Low-dose CT denoising method and device based on wavelet transform and state space model

Abstract

The invention discloses a low-dose CT denoising method and device based on wavelet transform and a state space model, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring a low-dose CT image and a corresponding full-dose CT image, and generating a data set; dividing the data set into a training set and a test set according to a preset proportion; constructing and training a CT denoising network, taking the training set as an input parameter, taking the predicted image as an output parameter, and taking the minimum error absolute value of the full-dose CT image and the predicted image as a loss function; and inputting the test set into the trained CT denoising network in the step S3 to obtain a prediction image. According to the low-dose CT denoising method, global dependence can be efficiently established, and details can be accurately reserved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a low-dose CT denoising method and device based on wavelet transform and state-space model. Background Technology

[0002] In clinical diagnosis, high-quality CT images are crucial for disease assessment. However, traditional methods of acquiring high-quality CT images require high doses of X-rays, and high-dose ionizing radiation can pose potential health risks. To reduce radiation risks, low-dose CT (LDCT) is widely used in clinical screening. However, due to the reduced number of photons during scanning, low-dose CT images generate significant quantum noise, leading to decreased tissue density contrast and blurred anatomical structure edges, directly impacting the accuracy of doctors' diagnoses.

[0003] Existing low-dose CT denoising techniques are mainly based on convolutional neural networks (CNNs) and Transformers, but these all have the following drawbacks: Methods based on convolutional neural networks are limited by the local receptive field of the convolutional kernel, making it difficult to capture long-distance dependencies in images, resulting in insufficient global structural consistency of the denoised images. Although Transformer-based methods have a global receptive field, their computational complexity increases quadratically with image resolution, making them inefficient for processing high-resolution CT images. Existing low-dose CT denoising methods ignore the distribution characteristics of noise in the frequency domain, making it difficult to balance noise suppression with the preservation of minute texture details, and easily resulting in the loss of details during denoising. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the technical problem to be solved by this application is: how to achieve low-dose CT denoising that can both efficiently establish global dependencies and accurately preserve details.

[0005] To achieve the above objectives, in a first aspect, embodiments of this application provide a low-dose CT denoising method based on wavelet transform and a state-space model, the method comprising the following steps: Acquire low-dose CT images and their corresponding full-dose CT images to generate a dataset; The dataset is divided into training and testing sets according to a preset ratio; A CT denoising network was constructed and trained, with the training set as the input parameter and the predicted image as the output parameter. The loss function was the minimum absolute value of the error between the full-dose CT image and the predicted image. Input the test set into the CT denoising network trained in S3 to obtain the predicted image.

[0006] In conjunction with the first aspect, in one embodiment, the CT denoising network adopts a U-shaped encoder-decoder architecture, including a multi-scale dilated convolutional fusion module, a hybrid spatial state module, and a convolutional reconstruction layer.

[0007] In conjunction with the first aspect, in one embodiment, the process of the CT denoising network outputting the predicted image includes: After inputting the test set into the multi-scale dilated convolution fusion module, the image processed by the multi-scale dilated convolution fusion module is obtained; The image processed by the multi-scale dilated convolutional fusion module is input into an encoder consisting of four stacked hybrid spatial state space modules. The feature images acquired step by step are then imported into a decoder consisting of four stacked hybrid spatial state space modules. The first three encoded feature images acquired by the encoder are then concatenated with the first three decoded feature images acquired by the decoder. The last set of decoded feature images from the decoder is input into the convolutional reconstruction layer. After 3×3 convolutional reconstruction, the final denoised CT image, i.e., the predicted image, is obtained.

[0008] In conjunction with the first aspect, in one embodiment, the number of each of the four stacked hybrid space state space modules in the encoder is set to 4, 6, 6, and 8 respectively. The number of each of the four stacked hybrid spatial state space modules in the decoder is set to 6, 6, 4, and 2 respectively.

