Low-dose SPECT image noise reduction method and system based on unsupervised strategy and generalized diffusion model

Abstract

The invention relates to a low-dose SPECT image noise reduction method and system based on an unsupervised strategy and a generalized diffusion model. The method comprises the following steps: carrying out slicing processing on a low-dose SPECT image to obtain a training image; in a forward propagation stage of the unsupervised diffusion training model, determining an attenuation coefficient degradation operator according to the body mass index of the patient, and carrying out image noise addition to obtain a random Poisson noise image; in a back propagation stage of the unsupervised diffusion training model, extracting optimization factors and time step length features, and performing denoising processing on the random Poisson noise image to obtain a predicted low-dose SPECT image; and loading the optimal weight into a back diffusion inference model, carrying out back diffusion by taking the predicted low-dose SPECT image as a diffusion end point, and carrying out noise reduction processing on the predicted low-dose SPECT image to obtain a predicted full-dose SPECT image. The data preparation cost can be reduced, artifacts are reduced, and the denoising effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and particularly to a low-dose SPECT image denoising method and system based on an unsupervised strategy and a generalized diffusion model. Background Art

[0002] SPECT is a mature medical imaging technology, widely used in the early diagnosis of coronary artery disease (CAD) and the risk assessment of myocardial infarction. Although it has diagnostic functions, the associated radiation exposure in SPECT imaging poses potential health risks to patients, so strategies for reducing the scanning dose are needed. However, if the scanning dose is reduced, while reducing the radiation risk, it will increase the noise in the image, potentially affecting the diagnostic accuracy. Low-dose SPECT images face the problems of increased noise and degraded image quality while reducing the radiation exposure of patients. With the improvement of hardware parallel computing capabilities and the availability of large-scale training data sets, the application of deep learning technology in medical imaging has made remarkable progress. However, in SPECT imaging, it is difficult to obtain real image pairs as training data. For example, it is impossible to obtain low-dose and full-dose image pairs of the same patient during the same scanning process; traditional filtering methods can suppress the noise of low-dose SPECT images to a certain extent, but the resulting denoised images are often over-smoothed and significantly different from full-dose SPECT images; the current mainstream deep learning algorithms focus on convolutional neural networks (CNNs), U-Nets, and generative adversarial networks (GANs), which focus on local feature processing and are difficult to capture global information and long-term dependencies, and traditional deep learning algorithms face challenges such as insufficient denoising ability, difficulty in capturing global information, long-term dependencies, and unstable training processes.

[0003] As an emerging denoising technology, diffusion models have demonstrated powerful denoising capabilities in medical imaging fields such as computed tomography (CT) and positron emission tomography (PET). Diffusion models define a forward diffusion process, adding noise to the input data in multiple steps and gradually perturbing it, and then recovering the original signal by learning the reverse process, which has significant advantages in the denoising task. Although the classical diffusion denoising probability model (DDPM) is easy to define and has a wide range of applications, it only adds random Gaussian noise and is still restricted to the classical diffusion model framework, making it difficult to directly apply to the denoising task of low-dose SPECT images, and it has disadvantages such as many sampling steps and long inference time. Moreover, the research on diffusion models for low-dose SPECT images is still in its infancy. Existing work has not fully considered the influence of degradation operators and time steps on different dose levels, and at the same time has not combined the unique characteristics of diffusion models to solve the practical problem of difficult-to-obtain training data pairs.

[0004] Therefore, the traditional low-dose SPECT image denoising method cannot obtain the matching pairs of low-dose and conventional-dose SPECT images of the same patient in the same process, resulting in a large consumption of model training time and computing resources, and there is a problem of poor denoising effect on low-dose SPECT images. Summary of the Invention

[0005] Based on this, in order to solve the above technical problems, a low-dose SPECT image denoising method and system based on an unsupervised strategy and a generalized diffusion model are provided, which can reduce the demand for high-quality clean images, reduce the consumption of training time and computing resources, and reduce the cost of data acquisition and preparation; at the same time, reduce artifacts and improve the denoising effect.

