Multi-parameter MRI medical image fusion method based on conditional diffusion probability model

Abstract

The embodiment of the invention provides a multi-parameter MRI medical image fusion method based on a conditional diffusion probability model, and the method comprises the steps: S101, splicing Gaussian noise of a current step number with a to-be-fused medical image pair to obtain a three-channel tensor, and inputting the three-channel tensor into a trained noise predictor to output predicted noise; the medical image pairs are MRI images with different parameters; s102, performing single-step reverse de-noising on the predicted noise to obtain a de-noised graph of the current step number; and S103, when judging that the current step number reaches a preset step number, outputting the de-noised image of the current step number to obtain a required medical fusion image, otherwise, setting the de-noised image as Gaussian noise of the next step number, and returning to the step S101. According to the method, image fusion can be effectively carried out on the prostate multi-parameter MRI images, recognition and observation of prostate structure information can be enhanced, and a suspected focus area can be accurately positioned.

Description

Technical Field

[0001] The present invention relates to the field of medical images, and in particular to a multi-parameter MRI medical image fusion method based on a conditional diffusion probability model. Background Art

[0002] Magnetic resonance imaging (MRI) is often used to examine prostate-related diseases because it can provide rich soft tissue information. Due to the physical limitations of imaging technology, a single sensor imaging mode can only reflect the characteristics of tissues or organs in certain aspects. According to the medical imaging modality, it can be divided into two categories: anatomical images and functional images. Different medical imaging modes can provide complementary information of different types of organ tissues, reduce the storage cost of multi-source data, and enhance decision-making execution capabilities. Therefore, the research on the fusion of multi-parameter information will be of great significance to medical imaging diagnosis and treatment. The fusion of anatomical imaging and functional images can not only effectively determine the patient's disease severity, but also accurately locate the lesions.

[0003] However, since medical image fusion is a process of reducing the image capacity by inputting dual-parameter or even multi-parameter images and outputting a single image, some image information will inevitably be lost after fusion. Therefore, the purpose of medical image fusion is to remove redundant information in each parameter image as much as possible, reduce feature aliasing at the same position of the image, retain complementary information and generate high-quality fused images.

[0004] From an application perspective, image fusion belongs to a branch of image generation. Generative adversarial networks (GANs) have been used as a method for image fusion since they were applied from image generation to image fusion. Although GAN-based image fusion models have produced satisfactory fused images, they are plagued by some problems, including unstable training processes and mode collapse. These challenges lead to unreasonable distribution and low quality of the generated fused images, which seriously affects the practicality of GAN-based methods.

[0005] In recent years, the denoising diffusion probability model (DDPM) has received extensive attention. It has good mathematical interpretability and high-quality generation results, and has achieved great success in image generation. However, in image fusion, the basic diffusion model cannot achieve direct image fusion due to the lack of image ground truth, which is a huge challenge for the current application of diffusion models in image fusion.

[0006] The existing DDPM-based image fusion method is Denoising Diffusion Model for Multi-Modality Image Fusion (DDFM), which uses an unconditional pre-trained diffusion model based on a scoring-based approach. This method relies on the priors provided by the pre-trained diffusion model, which means that this diffusion-based method does not need to be trained in the image fusion task. It achieves image fusion through a post-designed score matching optimization process, and the diffusion parameters pre-trained for image generation may be inappropriate and unsuitable for image fusion. Moreover, the application of a pre-trained network will lead to difficulties in modifying the network structure, further limiting the scope of application of this method. Summary of the invention

[0007] In view of this, the embodiments of the present disclosure provide a multi-parameter MRI medical image fusion method based on a conditional diffusion probability model, which at least partially solves the problems existing in the prior art.

[0008] The embodiment of the present invention provides a multi-parameter MRI medical image fusion method, which includes:

[0009] S101, concatenating the Gaussian noise of the current step number with the medical image pair to be fused to obtain a three-channel tensor, and inputting the three-channel tensor into a trained noise predictor to output predicted noise; wherein the medical image pair is an MRI image with different parameters; during the training process of the noise predictor, using a temporary true value image as the input of the noise predictor reduces the task difficulty of the entire framework from generating an expected fused image from Gaussian noise to shifting from a temporary true value image distribution to an expected fused image distribution;

[0010] S102, performing single-step reverse denoising on the predicted noise to obtain a denoised image of the current step number;

[0011] S103, when it is determined that the current number of steps reaches the preset number of steps, the denoised image of the current number of steps is output to obtain the required medical fusion image, otherwise the denoised image is set to the Gaussian noise of the next number of steps, and the process returns to step S101.

