Medical image super-resolution method and device based on edge-enhanced diffusion model

Abstract

The present invention discloses a medical image processing method, specifically, a medical image super-resolution method and device based on an edge-enhanced diffusion model. The present invention first extracts image edges from the predicted image, low-resolution image, and auxiliary image at the t-th time step to obtain target edge information, and then encodes the image and edge respectively to generate enhanced disentangled unique features and shared features; adaptively integrates effective auxiliary information of different modalities, and gives full play to the complementarity of multimodal data to enhance feature expression capabilities; uses a decoder to generate a noise image in a diffusion model, and reconstructs a high-resolution target image through a step-by-step denoising reverse diffusion process, thereby restoring more detail information. In the embodiment of the present invention, the importance of edge information is taken into account, and high-frequency edge information is injected into the corresponding image encoder to help capture details more effectively and obtain better super-resolution image reconstruction.

Description

Technical Field

[0001] The present invention relates to a medical image processing method, and in particular to a medical image super-resolution method and device based on an edge enhancement diffusion model. Background Art

[0002] Magnetic resonance imaging (MRI) is an advanced medical imaging technology. It has become an important tool in medical diagnosis due to its non-invasiveness, high soft tissue contrast, rich imaging parameters and no ionizing radiation. MRI not only has unique advantages in the diagnosis of lesions in the nervous system, cardiovascular, abdomen and joints, but also plays an indispensable role in basic research such as brain science and pathology. Compared with computed tomography (CT), MRI has obvious advantages in imaging resolution and soft tissue imaging ability. It can accurately display the lesion area and is widely used in non-invasive examinations of tissue damage, degenerative lesions, tumors, etc., especially in the diagnosis of head and heart lesions. It shows extremely high clinical value. However, MRI images often face the problem of low resolution, which not only affects the accuracy of diagnosis, but also increases the difficulty of post-processing. In order to obtain high-quality MRI images, it usually requires a longer acquisition time, higher magnetic field strength and higher cost, and also requires patients to maintain a fixed posture for a long time, causing discomfort.

[0003] At present, there are two main methods to improve the resolution of MRI images: one is to improve the signal-to-noise ratio and spatial resolution by hardware upgrades such as increasing the magnetic field strength and optimizing the gradient coil, but such methods are costly and have a long update cycle; the other is to enhance image details and clarity in a more flexible and low-cost way through image processing techniques such as super-resolution reconstruction. In recent years, with the development of deep learning and the emergence of diffusion models, MRI super-resolution technology has been significantly improved. However, existing methods often ignore high-frequency information in the image and fail to effectively integrate auxiliary information in the reference image, resulting in blurred edges and obvious artifacts in the generated high-resolution images. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the above-mentioned methods commonly used in the prior art often ignore the high-frequency information in the image and fail to effectively integrate the auxiliary information in the reference image, resulting in blurred edges and obvious artifacts in the generated high-resolution image. In order to solve the above-mentioned problem, the present invention provides a medical image super-resolution method and device based on an edge enhancement diffusion model.

[0005] The content of the present invention includes:

[0006] In a first aspect, an embodiment of the present invention provides a medical image super-resolution method based on an edge-enhanced diffusion model, comprising:

[0007] In the forward stage of the diffusion model, Gaussian noise is gradually added to the image to be processed to obtain the perturbation image at the tth time step, wherein the perturbation image at the 0th time step is the image to be processed, and the perturbation image at the Tth time step is a pure noise image. ;

[0008] In the reverse phase of the diffusion model, the pure noise image is subjected to a stepwise denoising process to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image;

[0009] The predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, including:

[0010] Performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information, wherein the target edge information includes edge information of the predicted image, the low-resolution image and the auxiliary image at the t-th time step;

[0011] Using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature;

[0012] Obtaining, by a detangling operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature;

[0013] Using K-selected multi-head attention mechanism and gating mechanism to fuse the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature;

[0014] The fused features are decoded using a noise image decoder to obtain the predicted image at the t-1th time step.

[0015] Optionally, performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information includes:

[0016] Using an edge information extractor to extract edge information from the predicted image at the t-th time step, the low-resolution image, and the auxiliary image, respectively, to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image;

[0017] The edge information extractor includes a first Sobel operator, a second Sobel operator, a third Sobel operator, a fourth Sobel operator, a first Laplacian operator and a second Laplacian operator.

[0018] Optionally, the extracting edge information of the predicted image at the t-th time step, the low-resolution image and the auxiliary image respectively using an edge information extractor to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image and edge information of the auxiliary image comprises:

[0019] Using the first Sobel operator to perform feature extraction on the input image along the vertical direction to obtain a first edge feature, using the second Sobel operator to perform feature extraction on the input image along the horizontal direction to obtain a second edge feature, using the third Sobel operator to perform feature extraction on the input image along the diagonal direction to obtain a third edge feature, using the fourth Sobel operator to perform feature extraction on the input image along the anti-diagonal direction to obtain a fourth edge feature, using the first Laplacian operator to perform feature extraction on the input image to obtain a fifth edge feature, and using the second Laplacian operator to perform feature extraction on the input image to obtain a sixth edge feature;

[0020] Calculate an amplitude feature and a direction feature based on the first edge feature and the second edge feature;

[0021] The first edge feature, the second edge feature, the third edge feature, the fourth edge feature, the fifth edge feature, the sixth edge feature, the amplitude feature and the direction feature are fused to obtain edge information of the input image, where the input image includes the predicted image at the t-th time step, the low-resolution image and the auxiliary image.