[0006] A low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model, the method comprising:

[0007] Collect low-dose SPECT images, perform slicing processing after preprocessing the low-dose SPECT images to obtain training images containing contextual slice information;

[0008] Input the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determine an attenuation coefficient degradation operator according to the patient's body mass index, and perform image denoising based on the attenuation coefficient degradation operator to obtain a random Poisson noise image;

[0009] In the backward propagation stage of the unsupervised diffusion training model, extract optimization factors and time step features, perform denoising processing on the random Poisson noise image to obtain a predicted low-dose SPECT image; and perform parameter adjustment according to the neural network algorithm to obtain optimal weights;

[0010] Load the optimal weights into the reverse diffusion inference model, perform reverse diffusion with the predicted low-dose SPECT image as the diffusion end point, and perform denoising processing on the predicted low-dose SPECT image to obtain a predicted full-dose SPECT image.

[0011] In one embodiment, collecting low-dose SPECT images, performing slicing processing after preprocessing the low-dose SPECT images to obtain training images containing contextual slice information, includes:

[0012] Collect low-dose SPECT images and perform image preprocessing operations to determine the first format type corresponding to the preprocessed low-dose SPECT images;

[0013] Perform slicing processing on the low-dose SPECT images of the first format type, and store the training images containing contextual slice information obtained after slicing according to the second format type.

[0014] In one embodiment, a decay coefficient degradation operator is determined according to the patient's body mass index, and random Poisson noise images are obtained by adding noise to the image based on the decay coefficient degradation operator, including:

[0015] Determine the patient's body mass index based on the low-dose SPECT image, and calculate a decay coefficient degradation operator according to the patient's body mass index;

[0016] Add Poisson noise to the low-dose SPECT image step by step according to the decay coefficient degradation operator to obtain random Poisson noise images.

[0017] In one embodiment, in the backpropagation stage of the unsupervised diffusion training model, an optimization factor and a time step feature are extracted, including:

[0018] In the backpropagation stage of the unsupervised diffusion training model, an optimization factor is estimated based on the current time step and the Poisson noise through a feature adjustment network in the unsupervised diffusion training model;

[0019] Extract the time step feature through a feature extractor in the unsupervised diffusion training model, and adjust the time step feature according to the optimization factor to obtain an adjusted time step feature.

[0020] In one embodiment, the neural network algorithm is a second-order neural network algorithm;

[0021] Adjust the parameters according to the neural network algorithm to obtain the optimal weights, including:

[0022] Obtain the network parameters in the second-order neural network algorithm, and based on the predicted low-dose SPECT image, use the network parameters to adjust the time step of the second-stage process in the second-order neural network algorithm to obtain a recovery operator parameter and a step size adjustment network parameter;

[0023] Perform model iterative training on the neural network algorithm according to the recovery operator parameter and the step size adjustment network parameter to obtain the optimal weights.

[0024] In one embodiment, the backbone network of the neural network algorithm is a U-Net, including an upsampling layer, an input layer, an output layer, and an intermediate layer;

[0025] The Adam optimizer is used during the training process of the neural network algorithm, and the network framework of the neural network algorithm is implemented by PyTorch.

[0026] In one embodiment, the method further includes:

[0027] Collect real full-dose SPECT images, compare the predicted full-dose SPECT images with the real full-dose SPECT images, and obtain a comparison result;

[0028] Evaluate the prediction effect of the predicted full-dose SPECT images according to the comparison result to obtain an evaluation result.

[0029] In one embodiment, evaluating the prediction effect of the predicted full-dose SPECT images according to the comparison result includes:

[0030] According to the comparison result, use methods such as normalized mean square error, structural similarity, peak signal-to-noise ratio, joint histogram, and linear regression to evaluate the prediction effect of the predicted full-dose SPECT images.