[0012] Preferably, in step S101, if the current step number is 1, the Gaussian noise is randomly generated Gaussian noise.

[0013] Preferably, the training process of the noise predictor includes: obtaining a source image pair as a training set, and using a simple combination method to generate a temporary true value image from the source image pair; adding noise to the temporary true value image through a forward diffusion process to generate a noisy image and corresponding noise; splicing the noisy image and the source image pair into a three-channel tensor, and inputting it into the noise predictor for training until the preset training conditions are met.

[0014] Preferably, the temporary true value image is represented as follows:

[0015]

[0016] in: is a temporary substitute for the true value image; Represents a simple combinatorial function; I DWI ,I T2 Represents a source image pair.

[0017] Preferably, a value t is randomly selected between 1 and T, and a Gaussian noise is randomly generated. The noise image generated by the forward diffusion process is expressed as:

[0018]

[0019] in, represents a noise image; represents the noise addition table, α t =1-β t , β t is the variable that adds the variance of the noise to the data at time t.

[0020] Preferably, when training the noise predictor, it also includes:

[0021] The predicted noise output by the noise predictor is subjected to loss calculation to obtain a total loss L, and the total loss L is subjected to gradient descent back propagation to optimize the noise predictor; wherein the loss function includes a denoising loss L diff With the guidance loss L guide ; The formula of total loss L is as follows:

[0022] L=L guide +L diff

[0023]

[0024] L guide =L SSIM +L int +L grad

[0025] in:

[0026] L SSIM is the structural similarity loss, and

[0027] L int is the strength loss, and

[0028] L gradis the edge contour loss, and

[0029] The desired fusion image The estimated value of

[0030] Preferably, in step S102, the denoised image obtained by the single-step reverse denoising process It is expressed as:

[0031]

[0032] in:

[0033] is the standard deviation at time t,

[0034] z:

[0035] The embodiment of the present invention further provides a multi-parameter MRI medical image fusion device, which includes:

[0036] A predicted noise output unit is used to concatenate the Gaussian noise of the current step number with the medical image pair to be fused to obtain a three-channel tensor, and input the three-channel tensor into a trained noise predictor to output predicted noise; wherein the medical image pair is an MRI image with different parameters; during the training process of the noise predictor, a temporary true value image is used as an input of the noise predictor to reduce the task difficulty of the entire framework from generating an expected fused image from Gaussian noise to shifting from the temporary true value image distribution to the expected fused image distribution;

[0037] A single-step reverse denoising unit, used for performing single-step reverse denoising on the predicted noise to obtain a denoised image of the current step number;

[0038] The judging unit is used to output a denoised image of the current number of steps to obtain the required medical fusion image when judging that the current number of steps reaches a preset number of steps; otherwise, the denoised image is set to the Gaussian noise of the next step and the predicted noise output unit is notified.

[0039] The embodiment of the present invention further provides a multi-parameter MRI medical image fusion device, which includes:

[0040] at least one processor; and,

[0041] a memory communicatively connected to the at least one processor; wherein,

[0042] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned multi-parameter MRI medical image fusion method.

[0043] In one of the above embodiments, a new image is constructed by combining MRI medical images of different parameters of the same patient. The new image retains the complementary information between the images of different parameters, and the fusion performance of the prostate image is improved by using a temporary true value image and an improved noise predictor, which can assist doctors to better identify and observe lesions, thereby improving the efficiency and accuracy of doctors' diagnosis and providing more timely and effective treatment for prostate patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0045] Figure 1 A schematic flow chart of a multi-parameter MRI medical image fusion method provided in the first embodiment of the present invention.

