A method for analyzing low-dose X-ray multi-contrast high-precision signals

Abstract

The present invention provides a method for analyzing low-dose X-ray multi-contrast high-precision signals, belonging to the technical fields of deep learning and X-ray grating contrast imaging. It includes: obtaining a grating stepped projection sequence under low-dose conditions; constructing a convolutional neural network model, which consists of a contrast analysis module and a U-net. The contrast analysis module is used to extract contrast signals, and the U-net is used for further image enhancement and noise removal; training the convolutional neural network model; using the trained convolutional neural network model to obtain high-quality absorption, phase, and dark-field projections from the low-dose grating stepped projection sequence. The present invention uses a convolutional neural network to extract physical information from the grating stepped projection sequence, which can reduce the influence of noise introduced by low-dose X-rays on contrast signals, effectively reduce the radiation dose received by the sample, and significantly improve the imaging quality, thus enhancing the application potential of X-ray grating differential phase contrast imaging.

Description

Technical Field

[0001] The present invention belongs to the technical fields of deep learning and X-ray grating contrast imaging, and particularly relates to a method for analyzing low-dose X-ray multi-contrast high-precision signals. Background Art

[0002] Compared with traditional absorption contrast imaging, the grating interference imaging technology based on the Talbot-Lau effect can obtain absorption, phase, and dark-field contrast images of an object. The absorption image has good imaging effects for heavy elements such as metals, bones, and teeth. Differential phase contrast imaging can obtain higher imaging contrast for light-element substances such as biological tissues. The dark-field contrast has the best imaging effect for low-density tissues, such as the human lungs. The grating interference imaging technology based on the Talbot-Lau effect can use an ordinary X-ray source to generate coherent X-rays through the source grating G0, pass through the phase grating G1 and freely propagate a certain distance, pass through the absorption grating G2, and finally be received by the detector. However, since the grating will absorb most of the X-rays and the absorption grating G2 needs to be moved multiple times (usually 4-8 times) to complete the extraction of absorption, phase, and dark-field contrast signals, the radiation dose during the imaging process is greatly increased.

[0003] Reducing the tube current is the simplest and easiest measure to reduce the radiation dose. However, reducing the radiation dose will introduce quantum noise into the projection data. And since the phase extraction of the stepped projection data needs to use the fast Fourier transform, as long as a pixel point in a certain projection sequence image is noisy, the noise will be introduced into the corresponding pixel point of the phase-contrast projection map, causing significant image degradation to the differential phase-contrast projection. Not only will the local fine structures be damaged by the noise, but the overall structure will also be affected by the noise until it disappears. Summary of the Invention

[0004] In order to overcome the above technical problems, the present invention provides a method for analyzing low-dose X-ray multi-contrast high-precision signals. By reducing the tube current to reduce the radiation dose, and at the same time using deep learning technology to achieve contrast analysis and image enhancement of low-dose projections. The convolutional neural network can learn the mapping between low-dose projection images and high-quality projection images. While completing the contrast extraction, the image details damaged by the noise are restored, ensuring the quality of contrast analysis under low-dose conditions and promoting the application of the grating differential phase-contrast imaging technology in the clinical field.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for analyzing low-dose X-ray multi-contrast high-precision signals, comprising the following steps:

[0007] Step S101: Use an X-ray grating differential phase contrast imaging device based on the Talbot-Lau effect to obtain at least three step-projection sequences including the sample and at least three step-projection sequences including the background under normal tube current conditions, and obtain at least three step-projection sequences including the sample and at least three step-projection sequences including the background under low tube current conditions;

[0008] Step S102: Construct a convolutional neural network model. The network model includes a cascaded contrast analysis module and a U-net module. The contrast analysis module is used to extract contrast signals, and the U-net module is used to enhance the extracted contrast signals and remove noise;

[0009] Step S103: Use the at least three step-projection sequences including the sample and at least three step-projection sequences including the background under the low tube current conditions obtained as the input of the network model, perform Fourier analysis on the at least three step-projection sequences including the sample and at least three step-projection sequences including the background under the normal tube current conditions, and use the obtained absorption, phase, and dark field images as labels to train the convolutional neural network model;

[0010] Step S104: Input the at least three step-projection sequences including the sample and at least three step-projection sequences including the background under the low tube current conditions into the trained network model to obtain high-precision absorption, phase, and dark field contrast images.

