A Reconfigurable X-ray Imaging Method Based on Two-dimensional Layered Materials

Abstract

This invention discloses a reconfigurable X-ray imaging method based on two-dimensional layered materials, capable of simultaneous image sensing and processing. This method, for the first time, applies two-dimensional layered materials to radiation imaging. Multiple layers of two-dimensional materials are combined via van der Waals (vdW) vertical heterostructures, maintaining van der Waals forces between each layer. In-plane atoms form a stable lattice structure through covalent bonding, ensuring atomic-level flatness and the absence of dangling bonds. The bias voltage of each pixel is then adjusted to achieve different light responses. Through image acquisition and processing, a reconfigurable visual sensor can be realized.

Description

Technical Field

[0001] The present invention relates to the field of radiation detection technology and radiation imaging, and in particular to a reconfigurable X-ray imaging method based on two-dimensional layered materials. Background Art

[0002] Existing X-ray imaging devices are generally as follows Figure 1 As shown, X-ray imaging is currently applicable in industrial inspection, medical imaging, security inspection, and other fields, and can achieve high spatial resolution. Current technical solutions for high-speed X-ray image detection all rely on first acquiring the X-ray image with a pixelated X-ray detector, followed by signal processing, image reconstruction, and image recognition. Although the time interval between an X-ray image entering the detector and being read out by the front-end electronics is less than 1 μs, the time delay increases to the millisecond level (under continuous measurement conditions) after the large number of pixelated signals undergo analog-to-digital conversion, digital signal processing, image reconstruction, and image recognition. Limited by signal processing bandwidth and PC computing power, the processing speed of X-ray image detection has reached the upper limit of current technology. New electronic technologies and software algorithms can optimize image processing speed, but substantial increases are inherently difficult. In summary, to address the current bottlenecks in high-speed X-ray image detection, we must explore new research approaches and approach the problem from a different perspective, hoping to find inspiration for solutions. Summary of the Invention

[0003] This method transplants the machine learning algorithm onto a two-dimensional layered material pixel array detector, which can simultaneously realize image sensing and processing. The main technical process is as follows:

[0004] First of all, the traditional radiation imaging process separates image sensing and processing. A large amount of image data information is converted from analog to digital, and then processed by the computer before image reconstruction can be performed. However, a large amount of data information that has undergone analog-to-digital conversion is redundant. Therefore, when processed on the software side, this redundant information will greatly waste computing resources, thereby greatly delaying the image processing speed. In addition, the implementation process of traditional computer neural networks is to pass the input data through layers of neurons in the middle to the output layer. In this process, the input data is continuously amplified and reduced, and nonlinearly processed through weights and activation functions between each neuron and between the neuron and the input data. Finally, after passing through multiple layers of neurons, the classification results are displayed on the output layer.

[0005] Therefore, this method attempts to convert the trained neural network weights originally stored in the computer into physical entities directly related to the input data, and realizes the simultaneous image perception and processing through a two-dimensional layered material detection matrix, shortening the processing speed on the computer side, thereby greatly improving imaging efficiency.

[0006] The specific method is as follows:

[0007] The first step is to simulate the X-ray imaging process using the Monte Carlo method. This method models the process of X-rays passing through objects of different fixed shapes. This method simulates the imaging process of objects of different shapes within a two-dimensional layered material detection plane, and obtains the different response matrices formed in the detector for objects of different shapes, materials, and structures. Using a Monte Carlo program to simulate the X-ray imaging process overcomes the difficulty of quickly acquiring large amounts of data samples under real-world measurement conditions. By simulating X-ray imaging processes under various conditions, deep learning can be given stronger learning and prediction capabilities.

[0008] Furthermore, the Monte Carlo program adopts one or more of MORSE, MCNP, EGS, GEANT4, FLUKA, SuperMC, Phits or GADRAS.

[0009] Furthermore, the two-dimensional layered material can be one or more of GaN, WSe2, MoS2, BN, SiC, and MoTe2.

[0010] Furthermore, its two-dimensional layered detection material interacts with each other through van der Waals forces, and forms a stable lattice structure through atomic covalent bonds within the plane, maintaining its atomic-level flatness and no dangling bonds between layers, and ensuring bipolar field effect characteristics.