[0009] In conjunction with the first aspect, in one embodiment, the multi-scale dilated convolutional fusion module includes a wavelet decomposition unit, a multi-scale dilated feature extraction unit, and a feature fusion unit; The process of obtaining the image processed by the multi-scale dilated convolutional fusion module includes: The test set is input into the wavelet decomposition unit to obtain the low-frequency approximate component and three high-frequency detail components; After the low-frequency approximation component and the three high-frequency detail components are concatenated in the channel dimension, they are input into the multi-scale hole feature extraction unit to obtain the concatenated feature map. The multi-scale hole feature extraction unit sends the stitched feature map to the feature fusion unit to obtain the upsampled feature map.

[0010] In conjunction with the first aspect, in one embodiment, the hybrid spatial state module includes a spatial state unit and a dual-path hybrid feedforward unit; The hybrid spatial state space module is used to acquire encoded feature images and decoded feature images, and the processes for acquiring encoded feature images and decoding feature images are the same; wherein the process for acquiring encoded feature images includes: The image processed by the multi-scale dilated convolutional fusion module is input into the spatial state space unit to obtain the residual stitching feature map; The residual stitched feature map is input into the dual-path hybrid feedforward unit, which then obtains the encoded feature image based on the residual stitched feature map.

[0011] In conjunction with the first aspect, in one embodiment, the process of obtaining the residual stitching feature map by the spatial state spatial unit includes: After performing layer normalization on the image processed by the multi-scale dilated convolution fusion module, the layer normalization result is obtained. After the layer normalization result is processed sequentially by linear projection, depth convolution and SiLU activation function, it is input into the SSM core operator to obtain the SSM modeling result; in parallel, the layer normalization result is processed sequentially by linear projection and SiLU activation function to generate a gated signal; The SSM modeling result is multiplied with the gated signal and then linearly projected. The linear projection result is then added to the image processed by the multi-scale dilated convolution fusion module to obtain the residual stitching feature map.

[0012] In conjunction with the first aspect, in one embodiment, the process by which the dual-path hybrid feedforward unit obtains the final residual feature map based on the residual stitched feature map includes: After performing layer normalization, 1×1 convolution, and GELU activation on the residual splicing feature map, a split operation is performed to divide it into an upper path and a lower path in the channel dimension; the upper path is processed by a 3×3 depthwise separable convolution; the lower path is processed by a 3×3 standard convolution and GELU activation. The processing results of the upper and lower paths are concatenated, fused by 1×1 convolution, and then added to the residual concatenated feature map to obtain the encoded feature image.

[0013] In conjunction with the first aspect, in one embodiment, the loss function expression is: ; In the formula, The total number of pixels in a single image. Indicates the first full-dose CT image The actual grayscale value of each pixel. The denoised image predicted by the model represents the first... The grayscale value of each pixel. Represents the pixel index, with a value range of ; Secondly, this application provides a third aspect, namely, a low-dose CT denoising device based on wavelet transform and state space model. The low-dose CT denoising device based on wavelet transform and state space model includes a processor, a memory, and a low-dose CT denoising program based on wavelet transform and state space model stored in the memory and executable by the processor. When the low-dose CT denoising program based on wavelet transform and state space model is executed by the processor, it implements the method provided in the first aspect.

[0014] Compared with the prior art, the advantages of this application are: (1) By using a multi-scale dilated convolution fusion module to separate noise and structure in the frequency domain and combining it with a hybrid spatial state space module to enhance features in the spatial domain, the problem of traditional methods being unable to distinguish between high-frequency noise and texture details is effectively solved, and the effect of denoising without losing details can be achieved.

[0015] (2) By using the state space unit to establish long-distance dependencies between pixels through the state space model, the consistency of the global anatomical structure is ensured, and the block effect caused by local denoising is avoided. By using the dual-path hybrid feedforward unit to extract fine features and spatial context through the dual-path parallel structure, key diagnostic information such as lesion edge and microstructure is restored, and the degree of lesion visualization is improved.

