A noise removal method for low-dose X-ray CT images

Abstract

The present application relates to a kind of low-dose X-ray CT image-oriented noise removal method, by the method of semi-supervised contrast learning, low-dose CT image is removed noise by teacher-student consistency network and prior knowledge transfer method, reduce effective information loss, retain more detail information, to provide high-quality input data for subsequent image reconstruction, realize the reconstruction of high-quality CT image under low-dose X-ray condition.Compared with the traditional method based on low-pass filter or image statistics, the present application can better remove noise, so that the image after denoising is more real and clear, and by adjusting its network structure and hyperparameter, it can be applied to various low-dose CT image denoising tasks, including different scanning modes, different organs, etc., with high versatility and scalability.

Description

Technical Field

[0001] This invention belongs to the field of computed tomography (CT) technology, and in particular to a noise removal method for low-dose X-ray CT images. Background Technology

[0002] Computed tomography (CT) imaging is a commonly used medical imaging technique. It utilizes the varying densities of human tissues and their different absorption rates of X-rays to generate two-dimensional or three-dimensional images, enabling the observation, diagnosis, and treatment of internal human structures. The working principle of CT involves passing an X-ray beam through the human body, measuring the amount of X-ray absorption by corresponding detectors, and transmitting this data to a computer for processing. The computer converts this data into images, allowing doctors to visually observe internal structures and make diagnoses. CT has wide applications in diagnosing tumors, fractures, and brain diseases. It is important to note that because CT uses X-rays, radiation dose is a crucial factor. During CT examinations, it is essential to minimize the radiation dose to avoid excessive radiation damage to the body. Low-dose imaging techniques can achieve this by reducing the radiation dose, but this can lead to a decrease in image quality, affecting the doctor's diagnostic judgment. Therefore, noise reduction processing is necessary for low-dose CT images to improve image quality.

[0003] Currently, commonly used low-dose CT image denoising methods include image post-processing, iterative reconstruction, and projection data filtering. For example, the wavelet-domain-based PWLS method can combine noise statistical models at different scales to change the local filtering intensity with the noise intensity, enhancing the denoising effect. Mean filtering methods smooth the image and eliminate noise by averaging the pixel values ​​around a given pixel and using this average as the current pixel value. Total variation-based denoising methods minimize the total variation of the image to remove noise and preserve image edges. Deep learning-based CT image denoising is a novel approach developed in recent years: It involves collecting a dataset of noisy CT images and dividing it into training and testing sets; designing and training a deep learning model, using noisy images as input and real noise-free images as output; and then using the trained model to denoise the noisy images in the testing set to obtain the denoised images. Some literature indicates the use of a deep residual network trained with mean squared error as the loss function, optimized using stochastic gradient descent, which avoids the vanishing gradient problem in deep network training and improves both training speed and denoising performance. Other literature proposes a noise reduction method based on DCNN, which first converts the image to the frequency domain and then inputs the frequency domain image into the DCNN for training. This method effectively removes Gaussian noise from images and can also be used for CT image denoising. However, currently, no method has been found that uses deep learning for CT image denoising to preserve image details and achieve good denoising results. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and propose a noise removal method for low-dose X-ray CT images. By using a teacher-student consensus network built based on image post-processing, prior knowledge learned from normal-dose CT images is applied to the denoising of low-dose CT images. At the same time, a variable autoencoder module is added to improve the denoising effect, thereby achieving high-quality CT reconstruction images for low-dose CT.

[0005] The technical problem solved by this invention is achieved through the following technical solution:

[0006] A noise removal method for low-dose X-ray CT images includes the following steps:

[0007] Step 1: Construct the dataset;

[0008] Step 2: Establish a denoising neural network model;

[0009] Step 3: Use the TensorFlow framework to build a variable autoencoder (VAE) module to extract prior knowledge from the normal dose CT images in Step 1;

[0010] Step 4: Construct the loss function for the teacher-student consistency network and the variable autoencoder module;

[0011] Step 5: Train the denoising neural network model using the dataset from Step 1;

[0012] Step 6: Input the noisy low-dose CT image into the trained student network SN, and use the trained variable autoencoder (VAE) as a supplementary feature extraction for the student network SN. The student network SN uses the low-dimensional feature representation of the NCT image learned during the VAE reconstruction process to obtain a denoised high-quality CT image at the output.

