A low-dose CT image denoising method

Abstract

This invention discloses a low-dose CT image denoising method, comprising the following steps: S1: establishing a dataset, defining a set of noisy LDCT images x, and a set of high-quality NDCT images y corresponding to the LDCT image set x; S2: constructing a denoising model with parameter θ. Use templates to build Mapping; S3: Training the denoising model using hybrid loss Set the learning rate and momentum parameters, and use error backpropagation to train the denoising model. S4: Image denoising; After preprocessing the LDCT images in the test set, the images are input into the trained denoising model to obtain the corresponding denoised images. The denoising method proposed in this application takes into account the denoising difficulty of different regions of the image, and adaptively balances them by adding weights. It also promotes the generation of image texture details by losing high-frequency information, thus avoiding the denoised image from becoming overly smooth.

Description

A method for denoising low-dose CT images Technical Field

[0001] This invention belongs to the field of image processing technology, and specifically relates to a method for denoising low-dose CT images. Background Technology

[0002] X-ray computed tomography (CT) scans are widely used in clinical screening, diagnosis, and intervention due to their rapid scanning speed and clear image quality. However, the high radiation levels associated with CT scans also expose patients to the risk of genetic damage and malignant tumors. Epidemiological studies have shown that even two to three CT scans can significantly increase the risk of cancer, especially in children. However, reducing the radiation dose can lead to noise and artifacts in the reconstructed CT images, potentially interfering with disease identification and analysis, and severely impairing diagnostic performance. Therefore, obtaining high-quality CT images by removing noise from low-dose CT (LDCT) images has significant medical and diagnostic implications.

[0003] With the rapid development of deep neural networks in recent years, the combination of CT and deep learning can not only achieve image analysis but also image reconstruction, showing great potential in LDCT image denoising. Deep learning methods consist of two important components: network model structure and loss function. Deep learning uses a network model to construct a mapping from LDCT images to normal-dose CT (NDCT) images. During training, the network model is updated through backpropagation of the loss function until the error is minimized. In other words, the loss function directly affects the model's performance. Although existing deep learning methods have made various innovations in model structure and achieved certain results, they typically use the MAE loss function as the training objective. The MAE loss function calculates the average difference between each pixel in the image. While this results in a high PSNR for the denoising results, it still has two drawbacks:

[0004] 1. It ignores the different denoising difficulties of different regions in LDCT images, resulting in the model ignoring more difficult-to-process regions to obtain higher average visual quality output.

[0005] 2. The texture information of an image exists in the high-frequency part, while the pixel information is in the low-frequency part of the image. Therefore, the MAE loss function has the problem of causing the reconstructed image to lose a lot of texture information. Summary of the Invention

[0006] To address the problems mentioned in the background art, the present invention provides a low-dose CT image denoising method to solve the problems in the prior art that ignore difficult-to-process areas to obtain higher average visual quality output and that the reconstructed image loses a large amount of texture information.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for denoising low-dose CT images includes the following steps:

[0009] S1: Create a dataset, specifying a set of LDCT images with noise. and LDCT image set Corresponding high-quality NDCT image set ;

[0010] S2: Setup parameters are Denoising model , build Mapping;

[0011] S3: Train the denoising model using hybrid loss. Set the learning rate and momentum parameters, and use error backpropagation to train the denoising model. Once the maximum number of iterations is reached, a trained denoised model is generated.

[0012] S4: Image denoising; After preprocessing the LDCT images in the test set, input them into the trained denoising model to obtain the corresponding denoised images.

[0013] Preferably, the specific steps for building the dataset are as follows:

[0014] S1.1: Convert the obtained CT image into a grayscale image;

[0015] S1.2: Image preprocessing, using the preprocessed LDCT image and the corresponding NDCT image as a group to obtain the training set;

[0016] S1.3: Randomly crop and rotate each set of images in the training set by a certain angle to expand the training set.

[0017] Preferably, in step S1.2, the image preprocessing involves sequentially converting the grayscale image to 224×224 and performing pixel normalization based on the mean and variance.