[0016] (3) The linear complexity state space model is used to replace the traditional Transformer. While ensuring the denoising effect, the computational resource consumption is significantly reduced. It does not require multi-step iterative sampling, and the computational efficiency is higher, which can better meet the actual needs of rapid imaging and batch processing of low-dose CT in clinical practice. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the low-dose CT noise reduction method in the embodiments of this application; Figure 2 This is a schematic diagram of the architecture of the CT denoising network in this application; Figure 3 This is a schematic diagram of the architecture of the multi-scale dilated convolutional fusion module in this application; Figure 4 This is a schematic diagram of the architecture of the hybrid space state space module of this application; Figure 5This is a schematic diagram of the hardware structure of a low-dose CT denoising device based on wavelet transform and state-space model involved in the embodiments of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] In a first aspect, embodiments of this application provide a low-dose CT denoising method based on wavelet transform and state-space model, referring to... Figure 1 The method includes the following steps: S1. Acquire low-dose CT images (LDCT) and their corresponding full-dose CT images (NDCT). After constructing the original image dataset X from the low-dose CT images and the label dataset Y from the full-dose CT images, normalize the original CT images to generate the dataset. S2. Divide the dataset into training and testing sets according to a preset ratio (e.g., 7:3); S3. Construct and train the CT denoising network (WSSM-Net), using the training set as the input parameter, the predicted image as the output parameter, and the minimum absolute value of the error between the full-dose CT image and the predicted image as the loss function. S4. Input the test set into the CT denoising network trained in S3, obtain the predicted image, and evaluate the accuracy of the CT denoising network.

[0023] By inputting low-dose CT images into the trained CT denoising network, denoising of low-dose CT images can be achieved.

[0024] In one embodiment, the above normalization process includes: ; In the formula, The normalized HU value, The HU (Hounsfield) value of the original CT image. and These are the minimum and maximum values ​​of the set cutoff window (e.g., -1000HU to 1000HU).

[0025] Furthermore, after normalization, data augmentation (such as random pruning and flipping) is performed to improve data quality and model generalization performance.

[0026] In one embodiment, the CT denoising network in the above method adopts a U-shaped encoder-decoder architecture, specifically including: The input consists of a multi-scale dilated convolutional fusion module (W-MDCF) based on discrete wavelet transform, a hybrid spatial state space module (HSSM) serving as the network backbone, and a convolutional reconstruction layer at the end.

[0027] Based on this, refer to Figure 2 , Figure 2 This is a schematic diagram of the architecture of the CT denoising network in this application. Figure 2 As shown, the process of the CT denoising network outputting the predicted image includes: The input image (taken from the training set or test set) is input into the multi-scale dilated convolutional fusion module, and the image processed by the multi-scale dilated convolutional fusion module is obtained. The image input after being processed by the multi-scale dilated convolutional fusion module is fed into an encoder consisting of four stacked hybrid spatial state space modules (the number of each stacked hybrid spatial state space module is set to 4, 6, 6, and 8 sequentially according to data transmission guidance). After gradually restoring the image resolution, the progressively acquired feature images are imported into a decoder consisting of four stacked hybrid spatial state space modules (the number of each stacked hybrid spatial state space module is set to 6, 6, 4, and 2 sequentially according to data transmission guidance). The first three encoded feature images acquired by the encoder are then concatenated with the first three decoded feature images acquired by the decoder. The last set of decoded feature images from the decoder is input into the convolutional reconstruction layer. After 3×3 convolutional reconstruction, the multi-channel features are mapped back to the pixel space of a single channel to obtain the final denoised CT image, i.e., the predicted image.

[0028] In one embodiment, the multi-scale dilated convolutional fusion module includes a wavelet decomposition unit, a multi-scale dilated feature extraction unit, and a feature fusion unit.