[0022] Optionally, the encoding of the predicted image at the t-th time step by a first image encoder to obtain a first image feature, encoding the low-resolution image by a second image encoder to obtain a second image feature, and encoding the auxiliary image by a third image encoder to obtain a third image feature includes:

[0023] Injecting an input image into an image encoder corresponding to the input image for encoding, the image encoder includes an L-layer network, the output feature of the L-th layer network is the image feature corresponding to the input image, the input of the j+1-th layer network of the image encoder is the output feature of the j-th layer network of the image encoder and the output feature of the j-th layer network of the edge encoder encoding the edge information of the input image, L is a positive integer, and j is a positive integer less than L;

[0024] Among them, the input image includes the predicted image of the t-th time step, the low-resolution image and the auxiliary image, the image encoder corresponding to the predicted image of the t-th time step is the first image encoder, the image feature corresponding to the predicted image of the t-th time step is the first image feature, the image encoder corresponding to the low-resolution image is the second image encoder, the image feature corresponding to the low-resolution image is the second image feature, the image encoder corresponding to the auxiliary image is the third image encoder, and the image feature corresponding to the auxiliary image is the third image feature.

[0025] Optionally, injecting the input image into an image encoder corresponding to the input image for encoding includes:

[0026] exist In the case of, the input image is injected into the image encoder corresponding to the input image for encoding, A preset positive integer.

[0027] Optionally, the K-selected multi-head attention mechanism and the gating mechanism are used to perform feature fusion on the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature, including:

[0028] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first shared feature, the second shared feature, the third shared feature, and the edge shared feature to obtain a fused shared feature;

[0029] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first unique feature, the second unique feature, the third unique feature and the edge unique feature to obtain a fused unique feature;

[0030] The fused feature is determined based on the fused shared feature and the fused unique feature.

[0031] Optionally, the decoding process of the fused features by using a noise image decoder to obtain the predicted image at the t-1th time step includes:

[0032] The fused features are input into the noisy image decoder, and the outputs of each layer of the target encoder are averaged and spliced ​​to the corresponding layer of the noisy image decoder through jump connections to obtain the predicted image of the t-1th time step; wherein the target encoder includes the first image encoder, the second image encoder, the third image encoder and the edge encoder.

[0033] In a second aspect, an embodiment of the present invention provides a medical image super-resolution device based on an edge-enhanced diffusion model, comprising:

[0034] The forward processing module is used to gradually add Gaussian noise to the image to be processed in the forward stage of the diffusion model to obtain the perturbed image at the t-th time step, wherein the perturbed image at the 0th time step is the image to be processed, and the perturbed image at the T-th time step is a pure noise image. ;

[0035] A reverse processing module, used for performing step-by-step denoising on the pure noise image in the reverse phase of the diffusion model to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image;

[0036] The predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, including:

[0037] Acquire target edge information, wherein the target edge information includes edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image;

[0038] Using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature;

[0039] Obtaining, by a detangling operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature;

[0040] Using K-selected multi-head attention mechanism and gating mechanism to fuse the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature;

[0041] The fused features are decoded using a noise image decoder to obtain the predicted image at the t-1th time step.

[0042] In a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is used to read the program in the memory to implement the steps in the medical image super-resolution method based on the edge-enhanced diffusion model as described in the first aspect.

[0043] In a fourth aspect, an embodiment of the present invention provides a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the medical image super-resolution method based on the edge-enhanced diffusion model as described in the first aspect.

[0044] The beneficial effect of the present invention is that, in the embodiment of the present invention, the predicted image, low-resolution image and auxiliary image at the t-th time step are firstly extracted to obtain the target edge information, and then the image and edge are respectively encoded to generate enhanced disentangled unique features and shared features; adaptively integrate effective auxiliary information of different modalities, give full play to the complementarity of multimodal data to improve the feature expression ability; use the decoder to generate the noise image in the diffusion model, and reconstruct the high-resolution target image through the step-by-step denoising reverse diffusion process, so as to restore more detail information. In the embodiment of the present invention, the importance of edge information is taken into account, and the high-frequency edge information is injected into the corresponding image encoder to help capture details more effectively and obtain better super-resolution image reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Attached Figure 1 A flowchart of a medical image super-resolution method based on an edge-enhanced diffusion model provided by an embodiment of the present invention;

[0046] Attached Figure 2 A flowchart of performing denoising on a predicted image at the t-th time step to obtain a predicted image at the t-1th time step provided by an embodiment of the present invention;

[0047] Attached Figure 3a A schematic diagram of a diffusion model framework based on edge enhancement for multi-contrast MRI super-resolution provided by an embodiment of the present invention;

