A method and system for removing noise from a CT reconstructed image

Abstract

The application discloses a CT reconstructed image noise removing method and system, wherein the CT reconstructed image noise removing method comprises the following steps: step (1), obtaining a plurality of low-dose high-noise image-high-dose low-noise image pairs; step (2), training a low-dose high-noise image-high-dose low-noise image pair image noise distribution estimation network obtained in step (1) to obtain an estimated noise distribution intensity; step (3), inputting the estimated noise distribution intensity obtained in step (2) and the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) into an image recovery network to obtain an optimized image recovery network; and step (4), inputting a CT image to be processed into the optimized image recovery network obtained in step (3) to obtain an optimized CT image. The removing method is simple in steps and can obviously remove the noise in the CT image, and a high-quality CT image is obtained.

Description

Technical Field

[0001] This invention relates to the field of image reconstruction technology, and in particular to a method and system for noise removal in CT reconstructed images. Background Technology

[0002] In recent years, with the continuous development of computed tomography (CT) technology, CT imaging has been increasingly widely used in medical diagnosis. However, the high-dose radiation from CT scans can be harmful to the human body, raising concerns. Therefore, it is necessary to reduce radiation dose while ensuring image quality meets clinical diagnostic requirements. The most common methods for reducing radiation dose are to reduce tube current and shorten the exposure time of the X-ray tube. However, this reduces the number of photons received by the detector, leading to noise contamination of the projection data. Consequently, the reconstructed CT images not only contain significant noise but also produce streak artifacts, greatly reducing image quality and visual effects, affecting diagnosticians' image analysis, and easily leading to misdiagnosis. Therefore, image denoising has always been a key research focus in the fields of computer vision and image processing.

[0003] Existing CT image denoising algorithms can be categorized into projection domain denoising algorithms, image reconstruction algorithms, and image domain denoising algorithms. Projection domain denoising algorithms directly apply the denoising to the projection data. Typical methods include structure-adaptive filtering, bilateral filtering, and penalized weighted least squares. The advantage of these methods is their ability to utilize the statistical patterns of noise distribution in the projection domain. However, when edges in the projection domain are not well preserved, these methods suffer from spatial resolution loss. Among image reconstruction algorithms, the most representative is the filtered back-projection algorithm, which offers high resolution and fast imaging speed. In recent years, improved FBP algorithms, as well as adaptive statistical iterative reconstruction (ASIR) and model-based iterative reconstruction (MBIR), have been proposed. While these methods effectively improve the quality of reconstructed images, they significantly increase computational complexity.

[0004] Projection domain denoising algorithms and image reconstruction algorithms both rely on the original projection data, while image domain denoising algorithms do not depend on projection data and can directly process the reconstructed CT images, thus becoming a research hotspot for low-dose CT image denoising. In recent years, CT image denoising algorithms based on artificial intelligence technologies (such as deep learning or neural network models) have achieved significant results compared to traditional methods. However, these algorithms still have shortcomings: they require a large amount of high-noise image-low-noise image data pairs for training, but in reality, it is difficult to obtain truly noisy images and corresponding normal-dose label images. While medium-dose label images are easier to obtain than normal-dose label images, they contain a certain amount of noise, which severely affects the denoising effect of the algorithm because deep learning models learn not only the structural information in the label images but also the noise information. How to utilize these low-quality image data to improve the denoising effect of the model is a challenging problem.

[0005] Therefore, it is essential to provide a method and system for noise removal in CT reconstructed images to address the shortcomings of existing technologies. Summary of the Invention

[0006] One objective of this invention is to overcome the shortcomings of existing technologies and provide a method for noise removal in CT reconstructed images. This method is simple in procedure and can significantly remove noise from CT images, resulting in high-quality CT images.

[0007] The above-mentioned objectives of the present invention are achieved through the following technical measures:

[0008] A method for noise removal from CT reconstructed images is provided, comprising the following steps:

[0009] Step (1): Obtain multiple pairs of low-dose high-noise images and high-dose low-noise images;

[0010] Step (2): Train the image noise distribution estimation network using the low-dose high-noise image-high-dose low-noise image obtained in step (1) to obtain the estimated noise distribution intensity;

[0011] Step (3): The estimated noise distribution intensity obtained in step (2) and the low-dose high-noise image-high-dose low-noise image obtained in step (1) are input into the image restoration network to train the optimized image restoration network.

[0012] Step (4): Input the CT image to be processed into the optimized image recovery network obtained in step (3) to obtain the optimized CT image.

