Conditional diffusion model magnetic resonance image generation method based on artificial intelligence

Abstract

The invention discloses a conditional diffusion model magnetic resonance image generation method based on artificial intelligence. The method comprises the following steps: acquiring a magnetic resonance 2D image and a corresponding label; constructing a conditional diffusion model training set; constructing a conditional diffusion model; constructing and generating a loss function; performing end-to-end training on the conditional diffusion model; and respectively inputting the random Gaussian noise and the image label into the trained conditional diffusion model, and outputting an image. The method makes full use of the advantage of gradual diffusion generation of the conditional diffusion model, improves the precision and fidelity of the generation effect, enables the output image to be used for expanding the magnetic resonance data set of the selected organism part, can make full use of parameters of different initialized Gaussian noises and different generation time steps, and improves the generation efficiency. Under the condition that the generation precision and fidelity are guaranteed, the diversity of output images is improved, mode collapse is avoided, and the method has practical application value; the generated image does not need to be annotated, the method can be directly used for an image segmentation task, and the sample number of a segmentation data set is increased.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to a conditional diffusion model magnetic resonance image generation method based on artificial intelligence, which is suitable for magnetic resonance images. Background Art

[0002] In recent years, with the development of deep learning-based image recognition algorithms, these algorithms have been proven to reduce people's work in identifying magnetic resonance images, such as magnetic resonance image segmentation, image classification and other tasks [Andre Esteve, et al., Nature Medicine, 2019]. Nevertheless, the performance and versatility of deep learning algorithms are related to the generalization and unbiasedness of the training set. Compared with natural images, magnetic resonance images have problems such as difficulty in acquisition, high cost, and strong privacy, making it difficult to form a large-scale data set. Especially for image segmentation tasks, it also requires imaging experts with professional knowledge to spend a lot of time to finely annotate the images. This makes it more difficult to coordinate the data set with data-driven neural network algorithms, and model underfitting often occurs in image segmentation tasks.

[0003] Generative models are the latest technology in data generation. Generating realistic MRI data is a promising application. The generated data can replace the real MRI dataset for training data-driven deep learning models, reducing the disadvantage of MRI acquisition being difficult to obtain due to its high cost. In recent years, there has been increasing attention paid to using generative models for data augmentation and simultaneous training of synthetic data, using architectures such as Generative Adversarial Networks (GAN) or Diffusion Model (DM) for data generation. Compared to natural images, MRI is complex and contains rich contextual information, which may make GAN difficult to train and may cause mode collapse, that is, the generator produces limited or repetitive outputs and fails to capture the diversity of MRI distribution, resulting in failure to improve the performance of data-driven deep learning models [Prafulla Dhariwal, et al, Nips, 2021]. On the contrary, DDPM is more resistant to perturbations and noise, and can provide better data distribution for generating datasets, thereby achieving high-fidelity diversified images and effectively generating magnetic resonance images [Xuan Xu, et al, MICCAI, 2023]. In addition, image synthesis can solve the problem of insufficient data to a certain extent, but the synthesized data still needs to be annotated by experienced imaging experts. Therefore, due to the lack of high-quality labels, it is impossible to directly use synthetic data as a training set for tasks such as image segmentation. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a conditional diffusion model magnetic resonance image generation method based on artificial intelligence.

[0005] The above-mentioned purpose of the present invention is achieved by the following technologies:

[0006] A method for generating magnetic resonance images using a conditional diffusion model based on artificial intelligence, comprising the following steps:

[0007] Step 1, obtaining a magnetic resonance 2D image and corresponding labels;

[0008] Step 2: construct a conditional diffusion model training set based on magnetic resonance 2D images as samples;

[0009] Step 3: construct a conditional diffusion model, which includes a forward process, a backward process and a generative network, and the generative network includes a generative encoder, a conditional encoder and a generative decoder;

[0010] Step 4: Construct and generate loss function;

[0011] Step 5: According to the generated loss function, the conditional diffusion model is trained end-to-end using the conditional diffusion model training set, and the parameters of the conditional diffusion model are saved;

[0012] Step 6: Input the random Gaussian noise and the image label into the generative encoder and conditional encoder of the conditional diffusion model trained in step 5 respectively. Through multi-step sampling, the generative decoder of the conditional diffusion model outputs the output image with the image label.

