Virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent
A deep learning-based GAN algorithm generates multi-phase liver MRI images from pre-enhancement scans, addressing safety, time, and cost issues in CE-MRI by eliminating the need for contrast agents and enhancing diagnostic efficiency.
Patent Information
- Application Number
- CN202411063492.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-12-13
AI Technical Summary
Existing contrast agent enhanced magnetic resonance imaging (CE-MRI) methods for liver diagnosis face challenges such as patient safety risks, prolonged scan times, resource inefficiency, and high costs, particularly in the diagnosis of hepatocellular carcinoma (HCC), due to the use of liver-specific contrast agents like Gd-EOB-DTPA.
A deep learning-based reconstruction algorithm using a generative adversarial network (GAN) to generate multi-phase liver MRI images from pre-enhancement scans, eliminating the need for contrast agents and reducing scan time to under a second.
The method provides safe, efficient, and cost-effective multi-phase liver imaging without contrast agents, reducing patient discomfort and medical resource strain while maintaining diagnostic accuracy.
Smart Images
Figure CN119131166A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction algorithm based on deep learning and without the need for contrast agent, which can realize the reconstruction generation of traditional multi-phase contrast agent enhanced liver magnetic resonance imaging without the need for contrast agent. Background Art
[0002] Contrast agent (CA) enhanced magnetic resonance imaging (CE-MRI) plays a very important role in clinical diagnosis, especially for hepatocellular carcinoma (HCC). First, the imaging features reflected by different enhancement phases (arterial phase, portal venous phase, delayed phase and / or hepatobiliary phase) are of great significance for the accurate evaluation and diagnosis of HCC. Secondly, in view of the great success of LI-RADS (Liver Imaging Reporting and Data System), more and more studies have revealed the diagnosis of liver by multi-phase CE-MRI. Furthermore, liver-specific contrast agents such as Gd-EOB-DTPA have been proposed in recent years, which can achieve high diagnostic efficacy for HCC even smaller than 1 cm.
[0003] However, Gd-EOB-DTPA multi-phase CE-MRI has the following disadvantages: 1. Patients need to receive injections of contrast agents. According to existing reports, patients with liver and kidney dysfunction and allergies have great biosafety risks; 2. Multi-phase CE-MRI of the liver requires multiple scans, and the liver and gallbladder specific phase scan needs to be collected 20 minutes after the contrast agent injection, which is unbearable for many patients with moderate to severe illnesses; 3. For areas with insufficient medical resources, the extension of a single scan time will further aggravate the shortage of medical resources; 4. Contrast agents are expensive. Research reports indicate that the annual cost of contrast agent injections in the United States is as high as more than one billion US dollars. Summary of the invention
[0004] The purpose of the present invention is to propose a reconstruction algorithm based on deep learning, which takes the pre-enhanced liver MRI image as input, and outputs multi-phase (arterial phase, portal venous phase, delayed phase and hepatobiliary specific phase) enhanced images of the liver efficiently and quickly without contrast agent for clinical diagnosis and evaluation.