[0048] Attached Figure 3b For attachment Figure 3a One of the partial enlarged schematic diagrams;

[0049] Attached Figure 3c For attachment Figure 3a The second partial enlarged schematic diagram;

[0050] Attached Figure 4A schematic diagram of an edge information extractor provided by an embodiment of the present invention;

[0051] Attached Figure 5a A block diagram of a K-selective attention and fine correction module provided in an embodiment of the present invention;

[0052] Attached Figure 5b For attachment Figure 5a One of the partial enlarged schematic diagrams;

[0053] Attached Figure 5c For attachment Figure 5a The second partial enlarged schematic diagram;

[0054] Attached Figure 6 A schematic diagram of a medical image super-resolution device based on an edge-enhanced diffusion model provided by an embodiment of the present invention;

[0055] Attached Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In the embodiments of the present application, the term "and / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. In the embodiments of the present application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first" and "second" are generally of one type, and the number of objects is not limited. For example, the first object can be one or more.

[0057] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0059] The present application embodiment provides a medical image super-resolution method and device based on edge enhancement diffusion model. Figure 1 , Figure 1 The present invention provides a schematic flow chart of a medical image super-resolution method based on an edge-enhanced diffusion model. The method specifically includes the following steps:

[0060] Step 101, in the forward phase of the diffusion model, gradually add Gaussian noise to the image to be processed to obtain a perturbed image at the t-th time step, wherein the perturbed image at the 0th time step is the image to be processed, and the perturbed image at the T-th time step is a pure noise image. .

[0061] Step 102, performing step-by-step denoising processing on the pure noise image in the reverse phase of the diffusion model to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image.

[0062] It should be understood that the diffusion model includes two core processes, namely the forward process and the reverse process. In the forward process, noise is gradually added to the data so that it eventually approaches a known distribution (usually a Gaussian distribution). This process is a Markov chain, that is, each step depends only on the state of the previous step. The reverse process starts with pure noise and gradually removes noise through the reverse Markov chain to generate samples consistent with the original data distribution.

[0063] In this embodiment, the image to be processed is processed based on the diffusion model. Figure 2 As shown, in order to effectively retain the high-frequency information in the image and make full use of the effective information between different MRI modalities, in the reverse process, the predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, which specifically includes the following steps:

[0064] Step 201, performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information, wherein the target edge information includes edge information of the predicted image, the low-resolution image and the auxiliary image at the t-th time step.

[0065] Step 202, using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature;

[0066] Step 203, obtaining, by a de-entanglement operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature;

[0067] Step 204, using a K-select multi-head attention mechanism and a gating mechanism to perform feature fusion on the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature, and the edge unique feature to obtain a fused feature;

[0068] Step 205: Use a noise image decoder to decode the fused features to obtain the predicted image at the t-1th time step.

[0069] This embodiment also provides a diffusion model based on edge enhancement for multi-contrast MRI super-resolution, the main structure of which can be seen in Figure 3a-3c , the following will combine Figure 3a-3c The specific process of performing denoising on the predicted image at the t-th time step to obtain the predicted image at the t-1th time step in the embodiment of the present invention is described.

[0070] In step 201, edge information of the predicted image at the tth time step, edge information of the low-resolution image, and edge information of the auxiliary image are first obtained. The edge information of the predicted image at the tth time step is the target high-resolution magnetic resonance imaging (High-Resolution MR, HR MR) image perturbed by noise, the low-resolution image represents the corresponding (Low-Resolution, LR) image, and the auxiliary image is the HR image used to assist MRI contrast. In the MRI super-resolution method based on the diffusion model, the low-resolution (LR) image can be used as a condition for reconstructing the high-resolution (HR) image, and the auxiliary image provides complementary information.

[0071] Optionally, in some embodiments, step 201 includes:

[0072] Using an edge information extractor to extract edge information from the predicted image at the t-th time step, the low-resolution image, and the auxiliary image, respectively, to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image;

[0073] The edge information extractor includes a first Sobel operator, a second Sobel operator, a third Sobel operator, a fourth Sobel operator, a first Laplacian operator and a second Laplacian operator.

[0074] In this embodiment, the edge information extractor includes four trainable Sobel operators (the first Sobel operator, the second Sobel operator, the third Sobel operator and the fourth Sobel operator) and two Laplacian operators (the first Laplacian operator and the second Laplacian operator). The Sobel operator is a commonly used first-order differential operator, which is simple, fast and moderately resistant to noise; while the second-order differential Laplacian operator is more sensitive to noise, but can detect edges more clearly and sharply, and has isotropic characteristics, which can evenly enhance edges and lines in all directions.

[0075] like Figure 4 As shown, the four trainable Sobel operators of the edge information extractor are used to extract the vertical direction (such as Figure 4 (a)), horizontal direction (as shown in Figure 4 (b)), diagonal direction (as shown in Figure 4 (c)) and anti-diagonal directions (as shown in Figure 4 (d)) and two Laplacian operators (as shown in Figure 4 Edge extraction is performed using (e) and (f) as shown in the figure.