[0013] Each pair of the low-dose high-noise image-high-dose low-noise image includes one low-dose high-noise image and one high-dose low-noise image, wherein the dose of the low-dose high-noise image is lower than the dose of the high-dose low-noise image.

[0014] Preferably, step (2) above specifically includes:

[0015] Step (2.1): Take the low-noise image high dose in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) as the label, and use the low-dose high-noise image as the input to train the image noise distribution estimation network, so that the objective function of the noise distribution estimation network converges, and finally obtain the image estimated by the pre-trained image noise distribution estimation network.

[0016] Step (2.2): The pre-trained image noise distribution estimation network estimates the image and compares it with the low-dose high-noise image in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) to obtain the estimated noise distribution intensity.

[0017] Preferably, the objective function of the noise distribution estimation network is the root mean square error measure function, expressed by equation (I).

[0018]

[0019] Where Y i For the high-dose low-noise image in a low-dose high-noise image-high-dose low-noise image pair, G i For Y i The trained image noise distribution estimation network estimates the image, where E is the objective function used by the image noise distribution estimation network, and Θ is the parameter in the objective function.

[0020] In step (2.2), the image estimated by the pre-trained image noise distribution estimation network is subtracted from the high dose of the low-noise image in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1), and the estimated noise distribution intensity is obtained by equation (II).

[0021]

[0022] Where X i For the high-dose low-noise image in a low-dose high-noise image-high-dose low-noise image pair, T i To train an image noise distribution estimation network to estimate images, D i Let represent the noise distribution intensity of the input image, where i is the index of the high-dose low-noise image in the low-dose high-noise image-high-dose low-noise image pair.

[0023] Preferably, step (3) specifically involves inputting the estimated noise distribution intensity and the low-dose high-noise image-high-dose low-noise image obtained in step (1) into the image restoration network for weighted fusion training, so that the image measurement objective function converges and an optimized image restoration network is obtained.

[0024] Preferably, the above image measurement objective function is the root mean square error measurement function, represented by equation (Ⅲ);

[0025]

[0026] H i For X i The output after image restoration network processing, N, is the total number of low-dose high-noise image-high-dose low-noise image pairs.

[0027] Preferably, the image noise distribution estimation network is a convolutional neural network, a generative adversarial network, or a diffusion generative network.

[0028] Preferably, the image restoration network is a convolutional neural network, a generative adversarial network, or a diffusion generative network.

[0029] Preferably, the objective function of the noise distribution estimation network is the root mean square error measure function, the L1 error measure function, or the distance divergence measure function.

[0030] Preferably, the above image measurement objective function is the root mean square error measurement function, the L1 error measurement function, or the distance divergence measurement function.

[0031] Preferably, step (1) above specifically involves acquiring multiple low-noise images, then adding Gaussian noise to the high-dose low-noise images to simulate and obtain corresponding low-dose high-noise images, and defining the high-dose low-noise images and the corresponding low-dose high-noise images as a pair of low-dose high-noise image-high-dose low-noise image pairs.

[0032] Another objective of this invention is to overcome the shortcomings of existing technologies and provide a CT reconstruction image noise removal system. This CT reconstruction image noise removal system significantly removes noise from the CT image to be processed, resulting in a high-quality CT image.

[0033] The above-mentioned objectives of the present invention are achieved through the following technical measures:

[0034] A CT reconstruction image noise removal system is provided, which employs the aforementioned CT reconstruction image noise removal method.

[0035] This invention discloses a method and system for noise removal in CT reconstructed images. The method includes the following steps: Step (1) acquiring multiple pairs of low-dose high-noise image-high-dose low-noise image pairs; Step (2) training an image noise distribution estimation network using the low-dose high-noise image-high-dose low-noise image pairs obtained in Step (1) to obtain an estimated noise distribution intensity; Step (3) inputting the estimated noise distribution intensity obtained in Step (2) and the low-dose high-noise image-high-dose low-noise image pairs obtained in Step (1) into an image restoration network for training to obtain an optimized image restoration network; Step (4) inputting the CT image to be processed into the optimized image restoration network obtained in Step (3) to obtain an optimized CT image. The noise removal method of this invention is simple and can significantly remove noise from CT images, resulting in high-quality CT images. Attached Figure Description

[0036] The invention will be further described with reference to the accompanying drawings, but the contents of the drawings do not constitute any limitation on the invention.

[0037] Figure 1 This is a flowchart of the CT reconstruction image noise removal method of the present invention.