[0013] Furthermore, the dataset in step 1 includes: the normal dose CT image dataset NCT and the low dose CT image dataset LCT. n sets of clinical images are randomly selected from the dataset as the training set, and the rest are used as the test set.

[0014] Moreover, the specific implementation method of step 2 is as follows: use ResNet as the backbone network for constructing the teacher network TN and the student network SN, and add corresponding upsampling layers, downsampling layers, convolutional layers, and activation functions to construct a complete network structure.

[0015] Furthermore, the variable autoencoder module (VAE) in step 3 includes an encoder and a decoder;

[0016] The encoder is constructed as follows:

[0017]

[0018] in, Here are the encoder parameters, and N is a Gaussian distribution. and Input normal dose CT image x NCT The corresponding mean and variance are used to obtain a latent vector z that follows a Gaussian distribution determined by the mean and variance through a mapping function. NCT ;

[0019] The decoder is constructed as follows:

[0020] p θ (x' NCT |z NCT )=N(μ θ (z NCT ),σ θ 2 (z NCT ))

[0021] Where θ is the decoder parameter, N is a Gaussian distribution, and μ θ (z NCT ) and σ θ 2 (z NCT ) are the potential vectors z NCT The corresponding mean and variance are used to obtain a reconstructed image x' that follows a Gaussian distribution determined by the mean and variance, through a mapping function. NCT ;

[0022] The encoder part of the Variable Autoencoder (VAE) module is connected to the encoder part of the SN by adding skip connections to share feature representations. At the same time, the decoder part of the VAE module is used as a supplementary feature extractor of the SN to extract richer feature information.

[0023] Furthermore, the method for calculating the teacher-student consensus network loss function L1 in step 4 is as follows:

[0024] L1 = L KD +λ1L PK +λ2L M

[0025]

[0026] L PK =||F SN -F TN ||1

[0027]

[0028] Among them, L KD For the loss of consistency between teachers and students, L PK For prior knowledge transfer loss, L M This is the mean squared error loss, where T is the temperature parameter, and y SN Let y be the output of SN on the i-th LCT image training sample. TN Let F be the output of TN on the i-th NCT image training sample. The prior knowledge transfer loss guides the features of SN to mimic the features of TN by reducing the L1 distance between SN and TN, where F SN and F TN These are the feature maps output from the last convolutional layer of SN and TN, respectively, and x in the mean squared error loss. i These are sample pixel values ​​from the NCT image dataset. λ1 and λ2 are the pixel values ​​generated by the decoder, where N is the number of pixels; λ1 and λ2 are both weighting coefficients.

[0029] Furthermore, the loss function L2 of the variable autoencoder module in step 4 is calculated as follows:

[0030] L2=αL R +βL KL

[0031]

[0032]

[0033] Moreover, L R It is the reconstruction error, L KL It is the Kurbak-Rebler divergence, x i These are sample pixel values ​​from the NCT image dataset. Let N be the pixel value generated by the decoder, and N be the number of pixels. The KL divergence is used to constrain the latent variables in the latent space to follow a standard normal distribution, where μ... i and Let α and β be the i-th components of the mean vector and variance vector, respectively, and let α and β be the adjustment weight coefficients.

[0034] The advantages and positive effects of this invention are:

[0035] This invention employs a semi-supervised contrastive learning method to remove noise from low-dose CT images using a teacher-student consensus network and prior knowledge transfer. This reduces the loss of effective information and preserves more detailed information, providing high-quality input data for subsequent image reconstruction and enabling the reconstruction of high-quality CT images under low-dose X-ray conditions. Compared to traditional methods based on low-pass filters or image statistics, this method can better remove noise, resulting in more realistic and clearer images. Furthermore, by adjusting its network structure and hyperparameters, it can be applied to various low-dose CT image denoising tasks, including different scanning methods and different organs, demonstrating high versatility and scalability. Attached Figure Description

[0036] Figure 1 This is a diagram illustrating the network calculation process of the variable autoencoder module of the present invention.