[0013] As mentioned above, the 2D magnetic resonance image is a 2D magnetic resonance image of the FLAIR modality, and the output image is an output image of the FLAIR sequence with image labels.

[0014] As described above, the generative encoder includes 1 input layer and multiple sequentially connected downsampling layers;

[0015] The conditional encoder includes an input layer and a plurality of sequentially connected downsampling layers;

[0016] The generative decoder includes multiple upsampling layers and an output layer.

[0017] As described above, each downsampling layer of the encoder includes a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence;

[0018] Each downsampling layer of the conditional encoder includes a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence;

[0019] Each upsampling layer of the decoder includes a normalization layer, a convolution layer, an activation function layer, and an upsampling convolution module which are connected in sequence.

[0020] As described above, when the conditional diffusion model is trained, the magnetic resonance 2D image is input into the generative encoder, the label corresponding to the magnetic resonance 2D image is input into the conditional encoder, and the output feature map of the generative encoder and the conditional encoder is merged and input into the first upsampling layer of the generative decoder.

[0021] The loss function generated as described above is based on the following formula:

[0022] L MSE =L2[G(x),x]

[0023] Among them, L MSE To generate the loss function, L2 is the mean square error operation, x is the magnetic resonance 2D image input to the conditional diffusion model, and G(x) is the output image of the generative decoder of the conditional diffusion model.

[0024] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned method for generating magnetic resonance images based on artificial intelligence conditional diffusion model when executing the computer program.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for generating magnetic resonance images based on artificial intelligence conditional diffusion model.

[0026] A computer program product comprises a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for generating magnetic resonance images based on artificial intelligence conditional diffusion model.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. The present invention generates data through a conditional diffusion model, making full use of the advantages of the conditional diffusion model's step-by-step diffusion generation, improving the accuracy and fidelity of the generation effect, so that the generated data can be used to expand the magnetic resonance image data set of the selected biological body part. Compared with the existing magnetic resonance image generation method based on deep learning, the present invention can make full use of the parameters of different initialization Gaussian noises and different generation time steps to generate, while ensuring the generation accuracy and fidelity, improve the diversity of the generated images, avoid mode collapse, and has practical application value.

[0029] 2. Compared with the traditional conditional diffusion model, the method of the present invention generates images corresponding to the label content without the need to annotate the generated images. It can be directly used for image segmentation tasks, expand the number of segmentation data sets, and change the data distribution of segmentation data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the process of the present invention;

[0031] Figure 2 It is a schematic diagram of the forward process and backward process of the conditional diffusion model based on artificial intelligence of the present invention;

[0032] Figure 3 Generate a network structure diagram for the artificial intelligence-based conditional diffusion model of the present invention;

[0033] Figure 4 This is a schematic diagram of the sampling layer structure of the artificial intelligence-based conditional diffusion model of the present invention;

[0034] Figure 5 This is a schematic diagram of the upsampling layer structure of the artificial intelligence-based conditional diffusion model of the present invention;

[0035] Figure 6 This is an embodiment of the present invention. A brain image generated by a magnetic resonance image generation method based on an artificial intelligence conditional diffusion model, wherein 0, 200, 400, 600, 800, and 1000 are sampling times. DETAILED DESCRIPTION

[0036] In order to facilitate those skilled in the art to understand and implement the present invention, the following embodiments, Figure 1-Figure 3 Table 1 further describes the present invention in detail, where the selected biological part is the subject's brain. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0037] Example 1

[0038] A method for generating magnetic resonance images using a conditional diffusion model based on artificial intelligence, such as Figure 1 As shown, the specific steps include:

[0039] Step 1. Obtain 2D FLAIR (Fluid-attenuated inversion recovery) MRI images of multiple subjects, use the 2D FLAIR MRI images as samples, and annotate the target categories of the 2D FLAIR MRI images. The target category annotations of the 2D FLAIR MRI images are used as labels corresponding to the target categories of the corresponding modality images.