[0076] Optionally, in some embodiments, the extracting edge information of the predicted image at the t-th time step, the low-resolution image, and the auxiliary image respectively using an edge information extractor to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image includes:

[0077] Using the first Sobel operator to perform feature extraction on the input image along the vertical direction to obtain a first edge feature, using the second Sobel operator to perform feature extraction on the input image along the horizontal direction to obtain a second edge feature, using the third Sobel operator to perform feature extraction on the input image along the diagonal direction to obtain a third edge feature, using the fourth Sobel operator to perform feature extraction on the input image along the anti-diagonal direction to obtain a fourth edge feature, using the first Laplacian operator to perform feature extraction on the input image to obtain a fifth edge feature, and using the second Laplacian operator to perform feature extraction on the input image to obtain a sixth edge feature;

[0078] Calculate an amplitude feature and a direction feature based on the first edge feature and the second edge feature;

[0079] The first edge feature, the second edge feature, the third edge feature, the fourth edge feature, the fifth edge feature, the sixth edge feature, the amplitude feature and the direction feature are fused to obtain edge information of the input image, where the input image includes the predicted image at the t-th time step, the low-resolution image and the auxiliary image.

[0080] In the embodiment of the present application, 4 Sobel operators and 2 Laplacian operators work in parallel to extract edge features in different directions. At the same time, the features extracted by the vertical Sobel operator (first Sobel operator) and the horizontal Sobel operator (second Sobel operator) are combined to calculate the amplitude feature and direction feature. Finally, these 8 features are spliced ​​together and edge feature fusion is performed through a convolution layer to obtain the target edge information.

[0081] like Figure 3b As shown in FIG. 1 , in a specific implementation, for the predicted image, low-resolution image, and auxiliary image at the t-th time step, edge information extractors need to be used to extract edge information respectively. Specifically, the edge information extractor is used to extract edge information from the predicted image at the t-th time step. Extract edge information and get Edge information ; Use edge information extractor to extract low-resolution images Extract edge information and get Edge information ; Use edge information extractor to extract auxiliary image Extract edge information and get Edge information .

[0082] In this embodiment, by combining the multi-directional Sobel operator and the Laplacian operator, a more comprehensive and accurate edge information extraction is achieved. This method overcomes the directional limitations and detail loss problems of the traditional method, and improves adaptability through trainable parameters, thereby further optimizing the edge detection effect.

[0083] In step 202, four encoders with the same architecture (a first image encoder, a second image encoder, a third image encoder, and an edge encoder) are used to encode the input image and its edge information respectively, so as to obtain representations of different information respectively.

[0084] Optionally, in some embodiments, encoding the predicted image at the t-th time step by using a first image encoder to obtain a first image feature, encoding the low-resolution image by using a second image encoder to obtain a second image feature, and encoding the auxiliary image by using a third image encoder to obtain a third image feature, comprises:

[0085] Injecting an input image into an image encoder corresponding to the input image for encoding, the image encoder includes an L-layer network, the output feature of the L-th layer network is the image feature corresponding to the input image, the input of the j+1-th layer network of the image encoder is the output feature of the j-th layer network of the image encoder and the output feature of the j-th layer network of the edge encoder encoding the edge information of the input image, L is a positive integer, and j is a positive integer less than L;

[0086] Among them, the input image includes the predicted image of the t-th time step, the low-resolution image and the auxiliary image, the image encoder corresponding to the predicted image of the t-th time step is the first image encoder, the image feature corresponding to the predicted image of the t-th time step is the first image feature, the image encoder corresponding to the low-resolution image is the second image encoder, the image feature corresponding to the low-resolution image is the second image feature, the image encoder corresponding to the auxiliary image is the third image encoder, and the image feature corresponding to the auxiliary image is the third image feature.

[0087] like Figure 3b As shown, in this embodiment, in order to enhance the feature representation capability, high-frequency edge information is injected into the corresponding image encoder to help capture details more effectively and further improve the encoding effect.

[0088] For the predicted image at time step t , low-resolution images and auxiliary images The following are similar operations:

[0089] The predicted image at the t-th time step is injected into the first image encoder for encoding, the first image encoder includes an L-layer network, wherein the input of the j+1-th layer network of the first image encoder is the output feature of the j-th layer network of the first image encoder and the output feature of the j-th layer network of the edge encoder for encoding edge information of the predicted image at the t-th time step.

[0090] is the first image encoder, is an edge encoder, and the edge information of all input images shares the same edge encoder. Indicates the encoder The j-th layer network, Indicates the encoder The j-th layer network output features, Represents edge encoder Yes Encoded j-th layer network output features.

[0091] The low-resolution image is injected into the second image encoder for encoding, the second image encoder includes an L-layer network, wherein the input of the j+1th layer network of the second image encoder is the output feature of the jth layer network of the second image encoder and the output feature of the jth layer network of the edge encoder for encoding edge information of the low-resolution image.

[0092] is the second image encoder, is an edge encoder, and the edge information of all input images shares the same edge encoder. Indicates the encoder The j-th layer network, Indicates the encoder The j-th layer network output features, Represents edge encoder Yes Encoded j-th layer network output features.

