Medical image reconstruction method and system based on consistency constraint diffusion model

Abstract

The invention discloses a medical image reconstruction method and system based on a consistency constraint diffusion model, and the method comprises the steps: carrying out the rapid reconstruction processing of a low-resolution pathological image based on a regression algorithm, and obtaining a to-be-optimized input pathological image; introducing a consistency feature intensifier module and a multi-scale feature mixer module, and constructing a consistency constraint diffusion model; detail optimization denoising processing is carried out on the to-be-optimized input pathological image through the consistency constraint diffusion model, and a super-resolution pathological image is obtained; and carrying out comparison loss calculation on the super-resolution pathology image and the target pathology image through an L1 loss function to obtain a super-resolution pathology reconstruction image. According to the embodiment of the invention, the super-resolution pathological reconstruction image with high fidelity can be output, and the characteristics of the super-resolution pathological reconstruction image and the characteristics of the input low-resolution image are kept consistent. The medical image reconstruction method and system based on the consistency constraint diffusion model can be widely applied to the technical field of medical image reconstruction.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image reconstruction, and in particular to a medical image reconstruction method and system based on a consistency constrained diffusion model. Background Art

[0002] Related image processing methods for generating high-resolution images from low-resolution images can be divided into two categories: regression-based methods and deep generative methods. In the field of pathological image super-resolution, most of the competitive methods are based on regression. Regression methods directly learn the mapping between low resolution (LR) and high resolution (HR) and tend to calculate the average of possible results, which leads to overly smooth output and lack of high-frequency details. Considering that pathological images have more complex structure and texture information than natural images, regression methods are not suitable for reconstructing pathological images. Deep generative methods represented by generative adversarial networks (GANs) have been shown to have the ability to learn complex distributions of images. However, due to the defects of optimization instability and mode collapse, the application of GANs in the field of pathological image super-resolution is limited. In addition, as an emerging technology, diffusion models have attracted widespread attention from researchers due to their powerful ability to generate high-fidelity images and stable optimization. However, the existing diffusion models cannot achieve high-fidelity restoration for complex spatial pathological images. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a medical image reconstruction method and system based on a consistency constrained diffusion model, which can output high-fidelity super-resolution pathology reconstructed images and maintain consistency with the input low-resolution image features.

[0004] The first technical solution adopted by the present invention is: a medical image reconstruction method based on a consistency constrained diffusion model, comprising the following steps:

[0005] Based on the regression algorithm, the low-resolution pathological image is quickly reconstructed to obtain the input pathological image to be optimized;

[0006] Introduce the consistency feature enhancer module and the multi-scale feature mixer module to build a consistency constrained diffusion model;

[0007] The input pathological image to be optimized is subjected to detail optimization and denoising through a consistency constrained diffusion model to obtain a super-resolution pathological image.

[0008] The super-resolution pathology image is compared with the target pathology image through the L1 loss function to obtain a super-resolution pathology reconstructed image.

[0009] Furthermore, the step of rapidly reconstructing the low-resolution pathological image based on the regression algorithm to obtain the input pathological image to be optimized specifically includes:

[0010] Acquire low-resolution pathology images;

[0011] Perform bicubic interpolation upsampling on the low-resolution pathological image to obtain an interpolated pathological image;

[0012] The low-resolution pathological image is reconstructed by a super-resolution model based on regression to obtain a preliminary pathological reconstructed image;

[0013] The interpolated pathological image and the preliminary pathological reconstruction image are integrated to obtain the input pathological image to be optimized.

[0014] Furthermore, the step of reconstructing the low-resolution pathological image by using a regression-based super-resolution model to obtain a preliminary pathological reconstructed image specifically includes:

[0015] The low-resolution pathology images are fed into a regression-based super-resolution model;

[0016] Perform coding mapping processing on low-resolution pathological images to obtain three-dimensional characteristic pathological images;

[0017] Performing feature fusion transformation processing on the three-dimensional characteristic pathology image to obtain a fused three-dimensional characteristic pathology image;

[0018] The spatial resolution of the fused three-dimensional characteristic pathological image is reconstructed to obtain a preliminary pathological reconstructed image.