[0037] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0038] The present invention will be further described in detail below with reference to the accompanying drawings.

[0039] A noise removal method for low-dose X-ray CT images includes the following steps:

[0040] Step 1: Construct a dataset by collecting CT images of various human tissues and organs from clinical practice as a dataset for training and testing the neural network. This includes, but is not limited to, using existing X-ray images for professional calibration and using professional software models for image synthesis.

[0041] The datasets include: the normal-dose CT image dataset NCT and the low-dose CT image dataset LCT, obtained by controlling scanning parameters such as scanning mode, scanning speed, focal spot size, X-ray tube voltage, and current. n sets of clinical images are randomly selected from the datasets as the training set, ideally including all human tissues and organs. The remaining portion of the datasets serves as the test set.

[0042] Step 2: Establish a denoising neural network model.

[0043] ResNet is used as the backbone network for constructing the teacher network TN and the student network SN. Appropriate upsampling and downsampling layers, convolutional layers, and activation functions are added to build the complete network structure. TN uses deeper layers and more parameters to obtain richer feature representations, extracting prior knowledge by learning the latent distribution of the data. SN, on the other hand, uses shallower layers and fewer parameters to improve computational efficiency, building a network with the same structure but different parameters by learning the features and generation process of TN.

[0044] Step 3: Use the TensorFlow framework to build a variable autoencoder (VAE) module to extract prior knowledge from the normal dose CT images in Step 1.

[0045] like Figure 1 As shown, the Variable Autoencoder (VAE) module comprises an encoder and a decoder. The encoder maps the normal-dose CT image to a low-dimensional vector containing the mean and variance in the latent space, enabling feature extraction of the input. These vectors are then used to sample from the latent space. The decoder maps the latent vector back to the image space, generating a reconstructed normal-dose CT image.

[0046] The encoder is constructed as follows:

[0047]

[0048] in, Here are the encoder parameters, and N is a Gaussian distribution. and Input normal dose CT image x NCT The corresponding mean and variance are used to obtain a latent vector z that follows a Gaussian distribution determined by the mean and variance through a mapping function. NCT .

[0049] The decoder is constructed as follows:

[0050] p θ (x' NCT |z NCT )=N(μ θ (z NCT ),σθ 2 (z NCT ))

[0051] Where θ is the decoder parameter, N is a Gaussian distribution, and μ θ (z NCT ) and σ θ 2 (z NCT ) are the potential vectors z NCT The corresponding mean and variance are used to obtain a reconstructed image x' that follows a Gaussian distribution determined by the mean and variance, through a mapping function. NCT .

[0052] The encoder part of the Variable Autoencoder (VAE) module is connected to the encoder part of the SN by adding skip connections to share feature representations. At the same time, the decoder part of the VAE module is used as a supplementary feature extractor of the SN to extract richer feature information.

[0053] Step 4: Construct the loss function for the teacher-student consistency network and the variable autoencoder module.

[0054] The main body of the teacher-student consistency network loss function L1 is the teacher-student consistency loss, and a weighted regularization term is added to better optimize the model's prediction results:

[0055] L1 = L KD +λ1L PK +λ2L M

[0056]

[0057] L PK =||F SN -F TN ||1

[0058]

[0059] Among them, L KD For the loss of consistency between teachers and students, L PK For prior knowledge transfer loss, L M This is the mean squared error loss, where T is the temperature parameter, and y SN Let y be the output of SN on the i-th LCT image training sample. TN Let F be the output of TN on the i-th NCT image training sample. The prior knowledge transfer loss guides the features of SN to mimic the features of TN by reducing the L1 distance between SN and TN, where F SN and F TN These are the feature maps output from the last convolutional layer of SN and TN, respectively, and x in the mean squared error loss.i These are sample pixel values ​​from the NCT image dataset. λ1 and λ2 are the pixel values ​​generated by the decoder, where N is the number of pixels. λ1 and λ2 are weighting coefficients, adjusted based on the relative magnitudes and trends of the prior knowledge transfer loss and mean squared error loss during training. A larger prior knowledge transfer loss indicates a greater focus on prior knowledge information, thus λ1 should be increased appropriately. Conversely, a larger mean squared error loss indicates a greater focus on TN's ability to extract NCT image data distribution features, thus λ2 should be increased appropriately.