[0029] Based on this, refer to Figure 3 , Figure 3 This is a schematic diagram of the architecture of the multi-scale dilated convolutional fusion module in this application. Figure 3 As shown, the process of obtaining the image after processing by the multi-scale dilated convolutional fusion module includes: S1. The input image is input to the wavelet decomposition unit, which obtains the low-frequency approximation component and three high-frequency detail components. Specifically, the wavelet decomposition unit performs two-dimensional discrete Haar wavelet transform (DWT) on the input image, decomposing the input image into the low-frequency approximation component Y. L and three high-frequency detail components Y HL Y LH Y HH ; For image coordinates (i, j), the Haar wavelet decomposition is calculated as follows: ; ; ; ; In the formula, Represents the spatial coordinate index of the image. Indicates row index, This represents the column index, used to locate pixel positions in the image. It represents the low-frequency approximate component, reflecting the overall contour and structural features of the image. It represents the horizontal high-frequency detail component, reflecting the high-frequency details of the image in the horizontal direction, such as edges and textures. It represents the high-frequency detail component in the vertical direction, reflecting the high-frequency details of the image such as edges and textures in the vertical direction. It represents the diagonal high-frequency detail component, reflecting the high-frequency details of the image such as edges and textures in the diagonal direction. Represents the spatial coordinates of the input image The pixel intensity value at that location. This represents the normalization coefficient, used to balance the energy of the wavelet transform, avoid the transformed value range being too large or too small, and ensure the reversibility of the transform. S2, the low-frequency approximation component and the three high-frequency detail components are concatenated in the channel dimension and then input into the multi-scale dilated feature extraction unit to obtain the concatenated feature map; the specific process of the multi-scale dilated feature extraction unit to obtain the concatenated feature map is as follows: the low-frequency approximation component and the three high-frequency detail components concatenated in the channel dimension are sent to four parallel convolution branches, and the output results of the four convolution branches are concatenated to obtain the concatenated feature map; The four convolutional branches include a 1×1 convolution (dilation rate D=1) and three 3×3 convolutions (dilation rates D=2, 4, and 6 respectively) to capture contextual information at different scales. S3. The multi-scale dilated feature extraction unit sends the stitched feature map to the feature fusion unit to obtain the upsampled feature map. The specific process for the feature fusion unit to obtain the upsampled feature map is as follows: the stitched feature map is input into the feature fusion unit and sequentially processed by a 3×3 convolution and a ReLU activation function. After processing and upsampling, the image is restored to its original resolution, and the image processed by the multi-scale dilated convolutional fusion module is obtained.

[0030] In one embodiment, the Hybrid Spatial State Space Module (HSSM) includes a Spatial State Space Element (SSM) and a Dual Hybrid Feedforward Element (DPHFN).

