Low-dose CT image denoising method and system based on conditional diffusion model

Abstract

The invention discloses a low-dose CT image denoising method and system based on a conditional diffusion model, and the method comprises the steps: constructing a denoising network model, only employing a standard-dose CT image in a training process, and simulating a generation process of the standard-dose CT image through the training of the denoising network model. And abundant image priori knowledge is constructed. Meanwhile, the de-noising network model uses a U-Net encoder-decoder structure, a multi-scale cross axis attention mechanism is added in a decoder and an encoder, important image details are reserved in the de-noising process, irrelevant noise components are removed at the same time, and smooth areas and continuous boundaries of the image can be recovered. In addition, the frequency compensation block in the jump connection can compensate the insufficiency of medium-high frequency signals in the training process, so that the original structure and texture information of the image are recovered.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image denoising, and in particular relates to a low-dose CT image denoising method and system based on a conditional diffusion model. Background Art

[0002] In the field of medical imaging, computed tomography (CT) is an important diagnostic tool, and its image quality directly impacts the accuracy of disease diagnosis. However, while CT scans provide high-resolution images, they are also associated with high radiation doses, which can pose a potential risk to patient health over the long term. To mitigate this risk, low-dose CT imaging technology has emerged, which reduces radiation dose and thus potential harm to patients. However, the attendant challenge is that low-dose CT images have a lower signal-to-noise ratio than standard-dose CT images, resulting in a decrease in image quality, particularly in terms of image contrast and clarity. This noise and artifacts can affect doctors' judgment of the images, leading to diagnostic uncertainty.

[0003] In recent years, extensive research has been conducted on the denoising of low-dose CT images. These methods can be roughly divided into three categories: sinusoidal filtering, iterative reconstruction, and image post-processing. Sinusoidal filtering performs denoising in the projection domain and reconstructs CT images using filtered back-projection techniques. Typical methods include structure-adaptive filtering and bilateral filtering. Iterative reconstruction constructs an objective function based on the noise statistical model of the projection data and the image prior distribution, and obtains the reconstructed image through optimization methods. Methods involved include dictionary learning, total variation, and non-local mean priors. Sinusoidal filtering and iterative reconstruction methods rely on projection data. Image post-processing methods, such as non-local mean filtering and block matching algorithms, act directly on CT images and do not require the original projection data. Compared with the first two categories of methods, image post-processing methods have faster reconstruction speeds.

[0004] At the same time, deep learning technology has made significant progress in the field of image denoising. Convolutional neural networks (CNNs), with their superior feature extraction capabilities, can effectively remove noise by learning noise patterns and image features. Residual networks offer new insights into image denoising, directly transferring and preserving information through cross-layer connections, allowing the network to stack layers to enhance feature expression. Diffusion models, as a new generative model, simulate the diffusion process of data and learn its distributional characteristics, thereby removing noise while preserving image detail as much as possible. Compared to traditional denoising techniques, methods based on diffusion models demonstrate significant advantages in handling complex noise and preserving image detail.

[0005] However, practical applications still face a number of challenges and issues: 1. Diffusion models typically require a large number of iterations to generate images, making training and inference time-consuming. 2. Generating high-quality denoised images directly from noise is difficult because of the need to capture and model the complex details of the image. 3. In clinical settings, obtaining paired CT images (i.e., low-dose and standard-dose CT images of the same patient) for model training is difficult. Summary of the Invention

[0006] The purpose of the present invention is to provide a low-dose CT image denoising method and system based on a conditional diffusion model.

[0007] In a first aspect, the present invention provides a low-dose CT image denoising method based on a conditional diffusion model, the method comprising: Construct a dataset containing standard-dose computed tomography images; use the standard-dose computed tomography images in the dataset as initial images, process the initial images through a diffusion process, and obtain a noisy image corresponding to each time step; add the noisy image and the corresponding time step to the dataset; Construct a denoising network model in reverse diffusion; the input of the denoising network model is the original feature map after the initial image and the noisy image are spliced ​​together, and the corresponding time step t ; The denoising network model was trained using the dataset, and the trained denoising network model was used to denoise low-dose computed tomography images.

