a first labeled dataset and a second unlabeled dataset of image data including the target organ. The device may acquire a third labeled dataset and a fourth unlabeled dataset for the area where the target organ is located among the image data, based on the data on which preprocessing has been performed. The device may input the third labeled dataset and the fourth unlabeled dataset into the AI model 400 to obtain an identification result of the target organization.

The operations performed by the above-described device will be described in detail with reference to FIGS. 2 to 8.

The hospital server 200 (eg, cloud server, etc.) may store image data (eg, computer tomography (CT) image) obtained by scanning the patient's body. The hospital server 200 may transmit stored image data to the device 100, the database 300, or the AI model 400.

The database 300 may store image data including the target organ obtained from the hospital server 200. Additionally, the database 300 may store various labeled data and unlabeled data generated by the device 100. Additionally, the database 300 can store various parameters necessary for learning and inference of the AI model 400.

The AI model 400 is a model learned to segment/identify specific organs in image data. The AI model may include a first AI model (eg, student model) and a second AI model (eg, teacher model).

Here, each of the first AI model and the second AI model may include a regression layer that performs a regression task and a segmentation layer that performs a segmentation task. That is, each of the first AI model and the second AI model may be composed of a plurality of layers that perform dual tasks. The configuration and operation method of each AI model will be described in detail with reference to FIGS. 2 to 8.

1 illustrates a case where the AI model 400 is implemented outside of the device 100 (e.g., implemented as cloud-based), but is not limited thereto and is a component of the device 100. It can be implemented as:

FIG. 2 is a block diagram illustrating the configuration of an apparatus 100 for identifying a target organ based on semi-supervised learning, according to an embodiment of the present disclosure.

As shown in FIG. 2 , device 100 may include memory 110, communication module 120, display 130, input module 140, and processor 150. However, it is not limited to this, and the software and hardware configuration of the device 100 may be modified/added/omitted depending on the required operation within the range obvious to those skilled in the art.

The memory 110 can store data supporting various functions of the device 100 and a program for the operation of the processor 150, and can store input/output data (e.g., music files, still images, video, etc.), a number of application programs (application programs or applications) running on the device, data for operation of the device 100, and commands can be stored. At least some of these applications may be downloaded from an external server via wireless communication.

The memory 110 is of a flash memory type, hard disk type, solid state disk type, SDD type (Silicon Disk Drive type), and multimedia card micro type. micro type), card type memory (e.g. SD or XD memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), EEPROM (electrically erasable) It may include at least one type of storage medium among programmable read-only memory (PROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, and optical disk.

Additionally, the memory 110 is separate from the device, but may include a database connected by wire or wirelessly. That is, the database shown in FIG. 1 may be implemented as a component of the memory 110.

The communication module 120 may include one or more components that enable communication with an external device, for example, at least one of a broadcast reception module, a wired communication module, a wireless communication module, a short-range communication module, and a location information module. may include.

Wired communication modules include various wired communication modules such as Local Area Network (LAN) modules, Wide Area Network (WAN) modules, or Value Added Network (VAN) modules, as well as USB (Universal Serial Bus) modules. ), HDMI (High Definition Multimedia Interface), DVI (Digital Visual Interface), RS-232 (recommended standard 232), power line communication, or POTS (plain old telephone service).

In addition to Wi-Fi modules and WiBro (Wireless broadband) modules, wireless communication modules include GSM (global System for Mobile Communication), CDMA (Code Division Multiple Access), WCDMA (Wideband Code Division Multiple Access), and UMTS (universal mobile telecommunications system). ), TDMA (Time Division Multiple Access), LTE (Long Term Evolution), 4G, 5G, 6G, etc. may include a wireless communication module that supports various wireless communication methods.

The display 130 displays (outputs) information processed by the device 100 (e.g., image data including target organs, various labeled datasets and unlabeled datasets, etc.). For example, the display may display execution screen information of an application (for example, an application) running on the device 100, or UI (User Interface) and GUI (Graphic User Interface) information according to such execution screen information. You can.

