A medical image denoising and reconstruction method based on unsupervised learning

Abstract

The present invention discloses a medical image denoising and reconstruction method based on unsupervised learning, which relates to the technical field of medical image processing and solves the technical problem that the existing diffusion model has unsatisfactory effect on detail recovery and noise reconstruction of medical images. The method comprises: constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image; inputting the first frequency domain subset and the second frequency domain subset into a trained generative parallel network to generate a preliminary reconstructed image, wherein the generative parallel network comprises at least two mutually independent first generative networks and second generative networks; mapping the features of the preliminary reconstructed image to a latent space, and performing diffusion reconstruction in the denoising process, encoding to obtain diffusion reconstruction features; decoding the diffusion reconstruction features by a decoder to obtain the final reconstructed image of the medical image. The present invention can achieve high-frequency detail recovery of medical images and reconstruct and generate high-quality medical images.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a medical image denoising and reconstruction method based on unsupervised learning. Background Art

[0002] Medical imaging technologies, such as computed tomography (CT) and magnetic resonance imaging (MRI), play an important role in modern medical diagnosis. CT and MRI each have unique advantages: CT imaging is fast and suitable for scanning high-contrast tissues such as bones and lungs; MRI can generate a variety of contrast-enhanced soft tissue images, does not involve radiation exposure, and is suitable for examinations of the nervous system and cardiovascular system. However, these imaging technologies still face challenges in noise reduction and image reconstruction.

[0003] In order to reduce the radiation dose, it is usually necessary to reduce the X-ray dose or the number of projections collected in CT scans, which will lead to large noise and artifacts in the reconstructed image, affecting the image quality. Similarly, in order to speed up MRI scans, the k-space is often undersampled to shorten the imaging time, but it will lead to image quality degradation and artifacts. Existing image reconstruction methods such as compressed sensing (CS) and super-resolution methods based on deep learning can improve image quality to a certain extent, but they are still insufficient in high-frequency detail recovery and low-noise image generation, and the computational complexity is high.

[0004] In recent years, diffusion models have performed well in the field of image generation. Diffusion models can generate high-quality images and maintain image detail structures by gradually adding noise to the image and learning the denoising process. However, traditional diffusion models are usually executed in high-dimensional space, resulting in high computational overhead and many iterations, making it difficult to meet real-time or near-real-time requirements. In addition, when processing low-dose or under-sampled data from CT and MRI, the existing diffusion models are still not ideal for restoring high-frequency details and reconstructing noise.

[0005] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:

[0006] Existing diffusion models do not perform well in detail restoration and noise reconstruction of medical images. Summary of the invention

[0007] The purpose of the present invention is to provide a medical image denoising and reconstruction method based on unsupervised learning to solve the technical problem that the diffusion model in the prior art has unsatisfactory effects on detail restoration and noise reconstruction of medical images. The preferred technical solutions among the many technical solutions provided by the present invention can produce many technical effects as described below.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The present invention provides a medical image denoising and reconstruction method based on unsupervised learning, comprising the following steps: constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image; inputting the first frequency domain subset and the second frequency domain subset into a trained generative parallel network to generate a preliminary reconstructed image, wherein the generative parallel network includes at least two mutually independent first generative networks and second generative networks; mapping the features of the preliminary reconstructed image to a latent space, and performing diffusion reconstruction in a denoising process, encoding to obtain diffusion reconstruction features; decoding the diffusion reconstruction features through a decoder to obtain a final reconstructed image of the medical image.

[0010] Optionally, in the step of inputting the first frequency domain subset and the second frequency domain subset into the trained generative parallel network to generate a preliminary reconstructed image, training the generative parallel network includes: inputting a first change image obtained by inverse Fourier transforming the first frequency domain subset into the first generative network to obtain a first reconstructed image estimate of the first change image; inputting a second change image obtained by inverse Fourier transforming the second frequency domain subset into the second generative network to obtain a second reconstructed image estimate of the second change image; respectively calculating a first reconstruction loss between the first reconstructed image estimate and the medical image, and a second reconstruction loss between the second reconstructed image estimate and the medical image; calculating a consistency loss between the first generative network and the second generative network; combining the first reconstruction loss, the second reconstruction loss and the consistency loss to form a total loss function; and using the total loss function to train the first generative network and the second generative network to obtain the generative parallel network.