[0046] Figure 2 A training process and reasoning flow chart of the multi-parameter MRI medical image fusion method provided in the first embodiment of the present invention;

[0047] Figure 3(a) shows Figure 2 The specific framework implementation diagram of the training process;

[0048] Figure 3(b) shows Figure 2 The specific framework implementation diagram of the reasoning process;

[0049] Figure 4 It is the structural diagram of the noise predictor;

[0050] Figure 5 Set up graphs for each network layer of the noise predictor;

[0051] Figure 6 The following is a comparison chart of the fusion results of three pairs of source images and the subjective effects of other methods;

[0052] Figure 7 The fusion results of three pairs of source images are compared with the objective evaluation indicators of other methods;

[0053] Figure 8 This is a schematic structural diagram of a multi-parameter MRI medical image fusion device provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0054] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0055] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0056] See also Figure 1 to Figure 3(b) The first embodiment of the present invention provides a multi-parameter MRI medical image fusion method, which can be executed by a multi-parameter MRI medical image fusion device (hereinafter referred to as a fusion device), and in particular, by one or more processors in the fusion device to implement the following steps:

[0057] S101, concatenating the Gaussian noise of the current step number with the medical image pair to be fused to obtain a three-channel tensor, and inputting the three-channel tensor into a trained noise predictor to output predicted noise.

[0058] In this embodiment, the fusion device may be a computing device with data processing capabilities, such as a laptop computer, a desktop computer, a tablet computer, a workstation or a server.

[0059] In this embodiment, for the Gaussian noise of the current step number, if the current step number is 1, the Gaussian noise is a randomly generated Gaussian noise that satisfies If the current step number is not 1, the Gaussian noise is the denoised image generated in the previous step number. For example, if the current step number is 5, the Gaussian noise is the denoised image generated in step number 4, and so on.

[0060] In this embodiment, the medical image pair is MRI images with different parameters, such as different scanning time, resolution, contrast, etc.

[0061] In this embodiment, after obtaining the Gaussian noise and the medical image pair to be fused, they are spliced ​​to obtain a three-channel tensor, and then the three-channel tensor is input into a trained noise predictor to output the predicted noise.

[0062] Before this, the noise predictor needs to be trained. The specific training process is as follows:

[0063] First, source image pairs as training sets are obtained, and the source image pairs are combined using a simple method to generate temporary truth images.

[0064] In this embodiment, since the temporary true value image can be regarded as a low-quality fusion image, the simple combination method used here should not be complicated, and efforts should be made to save computing power. For example, the corresponding pixel average value or maximum value retention can be used. This embodiment uses the average value retention, that is, each pixel retains the average value of the corresponding position pixel in the source image pair. The temporary true value image is expressed as follows:

[0065]

[0066] in: is a temporary substitute for the true value image; Represents a simple combinatorial function; I DWI ,I T2 Represents a source image pair.

[0067] Then, the temporary true value image is denoised by a forward diffusion process to generate a noisy image and corresponding noise;

[0068] Here, we set T = 1000, randomly select t between 1 and T, and randomly generate a Gaussian noise. The noise image can be expressed as follows according to the forward diffusion process formula:

[0069]

[0070] in, represents a noise image; represents the noise addition table, α t =1-β t , β t is the variable that adds the variance of the noise to the data at time t.

[0071] Finally, the noise-added image and the source image pair are concatenated into a three-channel tensor and input into the noise predictor for training until a preset training condition is met.

[0072] Specifically, the noise map and the source image pair are concatenated in the channel dimension into a three-channel tensor input to the noise predictor ∈ θ () for training, the noise predictor receives not only the three-channel tensor but also the time t and outputs the predicted noise

[0073] The structure of the noise predictor can be as follows: Figure 4As shown, the specific structure of the noise predictor of this embodiment is improved based on DiT in Scalable Diffusion Models with Transformers to adapt to the medical image fusion task. The specific change is that the label embedding vector used for classification in the original model that is irrelevant to the image fusion task is deleted, leaving only the time step vector t. In addition, the use of the original structure will cause the reverse denoising process to be confined to each patch block. Since the reverse denoising process is not a perfect inverse process of adding noise in the forward diffusion process, residual noise will inevitably remain. This causes the fused image to exhibit the same residual noise in each grid area of ​​the patch block size. To solve this problem, the Patchify layer is modified to use a vector that matches the number of elements in the patch block to receive information from each patch. And a residual convolution layer is added before the Patchify layer and after the corresponding Reshape layer to expand the receptive field. This ensures that the denoising process acts evenly on the entire image, rather than being confined to each block. See the settings of each network layer of the noise predictor for details. Figure 5 , where all Conv are convolutional layers with kernel size 3, stride 1, padding 1, and followed by batch normalization layer BN and activation function ReLU. The number of DiT-Blocks N = 4, and the multi-head self-attention Heads = 6. The embedding layer and other internal settings of DiT-Block are the same as the original settings.