[0031] A low-dose SPECT image denoising system based on an unsupervised strategy and a generalized diffusion model, the system includes:

[0032] An image processing module for collecting low-dose SPECT images, preprocessing the low-dose SPECT images and then performing slicing processing to obtain training images containing context slice information;

[0033] An attenuation coefficient degradation operator determination module for inputting the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determine the attenuation coefficient degradation operator according to the patient's body mass index, and perform image noise addition based on the attenuation coefficient degradation operator to obtain a random Poisson noise image;

[0034] A time step determination module for extracting optimization factors and time step features in the backward propagation stage of the unsupervised diffusion training model, denoising the random Poisson noise image to obtain a predicted low-dose SPECT image; and adjusting parameters according to the neural network algorithm to obtain optimal weights;

[0035] A denoising module for loading the optimal weights into a reverse diffusion inference model, performing reverse diffusion with the predicted low-dose SPECT image as the diffusion end point, and denoising the predicted low-dose SPECT image to obtain a predicted full-dose SPECT image.

[0036] In one embodiment, the image processing module is further configured to collect low-dose SPECT images and perform image preprocessing operations, determine the first format type corresponding to the preprocessed low-dose SPECT images; perform slicing processing on the low-dose SPECT images of the first format type, and store the training images containing context slice information obtained after slicing according to the second format type.

[0037] The above low-dose SPECT image denoising method and system based on unsupervised strategy and generalized diffusion model preprocess the low-dose SPECT image and then perform slicing to obtain training images. Through the forward and backward propagation of the unsupervised diffusion training model, it learns the features and noise distribution in the images, can automatically identify and remove noise, while retaining image details, and reduces the cost of data preparation. By combining the patient's body mass index to determine the attenuation coefficient degradation operator, personalized denoising according to patient information is achieved. Using the predicted low-dose SPECT image as the diffusion end point for reverse diffusion can enhance the ability to capture image details, reduce artifacts, and improve the denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is an application environment diagram of the low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model in an embodiment;

[0039] Figure 2 It is a schematic flowchart of the low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model in an embodiment;

[0040] Figure 3 It is a forward propagation diagram of the unsupervised diffusion training model in an embodiment, and the attenuation coefficient degradation operator is customized according to patient information;

[0041] Figure 4 It is a schematic diagram of the framework of the reverse propagation model of the unsupervised diffusion training model in an embodiment;

[0042] Figure 5 It is a reverse diffusion inference model diagram in an embodiment;

[0043] Figure 6 It is a schematic diagram of the process structure of the low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model in an embodiment;

[0044] Figure 7 It is a structural block diagram of the low-dose SPECT image denoising system based on unsupervised strategy and generalized diffusion model in an embodiment;

[0045] Figure 8 It is an internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0047] The low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model provided by the embodiments of the present application can be applied to, for example, Figure 1 the application environment shown as follows. As Figure 1 shown, the application environment includes a computer device 110. The computer device 110 can collect low-dose SPECT images, perform slicing processing on the low-dose SPECT images after preprocessing to obtain training images containing context slice information; the computer device 110 can input the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determine an attenuation coefficient degradation operator according to the patient's body mass index, and perform image denoising based on the attenuation coefficient degradation operator to obtain a random Poisson noise image; the computer device 110 can extract an optimization factor and a time step feature in the reverse propagation stage of the unsupervised diffusion training model, perform denoising processing on the random Poisson noise image to obtain a predicted low-dose SPECT image; and adjust parameters according to the neural network algorithm to obtain optimal weights; the computer device 110 can load the optimal weights into the reverse diffusion inference model, perform reverse diffusion with the predicted low-dose SPECT image as the diffusion end point, and perform denoising processing on the predicted low-dose SPECT image to obtain a predicted full-dose SPECT image. Among them, the computer device 110 can be but is not limited to various personal computers, laptop computers, robots, tablet computers and other devices.

[0048] In one embodiment, as Figure 2 shown, a low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model is provided, including the following steps:

[0049] Step 202, collect low-dose SPECT images, perform slicing processing on the low-dose SPECT images after preprocessing to obtain training images containing context slice information.

[0050] The computer device can collect low-dose SPECT images for processing to obtain training images for subsequent training operations of the diffusion model.

[0051] In one embodiment, a low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model provided can also include an image processing process. The specific process includes: collecting low-dose SPECT images and performing image preprocessing operations, determining the first format type corresponding to the preprocessed low-dose SPECT images; performing slicing processing on the low-dose SPECT images of the first format type, and storing the training images containing context slice information obtained after slicing according to the second format type.