[0040] In this embodiment, 3D MRI images of 369 subjects in FLAIR mode are obtained. The 3D MRI images in FLAIR mode are labeled with target categories, and then the 3D MRI images in FLAIR mode are divided into 2D image slices to obtain 2D MRI images in FLAIR mode.

[0041] Step 2: construct a conditional diffusion model training set based on the 2D FLAIR modality MRI images obtained in step 1 as samples. In this embodiment, the conditional diffusion model training set is constructed using 2D FLAIR modality MRI images of 369 subjects, and the size of each 2D FLAIR modality MRI image is 240×240.

[0042] Step 3: Construct a conditional diffusion model. In this embodiment, the conditional diffusion model includes a forward process, a backward process and a generating network. The forward process is a process of adding random Gaussian noise to the data, and the backward process is a process of denoising.

[0043] In this embodiment, the generative network includes a generative encoder, a conditional encoder, and a generative decoder in order. The generative encoder includes 1 input layer and multiple downsampling layers. In this embodiment, the generative encoder includes 1 input layer and 4 sequentially connected downsampling layers. The convolution kernel size of the input layer is 3×3, the step size is 1, the surrounding padding step size is 1, the input channel is 1, the output channel is 16, the input image size is 1×240×240, and the output feature map size is 16×240×240. The 4 downsampling layers all include a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence.

[0044] The input layer of the generative encoder takes the 2D magnetic resonance image to be generated as input and outputs it after cropping it to a size of 240 × 240. The output feature map of the input layer of the generative encoder is used as the input feature map of the normalization layer of the first downsampling layer of the generative encoder. In the first downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a step size of 1, a surrounding padding step size of 1, an input channel of 16, an output channel of 32, an input feature map size of 16×240×240, and an output feature map size of 32×240×240. The activation function layer uses the Relu function to perform nonlinear activation on the input feature map. The convolution kernel size of the pooling layer is 2×2, a step size of 1, no padding is used, the input channel is 32, the output channel is 32, the input feature map size is 32×240×240, and the output feature map size is 32×120×120. The output feature map of the pooling layer is used as the output feature map of the first downsampling layer.

[0045] The output feature map of the first downsampling layer of the encoder is used as the input feature map of the normalization layer of the second downsampling layer of the encoder. In the second downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, 32 input channels, 64 output channels, 32×120×120 input feature map size, and 64×120×120 output feature map size. The activation function layer uses the Relu function to perform nonlinear activation on the input feature map. The convolution kernel size of the pooling layer is 2×2, a stride of 1, no padding is used, 64 input channels, 64 output channels, 64 input feature map size, and 64×120×120 output feature map size. The output feature map of the pooling layer is used as the output feature map of the second downsampling layer.

[0046] The output feature map of the second downsampling layer of the generative encoder is used as the input feature map of the normalization layer of the third downsampling layer of the generative encoder. In the third downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, 64 input channels, 128 output channels, an input feature map size of 64×60×60, and an output feature map size of 128×60×60. The activation function layer uses the Relu function to perform nonlinear activation on the input feature map. The convolution kernel size of the pooling layer is 2×2, a stride of 1, no padding is used, 128 input channels, 128 output channels, an input feature map size of 128×60×60, and an output feature map size of 128×30×30. The output feature map of the pooling layer is used as the output feature map of the third downsampling layer.

[0047] The output feature map of the third downsampling layer of the generative encoder is used as the input feature map of the normalization layer of the fourth downsampling layer of the generative encoder. In the fourth downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution kernel size of the convolution layer is 3×3, the stride is 1, the surrounding padding step is 1, the input channel is 128, the output channel is 256, the input feature map size is 128×60×60, and the output feature map size is 256×60×60. The activation function layer uses the Relu function to perform nonlinear activation on the input feature map. The convolution kernel size of the pooling layer is 2×2, the stride is 1, no padding is used, the input channel is 256, the output channel is 256, the input feature map size is 256×60×60, and the output feature map size is 256×30×30. The output feature map of the pooling layer is used as the output feature map of the fourth downsampling layer.