[0074] In this embodiment, the loss calculation may be performed on the predicted noise to obtain the total loss, and then the total loss L may be subjected to gradient descent back propagation to optimize the noise predictor.

[0075] Specifically, the loss function is divided into two parts: denoising loss L diff With the guidance loss L guide . Thus, the total loss L is expressed as follows:

[0076] L=L guide +L diff ;

[0077] in:

[0078]

[0079] L guide =L SSIM +L int +L grad

[0080] L SSIM is the structural similarity loss, and

[0081] L intis the strength loss, and

[0082] L grad is the edge contour loss, and

[0083] The desired fusion image The estimated value of

[0084] In this embodiment, the above steps are repeatedly executed, and the training is stopped after a preset number of epochs on the training set, and the trained noise predictor ∈ θ ().

[0085] In this embodiment, during the training process of the noise predictor, a temporary true value image is used as the input of the noise predictor, so that the task difficulty of the entire framework is reduced from generating a desired fused image from Gaussian noise to shifting from the distribution of the temporary true value image to the distribution of the desired fused image, and the problem of being unable to achieve direct image fusion due to lack of groundtruth is solved.

[0086] In addition, by using the above-mentioned improved DiT as the noise predictor, the problems of insufficient receptive field and global feature loss of the existing diffusion model UNet noise predictor are solved.

[0087] S102, performing single-step reverse denoising on the predicted noise to obtain a denoised image of the current step number.

[0088] In this embodiment, the predicted noise is subjected to a single-step reverse denoising process to obtain a denoised image. The reverse denoising process formula is as follows:

[0089]

[0090] in:

[0091] is the standard deviation at time t,

[0092] z:

[0093] S103, when it is determined that the current number of steps reaches the preset number of steps, the denoised image of the current number of steps is output to obtain the required medical fusion image, otherwise the denoised image is set to the Gaussian noise of the next number of steps, and the process returns to step S101.

[0094] In this embodiment, the denoising image is used Instead of the Gaussian noise in step S102 Calculate again, and repeat this cycle until t=T, T-1, ..., 1, and stop at t=1. The output denoised image This is the required medical fusion image.

[0095] To further understand the present embodiment, the application of the present embodiment will be described below with practical examples.

[0096] 1. Experimental Dataset

[0097] Prostate MRI images of 552 patients provided by a city people's hospital were used as the original data set, and different parameter sequences were aligned. After all slices were extracted, 10,000 slices were randomly selected as the training set and 100 slices as the validation set. Each slice was a 320×320 8-bit bmp image.

[0098] 2. Evaluation indicators

[0099] This study uses five objective evaluation indicators to quantitatively evaluate the fusion performance of the algorithm, namely:

[0100] 2.1 Evaluation index based on mutual information and gradient similarity (Q AB / F ).

[0101] Q AB / F It is an evaluation index based on mutual information and gradient similarity. Gradient information is an important reflection of image structure information. By comparing the gradient information of the fused image with the source image, the degree to which the fused image retains structural information such as edges and textures can be evaluated. The retention of the brightness or intensity information of the fused image is also important. The performance of the fused image in terms of intensity information is measured by mutual information or correlation coefficient. Gradient similarity and intensity similarity are combined and weighted to calculate Q. AB / F , if Q AB / F The higher the value, the better the fused image is in preserving the source image information. AB / F The overall calculation formula is as follows:

[0102] Q AB / F =α·Q A / F +β·Q B / F

[0103] in:

[0104] Q A / F : The similarity between the fused image F and the source image A,

[0105] Q B / F : The similarity between the fused image F and the source image B,

[0106] α, β: Balance the weights of the two similarity indicators to satisfy α+β=1.