[0011] Optionally, the final optimization goal of using the total loss function to train the first generation network and the second generation network is:

[0012]

[0013] in, is the first reconstructed image estimate; for the second reconstructed image estimation; is the encoding matrix of the first changed image; is the first changed image; is the second changed image; is the encoding matrix of the second changed image; is a regularization term of the first generation network; is a regularization term of the second generation network; are the first regularization parameter, the second regularization parameter, and the third regularization parameter respectively; is the consistency loss of the first generation network and the second generation network.

[0014] Optionally, the using the total loss function to train the first generation network and the second generation network includes: using the Adam optimizer to perform gradient descent optimization on the first generation network and the second generation network, and adjusting the learning rate during the training process until the total loss function converges or meets a preset stopping condition, thereby completing the training.

[0015] Optionally, mapping the features of the preliminary reconstructed image to a latent space, performing diffusion reconstruction in a denoising process, and encoding to obtain diffusion reconstruction features, includes: forward diffusing the features of the preliminary reconstructed image in the latent space through a stochastic differential equation, injecting noise during the forward diffusion process, and then performing reverse diffusion in the process of gradually removing the noise in the latent space, and obtaining the diffusion reconstruction features after encoding.

[0016] Optionally, in the process of performing reverse diffusion in the process of gradually removing noise in the latent space, adjusting the diffusion denoising process through the large window attention model guided by the diffusion reconstruction feature, including: calculating the feature information of the preliminary reconstructed image through the cross attention mechanism ; Input the diffusion reconstruction feature into the multi-head self-attention layer of the large window attention model, and Update and get updated feature information ; Update the feature information Divide into multiple non-overlapping square windows, and then perform self-attention calculation to obtain dynamic adjustment parameters ; The dynamic adjustment parameters The feed-forward network of the large window attention model is input to guide the reconstruction process of the preliminary reconstructed image.

[0017] Optionally, in the process of adjusting the diffusion denoising by the large window attention model guided by the diffusion reconstruction feature, the large window attention model is optimized by using an adjustment loss function that is a combination of the reconstruction loss of the medical image and the consistency loss of the original frequency domain data of the medical image; the adjustment loss function for:

[0018] ;

[0019] ;

[0020] in, are the first weight parameter and the second weight parameter respectively; is the reconstruction loss of the medical image; A final reconstructed image of the medical image; for the preliminary reconstruction image; For fidelity; The frequency domain data of the next period after the original frequency domain data; is the consistency loss of the original frequency domain data.

[0021] Optionally, in the process of performing back diffusion in the process of gradually removing noise in the latent space, the original frequency domain data of the medical image in k-space is undersampled, and the diffusion reconstruction features are made consistent with the medical image through Bayesian reasoning.

[0022] Optionally, constructing at least two different first frequency domain subsets and second frequency domain subsets based on the medical image includes: randomly masking and sampling the original frequency domain data of the medical image in k-space to generate at least two different first frequency domain subsets and second frequency domain subsets; wherein each of the first frequency domain subset and the second frequency domain subset includes low-frequency information and high-frequency information of the medical image training sample, and the amount of information of the low-frequency information is greater than the amount of information of the high-frequency information.

[0023] Optionally, before constructing at least two different first frequency domain subsets and second frequency domain subsets based on the medical image, the method also includes: converting the medical image into a two-dimensional slice image, normalizing each of the two-dimensional slice images, and then performing Fourier transform to obtain the original frequency domain data of the medical image in k-space.

