Rapid bone scanning image noise reduction method based on state space unsupervised diffusion network

Abstract

The invention discloses a rapid bone scanning image noise reduction method based on a state space unsupervised diffusion network, and the method comprises the steps: collecting a noise-containing rapid bone scanning original image, carrying out the preprocessing, and constructing a non-diffusion module with a consistent period architecture; step-by-step noise adding is carried out on the conventional bone scanning image, a pure Gaussian noise image is generated, and the step-by-step noise adding process is based on a forward diffusion process of a Markov chain; fusing the pure Gaussian noise image and the rapid bone scanning image, and gradually denoising the fused image through a state space model; the denoising process is based on back diffusion of a Markov chain, and each back diffusion step is mapped through a state space model to output a high-quality conventional bone scanning image. According to the method, the rapid bone scanning image containing noise is converted into the high-quality image by utilizing the diffusion network based on the state space, the problem that the scanning time of a traditional whole-body bone scanning technology is long is solved, the application range is wide, and rapid, accurate, lossless and high-quality bone scanning image noise reduction can be achieved.

Description

TECHNICAL FIELD

[0001] The application relates to a state space unsupervised diffusion network-based fast bone scan image denoising method and belongs to the technical field of image processing. BACKGROUND

[0002] Similar to whole-body bone scanning as an important diagnostic technique, it plays an irreplaceable role in bone disease diagnosis and tumor bone metastasis detection. However, the traditional whole-body bone scanning technique usually requires a scanning process of 10 minutes or even longer, which not only increases the discomfort of patients but also limits the use efficiency of medical equipment. With the increasing demand for medical treatment, the demand for fast, efficient and accurate bone scanning technology is also increasingly urgent. Although accelerating the scanning speed is a direct solution, it often comes at the cost of image quality, which in turn affects the accuracy of diagnosis. Therefore, how to improve the scanning speed while maintaining or even improving the image quality has become the core problem of research in this field.

[0003] In recent years, although deep learning algorithms have made progress in medical image processing, there are still the following problems: 1. Most studies are based on simulation or limited clinical data, which cannot fully reflect the complexity of clinical practice; 2. Relies on a large number of paired image data, which is difficult to achieve in actual clinical practice; 3. Cannot improve image quality while maintaining diagnostic details. Therefore, developing a method to effectively improve the image quality of fast bone scanning images in real clinical environments, shorten the scanning time, and ensure diagnostic quality without relying on a large number of paired image data has become an important direction for speeding up bone scanning image research. SUMMARY

[0004] To solve the above problems, the application provides a state space unsupervised diffusion network-based fast bone scan image denoising method, which can improve the image quality of fast bone scanning.

[0005] In a first aspect, the application provides a fast bone scan image denoising method, comprising the following steps:

[0006] Step 1, collecting a noisy fast bone scan original image and performing preprocessing, the preprocessing including size standardization, pixel value normalization and contrast adjustment;

[0007] Step 2, constructing a non-diffusion module with a periodic consistent architecture, including two generators and two discriminators; one of the generators is used to convert the preprocessed noisy image into a denoised image, the other generator is used to reversely convert the denoised image into a noisy image, and the two discriminators are used for adversarial training of the images generated by the two generators;

[0008] Step 3, gradually add noise to the conventional bone scan image to generate a pure Gaussian noise image, the gradually adding noise process is based on a forward diffusion process of Markov chain;

[0009] Step 4, fuse the pure Gaussian noise image generated in step 3 with the fast bone scan image obtained in step 1, and gradually denoise the fused image through a state space model; the denoising process is based on a reverse diffusion of Markov chain, and each reverse diffusion step is mapped through a state space model to output a high-quality conventional bone scan image.

[0010] In an embodiment of the present application, the generator in step 2 is based on a U-Net structure, including an encoder-decoder architecture; the encoder is composed of several convolutional layers for extracting image features, and an IN regularization layer and a Leaky ReLU activation function are arranged after each convolutional layer to enhance the non-linear expression ability and stability of the model; several residual blocks are arranged between the encoder and the decoder, each residual block includes two convolutional layers, an IN regularization layer, and a ReLU activation function, the input information is preserved through cross-layer connection to enhance the effect of feature extraction; the decoder is composed of deconvolutional layers for gradually restoring the spatial resolution of the image, and an IN regularization layer and a ReLU activation function are arranged after each deconvolutional layer to ensure that the generated image is detailed and consistent; a ReflectionPad2d layer is used in the last layer to expand the image symmetrically along the edge to improve the image resolution, and finally a convolutional layer is used to restore the image to the original size.