[0008] Preferably, the denoising network model uses the U-Net network as the basic framework, which includes a position encoding module, an encoder, a decoder and a bottleneck layer; the position encoding module is used to convert the time step into a position vector and embed it into the original feature map to obtain an embedded image; the encoder includes multiple encoding layers, and the decoder includes multiple decoding layers; except for the last encoding layer, the output features of the remaining encoding layers are input to the next encoding layer after downsampling processing; the output features of the last encoding layer are input to the bottleneck layer. Except for the first decoding layer, the input features of the remaining decoding layers are the splicing results of the output features of the previous decoding layer after upsampling processing and the output features of the same encoding layer after frequency compensation block processing; the input features of the first decoding layer are the splicing results of the output features of the bottleneck layer and the output features of the same encoding layer; Preferably, both the encoding layer and the decoding layer adopt feature extraction modules; the bottleneck layer includes two feature extraction modules connected in sequence.

[0009] As an advantage, the feature extraction module includes a first sub-feature extraction module, a feature affine transformation block, a second sub-feature extraction module and a multi-scale cross-axis attention module connected in sequence; the first sub-feature extraction module and the second sub-feature extraction module have the same structure, and both perform group normalization, activation function and convolution operations on the input features in sequence to obtain output features. The feature affine transformation block is used to scale and translate the input features according to the learned weights and biases. The input features of the first sub-feature extraction module are processed by the convolution layer and then spliced ​​with the output features of the second sub-feature extraction module in the channel dimension. The spliced ​​results are normalized by the layers and input into the multi-scale cross-axis attention module.

[0010] Preferably, the multi-scale cross-axis attention module includes parallel horizontal branches and vertical branches; the horizontal branches and the vertical branches have the same structure, and both encode the input features respectively through three convolution layers, and perform convolution operations on the encoded results to obtain three intermediate features; the three intermediate features are added, and the results are rearranged through a dimension reorganization function to obtain a value matrix, a key matrix and a query matrix; the key matrix, the value matrix and the query matrix are normalized, and the query matrix and the key matrix are subjected to matrix multiplication to obtain attention weights, which are processed by an activation function and summed using the attention weights and the value matrix to obtain horizontal attention out1 and vertical attention out2; the horizontal and vertical attention are subjected to convolution operations, and the processed attention in the two directions is compared with the original feature map x Add them together to get the feature map output by the multi-scale cross-axis attention module.

[0011] Preferably, the three convolutional layers in the horizontal branch correspond one-to-one to the three convolutional layers in the vertical branch, and the corresponding convolutional layers have opposite processing directions.

[0012] Preferably, the frequency compensation block adopts different Gaussian blurs to obtain features of different ranges of the feature map of the input frequency compensation block, and generates blurred feature maps of different degrees; the residuals between the feature map of the input frequency compensation block and each blurred feature map are obtained respectively, and all the residuals are fused with the feature map of the input frequency compensation block in the channel dimension, and the fused result is subjected to matrix multiplication operation with the parameter matrix set in the frequency compensation block to obtain the feature map output by the frequency compensation block.

[0013] In a second aspect, the present invention provides a low-dose CT image denoising system based on a conditional diffusion model, which is used to execute the above-mentioned low-dose CT image denoising method; the low-dose computed tomography image denoising system includes an image acquisition module, a preprocessing module, a forward processing module and an image denoising module; the image acquisition module is used to acquire standard-dose computed tomography images and low-dose computed tomography images; the preprocessing module is used to preprocess the computed tomography images acquired by the image acquisition module; the forward processing module is used to convert the standard-dose computed tomography images into random noise; and the image denoising module is used to perform a denoising process on the low-dose computed tomography images.