[0024] Implementing one of the above technical solutions of the present invention has the following advantages or beneficial effects:

[0025] The medical image denoising and reconstruction method based on unsupervised learning of the present invention encodes the medical image to be reconstructed into a low-dimensional latent space through two parallel generative networks, thereby realizing compression of high-dimensional medical images and extraction of effective features; and then performing diffusion reconstruction in the latent space can generate higher quality medical images; the latent space learned by the generative network can capture the core structural information of the medical image, which helps the network model understand the image content, reduces the computational complexity, and improves the reconstruction accuracy of the medical image; in addition, both the latent space and diffusion reconstruction of the generative network can be performed under unsupervised conditions, without relying on manually labeled data, and image reconstruction can be completed through self-supervised learning, so that the method can be widely applied to unlabeled medical image data, reducing the data preparation cost and increasing the practicality and application scope of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. It is obvious that the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0027] Figure 1 is a flow chart of a medical image denoising and reconstruction method based on unsupervised learning according to an embodiment of the present invention;

[0028] Figure 2 It is a structural schematic diagram of a model of a medical image denoising and reconstruction method implemented in an embodiment of the present invention;

[0029] Figure 3 is a flowchart of step S2 of the medical image denoising and reconstruction method based on unsupervised learning in an embodiment of the present invention;

[0030] Figure 4 It is a schematic diagram of a structure of generating a parallel network according to an embodiment of the present invention;

[0031] Figure 5 is a schematic diagram of the structure of a diffusion model according to an embodiment of the present invention;

[0032] Figure 6 Flow chart of step S5 of the medical image denoising and reconstruction method based on unsupervised learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention clearer, the various exemplary embodiments to be described below will refer to the corresponding drawings, which constitute a part of the exemplary embodiments, wherein various exemplary embodiments that may be used to implement the present invention are described. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation methods described in the following exemplary embodiments do not represent all implementation methods consistent with the present disclosure. It should be understood that they are only examples of processes, methods, devices, etc. that are consistent with some aspects of the present disclosure as detailed in the attached claims, and other embodiments may also be used, or the embodiments listed herein may be modified in structure and function without departing from the scope and essence of the present invention.

[0034] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", etc. indicate the orientation or positional relationship based on the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the elements referred to must have a specific orientation, be constructed and operated in a specific orientation. The terms "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. The term "multiple" means two or more. The terms "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, an integral connection, a mechanical connection, an electrical connection, a communication connection, a direct connection, an indirect connection through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0035] In order to illustrate the technical solution of the present invention, a specific embodiment is used below for description, and only the parts related to the embodiment of the present invention are shown.

[0036] Embodiment 1:

[0037] like Figure 1 and Figure 2 As shown, the present invention provides a medical image denoising and reconstruction method based on unsupervised learning, comprising the following steps: S1, constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image; S2, inputting the first frequency domain subset and the second frequency domain subset into a trained generative parallel network to generate a preliminary reconstructed image , wherein the generated parallel network includes at least two mutually independent first generation networks and second generation networks; S3, the preliminary reconstructed image The features are mapped to the latent space, and diffusion reconstruction is performed in the denoising process, and the diffusion reconstruction features are encoded; S4, the diffusion reconstruction features are decoded by the decoder to obtain the final reconstructed image of the medical image .

[0038] The medical image denoising and reconstruction method based on unsupervised learning provided in this embodiment is applied to the field of medical imaging, especially suitable for medical image denoising and reconstruction. Through two parallel generative networks, the medical image to be reconstructed is encoded into a low-dimensional latent space to achieve compression of high-dimensional medical images and extraction of effective features; then diffusion reconstruction is performed in the latent space to generate higher quality medical images; the latent space learned by the generative network can capture the core structural information of the medical image, which helps the network model understand the image content, reduce the computational complexity, and improve the reconstruction accuracy of the medical image; in addition, the latent space and diffusion reconstruction of the generative network can be performed under unsupervised conditions, without relying on manually labeled data, and image reconstruction can be completed through self-supervised learning, so that the method can be widely applied to unlabeled medical image data, reducing the data preparation cost and increasing the practicality and application scope of the model.

[0039] Next, combine Figure 1-Figure 6 The specific implementation steps of the medical image denoising and reconstruction method based on unsupervised learning provided in this embodiment are introduced in detail:

[0040] First, execute step S1, and construct at least two different first frequency domain subsets and second frequency domain subsets according to the medical image; different first frequency domain subsets and second frequency domain subsets refer to different frequency information of the two subsets. The number of frequency domain subsets corresponds to the number of generated networks in the generated parallel network. If the training accuracy of the model is to be further improved, the generated networks can be set to more. Accordingly, the number of constructed frequency domain subsets should be consistent with the number of generated networks. For example, if 4 generated networks are set, 4 different frequency domain subsets need to be constructed accordingly.