The input module 140 is for receiving information from the user. When information is input through the user input unit, the processor 150 can control the operation of the device 100 to correspond to the input information.

The input module 140 includes hardware-type physical keys (e.g., buttons, dome switches, jog wheels, jog switches, etc. located on at least one of the front, back, and sides of the device) and software-type physical keys. May include touch keys. As an example, the touch key consists of a virtual key, soft key, or visual key displayed on the touch screen type display 130 through software processing, or the above It may consist of a touch key placed in a part other than the touch screen. Meanwhile, the virtual key or visual key can be displayed on the touch screen in various forms, for example, graphic, text, icon, video or these. It can be made up of a combination of .

The processor 150 may control the overall operation and functions of the device 100. Specifically, the processor 150 has a memory that stores data for an algorithm for controlling the operation of components within the device 100 or a program that reproduces the algorithm, and performs the above-described operations using the data stored in the memory. It may be implemented with at least one processor (not shown). At this time, the memory and processor may each be implemented as separate chips. Alternatively, the memory and processor may be implemented as a single chip.

In addition, the processor 150 can control any one or a combination of the above-described components in order to implement various embodiments according to the present disclosure described in FIGS. 3 to 8 below on the device 100. You can.

FIG. 3 is a flowchart illustrating a method of identifying a target organ based on semi-supervised learning performed by an apparatus, according to an embodiment of the present disclosure.

The device may perform preprocessing on a first labeled dataset and a first unlabeled dataset for image data including the target organ (S310).

Specifically, the device may acquire image data containing the target organ. For example, when the target organ is the pancreas, the device may acquire abdominal CT image data including the pancreas from another device (eg, a hospital server, etc.).

Referring to FIG. 4, the device includes a first labeled dataset 410-1 and a first unlabeled dataset for image data (e.g., abdominal CT image data) containing a target organ (e.g., pancreas). You can obtain (410-2). Additionally, the device may perform preprocessing 420 on each data set 410-1 and 410-2.

Here, the first labeled dataset 410-1 may include image data including the target organ and correct answer data (i.e., ground truth data) for the target organ.

And, as shown in FIG. 4, each dataset 410-1 and 410-2 may include image data with a size of 512x512x181 and may be composed of 466 pixels, but is not limited thereto.

The device can adjust the intensity and spacing values of the entire dataset by performing preprocessing. The type of preprocessing performed by the device can be divided into intensity normalization, interval normalization, and data augmentation.

As an example, (min-max) intensity normalization may be performed by adjusting the intensity values of the first labeled dataset and the abdominal image included in the first labeled dataset by a predefined value (e.g., -100HU). It may include an operation of cropping (ranging from 0 to 240 HU), removing regions unrelated to the target organ, and then performing normalization (e.g., performing normalization to 0 to 1).

Interval normalization may resample the central pixel interval value of all image data (e.g., by 0.8594 mm ² ) in order to make the pixel interval of the image data included in each dataset the same.

Data augmentation involves random rotation (e.g., -20° to 20°), random translation (up to 20px translation on the x and y axes), and arbitrary scaling (e.g., 0.8 to 1.2 times) for the image data included in each dataset. ) can be performed. Accordingly, the device can increase the amount of data included in each dataset (for example, by about 10 times).

As an example, as shown in FIG. 5(a), the device acquires image data (i.e., CT image data) for the original abdomen and configures each dataset based on the acquired image data. can do.

The device can obtain image data as shown in (b) of FIG. 5 by performing intensity normalization on the original image. As another example, the device may obtain image data as shown in (c) of FIG. 5 by performing interval normalization on the original image.

The device may acquire a second labeled dataset and a second unlabeled dataset for the area where the target organ is located among the image data, based on the preprocessed dataset (S320).

Here, the area where the target organ is located among the image data may mean an area within a bounding box that divides the area around the target organ (that is, an area that divides the surrounding area around the target organ). And, the second labeled dataset may mean a localized labeled data set, and the second unlabeled data set may mean a localized unlabeled data set.