[0052] The computer device can slice the low-dose SPECT images in TIF format and save them in IMA format. Then, the computer device can convert the sliced data into npy format for model training. During this process, a series of preprocessing operations such as normalization, image intensity adjustment, and histogram matching are also performed on the low-dose SPECT images to ensure the effectiveness and uniformity of the data, laying a solid foundation for subsequent model training.

[0053] Traditional supervised learning methods require a large number of clean images paired with noisy images as training data. However, in the actual process of SPECT image acquisition, it is often very difficult and time-consuming to obtain high-quality clean images. In this embodiment, through an unsupervised strategy, only low-dose images are used for training without the need for clean image pairing, greatly reducing the difficulty of data preparation; and the unsupervised strategy can quickly adapt to new data to meet the real-time requirements of the clinic. In practical applications, model training and image denoising can be quickly carried out without waiting for a large amount of data accumulation.

[0054] Step 204: Input the training images into the unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determine the attenuation coefficient degradation operator according to the patient's body mass index, and add noise to the image based on the attenuation coefficient degradation operator to obtain a random Poisson noise image.

[0055] In this embodiment, the generalized diffusion model can be composed of an unsupervised diffusion training model and a reverse diffusion inference model. Among them, the unsupervised diffusion training model includes a patient information-driven attenuation coefficient degradation operator module and a photon counting adaptive time step adjustment module in an unsupervised framework.

[0056] In the forward propagation stage of the unsupervised diffusion training model, as Figure 3 shown, customize the attenuation coefficient degradation operator through the patient information (height, weight, gender, age) corresponding to the input image, and gradually add noise to the low-dose SPECT image to obtain a random Poisson noise image.

[0057] Specifically, in one embodiment, a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model may further include a process of adding noise to the image. The specific process includes: determining the patient's body mass index based on the low-dose SPECT image, and calculating the attenuation coefficient degradation operator according to the patient's body mass index; gradually adding Poisson noise to the low-dose SPECT image according to the attenuation coefficient degradation operator to obtain a random Poisson noise image.

[0058] In low-dose SPECT images, the attenuation of photons passing through matter can be expressed by the exponential attenuation formula as: I = I0·e -ul, where I is the photon intensity received by the detector, I0 is the initial intensity of the photon source, u is the attenuation coefficient (usually in cm-1), and l is the thickness of the photon passing through the substance (usually in cm). In this embodiment, according to the thickness of the patient's body cross-section, i.e., l, can be estimated; the body fat percentage and lean body mass can be estimated based on BMI, thereby indirectly inferring the proportion of fat and non-fat tissues in the body, and the corresponding u is adopted according to different proportions; furthermore, the attenuation coefficient degradation operator D(x0,t)=(1 - I t )x0 + I t ρ t = x t is created, where x0 is the low-dose SPECT image, x t is the image after adding Poisson noise ρ t , I t is an arithmetic progression set of photon intensities at different training time steps and gradually increases with the time step t, and ρ t is the noise sampled from the standard Poisson distribution; during the forward propagation of the unsupervised diffusion training model, the low-dose SPECT image is gradually added with noise to a random Poisson noise distribution image, i.e., a random Poisson noise image, and the known prior distribution is converted into an unknown random distribution.

[0059] Step 206, in the backpropagation stage of the unsupervised diffusion training model, the optimization factor and the time step feature are extracted, the random Poisson noise image is denoised to obtain the predicted low-dose SPECT image; and the optimal weight is obtained by adjusting the parameters according to the neural network algorithm.

[0060] In the backpropagation stage of the unsupervised diffusion training model, as shown in Figure 4 , the random Poisson noise image is gradually denoised, and finally the predicted low-dose SPECT image is generated.

[0061] In one embodiment, a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model may further include a process of adjusting the time step feature. The specific process includes: in the backpropagation stage of the unsupervised diffusion training model, the optimization factor is estimated based on the current time step number and the Poisson noise through the feature adjustment network in the unsupervised diffusion training model; the time step feature is extracted by the feature extractor in the unsupervised diffusion training model, and the time step feature is adjusted according to the optimization factor to obtain the adjusted time step feature.