[0041] Specifically, step S1 includes: randomly masking and sampling the original frequency domain data of the medical image in the k-space to generate at least two different first frequency domain subsets and second frequency domain subsets; wherein each of the first frequency domain subset and the second frequency domain subset includes low-frequency information and high-frequency information of the medical image training samples, and the amount of information of the low-frequency information is greater than that of the high-frequency information. It is best to make each subset contain not only most of the low-frequency information, but also a small amount of high-frequency information. The constructed subsets must follow the above rules to ensure the restoration details and accuracy of the model.

[0042] Taking the construction of two different frequency domain subsets as an example, after obtaining the medical image, the original frequency domain data of the medical image is Random mask sampling is performed in the frequency domain to generate two different subsets, which are respectively recorded as the first frequency domain subsets. and the second frequency domain subset The construction rules need to be followed: first, the amount of low-frequency information in each frequency domain subset constructed should be much greater than the amount of high-frequency information; second, the frequency domain information between the two subsets constructed should be different. Each subset contains most of the low-frequency information to ensure that the reconstruction network can learn enough image structure information; each subset contains a small amount of high-frequency information to support the restoration of details in the reconstructed image; the first frequency domain subset and the second frequency domain subset The frequency domain information between them is distinguished to ensure that the two generating networks can learn from different subsets of information, so that the model can capture more detailed image features.

[0043] It should be noted that before step S1, the medical image needs to be preprocessed, specifically including: converting the medical image into a two-dimensional slice image, normalizing each two-dimensional slice image, and then performing Fourier transform to obtain the original frequency domain data of the medical image in k-space, which can be recorded as The input medical image data is converted into a two-dimensional slice image format, and the image intensity value is normalized to be in the range of [0, 1]. By eliminating the differences in dimension and value range between the data, the model can learn more effectively and improve the accuracy and generalization ability of the model.

[0044] Then, step S2 is performed to input the first frequency domain subset and the second frequency domain subset into the trained generative parallel network to generate a preliminary reconstructed image. , wherein the generative parallel network includes at least two independent first generative networks VAE1 and second generative networks VAE2. Since the number of generative networks and the number of frequency domain subsets are corresponding, during input, the first frequency domain subset is input into the first generative network for training, and the second frequency domain subset is input into the second generative network for training. The generative network may adopt a generative adversarial network (GAN), a variational autoencoder (VAE) or a diffusion model, etc. In this embodiment, two independent variational autoencoders are preferably used as the generative network. VAE has a stronger generation capability and can generate high-quality images in the latent space, which is beneficial to improving the reconstruction effect of medical images.

[0045] Specifically, in step S2, an unsupervised training method is used to generate a parallel network, so that there is no need to rely on manually labeled data, and image reconstruction is completed through self-supervised learning, so that the method of this embodiment can be applied to medical image data without labels, reducing the cost of data preparation. Figure 3 and Figure 4 As shown, the unsupervised training method for generating a parallel network specifically includes: S21, performing an inverse Fourier transform on the first frequency domain subset to obtain a first change image , input the first generation network VAE1, and obtain the first reconstructed image estimation of the first change image; the first frequency domain subset After inverse Fourier transform, the first change image is obtained ,Will Input to the first generation network VAE1 to restore the initial reconstruction result in the image space. S22, perform inverse Fourier transform on the second frequency domain subset to obtain a second changed image , input into the second generation network VAE2, obtain the second reconstructed image estimation of the second change image; the second frequency domain subset After inverse Fourier transform, the second change image is obtained , accordingly, Input it into the second generation network VAE2 to recover the second preliminary reconstruction result. S23, respectively calculate the first reconstruction loss between the first reconstructed image estimate and the medical image, and the second reconstruction loss between the second reconstructed image estimate and the medical image, and denote the first reconstruction loss and the second reconstruction loss as ; The first reconstruction loss and the second reconstruction loss represent the first frequency domain subset , the second frequency domain subset The reconstructed image estimate (the first frequency domain subset , the second frequency domain subset The reconstructed image estimates are recorded as ) and the original medical image. S24, calculate the consistency loss of the first generation network VAE1 and the second generation network VAE2, and record the consistency loss as ; The consistency loss constrains the difference between the two network outputs so that the network outputs remain consistent on unsampled data points; this consistency constraint helps the model to more accurately infer the unscanned frequency domain information. S25, the first reconstruction loss ( ), the second reconstruction loss ( ) and consistency loss ( ) are combined to form a total loss function; S26, the total loss function is used to train the first generation network and the second generation network to obtain a generation parallel network. By generating a parallel network to perform preliminary reconstruction of the medical image, the medical image can be restored as a whole.