[0038] Figure 2 This is the optimization result of Example 2. Detailed Implementation

[0039] The technical solution of the present invention will be further described in conjunction with the following embodiments.

[0040] Example 1

[0041] A method for noise removal in CT reconstructed images, such as Figure 1 As shown, it includes the following steps:

[0042] Step (1): Obtain multiple pairs of low-dose high-noise images and high-dose low-noise images;

[0043] Step (2): Train the image noise distribution estimation network using the low-dose high-noise image-high-dose low-noise image obtained in step (1) to obtain the estimated noise distribution intensity;

[0044] Step (3): The estimated noise distribution intensity obtained in step (2) and the low-dose high-noise image-high-dose low-noise image obtained in step (1) are input into the image restoration network to train the optimized image restoration network.

[0045] Step (4): Input the CT image to be processed into the optimized image recovery network obtained in step (3) to obtain the optimized CT image.

[0046] Each low-dose high-noise image-high-dose low-noise image pair includes one low-dose high-noise image and one high-dose low-noise image, with the dose of the low-dose high-noise image being lower than that of the high-dose low-noise image. In this embodiment, a normal dose high-dose high-noise image is, for example, 200 mAs; and a low dose low-dose high-noise image is, for example, 20 mAs.

[0047] Step (1) specifically involves acquiring multiple low-noise images, then adding Gaussian noise to the high-dose low-noise images to simulate and obtain the corresponding low-dose high-noise images. The high-dose low-noise images and the corresponding low-dose high-noise images are defined as a pair of low-dose high-noise images and high-dose low-noise images.

[0048] Step (2) specifically includes:

[0049] Step (2.1): Take the low-noise image high dose in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) as the label, and use the low-dose high-noise image as the input to train the image noise distribution estimation network, so that the objective function of the noise distribution estimation network converges, and finally obtain the image estimated by the pre-trained image noise distribution estimation network.

[0050] Step (2.2): The pre-trained image noise distribution estimation network estimates the image and compares it with the low-dose high-noise image in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) to obtain the estimated noise distribution intensity.

[0051] The objective function of the noise distribution estimation network is the root mean square error measure function, expressed by equation (Ⅰ).

[0052]

[0053] Where Y i For the high-dose low-noise image in a low-dose high-noise image-high-dose low-noise image pair, G i For Y i The trained image noise distribution estimation network estimates the image, where E is the objective function used by the image noise distribution estimation network, and Θ is the parameter in the objective function.

[0054] It should be noted that the objective function of the noise distribution estimation network of the present invention can be the root mean square error measure function, the L1 error measure function, or the distance divergence measure function. The image noise distribution estimation network of the present invention can be trained using supervised learning or unsupervised learning, and both supervised and unsupervised learning are conventional training techniques in the art, which should be known to those skilled in the art, and will not be described in detail here.

[0055] In step (2.2), the image estimated by the pre-trained image noise distribution estimation network is subtracted from the high dose of the low-noise image in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1), and the estimated noise distribution intensity is obtained by equation (II).

[0056]

[0057] Where X i For the high-dose low-noise image in a low-dose high-noise image-high-dose low-noise image pair, T i To train an image noise distribution estimation network to estimate images, D i Let represent the noise distribution intensity of the input image, where i is the index of the high-dose low-noise image in the low-dose high-noise image-high-dose low-noise image pair.

[0058] Specifically, step (3) involves inputting the estimated noise distribution intensity and the low-dose high-noise image-high-dose low-noise image obtained in step (1) into the image restoration network for weighted fusion training, so that the image measurement objective function converges and an optimized image restoration network is obtained.

[0059] The objective function for image measurement is the root mean square error measurement function, which is represented by equation (Ⅲ);

[0060]

[0061] H i For X i The output after image restoration network processing, N, is the total number of low-dose high-noise image-high-dose low-noise image pairs.

[0062] It should be noted that the image measurement objective function of the present invention can be the root mean square error measurement function, the L1 error measurement function, or the distance divergence measurement function.