[0011] In the second step, the training image dataset obtained in the first step is used as a sample to train and test the deep learning network model. First, the obtained training images are normalized and divided into training set and test set. The training set is input into the deep learning network model for training, and the training effect is tested with the test set, so as to obtain the optimal reconstructible parameters corresponding to each pixel of different radiation images.

[0012] Furthermore, the deep learning framework adopted by the deep learning network is one or more of DeepLearnToolbox, Caffe, CNTK, TensorFlow, Theano, Torch, Keras, Lasagne, DSSTNE, MXNet, DeepLearning4J, ConvNetJS, Chainer, and Scikit-Learn.

[0013] Furthermore, the deep learning network model uses one or more of the deep belief network, deep neural network, convolutional neural network, stacked autoencoder or convolutional autoencoder. The neural network needs to be built according to the actual detector pixel size and the classification content of the target. Since the final trained weights need to be converted into bias voltages of different pixel matrices, there can only be one layer of intermediate neurons at most. Otherwise, the correspondence between the imaged object and the detector pixel cannot be established. Therefore, when building the network, the following is finally selected: Figure 4 The structure shown in the figure is as follows, where P represents the weight parameter, C represents the classification output result, m is the number of categories, and n is the one-dimensional input data. All inputs and neurons are fully connected. The activation function of the neural network is softmax, and the loss function is softmaxloss. These two functions have good performance in multi-classification problems in machine learning. At the same time, the stochastic gradient descent method is used for training. After 5000 iterations, the loss value stabilized at around 0.75, and the final verification algorithm accuracy was 96%.

[0014] In the third step, the obtained reconfigurable parameters are converted into the matrix bias of the two-dimensional material through a formula, and the bias of the detector response matrix is set through specific equipment. Figure 4 The probe layout of the pixelated detector under this network is also shown. The figure shows the situation when the input data is divided into 4 categories. Based on this, four probe pixels need to be grouped into a group, that is, the four pixels of red, yellow, orange and green in the figure. At the same time, this group of four detector pixels corresponds to one pixel of the input image, so Figure 3 The left figures correspond to each other, and based on this, a one-to-one correspondence between weights and detector pixels can be constructed. It should be noted that the weights trained by the network must be normalized at the end. The blocking ability of a coefficient of 0 to rays is 100%, and the blocking ability of a coefficient of 1 is 0. And so on, different biases are set according to the size of the weights.

[0015] Furthermore, the reconfigurable detection matrix is characterized in that different responses to X-rays are achieved by applying different specific voltages at both ends of the detection material with the help of a specific structure.

[0016] The fourth step is accuracy verification. Objects of varying shapes and materials are randomly selected and irradiated with the same X-rays used in the Monte Carlo simulation process. A pixelated detection matrix with a biased voltage is used to perform preliminary image feature extraction. The preprocessed image data is then transferred to the software for image reconstruction. The reconstructed image is then registered with the original image to verify the accuracy of the imaging method.

[0017] Furthermore, the image data acquisition feature is characterized by: one or a combination of ASIC chip, FPGA chip, PCB board, ARM board, Windows motherboard, wireless radio frequency module, and GPRS data transmission module.

[0018] Furthermore, the image reconstruction module is characterized in that the image data is reconstructed using a trained network model.

[0019] Furthermore, the image registration work adopts one or more of the methods selected from the group consisting of a grayscale information-based method, a spatial transformation domain-based method, and an image feature-based method.

[0020] Furthermore, the computing platform used for image registration is one or more of OpenCV computer vision library, ArcGIS or MATLAB.

[0021] Compared with the existing technology, this method has the following advantages and effects:

[0022] (1) This method applies two-dimensional layered materials to radiation detection, giving full play to the excellent optoelectronic properties of two-dimensional layered materials and overcoming the disadvantage that images cannot be sensed and processed simultaneously in traditional radiation imaging.

[0023] (2) This method can form reconfigurable radiation detection devices with different functions by arranging and combining different two-dimensional layered materials.

[0024] (3) This method can achieve different image processing functions by setting the detector bias matrix, greatly reducing redundant image data and shortening imaging time.

