Method and apparatus for segmenting medical image

Abstract

A method for training a network by at least one processor comprises the steps of: obtaining a synthetic dataset including a synthetic image and segmentation data of the synthetic image; and training, using the synthetic dataset, a segmentation network to output the result of segmenting an input image in a clinical sense. The segmentation network is implemented to perform domain generalized semantic segmentation by learning a reliable synthetic image region by using a spatial weight related to the reliability of the synthetic image.

Description

Medical image segmentation method and device

[0001] The present disclosure relates to medical image analysis technology.

[0002] Cardiac image segmentation is an essential procedure for diagnosing cardiovascular diseases, through which clinically important biomarkers such as left ventricular (LV) thickness, LV volume, and ejection fraction can be obtained. Currently, this is a very cumbersome and time-consuming task that requires radiologists to work for several hours, leading to high interest in automated cardiac image segmentation technology.

[0003] Recently, deep learning technology has demonstrated the efficiency of image segmentation in various fields, and in cardiac image segmentation as well, fully automated segmentation using neural networks is expected to contribute to rapid and effective diagnosis. However, these data-driven cardiac image segmentation methods suffer from a domain shift problem, in which network performance degrades significantly for data with a distribution different from the training data.

[0004] In the case of medical ultrasound, image characteristics vary significantly depending on various factors, such as the equipment manufacturer, probe specifications, and ultrasound pulse frequency. Such image variation can severely impact diagnostic accuracy and lead to fatal consequences. Consequently, there is a growing demand for domain generalization-based semantic segmentation technologies.

[0005] The present disclosure relates to a medical image segmentation method and apparatus.

[0006] Specifically, the present disclosure relates to domain generalization and semantic image segmentation for medical image analysis.

[0007] According to some embodiments, a network training method by at least one processor comprises the steps of: acquiring a synthetic dataset including a synthetic image and segmentation data of said synthetic image; and training a segmentation network using said synthetic dataset to output a result of segmenting an input image into clinical semantics. The segmentation network is implemented to perform domain-generalized semantic segmentation by learning a reliable synthetic image region using spatial weights related to the reliability of said synthetic image.

[0008] Inaccurate regions of the synthetic image can be excluded from training by the above spatial weights.

[0009] The above spatial weights may be provided by a meta-network. The meta-network may be parameterized to perform spatial uncertainty-awareness on an input image based on meta-learning.

[0010] The above synthetic dataset may include synthetic images generated by varying at least one of the content condition and style condition input to the generative model.

[0011] The above content conditions may include segmented data representing an image Region of Interest (ROI) and anatomical features.

[0012] The above style condition may include at least one parameter representing image characteristics.

[0013] The above parameters may include at least one of the Nakagami scale parameter and the Nakagami distribution parameter.

[0014] The above segmentation network can be trained to segment cardiac ultrasound images into cardiac components including the left ventricular blood pool, the left ventricular wall, and the left atrium.

[0015] According to some embodiments, a medical image segmentation method by at least one processor comprises the steps of acquiring a medical image and generating segmentation data of the medical image using a segmentation network trained to output a result of segmenting the input image into clinical semantics. The segmentation network is implemented to perform domain-generalized semantic segmentation by learning reliable synthetic image regions using spatial weights related to the reliability of the synthetic image.

[0016] Inaccurate regions of the synthetic image can be excluded from training by the above spatial weights.

[0017] The above spatial weights may be provided by a meta-network. The meta-network may be parameterized to perform spatial uncertainty-awareness on an input image based on meta-learning.

[0018] The above segmentation network can be trained using synthetic images generated by varying at least one of the content condition and style condition input to the generative model.

[0019] The above medical image may include a cardiac ultrasound image.

[0020] The above segmented data may include the result of segmenting the cardiac ultrasound image into cardiac components including the left ventricular blood pool, the left ventricular wall, and the left atrium.

[0021] A medical image segmentation device according to some embodiments comprises a memory and a processor that executes instructions stored in the memory, wherein the processor is implemented to generate segmentation data of the medical image using a segmentation network trained to acquire a medical image and output a result of segmenting the input image into clinical significance. The segmentation network is implemented to perform domain-generalized semantic segmentation by learning reliable synthetic image regions using spatial weights related to the reliability of the synthetic image.