[0046] Further, in step S25, the first reconstruction loss ( ), the second reconstruction loss ( ) and consistency loss ( ) The total loss function formed by the combination for:

[0047]

[0048] Where N is the number of training samples; For the i The first change image of training samples; For the i A second variation image of a training sample; For the i The encoding matrix corresponding to the first change image of the training samples; For the i The encoding matrix corresponding to the second change image of the training samples; The consistency loss of the first generation network and the second generation network using the first change image and the second change image as calculation parameters; Other constraints for maintaining network stability; To control the weight parameters of different loss items, the total loss function can be used To optimize the network parameters, in order to improve the consistency and accuracy of the reconstruction results.

[0049] Furthermore, the total loss function is adopted The final optimization goal of training the first generation network and the second generation network is:

[0050]

[0051] in, is the first reconstructed image estimate; is estimated for the second reconstructed image; is the first changed image; is the second changed image; is the encoding matrix of the first change image; is the encoding matrix of the second change image; is the regularization term of the first generation network; is the regularization term of the second generation network; are the first regularization parameter, the second regularization parameter, and the third regularization parameter respectively; is the consistency loss between the first generated network and the second generated network.

[0052] Further, in step S26, the total loss function is used to train the first generation network and the second generation network, specifically including: using the Adam optimizer to perform gradient descent optimization on the first generation network and the second generation network, adjusting the learning rate during the training process until the total loss function converges or meets the preset stop condition, and completing the training. Adam (full name Adaptive Moment Estimation) optimizer is an optimization algorithm for training deep learning models. The algorithm has strong adaptive ability and can adaptively adjust the learning rate of each parameter, can converge quickly and reduce training time. In this embodiment, the learning rate is set to 0.0001, and gradient descent optimization is performed. The learning rate is adaptively adjusted during the training process. Whenever the performance on the verification set is stable, the learning rate is automatically reduced; the network is continuously trained until the total loss function converges or the preset stop condition is met. For example, the preset stop condition is set to stop training when the maximum training round is reached. Thus, a well-trained high-precision generation parallel network is obtained. This embodiment uses a variational autoencoder as a generative parallel network. Through the variational autoencoder, medical images can be encoded into a low-dimensional latent space to achieve compression and effective feature extraction of high-dimensional medical images. The latent representation learned by VAE can capture the core structural information of medical images, help the model understand the image content, reduce computational complexity, and improve reconstruction accuracy.

[0053] Next, execute step S3 to reconstruct the initial image The features of the image are mapped to the latent space, and diffusion reconstruction is performed in the denoising process, and the diffusion reconstruction features are obtained by encoding. This step is used to realize the diffusion reconstruction in the latent space. Specifically, step S3 includes: forward diffusion of the features of the preliminary reconstructed image in the latent space through a stochastic differential equation, and injecting noise in the forward diffusion process, and then reverse diffusion is performed in the process of gradually removing the noise in the latent space, and the diffusion reconstruction features are obtained after encoding. Figure 5 As shown in the figure, the forward diffusion and reverse diffusion of the preliminary reconstructed image can be realized by introducing the diffusion model. The diffusion model is introduced as part of the generative parallel network generation process, and a higher quality medical image is generated from the latent space through the reverse process of stepwise denoising. The diffusion model has a good effect in noise reduction and fine feature recovery, and can effectively remove noise in the reconstruction process to ensure the realism and accuracy of the final generated image. It is particularly suitable for detail reconstruction in medical images.