[0011] In an embodiment of the present application, the generator is constructed according to the following formula:

[0012]

[0013] wherein g layer represents the lth layer of the network, and L represents the total number of layers of the network; the generator reduces the complexity of the model by reducing the number and depth of convolutional layers;

[0014] The generator is trained by using a combination of adversarial loss and cycle consistency loss function, the cycle consistency loss function is used to ensure that the denoised image is consistent with the original image after reverse conversion, and follows the following formula:

[0015]

[0016] wherein G A is the first generation network, G B is the second generation network, I noisy is a noise image, is a cycle consistency loss function;

[0017] The goal of the generator is to minimize the following loss function:

[0018]

[0019] wherein, is a feature extraction loss for ensuring that the generated denoised image is consistent with the real image at the feature level, λ adv and λ FE are respective weight parameters, is an adversarial loss for the target domain.

[0020] In an embodiment of the present application, the discriminator in step 2 adopts a PatchGAN structure sensitive to local features, which is composed of multiple convolutional layers, each followed by an instance normalization layer and a LeakyReLU activation function; the PatchGAN discriminator enhances the detection capability of local features by dividing the input image into multiple small blocks and discriminating each small block; the input of the discriminator is a pair of images, including a real image or a generated image, for distinguishing whether the input image comes from the real distribution or the generated distribution; the input layer receives the input image and sends it to the subsequent network layer; the first convolutional layer uses several 3x3 convolutional kernels for feature extraction, and the number of convolutional kernels is gradually increased in the subsequent layers to capture more rich feature information; after multiple layers of convolution, instance normalization and activation function processing, a probability matrix is finally output through a Sigmoid activation function, which is used to distinguish whether each small block is a real image or a generated image.

[0021] In an embodiment of the present application, the noise adding process in step 3, the image x t at the current state t-1 The corresponding forward conversion probability of the previous moment x

[0022]

[0023] wherein, β t is the noise variance, ε is the added noise, N is the Gaussian distribution, and I is the identity covariance matrix; the backward diffusion is from x T to x0, and each step aims to gradually denoise the sample; in the case of large T and small β t , x t-1 and x t are approximately Gaussian distribution:

[0024] q(x t-1 |x t )=N(x t-1 ;μ(x t ,t),∑(x t ,t))。

[0025] In one embodiment of the application, step 4 specifically comprises:

[0026] Step 41: The fused noise image is inputted into a segmentation into two-dimensional patches, each patch is converted into a high-dimensional feature vector by a fully connected layer; different scanning modes are used on different model blocks to make each patch have a global receptive field, the scanning modes include: in the first block, row-first scanning is used, the image block sequence is scanned line by line, each line is scanned horizontally from left to right, and then moves to the next line; in the second module, the sequence order is reversed and the sequence is scanned in the same way; in the subsequent data blocks, column-first scanning is performed in forward and reverse order; after traversing all the scanning modes, the next module is cycled again;

[0027] Step 42: The state space model is used to process a noisy two-dimensional input, which contains time steps and class conditions; the noisy input is considered as a clean signal disturbed by Gaussian noise corresponding to the input time steps; the noisy input is first segmented into two-dimensional patches, each patch is converted into a high-dimensional feature vector by a fully connected layer; these patches are fed into a 3x3 deep convolutional layer, where local information is injected into the patches; these patches are also padded with learnable markers at the end of the row and column, so that the model can understand its two-dimensional spatial structure during one-dimensional sequential scanning; using one of the above scanning modes, the patch markers are flattened into a patch sequence; the time steps and class conditions are also converted into markers by a fully connected layer, and then attached to the sequence; the state space model aims to encode and decode one-dimensional sequential input; in a continuous-time SSM, the input signal x(i) is first encoded into a hidden state vector h(i), and then decoded into an output signal y(i); according to the following ordinary differential equation:

[0028] h'(i) = Ah(i) + Bx(i), y(i) = Ch(i) + Dx(i)

[0029] h'(i) represents the derivative of h, B, C, D represent the weights of the SSM; the input of natural language and two-dimensional vision is a discrete signal, so Mamba uses the zero-order hold rule for discretization; the above ordinary differential equation is repeatedly solved:

[0030] A = exp(ΔA), B = (ΔA) -1 (exp(ΔA)-I)·ΔB,

[0031] h i =Ah i-1 +Bx i ,y i =Ch i +Dx i

[0032] Subsequently, the sequence is fed into the Mamba block for scanning, and finally a high-quality bone scan image is output.