[0107] in, is the gradient strength of pixel (i, j) in images A and B, is the local gradient similarity, T is the smoothing factor to avoid the denominator being zero, usually T=1e -6 ,

[0108] 2.2 Sum of the Correlations of Differences (SCD)

[0109] SCD evaluates the preservation of source image detail information by the fused image by calculating the correlation of differential information between source images and the correlation of differential information between the fused image and each source image. If the fused image F has a high correlation with the differential information of both source images A and B, it means that the fused image effectively preserves the details and structural characteristics of the source images. The calculation formula of SCD is as follows:

[0110] SCD=r(ΔA,ΔF)+r(ΔB,ΔF)

[0111] in:

[0112] r(ΔA, ΔF), r(ΔB, ΔF): are the differential information correlations between the source images A and B and the fused image F. The differential correlation is usually defined as the Pearson correlation coefficient, and its formula is: Cov(ΔX,ΔY) is the covariance of the differential information ΔX and ΔY, σ ΔX ,σ ΔY is the standard deviation of the difference information.

[0113] 2.3 Structural Similarity Index SSIM (Structural Similarity Index Measure)

[0114] SSIM is a classic indicator for measuring image quality. It provides image similarity evaluation results that are consistent with human perception by combining brightness, contrast, and structural similarity. Its complete formula is:

[0115]

[0116] in:

[0117] μ x ,μ y : are the means of images x and y respectively.

[0118] σ x,σ y : are the standard deviations of images x and y respectively.

[0119] σ xy : Covariance of images x and y.

[0120] C 1 ,C 2 : Constant used to avoid zero denominator.

[0121] 2.4 Mutual Information

[0122] Mutual information (MI) is an image quality evaluation metric based on information theory, which is used to measure the amount of shared information between two images. It is often used in tasks such as image fusion and image registration to evaluate whether the fused image retains the information of the source image. If the mutual information between two images is high, it means that they share more information and the similarity between the images is strong; conversely, the mutual information is low, indicating that their correlation is weak. Mutual information is defined as the relationship between the joint probability distribution of two images and their respective marginal probability distributions. The specific formula is as follows:

[0123]

[0124] in:

[0125] X, Y: pixel values ​​of the two images.

[0126] p(x,y): joint probability distribution of images X and Y.

[0127] p(x): marginal probability distribution of image X.

[0128] p(y): marginal probability distribution of image Y.

[0129] 2.5 Correlation Coefficient CC

[0130] The correlation coefficient (CC) is a statistical indicator used to quantify the degree of similarity between two images. It reflects the degree of correlation between them by calculating the linear correlation between the corresponding pixel values ​​of the two images. The value range of CC is [-1,1]. CC = 1 means that the two images are completely positively correlated (the linear relationship is positive). CC = -1 means that the two images are completely negatively correlated (the linear relationship is negative). CC = 0 means that the two images are unrelated.

[0131] In image evaluation, a higher CC value (close to 1) usually indicates a higher similarity between two images. The calculation formula of the correlation coefficient is:

[0132]

[0133] in:

[0134] X={x 1 ,x 2 ,…,x N}: A set of pixel values ​​of the first image.

[0135] Y={y 1 ,y 2 ,…,y N}: A set of pixel values ​​of the second image.

[0136] N: total number of pixels.

[0137] μ X : The average pixel value of the first image X,

[0138] μ Y : The average pixel value of the second image Y,

[0139] 3. Experimental design and result analysis

[0140] This study conducts qualitative and quantitative experimental comparisons between this embodiment and seven commonly used image fusion methods, including NSST-PAPCNN, IFCNN, U2Fusion, DDcGAN, SwinFusion, Cddfuse, and DDFM. The fusion results of the three pairs of DWI-MRT2 images in the test set are compared as shown in the figure below. Figure 6 As shown. Here we are more concerned about the preservation of intensity, edge contours and structural features. Among them, NSST-PAPCNN over-emphasized the preservation of gradients, resulting in pseudo-structure enhancement on pair-2; IFCNN produced some blur at locations with large edge changes; U2Fusion lost a lot of texture details and almost only preserved the intensity; DDcGAN had high intensity on pair-1, low intensity on pair-2, and lost some texture details on pair-3; DDFM lost texture details on all three pairs of images, but the intensity was still preserved; It is worth noting that the difference between SwinFusion and CDDFuse and the method proposed in this embodiment is almost imperceptible visually, and there are differences only in quantitative analysis, such as Figure 7 shown.

[0141] In summary, this embodiment combines MRI medical images of different parameters of the same patient to construct a new image. The new image retains the complementary information between images of different parameters, and improves the fusion performance of prostate images by using temporary true value images and improved noise predictors, which can assist doctors to better identify and observe lesions, thereby improving the efficiency and accuracy of doctors' diagnosis and providing more timely and effective treatment for prostate patients.