[0022] According to the embodiment, through domain generalization, semantic segmentation of medical images can be performed precisely regardless of the medical image acquisition environment.

[0023] According to the embodiment, the segmentation performance of a network generalized to synthetic images can be improved through meta-learning based spatial weighting (MLSW) optimization.

[0024] According to the embodiment, clinical semantic segmentation of ultrasound images acquired from various types of ultrasound devices can be performed.

[0025] According to the embodiment, domain generalization for cardiac ultrasound images can be achieved by generating synthetic images representing a wide range of cardiac anatomical structures and ultrasound device characteristics through variation of image generation conditions.

[0026] FIG. 1 is a diagram conceptually illustrating a medical image segmentation device according to one embodiment.

[0027] FIG. 2 is a drawing illustrating a generation model according to one embodiment.

[0028] FIG. 3 is a diagram illustrating various content condition configurations according to one embodiment.

[0029] FIG. 4 is a drawing illustrating a style condition configuration according to one embodiment.

[0030] FIG. 5 is a flowchart of a network training method according to one embodiment.

[0031] FIG. 6 is a flowchart of an image segmentation method according to one embodiment.

[0032] Figure 7 is the result of comparing the splitting performance of the present disclosure.

[0033] Embodiments of the present disclosure are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present disclosure may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present disclosure in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0034] Throughout the specification, when a part is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "…part," "…unit," and "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware, software, or a combination of hardware and software.

[0035] The device of the present disclosure is a computing device configured and connected so as to perform the operation of the present disclosure by having at least one processor execute instructions. The computing device may include one or more processors, memory for loading a computer program executed by the processor, and storage for storing the computer program and various data.

[0036] A computer program may include instructions that cause a processor to perform a method / operation according to various embodiments of the present disclosure, and may be stored on a non-transitory computer-readable storage medium. The computer program may be downloaded over a network or sold in the form of a product.

[0037] The processor can perform methods / operations according to various embodiments of the present disclosure by executing instructions. The processor controls the overall operation of each component of the computing device. The processor may be configured to include at least one of a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), or any type of processor well known in the art of the present disclosure.

[0038] The model, neural network, or network of the present disclosure may be implemented as software / computer program executed on a computing device as an artificial intelligence model (AI model) that learns at least one task.

[0039] The medical images of the present disclosure may be images of various body parts captured by various modalities, for example, the modalities may be ultrasound, X-ray, MRI (magnetic resonance imaging), CT (computed tomography), etc. In the description, echocardiography images are used as examples, but are not limited thereto.

[0040] Accurate segmentation of echocardiogram images is essential for the diagnosis of cardiovascular diseases, and deep learning-based cardiac image segmentation is currently being researched. However, data-driven image segmentation methods suffer from a domain shift problem, where performance deteriorates for out-distribution data (OOD datasets) that the model does not observe. This is because image characteristics are significantly influenced by various factors, such as equipment manufacturers, probe specifications, and ultrasound pulse frequencies. Therefore, there is a need for semantic segmentation techniques based on Domain Generalization (DG) that are not affected by the domain in which the data was collected.

[0041] Domain generalization aims to maintain performance on out-of-distribution data different from the training data, and recently, data augmentation methods have been proposed to enhance the generalizability of neural networks. However, generative model-based data augmentation has the disadvantage that the fidelity of synthetic images is not guaranteed. Particularly in data augmentation for semantic segmentation, the generative model synthesizes images from segment data, and pairs of synthetic images and segment data are used as inputs and ground truths to train the neural network. However, there is a limitation in that accurate parameterization of the neural network is difficult because low-precision synthetic images are used.

[0042] In the following, we describe a method for generating medical images under various imaging conditions for domain generalization and training a segmentation network using synthetic images. By recognizing unreliable regions in the synthetic images and excluding inaccurate data from the training, the method ultimately enables precise image segmentation.

[0043] FIG. 1 is a diagram conceptually illustrating a medical image segmentation device according to one embodiment.