[0054] The method of this embodiment first generates the image to be reconstructed into a preliminary reconstructed image through a trained generative parallel network. , and use the generative parallel network to reconstruct the initial image Encode and map its features to a low-dimensional latent space; then forward diffusion is performed through the stochastic differential equation (SDE), and the form of forward diffusion is as follows:

[0055] ;

[0056] in, Reconstruct features for diffusion; Express Find the total differential; Add parameters to the noise to control the speed of noise addition during the diffusion process; is a standard Winer process (Wiener process, also known as Brownian motion, is an important continuous-time random process).

[0057] After the forward diffusion is completed, the features are then diffused backwards; specifically, the backward diffusion process is realized by gradually removing the noise in the latent space and restoring the high-frequency information of the latent features. The form of the backward diffusion is as follows:

[0058] ;

[0059] in, Indicates that under the given conditions y Next, the diffusion reconstruction feature The gradient of the log probability density of , is a gradient operator; in this embodiment, given the condition y That is, the input medical image has poor quality.

[0060] It should be noted that in the back diffusion process, the original frequency domain data of the undersampled medical image in k-space is combined to make the diffusion reconstruction features consistent with the medical image through Bayesian reasoning. Bayesian reasoning can ensure that the reconstructed potential features (i.e., diffusion reconstruction features) are consistent with the actual data.

[0061] As an optional implementation, in the process of gradually removing noise in the latent space, back diffusion is performed, and a denoising network is used to gradually remove noise in the latent space, and the diffusion denoising process is adjusted by a large window attention model guided by diffusion reconstruction features, wherein the large window attention model includes a multi-head self-attention layer and a forward feedback network. The diffusion reconstruction feature is a priori feature. Adjusting the denoising process through a large window attention layer guided by the prior feature can further improve the generation quality; the large window attention model expands the receptive field through a large window to capture long-distance dependencies in the image. The large window attention model guided by the prior feature consists of a multi-head self-attention layer (Feature-Guide Large WindowMulti-Head Self-Attention, referred to as FG-LMS) and a forward feedback network (Feature-Guide Feed-Forward Network, referred to as FG-FFN), wherein the multi-head self-attention layer is used to process input features and dynamically use the diffusion reconstruction features. To guide the calculation of self-attention; the forward feedback network processes spatial information and channel information through 1×1 convolution and 3×3 depth convolution to enhance the ability to restore details.

[0062] It should be noted that FG-LMS is an improved self-attention mechanism model that combines the ideas of large window attention and feature guidance, aiming to enhance the model's ability to capture global and local information, while using feature information to guide the allocation of attention. FG-FFN is an improved feedforward data network that enhances the feature extraction capability of the feedforward network by introducing a feature guidance mechanism, enabling it to dynamically adjust the network's behavior according to the characteristics of the input data.

[0063] Specifically, Figure 6 As shown, the diffusion denoising process is adjusted by the large window attention layer guided by the diffusion reconstruction feature, including: S51, calculating the feature information of the preliminary reconstructed image through the cross attention mechanism ; S52, input the diffusion reconstruction features into the multi-head self-attention layer of the large window attention model, and Update and get updated feature information ; S53, update feature information Divide into multiple non-overlapping square windows, and then perform self-attention calculation to obtain dynamic adjustment parameters ; S54, dynamically adjust the parameters The input is fed into a feed-forward network of a large window attention model to guide the reconstruction process of the preliminary reconstructed image.

[0064] Next, combine Figure 5 and Figure 6 Let's explain in detail the diffusion denoising process of adjusting the large window attention model guided by the diffusion reconstruction feature:

[0065] First, execute step S51 to calculate the feature information of the preliminary reconstructed image through the cross attention mechanism ; In this process, a preliminary reconstruction image generated by a parallel network will be generated Feature information is calculated through the cross-attention mechanism , which can effectively introduce the information in the reference image into the reconstruction of the target image, helping the model to better utilize the preliminary reconstructed image The high-frequency information in the image can be used to improve the reconstruction effect.

[0066] Then, step S52 is executed to input the diffusion reconstruction feature into the multi-head self-attention layer, and the feature information Update and get updated feature information ; Reconstruct the diffusion features Input to the multi-head self-attention layer, calculate dependencies, and update feature information The calculation formula is:

[0067] ;

[0068] in, express A linear function of Indicates that Normalization.