[0142] See also Figure 8 The second embodiment of the present invention further provides a multi-parameter MRI medical image fusion device, which includes:

[0143] The predicted noise output unit 210 is used to concatenate the Gaussian noise of the current step number with the medical image pair to be fused to obtain a three-channel tensor, and input the three-channel tensor into a trained noise predictor to output the predicted noise; wherein the medical image pair is an MRI image with different parameters; during the training process of the noise predictor, the temporary true value image is used as the input of the noise predictor to reduce the task difficulty of the entire framework from generating the expected fused image from the Gaussian noise to shifting from the temporary true value image distribution to the expected fused image distribution;

[0144] A single-step reverse denoising unit 220, configured to perform single-step reverse denoising on the predicted noise to obtain a denoised image of the current step number;

[0145] The judgment unit 230 is used to output a denoised image of the current number of steps to obtain the required medical fusion image when it is judged that the current number of steps reaches a preset number of steps; otherwise, the denoised image is set to the Gaussian noise of the next step number, and the predicted noise output unit is notified.

[0146] The third embodiment of the present invention further provides a multi-parameter MRI medical image fusion device, which includes:

[0147] at least one processor; and,

[0148] a memory communicatively connected to the at least one processor; wherein,

[0149] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned multi-parameter MRI medical image fusion method.

[0150] The technical features of the above-described embodiments may be arbitrarily combined, and the execution order of the method steps is not limited. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0151] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A multi-parameter MRI medical image fusion method, characterized in that: include: S101, concatenating the Gaussian noise of the current step number with the medical image pair to be fused to obtain a three-channel tensor, and inputting the three-channel tensor into a trained noise predictor to output predicted noise; wherein the medical image pair is an MRI image with different parameters; during the training process of the noise predictor, using a temporary true value image as the input of the noise predictor reduces the task difficulty of the entire framework from generating an expected fused image from Gaussian noise to shifting from a temporary true value image distribution to an expected fused image distribution; S102, performing single-step reverse denoising on the predicted noise to obtain a denoised image of the current step number; S103, when it is determined that the current number of steps reaches the preset number of steps, the denoised image of the current number of steps is output to obtain the required medical fusion image, otherwise the denoised image is set to the Gaussian noise of the next number of steps, and the process returns to step S101.

2. The multi-parameter MRI medical image fusion method according to claim 1, characterized in that: In step S101, if the current step number is 1, the Gaussian noise is randomly generated Gaussian noise.

3. The multi-parameter MRI medical image fusion method according to claim 1, characterized in that: The training process of the noise predictor includes: Obtaining source image pairs as training sets, and using a simple combination method to generate temporary truth images from the source image pairs; Noising the temporary true value image through a forward diffusion process to generate a noisy image and corresponding noise; The noise-added image and the source image pair are concatenated into a three-channel tensor and input into a noise predictor for training until a preset training condition is met.

4. The multi-parameter MRI medical image fusion method according to claim 1, characterized in that: The temporary true value image is expressed as follows: in: is a temporary substitute for the true value image; Represents a simple combination method; I DWI ,I T2 Represents a source image pair.

5. The multi-parameter MRI medical image fusion method according to claim 4, characterized in that: Suppose t is randomly selected between 1 and T, and a Gaussian noise is randomly generated According to the forward diffusion The noise image generated by the program is expressed as: in, represents a noise image; represents the noise addition table, α t =1-β t , β t is the variable that adds the variance of the noise to the data at time t.

6. The multi-parameter MRI medical image fusion method according to claim 5, characterized in that: When training the noise predictor, also include: The predicted noise output by the noise predictor is subjected to loss calculation to obtain a total loss L, and the total loss L is subjected to gradient descent back propagation to optimize the noise predictor; wherein the loss function includes a denoising loss L diff With the guidance loss L guide ; The formula of total loss L is as follows: L=L guide +L diff L guide =L SSIM +L int +L grad in: L SSIM is the structural similarity loss, and L int is the strength loss, and L grad is the edge contour loss, and The desired fusion image The estimated value of 7. The multi-parameter MRI medical image fusion method according to claim 6, characterized in that: In step S102, the denoised image obtained by the single-step reverse denoising process It is expressed as: in: is the standard deviation at time t, z: ift>1,elsez=0。

Images