[0019] Furthermore, the consistency constrained diffusion model includes a spatial attention mechanism module, a residual network module, a consistency feature enhancer module and a multi-scale feature mixer module, wherein:

[0020] The consistency feature enhancer module includes a first normalization layer, a multi-head cross attention mechanism module, a second normalization layer and a multi-layer perceptron layer;

[0021] The multi-scale feature mixer module includes a parallel hole convolution layer module, a hole convolution layer, a third normalization layer, a first channel MLP layer, a fourth normalization layer, a convolution layer, a fifth normalization layer and a second channel MLP layer.

[0022] Furthermore, the step of performing detail optimization and denoising processing on the input pathological image to be optimized by using the consistency constrained diffusion model to obtain a super-resolution pathological image specifically includes:

[0023] Inputting the input pathological image to be optimized into the consistency constrained diffusion model;

[0024] The residual network module based on the consistency constrained diffusion model performs feature extraction processing on the input pathological image to be optimized to obtain the input pathological image features to be optimized;

[0025] A consistency feature enhancer module based on a consistency constrained diffusion model performs consistency constrained feature enhancement processing on the input pathology image features to be optimized to obtain enhanced input pathology image features;

[0026] Based on the spatial attention mechanism module of the consistency constrained diffusion model, the spatial feature extraction process is performed on the enhanced input pathological image features to obtain the input pathological image spatial features;

[0027] A multi-scale feature mixer module based on the consistency constrained diffusion model performs multi-scale feature fusion processing on the spatial features of the input pathological image to obtain a super-resolution pathological image.

[0028] Furthermore, the expression of the detail optimization denoising process is specifically as follows:

[0029]

[0030] In the above formula, ∈ θ represents the noise predictor, x rc represents the preliminary pathological reconstruction image, x up represents the interpolated pathological image, t represents the current iteration number, y t represents the denoised image outputted by the t-th iteration, y t-1 represents the denoised image outputted by the iterative step t-1, α t represents the hyperparameter, represents the variance of the prediction noise distribution.

[0031] Furthermore, the consistency feature enhancer module based on the consistency constraint diffusion model performs consistency constraint feature enhancement processing on the input pathological image features to be optimized to obtain the enhanced input pathological image features, which specifically includes:

[0032] Inputting the input pathological image features to be optimized into the consistency feature enhancer module of the consistency constrained diffusion model;

[0033] Based on the first normalization layer of the consistency feature enhancer module, the input pathology image features to be optimized are normalized to obtain the normalized input pathology image features;

[0034] Based on the multi-head cross attention mechanism module of the consistency feature enhancer module, the normalized input pathological image features are modeled for correlation, and the input pathological image features with correlation structure are obtained;

[0035] Based on the second normalization layer of the consistency feature enhancer module, the input pathology image features with the correlation structure are subjected to secondary normalization to obtain the standard input pathology image features;

[0036] Based on the multi-layer perceptron layer of the consistency feature enhancer module, feature mapping is performed on the standard input pathology image features to obtain enhanced input pathology image features.

[0037] Furthermore, the multi-scale feature mixer module based on the consistency constrained diffusion model performs multi-scale feature fusion processing on the input pathological image spatial features to obtain a super-resolution pathological image, which specifically includes:

[0038] Inputting the input pathological image spatial features into the multi-scale feature mixer module of the consistency constrained diffusion model;

[0039] Based on the parallel hole convolution layer module of the multi-scale feature mixer module, multi-scale feature extraction processing is performed on the input pathological image spatial features to obtain multi-scale input pathological image spatial features;

[0040] Based on the dilated convolution layer of the multi-scale feature mixer module, the multi-scale input pathology image spatial features are processed with an expanded receptive field to obtain the expanded multi-scale input pathology image spatial features;