[0033] In an embodiment of the present application, the method further comprises:

[0034] Step 5: At the end of each training cycle, the quality of the generated images is evaluated using quantitative indicators and qualitative analysis; early stopping is performed on the model to avoid overfitting, and the model with the best performance on the validation set is selected for inspection;

[0035] Step 6: Deploy the trained model in the actual fast bone scan image denoising process, evaluate the effectiveness and user satisfaction of the model in the real clinical environment, and ensure that the model can improve the scanning efficiency and maintain or improve the diagnostic quality;

[0036] Step 7: Perform extensive stability and robustness testing on the model, including performance evaluation under different noise levels, image quality, and device difference conditions, to ensure that the model maintains stable denoising effect in diverse clinical environments.

[0037] In a second aspect, the present application provides a fast bone scan image denoising system, applied to the fast bone scan image denoising method, the system comprising:

[0038] An image acquisition and preprocessing module acquires a noisy fast bone scan original image and performs preprocessing, the preprocessing including size standardization, pixel value normalization, and contrast adjustment;

[0039] A construction module is used to construct a non-diffusion module with a periodic consistent architecture, including two generators and two discriminators; one of the generators is used to convert the preprocessed noisy image into a denoised image, the other generator is used to convert the denoised image into a noisy image, and the two discriminators are used for adversarial training of the images generated by the two generators;

[0040] A noise adding module is used to add noise to the conventional bone scan image step by step to generate a pure Gaussian noise image, and the step-by-step noise adding process is based on the forward diffusion process of Markov chain;

[0041] A fusion output module is used to fuse the generated pure Gaussian noise image with the obtained fast bone scan image, and to denoise the fused image step by step through a state space model; the denoising process is based on the reverse diffusion of Markov chain, and each reverse diffusion step is mapped through a state space model to output a high-quality conventional bone scan image.

[0042] In a third aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the fast bone scan image denoising method when executing the computer program.

[0043] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the fast bone scan image denoising method when executed by a processor.

[0044] The present application has the following beneficial effects:

[0045] The fast bone scan image denoising method based on a state space unsupervised diffusion network provided by the present application significantly reduces the computational complexity of the model by using a lightweight generation network, improves the image processing speed, and thus ensures the efficiency in the fast response clinical environment. The non-diffusion module can generate a source image matched with the target image in the training set, eliminating the need to collect paired source-target images from a large number of subjects. Compared with other generative networks, the diffusion model performs better in the reliability of network mapping, thereby improving the quality and diversity of samples. The sequence model backbone based on the state space model has shown significant effectiveness and efficiency in various image tasks. In addition, Mamba achieves comparable performance to Transformer while being more efficient in inference time. The application of the cycle-consistency loss function ensures the consistency of the denoised image in structure, effectively avoiding the loss of important diagnostic information during processing. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0047] Figure 1 The basic flowchart of the fast bone scan image denoising method based on a state space unsupervised diffusion network provided by the present application.

[0048] Figure 2 The structure diagram of the forward noise addition of the fast bone scan image denoising method based on a state space unsupervised diffusion network provided by the present application.

[0049] Figure 3 The structure diagram of the reverse denoising of the fast bone scan image denoising method based on a state space unsupervised diffusion network provided by the present application.

[0050] Figure 4 The generator structure schematic diagram of the non-diffusion module of the fast bone scan image denoising method based on the state space unsupervised diffusion network provided by the present application.

[0051] Figure 5 The local structure diagram in the generator of the non-diffusion module of the fast bone scan image denoising method based on the state space unsupervised diffusion network provided by the present application.

[0052] Figure 6 The discriminator structure schematic diagram of the non-diffusion module of the fast bone scan image denoising method based on the state space unsupervised diffusion network provided by the present application.

[0053] Figure 7 The overall structure schematic diagram of the diffusion module of the fast bone scan image denoising method based on the state space unsupervised diffusion network provided by the present application.

