Fat-suppressed magnetic resonance image generation method and device

Abstract

The disclosed technology relates to a method and device for generating a fat-suppressed magnetic resonance image, the method comprising: a step of inputting a magnetic resonance image into an encoder of a neural network to extract features of the magnetic resonance image by an image device; and a generator of the neural network generating a T2-weighted fat-suppressed image based on the features, wherein, before inputting the magnetic resonance image, the neural network learns based on a result of a discriminator of the neural network discriminating a loss caused by the generation of the T2-weighted fat-suppressed image and a result of a decoder reconstructing the magnetic resonance image input to the encoder using Bloch equations.

Description

Technical Field

[0001] The disclosed technology relates to a method and apparatus for generating fat-suppressed magnetic resonance images using a generative adversarial neural network (GAN) based on the Bloch equation. Background Art

[0002] Magnetic resonance imaging (MRI) involves transmitting high frequencies to the human body through a device composed of magnets, causing hydrogen nuclei in specific areas of the body to resonate. The resulting signal differences between tissues are converted into digital data and constructed as an image. Soft tissues such as muscles and ligaments generally have higher resolution and contrast, allowing high-resolution images to be obtained without the use of special substances such as contrast agents.

[0003] Magnetic resonance images are among the most widely used medical images, and image analysis technology using neural networks is being used to analyze them more objectively. Medical professionals can face the problem of subjective or unintentional bias in their diagnoses. This is particularly true when the distinction between fat and disease is ambiguous, relying on the subjective judgment and experience of medical professionals, making this a common problem. However, the use of neural networks that learn from a wide range of data allows for more objective analysis of medical images.

[0004] On the other hand, in existing technologies, when learning neural networks for analyzing medical images, network normalization is performed using only a loss function suitable for the neural network. This leads to the problem of missing subtle features included in the original image. Therefore, a technology is needed to generate more accurate images by preserving these subtle features.

[0005] Prior art literature

[0006] Patent document: Korean Patent Publication No. 10-2021-0017290 Summary of the Invention

[0007] Technical problems to be solved

[0008] The disclosed technology aims to provide a method and apparatus for generating fat-suppressed magnetic resonance images using a generative adversarial neural network based on the Bloch equation.

[0009] Workaround

[0010] A first aspect of the technology disclosed to solve the above-mentioned technical problem provides a method for generating a fat-suppressed magnetic resonance image, the method comprising: a step in which an image device inputs a magnetic resonance image to an encoder of a neural network to extract features of the magnetic resonance image; and a generator of the neural network generates a T2-weighted fat-suppressed image based on the features, wherein, before the magnetic resonance image is input, the neural network learns based on a result of a discriminator of the neural network discriminating a loss caused by the generation of the T2-weighted fat-suppressed image and a result of a decoder reconstructing the magnetic resonance image input to the encoder using Bloch equations.

[0011] A second aspect of the technology disclosed to solve the above-mentioned technical problems provides a fat-suppressed magnetic resonance image generation device, the device comprising: an input device for receiving input of a T1 magnetic resonance image and a T2 magnetic resonance image; a storage device for storing a generative adversarial neural network based on Bloch equations, the generative adversarial neural network based on Bloch equations comprising an encoder, a decoder, a generator, and a discriminator as lower-level networks; and a computing device for extracting features of the magnetic resonance image using the encoder and generating a T2-weighted fat-suppressed image based on the extracted features using the generator, wherein, before the T1 magnetic resonance image and the T2 magnetic resonance image are input, the generative adversarial neural network based on Bloch equations learns based on a result of the discriminator discriminating a loss caused by the generation of the T2-weighted fat-suppressed image and a result of the decoder reconstructing the magnetic resonance image input to the encoder using Bloch equations.

[0012] Beneficial effects

[0013] The embodiments of the disclosed technology may have the following advantages. However, this does not mean that the embodiments of the disclosed technology should include all of the following effects, and thus the scope of rights of the disclosed technology should not be understood as being limited thereto.

[0014] According to an embodiment of the disclosed technology, a method and apparatus for generating fat-suppressed magnetic resonance images using a generative adversarial neural network based on the Bloch equation can generate magnetic resonance images having different contrasts without performing additional imaging.

[0015] Furthermore, there is an effect of preserving subtle features in the image by reconstructing the image using the Bloch equation suitable for the image of each contrast.