[0093] The auxiliary image is injected into the first image encoder for encoding, and the third image encoder includes an L-layer network, wherein the input of the j+1th layer network of the third image encoder is the output feature of the jth layer network of the third image encoder and the output feature of the jth layer network of the edge encoder for encoding the edge information of the auxiliary image.

[0094] is the third image encoder, is an edge encoder, and the edge information of all input images shares the same edge encoder. Indicates the encoder The j-th layer network, Indicates the encoder The j-th layer network output features, Represents edge encoder Yes Encoded j-th layer network output features.

[0095] like Figure 3b As shown by the dashed line, the above process can be specifically expressed as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] .

[0101] According to the above content, the diffusion model is divided into a forward process and a reverse process. In the forward process Can be based on Adding noise, their features are relatively clear in the short period of time, exceeding a certain threshold ( ) will be occupied by more noise information. In the training process of practical applications, it is necessary to train the noise network for all time steps. When the noise in the Yes, but This is consistent with the above characteristics of time-varying. According to the characteristics of diffusion distribution, when t is relatively small, the image to be processed The structure and edges of the image are clear; however, as t increases, the noise begins to increase and quickly becomes dominant, the edges of the image become blurred, and the details are gradually lost. , middle The edge information is no longer recognizable until the final stage when the image becomes pure noise. For more reasonable use The edge features of When The edge information of the image is continuously injected into its corresponding image encoder. When , stop the injection operation and freeze the corresponding operation network layer to reduce noise interference. Since the reverse diffusion process follows a similar change law as the diffusion process, the reverse diffusion process also follows Without injecting edge information, Optionally, in some embodiments, injecting the edge information of the input image into an image encoder corresponding to the input image for encoding comprises:

[0102] exist In the case of, the input image is injected into the image encoder corresponding to the input image for encoding, A preset positive integer.

[0103] Injecting noise into the image encoding process will interfere with the model and even cause distortion of the generated results. The distribution characteristics of The high-frequency information in the image can avoid the hindrance of noise to edge information extraction and improve the accuracy of edge information extraction.

[0104] like Figure 3c As shown, in step 203, after the encoder completes encoding, the shared features and unique features of each encoder are obtained through a de-entanglement operation, and the first shared features after de-entanglement are obtained respectively. , the first unique feature , the second shared feature , the second unique feature , the third shared feature , the third unique feature , edge sharing features and unique edge features .in, represents the output feature of the image encoder, S represents the shared feature identifier, I represents the unique feature identifier, express The convolution operation can be expressed as follows:

[0105] ;

[0106] ;

[0107] .

[0108] In the specific implementation, the loss of unique features and shared features after disentanglement is as follows:

[0109] ;

[0110] ;

[0111] in, Respectively .

[0112] Different reference condition information plays a key role in the generation of super-resolution images. In order to extract more effective knowledge from multimodal information, it is necessary to aggregate reference information of different modalities to improve the generation quality of high-resolution images. Figure 5a-5c As shown, specifically, in step 204, effective multimodal information is flexibly extracted from shared features and unique features of different data types through the K-selection attention mechanism and the fine correction module.

[0113] Optionally, in some embodiments, step 204 includes:

[0114] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first shared feature, the second shared feature, the third shared feature, and the edge shared feature to obtain a fused shared feature;

[0115] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first unique feature, the second unique feature, the third unique feature and the edge unique feature to obtain a fused unique feature;

[0116] The fused feature is determined based on the fused shared feature and the fused unique feature.

[0117] like Figure 5b As shown, specifically, first calculate the attention h, integrate the generated from the reference condition Corresponding to the effective information of the super-resolution image. and unique features (hereinafter referred to as ), for every pair ( ). K selection attention is calculated as follows: express The i-th query pair in The similarity degree of the j-th key in Represents the first K largest values. express The normalized weight values ​​of the top-K in the middle. Indicates the index corresponding to the key that obtains the corresponding weight value. is the attention result of the i-th query. The specific calculation formula is as follows:

[0118] ;

[0119] ;

[0120] ;

[0121] .

[0122] Through the above method, the K-selective attention mechanism is used to flexibly extract effective multimodal features from shared features and unique features, ensuring efficient aggregation of different reference information to improve the quality of super-resolution images. However, although the above operation effectively captures and integrates highly relevant reference information from multiple conditions, the sum of the first K weights is crucial to represent the overall confidence. By correcting this part, the deviation can be effectively corrected, thereby improving the accuracy and generation quality of the model.

[0123] In this embodiment, if Figure 5c As shown, h is corrected using the fine correction module. In order to correct h, it is necessary to calculate The sum of the weights C, Represents the i-th value in C. Attention h and the original input features After stacking, correction is performed. FFD represents the standard feed-forward layer. The specific calculation method is as follows:

[0124] ;

[0125] ;

[0126] For each pair ( ) can be calculated to get the corresponding results :Different conditions represent conditional information from different angles, which have different effects on the generated super-resolution results. Gating is used to give different reference information different weights . The disentangled information represented by is shown as follows:

[0127] ;

[0128] ;

[0129] ;

[0130] Finally, the features are obtained through a convolutional block attention module (CBAM). Deeply refine the spatial and channel dimensions to obtain the final disentangled information , as shown below:

[0131] ;

[0132] in, To fuse shared features, To combine unique features.