[0069] Next, execute step S53 to update the feature information Divide into multiple non-overlapping square windows, and then perform self-attention calculation to obtain dynamic adjustment parameters ; Updated features First, it is divided into N non-overlapping square windows , which is then mapped to query through a linear layer ,key ,value , perform self-attention calculations, and accumulate to obtain dynamic adjustment parameters . Dynamic adjustment parameters The calculation formula is:

[0070] Fp= ;

[0071] in, is the query vector, indicating the location of the current feature; is the key vector, representing the features of all other positions; express The transpose of is a value vector containing the values ​​associated with ; is the relative position offset, used to capture the position encoding; is a scaling factor, usually the size of the query key dimension, used to keep the value stable.

[0072] Finally, step S54 is executed to adjust the dynamic adjustment parameters Input the forward feedback network to guide the reconstruction process of the preliminary reconstructed image. In the forward feedback network, the dynamic adjustment parameters are first Linearization and Normalization: ;in, is the dynamic adjustment parameter after linearization and normalization. Then guide the reconstruction process: first use 1×1 convolution to aggregate information from different channels; then use 3×3 deep convolution to aggregate information from spatially adjacent pixels, and use the gating mechanism to enhance information encoding to obtain the final reconstruction features; the adjustment process is as follows:

[0073] ;

[0074] in, Express conduct Convolution, GELU (Gaussian Error Linear Unit) represents non-linear activation.

[0075] In order to generate more reliable prior knowledge, in the process of adjusting the diffusion denoising by the large window attention layer guided by the diffusion reconstruction feature, the large window attention layer is optimized by using the adjustment loss function of the combination of the reconstruction loss of the medical image and the consistency loss of the original frequency domain data of the medical image; the adjustment loss function for:

[0076] ;

[0077] ;

[0078] in, are the first weight parameter and the second weight parameter respectively; Reconstruction loss for medical images; The final reconstructed image of the medical image; To initially reconstruct the image; For fidelity; is the frequency domain data of one cycle after the original frequency domain data; is the consistency loss of the original frequency domain data. You can set and , to balance the contribution of the two losses.

[0079] A cross-attention mechanism is introduced in the denoising process to align the potential representation of the generative parallel network with the generation process mechanics of the diffusion model, so that the generated images more accurately reflect the structural features in real medical images.

[0080] Finally, step S4 is executed to decode the expanded reconstruction features through the decoder to obtain the final reconstructed image of the medical image. After the diffusion process is completed, the decoder network is used to decode the latent features back to the high-dimensional image space to generate the final reconstructed image . The method of this embodiment combines the generative model and the diffusion model to generate high-quality images in the latent space, and introduces a cross-attention mechanism to improve the reconstruction capability of high-frequency details. Through this method, noise can be effectively reduced, artifacts can be removed, and efficient and accurate image reconstruction can be achieved in low-dose CT and undersampled MRI imaging, thereby meeting the clinical needs of fast and high-quality imaging. In addition, this method is based on a self-supervised learning framework and can adapt to different types of medical images and imaging tasks, such as CT, MRI, etc. Because it does not rely on a specific imaging mode or structure, it has strong versatility and is suitable for a variety of medical image reconstruction and denoising tasks, and has broad application prospects.

[0081] The embodiment is only a specific example and does not represent only one implementation mode of the present invention.