[0018] Preferably, in S2, the denoising model uses the original network architecture, which includes pix2pix, RED-CNN, SPARNet, and ldct-nonlocal.

[0019] Preferably, in S3, LDCT images Input into the denoising model Then, the denoised image is obtained. Where N represents the amount of data, the hybrid loss function used in hybrid loss training consists of two parts: WPLoss loss and HFLoss loss. The specific method for obtaining the WPLoss loss function is as follows:

[0020] After dividing the generated denoised image and the corresponding NDCT image into multiple non-overlapping sub-blocks, the MAE loss of each sub-block is calculated. The loss weights for each sub-block are defined as follows:

[0021] ;

[0022] in, The first image represents the denoised image and the second image of the NDCT image. Each block is correct. It is a positive number, and it is a set. One of the values;

[0023] The WPLoss loss function is defined as follows:

[0024] ;

[0025] The specific method for obtaining the HFLoss loss function is as follows:

[0026] The image is transformed from the spatial domain to the frequency domain using Fourier transform. Then, a high-frequency filter is used to retain only the high-frequency components. Finally, the corresponding result of the high-frequency components is obtained through inverse Fourier transform. The difference in high-frequency information is used as the HFLoss loss. The discrete Fourier transform of the two-dimensional image is expressed as follows:

[0027] ;

[0028] The discrete inverse Fourier transform of a two-dimensional image is expressed as:

[0029] ;

[0030] In the formulas for Discrete Fourier Transform and Inverse Fourier Transform, m and n represent the length and width of the image, respectively.

[0031] Therefore, the loss function HPLoss is calculated as follows:

[0032] ;

[0033] Where G represents a high-pass filter;

[0034] Combining the WPLoss and HFLoss loss functions described above, the hybrid loss function is expressed as follows:

[0035] ;

[0036] in, and for and The weighting coefficients.

[0037] Preferably, in S4, the image preprocessing is the same as that described in S1.2.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This application first inputs training data into a denoising model to obtain a denoised image. Then, it calculates the MAE loss for different sub-blocks of the denoised image and assigns weights to obtain WPLoss. After transforming the image from the spatial domain to the frequency domain using Fourier transform, a high-pass filter is used to filter out low-frequency information. Then, an inverse Fourier transform is used to transform the image back to the spatial domain, and the difference in high-frequency information is calculated to obtain HFLoss. Finally, a hybrid loss consisting of WPLoss and HFLoss is used to update the denoising model in reverse.

[0040] The denoising method proposed in this invention takes into account the denoising difficulty of different regions of the image, and adaptively balances them by adding weights. It also promotes the generation of image texture details by losing high-frequency information, thus avoiding the image from becoming overly smooth after denoising. Attached Figure Description

[0041] Figure 1 is a schematic diagram of the method flow of this application;

[0042] Figure 2 shows the effect of the hybrid loss function on the denoising effect of each model on selected abdominal CT images in the Mayo test set.

[0043] Figure 3 is a schematic diagram comparing the denoising effects of using different loss functions on pix2pix;

[0044] Figure 4 shows the effect of the mixture loss function on the denoising effect of each model on selected abdominal CT images in the simulated dataset test set.

[0045] Figure 5 is a schematic diagram comparing the denoising effects of different loss functions on ldct-nonlocal;

[0046] Figure 6 shows the quantization results of denoising each model using different loss functions on the Mayo test set;

[0047] Figure 7 shows the quantization results of each model using different loss functions on the simulated dataset test set after denoising. Detailed Implementation

[0048] To facilitate understanding of the technical content of this invention by those skilled in the art, the invention will be further described in detail below with reference to the accompanying drawings and specific examples. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0049] Example 1

[0050] As shown in Figure 1, a low-dose CT image denoising method includes the following steps:

[0051] S1: Create a dataset, specifying a set of LDCT images with noise. and LDCT image set Corresponding high-quality NDCT image set ;

[0052] S2: Setup parameters are Denoising model , build Mapping;

[0053] S3: Train the denoising model using hybrid loss. Set the learning rate and momentum parameters, and use error backpropagation to train the denoising model. Once the maximum number of iterations is reached, a trained denoised model is generated.