[0133] Different reference conditions provide information from multiple perspectives and play different roles in super-resolution image generation. Therefore, compared with simply averaging all reference information, it is more efficient to dynamically adjust the effect of each condition on the image quality. Therefore, a weighted gating mechanism is introduced in this embodiment to assign an adaptive weight to each reference condition according to its contribution, thereby more accurately controlling the influence of each condition on the generation process.

[0134] In step 205, the fusion feature is decoded using a noise image decoder to obtain the predicted image at the t-1 time step. Optionally, in some embodiments, step 205 includes:

[0135] The fused features are input into the noisy image decoder, and the outputs of each layer of the target encoder are averaged and spliced ​​to the corresponding layer of the noisy image decoder through jump connections to obtain the predicted image of the t-1th time step; wherein the target encoder includes the first image encoder, the second image encoder, the third image encoder and the edge encoder.

[0136] like Figure 3a Specifically, the fused features are input into the decoder, and at the same time, the average of all encoder features at each layer is spliced ​​to the corresponding layer of the decoder through a jump connection and combined with the fused features to generate the expected noise image. , the disentangled information obtained above , After stacking, through a The convolution operation is performed and input into the decoder D. At the same time, the output of each layer of each encoder is averaged and spliced ​​to the corresponding layer of the decoder through a jump connection, which works together with the decoder features to finally generate the expected noise image. , and according to the noise image Get predictions .

[0137] exist Figure 3a The main architecture of the entire network shown in the figure is a noise network A loss function for this noise is defined as follows, where is a very small real number:

[0138] ;

[0139] In the embodiment of the present application, the predicted image, low-resolution image and auxiliary image at the tth time step are firstly extracted to obtain the target edge information, and then the image and edge are respectively encoded to generate enhanced disentangled unique features and shared features; the effective auxiliary information of different modalities is adaptively integrated to give full play to the complementarity of multimodal data to improve the feature expression ability; the decoder is used to generate the noise image in the diffusion model, and the high-resolution target image is reconstructed through the step-by-step denoising reverse diffusion process to restore more detail information. In the embodiment of the present invention, the importance of edge information is taken into account, and the high-frequency edge information is injected into the corresponding image encoder to help capture details more effectively and obtain better super-resolution image reconstruction.

[0140] like Figure 6 As shown, the embodiment of the present invention further provides a medical image super-resolution device 600 based on an edge enhancement diffusion model, comprising:

[0141] The forward processing module 601 is used to gradually add Gaussian noise to the image to be processed in the forward stage of the diffusion model to obtain a perturbed image at the t-th time step, wherein the perturbed image at the 0th time step is the image to be processed, and the perturbed image at the T-th time step is a pure noise image. ;

[0142] A reverse processing module 602 is used to perform step-by-step denoising on the pure noise image in the reverse phase of the diffusion model to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image;

[0143] The predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, including:

[0144] Acquire target edge information, wherein the target edge information includes edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image;

[0145] Using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature;

[0146] Obtaining, by a detangling operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature;

[0147] Using K-selected multi-head attention mechanism and gating mechanism to fuse the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature;

[0148] The fused features are decoded using a noise image decoder to obtain the predicted image at the t-1th time step.

[0149] Optionally, performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information includes:

[0150] Using an edge information extractor to extract edge information from the predicted image at the t-th time step, the low-resolution image, and the auxiliary image, respectively, to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image;

[0151] The edge information extractor includes a first Sobel operator, a second Sobel operator, a third Sobel operator, a fourth Sobel operator, a first Laplacian operator and a second Laplacian operator.

[0152] Optionally, the extracting edge information of the predicted image at the t-th time step, the low-resolution image and the auxiliary image respectively using an edge information extractor to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image and edge information of the auxiliary image comprises:

[0153] Using the first Sobel operator to perform feature extraction on the input image along the vertical direction to obtain a first edge feature, using the second Sobel operator to perform feature extraction on the input image along the horizontal direction to obtain a second edge feature, using the third Sobel operator to perform feature extraction on the input image along the diagonal direction to obtain a third edge feature, using the fourth Sobel operator to perform feature extraction on the input image along the anti-diagonal direction to obtain a fourth edge feature, using the first Laplacian operator to perform feature extraction on the input image to obtain a fifth edge feature, and using the second Laplacian operator to perform feature extraction on the input image to obtain a sixth edge feature;

[0154] Calculate an amplitude feature and a direction feature based on the first edge feature and the second edge feature;

[0155] The first edge feature, the second edge feature, the third edge feature, the fourth edge feature, the fifth edge feature, the sixth edge feature, the amplitude feature and the direction feature are fused to obtain edge information of the input image, where the input image includes the predicted image at the t-th time step, the low-resolution image and the auxiliary image.