[0041] Based on the third normalization layer and the first channel MLP layer of the multi-scale feature mixer module, the expanded multi-scale input pathology image spatial features are normalized and reconstructed to obtain preliminary multi-scale input pathology image spatial reconstruction features;

[0042] Based on the fourth normalization layer and convolution layer of the multi-scale feature mixer module, the preliminary multi-scale input pathology image spatial reconstruction features are normalized and feature extracted to obtain the multi-scale input pathology image spatial reconstruction features;

[0043] Based on the fifth normalization layer and the second channel MLP layer of the multi-scale feature mixer module, the spatial reconstruction features of the multi-scale input pathology image are standardized and fused to obtain a super-resolution pathology image.

[0044] Furthermore, it also includes setting the denoising process function of the consistency constraint diffusion model, and its expression is:

[0045]

[0046] p(y T )=N(y T ; 0, 1)

[0047] p θ (yt-1 |y t , x) = N(y t -1|μ θ (x,y t , t), σ 2 I)

[0048] In the above formula, p θ (y t-1 |y t ,x)、p θ (y 0：T |x), p(y T ) represents the conditional distribution of pathological image reconstruction learning, T represents the length of the fixed Markov chain, I represents the identity matrix, y T represents random Gaussian noise, N(·) represents normal distribution, μ θ and σ represent the mean and variance of this normal distribution.

[0049] The second technical solution adopted by the present invention is: a medical image reconstruction system based on a consistency constrained diffusion model, comprising:

[0050] A reconstruction module, used for quickly reconstructing low-resolution pathological images based on a regression algorithm to obtain an input pathological image to be optimized;

[0051] A construction module is used to introduce a consistency feature enhancer module and a multi-scale feature mixer module to construct a consistency constraint diffusion model;

[0052] An optimization module, used for performing detail optimization processing on the input pathological image to be optimized through a consistency constrained diffusion model to obtain a super-resolution pathological image;

[0053] The optimization module is used to calculate the contrast loss between the super-resolution pathology image and the target pathology image through the L1 loss function to obtain a super-resolution pathology reconstructed image.

[0054] The beneficial effects of the method and system of the present invention are as follows: the present invention rapidly reconstructs low-resolution pathological images based on a regression algorithm, and based on the powerful ability of the regression algorithm to accurately represent continuous images and allow the input image to be reconstructed to any desired resolution, further introduces a consistency feature enhancer module and a multi-scale feature mixer module, constructs a consistency constrained diffusion model, and performs detail optimization and denoising processing on the input pathological image to be optimized. Through the consistency feature enhancer module, a consistency constraint is imposed in the back diffusion process, so that the consistency constrained diffusion model can generate results that are highly consistent with the input low-resolution image. The multi-scale feature mixer module can effectively capture spatial features with long-term dependencies, and simultaneously fuse spatial and channel domain features, thereby enhancing the representation capability of the consistency constrained diffusion model, promoting image reconstruction, and being able to output a super-resolution pathological reconstructed image with high fidelity and maintaining consistency with the input low-resolution image features. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the steps of a medical image reconstruction method based on a consistency constrained diffusion model of the present invention;

[0056] Figure 2 It is a structural block diagram of a medical image reconstruction system based on a consistency constrained diffusion model of the present invention;

[0057] Figure 3 is a structural schematic diagram of a medical image reconstruction model provided by a specific embodiment of the present invention;

[0058] Figure 4 It is a structural diagram of a consistency-constrained diffusion model provided by a specific embodiment of the present invention;

[0059] Figure 5 is a structural diagram of a consistency feature enhancer module provided in a specific embodiment of the present invention;

[0060] Figure 6 It is a schematic diagram of the structure of a multi-scale feature mixer module provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0062] First, the technical terms of the embodiments of the present invention need to be explained:

[0063] Whole Slide Imaging is a digital imaging technology used to convert tissues on traditional glass slides into digital formats, namely Whole Slide Images (WSIs). Digitized pathology images can be stored, shared, and accessed online, helping pathologists with remote diagnosis, education, and collaboration. However, high-resolution WSIs need to be acquired through high-magnification scanners, which are usually very expensive and time-consuming. In addition, the huge file size of high-resolution WSIs (ranging from several GB to tens of GB) also increases the difficulty of data transmission, storage, and management. These challenges limit the application of WSIs in some medical institutions with limited resources.