[0005] In order to achieve the above object, the technical solution of the present invention is to provide a virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent, characterized in that it comprises the following steps:
[0006] Step 1: Construct a neural network model based on a generative adversarial network, including the design of a generator and a discriminator, and further including the following steps:
[0007] Step 101, design a generator, the U-Net-based generator consists of an encoder and a decoder, wherein the encoder includes eight encoder blocks, respectively defined as ECoV Block 1-8, and the decoder includes eight decoder blocks, respectively defined as DCoV Block 1-8, wherein the encoder block ECoV Block 2-8 and the decoder block DCoV Block 2-8 are correspondingly jump-connected, and in the encoder and the decoder, the feature map output by the current encoder block or decoder block is input into the next encoder block or decoder block;
[0008] For the encoder:
[0009] In the encoder block ECoV Block 1: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3 and a 2D convolution layer 3, and the feature map of the input encoder block ECoV Block 1 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 1;
[0010] The structure of encoder blocks ECoV Block 2-3 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3 and a 2D convolution layer 3, and the feature map of the input encoder block ECoV Block 2 or 3 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 2 or 3;
[0011] The structure of encoder blocks ECoV Block 4-7 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 4, 5, 6 or 7 is connected with the residual of the feature map output by the Dropout layer, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 4, 5, 6 or 7;
[0012] In the encoder block ECoV Block 8: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 8 is connected with the residual of the feature map output by the Dropout layer, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 8;
[0013] For the decoder:
[0014] In the decoder block DCoV Block 1: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 8 is connected with the residual of the feature map output by the Dropout layer, and then a transposed upsampling convolution is performed through a convolution layer 5 to obtain the feature map output by the decoder block DCoV Block 1;
[0015] The structures of decoder blocks DCoV Block 2-4 are the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input decoder block DCoV Block 2, 3 or 4 is connected with the residual of the feature map output by the Dropout layer, and then a 2D convolution layer 5 is used for transposed upsampling convolution to obtain the feature map output by the decoder block DCoVBlock 2, 3 or 4;
[0016] The structure of decoder blocks DCoV Block 5-7 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3 and a 2D convolution layer 3, and the feature map of the input decoder block DCoV Block 5, 6 or 7 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a transposed upsampling convolution is performed through a 2D convolution layer 5 to obtain the feature map output by the decoder block DCoV Block 5, 6 or 7;
[0017] In the decoder block DCoV Block 8: the input feature map is processed by an upsampling layer, a zero padding layer, a 2D convolutional layer 6, and a Tanh activation layer in sequence to obtain the final output;
[0018] Step 102: Design the discriminator
[0019] The discriminator includes six discriminator blocks, which are defined as discriminator blocks 1-6. The output of the current discriminator block is input to the next discriminator block, where:
[0020] In the discriminator block Block 1: the input feature map is processed by a 2D convolution layer and a LeakyReLU layer respectively;
[0021] The structures of discriminator blocks 2-5 are the same: the input feature maps are processed by a 2D convolution layer, an instance normalization layer, and a Leaky ReLU layer respectively;
[0022] In the discriminator block 6: the input feature map is processed by a 2D convolution layer and a Tanh activation layer respectively;
[0023] Step 2: deploy the neural network model constructed in step 1, and train the neural network model. During the training, four neural network models are trained respectively, and the four neural network models synthesize arterial phase, portal venous phase, transition phase and hepatobiliary specific phase images from the plain scan MRI respectively;
[0024] Step 3: Input the plain scan MRI acquired in real time into the four trained neural network models to obtain the arterial phase, portal venous phase, transition phase and hepatobiliary specific phase images inferred by the generators of the four neural network models.
[0025] Preferably, in step 2, when training the neural network model, the optimization goal of the discriminator is to minimize the loss based on binary cross entropy, and the optimization goal of the generator is to minimize the loss combination based on BCE and mean square error.
[0026] Preferably, in step 2, when the neural network model is trained, the four neural network models are trained independently using the same hyperparameter architecture.
[0027] The present invention establishes a deep learning model based on Generative Adversarial Networks, takes the pre-enhancement MRI image as input, and uses the enhanced images of each phase as the gold standard to train the model. After the model training is completed, any pre-enhancement image can be input, so that the model can output four-phase liver enhancement images.