[0054] S4: Image denoising; After preprocessing the LDCT images in the test set, input them into the trained denoising model to obtain the corresponding denoised images.

[0055] In this embodiment, the application first inputs training data into a denoising model to obtain a denoised image, wherein a hybrid loss is used to train the denoising model. Set the learning rate and momentum parameters, and use error backpropagation to train the denoising model. Once the maximum number of iterations is reached, a trained denoised model is generated, and the denoised model is updated in reverse using a hybrid loss.

[0056] The denoising method proposed in this invention takes into account the denoising difficulty of different regions of the image, and adaptively balances them by adding weights. It also promotes the generation of image texture details by losing high-frequency information, thus avoiding the image from becoming overly smooth after denoising.

[0057] Example 2

[0058] The difference between this embodiment and Embodiment 1 is that the specific steps for building the dataset are as follows:

[0059] S1.1: Convert the obtained CT image into a grayscale image;

[0060] S1.2: Image preprocessing, using the preprocessed LDCT image and the corresponding NDCT image as a group to obtain the training set;

[0061] S1.3: Randomly crop and rotate each set of images in the training set by a certain angle to expand the training set.

[0062] Example 3

[0063] The difference between this embodiment and embodiment 2 is that in step S1.2, the image preprocessing is to sequentially convert the grayscale image to 224×224 and perform pixel normalization based on mean and variance.

[0064] Example 4

[0065] The difference between this embodiment and embodiment 1 is that in S2, the denoising model uses the original network architecture, which includes pix2pix, RED-CNN, SPARNet and ldct-nonlocal.

[0066] Example 5

[0067] The difference between this embodiment and embodiment 4 is that in S3, the LDCT image... Input into the denoising model Then, the denoising result is obtained. Where N represents the data size. The hybrid loss function proposed in this invention consists of two parts: WPLoss loss and HFLoss loss. It can be expressed as follows:

[0068]

[0069] in, and These represent WPLoss and HFloss, respectively. and These are their weighting coefficients.

[0070] WPLoss:

[0071] The generated denoised image and the corresponding NDCT image are divided into After identifying non-overlapping sub-blocks, the MAE loss for each sub-block is calculated. In regions with higher denoising difficulty, where the generated images from training differ significantly from the target images, larger weights are assigned to increase the model's focus on these regions. Specifically, the th... The loss weights for each sub-block are defined as follows:

[0072]

[0073] in, The first denoised image and the second NDCT image represent the... Each block is correct. It is a positive number, and in this invention it is defined as a set. One of the values. It's important to note that, to prevent model instability due to excessively large local loss weights, the weights should be adjusted accordingly. The range of values ​​is limited, for example [0.25, 4].

[0074] Therefore, the definition of WPLoss proposed in this invention is as follows:

[0075]

[0076] Experiments revealed that the denoising model's denoising ability was too weak in some areas and too strong in others. Generally, the more complex the texture, the more difficult the denoising, but these are often the locations of key pathological information. Here, by adjusting the weight coefficients of the local MAE loss in the image, the model can be made to focus more on regions with complex textures, which requires almost no additional computation.

[0077] HFLoss:

[0078] Fourier transform can be used to transform an image from the spatial domain to the frequency domain. Then, a high-frequency filter is used to retain only the high-frequency components. Finally, an inverse Fourier transform is performed to obtain the corresponding results for the high-frequency components of the image, and the difference in high-frequency information is used as the HFLoss loss.

[0079] The discrete Fourier transform and inverse Fourier transform of a two-dimensional image are shown below:

[0080]

[0081]

[0082] Where m and n represent the length and width of the image.

[0083] Therefore, the loss function HPLoss defined in this invention is calculated as follows:

[0084]

[0085] Where G represents a high-pass filter.

[0086] Example 6

[0087] The difference between this embodiment and embodiment 2 is that in S4, the image preprocessing is the same as that described in S1.2.