[0064] Single Image Super-Resolution (SISR) is an image processing technique that aims to generate a high-resolution image from a low-resolution image. SISR improves the clarity of an image by inferring and supplementing missing high-frequency details, making the image look more realistic.

[0065] Reference Figure 1 The present invention provides a medical image reconstruction method based on a consistency constrained diffusion model, the method comprising the following steps:

[0066] S100, quickly reconstructing the low-resolution pathological image based on a regression algorithm to obtain an input pathological image to be optimized;

[0067] Specifically, a low-resolution pathological image is acquired; a bicubic interpolation upsampling process is performed on the low-resolution pathological image to obtain an interpolated pathological image; the low-resolution pathological image is reconstructed through a regression-based super-resolution model to obtain a preliminary pathological reconstructed image; the interpolated pathological image and the preliminary pathological reconstructed image are integrated to obtain an input pathological image to be optimized.

[0068] Among them, it should be noted that the low-resolution pathological image is input into the regression-based super-resolution model; the low-resolution pathological image is coded and mapped to obtain a three-dimensional characteristic pathological image; the three-dimensional characteristic pathological image is feature fused and transformed to obtain a fused three-dimensional characteristic pathological image; the fused three-dimensional characteristic pathological image is spatially reconstructed to obtain a preliminary pathological reconstructed image.

[0069] S200, introduce the consistency feature enhancer module and the multi-scale feature mixer module to build a consistency constraint diffusion model;

[0070] Specifically, the consistency constrained diffusion model includes a spatial attention mechanism module, a residual network module, a consistency feature enhancer module and a multi-scale feature mixer module, where Figure 5As shown in , the consistency feature enhancer module includes a first normalization layer, a multi-head cross attention mechanism module, a second normalization layer and a multi-layer perceptron layer, as shown in Figure 6 As shown, the multi-scale feature mixer module includes a parallel hole convolution layer module, a hole convolution layer, a third normalization layer, a first channel MLP layer, a fourth normalization layer, a convolution layer, a fifth normalization layer and a second channel MLP layer.

[0071] It should be noted that, considering the importance of pathological images in clinical diagnosis, the consistency of the reconstructed pathological images will directly affect the diagnosis of pathologists. Therefore, the embodiment of the present invention hopes that the reconstructed image has high fidelity while maintaining a high degree of consistency with the input low-resolution image. To achieve this goal, a consistency constrained diffusion (CorD) model is further developed. The model structure is as follows Figure 4 As shown, its core is the consistency feature enhancer (CFEnhancer) module. The CFEnhancer module imposes consistency constraints during the back-diffusion process, enabling CorD to generate results that are highly consistent with the input low-resolution image. In addition, an embodiment of the present invention also proposes a module based on convolution and multilayer perceptron (MLP), called a multi-scale feature mixer (MRFMixer), and integrates it into CorD. MRFMixer can effectively capture spatial features with long-term dependencies, while fusing spatial and channel domain features, thereby enhancing the representation ability of CorD and promoting image reconstruction.

[0072] S300, performing detail optimization and denoising processing on the input pathological image to be optimized by using a consistency constrained diffusion model to obtain a super-resolution pathological image;

[0073] Specifically, the input pathological image to be optimized is input into the consistency constrained diffusion model; the residual network module based on the consistency constrained diffusion model performs feature extraction processing on the input pathological image to be optimized to obtain the input pathological image features to be optimized; the consistency feature enhancer module based on the consistency constrained diffusion model performs consistency constrained feature enhancement processing on the input pathological image features to be optimized to obtain enhanced input pathological image features; the spatial attention mechanism module based on the consistency constrained diffusion model performs spatial feature extraction processing on the enhanced input pathological image features to obtain the input pathological image spatial features; the multi-scale feature mixer module based on the consistency constrained diffusion model performs multi-scale feature fusion processing on the spatial features of the input pathological image to obtain a super-resolution pathological image.