[0028] Compared with the prior art solutions, the present invention has the following beneficial effects:
[0029] 1) No contrast agent injection is required; 2) High time efficiency: multi-phase enhanced MRI acquisition that takes conventional methods half an hour can be completed in 1 second; 3) Insensitive to respiratory motion artifacts and strong robustness; 4) The generated multi-phase images are mutually registered, which is conducive to disease diagnosis and evaluation. Ultimately, it brings clinical benefits to patients and clinical medical institutions and reduces medical costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 For the overview diagram;
[0031] Figure 2 The structure of the neural network model constructed by the embodiment of the present invention is illustrated;
[0032] Figure 3 The comparison between each phase MRI generated by the model disclosed in the embodiment of the present invention and each phase MRI acquired by actual magnetic resonance scanning is illustrated;
[0033] Figure 4 The diagram illustrates the comparison between each phase MRI generated by the model disclosed in the embodiment of the present invention and each phase MRI acquired by actual magnetic resonance scanning. DETAILED DESCRIPTION
[0034] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0035] like Figure 1 As shown, a virtual multi-phase contrast-enhanced magnetic resonance imaging reconstruction method without the need for contrast agents disclosed in an embodiment of the present invention aims to combine a pre-trained neural network model to achieve reconstruction and generation of traditional multi-phase contrast-enhanced magnetic resonance imaging without the need for contrast agents, thereby bringing benefits to patients and clinical medical institutions.
[0036] A virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent disclosed in an embodiment of the present invention comprises the following steps:
[0037] Step 1: Build Figure 2 The neural network model based on the generative adversarial network shown includes the design of the generator and the discriminator, and further includes the following steps:
[0038] Step 101, design a generator.
[0039] In the embodiment of the present invention, the U-Net-based generator consists of an encoder and a decoder.
[0040] The encoder contains 8 blocks, defined as ECoV Block 1-8. The decoder also contains 8 blocks, defined as DCoV Block 1-8. ECoV Block 2-8 and DCoV Block 2-8 are skipped and connected to each other, and in the encoder and decoder, the feature map output by the current block is input to the next block.
[0041] For the encoder: In ECoV Block 1, the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1 (convolution kernel size: 1×1), a Leaky ReLU activation layer 2, a 2D convolution layer 2 (convolution kernel size: 3×3), a Leaky ReLU activation layer 3 and a 2D convolution layer 3 (convolution kernel size: 1×1), and the feature map of the input ECoV Block 1 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a 2D convolution layer 4 (convolution kernel size: 4×4, step size: 2) is used for strided downsampling convolution to obtain the feature map output by ECoV Block 1.
[0042] The difference between ECoV Block 2-3 and ECoV Block 1 is that an instance normalization layer is inserted between the leaky ReLU activation layer 1 and the 2D convolution layer 1, between the leaky ReLU activation layer 2 and the 2D convolution layer 2, and between the leaky ReLU activation layer 3 and the 2D convolution layer 3, respectively. The other modules are the same as ECoV Block 1.
[0043] The difference between ECoV Block 4-7 and ECoV Block 2-3 is that a Dropout layer (ratio = 0.3) is inserted after the 2D convolution layer 3, the feature map input to ECoV Block 4, 5, 6 or 7 is connected to the residual feature map output by the Dropout layer, and the other modules are the same as ECoV Block 2-3.
[0044] The difference between ECoV Block 8 and ECoV Block 1 is that a Dropout layer (ratio = 0.3) is inserted after the 2D convolution layer 3, and the feature map input to ECoV Block 8 is connected to the residual feature map output by the Dropout layer. The other modules are the same as ECoV Block 1.
[0045] In this embodiment of the present invention, the output feature map sizes of ECoV Block 1-8 are 64, 128, 256, 512, 512, 512, 512 and 512 respectively.
[0046] For the decoder: The structure of DCoV Block 1 is basically the same as that of ECoV Block 8, except that the 2D convolution layer 4 for implementing strided downsampling is replaced by the convolution layer 5 for implementing transposed upsampling (convolution kernel size: 1×1, stride: 2).
[0047] The structure of DCoV Block 2-4 is basically the same as that of ECoV Block 4-7, except that the 2D convolution layer 4 for implementing strided downsampling is replaced by the convolution layer 5 for implementing transposed upsampling (convolution kernel size: 1×1, stride: 2).
[0048] The structure of DCoV Block 5-7 is basically the same as that of ECoV Block 2-3, except that the 2D convolution layer 4 for implementing strided downsampling is replaced by the convolution layer 5 for implementing transposed upsampling (convolution kernel size: 1×1, stride: 2).
