A digital X-ray image denoising method based on trainable joint bilateral filter

Abstract

The present invention discloses a plug-and-play, trainable joint bilateral filtering processing method for low-dose digital X-ray imaging. First, a guide image of the joint bilateral filter is generated through a deep neural network, and then the joint bilateral filter is trained with the obtained guide image and the original noise image to finally obtain a denoised image. The present invention can achieve rapid generation of high-quality images while significantly reducing the X-ray dose, providing an effective denoising method for low-dose digital X-ray images.

Description

Technical Field

[0001] The invention relates to a digital X-ray image denoising method based on a trainable joint bilateral filter, and belongs to the field of computer image processing. Background Art

[0002] Digital X-ray imaging (DR) is an advanced X-ray photography technology formed by the combination of computer digital image processing technology and X-ray radiation technology. X-rays are ionizing radiation, and their excessive or inappropriate use may also damage normal organs. The rationalization of the use of X-ray imaging and the optimization of related radiation doses are essential to protecting the health of patients. However, there is a clear positive correlation between image quality and radiation dose. A reduction in radiation dose usually introduces noise and artifacts, reducing image quality; in addition, these noises will also affect the subsequent processing of DR images, such as image segmentation and recognition.

[0003] Commonly used DR image denoising methods, such as least squares filtering, mean filtering and bilateral filtering, can maintain the overall quality of the image under low-dose conditions, but it is difficult to make substantial changes to local appearance, texture and other details.

[0004] Recently, as deep learning technology has demonstrated powerful feature learning capabilities in various fields, the present invention combines traditional joint bilateral filtering with deep learning to address the existing problems and proposes a digital X-ray image denoising method based on a trainable joint bilateral filter. While performing fast filtering, it can restore the rich texture and edges of low-dose DR images. Summary of the invention

[0005] The object of the present invention is to provide a digital X-ray image denoising method based on a trainable joint bilateral filter, which is used for DR image denoising under low-dose conditions.

[0006] The present invention adopts the following technical solutions to solve the above problems:

[0007] The present invention provides a digital X-ray image denoising method based on a trainable joint bilateral filter, and the specific steps are as follows:

[0008] Step 1: Randomly read low-dose DR images of size 128×128 as the input of the convolutional neural network EDCNN, and the corresponding normal-dose images as labels to jointly train the EDCNN network;

[0009] Step 2: Use the trained EDCNN network to remove noise from the low-dose image to obtain a clearer image, i.e., the guidance map of the joint bilateral filter;

[0010] Specifically, the EDCNN network consists of 1 edge enhancement module and 8 convolutional blocks, such asFigure 2 As shown in the figure, the convolution block consists of the sequence of "1*1 convolution + LeakyReLU activation layer + 3*3 convolution + LeakyReLU activation layer", as shown in the figure. Figure 3 As shown. The edge enhancement module consists of 4 Sobel operators, including vertical, horizontal and 2 diagonal directions, namely:

[0011]

[0012] The EDCNN network optimizes parameters through the following loss function:

[0013]

[0014] Among them, F(x i ,θ) is the operation of the EDCNN model; x i is the pixel at the i-th position of the current low-dose image; θ is the trainable parameter of the model; y i is the pixel at the ith position corresponding to the normal dose label; N is the total number of pixels on the image.

[0015] Step 3: Use the obtained guide image and the original low-dose image as inputs of the trainable joint bilateral filter, and use the normal-dose image as a label to jointly train the joint bilateral filter.

[0016] Specifically, the proposed trainable joint bilateral filter includes a grayscale domain kernel, a trainable spatial domain kernel, and a traditional bilateral filter. The spatial domain kernel is a 3*3 matrix with an initial value of 1; the grayscale domain kernel is a constant with a value of 0.8. The calculation method of the joint bilateral filter is:

[0017]

[0018] Among them, I n is the original low-dose image; I f is the image after the joint bilateral filter; x is the pixel point of the current image; N(x) is the neighborhood point of x considered in the filtering operation; I g is the guide image; F(·) is the filter function used in the entire neighborhood; G(·) is the function used to calculate the spatial domain kernel, W g is the weight of G(·).