[0044] Referring to FIG. 1, a medical image segmentation device (simply referred to as the ‘device’) (1), driven by at least one processor, is implemented to output a result of segmenting a medical image in clinical sense. In the description, an echocardiography image is used as an example, but the medical image to which the present disclosure applies is not limited thereto. An echocardiography image may be referred to as an echo image, an ultrasound image, a B-mode image, etc. The medical image is segmented into anatomical features that have clinical significance; for example, an echocardiography image may be segmented into a left ventricle (LV) blood pool, a left ventricle wall, a left atrium (LA), etc., which are clinically important cardiac components. The corresponding area may be annotated with a label in the area segmented into the left ventricle blood pool, left ventricle wall, left atrium, etc.

[0045] The device (1) may include a neural network trained using synthetic images generated under various conditions for Domain Generalization (DG). Here, it may be implemented to perform precise image segmentation by recognizing unreliable regions in the synthetic images through meta-learning and excluding inaccurate data from the unreliable regions from the neural network training.

[0046] The device (1) outputs a segmentation network (θ) that segments a medical image (10) in a clinical sense (20). seg )(100), a meta network (Meta network, θ) that outputs a spatial weight (W) (30) related to the reliability of the synthetic image through spatial uncertainty-awareness of the image. metaIt may include )(200). A generative model (300) that synthesizes datasets of various conditions may be used for domain generalization. The partition network (100) and the meta network (200) may be implemented in various structures, for example, they may be implemented based on U-net.

[0047] In the present disclosure, training data may consist of a synthetic dataset generated by a generative model (300) and a validation dataset. The synthetic dataset is a machine-generated dataset and may be denoted as 'MG'. The validation dataset is a more precise dataset than the synthetic dataset and may be denoted as 'HM' (Human-made) in the description to indicate that it is a dataset created through annotation by a radiologist. Machine-generated synthetic dataset D MG is a synthesized image X generated under various image conditions MG and its segmented data (segments) Y MG It can consist of pairs. Validation dataset D HM video X HM and the data Y that was manually partitioned from it HM It can be composed of pairs. Here, X is the input image and Y is the segmented data corresponding to the ground truth. The ground truth Y of the echocardiogram image may include, for example, data segmented into the left ventricular blood pool, the left ventricular wall, and the left atrium.

[0048] Image characteristics are determined by content characteristics and style characteristics. Content characteristics are factors that determine the spatial composition of the image and can be determined, for example, by the shape of the heart and the Region of Interest (ROI). Meanwhile, even when examining the same patient, differences in images may occur depending on the measurement equipment, which can be referred to as style characteristics. Style characteristics can be determined, for example, by the equipment manufacturer, probe specifications, and ultrasound pulse frequency.

[0049] To obtain a synthetic image having various image characteristics, the generation model (300) has content condition y content and style condition y style It can be implemented to generate a synthetic image by utilizing it as input.

[0050] A segmentation network (100) can be trained to receive a medical image (10) as input and output segmentation data (20) that segments the image in clinical sense. The segmentation network (100) uses training data to produce segmentation data θ inferred from the medical image X. seg It is trained to minimize the loss of (X) and the correct value Y, and to improve generalization performance, synthetic images generated under various conditions by the generative model (300) can be used as training data. Meanwhile, since the precision of synthetic images is generally low, it is difficult to accurately parameterize the segmentation network (100) trained on synthetic images. The present disclosure may utilize a meta-network (200) that captures inaccurate regions in synthetic images used for domain generalization.

[0051] Through meta-learning based spatial weighting (MLSW) that prevents the learning of unreliable regions (inaccurate regions) captured by the meta-network (200), the segmentation network (100) can learn fine details from human-generated data while learning the diversity of machine-generated data.

[0052] The meta network (200) is parameterized to perform spatial uncertainty-awareness on images based on meta learning, and a precise verification dataset D for meta optimization HM It can use. That is, the meta-network (200) is a precise verification dataset D annotated by humans. HM It can be used to learn the perception of spatial uncertainty regarding images and trained to capture low-reliance spatial regions in synthetic images. For example, the meta-network θ meta (200) is image X as in Equation 1 HM Partition data θ inferred from seg (X HM ) and the correct value Y HM It can be parameterized to minimize the loss.

[0053]

[0054] In Equation 1, L(x,y) can represent the cross-entropy loss between x and y. θ' seg is the spatial weight (W MG Composite image X with ) applied MG It can refer to adapted parameters using the loss, and It can be expressed as.