[0156] Optionally, the encoding of the predicted image at the t-th time step by a first image encoder to obtain a first image feature, encoding the low-resolution image by a second image encoder to obtain a second image feature, and encoding the auxiliary image by a third image encoder to obtain a third image feature includes:

[0157] Injecting an input image into an image encoder corresponding to the input image for encoding, the image encoder includes an L-layer network, the output feature of the L-th layer network is the image feature corresponding to the input image, the input of the j+1-th layer network of the image encoder is the output feature of the j-th layer network of the image encoder and the output feature of the j-th layer network of the edge encoder encoding the edge information of the input image, L is a positive integer, and j is a positive integer less than L;

[0158] Among them, the input image includes the predicted image of the t-th time step, the low-resolution image and the auxiliary image, the image encoder corresponding to the predicted image of the t-th time step is the first image encoder, the image feature corresponding to the predicted image of the t-th time step is the first image feature, the image encoder corresponding to the low-resolution image is the second image encoder, the image feature corresponding to the low-resolution image is the second image feature, the image encoder corresponding to the auxiliary image is the third image encoder, and the image feature corresponding to the auxiliary image is the third image feature.

[0159] Optionally, injecting the input image into an image encoder corresponding to the input image for encoding includes:

[0160] exist In the case of, the input image is injected into the image encoder corresponding to the input image for encoding, A preset positive integer.

[0161] Optionally, the K-selected multi-head attention mechanism and the gating mechanism are used to perform feature fusion on the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature, including:

[0162] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first shared feature, the second shared feature, the third shared feature, and the edge shared feature to obtain a fused shared feature;

[0163] Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first unique feature, the second unique feature, the third unique feature and the edge unique feature to obtain a fused unique feature;

[0164] The fused feature is determined based on the fused shared feature and the fused unique feature.

[0165] Optionally, the decoding process of the fused features by using a noise image decoder to obtain the predicted image at the t-1th time step includes:

[0166] The fused features are input into the noisy image decoder, and the outputs of each layer of the target encoder are averaged and spliced ​​to the corresponding layer of the noisy image decoder through jump connections to obtain the predicted image of the t-1th time step; wherein the target encoder includes the first image encoder, the second image encoder, the third image encoder and the edge encoder.

[0167] The medical image super-resolution device 600 based on the edge-enhanced diffusion model provided in the embodiment of the present application can execute the above method embodiment, and its implementation principle and technical effects are similar, which will not be repeated in this embodiment.

[0168] It should be noted that the division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0169] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0170] like Figure 7 As shown, an embodiment of the present application provides an electronic device 700, including: a memory 702, a processor 701, and a program stored in the memory 702 and executable on the processor 701; the processor 701 is used to read the program in the memory 702 to implement the steps in the medical image super-resolution method based on the edge enhancement diffusion model as described above.

[0171] An embodiment of the present application also provides a readable storage medium, on which a program is stored. When the program is executed by a processor, the various processes of the above-mentioned medical image super-resolution method embodiment based on the edge enhancement diffusion model are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here. Among them, the readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical storage (such as compact disk (CD), digital video disk (DVD), Blu-ray Disc (BD), high-definition versatile disc (HVD), etc.), and semiconductor memory (such as read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read only memory (EEPROM), non-volatile memory (NAND FLASH), solid-state drive (Solid State Disk or Solid State Drive, SSD)), etc.

[0172] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0173] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), and includes a number of instructions for a terminal (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) to execute the methods described in each embodiment of the present application.

[0174] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A medical image super-resolution method based on edge-enhanced diffusion model, characterized in that: include: In the forward stage of the diffusion model, Gaussian noise is gradually added to the image to be processed to obtain the perturbation image at the tth time step, wherein the perturbation image at the 0th time step is the image to be processed, and the perturbation image at the Tth time step is a pure noise image. ; In the reverse phase of the diffusion model, the pure noise image is subjected to a stepwise denoising process to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image; The predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, including: Performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information, wherein the target edge information includes edge information of the predicted image, the low-resolution image and the auxiliary image at the t-th time step, and the auxiliary image is a high-resolution image used to assist super-resolution comparison; Using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature; Obtaining, by a detangling operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature; Using K-selected multi-head attention mechanism and gating mechanism to fuse the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature; The fused features are decoded using a noise image decoder to obtain the predicted image at the t-1th time step.

2. The method according to claim 1, characterized in that: The performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information includes: Using an edge information extractor to extract edge information from the predicted image at the t-th time step, the low-resolution image, and the auxiliary image, respectively, to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image; The edge information extractor includes a first Sobel operator, a second Sobel operator, a third Sobel operator, a fourth Sobel operator, a first Laplacian operator and a second Laplacian operator.