[0074] Among them, it should be noted that the input pathological image features to be optimized are input into the consistency feature enhancer module of the consistency constrained diffusion model; based on the first normalization layer of the consistency feature enhancer module, the input pathological image features to be optimized are normalized to obtain the normalized input pathological image features; based on the multi-head cross attention mechanism module of the consistency feature enhancer module, the normalized input pathological image features are subjected to correlation modeling processing to obtain the input pathological image features with correlation structure; based on the second normalization layer of the consistency feature enhancer module, the input pathological image features with correlation structure are subjected to secondary normalization processing to obtain the standard input pathological image features; based on the multi-layer perceptron layer of the consistency feature enhancer module, the standard input pathological image features are subjected to feature mapping processing to obtain the enhanced input pathological image features.

[0075] Among them, it should be noted that the input pathological image spatial features are input into the multi-scale feature mixer module of the consistency constrained diffusion model; based on the parallel hole convolution layer module of the multi-scale feature mixer module, the input pathological image spatial features are subjected to multi-scale feature extraction processing to obtain the multi-scale input pathological image spatial features; based on the hole convolution layer of the multi-scale feature mixer module, the multi-scale input pathological image spatial features are subjected to extended receptive field processing to obtain the extended multi-scale input pathological image spatial features; based on the third normalization layer and the first channel MLP layer of the multi-scale feature mixer module, the extended multi-scale input pathological image spatial features are subjected to standardization and feature reconstruction processing to obtain the preliminary multi-scale input pathological image spatial reconstruction features; based on the fourth normalization layer and convolution layer of the multi-scale feature mixer module, the preliminary multi-scale input pathological image spatial reconstruction features are subjected to standardization and feature extraction processing to obtain the multi-scale input pathological image spatial reconstruction features; based on the fifth normalization layer and the second channel MLP layer of the multi-scale feature mixer module, the multi-scale input pathological image spatial reconstruction features are subjected to standardization and feature fusion processing to obtain the super-resolution pathological image.

[0076] In this embodiment, in order to use CorD more efficiently, the embodiment of the present invention proposes a simple and effective two-stage framework, namely, the pathology image reconstruction with consistency constraints (CorPiR) framework, for super-resolution pathology image reconstruction. Figure 3 As shown in Figure 2, CorPiR integrates regression-based methods with CorD into an end-to-end framework. In the first stage, the input low-resolution image is quickly reconstructed using a regression-based method to reduce the difference between the input distribution and the target distribution. By doing so, CorD can focus on recovering the detail information of the image, thereby generating realistic and consistent images more efficiently in the second stage.

[0077] S400 , performing a comparison loss calculation on the super-resolution pathology image and the target pathology image using an L1 loss function to obtain a super-resolution pathology reconstructed image.

[0078] Specifically, first, given an input low-resolution image x, the regression-based super-resolution method LIIF is used in the first stage to perform a preliminary reconstruction of x and obtain the preliminary reconstructed image x rc LIIF is chosen because of its strong ability to accurately represent continuous images and allows the input image x to be reconstructed to any desired resolution. Then in the second stage, we use the CorD model to further optimize the preliminary reconstructed image x rc CorD uses a fixed Markov chain of length T to learn an approximate distribution of the data distribution p(y|x), where x represents the preliminary reconstructed image x rc , and y represents the output super-resolution image. We define the forward Markov diffusion process q as the process of continuously adding Gaussian noise to the true high-resolution image, which is as follows:

[0079]