[0055] The meta network (200) is a synthetic image X MG Partition data θ inferred fromseg (X MG ) and the correct value Y MG Through this, the spatial weight (Spatial Weight, W) related to the reliability of the synthetic image MG It can be implemented to output )(30). Spatial weight W MG can be defined as in Equation 2. Spatial weight W MG The pixel regions used for training the segmentation network (100) can be filtered by this, and thereby the segmentation network (100) can train the segmentation task through reliable pixels in the synthetic region.

[0056]

[0057] Medical image X, segmentation data θ inferred from medical image X seg (X MG ), and space loss L(Y, θ' seg (X)) This space weight W MG It can be used as an input for the generation of (30).

[0058] Spatial weight W MG (30) is applied to the segmentation loss associated with the synthetic image, so that the segmentation network (100) can be trained using reliable pixels while excluding inaccurate pixels from the synthetic image. Through this meta-learning-based spatial weighting method, the segmentation network (100) is prevented from learning unreliable pixels of the synthetic image, thereby enabling precise image segmentation.

[0059] θ seg In the parameterization of D MG and D HM All of these can be used, ultimately enabling domain generalization and precise semantic partitioning. Machine-generated synthetic dataset D MG While training with, θ seg is the spatial weight W for the synthetic image MGIt is optimized using trusted pixels filtered by. Meanwhile, dataset D annotated by radiology experts HM Since it can be considered precise, D HM In the case of training with, D without spatial weight-based pixel filtering HM You can learn the details of.

[0060] θ seg It can be trained according to an objective function like Equation 3. θ seg and θ meta It can be optimized through the algorithm in Table 1.

[0061]

[0062]

[0063] In this way, the meta network (200) is a precise verification dataset D annotated by a human HM It can be trained to capture spatial regions with low reliability within synthetic data using this. By using only reliable pixels provided by the meta-network (200), the image segmentation process of the segmentation network (100) can be optimized. Through this, the segmentation network trained with synthetic images generated under various conditions can perform domain-generalized semantic segmentation.

[0064] FIG. 2 is a drawing illustrating a generation model according to one embodiment, FIG. 3 is a drawing illustrating various content condition configurations according to one embodiment, and FIG. 4 is a drawing illustrating style condition configurations according to one embodiment.

[0065] Referring to FIG. 2, the generation model (300) has a content condition y content and style condition y style It is implemented to generate a synthetic image by utilizing it as input. Through variation of image conditions, it is possible to generate echocardiogram images that exhibit a wide range of cardiac anatomical structures and ultrasound device characteristics.

[0066] The generation model (300) can be implemented as a diffusion probabilistic model that uses an iterative denoising process for precise image generation. Since iterative denoising requires significant computational cost, a latent diffusion model framework that applies a denoising scheme to latent regions rather than image regions may be used.

[0067] Content conditions can be composed, for example, of an image region of interest (ROI) and segments representing anatomical features. The image region of interest can be composed, for example, of a general angle range commonly used by radiologists (e.g., 60° < θ < 150°). Segment data can be composed of clinically important components in the image and can be displayed within the region of interest. The configuration of various content conditions can vary depending on the combination of anatomical features in images from available datasets (e.g., Echonet, HMC-QU, CAMUS).

[0068] Referring to Fig. 3, segmented data of an echocardiogram image can be generated in various ways depending on different combinations of clinically important cardiac components, such as the left ventricular blood pool, the left ventricular wall, and the left atrium. For example, the content conditions can be configured in various ways, such as segmented data containing only the left atrium segmented in the ROI of the ultrasound image (B-mode image), segmented data containing the left ventricular blood pool and the left atrium, or segmented data containing the left ventricular blood pool, the left ventricular wall, and the left atrium.

[0069] Style conditions represent image characteristics such as brightness, contrast, and scattering, and can be determined according to the specifications of the image acquisition device (e.g., ultrasound device). To accurately quantify the style conditions of an image, for example, the Nakagami distribution can be used. Here, the Nakagami distribution is a statistical model representing the distribution of an ultrasound signal and can be defined as shown in Equation 4.