[0025] (4) This method is expected to contribute to the field of high-speed X-ray imaging and reveal more dynamic scientific problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 For existing X-ray imaging devices and imaging processes

[0027] Figure 2 This is the main technical process of the present invention

[0028] Figure 3 The target object of X-ray imaging

[0029] Figure 4 The neural network structure (left) and the corresponding relationship between the detection pixel matrix and the weight (right) DETAILED DESCRIPTION

[0030] The specific embodiments of the present invention are further described below in conjunction with the accompanying drawings. The following embodiments are intended to explain the present invention, but the present invention is not limited to the following embodiments.

[0031] See also Figure 2 , an X-ray imaging method based on two-dimensional layered materials, comprising the following steps:

[0032] Step 101: Use the Monte Carlo method to simulate the X-ray imaging process to obtain a sufficient number of sample images.

[0033] Specifically, the Monte Carlo method is used to model the X-ray imaging process, simulating the process of objects of different shapes, materials, and structures being irradiated by radiation, thereby obtaining radiation images of different types of objects. The Monte Carlo method, also known as random sampling or statistical experimental methods, is a branch of computational mathematics that can realistically simulate actual physical processes, solving problems that are highly consistent with reality. Monte Carlo programs are general-purpose software packages based on the Monte Carlo method for calculating neutron, photon, electron, or coupled neutron / photon / electron transport problems in complex three-dimensional geometric structures. Based on parameters such as actual detector size and material composition, Monte Carlo programs can be used to model these problems and determine their responses to neutrons, photons, and electrons. Using Monte Carlo programs to simulate the coded aperture camera imaging process can overcome the difficulty of obtaining a large number of data samples in a short period of time under real-world conditions. By simulating the coded imaging process under various conditions, deep learning can be enhanced in terms of learning and prediction capabilities. The Monte Carlo program uses one or more of the following: MORSE, MCNP, EGS, GEANT4, FLUKA, SuperMC, Phits, or GADRAS.

[0034] Step 102: Use the radiation image described in step 101 as a sample to train and test the deep learning network model.

[0035] Specifically, the obtained training images are first normalized and divided into training sets and test sets. The training set is input into the deep learning network model for training, and the training effect is tested with the test set to obtain the optimal reconstructible parameters corresponding to different radiation images and each pixel. The deep learning framework used by the deep learning network is one or more of DeepLearnToolbox, Caffe, CNTK, TensorFlow, Theano, Torch, Keras, Lasagne, DSSTNE, MXNet, DeepLearning4J, ConvNetJS, Chainer, and Scikit-Learn. The deep learning network model uses one or more of deep belief networks, deep neural networks, convolutional neural networks, stacked autoencoders, or convolutional autoencoders. The neural network needs to be built according to the pixel size of the actual detector and the classification content of the target. Since the final trained weights need to be converted into biases of different pixel matrices, there can only be one layer of neurons in the middle layer, otherwise the correspondence between the imaged object and the detector pixels cannot be established. Therefore, when constructing the network, the following is finally selected: Figure 4 The structure shown in the figure is as follows, where P represents the weight parameter, C represents the classification output result, m is the number of categories, and n is the one-dimensional input data. All inputs and neurons are fully connected. The activation function of the neural network is softmax, and the loss function is softmaxloss. These two functions have good performance in multi-classification problems in machine learning. At the same time, the stochastic gradient descent method is used for training. After 5000 iterations, the loss value stabilized at around 0.75, and the final verification algorithm accuracy was 96%.

[0036] Step 103: Convert the trained weight parameters into bias voltages of the pixel detection matrix.

[0037] Specifically, the conversion formula is as follows

[0038]

[0039] Where n = 0.1; β = 0.5V; P: input light intensity; δ k Update weights of neural network direction propagation; g i,j The pixel position in row i and column j is used to set the bias voltage of the detector response matrix through a specific structure. Figure 4 The probe layout of the pixelated detector under this network is also shown. The figure shows the situation when the input data is divided into 4 categories. Based on this, four probe pixels need to be grouped into a group, that is, the four pixels of red, yellow, orange and green in the figure. At the same time, this group of four detector pixels corresponds to one pixel of the input image, so Figure 3The left figures correspond to each other, and based on this, a one-to-one correspondence between weights and detector pixels can be constructed. It should be noted that the weights trained by the network must be normalized at the end. The blocking ability of a coefficient of 0 to rays is 100%, and the blocking ability of a coefficient of 1 is 0. And so on, different biases are set according to the size of the weights.