3. The method according to claim 2, characterized in that: The step of extracting edge information from the predicted image at the t-th time step, the low-resolution image, and the auxiliary image respectively using an edge information extractor to obtain edge information of the predicted image at the t-th time step, edge information of the low-resolution image, and edge information of the auxiliary image comprises: Using the first Sobel operator to perform feature extraction on the input image along the vertical direction to obtain a first edge feature, using the second Sobel operator to perform feature extraction on the input image along the horizontal direction to obtain a second edge feature, using the third Sobel operator to perform feature extraction on the input image along the diagonal direction to obtain a third edge feature, using the fourth Sobel operator to perform feature extraction on the input image along the anti-diagonal direction to obtain a fourth edge feature, using the first Laplacian operator to perform feature extraction on the input image to obtain a fifth edge feature, and using the second Laplacian operator to perform feature extraction on the input image to obtain a sixth edge feature; Calculate an amplitude feature and a direction feature based on the first edge feature and the second edge feature; The first edge feature, the second edge feature, the third edge feature, the fourth edge feature, the fifth edge feature, the sixth edge feature, the amplitude feature and the direction feature are fused to obtain edge information of the input image, where the input image includes the predicted image at the t-th time step, the low-resolution image and the auxiliary image.

4. The method according to claim 1, characterized in that: The method of encoding the predicted image at the t-th time step by using the first image encoder to obtain the first image feature, encoding the low-resolution image by using the second image encoder to obtain the second image feature, and encoding the auxiliary image by using the third image encoder to obtain the third image feature comprises: Injecting an input image into an image encoder corresponding to the input image for encoding, the image encoder includes an L-layer network, the output feature of the L-th layer network is the image feature corresponding to the input image, the input of the j+1-th layer network of the image encoder is the output feature of the j-th layer network of the image encoder and the output feature of the j-th layer network of the edge encoder encoding the edge information of the input image, L is a positive integer, and j is a positive integer less than L; Among them, the input image includes the predicted image of the t-th time step, the low-resolution image and the auxiliary image, the image encoder corresponding to the predicted image of the t-th time step is the first image encoder, the image feature corresponding to the predicted image of the t-th time step is the first image feature, the image encoder corresponding to the low-resolution image is the second image encoder, the image feature corresponding to the low-resolution image is the second image feature, the image encoder corresponding to the auxiliary image is the third image encoder, and the image feature corresponding to the auxiliary image is the third image feature.

5. The method according to claim 4, characterized in that: The step of injecting the input image into an image encoder corresponding to the input image for encoding comprises: exist In the case of, the input image is injected into the image encoder corresponding to the input image for encoding, A preset positive integer.

6. The method according to claim 1, characterized in that: The K-selected multi-head attention mechanism and the gating mechanism are used to perform feature fusion on the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature, including: Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first shared feature, the second shared feature, the third shared feature, and the edge shared feature to obtain a fused shared feature; Using a K-selection multi-head attention mechanism and a gating mechanism to perform feature fusion on the first unique feature, the second unique feature, the third unique feature and the edge unique feature to obtain a fused unique feature; The fused feature is determined based on the fused shared feature and the fused unique feature.

7. The method according to claim 1, characterized in that: The step of decoding the fused features using a noise image decoder to obtain the predicted image at the t-1th time step includes: The fused features are input into the noisy image decoder, and the outputs of each layer of the target encoder are averaged and spliced ​​to the corresponding layer of the noisy image decoder through jump connections to obtain the predicted image of the t-1th time step; wherein the target encoder includes the first image encoder, the second image encoder, the third image encoder and the edge encoder.

8. A medical image super-resolution device based on an edge-enhanced diffusion model, characterized in that: include: The forward processing module is used to gradually add Gaussian noise to the image to be processed in the forward stage of the diffusion model to obtain the perturbed image at the t-th time step, wherein the perturbed image at the 0th time step is the image to be processed, and the perturbed image at the T-th time step is a pure noise image. ; A reverse processing module, used for performing step-by-step denoising on the pure noise image in the reverse phase of the diffusion model to obtain a predicted image at the t-th time step, wherein the predicted image at the 0-th time step is a super-resolution image; The predicted image at the t-th time step is denoised to obtain the predicted image at the t-1th time step, including: Performing image edge extraction on the predicted image, the low-resolution image and the auxiliary image at the t-th time step to obtain target edge information, wherein the target edge information includes edge information of the predicted image, the low-resolution image and the auxiliary image at the t-th time step, and the auxiliary image is a high-resolution image used to assist super-resolution comparison; Using a first image encoder to encode the predicted image at the t-th time step to obtain a first image feature, using a second image encoder to encode the low-resolution image to obtain a second image feature, using a third image encoder to encode the auxiliary image to obtain a third image feature, and using an edge encoder to encode the target edge information to obtain an edge feature; Obtaining, by a detangling operation, a first shared feature and a first unique feature corresponding to the first image feature, a second shared feature and a second unique feature corresponding to the second image feature, a third shared feature and a third unique feature corresponding to the third image feature, and an edge shared feature and an edge unique feature corresponding to the edge feature; Using K-selected multi-head attention mechanism and gating mechanism to fuse the first shared feature, the first unique feature, the second shared feature, the second unique feature, the third shared feature, the third unique feature, the edge shared feature and the edge unique feature to obtain a fused feature; The fused features are decoded using a noise image decoder to obtain the predicted image at the t-1th time step.

9. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; wherein the processor is used to read the program in the memory to implement the steps in the medical image super-resolution method based on the edge-enhanced diffusion model as described in any one of claims 1 to 7.

10. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps in the medical image super-resolution method based on the edge-enhanced diffusion model as described in any one of claims 1 to 7 are implemented.

Images