[0063] It should be further explained that the objective function used in existing CT image restoration networks is the root mean square error measure function (i.e., removing D from equation (Ⅲ), without fusing noise intensity). Network training aims to learn a nonlinear function mapping from "low-dose, high-noise images" to "high-dose, low-noise images." Using the root mean square error measure function means that all image pairs have the same weight during network training. When the training data contains low-quality labeled data and artifacts caused by noise, the labeled images in the training image pairs may have different noise distributions and levels. In this case, using a simple root mean square error measure function can lead to biased and unstable network learning, because supervised learning networks aim to learn a nonlinear function mapping from "low-dose, high-noise images" to "high-dose, low-noise images." Low-quality labeled images may cause the network to overfit, and the network may learn artifact noise patterns in the labeled image data instead of the desired structural patterns of clean images. This will severely affect the performance of the image restoration network. Therefore, this invention redesigns the loss function, appropriately weighting the labeled data—that is, reducing the weight of labeled image data with high noise and increasing the weight of clean labeled image data—to improve the network's image restoration performance and stability. One feasible approach is to pre-train an image noise distribution estimation network using a subset of high-quality data, specifically low-dose high-noise image-normal-dose low-noise image pairs. This pre-trained network is then used to estimate the noise distribution in the target image, thereby weighting the training data—reducing the weight of labeled image data with high noise and increasing the weight of clean labeled image data. This design effectively addresses the problem of network degradation caused by low-quality training sets. It should be noted that the pre-trained image noise distribution estimation network may overestimate or underestimate the noise distribution in the target CT image, but due to the network's complex learning capabilities, the estimated noise distribution (Di) will be close to the actual noise distribution in the noisy image.

[0064] The image noise distribution estimation network of this invention is a convolutional neural network, a generative adversarial network, or a diffusion generative network. The image restoration network is a convolutional neural network, a generative adversarial network, or a diffusion generative network.

[0065] It should be noted that in the training of the noise distribution estimation network, the convergence of the objective function of the noise distribution estimation network and the convergence of the objective function of the image measurement network are both directly processed by the calculator. Moreover, the calculator automatically determines the state of convergence of the objective function of the noise distribution estimation network and the objective function of the image measurement network. Therefore, the convergence of the objective function of the noise distribution estimation network and the convergence of the objective function of the image measurement network are conventional techniques for those skilled in the art, and will not be described in detail here.

[0066] This CT reconstruction image noise removal method uses a low-dose high-noise image-high-dose low-noise image pair to estimate the noise distribution intensity and finally obtains an optimized image restoration network. By inputting the CT image into the optimized image restoration network, noise in the CT image is automatically and significantly removed, resulting in a high-quality CT image. Moreover, the CT reconstruction image noise removal method of this invention also has the advantage of simple steps.

[0067] Example 2

[0068] Application of a noise removal method for CT reconstructed images.

[0069] (1) Number of training images:

[0070] The image noise distribution estimation network uses 100 high-noise and low-noise images of each of the 5 groups of patients, each with a size of 512x512 pixels, which is a total of 250 pairs of low-dose high-noise image-high-dose low-noise image pairs, 500 images, where the high-dose low-noise images are normal-dose images.

[0071] The image restoration network used 250 pairs of low-dose high-noise images and high-dose low-noise images from the above 5 groups of patients; an additional 150 pairs of low-dose high-noise images and high-dose low-noise images from 3 groups of patients were added for training, and the high-dose low-noise images from the 3 groups of patients were medium-dose images, for a total of 400 image pairs and 800 images.

[0072] The 512×512 image is cropped into 128×128 image blocks. Network training is based on these 128×128 image blocks. The batch size of each input to the network is 4 blocks, meaning that the network uses 4 128×128 image blocks for training each iteration.

[0073] It should be noted that the image noise distribution estimation network in this embodiment needs to be trained in step (2) of embodiment 1. The pre-trained image noise distribution estimation network is directly used to perceive the noise in the training data (labeled image).

[0074] (2) Process parameter settings:

[0075] The total data size in the image noise distribution estimation network was 512×512×100×5 (i.e., 250 pairs of low-dose high-noise images and high-dose low-noise images from 5 patient groups). This data was cropped to a size of 128×128×16×100×5, where 16 of the 512×512 images were cropped into 16 128×128 image patches. The network was trained for 100 epochs, with 2000 iterations per epoch. The Adam optimizer was used, with an initial learning rate of 1e-4, which was reduced to 50% of its original value every 20 epochs.

[0076] The total data size in the image restoration network was 512×512×100×8 (i.e., 400 pairs of low-dose high-noise images and high-dose low-noise images from 8 patient groups). This data was cropped to a size of 128×128×16×100×8, where 16 of the 512x512 images were cropped into 16 128x128 image blocks. The network was trained for 100 epochs, with 3200 iterations per epoch, using the Adam optimizer. The initial learning rate was 1e-4, and the learning rate was reduced to 50% of its original value every 20 epochs. In total, data from 8 patient groups was used to train the network.

[0077] During the network testing in step (4), optimized CT images were obtained using two sets of patient head data, such as... Figure 2 .