[0014] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the above-mentioned low-dose CT image denoising method based on the conditional diffusion model.

[0015] In a fourth aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the above-mentioned low-dose CT image denoising method based on the conditional diffusion model.

[0016] The present invention has the following beneficial effects: 1. The present invention constructs a dataset by constructing standard-dose CT images and splicing standard-dose CT images and noisy images as input features, thereby solving the problem of high cost of obtaining paired data of low-dose CT and standard-dose CT images in clinical settings. At the same time, the present invention uses standard-dose CT images as priors, thereby simplifying the learning process and providing a good foundation for the generation of high-quality denoised images, which helps to improve the stability and efficiency of training.

[0017] 2. The present invention fuses multi-scale features in each axial attention path by using strip convolution kernels of different kernel sizes to improve the efficiency of encoding spatial information; establishes dual cross attention between horizontal and vertical axial attention to better utilize multi-scale features and capture global information; the lightweight design of the multi-scale cross-axis attention module aims to reduce computational complexity, making it more suitable for practical application scenarios.

[0018] 3. By incorporating a multi-scale cross-axis attention mechanism into the decoder and encoder, this method preserves important image details during the denoising process while removing irrelevant noise components, helping to restore smooth regions and continuous boundaries in the image. Furthermore, by incorporating frequency compensation blocks into skip connections, it compensates for deficiencies in mid- and high-frequency signals during training, thereby restoring the image's original structure and texture information and preventing over-smoothing that can lead to loss of local information, thus facilitating clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is the overall flow chart of the present invention.

[0020] Figure 2 Schematic diagram of the denoising network model in the present invention.

[0021] Figure 3 Schematic diagram of the feature extraction module in the present invention.

[0022] Figure 4 Schematic diagram of the multi-scale cross-axis attention module in the present invention.

[0023] Figure 5 Schematic diagram of the frequency compensation block in the present invention.

[0024] Figure 6 Schematic diagram of low-dose computed tomography image.

[0025] Figure 7 Schematic diagram of low-dose computed tomography image after denoising. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings.

[0027] A low-dose CT image denoising method based on a conditional diffusion model. The low-dose CT image denoising system adopted includes an image acquisition module, a preprocessing module, a forward processing module and an image denoising module; the image acquisition module is used to acquire standard-dose CT images and low-dose CT images; the preprocessing module is used to preprocess the CT images acquired by the image acquisition module; the forward processing module is used to convert the standard-dose CT images into random noise; and the image denoising module is used to perform a denoising process on the low-dose CT images, thereby converting the low-dose CT images into standard-dose CT images.

[0028] like Figure 1 As shown, the low-dose CT image denoising method includes the following steps: Step 1: Build a dataset Obtain standard-dose CT images of multiple subjects at different locations and normalize each image to the range [0, 1]. Use the processed standard-dose CT images as the initial images to construct a dataset. Gradually add Gaussian noise to the initial images, with each step generating a noisy image that is only related to the previous step's result until a completely noisy image is obtained. The noisy images obtained at each time step are then normalized to the range [0, 1]. and the corresponding time steps are added to the dataset.

[0029] Step 2: Build a denoising network model. The goal of the denoising network model is to reconstruct the original data from the noisy data, which is a process of continuously removing noise. The denoising network model learns how to predict the noise added at each step, thereby gradually removing the noise and restoring the data. The input of the denoising network model is the conditional image (initial image). and noisy image Original feature map after splicing x and the corresponding time steps t The conditional image is used to guide the generation process as prior information for image generation to generate images that better meet specific requirements, reducing the randomness and uncontrollability of the generated images.