[0049] In DCoV Block 8: The input feature map is processed in sequence by an upsampling layer (amplification factor = 2), a zero padding layer, a 2D convolutional layer (convolution kernel size: 4×4, padding = 1), and a Tanh activation layer to obtain the final output.
[0050] In this embodiment of the present invention, the output feature size maps of DCoV Block 1-8 are 512, 512, 512, 512, 256, 128, 64 and 1 respectively.
[0051] Step 102: Design the discriminator
[0052] The discriminator consists of six discriminator blocks, defined as Block 1-6, and the output of the current discriminator block is input into the next discriminator block, where:
[0053] In Block 1: The input feature map is processed by a 2D convolution layer (kernel size: 4×4, stride: 2, padding: 1) and a Leaky ReLU layer.
[0054] In Block 2-5: The input feature map is processed by a 2D convolution layer (kernel size: 4×4, stride: 2, padding: 1), an instance normalization layer, and a Leaky ReLU layer.
[0055] In Block 6: The input feature map is processed by a 2D convolution layer (kernel size: 3×3, stride: 1, padding: 1) and a Tanh activation layer.
[0056] In the embodiment of the present invention, the output feature map sizes of Blocks 1-6 are 8, 16, 32, 64, 128 and 1 respectively.
[0057] Step 2: Deployment and training of neural network model based on generative adversarial network.
[0058] All sample data are randomly divided into a training group and a test group (in the embodiment of the present invention, the training group: the test group = 8:2), and four neural network models are trained respectively. These four neural network models synthesize arterial phase (AP), portal venous phase (VP), transitional phase (TP) and hepatobiliary phase (HBP) images from plain scan MRI (Pre). The Pre image is used as the input of the four neural network models, and the AP, VP, TP and HBP images collected by the actual scan are used as the true value for training the neural network model. The four neural network models are trained independently, using the same hyperparameters and configurations (loss function, number of training rounds, optimizer, etc.).
[0059] In the embodiment of the present invention, the optimization goal of the discriminator is to minimize the loss based on binary cross entropy (BCE), as shown in the following formula:
[0060] Loss = BCE real (inputs: Real images, Targets: 1)+BCE Syn (inputs: Synthesizedimages, Targets: 0)
[0061] In the formula, “1” means that the real CE-MRI image is judged as true, and “0” means that the generated CE-MRI image is judged as false.
[0062] The optimization goal of the generator is to minimize the loss combination based on BCE and mean squared error (MSE):
[0063] Loss = BCE Syn (inputs: Synthesized images, Targets: 1) + α × MSE (inputs: Synthesized images, Targets: Real images)
[0064] In the formula, "1" indicates that the generated CE-MRI image is judged to be true, and α is an adjustment parameter, which is set to 10 in the embodiment of the present invention.
[0065] The Adam optimizer is used, and the learning rates of the discriminator's Adam optimizer and the generator's Adam optimizer are set to 0.0001 and 0.00005 respectively. The total number of training rounds is 150. All neural network models are implemented using Pytorch (version: 1.9.0) and trained on an Nvidia GeForce RTX4090 graphics processing unit (GPU).
[0066] Step 3: Real-time reasoning of neural network model:
[0067] The plain scan images acquired by MRI scanning are input into the generators of the four trained neural network models, and the generators can independently complete the image prediction outputs of AP, VP, TP and HBP. Figure 3-4 These are the virtual AP, VP, TP and HBP generated by the trained neural network and the actual AP, VP, TP and HBP images acquired through magnetic resonance scanning.