[0054] Figure 8 The state space model local structure diagram of the diffusion module of the fast bone scan image denoising method based on the state space unsupervised diffusion network provided by the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0056] Embodiment one:

[0057] The present embodiment provides a fast bone scan image denoising method, comprising the following steps:

[0058] Step 1, collecting a noisy fast bone scan original image and performing preprocessing, the preprocessing including size standardization, pixel value normalization and contrast adjustment;

[0059] Step 2, constructing a non-diffusion module with a consistent architecture, including two generators and two discriminators; one of the generators is used to convert the preprocessed noisy image into a denoised image, the other generator is used to reversely convert the denoised image into a noisy image, and the two discriminators are respectively used for adversarial training of the images generated by the two generators;

[0060] Step 3, gradually adding noise to the conventional bone scan image to generate a pure Gaussian noise image, the gradual noise adding process is based on the forward diffusion process of Markov chain;

[0061] Step 4, fuse the pure Gaussian noise image generated in step 3 with the fast bone scan image obtained in step 1, and perform step-by-step denoising on the fused image through a state space model (MambaSSM); the denoising process is based on the back diffusion of Markov chains, and each back diffusion step is mapped through a state space model to output a high-quality conventional bone scan image.

[0062] Embodiment Two:

[0063] Referring to Figures 1-8 , the embodiment provides a fast bone scan image denoising method based on a state space unsupervised diffusion network, and the denoising process is shown in Figure 1 , the embodiment also details the training process of the non-diffusion module and the diffusion module based on the state space unsupervised diffusion network. The fast bone scan image denoising method comprises the following steps:

[0064] Step 1, collect a noisy fast bone scan original image, and perform size standardization, pixel value normalization and contrast adjustment preprocessing on the image;

[0065] In this embodiment, a series of noisy fast bone scan original images are collected, which are obtained in the actual scanning process and may contain noise caused by device performance limitations or insufficient acquisition time; the images are carefully preprocessed, including but not limited to image size standardization and pixel value normalization, to ensure that the image data is suitable for subsequent model processing; it also includes adjusting the image contrast to highlight the bone structure features and normalizing the pixel values to adapt to the model input.

[0066] Step 2, construct a non-diffusion module in a cycle-consistent architecture, use a cycle-consistent architecture containing two generators and two discriminators to input the preprocessed noisy image into one of the generators to generate a pseudo-target image, and simultaneously realize the inverse mapping of the pseudo-target image to the source image through the other generator, and the discriminators are locally feature sensitive PatchGAN structures for adversarial training;

[0067] In order to increase the paired samples, the non-diffusion module in the cycle-consistent architecture is used in this embodiment, including two generators and two discriminators, to generate source images paired with the target images in the training set processed in step 1. The structure of the generator is shown in Figure 4 and Figure 5 One of the generators is used to convert the preprocessed noisy low-quality image into a denoised high-quality image, and the other generator is used to reversely convert the denoised high-quality image into a noisy low-quality image to realize cycle consistency; two discriminators are designed for the source domain and the target domain respectively, and the two discriminators are respectively used to evaluate the authenticity of the generated denoised high-quality image and the noisy low-quality image, and provide feedback to guide the optimization of the generation. The discriminator is as shown inFigure 6 As shown, the PatchGAN structure is adopted, focusing on local feature discrimination and improving sensitivity to details.

[0068] Step 21, the generator is constructed based on the U-Net structure with an encoder-decoder architecture; the input layer of the generator receives the input image and sends it to the subsequent network layers; the encoder part is composed of several convolutional layers (Convolutional Layers) for extracting image features, followed by an IN normalization layer (Instance Normalization) and a Leaky ReLU activation function after each convolutional layer to enhance the non-linear expression ability and stability of the model; between the encoder and the decoder, several residual blocks (Residual Blocks) are used, each including two convolutional layers, an IN normalization layer, and a ReLU activation function, which preserves input information through skip connections to enhance feature extraction; the decoder part is composed of transposed convolutional layers (Transposed Convolutional Layers) for gradually restoring the spatial resolution of the image, followed by an IN normalization layer and a ReLU activation function after each transposed convolutional layer to ensure that the generated image is detailed and consistent; the last layer uses the ReflectionPad2d layer, which is an image enhancement technique that expands the image symmetrically along the edges to improve the image resolution, and finally restores the image to the original size through a convolutional layer to ensure that the generated image has the same size as the input image and effectively solves the problem of object edge information;

[0069] The construction of the generator follows the following formula:

[0070]

[0071] where g layer represents the lth layer of the network, and L represents the total number of layers of the network; the generator reduces the complexity of the model by reducing the number and depth of convolutional layers.