[0030] like Figure 2 As shown in Figure 1, the denoising network model adopts a U-Net encoder-decoder symmetrical structure, which includes a positional encoding module (PositionalEncoding), an encoder, a decoder, and a bottleneck layer. The positional encoding module converts the time step into t Convert to position vector and embed the position vector into the original feature map x In this way, the denoising network model can better understand the current noise level and image state in the inverse diffusion process, so as to more accurately predict and remove noise. The specific process is as follows: t Perform sine and cosine position encoding and calculate half the dimension of the position encoding. Since each position encoding will contain sine and cosine values, the dimension needs to be halved. Create a sequence of integer step lengths and normalize it to the range [0, 1]. Based on the set noise level, calculate the noise level encoding and obtain the sine and cosine values ​​based on the noise level encoding. Then use the time step embedding and the original feature map to x Directly concatenate them to get the embedded image.

[0031] The embedded image is segmented using the to_patches method: For the embedded image (B, C, H, W), the height H and width W are split into p×p patches, respectively. The segmented patches are reshaped (the reshaped patches are of size B, H / p, W / p, and C×p×p), and concatenated along the channel dimension. This improves computational efficiency and enhances the model's local perception. The segmented embedded image is downsampled from the input image (2×512×512) to 2×256×256 after convolution with a kernel of 3×3, a stride of 2, and padding of 1. This is then fed into the encoder.

[0032] The encoder consists of four sequentially connected encoding layers, with adjacent encoding layers connected by downsampling. The decoder consists of four sequentially connected decoding layers, with adjacent decoding layers connected by upsampling. Skip connections are used between encoding and decoding layers at the same level (same depth); a frequency compensation block is connected in series with the skip connections between the first three encoding layers and their corresponding decoding layers. With the exception of the last encoding layer, the output features of the remaining encoding layers are downsampled and then input to the next encoding layer. The output features of the last encoding layer are input to the bottleneck layer. With the exception of the first decoding layer, the input features of the remaining decoding layers are the concatenation of the upsampled output features of the previous decoding layer and the frequency compensation block-processed output features of the same encoding layer. The input features of the first decoding layer are the concatenation of the output features of the bottleneck layer and the output features of the same encoding layer.

[0033] like Figure 3 As shown in the figure, both the encoding layer and the decoding layer use feature extraction modules; the bottleneck layer includes two feature extraction modules connected in sequence. The feature extraction module includes a first sub-feature extraction module, a feature affine transformation block (Feature Wise Affine), a second sub-feature extraction module, and a multi-scale cross-axis attention module (MCA) connected in sequence; the first sub-feature extraction module and the second sub-feature extraction module have the same structure, and both group and normalize the input features through a group normalization layer (GroupNorm) to reduce internal covariate shift, then use the activation function Swish to increase nonlinearity, and then use a 3×3 convolution kernel and a convolution operation with padding of 1 to identify local features and obtain output features. , which is expressed as: in, Represents the convolution operation; represents the activation function; represents the group normalization layer; Represents the features of the input sub-feature extraction module.

[0034] In this embodiment, the size of the output feature of the second sub-feature extraction module is 32×256×256.

[0035] The feature affine transformation block is used to scale and translate the input features according to the learned weights and biases. The affine transformation enhances the expressiveness of the features, helps the model learn more complex data representations, and prevents the model from overfitting.

[0036] The input features of the first sub-feature extraction module are processed by a convolution layer with a convolution kernel of 1×1, and then spliced ​​with the output features of the second sub-feature extraction module in the channel dimension. The spliced ​​results are normalized by the layers and input into the multi-scale cross-axis attention module.

[0037] like Figure 4 As shown in the figure, the multi-scale cross-axis attention module consists of parallel horizontal and vertical branches. The horizontal and vertical branches have the same structure, encoding the input features through three convolutional layers and performing a 1×1 convolution on the encoded results to obtain three intermediate features. The convolution kernel sizes of the three convolutional layers in the horizontal branch are 1×7, 1×11, and 1×21, respectively, with padding of 3, 5, and 10, respectively; the convolution kernel sizes of the three convolutional layers in the vertical branch are 7×1, 11×1, and 21×1, respectively, with padding of 3, 5, and 10, respectively. The three intermediate features are added together and rearranged using the rearrange function (dimensionality reorganization function) to obtain the value matrix, key matrix, and query matrix, which are expressed as: in, Represent the key matrix, value matrix and query matrix in the horizontal direction respectively; Represent the key matrix, value matrix and query matrix in the vertical direction respectively.