[0016] In addition, it has the effect of more accurately identifying inflammatory diseases such as bone marrow edema. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 FIG2 is a diagram illustrating a process of generating a fat-suppressed magnetic resonance image using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology.

[0018] Figure 2 The present invention is a flowchart of a method for generating fat-suppressed magnetic resonance images using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology.

[0019] Figure 3 This is a block diagram of a device for generating a fat-suppressed magnetic resonance image using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology.

[0020] Figure 4 FIG. 1 is a diagram illustrating the structure of a generative adversarial neural network based on Bloch equations according to an embodiment of the disclosed technology.

[0021] Figure 5 FIG. 1 is a graph comparing the performance of a generative adversarial neural network based on Bloch equations according to an embodiment of the disclosed technology. DETAILED DESCRIPTION

[0022] The present invention is susceptible to various modifications and embodiments, and specific embodiments will be illustrated in the accompanying drawings and described in detail. However, it is not intended that the present invention be limited to specific embodiments, but rather should be understood to include all modifications, equivalents, and alternatives within the scope of the present invention.

[0023] Terms such as "first," "second," "A," and "B" may be used to describe various components, but the related components are not limited to these terms. Instead, they are used solely to distinguish one component from another. For example, a first component may be named a second component, and similarly, a second component may be named a first component, without exceeding the scope of the present invention. The term "and / or" includes any combination of the multiple related items or any one of the multiple related items.

[0024] In the terms used in this specification, singular expressions should be understood to include plural expressions, unless the context clearly indicates a different interpretation. In addition, it should be understood that the term "comprising" means the presence of the described features, quantities, steps, operations, constituent elements, components, or a combination thereof, and does not exclude the possibility of the presence or addition of one or more other features, quantities, steps, operations, constituent elements, components, or a combination thereof.

[0025] Before describing the drawings in detail, it should be understood that in this specification, components are distinguished solely based on their primary functions. That is, two or more components described below may be combined into a single component, or a single component may be divided into two or more components based on more specific functions.

[0026] In addition, each of the components described below can not only perform its own primary functions, but also perform part or all of the functions performed by other components, and part of the primary functions performed by each component can obviously also be performed by other components. Therefore, the presence or absence of each component described in this specification should be interpreted functionally.

[0027] Figure 1 FIG is a diagram illustrating a process of generating a fat-suppressed magnetic resonance image using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology. Figure 1 The imaging device may receive input of two magnetic resonance images having different contrast ratios and, based thereon, generate a magnetic resonance image having a different contrast ratio. The two magnetic resonance images input to the imaging device may be a T1-weighted image and a T2-weighted image obtained by imaging the same body tissue at the same time. Furthermore, the image generated based on these T1-weighted and T2-weighted images may be a T2-weighted fat-suppressed image.

[0028] To generate T2-weighted fat-suppressed images, the imaging device utilizes a generative adversarial neural network (Bloch-GAN) based on the Bloch equation. Furthermore, the GAN can be fed with a magnetic resonance image to determine the loss associated with generating the T2-weighted fat-suppressed image and calculate the loss associated with reconstructing the magnetic resonance image. The imaging device can cause the neural network to learn based on the results of determining the loss associated with image generation and calculating the loss associated with image reconstruction.

[0029] The image input to the imaging device may be a magnetic resonance image obtained by photographing a specific tissue or organ of the patient. The imaging device may utilize images photographed at the same time with different contrasts. For example, a T1-weighted image and a T2-weighted image of the patient's spine may be input. These images may be photographed by a photographing device included in the imaging device or may be obtained using another photographing device connected to the imaging device. The obtained magnetic resonance image may be input to a neural network stored in the imaging device and reconstructed into an image having a different contrast from the photographed image.

[0030] On the other hand, the imaging device can essentially use the T1 image to generate a T2 fat-suppressed image. To this end, a generative adversarial neural network based on the Bloch equation can be used to generate an image of a contrast not previously captured, leveraging the relationship between images of various contrasts. Before generating the image, the imaging device can perform a learning process on the generative adversarial neural network. The generative adversarial neural network includes an encoder, a decoder, a generator, and a discriminator as lower-level networks. During the neural network learning process, the imaging device can use the discriminator and decoder to normalize the entire network.