[0072] Step 22, the generator is trained using a combination of adversarial loss and cycle consistency loss function; the cycle consistency loss function is used to ensure that the denoised image remains consistent after being converted back to the original image, following the formula:

[0073]

[0074] where G A is the first generation network, G B is the second generation network, I noisy is the noisy image, is a cycle-consistent loss function;

[0075] The objective of the generator is to minimize the following loss function:

[0076]

[0077] wherein, is a feature extraction loss, used to ensure that the generated denoised image is consistent with the real image at the feature level, λ adv and λ FE are the corresponding weight parameters, is an adversarial loss for the target domain.

[0078] Step 23, the training of the discriminator is alternately performed to maintain the balance between the generator and the discriminator. As shown in Figure 5 , the discriminator adopts a PatchGAN architecture, composed of multiple convolutional layers, each followed by an instance normalization layer and a Leaky ReLU activation function. Unlike traditional global discriminators, the PatchGAN discriminator enhances the detection capability of local features by dividing the input image into multiple small patches and discriminating each patch. The input of the discriminator is a pair of images (real image or generated image), and the main purpose is to distinguish whether the input image is from the real distribution or the generated distribution. The input layer receives the input image and sends it to the subsequent network layer; the first convolutional layer uses several 3x3 convolutional kernels for feature extraction, and the number of convolutional kernels is gradually increased in the subsequent layers to capture more rich feature information. Each convolutional layer is followed by an instance normalization layer and a Leaky ReLU activation function to enhance the non-linear expression ability and stability of the model. After multiple layers of convolution, instance normalization and activation function processing, a probability matrix is finally output through a Sigmoid activation function to distinguish whether each small block is a real image or a generated image. In the training process, the generator tries to generate images that are more and more difficult to be identified by the discriminator, while the discriminator constantly learns to better distinguish real and fake images.

[0079] Step 3, the image generated in step 2 is gradually added with noise to the conventional bone scan image obtained in step 1. As shown in Figure 2 , the conventional diffusion model maps between pure noise samples and actual images through a gradual process for T time steps. In the forward direction, a small amount of Gaussian noise is repeatedly added to the input image, i.e., the bone scan image x0~q(x0) of the conventional scan time; the sample x T is obtained from an isotropic Gaussian distribution, and the forward diffusion forms a Markov chain; in the process of adding noise, the image xt with the previous time x t-1 The corresponding forward transition probability is:

[0080]

[0081] where β t is the noise variance, ε is the added noise, N is the Gaussian distribution, and I is the identity covariance matrix; the backward diffusion goes from x T to x0, and each step aims to gradually denoise the samples; in the case of a large T and a small β t , x t-1 and x t can be approximated as Gaussian distributions:

[0082] q(x t-1 |x t ) = N(x t-1 ; μ(x t , t),∑(x t , t))

[0083] Step 4, the pure Gaussian noise image of the conventional bone scan obtained in step 3 is fused with the fast bone scan image paired with it obtained in step 1 in a direct superposition manner, and the fused image is gradually denoised. The denoising process is shown in Figure 3 , which is also a Markov chain, specifically each backward diffusion step is mapped through a state space model (MambaSSM), and finally a high-quality conventional bone scan image is obtained.

[0084] Step 41, the fused noise input is divided into two-dimensional patches, and each patch is converted into a high-dimensional feature vector by a fully connected layer. In order to make each patch have a global receptive field, different scanning modes are used on different model blocks. Specifically, four scanning methods are given, in the first block, row-first scanning is used, that is, the image block sequence is scanned row by row, and each row is scanned horizontally from left to right, and then moved to the next row. In the second module, the sequence order is reversed and the sequence is scanned in the same way. In subsequent data blocks, column-first scanning is performed in forward and reverse order. After traversing all scanning modes, they are cycled again in the next module.