[0060] The loss function L2 of the variable autoencoder module is calculated as follows:

[0061] L2=αL R +βL KL

[0062]

[0063]

[0064] Moreover, L R It is the reconstruction error, L KL It is the Kurbak-Rebler divergence, x i These are sample pixel values ​​from the NCT image dataset. Let N be the pixel value generated by the decoder, and N be the number of pixels. The KL divergence is used to constrain the latent variables in the latent space to follow a standard normal distribution, where μ... i and The weights α and β are the i-th components of the mean and variance vectors, respectively. Adjusting these weights is achieved by training the variable autoencoder module. During training, the impact of reconstruction error and KL divergence on the overall L2 loss function is observed, and the weights are adjusted based on their trends. If the reconstruction error is large, it indicates a greater emphasis on data reconstruction or restoration, so α is increased appropriately. If the KL divergence is large, it indicates a greater focus on the expression and representation learning of latent variables, so β ​​is increased appropriately.

[0065] Step 5: Train the denoising neural network model using the dataset from Step 1.

[0066] The neural network model is trained using the standard backpropagation algorithm: First, NCT images are input into the TN and variable autoencoder respectively for training. The TN learns information and details from normal-dose CT images, generating a high-quality reference image as the standard to which the output of the SN approximates during training. The variable autoencoder learns prior knowledge from the NCT images, extracting the features of the data distribution, i.e., the distribution of the latent vectors, thereby generating new reconstructed images with a similar distribution to the training data. When the value of the loss function L2 is less than or equal to 10... -4Model training is completed on time; then, LCT images are input into SN for training. The loss function is optimized by comparing the error between its output and the reference image generated by TN, the model parameters are updated, and it is observed whether the loss function can converge to a minimum. If it cannot converge, the parameters involved in the network, such as λ1, are changed and adjusted, and appropriate hyperparameter values ​​are determined through methods such as cross-validation. Multiple models are trained according to different hyperparameter combinations and evaluated on the validation set. The best-performing combination is selected as the final hyperparameter values, and training is repeated until the value of the loss function L1 is less than or equal to 10. -3 Complete model training.

[0067] Step 6: Input the noisy low-dose CT image into the trained student network SN, and use the trained variable autoencoder (VAE) as a supplementary feature extraction for the student network SN. The student network SN uses the low-dimensional feature representation of the NCT image learned during the VAE reconstruction process to obtain a denoised high-quality CT image at the output.

[0068] Based on the above-described noise removal method for low-dose X-ray CT images, the effectiveness of the present invention was verified through testing.

[0069] like Figure 2 As shown, this invention first scans the head, chest, abdomen, pelvis, and spine with a radiation dose of 3 mSv or lower to collect information from different tissues and organs of different patients, resulting in a dataset of, but not limited to, 2000 widely distributed LCT images. 1500 or more of these noisy LCT images are used as the training set, and the remaining 500 are used as the test set. Similarly, the same areas are scanned with a radiation dose of 8 mSv or higher, resulting in a dataset of, but not limited to, 2000 widely distributed NCT images. 1500 or more of these NCT images are used as the training set, and the remaining 500 are used as the test set. A ResNet-based teacher-student consensus denoising network model is established, and a variable autoencoder is built within the TensorFlow framework to extract prior knowledge from normal dose CT images. A skip connection structure is used to establish the relationship between the variable autoencoder and the signal generator (SN), enabling the SN to better learn the latent distribution of the NCT data. 1500 NCT images were input into TN and VAE respectively. A loss function was constructed, and the network was trained using the backpropagation algorithm. Parameters such as α and β were adjusted to make the output of SN continuously approximate the reference image output by TN. When the loss function is less than or equal to 10... -3 Training is completed in a timely manner. The LCT image is input into the trained SN, and calculations are performed under the joint guidance of TN and VAE. A high-quality denoised CT image can be obtained from the output of the SN.