[0011] The beneficial effects of the present invention are as follows:

[0012] Compared with the existing X-ray grating contrast signal analysis methods, the present invention can effectively reduce the radiation dose received by the sample and enhance the application potential of the imaging technology based on the Talbot-Lau effect; at the same time, it fully exploits the continuous physical information in the step-projection sequences, thereby effectively enhancing the contrast images affected by low-dose noise. It avoids the loss of details and local structure damage of the phase-contrast images caused by using the fast Fourier transform to analyze the low-dose step-projection sequences, and improves the final imaging quality.

[0013] The low-dose X-ray multi-contrast high-precision signal analysis method in the present invention does not require modification of the existing imaging device, has strong scalability, and can effectively reduce the radiation dose. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flowchart of a low-dose X-ray multi-contrast high-precision signal analysis method of the present invention;

[0015] Figure 2 is a schematic diagram of the imaging principle based on the Talbot-Lau grating interferometer;

[0016] Figure 3 Three-dimensional reconstruction diagram of the mice used in the experiment;

[0017] Figure 4 Convolutional neural network structure diagram adopted in the embodiment of the present invention;

[0018] Figure 5 Comparison diagram of the implementation effects of the present invention and the prior art, where (a) is the absorption, phase, and dark-field contrast diagrams of Fourier analysis at the standard dose, (b) is the absorption, phase, and dark-field contrast diagrams of Fourier analysis at one-eighth dose, and (c) is the absorption, phase, and dark-field contrast diagrams obtained by the present invention at one-eighth dose. Specific implementation manners

[0019] Next, the technical solutions in this embodiment will be clearly and completely described in conjunction with the accompanying drawings in this embodiment. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0020] As Figure 1 shown, this embodiment discloses a low-dose X-ray multi-contrast high-precision signal analysis method, including the following steps:

[0021] Step S101, obtaining a stepped projection sequence: using an X-ray grating differential phase contrast imaging device based on the Talbot-Lau effect, obtaining at least three stepped projection sequences including samples and at least three stepped projection sequences including backgrounds under normal tube current conditions, and obtaining at least three stepped projection sequences including samples and at least three stepped projection sequences including backgrounds under low tube current conditions, as follows:

[0022] As Figure 2 shown, when X-rays penetrate an object, absorption attenuation, phase shift, and scattering will occur. The X-ray grating differential phase contrast imaging device includes six parts: an X-ray source, a source grating G0, a sample, a phase grating G1, an absorption grating G2, and a detector;

[0023] Among them, the source grating G0 is used to generate coherent X-rays; the duty cycle of the phase grating G1 is 50%, and it uses X-rays to generate a phase shift of π; the duty cycle of the absorption grating G2 is 50%, where a part can completely absorb X-rays and another part can transmit X-rays, and the detector is used to convert X-ray energy into a digital image;

[0024] The corresponding relationship between the imaging experimental parameters of the X-ray grating differential phase contrast imaging device is as follows:

[0025] (1)

[0026] (2)

[0027] (3)

[0028] (4)

[0029] Wherein, d is the distance between the phase grating G1 and the absorption grating G2, m is an integer representing m times the fractional Talbot distance, k = (L + d) / L is the magnification ratio, g1 is the period of the phase grating G1, λ is the wavelength of the X-ray, g2 is the period of the absorption grating G2, g0 is the period of the source grating G0, L is the distance between the source grating G0 and the phase grating G1, and s is the width that allows the X-ray to pass through in each period of the source grating G0;

[0030] During the imaging process, at least three projection sequences are obtained by stepping the absorption grating, and then the curve of the light intensity varying with the position is obtained through fast Fourier transform, which is called the stepping curve. The grating stepping curve can be approximated as:

[0031] (5)

[0032] Wherein, x is the independent variable representing the position of the absorption grating, f(x) represents the X-ray intensity corresponding to the absorption grating at the position x, a0 is the constant term, a1 is the amplitude, and φ1 is the initial phase. Subsequently, using the formula:

[0033] (6)

[0034] (7)

[0035] (8)

[0036] In the formula, A represents the absorption contrast image, and respectively represent the constant terms of the sample and the background in formula (5), φ represents the phase contrast image, and respectively represent the initial phases of the sample and the background in formula (5), V represents the dark field contrast image, v s and v r are respectively the dark field contrast images of the sample and the background, and respectively represent the amplitudes of the sample and the background in formula (5).