[0048] The input layer of the conditional encoder takes the label as input and outputs it after cropping it to a size of 240×240. The conditional encoder includes 1 input layer and multiple downsampling layers connected in sequence. In this embodiment, the conditional encoder includes 1 input layer and 4 downsampling layers connected in sequence. The convolution kernel size of the input layer is 3×3, the step size is 1, the surrounding padding step size is 1, the input channel is 1, the output channel is 16, the input image size is 1×240×240, and the output feature map size is 16×240×240. The 4 downsampling layers all include a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence.

[0049] The output feature map of the input layer of the conditional encoder is used as the input feature map of the normalization layer of the first downsampling layer of the conditional encoder. In the first downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, 16 input channels, 32 output channels, an input feature map size of 16×240×240, and an output feature map size of 32×240×240. The activation function layer uses the Relu function to nonlinearly activate the input, the convolution kernel size of the pooling layer is 2×2, a stride of 1, no padding is used, 32 input channels, 32 output channels, an input feature map size of 32×240×240, and an output feature map size of 32×120×120. The output feature map of the pooling layer is used as the output feature map of the first downsampling layer.

[0050] The output feature map of the first downsampling layer of the conditional encoder is used as the input feature map of the normalization layer of the second downsampling layer of the conditional encoder. In the second downsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, 32 input channels, 64 output channels, an input feature map size of 32×120×120, and an output feature map size of 64×120×120. The activation function layer uses the Relu function to nonlinearly activate the input. The convolution kernel size of the pooling layer is 2×2, a stride of 1, no padding is used, 64 input channels, 64 output channels, an input feature map size of 64×120×120, and an output feature map size of 64×60×60. The output feature map of the pooling layer is used as the output feature map of the second downsampling layer.

[0051] The output feature map of the second downsampling layer of the conditional encoder is used as the input feature map of the normalization layer of the third downsampling layer of the conditional encoder. In the third downsampling layer: the normalization layer uses the batch normalization function to normalize the input, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, 64 input channels, 128 output channels, an input feature map size of 64×60×60, and an output feature map size of 128×60×60. The activation function layer uses the Relu function to nonlinearly activate the input, the convolution kernel size of the pooling layer is 2×2, a stride of 1, no padding is used, 128 input channels, 128 output channels, an input feature map size of 128×60×60, and an output feature map size of 128×30×30. The output feature map of the pooling layer is used as the output feature map of the third downsampling layer.

[0052] The output feature map of the third downsampling layer of the conditional encoder is used as the input feature map of the normalization layer of the fourth downsampling layer of the conditional encoder. In the fourth downsampling layer: the normalization layer uses the batch normalization function to normalize the input, the convolution kernel size of the convolution layer is 3×3, the stride is 1, the surrounding padding step is 1, the input channel is 128, the output channel is 256, the input feature map size is 128×60×60, and the output feature map size is 256×60×60. The activation function layer uses the Relu function to nonlinearly activate the input, the convolution kernel size of the pooling layer is 2×2, the stride is 1, no padding is used, the input channel is 256, the output channel is 256, the input feature map size is 256×60×60, and the output feature map size is 256×30×30. The output feature map of the pooling layer is used as the output feature map of the fourth downsampling layer.

[0053] In this embodiment, the generative decoder includes multiple upsampling layers and 1 output layer. In this embodiment, the generative decoder includes 4 upsampling layers connected in sequence and 1 output layer. Among them, the 4 upsampling layers each include a normalization layer, a convolution layer, an activation function layer, and an upsampling convolution module connected in sequence.