[0070] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A noise removal method for low-dose X-ray CT images, characterized in that: Includes the following steps: Step 1: Construct the dataset; Step 2: Establish a denoising neural network model; Step 3: Use the TensorFlow framework to build a variable autoencoder (VAE) module to extract prior knowledge from the normal dose CT images in Step 1; In step 3, the Variable Autoencoder (VAE) module includes an encoder and a decoder; The encoder is constructed as follows: ； Where Ø represents the encoder parameters, and N is a Gaussian distribution. and Input normal dose CT images respectively The corresponding mean and variance are used to obtain a latent vector that follows a Gaussian distribution determined by the mean and variance through a mapping function. ; The decoder is constructed as follows: ； Where θ represents the decoder parameters, and N is a Gaussian distribution. and They are the potential vectors The corresponding mean and variance are used to obtain a reconstructed image that follows a Gaussian distribution determined by the mean and variance, through a mapping function. ; The encoder part of the Variable Autoencoder (VAE) module is connected to the encoder part of the SN by adding skip connections to share feature representations. At the same time, the decoder part of the VAE module is used as a supplementary feature extractor of the SN to extract richer feature information. Step 4: Construct the loss function for the teacher-student consistency network and the variable autoencoder module; Step 4: Teacher-Student Consistency Network Loss Function The calculation method is as follows: = + + ； ； = ； = ； in, For the loss of consistency between teachers and students, Loss due to prior knowledge transfer. This is the mean square error loss, where T is the temperature parameter. Let SN be the output of the i-th LCT image training sample. Let TN be the output of the i-th NCT image training sample. The prior knowledge transfer loss guides the features of SN to mimic the features of TN by reducing the L1 distance between SN and TN. and These are the feature maps output from the last convolutional layer of SN and TN, respectively, and the mean squared error loss is... These are sample pixel values ​​from the NCT image dataset. N represents the pixel values ​​generated by the decoder, where N is the number of pixels. , All are weighting coefficients; The loss function of the variable autoencoder module in step 4 The calculation method is as follows: =α +β ； = +(1- ) ； = ； and, It is a reconstruction error. It is the Kurbak-Rebler divergence. These are sample pixel values ​​from the NCT image dataset. Let N be the pixel value generated by the decoder, and N be the number of pixels. The KL divergence is used to constrain the latent variables in the latent space to follow a standard normal distribution. and Let be the i-th component of the mean vector and the variance vector, and α and β be the adjustment weight coefficients; Step 5: Train the denoising neural network model using the dataset from Step 1; The neural network model is trained using the standard backpropagation algorithm: First, NCT images are input into the TN and variable autoencoder respectively for training. The TN learns information and details from normal-dose CT images, generating a high-quality reference image as the standard approximated by the output during SN training. The variable autoencoder learns prior knowledge from the NCT images, extracting the features of the data distribution as the distribution of the latent vectors, thereby generating new reconstructed images with a similar distribution to the training data. When the loss function... The value is less than or equal to The model training is completed in one step; then, the LCT image is input into the SN for training. The loss function is optimized by comparing the error between its output and the reference image generated by TN, the model parameters are updated, and it is observed whether the loss function can converge to a minimum. If it cannot converge, the parameters involved in the network are modified and adjusted. The parameters are determined using methods such as cross-validation. Multiple models are trained based on different hyperparameter combinations, evaluated on a validation set, and the best-performing combination is selected as the final hyperparameter values. The model is then trained again until the loss function is achieved. The value is less than or equal to Complete model training; Step 6: Input the noisy low-dose CT image into the trained student network SN, and use the trained variable autoencoder (VAE) as a supplementary feature extraction for the student network SN. The student network SN uses the low-dimensional feature representation of the NCT image learned during the VAE reconstruction process to obtain a denoised high-quality CT image at the output.

2. The noise removal method for low-dose X-ray CT images according to claim 1, characterized in that: The dataset in step 1 includes: the normal dose CT image dataset NCT and the low dose CT image dataset LCT. n sets of clinical images are randomly selected from the dataset as the training set, and the rest are used as the test set.

3. The noise removal method for low-dose X-ray CT images according to claim 2, characterized in that: The specific implementation method of step 2 is as follows: use ResNet as the backbone network for constructing the teacher network TN and the student network SN, and add corresponding upsampling layers, downsampling layers, convolutional layers, and activation functions to construct a complete network structure.

Images