[0088] Experimental Data: Experiment 1: Mayo Dataset

[0089] The Mayo dataset is a publicly available dataset used in the "2016-NIH-AAPM-Mayo Clinic low-dose CT challenge". This dataset consists of 2378 3 mm thick NDCT images and corresponding LDCT images from 10 patients. The NDCT images were generated under conditions of tube potential and effective mA of 120 kV, 200 mAs. The dataset was divided into training and test sets in a 9:1 ratio based on the number of patients. In the experiments, images of a uniform size of 224×224 were used.

[0090] Figure 6 shows the quantitative metrics of denoising results for different models using different loss functions on the Miao test set. It can be seen that the hybrid loss function proposed in this invention can significantly improve the denoising performance of all models. On RED-CNN, pix2pix, SPARNet, and ldct-nonlocal, PSNR is improved by 0.0017, 0.7376, 0.2739, and 0.1788, respectively, and SSIM is improved by 0.001, 0.003, 0.0014, and 0.0012, respectively.

[0091] A representative abdominal CT image was selected from the Mayo test set, and the impact of the hybrid loss function on each model was visually verified. Figure 2 shows the denoising results of each model before and after using the hybrid loss function. Figures 2(a1) to (e1) represent LDCT images and the denoising results of RED-CNN, pix2pix, SPARNet, and ldct-nonlocal before adding the hybrid loss function, respectively. Figures 2(a2) to (e2) represent NDCT images and the denoising results of each model after adding the hybrid loss function, respectively. It can be seen that strong quantum noise exists in LDCT images, especially near structural areas with high attenuation coefficients (such as bones). All models show denoising effects to varying degrees. However, in the first row, RED-CNN leads to over-smoothing and loses a large amount of texture information because it only uses the average loss per pixel as its loss function. In contrast, pix2pix and SPARNet better preserve structural features, but still retain a large amount of noise and artifacts. Although ldct-nonlocal eliminates most of the noise and artifacts and has good structural fidelity, some blurring remains near the bones. In contrast, it can be seen that all models show better denoising performance and retain more details after using the hybrid loss function proposed in this invention.

[0092] Figure 2, rows 2 and 4, shows magnified images of rectangular regions of interest (ROIs) in LDCT. Observing the areas indicated by the arrows, RED-CNN smooths some important structures. pix2pix, SPARNet, and ldct-nonlocal can preserve these details to varying degrees without blurring, but localized block artifacts appear, such as near bones. For the circled low-contrast liver lesion areas, after using the hybrid loss function proposed in this invention, the denoising results of all models reduce artifacts, better preserve edges, and improve the detectability of lesions.

[0093] Figure 3 further illustrates the denoising effects of using different loss functions on a separate abdominal LDCT image using pix2pix, verifying the effectiveness of each part of the proposed hybrid loss function. The first row shows the LDCT image, the denoising results of pix2pix using different loss functions, and the corresponding NDCT image. The second row represents a magnified image of the matrix region in the LDCT image. By observing the area indicated by the arrow, it can be observed that the original pix2pix denoising result exhibits severe blocky artifacts and is blurred near the bones. Although the artifacts in the pix2pix denoising result are reduced after adding WPLoss, the problem of texture loss still exists. In contrast, after using the hybrid loss function, the denoising result of pix2pix shows the smallest difference from the NDCT image and retains most of the details.

[0094] Experiment 2: Simulated Dataset

[0095] The simulated dataset used 4000 NDCT images from 34 patients downloaded from TCIA, and corresponding NDCT images were obtained by adding Poisson noise to the projection domain. Of these, 3600 images were used for training, and the remainder served as the test set. In the experiments, images of a uniform size of 224×224 were used. The Poisson noise model can be represented as:

[0096]

[0097] in Indicates the first X-ray along the path of the X-ray. The measurement of each path represents the conversion gain from X-ray photons to electrons, and the initial incident intensity of the X-rays is used to constrain the noise intensity. Compared to the Mayo dataset, the TCIA dataset contains more body parts.