[0019] Step 4: Use the trained joint bilateral filter to remove noise from the low-dose image, and finally obtain a denoised high-quality image.

[0020] Specifically, the input of the trainable joint bilateral filter is the original low-dose image and the guide image, and the normal-dose image is used as a label. Its loss function is the same as the loss function of the EDCNN model. The value of the spatial domain kernel is iteratively updated by the optimizer. After the model parameter training is completed, when a low-dose DR image is input, a clear, high-quality DR image can be obtained.

[0021] Compared with the existing technology, the present invention utilizes the characteristics of deep neural networks that can extract rich image information, extracts key information from low-dose DR images, and generates clearer guide images, thereby improving the sensitivity of the joint bilateral filter to the details and texture of the image, helping to quickly and effectively remove noise and generate high-quality DR images.

[0022] Figure 1 It is a schematic diagram of the entire process of the present invention.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The method of the present invention first introduces the EDCNN network based on deep learning under digital X-ray images, and obtains an inference model with strong generalization ability after training with a large amount of data. Even if the new image is significantly different from the existing data, a preliminary denoised image, i.e., the guided image of the joint bilateral filter, can be obtained without manually setting any parameters.

[0025] 2. The method of the present invention constructs a denoising method based on a trainable joint bilateral filter for digital X-ray images. Compared with the traditional joint bilateral filter, the proposed trainable joint bilateral filter is combined with deep learning, which not only utilizes deep feature information, but also improves the restoration effect of details such as edges and textures while ensuring the overall denoising effect.

[0026] 3. Making the spatial domain kernel of the joint bilateral filter trainable can not only effectively utilize the advantages of the joint bilateral filter, but also iteratively update the parameters through the deep learning optimizer to find the most reasonable values, making the denoising algorithm of low-dose digital X-ray images more robust.

[0027] 4. After setting the initial value, the present invention does not need to adjust other parameters. It is an end-to-end, plug-and-play digital X-ray image denoising algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the entire process of the present invention.

[0029] Figure 2 Schematic diagram of the structure of the EDCNN neural network.

[0030] Figure 3Schematic diagram of the structure of the convolution block in the EDCNN neural network.

[0031] Figure 4 The noise reduction result of the present invention on a quarter-dose digital X-ray image (head) is shown.

[0032] Figure 5 This is the noise reduction result of the present invention on a quarter-dose digital X-ray image (ankle).

[0033] Figure 6 The overall flow chart of this method is shown in FIG. DETAILED DESCRIPTION

[0034] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0035] Example 1: See Figure 1 , a digital X-ray image denoising method based on a trainable joint bilateral filter, the specific steps are as follows:

[0036] Step 1: Randomly read low-dose digital X-ray images of size 128×128 as the input of the EDCNN network, and the corresponding normal-dose images as labels to jointly train the EDCNN network.

[0037] Step 2: Use the trained EDCNN network to preliminarily remove noise from the image and generate a guidance image for training the parameters of the joint bilateral filter.

[0038] Specifically, the EDCNN network consists of 1 edge enhancement module and 8 convolutional blocks, such as Figure 2 As shown in the figure, the convolution block consists of the sequence of "1*1 convolution + LeakyReLU activation layer + 3*3 convolution + LeakyReLU activation layer", as shown in the figure. Figure 3 As shown. The edge enhancement module consists of four Sobel operators in vertical, horizontal and two diagonal directions, namely:

[0039]

[0040] The EDCNN network optimizes parameters through the following loss function:

[0041]

[0042] Among them, F(x i ,θ) is the operation of the EDCNN model; x iis the pixel at the i-th position of the current low-dose image; θ is the trainable parameter of the model; y i is the pixel at the ith position corresponding to the normal dose label; N is the total number of pixels on the image.

[0043] In step 3, the obtained guidance image and the original low-dose image are used as inputs of the trainable joint bilateral filter, and the normal-dose image is used as a label to jointly train the joint bilateral filter.