[0054] The output feature maps of the fourth upsampling layer of the generative encoder and conditional encoder are merged into channels as the input of the normalization layer of the first upsampling layer of the generative decoder. In the first upsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, and a surrounding padding stride of 1. The input channel is 512, the output channel is 128, the input feature map size is 512×30×30, and the output feature map size is 128×30×30. The activation function layer uses the Relu function to nonlinearly activate the input. The convolution kernel size of the upsampling convolution module is 2×2, the stride is 1, and no padding is used. The input channel is 128, the output channel is 128, the input feature map size is 128×30×30, and the output feature map size is 128×60×60. The output feature map of the upsampling convolution module is used as the output feature map of the generative decoder output.

[0055] The output feature map of the first upsampling layer of the generative decoder is used as the input feature map of the normalization layer of the second upsampling layer of the generative decoder. In the second upsampling layer: the normalization layer uses the batch normalization function to normalize the input feature map, the convolution layer has a convolution kernel size of 3×3, a stride of 1, a surrounding padding stride of 1, an input channel of 128, an output channel of 64, an input feature map size of 128×30×30, and an output feature map size of 64×30×30. The activation function layer uses the Relu function to nonlinearly activate the input. The convolution kernel size of the upsampling convolution module is 2×2, a stride of 1, no padding is used, the input channel is 64, the output channel is 64, the input feature map size is 64×30×30, and the output feature map size is 64×60×60. The output feature map of the upsampling convolution module is used as the output feature map of the second upsampling layer of the generative decoder.

[0056] The output feature map of the second upsampling layer of the generative decoder is used as the input feature map of the normalization layer of the third upsampling layer of the generative decoder. In the third upsampling layer: the normalization layer uses the batch normalization function to normalize the input, the convolution kernel size of the convolution layer is 3×3, the stride is 1, the surrounding padding step is 1, the input channel is 64, the output channel is 32, the input feature map size is 64×60×60, and the output feature map size is 32×60×60. The activation function layer uses the Relu function to nonlinearly activate the input, the convolution kernel size of the upsampling convolution module is 2×2, the stride is 1, no padding is used, the input channel is 32, the output channel is 32, the input feature map size is 32×60×60, and the output feature map size is 32×120×120. The output feature map of the upsampling convolution module is used as the output feature map of the third upsampling layer of the generative decoder.

[0057] The output feature map of the third upsampling layer of the generative decoder is used as the input feature map of the normalization layer of the fourth upsampling layer of the generative decoder. In the fourth upsampling layer: the normalization layer uses the batch normalization function to normalize the input, the convolution kernel size of the convolution layer is 3×3, the stride is 1, the surrounding padding step is 1, the input channel is 512, the output channel is 32, the input feature map size is 32×120×120, and the output feature map size is 16×120×120. The activation function layer uses the Relu function to nonlinearly activate the input, the convolution kernel size of the upsampling convolution module is 2×2, the stride is 1, no padding is used, the input channel is 16, the output channel is 16, the input feature map size is 16×120×120, and the output feature map size is 16×240×240. The output feature map of the upsampling convolution module is used as the output feature map of the fourth upsampling layer of the generative decoder.

[0058] The convolution kernel size of the output layer is 3×3, the stride is 1, the surrounding padding stride is 1, the input channel is 16, the output channel is 1, the input feature map size is 16×240×240, and the output feature map size is 1×240×240.

[0059] Step 4: Set the generated loss function L MSE :

[0060] L MSE =L2[G(x),x]

[0061] Wherein, L2 is the mean square error operation, x is the magnetic resonance 2D image input by the conditional diffusion model, and G(x) is the output image of the generative decoder of the conditional diffusion model.

[0062] Step 5. According to the generated loss function set in step 4, the conditional diffusion model constructed in step 3 is trained end-to-end using the conditional diffusion model training set generated in step 2. The learning rate of the network is initialized to 0.001, the batch size is set to 8, and the network is trained on the PyTorch platform using the ADAM optimizer. Each iteration includes 1000 samples. When the total number of iterations reaches 1000, the training is stopped and the parameters of the conditional diffusion model are saved.