[0040] Step 104: Randomly select objects of different shapes, use the same rays and position distribution as the Monte Carlo simulation, and minimize environmental interference, and image them through the set two-dimensional material detection matrix. After the image is collected, it is input into the trained neural network model for image reconstruction, and the reconstructed image is aligned with the original image to verify the feasibility and accuracy of the method.

[0041] Specifically, image data acquisition utilizes one or a combination of ASIC chips, FPGA chips, PCB boards, ARM boards, Windows motherboards, wireless radio frequency modules, and GPRS data transmission modules; the image reconstruction module is to reconstruct image data using a trained network model; the image registration work uses one or more of the methods based on grayscale information, the methods based on spatial transformation domain, and the methods based on image features; the computing platform used for image registration is one or more of the OpenCV computer vision library, ArcGIS, or MATLAB.

[0042] The above contents described in this specification are merely examples of the present invention. Those skilled in the art may make various modifications, additions, or substitutions to the described embodiments, without departing from the contents of this specification or exceeding the scope defined by the claims, and such modifications, additions, or substitutions may be made to the described embodiments. Such modifications, additions, or substitutions may be made by persons skilled in the art. Such modifications, additions, or substitutions may be made to the described embodiments without departing from the contents of this specification or exceeding the scope defined by the claims, and such modifications shall fall within the scope of protection of the present invention.

Claims

1. A reconfigurable X-ray imaging method based on two-dimensional layered materials, characterized by The following steps are involved: Step 1: Use the Monte Carlo method to model the imaging process of a two-dimensional material pixelated detector, simulating the imaging process after rays pass through an object of fixed shape, and obtain radiation images of objects of different shapes; Step 2: The radiation images are preprocessed and used as samples to train and test the deep learning network model to obtain the optimal reconfigurable parameters for different radiation images; Step 3: Based on the optimal reconfigurable parameters, the detector's response sensitivity is optimized by dynamically adjusting the bias voltage of the 2D material pixel detection array and utilizing the bipolar field effect characteristics under the action of the interlayer van der Waals force. Step 4: Use rays to illuminate an object of fixed shape, use the set two-dimensional material detection array to image the object, collect the preprocessed image, input the trained neural network for image reconstruction, and perform correlation registration with the original image to verify the accuracy and feasibility of the method.

2. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, wherein the two-dimensional layered material is one or more of GaN, WSe2, MoS2, BN, SiC, and MoTe2.

3. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: The two-dimensional layered material interacts with each other through van der Waals forces, and forms a stable lattice structure through atomic covalent bonds within the plane, maintaining its atomic-level flatness and having no dangling bonds between layers, and ensuring bipolar field effect characteristics.

4. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: The Monte Carlo method is a Monte Carlo program, and the Monte Carlo program adopts one or more of MCNP, EGS, GEANT4, FLUKA, SuperMC, and Phits.

5. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, wherein the deep learning framework used in its deep learning network is one or more of DeepLearnToolbox, Caffe, CNTK, TensorFlow, Theano, Torch, Keras, Lasagne, DSSTNE, MXNet, DeepLearning4J, ConvNetJS, Chainer, and Scikit-Learn.

6. According to the reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, the deep learning network model adopts one or more of a deep belief network, a deep neural network, a convolutional neural network, a stacked autoencoder or a convolutional autoencoder.

7. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: By applying different specific voltages at both ends of the detection material, different responses to radiation can be achieved.

8. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: Image data acquisition uses one or a combination of ASIC chips, FPGA chips, PCB boards, ARM boards, Windows motherboards, wireless radio frequency modules, and GPRS data transmission modules.

9. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: The image reconstruction module uses the trained network model to reconstruct image data.

10. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, wherein the image registration method is one or more of a grayscale information-based method, a spatial transform domain-based method, and an image feature-based method.

11. The reconfigurable X-ray imaging method based on two-dimensional layered materials according to claim 1, characterized in that: The computing platform used for the image registration in step 4 is one or more of the OpenCV computer vision library, ArcGIS or MATLAB.

Images