[0070]

[0071] In Equation 4, Γ and U represent the gamma function and the unit-step function, respectively. The Nakagami distribution can be determined by the Nakagami scale parameter Ω and the Nakagami distribution parameter m.

[0072] Referring to Fig. 4, the ultrasound image can be divided into K partitions according to the travel time of the ultrasound signal, and style conditions can be constructed by quantifying Ω and m of the ultrasound signal within each sub-partition. Through this, different style parameters that simulate different measurement environments, such as the Nakagami scale parameter Ω and the Nakagami distribution parameter m, can be constructed to represent a wide range of image features.

[0073] FIG. 5 is a flowchart of a network training method according to one embodiment.

[0074] Referring to FIG. 5, the device (1) generates synthetic images according to content conditions and style conditions using a generation model (300) (S110). The device (1) can generate synthetic images for domain generalization using the generation model (300). The generation model (300) can be implemented as a diffusion probabilistic model using an iterative noise removal process, and a latent diffusion model framework can be used. The content conditions may consist of conditions related to segmented data, for example, segmented data representing image regions of interest and anatomical features. The style conditions represent conditions related to the image acquisition device, for example, image characteristics such as brightness, contrast, and scattering, and may consist of various specifications of the image acquisition device (e.g., ultrasound device). To accurately quantify the style conditions of the image, for example, a Nakagami distribution may be used. The Nakagami distribution can be determined by the Nakagami scale parameter Ω and the Nakagami distribution parameter m. Various synthetic videos for domain generalization can be generated by changing content and style conditions in various ways.

[0075] In the process of training a segmentation network (100) to output segmentation data that segments an input image into clinical significance using a synthetic dataset containing a synthetic image and segmentation data, the device (1) provides spatial weights related to the reliability of the synthetic image to the segmentation network (100) to learn reliable synthetic image regions (S120). Inaccurate regions of the synthetic image may be excluded from training by the spatial weights. Spatial weights related to the reliability of the synthetic image may be provided by a meta-network (200). The meta-network (200) may meta-learn to output spatial weights related to the reliability of the synthetic image using a precise verification dataset containing a verification image and data segmented by a person.

[0076] FIG. 6 is a flowchart of an image segmentation method according to one embodiment.

[0077] Referring to FIG. 6, the device (1) acquires a medical image (S210). The medical image may be, for example, a cardiac ultrasound image.

[0078] The device (1) generates segmentation data of a medical image using a segmentation network trained to output a result of segmenting an input image into clinical meaning (S220). The segmentation network can be implemented to perform domain-generalized semantic segmentation by learning reliable synthetic image regions using spatial weights related to the reliability of the synthetic image. The device (1) can output a result of segmenting an echocardiogram image into cardiac components (left ventricular blood pool, left ventricular wall, left atrium, etc.) using a segmentation network trained to perform generalized semantic segmentation. Here, the segmentation network can output precise segmentation data regardless of the medical image acquisition environment.

[0079] Figure 7 is the result of comparing the splitting performance of the present disclosure.

[0080] Referring to FIG. 7, the dataset used for the partition network (100) may consist of an in-distribution dataset (ID dataset) and an out-distribution dataset (OOD dataset).

[0081] The dataset within the distribution may consist of public datasets (Echonet, HMC-QU, CAMUS) and synthetic datasets generated from generative models (300). The dataset within the distribution may be configured to verify the performance of the partitioned network without domain movement.

[0082] An out-of-distribution dataset consists of unobserved data from a distribution different from the training dataset and is used to evaluate the robustness of a model to data with new attributes. The out-of-distribution dataset may include medical images of randomly selected patients, and, for example, may include ultrasound images acquired with ultrasound devices of various specifications, ranging from portable to high-performance console-type ultrasound devices.

[0083] Human-labeled annotated dataset D HM or synthetic dataset D MG An existing baseline network trained with, and without the meta-learning based spatial weighting (MLSW) optimization of the present disclosure, D HM and D MG The performance of the segmented network of the present disclosure (corresponding to 'proposed') can be verified through an existing reference network trained with.

[0084] Looking at the splitting results for the in-distribution and out-of-distribution datasets, D HM It can be seen that a reference network trained solely on the standard network has limited generalization ability and exhibits serious defects such as disconnected LVs.