[0031] Meanwhile, during the learning process, the imaging device may receive inputs of T1-weighted and T2-weighted images. A T2-weighted fat-suppressed image may be generated based on the input images having different contrast ratios. As one embodiment, an encoder using a generative adversarial neural network based on the Bloch equations may be used to extract features from a magnetic resonance image, and a generator may be used to generate a T2-weighted fat-suppressed image based on the extracted features. A discriminator may then determine the loss caused by generating the fat-suppressed image, and a decoder may reconstruct the magnetic resonance image input to the encoder using the Bloch equations. When learning is performed using this process, subsequent image analysis may utilize only the encoder and generator to generate images. Specifically, the discriminator and decoder are lower-level networks utilized solely during the learning process. After the entire network has completed learning using the image loss determined by the discriminator and the reconstruction loss calculated by the decoder, the encoder and generator may be used solely to extract features from the input image and then generate a T2-weighted fat-suppressed image based on these features.

[0032] Meanwhile, during the learning process, the decoder can generate hypothetical magnetic resonance parameter maps and reconstruct images using the Bloch equations appropriate for each contrast image. Here, the magnetic resonance parameter maps generated by the decoder include a T1 magnetic relaxivity map, a T2 magnetic relaxivity map, and a proton density map. That is, the decoder can use these quantized parameter maps to reconstruct images with a different contrast than the image input to the neural network.

[0033] Existing neural network-based magnetic resonance image analysis methods only provide a loss function for image comparison, leading to the problem of missing subtle features in the image when reconstructing it. While this existing method can certainly provide somewhat objective image analysis, to more accurately identify lesions, the image should be reconstructed by reflecting the subtle features in the image. The disclosed technology uses the Bloch equation to normalize the network and uses a wavelet transform to replace the pooling and non-pooling layers included in existing networks, thereby preserving subtle information lost during the network's data processing. By reconstructing images in this way, it is possible to achieve higher accuracy than existing technologies in identifying lesions such as bone marrow edema.

[0034] Figure 2 This is a flowchart of a method for generating fat-suppressed magnetic resonance images using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology. Figure 2 The magnetic resonance image generation method 200 may include a feature extraction step 210 and a T2-weighted fat-suppressed image generation step 220. Steps 210 and 220 may be sequentially performed by a generative adversarial neural network based on the Bloch equations stored in an imaging device. The imaging device may receive a magnetic resonance image and input it into the neural network, and the neural network may generate a T2-weighted fat-suppressed image based on the input image.

[0035] In step 210, the imaging device inputs a magnetic resonance image into the neural network encoder. The neural network encoder then extracts features from the input magnetic resonance image. The imaging device stores a previously learned generative adversarial neural network based on the Bloch equations for image reconstruction. Then, when the magnetic resonance image is input, the image is input into the neural network encoder to extract features. Features included in the image can be extracted using multiple layers included in the encoder.

[0036] In step 220, the generator of the neural network generates a fat-suppressed image based on the extracted features. The magnetic resonance images input in step 210 may be T1-weighted images and T2-weighted images, and the generator may generate a T2-weighted fat-suppressed image based on these T1 and T2 images.

[0037] On the other hand, before executing steps 210 and 220, the neural network performs a learning process. As one embodiment, the neural network's discriminator can determine the loss caused by the generation of the fat-suppressed image. The discriminator can determine whether the generated fat-suppressed image is real or fake. Similar to existing generative adversarial neural networks, the neural network's generator can be trained by using the discriminator's judgment result as the adversarial loss for the generated image.

[0038] Furthermore, the neural network decoder can be used to calculate losses related to image reconstruction. For example, the Bloch equation can be used to reconstruct the magnetic resonance image input to the encoder. The decoder can use features extracted from the encoder to generate a magnetic resonance parameter map and reconstruct the image using the Bloch equation function. The decoder can identify values ​​representing the relationship between the T1 image, T2 image, and T2 fat-suppressed image to prevent overfitting of the encoder and effectively learn, and learn to reconstruct the T1 image and T2 image input to the encoder using the Bloch equation. After this learning process is fully executed, steps 210 and 220 can be performed.