As an example, the device may acquire a second labeled data set based on a result of comparing the first labeled data set on which preprocessing was performed and ground truth data for the target organ. Specifically, the device may obtain a second labeled dataset by applying a crop and resize algorithm to the image data based on the area where the target organ is located.

As another example, the device may obtain a second unlabeled data set using a pseudo label based on the first unlabeled data set. Specifically, the device may obtain a second unlabeled dataset by applying a cropping and resizing algorithm to the capital label based on the first unlabeled dataset.

Referring to FIG. 4, the device creates a specific model by comparing the first labeled dataset on which preprocessing was performed and ground truth data for the target organ (e.g., ground truth data included in the first labeled dataset) 450. (430) can be learned. And, the device may acquire a second labeled dataset 410-3 based on the comparison result (eg, probability of dividing/identifying a specific organ) 440-1.

Then, the device applies a crop and resize algorithm to the number label 440-2 based on the first unlabeled dataset 410-2 to create a second unlabeled dataset 410-2. 4) can be obtained.

The device may input the second labeled data set and the second unlabeled data set into the first AI model and the second AI model to obtain an identification result of the target organization (S330).

Here, the first AI model may be implemented as a student model, and the second AI model may be implemented as a teacher model. That is, a mean-teacher model may be constructed based on the first AI model and the second AI model.

Referring to FIG. 4, each of the first AI model 460-1 and the second AI model 460-2 includes regression layers 470-2 and 470-3 that perform a regression task and segmentation. It may include segmentation layers 470-1 and 470-4 that perform a segmentation task. That is, the first AI model 460-1 and the second AI model 460-2 may be composed of layers that perform dual tasks.

And, output through the regression layers (470-2, 470-3) and division layers (470-1, 470-4) of the first AI model (460-1) and the second AI model (460-2), respectively. A loss function (or cost function) may be configured to obtain an identification result of the target organization based on multi-scale data.

Here, the loss function refers to a function for outputting the error between the value output using the first AI model 460-1 and the second AI model 460-2 and the output value desired by the user of the device.

The loss function (L _total ) for obtaining the identification result of the target organization can be implemented as shown in Equation 1. In Equation 1, α may be 0.1, but is not limited thereto.

Here, the supervised learning error data (L _supervised ) is the difference between the first multi-scale probability data (480-3) and the ground truth data (480-5) output through the division layer (470-1) of the first AI model. It may be a value obtained through Supervised learning error data (L _supervised (X,Y)) can be configured as shown in Equation 2.

는 제1 AI 모델의 파라미터(또는, 가중치)를 의미한다. 는 바이너리 크로스 엔트로피 손실을 의미한다. S는 멀티 스케일 확률 값의 개수를 의미한다. N 및 M 각각은 레이블된 데이터 셋 및 레이블되지 않은 데이터셋의 개수를 의미한다. ₀, ₁, ₂, _s는 각각 0.4, 0.3, 0.2, 0.1일 수 있으나, 이에 제한되는 것은 아니다.In Equation 2, L _DICE means DICE loss, and can be used in a semantic segmentation network structure where data imbalance characteristics exist. f _seg refers to the output function of the split layer. It is not limited to this.

means the parameters (or weights) of the first AI model. stands for binary cross entropy loss. S refers to the number of multi-scale probability values. N and M refer to the number of labeled and unlabeled data sets, respectively. ₀ , _One , ₂ , _s may be 0.4, 0.3, 0.2, and 0.1, respectively, but is not limited thereto.

Inter-model error data (L _inter-model ) is the first multi-scale distance map (480-4) output through the regression layer (470-2) of the first AI model and the regression layer (470-4) of the second AI model. It may be a value obtained through the difference between the second multi-scale distance map 480-1 output through 3). Inter-model error data (L _inter-model (X)) can be structured as shown in Equation 3.

tea는 제2 AI 모델의 파라미터(또는, 가중치)를 의미한다.In Equation 3, L _MSE means the mean square error loss value. f _reg refers to the output function of the regression layer.

tea means the parameters (or weights) of the second AI model.