[0080] where β t is a hyperparameter that is preset to a constant and decays continuously, β t ∈(0,1) and β1<β2<…<β T . I is the identity matrix. Let α t : =1-β t , given y0, y t The distribution of can be expressed as:

[0081]

[0082] in During the inverse diffusion process, CorPiR learns the conditional distribution p θ (y t-1 |y t , x) to the random Gaussian noise y T ∈N(0, I) denoising to get y0, the denoising process can be defined as:

[0083]

[0084] p(y T )=N(y T ; 0, 1)

[0085] p θ (y t-1 |y t , x) = N(y t-1 |μ θ (x,y t, t), σ 2 I)

[0086] where μ θ is a t , t-related functions, and σ is a preset value, which is set to (1-α t ), where α t : =1-β t , while β t is a hyperparameter that is preset to a constant and decays continuously, β t ∈(0,1) and β1<β2<…<β T .

[0087] We obtain the output y0 through the above denoising process and compare it with the true value using the L1 loss function. Generally speaking, the second stage of detail optimization will go through a series of iterative denoising steps, starting from the conditional input x rc and consistency constraint input x up Recover a high-fidelity and consistent super-resolution image y0 from the Gaussian noise of rc Detail optimization.

[0088] The process of training CorPiR is equivalent to optimizing CorD, because the first stage uses a pre-trained model and the parameters remain unchanged during the training process. Our goal is to The target image y0 is restored by optimizing the conditional noise predictor ∈ θ To achieve this goal, we typically optimize the diffusion model by optimizing the negative log-likelihood of the variational lower bound obtained by variational inference, as follows:

[0089]

[0090] where ∈~N(0,I) is the noise added in the t-th step of the diffusion process, and t∈{1,...,T}.

[0091] The reasoning process of our framework is divided into two stages according to its structure. In the first stage, we use the pre-trained LIIF model to perform preliminary reconstruction of the low-resolution image x to obtain the target resolution image x rc In the second stage, our inference process is based on the Gaussian conditional distribution p learned during the optimization process. θ (y t-1 |y t , x), defined as p θ (y t-1 |y t , x) formula. According to the previous related methods, given input y t , we use the trained conditional noise predictor ∈θ To estimate ∈, and obtain p θ (y t-1 |y t , the mean of x):

[0092]

[0093] And p θ (y t-1 |y t , the variance of x) is set to (1-α t ).

[0094] Based on the above parameterization, each iterative step in the second stage reasoning process can be expressed as follows:

[0095]

[0096] In the above formula, let α t : =1-β t , where β t is a hyperparameter that is preset to a constant and decays continuously, β t ∈(0,1) and β1<β2<…<β T . I is the identity matrix.

[0097] where ∈ t ~N(0,I). After t steps of denoising, the super-resolution image y0 is finally obtained.

[0098] In summary, in order to verify the effectiveness of the super-resolution framework proposed in the embodiment of the present invention, we evaluated the image reconstruction results of CorPiR on pathological image datasets with different magnifications. Table 1 shows the reconstruction result indicators of CorPiR at different magnifications and upsampling factors. We also compared CorPiR with other competitive methods on the test set with 40x magnification, and the results are shown in Table 2.

[0099] Table 1 CorPiR at different magnifications (40x, 30x, 20x, 10x, and 5x) and upsampling factors (X2, X4, and

[0100] X8) super-resolution reconstruction performance data table

[0101]

[0102] It should be noted that P, S, L, and C represent PSNR, SSIM, LPIPS, and Consistency indicators, respectively. ↑ means the higher the better, and ↓ means the lower the better.

[0103] Table 2 Comparison of reconstruction performance of three upsampling factors (X2, X4, X8) on the 40x test set

[0104]

[0105]

[0106] Where * represents the deep generation method.