[0039] Figure 3 This is a block diagram of a device for generating a fat-suppressed magnetic resonance image using a generative adversarial neural network based on the Bloch equation according to an embodiment of the disclosed technology. Figure 3 The magnetic resonance image generating apparatus 300 includes an input device 310 , a storage device 320 , and a computing device 330 . Furthermore, an output device 340 may also be included.

[0040] The input device 310 receives input regarding T1 and T2 magnetic resonance images. The input device 310 can receive T1-weighted and T2-weighted images transmitted from the MRI imaging device as magnetic resonance images. Of course, medical personnel can also directly input images captured by the MRI imaging device into the input device 310. To this end, the input device 310 may include an interface capable of receiving data corresponding to the magnetic resonance images or receiving input from medical personnel. Furthermore, input regarding the magnetic resonance images can be received via such an interface. For example, input regarding the magnetic resonance images can be received via an input interface such as a keyboard or mouse.

[0041] Storage device 320 stores a generative adversarial neural network based on the Bloch equations. Storage device 320 can store a previously learned generative adversarial neural network. Of course, it can also store an initial neural network that has not been previously learned, followed by a neural network updated under the control of computing device 330. Storage device 320 can be implemented as a memory with a capacity capable of storing a generative adversarial neural network. Furthermore, in addition to the neural network, it can also store additional data. For example, it can also store learning data used to learn the neural network.

[0042] The computing device 330 can use a generative adversarial neural network to extract features of the magnetic resonance image and generate a T2-weighted fat-suppressed image based on the extracted features. Figure 1 and Figure 2 As described above, the neural network undergoes a learning process before generating T2-weighted fat-suppressed images. Specifically, the neural network's discriminator determines the loss incurred by image generation, while the decoder calculates the loss incurred by reconstructing the magnetic resonance image according to the Bloch equation. A generative adversarial neural network based on the Bloch equation includes lower-level networks, including an encoder, a generator, a discriminator, and a decoder. The computing device 330 can input an input magnetic resonance image into the neural network's encoder to extract features. The generator can then generate a T2-weighted fat-suppressed image. The discriminator can then determine the loss incurred by generating the fat-suppressed image, and the decoder can reconstruct the image. During the learning process of the generative adversarial neural network, the computing device 330 can utilize all four lower-level networks. For example, the discriminator's loss and the decoder's loss can be used to normalize the entire network. After learning is complete, images can be generated using only the encoder and generator.

[0043] On the other hand, the magnetic resonance image generating apparatus 300 may further include an output device 340 for outputting the generated T2-weighted fat-suppressed image. The output device 340 may be implemented as a device such as a display, and may output the T2-weighted fat-suppressed image to a screen, thereby providing information for medical personnel to determine lesions.

[0044] On the other hand, the magnetic resonance image generating apparatus 300 described above may also be implemented as a program (or application) including an executable algorithm that can be executed by a computer. The program may be provided by being stored in a temporary or non-transitory computer readable medium.

[0045] A non-transitory readable medium is not a medium that stores data for a short period of time, such as a register, cache, or memory, but rather a medium that stores data semi-permanently and can be read by a device. Specifically, the various applications or programs described above can be provided by being stored in a non-transitory readable medium such as a CD, DVD, hard disk, Blu-ray disc, USB, memory card, ROM (read-only memory), PROM (programmable read-only memory), EPROM (erasable programm ...

[0046] The temporary readable medium means various RAMs such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronized DRAM (SLDRAM) and direct RAM RAM (DRRAM).

[0047] Figure 4 is a diagram showing the structure of a generative adversarial neural network based on Bloch equations according to an embodiment of the disclosed technology. Figure 4 ,The generative adversarial neural network based on the Bloch equation includes an encoder, a generator, a discriminator, and a decoder.

[0048] The generative adversarial neural network based on the Bloch equation has a structure that combines the existing generative adversarial neural network with the autoencoder. First, the encoder can extract the features of the input image. Figure 4 As shown, the encoder can include a residual block. The features input through the residual block can be instance normalized through a 3×3 convolutional layer. Unlike batch normalization, which calculates the mean and standard deviation of a batch, instance normalization only calculates one image for each mini-batch and utilizes the distribution of each individual image. Features can then be extracted by performing instance normalization and convolution in reverse through a ReLU process.