[0063] Step 6: Input the random Gaussian noise and the image label into the input layer of the generative encoder and the input layer of the conditional encoder of the conditional diffusion model trained in step 5 respectively. After multiple samplings, the generative decoder of the conditional diffusion model outputs the output image of the FLAIR sequence with the image label. Figure 6This is a brain image (output image) effect generated by the conditional diffusion model magnetic resonance image generation method based on artificial intelligence, where each column is 0, 200, 400, 600, 800, 1000, which is the generation effect obtained by sampling times. The generation effect shows that the present invention can generate magnetic resonance images with high fidelity, that is, the generated images can accurately reflect the brain structure (such as gray matter, white matter, cerebrospinal fluid, etc.) and the morphology of the lesion, and can be used to expand the data set.

[0064] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.

[0065] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0066] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0067] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0068] The specific embodiments described herein are merely examples of the spirit of the present invention. A person skilled in the art of the present invention may make various modifications or additions to the specific embodiments described or replace them in a similar manner, but this will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for generating magnetic resonance images based on a conditional diffusion model based on artificial intelligence, characterized in that: The following steps are involved: Step 1, obtaining a magnetic resonance 2D image and corresponding labels; Step 2: construct a conditional diffusion model training set based on magnetic resonance 2D images as samples; Step 3: construct a conditional diffusion model, which includes a forward process, a backward process and a generative network, and the generative network includes a generative encoder, a conditional encoder and a generative decoder; Step 4: Construct and generate loss function; Step 5: According to the generated loss function, the conditional diffusion model is trained end-to-end using the conditional diffusion model training set, and the parameters of the conditional diffusion model are saved; Step 6: Input the random Gaussian noise and the image label into the generative encoder and conditional encoder of the conditional diffusion model trained in step 5 respectively, and the generative decoder of the conditional diffusion model outputs the output image with the image label.

2. The method for generating magnetic resonance images using a conditional diffusion model based on artificial intelligence according to claim 1, characterized in that: The magnetic resonance 2D image is a magnetic resonance 2D image of the FLAIR modality, and the output image is an output image of a FLAIR sequence with an image label.

3. The method for generating magnetic resonance images based on a conditional diffusion model using artificial intelligence according to claim 1, characterized in that: The generative encoder comprises an input layer and a plurality of sequentially connected downsampling layers; The conditional encoder includes an input layer and a plurality of sequentially connected downsampling layers; The generative decoder includes multiple upsampling layers and an output layer.

4. The method for generating magnetic resonance images using a conditional diffusion model based on artificial intelligence according to claim 3, characterized in that: Each downsampling layer of the generative encoder includes a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence; Each downsampling layer of the conditional encoder includes a normalization layer, a convolution layer, an activation function layer, and a pooling layer connected in sequence; Each upsampling layer of the decoder includes a normalization layer, a convolution layer, an activation function layer, and an upsampling convolution module which are connected in sequence.

5. The method for generating magnetic resonance images using a conditional diffusion model based on artificial intelligence according to claim 4, characterized in that: When the conditional diffusion model is trained, the magnetic resonance 2D image is input into the generative encoder, the label corresponding to the magnetic resonance 2D image is input into the conditional encoder, and the output feature map of the generative encoder and the conditional encoder is merged and input into the first upsampling layer of the generative decoder.

6. The method for generating magnetic resonance images based on a conditional diffusion model using artificial intelligence according to claim 1, characterized in that: The generation loss function is based on the following formula: L MSE =L2[G(x),x] Among them, L MSE To generate the loss function, L2 is the mean square error operation, x is the magnetic resonance 2D image input to the conditional diffusion model, and G(x) is the output image of the generative decoder of the conditional diffusion model.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for generating magnetic resonance images based on the conditional diffusion model based on artificial intelligence according to any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for generating magnetic resonance images based on a conditional diffusion model based on artificial intelligence according to any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method for generating magnetic resonance images based on a conditional diffusion model based on artificial intelligence according to any one of claims 1 to 6 are implemented.

Images