[0062] Among them, in the reverse diffusion process of the unsupervised diffusion training model, a feature adjustment network A δ is designed, that is, predicted based on the current time step number and the initial addition of Poisson noise ρ tTo estimate the optimization factors σ t-1 and τ t-1 . Then, the time-step features are extracted by a feature extractor, and the optimization factors are combined to adjust and map the time-step features. The specific calculation formula can be expressed as: where SinPE(t - 1) represents the sine position encoding of time step t - 1.

[0063] In one embodiment, a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model may further include a process of obtaining the optimal weights. The specific process includes: obtaining the network parameters in the second-order neural network algorithm, based on the predicted low-dose SPECT image, using the network parameters to adjust the time steps in the second stage of the second-order neural network algorithm to obtain the recovery operator parameters and the step-size adjustment network parameters; performing model iterative training on the neural network algorithm according to the recovery operator parameters and the step-size adjustment network parameters to obtain the optimal weights.

[0064] When adjusting the time-step features, the training code and mode in the reverse diffusion process are changed, and the first-order process of the original generalized diffusion model is improved to a second-order process. The specific formula can be expressed as:

[0065]

[0066] where R θ and A δ are the recovery operator parameters and the step-size adjustment network parameters learned in the reverse diffusion process, that is, the optimal weights; θ, δ are the parameters of the overall second-order neural network. The second-order neural network process combines the annealing sampling strategy of "recovery - degradation - recovery". The first-stage process is the same as that of the classical diffusion model, and the second-stage process combines the time-step adjustment strategy depending on the dose level to reduce the step-size cumulative error problem as much as possible and improve the denoising ability of the overall network model. The overall details of the second-order neural network are as Figure 5 shown, where x T is the low-dose SPECT image, which replaces the end point of the diffusion process; x0 is the predicted full-dose SPECT image.

[0067] In one embodiment, the backbone network of the neural network algorithm is U-Net, including an upsampling layer, a downsampling layer, an input layer, an output layer, and an intermediate layer; the Adam optimizer is used during the training process of the neural network algorithm, and the network framework of the neural network algorithm is implemented by PyTorch.

[0068] Specifically, the backbone network of the neural network algorithm is U-Net, which consists of two upsampling and downsampling layers, an input layer, an output layer, and an intermediate layer; the input size is 3×64×64, where the 3-channel image is composed of the low-dose SPECT image x combined with the upper and lower slice data x -1 and x +1 The Adam optimizer is used in the training process, the learning rate is set to 2e-4, and 150k iterations of training are set. The overall network framework is implemented by PyTorch and trained on an NVIDIA RTX 3090 GPU (24GB), with the batch size set to 8.

[0069] In this embodiment, a personalized attenuation coefficient degradation operator is customized by combining patient information, so that the degradation process of the SPECT image conforms to physical phenomena, and the robustness and noise reduction effect of the model are further consolidated and improved according to patient information; a feature adjustment network is used to extract optimization factors, and a time step feature is extracted by a feature extractor, and the adjustment and mapping of the time step feature are realized by combining the optimization factors, so that the time step and the photon count level are equivalent; in the unsupervised diffusion training model and the reverse diffusion inference model, a random Poisson noise image and a low-dose SPECT image are used as the end points of the diffusion process respectively, jumping out of the classical diffusion model framework and generalizing it to a generalized diffusion model, which greatly shortens the sampling time in training and provides the network with richer image information features, thereby enhancing the model's ability to capture image details.

[0070] Step 208: Load the optimal weights into the reverse diffusion inference model, perform reverse diffusion with the predicted low-dose SPECT image as the diffusion end point, and perform noise reduction processing on the predicted low-dose SPECT image to obtain the predicted full-dose SPECT image.

[0071] The computer device can put the trained R θ and A δ weights into the reverse diffusion inference model, use the low-dose SPECT image as the end point of the diffusion process, directly perform the reverse diffusion process on it, predict its denoised high-dose image and compare it with the full-dose image to evaluate the prediction effect. Among them, the network structure is as Figure 5 shown.