[0098] Figure 7 shows the quantitative metrics of the denoising results of different models on the simulated dataset test set. The results show that the denoising performance of each model is improved after using the hybrid loss function proposed in this invention, which provides quantitative evaluation support for visual observation.

[0099] Figure 4 illustrates the impact of the hybrid loss function on the denoising performance of each model. The second and fourth rows are magnified images of the matrix regions on the LDCT images. By observing the parts marked with arrows, it is clear that all models improve in both noise suppression and structure preservation after incorporating the hybrid loss. Conversely, in the original models, RED-CNN and pix2pix smooth out some details. While SPARNet and ldct-nonloacl retain more details, localized blocky artifacts appear.

[0100] Figure 5 shows the denoising results of ldct-nonlocal on another abdominal CT image in the simulated dataset test set using different loss functions. Obvious block artifacts and blurring appeared after ldct-nonlocal denoising. Visual inspection of the magnified image revealed that the denoising result after adding WPLoss was clearer, but the edge sharpness remained, while adding HFLoss compensated for the edge texture blurring.

Claims

1. A method for denoising low-dose CT images, characterized in that, Includes the following steps: S1: Create a dataset, specifying a set of LDCT images with noise. and LDCT image set Corresponding high-quality NDCT image set S2: Setup parameters are as follows Denoising model , build Mapping; S3: Training the denoising model using hybrid loss Set the learning rate and momentum parameters, and use error backpropagation to train the denoising model. Reaching the maximum number of iterations, a well-trained denoising model is generated. In S3, the LDCT image... Input into the denoising model Then, the denoised image is obtained. Where N represents the data size, the hybrid loss function used in hybrid loss training consists of two parts: WPLoss loss and HFLoss loss. The specific method for obtaining the WPLoss loss function is as follows: After dividing the generated denoised image and the corresponding NDCT image into multiple non-overlapping sub-blocks, the MAE loss of each sub-block is calculated. The loss weights for each sub-block are defined as follows: ;in, The first image represents the denoised image and the second image of the NDCT image. Each block is correct. It is a positive number, and it is a set. If a value is found in the given information, then the WPLoss loss function is defined as follows: The specific method for obtaining the HFLoss loss function is as follows: The image is transformed from the spatial domain to the frequency domain using Fourier transform. Then, a high-frequency filter is used to retain only the high-frequency components. Finally, the corresponding result for the high-frequency components of the image is obtained through inverse Fourier transform. The difference in the high-frequency information is used as the HFLoss loss. The discrete Fourier transform of the two-dimensional image is expressed as: The discrete inverse Fourier transform of a two-dimensional image is expressed as: In the formulas for Discrete Fourier Transform and Inverse Fourier Transform, m and n represent the length and width of the image, respectively. Therefore, the loss function HPLoss is calculated as follows: Where G represents a high-pass filter; combining the above WPLoss and HFLoss loss functions, the hybrid loss function is expressed as follows: ;in, and for and S4: Image denoising; After preprocessing the LDCT images in the test set, input them into the trained denoising model to obtain the corresponding denoised images.

2. The low-dose CT image denoising method according to claim 1, characterized in that, The specific steps for building the dataset are as follows: S1.1: Convert the obtained CT images into grayscale images; S1.2: Image preprocessing, taking the preprocessed LDCT images and the corresponding NDCT images as a group to obtain the training set; S1.3: Randomly crop and randomly rotate each group of images in the training set to expand the training set.

3. The low-dose CT image denoising method according to claim 2, characterized in that, In step S1.2, the image preprocessing involves sequentially converting the grayscale image to 224×224 and performing pixel normalization based on the mean and variance.

4. The low-dose CT image denoising method according to claim 1, characterized in that, In S2, the denoising model uses the original network architecture, which includes pix2pix, RED-CNN, SPARNet, and ldct-nonlocal.

5. A low-dose CT image denoising method according to claim 2, characterized in that, In S4, the image preprocessing is the same as that described in S1.2.

Images