[0078] from Figure 2 As can be seen, the results of the embodiment of noise removal of low-dose reconstructed image data of the human abdomen using the method of the present invention show that, compared with the results of the traditional filtered back projection method, the image quality of the present invention is superior and the noise removal effect is significant. The demonstrated example results show that the calculated peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) between the present invention and the reference image are significantly improved compared to the traditional filtered back projection method.

[0079] Example 3

[0080] A CT reconstruction image noise removal system employs the CT reconstruction image noise removal method as described in Example 1.

[0081] This CT reconstruction image noise removal system uses a low-dose high-noise image-high-dose low-noise image pair to estimate the noise distribution intensity and finally obtains an optimized image restoration network. By inputting the CT image into the optimized image restoration network, noise in the CT image is automatically and significantly removed, resulting in a high-quality CT image.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A CT reconstructed image noise removal method, characterized by, It comprises the following steps: Step (1), obtaining a plurality of pairs of low-dose high-noise images and high-dose low-noise images; Step (2), training the image noise distribution estimation network obtained in step (1) to obtain an estimated noise distribution intensity; Step (3), inputting the estimated noise distribution intensity obtained in step (2) and the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) into the image restoration network for training to obtain an optimized image restoration network; Step (4), inputting the CT image to be processed into the optimized image restoration network obtained in step (3) to obtain an optimized CT image; The step (2) specifically comprises: Step (2.1), training the image noise distribution estimation network by taking the low-noise image high dose in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) as a label and taking the low-dose high-noise image as an input, so that the noise distribution estimation network objective function converges, and finally an image estimated by a pre-trained image noise distribution estimation network is obtained; Step (2.2), obtaining an estimated noise distribution intensity by subtracting the pre-trained image noise distribution estimation network estimated image from the low-dose high-noise image-high-dose low-noise image pair obtained in step (1); The noise distribution estimation network objective function is a root mean square error measure function, which is represented by formula (I), … Formula (I); wherein is a high dose low noise image in a pair of a low dose high noise image - high dose low noise image, and estimates an image by the trained image noise distribution estimation network, is an objective function used by the image noise distribution estimation network, is a parameter in the objective function; In the step (2.2), the pre-trained image noise distribution estimation network estimated image is subtracted from the low-noise image high dose in the low-dose high-noise image-high-dose low-noise image pair obtained in step (1), which is represented by formula (II), to obtain the estimated noise distribution intensity; … Formula (II); wherein is a high dose low noise image of a low dose high noise image - high dose low noise image pair, is an image estimated by the training image noise distribution estimation network, is a noise distribution strength of the input image, wherein i is an index of a low noise image high dose of a low dose high noise image - high dose low noise image pair.

2. The CT reconstructed image noise removal method according to claim 1, characterized in that: Each pair of the low-dose high-noise image-high-dose low-noise image pair comprises a low-dose high-noise image and a high-dose low-noise image, and the dose of the low-dose high-noise image is lower than that of the high-dose low-noise image.

3. The CT reconstructed image noise removal method of claim 2, wherein: The step (3) specifically comprises inputting the estimated noise distribution intensity and the low-dose high-noise image-high-dose low-noise image pair obtained in step (1) into the image restoration network for weighted fusion training, so that the image measure objective function converges, and an optimized image restoration network is obtained.

4. The CT reconstructed image noise removal method of claim 3, wherein: The image measure objective function is a root mean square error measure function, which is represented by formula (III); … Formula (III); is the output after processing by the image restoration network, N is the total number of low-dose high-noise image - high-dose low-noise image pairs.

5. The CT reconstructed image noise removal method of claim 1, wherein: The image noise distribution estimation network is a convolutional neural network, a generative adversarial network or a diffusion generative network; The image restoration network is a convolutional neural network, a generative adversarial network or a diffusion generative network; The noise distribution estimation network objective function is a root mean square error measure function, L 1 error measure function or a distance divergence measure function; The image measure objective function is a root mean square error measure function, L 1 error measure function or a distance divergence measure function.

6. The CT reconstructed image noise removal method of claim 1, wherein: The step (1) specifically comprises obtaining a plurality of low-noise images, then adding Gaussian noise to the high-dose low-noise images to simulate corresponding low-dose high-noise images, and defining the high-dose low-noise images and the corresponding low-dose high-noise images as a pair of low-dose high-noise image-high-dose low-noise image pair.

7. A CT reconstructed image noise removal system characterized by: The CT reconstructed image noise removal method according to any one of claims 1 to 6 is adopted.

Images