[0072] In this embodiment, the combination of the unsupervised strategy and the generalized diffusion model not only solves the problem of difficult data acquisition, but also further improves the denoising effect through the generation ability of the generalized diffusion model. This synergistic relationship enables the model to achieve excellent performance even with limited data. By introducing the context information between adjacent slices, the error adjustment network in the generalized diffusion model can better process the structural information in the image, reduce artifacts, and improve the denoising effect. The utilization of this context information in combination with the generalized diffusion model further enhances the quality of the image.

[0073] In one embodiment, a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model may further include a process of evaluating the prediction effect. The specific process includes: collecting real full-dose SPECT images, comparing the predicted full-dose SPECT images with the real full-dose SPECT images to obtain a comparison result; evaluating the prediction effect of the predicted full-dose SPECT images according to the comparison result to obtain an evaluation result.

[0074] Specifically, in one embodiment, the computer device may evaluate the prediction effect of the predicted full-dose SPECT images by using methods such as normalized mean square error, structural similarity, peak signal-to-noise ratio, joint histogram, and linear regression according to the comparison result.

[0075] The computer device can evaluate the PSNR, SSIM, NMSE, joint histogram, and linear regression of the denoised SPECT images, and compare the denoised result images generated by different models with the full-dose SPECT images as a reference, so as to judge the denoising effect of the low-dose SPECT image denoising method based on the unsupervised strategy and the generalized diffusion model in this application.

[0076] This application realizes an innovative technical integration in the framework of the generalized diffusion model, providing a new solution for the denoising task of low-dose SPECT images. The unsupervised strategy reduces the need for high-quality clean images and lowers the cost of data acquisition and preparation. At the same time, the efficiency of the generalized diffusion model reduces the consumption of training time and computing resources, further reducing the overall cost. By introducing the collaborative application of multiple technologies, the denoising ability for low-dose SPECT images can be significantly improved, and the sampling steps and inference time of the diffusion model can be shortened.

[0077] In one embodiment, a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model may also be as Figure 6 shown, and the specific process includes:

[0078] First, slice the low-dose SPECT images in TIF format, save them in IMA format, and convert the slice data to npy format for easy model training. During this process, a series of preprocessing operations such as normalizing the low-dose SPECT images, adjusting the image intensity, and histogram matching are also performed to ensure the effectiveness and uniformity of the data, laying a solid foundation for subsequent model training.

[0079] Next, input the images combined with context slice information into the unsupervised diffusion training model. Through overall training, continuously optimize the model parameters, and finally save the optimal weights. This enables the model to fully learn the noise characteristics and image structure information in the low-dose SPECT images, providing strong model support for subsequent noise reduction processing.

[0080] Load the saved optimal weights into the reverse diffusion inference model, and use this model to perform noise reduction processing on the low-dose SPECT images to predict the full-dose SPECT images. By means of the reverse process of the diffusion model, gradually remove the noise in the images and restore high-quality images closer to the real full-dose SPECT images.

[0081] By comparing with the real full-dose SPECT images, adopt objective evaluation indicators such as PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), and NMSE (Normalized Mean Square Error) to comprehensively evaluate the noise reduction effect. These indicators measure the similarity and quality difference between the denoised images and the real full-dose images from different angles, thus accurately reflecting the noise reduction performance of the proposed algorithm.

[0082] It should be understood that although the steps in each of the above flowcharts are shown in sequence according to the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in each of the above flowcharts may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily need to be executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0083] In one embodiment, as Figure 7 shown, a low-dose SPECT image noise reduction system based on an unsupervised strategy and a generalized diffusion model is provided, including: an image processing module 710, an attenuation coefficient degradation operator determination module 720, a time step determination module 730, and a noise reduction module 740, where:

[0084] The image processing module 710 is used to collect low-dose SPECT images, perform preprocessing on the low-dose SPECT images, and then perform slicing processing to obtain training images containing contextual slice information.

[0085] The attenuation coefficient degradation operator determination module 720 is used to input the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determine the attenuation coefficient degradation operator according to the patient's body mass index, and add noise to the images based on the attenuation coefficient degradation operator to obtain random Poisson noise images.

[0086] The time step determination module 730 is used to extract the optimization factor and time step features in the backward propagation stage of the unsupervised diffusion training model, perform denoising processing on the random Poisson noise images to obtain predicted low-dose SPECT images, and adjust the parameters according to the neural network algorithm to obtain the optimal weights.