Uncertainty data (L _uncertainty ) about the boundary of the target organization is a reliability map based on the first multi-scale probability data (480-3) output through the division layer (470-2) of the first AI model ( It may be a value output based on the confidence map) 480-6 and the second multi-scale probability data 480-2 output through the division layer 470-4 of the second AI model.

Here, the reliability map may refer to plot data indicating the probability of being a specific institution within the image data.

Uncertainty data (L _uncertainty (X, C, Y)) about the boundary of the target organization can be configured as shown in Equation 4.

In Equation 4, C may mean a reliability map.

Inter-layer error data (or inter-task error data (L _inter-task ) is obtained through the difference between the first multi-scale probability data 480-3 and the second multi-scale probability data 480-2. The inter-layer error data may be structured as shown in Equation 5.

The device can learn the first AI model and the second AI model again using the acquired identification result of the target organization (490-1) and the separate capital label (490-2).

Figures 6 and 7 are diagrams for explaining the configuration and operation of the first AI model and the second AI model according to an embodiment of the present disclosure.

Referring to FIG. 6, the first AI model (student model) may be learned based on the input data 610-1 and the label 620-2. For example, the first AI model may be trained so that the error (eg, classification loss) between the input data 610-1 and the label 620-2 is reduced. At this time, specific noise (η) may be applied on the hidden layer included in the first AI model.

The second AI model (teacher model) may use the EMA (exponential moving average) value (θ') of the weight (θ) used in the first AI model as a weight. In addition, the first AI model and the second AI model may be learned so that the error value (e.g., consistency cost) of the output value of each of the first AI model and the second AI model is reduced.

The first AI model and the second AI model may be composed of an uncertainty aware multi-scale measurement network as shown in FIG. 7.

Figure 8 is a diagram showing the results of identifying a target organ in image data, according to an embodiment of the present disclosure. As shown in FIG. 8, the device can output the results of identifying the target organ in a total of four cases (i.e., four types of image data).

At this time, the device has 1) a model learned based on image data and ground truth, 2) a model learned based on the mean-teacher model, 3) a model learned based on the mean-teacher model and the UA-MT network, 4 ) a model learned based on the mean-teacher model, UA-MT network, and dual-task, 5) a model learned based on the mean-teacher model, UA-MT network, dual-task, and refinement, respectively. You can output the results of identifying the target organization using .

Meanwhile, the disclosed embodiments may be implemented in the form of a recording medium that stores instructions executable by a computer. Instructions may be stored in the form of program code, and when executed by a processor, may create program modules to perform operations of the disclosed embodiments. The recording medium may be implemented as a computer-readable recording medium.

Computer-readable recording media include all types of recording media storing instructions that can be decoded by a computer. For example, there may be read only memory (ROM), random access memory (RAM), magnetic tape, magnetic disk, flash memory, optical data storage device, etc.

As described above, the disclosed embodiments have been described with reference to the attached drawings. A person skilled in the art to which this disclosure pertains will understand that the present disclosure may be practiced in forms different from the disclosed embodiments without changing the technical idea or essential features of the present disclosure. The disclosed embodiments are illustrative and should not be construed as limiting.

100: device
110: memory
120: communication module
130: display
140: input module
150: processor

Claims (18)