[0085] Step 42, as Figure 7As shown, the state space model (Mamba SSM) is used to process a noisy two-dimensional (2D) input, such as an image or latent features, containing time steps and class conditions. This noisy input can be considered as a clean signal perturbed by a certain level of Gaussian noise corresponding to the input time steps. The noisy input is first segmented into 2D patches, each of which is converted to a high-dimensional feature vector by a fully connected layer. Next, these patches are fed into a 3x3 deep convolutional layer, where local information is injected into the patches. The patches are also padded with learnable markers at the end of the rows and columns, enabling the model to understand its 2D spatial structure during the one-dimensional sequential scan. Then, the patch markers are flattened into a patch sequence using one of the four scan patterns described above. The time steps and class conditions are also converted to markers by a fully connected layer and then appended to the sequence. The state space model (SSM) is designed to encode and decode one-dimensional sequential inputs, with its specific structure as shown in Figure 8 In a continuous-time SSM, the input signal x(i) is first encoded into a hidden state vector h(i) and then decoded into an output signal y(i); according to the following ordinary differential equation (ODE):

[0086] h'(i) = Ah(i) + Bx(i), y(i) = Ch(i) + Dx(i)

[0087] h'(i) represents the derivative of h, B, C, D represent the weights of the SSM; usually, the input of natural language and two-dimensional vision is a discrete signal, therefore Mamba utilizes the zero-order hold (ZOH) rule for discretization; thus, the above ordinary differential equation can be solved iteratively:

[0088] A = exp(ΔA), B = (ΔA) -1 (exp(ΔA) - I) · ΔB,

[0089] h i = Ah i-1 +Bx i ,y i = Ch i +Dx i

[0090] Subsequently, the sequence is fed into the Mamba block for scanning, and finally outputs high-quality bone scan images.

[0091] Step 5, at the end of each training cycle, evaluate the quality of the generated images using quantitative indicators (such as PSNR and SSIM) and qualitative analysis (such as visual inspection); early stop the model to avoid overfitting, and select the best performance model on the validation set for inspection.

[0092] Step 6, deploy the trained model in the actual fast bone scan image denoising process, cooperate with medical professionals, evaluate the effectiveness and user satisfaction of the model in the real clinical environment, and ensure that the model can improve scanning efficiency and maintain or improve diagnostic quality.

[0093] Step 7, extensive stability and robustness testing of the model, including performance evaluation under different noise levels, image quality, device differences and other conditions, to ensure that the model can maintain stable denoising effect in diversified clinical environment.

[0094] In summary, the method is used to obtain fast bone scan images, and the paired sample number is increased using the generative network to solve the problem of difficulty in obtaining medical images. The diffusion network based on state space is used to convert the fast bone scan image containing noise to a high-quality image, solving the problem of long scanning time of traditional whole body bone scan technology, and the application range is wide. The fast bone scan image denoising method can realize fast, accurate, non-destructive and high-quality bone scan image denoising.

[0095] Example Three

[0096] The embodiment provides a fast bone scan image denoising system, which is applied to the fast bone scan image denoising method as described above, and the system comprises:

[0097] An image acquisition and preprocessing module acquires a fast bone scan original image containing noise and performs preprocessing, and the preprocessing includes size standardization, pixel value normalization and contrast adjustment.

[0098] A construction module is used to construct a non-diffusion module with a consistent architecture, including two generators and two discriminators. One generator is used to convert the preprocessed image containing noise to a denoised image, and the other generator is used to reversely convert the denoised image to an image containing noise. The two discriminators are used for adversarial training of the images generated by the two generators.

[0099] A noise adding module is used to add noise to the conventional bone scan image step by step to generate a pure Gaussian noise image, and the step-by-step noise adding process is based on the forward diffusion process of Markov chain.

[0100] A fusion output module is used to fuse the generated pure Gaussian noise image with the obtained fast bone scan image, and to denoise the fused image step by step through a state space model. The denoising process is based on the reverse diffusion of Markov chain, and each reverse diffusion step is mapped through a state space model to output a high-quality conventional bone scan image.

[0101] The specific definitions and beneficial effects of the above quick bone scan image denoising system can refer to the definitions of the quick bone scan image denoising method, which will not be repeated here. The above modules can be realized by software, hardware and their combinations. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory in the computer device in software form, so as to call and execute the operations corresponding to the above modules by the processor.

[0102] In addition, the present application also provides a computer device, which can include a processor, a memory, a network interface and a database connected by a system bus. 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, a computer program and a database. 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 the external terminal through the network connection. The computer program is executed by the processor to make the processor execute the steps of the quick bone scan image denoising method of any one of the above embodiments.

[0103] The working process, working details and technical effects of the computer device provided in the embodiment can refer to the above embodiments of the quick bone scan image denoising method, which will not be repeated here.

[0104] In addition, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by the processor to realize the steps of the quick bone scan image denoising method of any one of the above embodiments. The computer readable storage medium refers to a carrier for storing data, which can include but is not limited to floppy disk, optical disk, hard disk, flash memory, USB flash disk and / or memory stick, etc. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices.