[0087] The denoising module 740 is used to load the optimal weights into the reverse diffusion inference model, perform reverse diffusion with the predicted low-dose SPECT images as the diffusion end point, and perform denoising processing on the predicted low-dose SPECT images to obtain predicted full-dose SPECT images.

[0088] In one embodiment, the image processing module 710 is further used to collect low-dose SPECT images and perform image preprocessing operations, determine the first format type corresponding to the preprocessed low-dose SPECT images; perform slicing processing on the low-dose SPECT images of the first format type, and store the training images containing contextual slice information obtained after slicing according to the second format type.

[0089] In one embodiment, the attenuation coefficient degradation operator determination module 720 is further used to determine the patient's body mass index based on the low-dose SPECT images, and calculate the attenuation coefficient degradation operator according to the patient's body mass index; gradually add Poisson noise to the low-dose SPECT images according to the attenuation coefficient degradation operator to obtain random Poisson noise images.

[0090] In one embodiment, the time step determination module 730 is further used to estimate the optimization factor based on the current time step and Poisson noise through the feature adjustment network in the unsupervised diffusion training model in the backward propagation stage of the unsupervised diffusion training model; extract the time step features through the feature extractor in the unsupervised diffusion training model, and adjust the time step features according to the optimization factor to obtain the adjusted time step features.

[0091] In one embodiment, the neural network algorithm is a second-order neural network algorithm; the time step determination module 730 is further configured to obtain the network parameters in the second-order neural network algorithm, and based on the predicted low-dose SPECT image, use the network parameters to adjust the time step of the two-stage process in the second-order neural network algorithm to obtain the restoration operator parameters and the step adjustment network parameters; perform model iterative training on the neural network algorithm according to the restoration operator parameters and the step adjustment network parameters to obtain the optimal weights.

[0092] In one embodiment, the backbone network of the neural network algorithm is a U-Net, including an upsampling and downsampling layer, an input layer, an output layer, and an intermediate layer; the Adam optimizer is used during the training process of the neural network algorithm, and the network framework of the neural network algorithm is implemented by PyTorch.

[0093] In one embodiment, a low-dose SPECT image denoising system based on an unsupervised strategy and a generalized diffusion model may further include an evaluation module, configured to collect real full-dose SPECT images, compare the predicted full-dose SPECT images with the real full-dose SPECT images to obtain a comparison result; evaluate the prediction effect of the predicted full-dose SPECT images according to the comparison result to obtain an evaluation result.

[0094] In one embodiment, the evaluation module is further configured to evaluate the prediction effect of the predicted full-dose SPECT images according to the comparison result by using methods such as normalized mean square error, structural similarity, peak signal-to-noise ratio, joint histogram, and linear regression.

[0095] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as Figure 8 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0096] Those skilled in the art can understand,Figure 8 The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component layout.

[0097] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps of a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model are implemented.

[0098] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of a low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model are implemented.

[0099] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or an external cache. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0100] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0101] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model, characterized in that, The method includes: Collecting low-dose SPECT images, preprocessing the low-dose SPECT images, and then performing slicing processing to obtain training images containing contextual slice information; Inputting the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determining an attenuation coefficient degradation operator according to the patient's body mass index, and adding noise to the images based on the attenuation coefficient degradation operator to obtain a random Poisson noise image; In the backward propagation stage of the unsupervised diffusion training model, extracting an optimization factor and time step features, denoising the random Poisson noise image to obtain a predicted low-dose SPECT image; and adjusting the parameters according to the neural network algorithm to obtain the optimal weights; Loading the optimal weights into a reverse diffusion inference model, performing reverse diffusion with the predicted low-dose SPECT image as the diffusion end point, and denoising the predicted low-dose SPECT image to obtain a predicted full-dose SPECT image.

2. The low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model according to claim 1, characterized in that, Collecting low-dose SPECT images, preprocessing the low-dose SPECT images, and then performing slicing processing to obtain training images containing contextual slice information, including: Collecting low-dose SPECT images and performing image preprocessing operations to determine the first format type corresponding to the preprocessed low-dose SPECT images; Performing slicing processing on the low-dose SPECT images of the first format type, and storing the training images containing contextual slice information obtained after slicing according to the second format type.