[0031] Based on this, refer to Figure 4 , Figure 4 This is a schematic diagram of the architecture of the hybrid space state space module in this application. Figure 4 As shown, the process of acquiring encoded feature images and decoding feature images in this hybrid spatial state space module is the same. Taking the acquisition of encoded feature images as an example, the process of acquiring encoded feature images includes: S1. The image processed by the multi-scale dilated convolutional fusion module is input into the spatial state spatial unit for global feature extraction to obtain the residual stitched feature map; the process of obtaining the residual stitched feature map by the spatial state spatial unit includes: The image processed by the multi-scale dilated convolutional fusion module is then subjected to layer normalization (LayerNorm). The expression for this layer normalization is: ; In the formula, To standardize the output results, For the input sample, The mean of the currently processed samples Represents variance. It is a constant (take it as small as possible to avoid a denominator of 0). and These are the scaling factor and the translation factor, respectively, both of which are learnable parameters; The standardized result is then sequentially processed through linear projection, depthwise convolution (DW Conv), and the SiLU activation function. After processing, the SSM core operator is input to obtain the SSM modeling results; in parallel, the standardized processing results are processed sequentially through linear projection and SiLU activation functions to generate a gated signal; The SSM modeling result is multiplied with the gated signal and then linearly projected. The linear projection result is then added to the image processed by the multi-scale dilated convolution fusion module to obtain the residual stitching feature map. The SSM core operator is based on continuous-time state equations: ; In the formula, Hidden state vector Regarding time The derivative of represents the rate of change of the state. This represents the state transition matrix, which describes the dynamic evolution of the state itself. The hidden state vector in continuous time represents the system's state over time. The internal state of. Represents the input matrix, which takes the external input... Mapped to the state space. This represents the input vector in continuous time, corresponding to the model's time... Input features. The output vector represents the output vector in continuous time, which is determined by the hidden state. Obtained by linear transformation. The output matrix represents the hidden state. Mapped to the output space; Discretize using the zero-order hold (ZOH) principle as follows: ; ; In the formula, Represents discrete time steps The hidden state vector. This represents the state transition matrix after ZOH discretization, which is derived from the continuous-time matrix. It is derived. Indicates the previous time step The hidden state vector. This represents the input matrix after ZOH discretization, which is a continuous-time matrix. It is derived. Represents discrete time steps The input vector. Represents discrete time steps The output vector. The output matrix is ​​represented by the one in the continuous-time equation. Consistent. This indicates the time step of the discretization. This represents the state transition matrix in the continuous-time state equations, describing the state of the continuous system. The dynamic evolution law. The matrix exponentiation operation is the core mathematical operation for converting a continuous-time linear system into a discrete-time system. S2. Input the residual stitched feature map into the dual-path hybrid feedforward unit. The dual-path hybrid feedforward unit obtains the encoded feature image based on the residual stitched feature map. The specific process includes: After performing layer normalization (LayerNorm), 1×1 convolution, and GELU activation on the residual splicing feature map, a split operation is performed to divide it into an upper path and a lower path in the channel dimension. The upper path is processed by a 3×3 depthwise separable convolution (DW Conv) to capture the spatial context; the lower path is processed by a 3×3 standard convolution and GELU activation. The processing results of the upper and lower paths are concatenated (Concat), fused by 1×1 convolution, and then added to the residual concatenated feature map to obtain the encoded feature image; In one embodiment, the loss function in the above method The expression is: ; In the formula, The total number of pixels in a single image. Indicates the first full-dose CT image The actual grayscale value of each pixel. The denoised image predicted by the model represents the first... The grayscale value of each pixel. Represents the pixel index, with a value range of ; In one embodiment, the evaluation metrics for assessing the accuracy of the CT denoising network include peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean square error (RMSE). A higher PSNR indicates better accuracy of the CT denoising network, a SSIM closer to 1 indicates better preservation of the CT denoising network structure, and a smaller RMSE indicates that the denoising result is closer to the reference full-dose CT image.

[0032] The peak signal-to-noise ratio (PSNR) mentioned above is used to measure image fidelity. If the image is normalized to the [0,1] interval, its calculation formula is: ; In the formula, The mean square error is calculated as follows: ; In the formula, This represents the height of the image (number of rows), which is the total number of pixels in the vertical direction. The width of the image (number of columns), which is the total number of pixels in the horizontal direction. Represents the row index of the image, traversing from 1 to... All lines. The column index representing the image, traversing from 1 to... All columns. Represents the coordinates of a true full-dose CT image. The pixel intensity value at that location. This indicates the model's predicted denoised image in coordinates. The pixel intensity value at that location.

[0033] The Structural Similarity Index (SSIM) is used to measure the consistency of structural information, and its calculation formula is as follows: ; In the formula, and They are respectively X and Y The average value, and They are respectively X and Y Image variance for X and Y covariance, and It is a constant. ; Root mean square error (RMSE) measures the overall pixel deviation, and its calculation formula is: ; In the formula, This represents the height of the image (number of rows), which is the total number of pixels in the vertical direction. This represents the width of the image (number of columns), which is the total number of pixels in the horizontal direction. Represents the row index of the image, traversing from 1 to... All lines. The column index representing the image, traversing from 1 to... All columns. Represents the coordinates of a true full-dose CT image. The pixel intensity value at that location. This indicates the model's predicted denoised image in coordinates. Pixel intensity value at; The effectiveness of the method provided in this application will be verified below using specific experimental data.

[0034] Data with a slice thickness of 3 mm was selected from the publicly available dataset provided by the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge, totaling 2378 pairs of CT images. 211 pairs of images from patient L506 were used as the test set, and 2167 pairs of images from the remaining 9 patients were used as the training set. All images were 256×256 pixels in size. To enhance the diversity of training samples, the 256×256 CT images were randomly and adaptively cropped into 48×48 patches for training. The target NDCT images were cropped in the same way.