[0105] The working process, working details and technical effects of the computer readable storage medium provided in the embodiment can refer to the above embodiments of the quick bone scan image denoising method, which will not be repeated here.

[0106] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. 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 external cache memory. As an illustration but 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).

[0107]

[0108] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions recorded in the foregoing embodiments, or replace some of the technical features with equivalent ones, as long as the modifications, replacements, improvements, etc. are within the spirit and principles of the present application.​

Claims

1. A fast bone scan image noise reduction method, characterized in that, Includes the following steps: Step 1: Acquire noisy, rapid bone scan raw images and perform preprocessing, including size normalization, pixel value normalization, and contrast adjustment. Step 2: Construct a non-diffusion module with a periodically consistent architecture, including two generators and two discriminators; One generator is used to convert a preprocessed noisy image into a denoised image, and another generator is used to convert the denoised image back into a noisy image. Two discriminators are used to perform adversarial training on the images generated by the two generators, respectively. Step 3: Gradually add noise to the conventional bone scan image to generate a pure Gaussian noise image. The gradual noise addition process is based on the forward diffusion process of a Markov chain. Step 4: Fuse the pure Gaussian noise image generated in Step 3 with the fast bone scan image obtained in Step 1. Perform stepwise denoising on the fused image using a state-space model. The denoising process is based on the back diffusion of a Markov chain. Each back diffusion step is mapped through a state-space model to output a high-quality conventional bone scan image.

2. The method for rapid bone scan image noise reduction according to claim 1, characterized in that, The generator in step 2 is based on the U-Net architecture, including an encoder-decoder architecture; The encoder consists of several convolutional layers for extracting image features. Each convolutional layer is followed by an IN regularization layer and a Leaky ReLU activation function to enhance the model's nonlinear expressiveness and stability. Several residual blocks are placed between the encoder and decoder. Each residual block includes two convolutional layers, an IN regularization layer, and a ReLU activation function. These cross-layer connections preserve input information and enhance feature extraction. The decoder consists of deconvolutional layers to progressively restore the image's spatial resolution. Each deconvolutional layer is followed by an IN regularization layer and a ReLU activation function to ensure rich and consistent detail in the generated image. The final layer uses a ReflectionPad2d layer to symmetrically expand the image along its edges, improving image resolution. Finally, a convolutional layer restores the image to its original size.

3. The method for rapid bone scan image noise reduction according to claim 2, characterized in that, The generator is constructed according to the following formula: Among them, g layer The generator represents the l-th layer of the network, and L represents the total number of layers in the network. The generator reduces the complexity of the model by reducing the number and depth of convolutional layers. The generator is trained using a combination of adversarial loss and cycle consistency loss. The cycle consistency loss function is used to ensure that the denoised image remains consistent with the original image after inverse transformation, following the formula: Among them, G A It is the first-generation network, G B It is the second generative network, I noisy It's a noisy image. It is the cycle consistency loss function; The goal of the generator is to minimize the following loss function: in, It is the feature extraction loss, used to ensure that the generated denoised image is consistent with the real image at the feature level, λ adv and λ FE These are the corresponding weight parameters. It is an adversarial loss targeting the target domain.

4. The method for rapid bone scan image noise reduction according to claim 3, characterized in that, The discriminator in step 2 employs a local feature-sensitive PatchGAN structure, consisting of multiple convolutional layers. Each convolutional layer is followed by an instance normalization layer and a Leaky ReLU activation function. The PatchGAN discriminator enhances its ability to detect local features by dividing the input image into multiple small blocks and discriminating between each block. The input to the discriminator is a pair of images, including real or generated images, used to distinguish whether the input image comes from a real or generated distribution. The input layer receives the input image and feeds it into subsequent network layers. The first convolutional layer uses several 3×3 convolutional kernels for feature extraction, and the number of convolutional kernels is gradually increased in subsequent layers to capture richer feature information. After multiple convolutions, instance normalization, and activation function processing, a probability matrix is ​​finally output through a Sigmoid activation function to determine whether each block is a real or generated image.