[0049] On the other hand, the generator can generate a fat-suppressed image based on the features extracted in this manner. The generator can generate a T2-weighted fat-suppressed image based on the features of each of the T1-weighted image and the T2-weighted image. When generating images in this manner, the discriminator can discriminate the loss associated with the generated image.

[0050] The decoder, on the other hand, generates hypothetical quantized magnetic resonance parameter maps and reconstructs the image using the Bloch equations appropriate for each contrast ratio. The decoder identifies the value representing the relationship between the T1-weighted image, the T2-weighted image, and the T2-weighted fat-suppressed image, and reconstructs the T1-weighted and T2-weighted images input to the encoder into a T2-weighted fat-suppressed image using the Bloch equations. This process prevents overfitting in the encoder and allows subtle features to be reflected in the reconstructed image.

[0051] On the other hand, in the Bloch equation-based generative adversarial neural network according to the disclosed technology, the adversarial loss function for learning is expressed by the following Equation 1:

[0052] [Equation 1]

[0053]

[0054] Here, Ge represents the generator of the neural network, and En represents the encoder of the neural network. Furthermore, x represents the input multi-contrast image. The decoder of the neural network learns to reconstruct an image by identifying the T1, T2, and S0 values ​​that are the basis for the relationship between the T1 and T2 images input to the encoder and the T2 fat-suppressed image generated by the generator. S0 represents the signal amplitude value without events and is proportional to the proton density value, voxel size, data sampling period, phase encoding step number, and magnetic field strength.

[0055] On the other hand, the normalization function used in the neural network is represented by the following Equation 2.

[0056] [Equation 2]

[0057]

[0058]

[0059] Where TR = TR T1-w , TR T2-w TE=TE T1-w ,TE T2-w and Denote images of the assumed quantized T1, T2, and signal amplitude values ​​without events, respectively. Furthermore, De denotes the decoder of the neural network. According to Equation 2, the decoder can be trained to generate images of the assumed quantized T1, T2, and S0 values, and can generate using given TR, TE, x values, and equations

[0060] In the case where the input source image and the target image are perfectly aligned, the learning process of most image transformation networks relies on a pixel-by-pixel loss function. However, this pixel-by-pixel loss function is designed to enable the network to learn to generate mathematically similar images, rather than to generate semantically similar images. That is, from a mathematical point of view, they are similar, but the subtle features of the actual original image may be missed. Therefore, in order to reflect the subtle features of the original image, other loss functions that can generate semantically similar images should be used. In the learning process of the generative adversarial neural network based on the Bloch equation, the pixel-by-pixel l1 loss function is used. and the perceptual loss function Both. The two loss functions are expressed by Equation 3 below.

[0061] [Equation 3]

[0062]

[0063]

[0064] Among them, VGG k (y) is an image network dataset for input y, representing the features of the kth convolutional layer of the pre-trained VGG-16 network, and n k 、w k 、h k denote the number of channels, width, and height, respectively. As a result, the total loss function for the encoder, decoder, and generator that constitute the image generation part of the generative adversarial neural network based on the Bloch equation is It is defined by the following equation.

[0065] [Equation 4]

[0066]

[0067] Among them, λ adv ,λ R ,λ l1 and λ VGG Represents the relative weights of each loss function. As a result, the generative adversarial neural network based on the Bloch equation can make the adversarial loss multiplied by the weights Normalized loss Pixel-by-pixel loss and cognitive loss function Learning is performed in a way that minimizes the sum of .

[0068] Figure 5 is a graph comparing the performance of a generative adversarial neural network based on Bloch equations according to an embodiment of the disclosed technology. Figure 5 In

[15] , Bloch-GAN represents the case where normalization is applied to the autoencoder part through appropriate learning, Bloch-GAN (NAE) represents the case where no autoencoder normalization is applied, and Bloch-GAN (AE) represents the case where the autoencoder part is normalized using an existing method (instead of the autoencoder normalization based on the Bloch equation). Figure 5 The images in the first and second rows are obtained from the first data set, and the images in the third and fourth rows are obtained from the second data set. In addition, the images in the second and fourth rows represent a case where the area indicated by the red arrow in the image in the first row is magnified.