[0038] Normalize the key matrix, value matrix, and query matrix, obtain the attention weight by matrix multiplication of the query matrix and key matrix, increase the nonlinear representation capability through the Softmax layer, and sum the attention weight and value matrix to obtain the horizontal attention out1 and the vertical attention out2, which are expressed as follows: The attention in the horizontal and vertical directions is processed through 1×1 convolution operation, and the attention after processing in the two directions is combined with the original feature map x Add them together to get the feature map output by the multi-scale cross-axis attention module.

[0039] like Figure 5 As shown in Figure 1, the frequency compensation block uses different Gaussian blurs (convolution kernels are 3×3, 5×5, 7×7, and standard deviations are 1, 2, and 4) to capture features in different ranges of the feature map of the input frequency compensation block and generate blurred feature maps x1, x2, and x3 with different degrees of blur, which are expressed as: Get the residual between the original feature map and the blurred feature map 、 and , which is expressed as: The residuals R1, R2, R3 and the original feature map are fused in the channel dimension. The fused result is multiplied by the parameter matrix learned in the frequency compensation block to obtain the feature map output by the frequency compensation block. , which is expressed as: in, Represents channel splicing operation; Represents a matrix multiplication operation; It is a learnable parameter matrix used to improve the model's ability to capture and process image features.

[0040] Step 3: Training the model In order to make the denoising network model have a better effect, construct a loss function To optimize the denoising network model to further improve the performance of the model. Use the mean square error (MSE) to converge the error to the minimum value, the loss function The expression is: in, This is a standard-dose CT image; is the predicted denoised image.

[0041] This implementation uses the Adamw optimizer for training and updating network weights. Decoupling weight decay from gradient updates avoids incorrect regularization in the Adam optimizer. Adding an additional L2 regularization term to implement weight decay, rather than applying it directly to gradient updates, effectively reduces the weight parameter value and improves the model's generalization. By revising the weight decay implementation and combining it with an adaptive learning rate and momentum mechanism, the model is optimized by continuously adjusting the learning rate and adjustable hyperparameters. The denoising performance of the trained model is evaluated across multiple dimensions. The denoising model parameters that perform best in the training set are retained.

[0042] Step 4: Denoise the low-dose CT images; use the trained denoising network model to Figure 6 The low-dose CT image shown in the figure is denoised to generate the following Figure 7 The CT image shown has good denoising effect.

[0043] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The scope of protection of the present invention should not be regarded as limited to the specific forms described in the embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A low-dose CT image denoising method based on a conditional diffusion model, characterized by: The method includes: Construct a dataset containing standard-dose computed tomography images; use the standard-dose computed tomography images in the dataset as initial images, process the initial images through a diffusion process, and obtain a noisy image corresponding to each time step; add the noisy image and the corresponding time step to the dataset; Construct a denoising network model in reverse diffusion; the input of the denoising network model is the original feature map after the initial image and the noisy image are spliced ​​together, and the corresponding time step t ; The denoising network model was trained using the dataset, and the trained denoising network model was used to denoise low-dose computed tomography images.

2. The low-dose CT image denoising method based on the conditional diffusion model according to claim 1, characterized in that: The denoising network model uses the U-Net network as its basic framework, which includes a position encoding module, an encoder, a decoder, and a bottleneck layer. The position encoding module is used to convert time steps into position vectors and embed them into the original feature map to obtain an embedded image. The encoder includes multiple encoding layers, and the decoder includes multiple decoding layers. In the encoder, except for the last coding layer, the output features of the remaining coding layers are downsampled and input to the next coding layer; the output features of the last coding layer are input to the bottleneck layer; In the decoder, except for the first decoding layer, the input features of the remaining decoding layers are the concatenation of the output features of the previous decoding layer after upsampling processing and the output features of the same coding layer after frequency compensation block processing; the input features of the first decoding layer are the concatenation of the output features of the bottleneck layer and the output features of the same coding layer.