5. The method for rapid bone scan image noise reduction according to claim 4, characterized in that, The noise addition process in step 3, the current state of the image x t Compared to the previous moment x t-1 The corresponding forward transition probability is: Where, β t Here, ε is the noise variance, ε is the added noise, N is a Gaussian distribution, and I is the identity covariance matrix; backdiffusion from x T Up to x0, each step aims to progressively reduce noise in the samples; at larger T and β... t In the case of a smaller value, x t-1 and x t Approximately Gaussian distribution: q(x t-1 |x t )=N(x t-1 ;μ(x t ,t),∑(x t ,t))。 6. The method for rapid bone scan image noise reduction according to claim 5, characterized in that, Step 4 specifically includes: Step 41: The fused noisy image input is segmented into two-dimensional patches, each patch being converted into a high-dimensional feature vector by a fully connected layer; different scanning modes are used on different model blocks to ensure that each patch has a global receptive field. The scanning modes include: in the first block, row-first scanning is used, scanning the image patch sequence line by line, scanning horizontally from left to right in each line, and then moving to the next line; in the second block, the sequence order is reversed and the sequence is scanned in the same way; in subsequent data blocks, column-first scanning is performed in both forward and reverse order; after traversing all scanning modes, the process is repeated in the next block. Step 42: The state-space model is used to process a noisy two-dimensional input containing a time step and class conditions; the noisy input is considered a clean signal perturbed by Gaussian noise corresponding to the input time step; the noisy input is first segmented into two-dimensional patches, each patch being transformed into a high-dimensional feature vector by a fully connected layer; these patches are fed into a 3×3 deep convolutional layer, where local information is injected into the patches; these patches are also padded with learnable labels at the end of rows and columns, enabling the model to understand its two-dimensional spatial structure during a one-dimensional sequential scan; using one of the scan patterns described above, the patch labels are flattened into a patch sequence; the time step and class conditions are also transformed into labels by a fully connected layer and then appended to the sequence; the state-space model is designed to encode and decode the one-dimensional sequential input; in the continuous-time SSM, the input signal x(i) is first encoded into a hidden state vector h(i) and then decoded into an output signal y(i); according to the following ordinary differential equation: h′(i)=Ah(i)+Bx(i), y(i)=Ch(i)+Dx(i) h′(i) represents the derivative of h, and B, C, and D represent the weights of the SSM. The inputs to natural language and 2D vision are discrete signals, therefore Mamba uses the zero-order preservation rule for discretization. The above ordinary differential equations are solved repeatedly: A=exp(ΔA),B=(ΔA) -1 (exp(ΔA)-I)·ΔB, h i =Ah i-1 +Bx i ,y i =Ch i +Dx i The sequence is then fed into a Mamba block for scanning, ultimately outputting high-quality bone scan images.

7. A method for rapid bone scan image noise reduction according to claim 6, characterized in that, The method further includes: Step 5: At the end of each training cycle, evaluate the quality of the generated images using quantitative metrics and qualitative analysis; perform early stopping on the model to avoid overfitting, and select the best-performing model on the validation set for verification; Step 6: Deploy the trained model in a real-world fast bone scan image denoising process, evaluate the model's effectiveness and user satisfaction in a real clinical setting, and ensure that the model can improve scanning efficiency while maintaining or enhancing diagnostic quality. Step 7: Conduct extensive stability and robustness tests on the model, including performance evaluation under different noise levels, image quality, and device differences, to ensure that the model can maintain stable noise reduction performance in diverse clinical environments.

8. A rapid bone scan image noise reduction system, characterized in that, The system, applied to a rapid bone scan image noise reduction method as described in any one of claims 1-7, comprises: The image acquisition and preprocessing module acquires noisy, rapid bone scan raw images and performs preprocessing, including size normalization, pixel value normalization, and contrast adjustment. The building module is used to construct a non-diffusion module with a periodic consistent architecture, including two generators and two discriminators. One generator is used to convert a preprocessed noisy image into a denoised image, and the other generator is used to inversely convert the denoised image into a noisy image. The two discriminators are used to perform adversarial training on the images generated by the two generators, respectively. The noise-adding module is used to progressively add noise to a conventional bone scan image to generate a pure Gaussian noise image. The progressive noise-adding process is based on the forward diffusion process of a Markov chain. The fusion output module is used to fuse the generated pure Gaussian noise image with the obtained fast bone scan image, and to perform stepwise denoising on the fused image through a state space model. The denoising process is based on the back diffusion of a Markov chain, and each back diffusion step is mapped through the state space model to output a high-quality conventional bone scan image.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a fast bone scan image noise reduction method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of a fast bone scan image noise reduction method as described in any one of claims 1-7.

Images