In a method for identifying a target organ based on semi-supervised learning, performed by a device, the method includes:
performing preprocessing on a first labeled dataset and a first unlabeled dataset for image data including the target organ;
Obtaining a second labeled dataset and a second unlabeled dataset for a region where the target organ is located among the image data, based on the preprocessed dataset; and
Inputting the second labeled data set and the second unlabeled data set into a first AI model and a second AI model to obtain an identification result of the target organization,
Each of the first AI model and the second AI model includes a regression layer that performs a regression task and a segmentation layer that performs a segmentation task,
The second AI model uses the EMA (exponential moving average) value of the weight used in the first AI model as a weight,
A loss function for obtaining an identification result of the target organization based on multi-scale data output through the regression layer and division layer of each of the first AI model and the second AI model. How it is constructed. According to paragraph 1,
The step of performing the preprocessing is,
The method comprising performing intensity normalization, spacing normalization, and data augmentation on the first labeled dataset and the first labeled dataset. According to paragraph 2,
The intensity normalization is,
An operation of cropping the first labeled dataset and intensity values of the abdominal image included in the first labeled dataset by a predefined value, removing regions unrelated to the target organ, and then performing normalization. Method, including. According to paragraph 1,
Obtaining the second labeled dataset and the second unlabeled dataset includes:
A method comprising obtaining the second labeled dataset based on a result of comparing the first labeled dataset on which the preprocessing was performed and ground truth data for the target organ. According to paragraph 4,
Obtaining the second labeled dataset and the second unlabeled dataset includes:
Obtaining the second unlabeled data set using a pseudo label based on the first unlabeled data set. According to paragraph 1,
The loss function is,
Based on a confidence map based on the first multi-scale probability data output through the division layer of the first AI model and the second multi-scale probability data output through the division layer of the second AI model A method based on uncertainty data about the output interface of the target organ. According to clause 6,
The loss function is,
A method based on supervised learning error data obtained through differences between the first multi-scale probability data and ground truth data. In clause 7,
The loss function is,
Based on inter-model error data obtained through the difference between the first multi-scale distance map output through the regression layer of the first AI model and the second multi-scale distance map output through the regression layer of the second AI model. , method. According to clause 8,
The loss function is,
A method based on inter-layer error data obtained through a difference between the first multi-scale probability data and the second multi-scale probability data. In the device for identifying a target organ based on semi-supervised learning, the device includes:
One or more memories; and
Contains one or more processors,
The one or more processors:
perform preprocessing on a first labeled dataset and a first unlabeled dataset for image data containing the target organ;
Based on the preprocessed dataset, obtain a second labeled dataset and a second unlabeled dataset for a region where the target organ is located among the image data; and
Input the second labeled data set and the second unlabeled data set into a first AI model and a second AI model to obtain an identification result of the target organization,
Each of the first AI model and the second AI model includes a regression layer that performs a regression task and a segmentation layer that performs a segmentation task,
The second AI model uses the EMA (exponential moving average) value of the weight used in the first AI model as a weight,
A loss function for obtaining an identification result of the target organization based on multi-scale data output through the regression layer and division layer of each of the first AI model and the second AI model. The device consists of: According to clause 10,
The one or more processors:
An apparatus for performing intensity normalization, spacing normalization, and data augmentation on the first labeled dataset and the first labeled dataset. According to clause 11,
The intensity normalization is,
An operation of cropping the first labeled dataset and intensity values of the abdominal image included in the first labeled dataset by a predefined value, removing regions unrelated to the target organ, and then performing normalization. Device, including. According to clause 10,
The one or more processors:
Apparatus for obtaining the second labeled dataset based on a result of comparing the first labeled dataset on which the preprocessing was performed and ground truth data for the target organ. According to clause 13,
The one or more processors:
Apparatus, obtaining the second unlabeled data set using a pseudo label based on the first unlabeled data set. According to clause 10,
The loss function is,
Based on a confidence map based on the first multi-scale probability data output through the division layer of the first AI model and the second multi-scale probability data output through the division layer of the second AI model A device based on the output uncertainty data about the boundary surface of the target organ. According to clause 15,
The loss function is,
An apparatus based on supervised learning error data obtained through differences between the first multi-scale probability data and ground truth data. According to clause 16,
The loss function is,
Based on inter-model error data obtained through the difference between the first multi-scale distance map output through the regression layer of the first AI model and the second multi-scale distance map output through the regression layer of the second AI model. , Device. According to clause 17,
The loss function is,
Based on inter-layer error data obtained through the difference between the first multi-scale probability data and the second multi-scale probability data.

Images