[0044] Specifically, the proposed trainable joint bilateral filter includes a grayscale domain kernel, a trainable spatial domain kernel, and a traditional bilateral filter. The spatial domain kernel is a 3*3 matrix with an initial value of 1; the grayscale domain kernel is a constant with a value of 0.8. The calculation method of the joint bilateral filter is:

[0045]

[0046] Among them, I n is the original low-dose image; I f is the image after the joint bilateral filter; x is the pixel point of the current image; N(x) is the neighborhood point of x considered in the filtering operation; I g is the guide image; F(·) is the filter function used in the entire neighborhood; G(·) is the function used to calculate the spatial domain kernel, W g is the weight of G(·).

[0047] Step 4: Use the trained joint bilateral filter to remove noise from the low-dose image and finally obtain a high-quality image.

[0048] Specifically, the input of the trainable joint bilateral filter is the original low-dose image and the guide image, and the normal-dose image is used as a label. Its loss function is the same as the loss function of the EDCNN model. The value of the spatial domain kernel is iteratively updated by the optimizer. After the model parameter training is completed, when a low-dose DR image is input, a clear, high-quality DR image can be obtained.

[0049] Effect evaluation:

[0050] The present invention invents a digital X-ray image denoising method based on a trainable joint bilateral filter, which can quickly generate high-quality images while significantly reducing the X-ray dose. Compared with the existing methods, the denoising effect of the present invention has been improved in terms of PSNR and SSIM indicators, and the denoised image has a better visual effect.

[0051] Table 1. Denoising effect of the present invention on the ChestX-ray14 dataset

[0052] method PSNR(dB) SSIM(%) BM3D (Existing Method 1) 32.46 0.9205 DnCNN (Existing Method 2) 32.97 0.9253 Method of the present invention 33.48 0.9389

[0053] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Equivalent replacements or substitutions made on the basis of the above technical solutions all fall within the protection scope of the present invention.

Claims

1. A digital X-ray image denoising method based on a trainable joint bilateral filter, characterized in that: The specific steps of the method are as follows: Step 1: randomly read low-dose images and corresponding normal-dose images of size 128×128, where the low-dose images are used as inputs of the convolutional neural network EDCNN, and the normal-dose images are used as labels to jointly train the EDCNN network; Step 2: Use the trained EDCNN network to remove noise from the low-dose image to obtain a clearer image, i.e., the guidance map of the joint bilateral filter; Step 3, using the obtained guide image and the original low-dose image as inputs of the trainable joint bilateral filter, and the normal-dose image as a label, to jointly train the joint bilateral filter; Step 4: Use the trained joint bilateral filter to remove noise from the low-dose image, and finally obtain a denoised high-quality image; Among them, the EDCNN network that needs to be trained in step 1 includes 1 edge enhancement module and 8 convolution blocks, among which, The convolution block is composed of "1*1 convolution + LeakyReLU activation layer + 3*3 convolution + LeakyReLU activation layer" in sequence; the edge enhancement module is composed of 4 Sobel operators, including vertical, horizontal and 2 diagonal directions; The loss function L of the EDCNN network to be trained in step 1 mse for: Among them, F(x i ,θ) is the operation of the EDCNN model; x i is the pixel at the i-th position of the current low-dose image; θ is the trainable parameter of the model; y i is the pixel at the ith position corresponding to the normal dose label; N is the total number of pixels on the image; The EDCNN model parameters in step 2 are iteratively updated by the optimizer. The joint bilateral filter in step 3 includes a gray domain kernel, a trainable spatial domain kernel and a traditional bilateral filter. The spatial domain kernel is a 3*3 matrix with an initial value of 1; the gray domain kernel is a constant with a value of 0.

8. The calculation method of the joint bilateral filter is: Among them, I n is the original low-dose image; I f is the image after the joint bilateral filter; x is the pixel point of the current image; N(x) is the neighborhood point of x considered in the filtering operation; I g is the guide image; F(·) is the filter function used in the entire neighborhood; G(·) is the function used to calculate the spatial domain kernel, W g is the weight of G(·), The input of the joint bilateral filter in step 4 is the original low-dose image and the guidance image, and the normal-dose image is used as the label. Its loss function is The value of its spatial domain kernel is iteratively updated by the optimizer.

Images