[0069] Although the results from the three neural networks appear similar, it can be confirmed that edema within the spinal body is not accurately represented in the Bloch-GAN (NAE) and Bloch-GAN (AE) models. In contrast, the results from the normalized Bloch-GAN model based on the Bloch equation show that areas corresponding to edema appear brighter. This means that by reflecting subtle features in the original image, the contrast value of the lesion area is higher than that obtained using the other neural networks, allowing for more accurate lesion analysis.

[0070] While the embodiments shown in the accompanying drawings are used to illustrate a method and apparatus for generating fat-suppressed magnetic resonance images using a generative adversarial neural network based on the Bloch equation according to one embodiment of the disclosed technology, this is merely exemplary. Those skilled in the art will appreciate that various modifications and equivalent embodiments may be implemented. Therefore, the true scope of protection for the disclosed technology should be determined by the appended claims.

Claims

1. A method for generating a fat-suppressed magnetic resonance image, comprising: The imaging device inputs two magnetic resonance images having different contrasts to an encoder of a neural network to extract features of the magnetic resonance images; as well as The generator of the neural network generates a magnetic resonance image having a different contrast from the two magnetic resonance images based on the features, In which, before the magnetic resonance image is input, the neural network learns based on the results of the discriminator of the neural network distinguishing the loss caused by the generation of the T2-weighted fat-suppressed image and the results of the decoder reconstructing the magnetic resonance image input to the encoder using the Bloch equation.

2. The fat-suppressed magnetic resonance image generation method according to claim 1, wherein: The two magnetic resonance images input to the encoder are a T1-weighted image and a T2-weighted image obtained by photographing a specific tissue in the body at the same time, and The magnetic resonance image generated by the generator is a T2-weighted fat-suppressed magnetic resonance image.

3. The fat-suppressed magnetic resonance image generation method according to claim 1, wherein: The neural network is a generative adversarial neural network based on the Bloch equation.

4. The fat-suppressed magnetic resonance image generation method according to claim 3, wherein: The generative adversarial neural network includes an adversarial loss function, a normalized loss function, a pixel-by-pixel loss function, and a cognitive loss function as a total loss function, and learns in a manner that minimizes the total loss function.

5. The fat-suppressed magnetic resonance image generation method according to claim 1, wherein: The decoder generates a hypothetical magnetic resonance parameter map and reconstructs the magnetic resonance image based on the magnetic resonance parameter map.

6. The fat-suppressed magnetic resonance image generation method according to claim 5, wherein: The magnetic resonance parameter maps include a T1 magnetic relaxation rate map, a T2 magnetic relaxation rate map and a proton density map.

7. A fat-suppressed magnetic resonance image generating apparatus, comprising: an input device that receives an input regarding two magnetic resonance images having contrasts different from each other; A storage device storing a generative adversarial neural network based on the Bloch equation, wherein the generative adversarial neural network based on the Bloch equation includes an encoder, a decoder, a generator, and a discriminator as lower-level networks; as well as a computing device that extracts features of the two magnetic resonance images using the encoder, and generates a magnetic resonance image having a different contrast from the two magnetic resonance images using the generator based on the extracted features, In which, before inputting the two magnetic resonance images, the generative adversarial neural network based on the Bloch equation is learned according to the result of the discriminator discriminating the loss caused by the generation of the magnetic resonance image and the result of the decoder reconstructing the magnetic resonance image input to the encoder using the Bloch equation.

8. The fat-suppressed magnetic resonance image generating apparatus according to claim 7, wherein: The two magnetic resonance images are a T1-weighted image and a T2-weighted image obtained by photographing a specific tissue in the body at the same time, and The magnetic resonance image generated by the generator is a T2-weighted fat-suppressed magnetic resonance image.

9. The fat-suppressed magnetic resonance image generating apparatus according to claim 7, wherein: The generative adversarial neural network includes an adversarial loss function, a normalized loss function, a pixel-by-pixel loss function, and a cognitive loss function as a total loss function, and learns in a manner that minimizes the total loss function.

10. The fat-suppressed magnetic resonance image generating apparatus according to claim 7, wherein: The generative adversarial neural network generates a hypothetical magnetic resonance parameter map and reconstructs the magnetic resonance image based on the magnetic resonance parameter map.

11. The fat-suppressed magnetic resonance image generating apparatus according to claim 10, wherein: The magnetic resonance parameter maps include a T1 magnetic relaxation rate map, a T2 magnetic relaxation rate map and a proton density map.

Images