[0107] Reference Figure 2 , a medical image reconstruction system based on consistency constrained diffusion model, comprising:

[0108] A reconstruction module 201 is used to quickly reconstruct the low-resolution pathological image based on a regression algorithm to obtain an input pathological image to be optimized;

[0109] A construction module 202 is used to introduce a consistency feature enhancer module and a multi-scale feature mixer module to construct a consistency constraint diffusion model;

[0110] An optimization module 203 is used to perform detail optimization processing on the input pathological image to be optimized by using a consistency constrained diffusion model to obtain a super-resolution pathological image;

[0111] The optimization module 204 is used to perform a comparison loss calculation between the super-resolution pathology image and the target pathology image through an L1 loss function to obtain a super-resolution pathology reconstructed image.

[0112] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0113] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A medical image reconstruction method based on a consistency constrained diffusion model, characterized in that: The following steps are involved: Based on the regression algorithm, the low-resolution pathological image is quickly reconstructed to obtain the input pathological image to be optimized; Introduce the consistency feature enhancer module and the multi-scale feature mixer module to build a consistency constrained diffusion model; The input pathological image to be optimized is subjected to detail optimization and denoising through a consistency constrained diffusion model to obtain a super-resolution pathological image. The super-resolution pathology image is compared with the target pathology image through the L1 loss function to obtain a super-resolution pathology reconstructed image.

2. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 1, characterized in that: The step of rapidly reconstructing the low-resolution pathological image based on the regression algorithm to obtain the input pathological image to be optimized specifically includes: Acquire low-resolution pathology images; Perform bicubic interpolation upsampling on the low-resolution pathological image to obtain an interpolated pathological image; The low-resolution pathological image is reconstructed by a super-resolution model based on regression to obtain a preliminary pathological reconstructed image; The interpolated pathological image and the preliminary pathological reconstruction image are integrated to obtain the input pathological image to be optimized.

3. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 2, characterized in that: The step of reconstructing the low-resolution pathological image by using a regression-based super-resolution model to obtain a preliminary pathological reconstructed image specifically includes: The low-resolution pathology images are fed into a regression-based super-resolution model; Perform coding mapping processing on low-resolution pathological images to obtain three-dimensional characteristic pathological images; Performing feature fusion transformation processing on the three-dimensional characteristic pathology image to obtain a fused three-dimensional characteristic pathology image; The spatial resolution of the fused three-dimensional characteristic pathological image is reconstructed to obtain a preliminary pathological reconstructed image.

4. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 3, characterized in that: The consistency constrained diffusion model includes a spatial attention mechanism module, a residual network module, a consistency feature enhancer module and a multi-scale feature mixer module, wherein: The consistency feature enhancer module includes a first normalization layer, a multi-head cross attention mechanism module, a second normalization layer and a multi-layer perceptron layer; The multi-scale feature mixer module includes a parallel hole convolution layer module, a hole convolution layer, a third normalization layer, a first channel MLP layer, a fourth normalization layer, a convolution layer, a fifth normalization layer and a second channel MLP layer.

5. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 4, characterized in that: The step of performing detail optimization and denoising processing on the input pathological image to be optimized by using the consistency constrained diffusion model to obtain a super-resolution pathological image specifically includes: Inputting the input pathological image to be optimized into the consistency constrained diffusion model; The residual network module based on the consistency constrained diffusion model performs feature extraction processing on the input pathological image to be optimized to obtain the input pathological image features to be optimized; A consistency feature enhancer module based on a consistency constrained diffusion model performs consistency constrained feature enhancement processing on the input pathology image features to be optimized to obtain enhanced input pathology image features; Based on the spatial attention mechanism module of the consistency constrained diffusion model, the spatial feature extraction process is performed on the enhanced input pathological image features to obtain the input pathological image spatial features; A multi-scale feature mixer module based on the consistency constrained diffusion model performs multi-scale feature fusion processing on the spatial features of the input pathological image to obtain a super-resolution pathological image.

6. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 5, characterized in that: The expression of the detail optimization denoising process is specifically as follows: In the above formula, ∈ θ represents the noise predictor, x rc represents the preliminary pathological reconstruction image, x up represents the interpolated pathological image, t represents the current iteration number, y t represents the denoised image outputted by the t-th iteration, y t-1 represents the denoised image outputted by the iterative step t-1, α t represents the hyperparameter, represents the variance of the prediction noise distribution.

7. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 6, characterized in that: The consistency feature enhancer module based on the consistency constraint diffusion model performs consistency constraint feature enhancement processing on the input pathology image features to be optimized to obtain the enhanced input pathology image features. This step specifically includes: Inputting the input pathological image features to be optimized into the consistency feature enhancer module of the consistency constrained diffusion model; Based on the first normalization layer of the consistency feature enhancer module, the input pathology image features to be optimized are normalized to obtain the normalized input pathology image features; Based on the multi-head cross attention mechanism module of the consistency feature enhancer module, the normalized input pathological image features are modeled for correlation, and the input pathological image features with correlation structure are obtained; Based on the second normalization layer of the consistency feature enhancer module, the input pathology image features with the correlation structure are subjected to secondary normalization to obtain the standard input pathology image features; Based on the multi-layer perceptron layer of the consistency feature enhancer module, feature mapping is performed on the standard input pathology image features to obtain enhanced input pathology image features.

8. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 7, characterized in that: The multi-scale feature mixer module based on the consistency constrained diffusion model performs multi-scale feature fusion processing on the input pathological image spatial features to obtain a super-resolution pathological image, which specifically includes: Inputting the input pathological image spatial features into the multi-scale feature mixer module of the consistency constrained diffusion model; Based on the parallel hole convolution layer module of the multi-scale feature mixer module, multi-scale feature extraction processing is performed on the input pathological image spatial features to obtain multi-scale input pathological image spatial features; Based on the dilated convolution layer of the multi-scale feature mixer module, the multi-scale input pathology image spatial features are processed with an expanded receptive field to obtain the expanded multi-scale input pathology image spatial features; Based on the third normalization layer and the first channel MLP layer of the multi-scale feature mixer module, the expanded multi-scale input pathology image spatial features are normalized and reconstructed to obtain preliminary multi-scale input pathology image spatial reconstruction features; Based on the fourth normalization layer and convolution layer of the multi-scale feature mixer module, the preliminary multi-scale input pathology image spatial reconstruction features are normalized and feature extracted to obtain the multi-scale input pathology image spatial reconstruction features; Based on the fifth normalization layer and the second channel MLP layer of the multi-scale feature mixer module, the spatial reconstruction features of the multi-scale input pathology image are standardized and fused to obtain a super-resolution pathology image.

9. The medical image reconstruction method based on the consistency constrained diffusion model according to claim 8, characterized in that: It also includes the denoising process function for setting the consistency constraint diffusion model, which is expressed as: p(and T )=N(and T ;0 ，I) p θ (y t-1 |y t ,x)=N(y t-1 |μ θ (x,y t ,t),σ 2 I) In the above formula, p θ (y t-1 |y t ,x)、p θ (y 0：T |x), p(y T ) represents the conditional distribution of pathological image reconstruction learning, T represents the length of the fixed Markov chain, I represents the identity matrix, y T represents random Gaussian noise, N(·) represents normal distribution, μ θ and σ represent the mean and variance of this normal distribution.

10. A medical image reconstruction system based on a consistency constrained diffusion model, characterized in that: Includes the following modules: A reconstruction module, used for quickly reconstructing low-resolution pathological images based on a regression algorithm to obtain an input pathological image to be optimized; A construction module is used to introduce a consistency feature enhancer module and a multi-scale feature mixer module to construct a consistency constraint diffusion model; An optimization module, used for performing detail optimization processing on the input pathological image to be optimized through a consistency constrained diffusion model to obtain a super-resolution pathological image; The optimization module is used to calculate the contrast loss between the super-resolution pathology image and the target pathology image through the L1 loss function to obtain a super-resolution pathology reconstructed image.

Images