[0037] The data is obtained by taking three steps through an absorption grating under low tube current and normal tube current conditions respectively, and four projection images of the sample and the background and corresponding high-quality absorption, phase, and dark-field contrast projections are obtained respectively. Figure 3 It is a three-dimensional reconstruction diagram of the mouse used in the experiment.

[0038] Step S102, constructing a convolutional neural network model: The network model includes a contrast parsing module and a U-net module connected in series. The contrast parsing module is used to extract contrast signals, and the U-net module is used to enhance the extracted contrast signals and remove noise.

[0039] As Figure 4 shown, the contrast parsing network consists of the following three parts: 1) Inception block 1, a series of convolutional layers form the first Inception block. Most of the convolutional layers use 1×1Conv, which can introduce non-linear transformation while reducing the channel dimension and improve the expression ability of the network. It can be observed that this Inception block receives features from 5 branches, which are the 8-channel feature layer of the first layer, the 16-channel feature layer of the second layer, the 32-channel feature layer of the third layer, the 32-channel 1×1Conv feature layer and 32-channel 1×3Conv feature layer of the fourth layer, and the 64-channel 1×1 convolutional layer and 64-channel 3×1 convolutional layer of the fifth layer, and merges them. 2) ResNet, the role of introducing ResNet into two Inception blocks is to introduce residual connections, allowing information to directly skip some levels in the network, so as to better spread gradients and information. This can solve the problem of gradient disappearance, enable the network to be stacked deeper, and better capture complex features in the image. The residual block consists of a 128-channel 3×3 convolutional layer and a 64-channel 1×1 convolutional layer. 3) Inception block 2, further expanding the representation ability on the basis of the first Inception block. Similar to the first Inception block, it contains multiple parallel convolutional layers. Specifically, this Inception block receives features from 3 branches, which are three 64-channel 1×1Conv layers of the 11th layer, two 32-channel 1×1Conv layers and a 16-channel 1Conv layer of the 12th layer and merges them, and finally outputs a preliminary differential phase contrast projection image. The contrast parsing module and the U-net are connected in series to construct a convolutional neural network, where the contrast parsing module consists of an Inception module and a residual module, and is used for the extraction of phase contrast.

[0040] Furthermore, the U-net described in step 2 mainly includes: 1) An encoder. The main function of the encoder is to convert the input image into a low-resolution, high-level feature representation and capture the global context information in the image. As shown in the purple rectangular box in the figure, the input image reduces the resolution through a series of convolutional layers and extracts increasingly abstract feature representations. It is cascaded by 5 consecutive Conv layers with a convolution kernel of 3×3 and a channel number of 16, 32, 64, 128, and 256, and a convolution layer with a stride of 2, resulting in five feature representations of different scales. 2) A decoder. The main function of the decoder is to utilize the multi-scale features extracted by the encoder and retain more detailed information while restoring the resolution to obtain a more accurate image processing result. As shown in the figure, the decoder accepts the output feature map of the encoder part and gradually restores the resolution of the projected image through a series of upsampling and convolution operations. It is cascaded by 4 consecutive Conv layers with a convolution kernel of 3×3 and a channel number of 128, 64, 32, and 16, and an upsampling convolution layer with a stride of 2. 3) Skip connections. Skip connections help solve the problems of local structure loss and resolution loss in the image enhancement task and improve the context awareness ability and accuracy of the projection enhancement network. The skip connections act between the encoder and the decoder, connecting the features of the encoder with the corresponding layers of the decoder, which can transfer low-level image feature information to the decoder to assist it in further restoring the local structure of the image and improving the image enhancement ability.

[0041] It should be noted that the contrast analysis network and the projection enhancement network form an overall deep convolutional network and are trained jointly, rather than a cascade of separately trained network models. Therefore, there is no logical sequence between contrast analysis and projection enhancement, but they jointly serve the goal of outputting a differential phase-contrast projection image with correct structure and complete information.

[0042] Step S103, training a convolutional neural network model based on the step projection sequence: Using at least three step projection sequences including the sample under the low tube current condition and at least three step projection sequences including the background as the input of the network model, performing Fourier analysis on at least three step projection sequences including the sample under the normal tube current condition and at least three step projection sequences including the background, and using the obtained absorption, phase, and dark field images as labels to train the convolutional neural network model.

[0043] Step S104, obtaining a high-precision contrast image using the trained network model: Inputting at least three step projection sequences including the sample under the low tube current condition and at least three step projection sequences including the background into the trained network model to obtain high-precision absorption, phase, and dark field contrast images.