Claims
1. A virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent, characterized in that: The following steps are involved: Step 1: Construct a neural network model based on a generative adversarial network, including the design of a generator and a discriminator, and further including the following steps: Step 101, design a generator. The U-Net-based generator consists of an encoder and a decoder, wherein the encoder includes eight encoder blocks, which are defined as ECoV Block 1-8 respectively, and the decoder includes eight decoder blocks, which are defined as DCoVBlock 1-8 respectively, wherein the encoder block ECoV Block 2-8 and the decoder block DCoV Block 2-8 are correspondingly jump-connected, and in the encoder and the decoder, the feature map output by the current encoder block or decoder block is input into the next encoder block or decoder block; For the encoder: In the encoder block ECoV Block 1: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3 and a 2D convolution layer 3, and the feature map of the input encoder block ECoV Block 1 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 1; The structure of encoder blocks ECoV Block 2-3 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3 and a 2D convolution layer 3, and the feature map of the input encoder block ECoV Block 2 or 3 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 2 or 3; The structure of encoder blocks ECoV Block 4-7 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 4, 5, 6 or 7 is connected with the residual of the feature map output by the Dropout layer, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoVBlock 4, 5, 6 or 7; In the encoder block ECoV Block 8: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 8 is connected with the residual of the feature map output by the Dropout layer, and then a strided downsampling convolution is performed through a 2D convolution layer 4 to obtain the feature map output by the encoder block ECoV Block 8; For the decoder: In the decoder block DCoV Block 1: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input encoder block ECoV Block 8 is connected with the residual of the feature map output by the Dropout layer, and then a transposed upsampling convolution is performed through a convolution layer 5 to obtain the feature map output by the decoder block DCoVBlock 1; The structures of decoder blocks DCoV Block 2-4 are the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3, a 2D convolution layer 3 and a Dropout layer, and the feature map of the input decoder block DCoV Block 2, 3 or 4 is connected with the residual of the feature map output by the Dropout layer, and then a 2D convolution layer 5 is used for transposed upsampling convolution to obtain the feature map output by the decoder block DCoVBlock 2, 3 or 4; The structure of decoder blocks DCoV Block 5-7 is the same: the input feature map is processed in sequence by a Leaky ReLU activation layer 1, an instance normalization layer 1, a 2D convolution layer 1, a Leaky ReLU activation layer 2, an instance normalization layer 2, a 2D convolution layer 2, a Leaky ReLU activation layer 3, an instance normalization layer 3 and a 2D convolution layer 3, and the feature map of the input decoder block DCoV Block 5, 6 or 7 is connected with the residual of the feature map output by the 2D convolution layer 3, and then a transposed upsampling convolution is performed through a 2D convolution layer 5 to obtain the feature map output by the decoder block DCoV Block 5, 6 or 7; In the decoder block DCoV Block 8: the input feature map is processed by an upsampling layer, a zero padding layer, a 2D convolution layer 6 and a Tanh activation layer in turn to obtain the final output; Step 102, design the discriminator The discriminator includes six discriminator blocks, which are defined as discriminator blocks 1-6. The output of the current discriminator block is input to the next discriminator block, where: In the discriminator block Block 1: the input feature map is processed by a 2D convolution layer and a Leaky ReLU layer respectively; The structures of discriminator blocks 2-5 are the same: the input feature maps are processed by a 2D convolution layer, an instance normalization layer, and a Leaky ReLU layer respectively; In the discriminator block 6: the input feature map is processed by a 2D convolution layer and a Tanh activation layer respectively; Step 2: deploy the neural network model constructed in step 1, and train the neural network model. During the training, four neural network models are trained respectively, and the four neural network models synthesize arterial phase, portal venous phase, transition phase and hepatobiliary specific phase images from the plain scan MRI respectively; Step 3: Input the plain scan MRI acquired in real time into the four trained neural network models to obtain the arterial phase, portal venous phase, transition phase and hepatobiliary specific phase images inferred by the generators of the four neural network models.
2. A virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent as claimed in claim 1, characterized in that: In step 2, when training the neural network model, the optimization goal of the discriminator is to minimize the loss based on binary cross entropy, and the optimization goal of the generator is to minimize the loss combination based on BCE and mean square error.
3. The virtual multi-phase contrast agent enhanced magnetic resonance imaging reconstruction method without contrast agent as claimed in claim 1, characterized in that: In step 2, when training the neural network model, the four neural network models are trained independently and use the same hyperparameter architecture.