[0035] On an Ubuntu 22.04 system, a CT denoising network was built using Python 3.10 and the PyTorch 2.2.0 framework. Training and testing of the CT denoising network were completed on an NVIDIA RTX 4090 GPU (24GB VRAM). The Adam optimizer was used during training, with a learning rate of 5e-5 and a momentum parameter... =0.5、 =0.999, Quantitative comparative experiments were conducted using four mainstream algorithms: RED-CNN, EDCNN, WGAN-VGG, and CoreDiff, with a learning rate of 1×10⁻⁶. -4 The batch size is 1, and the training lasts for 100 epochs until convergence.

[0036] The comparison results are shown in Table 1:

[0037] Table 1 As shown in Table 1, the WSSM-Net proposed in this invention outperforms the comparative methods in terms of peak signal-to-noise ratio, structural similarity index, and root mean square error. Specifically, WSSM-Net achieves a PSNR of 34.1562dB, a significant improvement over the classic RED-CNN (31.5624dB) and EDCNN (32.1487dB), validating the advantages of state-space models in overcoming the limitations of convolutional local receptive fields and capturing long-range dependencies. Compared to the generative adversarial network WGAN-VGG (32.8951dB), this method achieves a SSIM score of 0.9038, demonstrating superior anatomical structure preservation and effectively avoiding texture distortion that may occur with generative models. Compared to the current state-of-the-art but computationally expensive diffusion model CoreDiff (34.0273dB), this method achieves the same level of denoising performance (slightly better PSNR by 0.13dB and SSIM by 0.0023), while possessing higher computational efficiency due to the linear computational complexity of the state-space model. It eliminates the need for multi-step iterative sampling and better meets the needs of rapid low-dose CT imaging in clinical settings.

[0038] Secondly, embodiments of this application provide a low-dose CT denoising device based on wavelet transform and state space model. The low-dose CT denoising device based on wavelet transform and state space model can be a personal computer (PC), laptop computer, server or other device with data processing capabilities.

[0039] Reference Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of a low-dose CT denoising device based on wavelet transform and state-space model involved in the embodiments of this application. In the embodiments of this application, the low-dose CT denoising device based on wavelet transform and state-space model may include a processor, a memory, a communication interface, and a communication bus.

[0040] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0041] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces enable interconnection of devices within the low-dose CT denoising equipment based on wavelet transform and state-space models, and also enable interconnection between the low-dose CT denoising equipment and other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0042] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0043] The processor can be a general-purpose processor, which can call a low-dose CT denoising program based on wavelet transform and state space model stored in memory and execute the low-dose CT denoising method based on wavelet transform and state space model provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the low-dose CT denoising program based on wavelet transform and state space model is called can refer to the various embodiments of the low-dose CT denoising method based on wavelet transform and state space model of this application, and will not be repeated here.

[0044] Those skilled in the art will understand that Figure 5 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0045] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0046] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0047] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0048] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0049] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0050] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0051] The above are merely specific embodiments of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the scope of the claims.

Claims

1. A low-dose CT denoising method based on wavelet transform and state-space model, characterized in that, The method includes the following steps: Acquire low-dose CT images and their corresponding full-dose CT images to generate a dataset; The dataset is divided into training and testing sets according to a preset ratio; A CT denoising network was constructed and trained, with the training set as the input parameter and the predicted image as the output parameter. The loss function was the minimum absolute value of the error between the full-dose CT image and the predicted image. Input the test set into the CT denoising network trained in S3 to obtain the predicted image.

2. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 1, characterized in that, The CT denoising network adopts a U-shaped encoder-decoder architecture, including a multi-scale dilated convolutional fusion module, a hybrid spatial state module, and a convolutional reconstruction layer.

3. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 2, characterized in that, The process of the CT denoising network outputting the predicted image includes: After inputting the test set into the multi-scale dilated convolution fusion module, the image processed by the multi-scale dilated convolution fusion module is obtained; The image processed by the multi-scale dilated convolutional fusion module is input into an encoder consisting of four stacked hybrid spatial state space modules. The feature images acquired step by step are then imported into a decoder consisting of four stacked hybrid spatial state space modules. The first three encoded feature images acquired by the encoder are then concatenated with the first three decoded feature images acquired by the decoder. The last set of decoded feature images from the decoder is input into the convolutional reconstruction layer. After 3×3 convolutional reconstruction, the final denoised CT image, i.e., the predicted image, is obtained.

4. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 3, characterized in that, The number of each of the four stacked hybrid spatial state space modules in the encoder is set to 4, 6, 6, and 8 respectively. The number of each of the four stacked hybrid spatial state space modules in the decoder is set to 6, 6, 4, and 2 respectively.

5. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 3, characterized in that, The multi-scale dilated convolutional fusion module includes a wavelet decomposition unit, a multi-scale dilated feature extraction unit, and a feature fusion unit. The process of obtaining the image processed by the multi-scale dilated convolutional fusion module includes: The test set is input into the wavelet decomposition unit to obtain the low-frequency approximate component and three high-frequency detail components; After the low-frequency approximation component and the three high-frequency detail components are concatenated in the channel dimension, they are input into the multi-scale hole feature extraction unit to obtain the concatenated feature map. The multi-scale hole feature extraction unit sends the stitched feature map to the feature fusion unit to obtain the upsampled feature map.

6. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 3, characterized in that, The hybrid spatial state module includes a spatial state unit and a dual-path hybrid feedforward unit; The hybrid spatial state space module is used to acquire encoded feature images and decoded feature images, and the processes for acquiring encoded feature images and acquiring decoded feature images are the same. The process of obtaining the encoded feature image includes: The image processed by the multi-scale dilated convolutional fusion module is input into the spatial state space unit to obtain the residual stitching feature map; The residual stitched feature map is input into the dual-path hybrid feedforward unit, which then obtains the encoded feature image based on the residual stitched feature map.

7. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 6, characterized in that, The process of obtaining the residual stitching feature map from the spatial state spatial unit includes: After performing layer normalization on the image processed by the multi-scale dilated convolution fusion module, the layer normalization result is obtained. After the layer normalization result is processed sequentially by linear projection, depth convolution and SiLU activation function, it is input into the SSM core operator to obtain the SSM modeling result; in parallel, the layer normalization result is processed sequentially by linear projection and SiLU activation function to generate a gated signal; The SSM modeling result is multiplied with the gated signal and then linearly projected. The linear projection result is then added to the image processed by the multi-scale dilated convolution fusion module to obtain the residual stitching feature map.

8. The low-dose CT denoising method based on wavelet transform and state-space model according to claim 7, characterized in that, The process by which the dual-path hybrid feedforward unit obtains the final residual feature map based on the residual stitched feature map includes: After performing layer normalization, 1×1 convolution, and GELU activation on the residual splicing feature map, a split operation is performed to divide it into an upper path and a lower path in the channel dimension; the upper path is processed by a 3×3 depthwise separable convolution; the lower path is processed by a 3×3 standard convolution and GELU activation. The processing results of the upper and lower paths are concatenated, fused by 1×1 convolution, and then added to the residual concatenated feature map to obtain the encoded feature image.

9. A low-dose CT denoising method based on wavelet transform and state-space model according to claim 1, characterized in that, The loss function expression is as follows: ； In the formula, The total number of pixels in a single image. Indicates the first full-dose CT image The actual grayscale value of each pixel. The denoised image predicted by the model represents the first... The grayscale value of each pixel. Represents the pixel index, with a value range of ; 10. A low-dose CT denoising device based on wavelet transform and state-space model, characterized in that, The low-dose CT denoising device based on wavelet transform and state space model includes a processor, a memory, and a low-dose CT denoising program based on wavelet transform and state space model stored in the memory and executable by the processor. When the low-dose CT denoising program based on wavelet transform and state space model is executed by the processor, it implements the steps of the low-dose CT denoising method based on wavelet transform and state space model as described in any one of claims 1 to 9.

Images