3. The low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model according to claim 1, characterized in that, Determining an attenuation coefficient degradation operator according to the patient's body mass index, and adding noise to the images based on the attenuation coefficient degradation operator to obtain a random Poisson noise image, including: Determining the patient's body mass index based on the low-dose SPECT image, and calculating an attenuation coefficient degradation operator according to the patient's body mass index; Gradually adding Poisson noise to the low-dose SPECT image according to the attenuation coefficient degradation operator to obtain a random Poisson noise image.

4. The low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model according to claim 3, wherein In the backward propagation stage of the unsupervised diffusion training model, extracting an optimization factor and time step features, including: In the backward propagation stage of the unsupervised diffusion training model, estimating an optimization factor based on the current time step number and the Poisson noise through a feature adjustment network in the unsupervised diffusion training model; Extracting time step features through a feature extractor in the unsupervised diffusion training model, and adjusting the time step features according to the optimization factor to obtain adjusted time step features.

5. The low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model according to claim 4, wherein The neural network algorithm is a second-order neural network algorithm; Adjusting the parameters according to the neural network algorithm to obtain the optimal weights, including: Obtaining the network parameters in the second-order neural network algorithm, and based on the predicted low-dose SPECT image, adjusting the time step of the second-stage process in the second-order neural network algorithm using the network parameters to obtain a recovery operator parameter and a step size adjustment network parameter; Performing model iterative training on the neural network algorithm according to the recovery operator parameter and the step size adjustment network parameter to obtain the optimal weights.

6. The low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model according to claim 5, characterized in that, The backbone network of the neural network algorithm is U-Net, which includes upsampling and downsampling layers, an input layer, an output layer, and an intermediate layer; During the training process of the neural network algorithm, the Adam optimizer is adopted, and the network framework of the neural network algorithm is implemented by PyTorch.

7. The low-dose SPECT image denoising method based on unsupervised strategy and generalized diffusion model according to claim 1, characterized in that The method further includes: Collecting real full-dose SPECT images, and comparing the predicted full-dose SPECT images with the real full-dose SPECT images to obtain a comparison result; Evaluating the prediction effect of the predicted full-dose SPECT images according to the comparison result to obtain an evaluation result.

8. The low-dose SPECT image denoising method based on an unsupervised strategy and a generalized diffusion model according to claim 7, characterized in that Evaluating the prediction effect of the predicted full-dose SPECT images according to the comparison result includes: According to the comparison result, using methods such as normalized mean square error, structural similarity, peak signal-to-noise ratio, joint histogram, and linear regression to evaluate the prediction effect of the predicted full-dose SPECT images.

9. A low-dose SPECT image denoising system based on an unsupervised strategy and a generalized diffusion model, characterized in that, The system includes: An image processing module for collecting low-dose SPECT images, performing preprocessing on the low-dose SPECT images and then slicing them to obtain training images containing context slice information; An attenuation coefficient degradation operator determination module for inputting the training images into an unsupervised diffusion training model. In the forward propagation stage of the unsupervised diffusion training model, determining the attenuation coefficient degradation operator according to the patient's body mass index, and adding noise to the images based on the attenuation coefficient degradation operator to obtain random Poisson noise images; A time step determination module for extracting optimization factors and time step features in the backward propagation stage of the unsupervised diffusion training model, denoising the random Poisson noise images to obtain predicted low-dose SPECT images; and adjusting parameters according to the neural network algorithm to obtain optimal weights; A noise reduction module for loading the optimal weights into a reverse diffusion inference model, performing reverse diffusion with the predicted low-dose SPECT images as the diffusion end point, and denoising the predicted low-dose SPECT images to obtain predicted full-dose SPECT images.

10. The low-dose SPECT image denoising system based on an unsupervised strategy and a generalized diffusion model according to claim 9, wherein The image processing module is further used for collecting low-dose SPECT images and performing image preprocessing operations, determining the first format type corresponding to the preprocessed low-dose SPECT images; slicing the low-dose SPECT images of the first format type, and storing the training images containing context slice information obtained after slicing according to the second format type.

Images