[0085] Synthetic dataset DMG The reference network trained with is D HM Although generalization performance is improved compared to a reference network trained solely with the present disclosure, it can be seen that the generalization performance and segmentation accuracy of the segmentation network proposed through meta-learning-based spatial weighting (MLSW) optimization of the present disclosure are the best.

[0086] In this way, according to the embodiment, semantic segmentation of medical images can be performed precisely regardless of the medical image acquisition environment through domain generalization.

[0087] According to the embodiment, the segmentation performance of a network generalized to synthetic images can be improved through meta-learning based spatial weighting (MLSW) optimization.

[0088] According to the embodiment, clinical semantic segmentation of ultrasound images acquired from various types of ultrasound devices can be performed.

[0089] According to the embodiment, domain generalization for cardiac ultrasound images can be achieved by generating synthetic images representing a wide range of cardiac anatomical structures and ultrasound device characteristics through variation of image generation conditions.

[0090] The embodiments of the present disclosure described above are not implemented only through devices and methods, but may also be implemented through a program that realizes a function corresponding to the configuration of the embodiments of the present disclosure or a recording medium on which such program is recorded.

[0091] Although embodiments of the present disclosure have been described in detail above, the scope of the present disclosure is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concepts of the present disclosure as defined in the following claims also fall within the scope of the present disclosure.

Claims

1. A network training method using at least one processor, A step of obtaining a synthetic dataset including a synthetic image and segmented data of the synthetic image, and The method includes the step of training a segmentation network to output a result of segmenting an input image into clinical significance using the above synthetic dataset, and The above partitioned network is A training method implemented to perform domain-generalized semantic segmentation by learning a reliable synthetic image region using spatial weights related to the reliability of the synthetic image.

2. In Paragraph 1, A training method in which inaccurate regions of the synthetic image are excluded from training by the above spatial weights.

3. In Paragraph 1, The above spatial weights are provided by a meta-network, and The above meta-network is A training method parameterized to perform spatial uncertainty-awareness on input images based on meta-learning.

4. In Paragraph 1, The above synthetic dataset is A training method comprising synthetic images generated by varying at least one of the content condition and style condition input to a generative model.

5. In Paragraph 4, The above content conditions are A training method comprising segmented data representing an image region of interest (ROI) and anatomical features.

6. In Paragraph 4, The above style conditions are A training method comprising at least one parameter representing image characteristics.

7. In Paragraph 6, The above parameters are A training method comprising at least one of a Nakagami scale parameter and a Nakagami distribution parameter.

8. In Paragraph 1 The above partitioned network is A training method trained to segment cardiac ultrasound images into cardiac components including the left ventricular blood pool, the left ventricular wall, and the left atrium.

9. A medical image segmentation method using at least one processor, wherein The step of acquiring medical images, and The method includes the step of generating segmentation data of the medical image using a segmentation network trained to output a result of segmenting the input image into clinical meanings, and The above partitioned network is A medical image segmentation method implemented to perform domain-generalized semantic segmentation by learning a reliable synthetic image region using spatial weights related to the reliability of the synthetic image.

10. In Paragraph 9, A medical image segmentation method in which inaccurate regions of the synthetic image are excluded from training by the spatial weights above.

11. In Paragraph 9, The above spatial weights are provided by a meta-network, and The above meta-network is A medical image segmentation method parameterized to perform spatial uncertainty-awareness on input images based on meta-learning.

12. In Paragraph 9, The above partitioned network is A medical image segmentation method trained using synthetic images generated by varying at least one of the content condition and style condition input to a generative model.

13. In Paragraph 9, The above medical image is a medical image segmentation method including a cardiac ultrasound image.

14. In Paragraph 13, A medical image segmentation method comprising the result of segmenting the above-mentioned cardiac ultrasound image into cardiac components including a left ventricular blood pool, a left ventricular wall, and a left atrium.

15. Memory, and It includes a processor that executes instructions stored in the above memory, and The above processor It is implemented to generate segmentation data of said medical image using a segmentation network trained to acquire a medical image and output a result of segmenting the input image into clinical meanings, and The above partitioned network is A medical image segmentation device implemented to perform domain-generalized semantic segmentation by learning a reliable synthetic image region using spatial weights related to the reliability of the synthetic image.

Images