[0082] The above description is only the preferred embodiment of the present invention. It is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. A medical image denoising and reconstruction method based on unsupervised learning, characterized in that: The following steps are involved: constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image; Inputting the first frequency domain subset and the second frequency domain subset into a trained generative parallel network to generate a preliminary reconstructed image, wherein the generative parallel network includes at least two mutually independent first generative networks and second generative networks; Mapping the features of the preliminary reconstructed image to a latent space, performing diffusion reconstruction in a denoising process, and encoding to obtain diffusion reconstruction features; Decoding the diffusion reconstruction feature by a decoder to obtain a final reconstructed image of the medical image; Mapping the features of the preliminary reconstructed image to a latent space, performing diffusion reconstruction in a denoising process, and encoding to obtain diffusion reconstruction features include: By using a stochastic differential equation, the features of the preliminary reconstructed image are forward diffused in the latent space, noise is injected during the forward diffusion process, and then reversely diffused in the process of gradually removing the noise in the latent space, and the diffused reconstruction features are obtained after encoding; In the reverse diffusion in the process of gradually removing noise in the latent space, the diffusion denoising process is adjusted by the large window attention model guided by the diffusion reconstruction feature, including: The preliminary reconstructed image is calculated through the cross attention mechanism to obtain feature information ; The diffusion reconstruction feature is input into the multi-head self-attention layer of the large window attention model, and the feature information Update and get updated feature information ; The update feature information Divide into multiple non-overlapping square windows, and then perform self-attention calculation to obtain dynamic adjustment parameters ; The dynamic adjustment parameter Inputting the feedforward network of the large window attention model to guide the reconstruction process of the preliminary reconstructed image; In the process of adjusting the diffusion denoising by the large window attention model guided by the diffusion reconstruction feature, the large window attention model is optimized by using an adjustment loss function that is a combination of the reconstruction loss of the medical image and the consistency loss of the original frequency domain data of the medical image; the adjustment loss function for: ； ； in, are the first weight parameter and the second weight parameter respectively; is the reconstruction loss of the medical image; A final reconstructed image of the medical image; for the preliminary reconstruction image; For fidelity; The frequency domain data of the next period after the original frequency domain data; is the consistency loss of the original frequency domain data.

2. The medical image denoising and reconstruction method based on unsupervised learning according to claim 1, characterized in that: In the step of inputting the first frequency domain subset and the second frequency domain subset into a trained generative parallel network to generate a preliminary reconstructed image, training the generative parallel network includes: Inputting a first change image obtained by performing an inverse Fourier transform on the first frequency domain subset into the first generation network to obtain a first reconstructed image estimate of the first change image; Inputting a second change image obtained by inverse Fourier transforming the second frequency domain subset into the second generation network to obtain a second reconstructed image estimate of the second change image; respectively calculating a first reconstruction loss between the first reconstructed image estimate and the medical image, and a second reconstruction loss between the second reconstructed image estimate and the medical image; Calculating the consistency loss of the first generation network and the second generation network; Combining the first reconstruction loss, the second reconstruction loss, and the consistency loss to form a total loss function; The first generation network and the second generation network are trained using the total loss function to obtain the generation parallel network.

3. The medical image denoising and reconstruction method based on unsupervised learning according to claim 2, characterized in that: The final optimization goal of training the first generation network and the second generation network using the total loss function is: in, is the first reconstructed image estimate; for the second reconstructed image estimation; is the first changed image; is the second changed image; is the encoding matrix of the first changed image; is the encoding matrix of the second changed image; is a regularization term of the first generation network; is a regularization term of the second generation network; are the first regularization parameter, the second regularization parameter, and the third regularization parameter respectively; is the consistency loss of the first generation network and the second generation network.

4. The medical image denoising and reconstruction method based on unsupervised learning according to claim 3, characterized in that: The adopting the total loss function to train the first generation network and the second generation network includes: The Adam optimizer is used to perform gradient descent optimization on the first generation network and the second generation network, and the learning rate is adjusted during the training process until the total loss function converges or meets the preset stopping condition, and the training is completed.

5. The medical image denoising and reconstruction method based on unsupervised learning according to claim 1, characterized in that: In the reverse diffusion during the process of gradually removing noise in the latent space, the diffusion reconstruction features are made consistent with the medical image through Bayesian reasoning in combination with the original frequency domain data of the medical image in the k-space that is undersampled.

6. The medical image denoising and reconstruction method based on unsupervised learning according to claim 1, characterized in that: The step of constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image comprises: Randomly performing mask sampling from original frequency domain data of the medical image in k-space to generate at least two different first frequency domain subsets and second frequency domain subsets; Each of the first frequency domain subset and the second frequency domain subset includes low-frequency information and high-frequency information of the medical image training sample, and the information amount of the low-frequency information is greater than the information amount of the high-frequency information.

7. The medical image denoising and reconstruction method based on unsupervised learning according to claim 6, characterized in that: Before constructing at least two different first frequency domain subsets and second frequency domain subsets according to the medical image, the method further includes: The medical image is converted into a two-dimensional slice image, each of the two-dimensional slice images is normalized, and then Fourier transformed to obtain the original frequency domain data of the medical image in k-space.

Images