3. The low-dose CT image denoising method based on the conditional diffusion model according to claim 2, characterized in that: The encoding layer and the decoding layer both adopt feature extraction modules; the bottleneck layer includes two feature extraction modules connected in sequence.

4. The low-dose CT image denoising method based on the conditional diffusion model according to claim 3, characterized in that: The feature extraction module includes a first sub-feature extraction module, a feature affine transformation block, a second sub-feature extraction module and a multi-scale cross-axis attention module connected in sequence; the first sub-feature extraction module and the second sub-feature extraction module have the same structure, and both perform group normalization, activation function and convolution operations on the input features in sequence to obtain output features. The feature affine transformation block is used to scale and translate the input features according to the learned weights and biases. The input features of the first sub-feature extraction module are processed by the convolution layer and then spliced ​​with the output features of the second sub-feature extraction module in the channel dimension. The spliced ​​results are normalized by the layers and input into the multi-scale cross-axis attention module.

5. The low-dose CT image denoising method based on the conditional diffusion model according to claim 4, characterized in that: The multi-scale cross-axis attention module includes parallel horizontal and vertical branches. The horizontal and vertical branches have the same structure, and both encode input features through three convolutional layers, and perform convolution operations on the encoded results to obtain three intermediate features. The three intermediate features are added and the result of the addition is rearranged through a dimension reorganization function to obtain a value matrix, a key matrix, and a query matrix. Normalize the key matrix, value matrix, and query matrix, obtain the attention weight by matrix multiplication of the query matrix and key matrix, process them through the activation function, and sum them with the attention weight and value matrix to obtain the horizontal attention out1 and the vertical attention out2; The attention in the horizontal and vertical directions is processed through convolution operations, and the attention after processing in the two directions is combined with the original feature map x Add them together to get the feature map output by the multi-scale cross-axis attention module.

6. The low-dose CT image denoising method based on the conditional diffusion model according to claim 5, characterized in that: The three convolutional layers in the horizontal branch correspond one-to-one to the three convolutional layers in the vertical branch, and the corresponding convolutional layers have opposite processing directions.

7. The low-dose CT image denoising method based on the conditional diffusion model according to claim 2, characterized in that: The frequency compensation block uses different Gaussian blurs to obtain features in different ranges of the feature map of the input frequency compensation block, generating feature maps with different degrees of blur; respectively obtains the residuals between the feature map of the input frequency compensation block and each blurred feature map, and fuses all the residuals with the feature map of the input frequency compensation block in the channel dimension, and performs a matrix multiplication operation on the fused result and the parameter matrix set in the frequency compensation block to obtain the feature map output by the frequency compensation block.

8. A low-dose CT image denoising system based on a conditional diffusion model, characterized by: Used to execute the low-dose CT image denoising method based on the conditional diffusion model according to claim 1; the low-dose computed tomography image denoising system comprises an image acquisition module, a preprocessing module, a forward processing module and an image denoising module; the image acquisition module is used to acquire standard-dose computed tomography images and low-dose computed tomography images; The preprocessing module is used to preprocess the computed tomography images acquired by the image acquisition module; the forward processing module is used to convert the standard-dose computed tomography images into random noise; and the image denoising module is used to perform denoising on the low-dose computed tomography images.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The memory stores a computer program; the processor executes a low-dose CT image denoising method based on a conditional diffusion model as described in any one of claims 1 to 7.

10. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement a low-dose CT image denoising method based on a conditional diffusion model as described in any one of claims 1 to 7.

Images