[0044] Figure 5Among them, (a) is the absorption, phase, and dark-field contrast maps of the standard-dose Fourier analysis, (b) is the absorption, phase, and dark-field contrast maps of the one-eighth-dose Fourier analysis, and (c) is the absorption, phase, and dark-field contrast maps obtained by the one-eighth-dose of the present invention. As can be seen from the figure, the examples of the present invention have absolute advantages in terms of noise reduction, artifact suppression, or structure restoration. This is due to the presence of the contrast analysis module, which enables the network to effectively extract the continuous physical information in the step projection sequence, form the mutual enhancement of the effective information between the projection sequences, and thus restore the detailed information damaged due to the low tube current.

[0045] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for analyzing low-dose X-ray multi-contrast high-precision signals, characterized in that It includes the following steps: Step S101: Use an X-ray grating differential phase contrast imaging device based on the Talbot-Lau effect to obtain at least three step projection sequences including the sample and at least three step projection sequences including the background under normal tube current conditions, and obtain at least three step projection sequences including the sample and at least three step projection sequences including the background under low tube current conditions; Step S102: Construct a convolutional neural network model. The network model includes a serially connected contrast parsing module and a U-net module. The contrast parsing module is used to extract contrast signals, and the U-net module is used to enhance and remove noise from the extracted contrast signals; the contrast parsing module includes a serially connected first Inception block, a ResNet block, and a second Inception block. The first Inception block includes several parallel convolutional layers; the ResNet block is used to introduce residual connections; the second Inception block includes several parallel convolutional layers; Step S103: Use the at least three step projection sequences including the sample and at least three step projection sequences including the background under the low tube current conditions obtained as the input of the network model, perform Fourier analysis on the at least three step projection sequences including the sample and at least three step projection sequences including the background under the normal tube current conditions, and use the obtained absorption, phase, and dark field images as labels to train the convolutional neural network model; Step S104: Input the at least three step projection sequences including the sample and at least three step projection sequences including the background under the low tube current conditions into the trained network model to obtain high-precision absorption, phase, and dark field contrast images.

2. The method for analyzing a low-dose X-ray multi-contrast high-precision signal according to claim 1, wherein In step S101, the X-ray grating differential phase contrast imaging device includes an X-ray source, a source grating, a sample, a phase grating, an absorption grating, and a detector arranged in sequence along the optical path; The corresponding relationships between the imaging experimental parameters of the X-ray grating differential phase contrast imaging device are as follows: (1) (2) (3) (4) Where d is the distance between the phase grating and the absorption grating, m is an integer representing m times the fractional Talbot distance, k = (L + d) / L is the magnification ratio, g1 is the period of the phase grating, λ is the wavelength of the X-ray, g2 is the period of the absorption grating, g0 is the period of the source grating, L is the distance between the source grating and the phase grating, and s is the width of the source grating that allows the X-ray to pass through in each period; Under the conditions of low tube current and normal tube current respectively, step through the absorption grating to obtain at least three step projection sequences including the sample and at least three step projection sequences including the background.

3. A method for analyzing a low-dose X-ray multi-contrast high-precision signal according to claim 1, characterized in that, The U-net module includes an encoder and a decoder, and the encoder and the decoder connect their corresponding layers in a skip connection manner.

4. A method for analyzing low-dose X-ray multi-contrast high-precision signals according to claim 1, characterized in that, In step S103, during the imaging process, step through the absorption grating to obtain at least three step projection sequences including the sample and at least three step projection sequences including the background under normal tube current conditions, and obtain a step curve of the light intensity changing with position through fast Fourier transform: (5) Among them, x is the independent variable, representing the position of the absorption grating, f(x) represents the X-ray intensity corresponding to the absorption grating at position x, a0 is the constant term, a1 is the amplitude, and φ1 is the initial phase. The absorption, phase, and dark-field contrast images of the sample and the background are obtained using the following formulas respectively: (6) (7) (8) In the formula, A represents the absorption contrast image, and represent the constant terms of the sample and the background in formula (5) respectively, φ represents the phase contrast image, and represent the initial phases of the sample and the background in formula (5) respectively, V represents the dark field contrast image, v s and v r are the dark field contrast images of the sample and the background respectively, and represent the amplitudes of the sample and the background in formula (5) respectively; Taking the obtained absorption contrast image A, phase contrast image φ, and dark-field contrast image V as labels, the convolutional neural network model is trained.

Images