Method for generating CT image by using quantum algorithm
The quantum algorithm-based method optimizes CT image reconstruction by minimizing sinogram discrepancies, enhancing image resolution and diagnostic accuracy, and reducing the reliance on contrast agents, addressing the limitations of existing CT image reconstruction techniques.
Patent Information
- Application Number
- PCT/KR2023/008927
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-06-02
- Filing Date
- 2023-06-27
- Publication Date
- 2024-12-05
AI Technical Summary
Current CT image reconstruction methods face challenges in producing clear and high-contrast images, leading to difficulties in accurate lesion diagnosis, and the use of contrast agents is uncomfortable and poses side effects, limiting their application.
A quantum algorithm-based method using a QUBO model is applied to optimize the sinogram data, minimizing discrepancies between experimentally obtained and quantized sinograms, and reconstructing CT images with improved resolution through quantum computing, specifically utilizing D-Wave's hybrid solver for global minimum energy calculation.
This approach enables the generation of clear and high-resolution CT images with reduced computational complexity, improving diagnostic accuracy and minimizing the need for contrast agents by leveraging quantum optimization for image segmentation and reconstruction.
Smart Images

Figure KR2023008927_05122024_PF_FP_ABST
Abstract
Description
CT image generation method using quantum algorithm
[0001] The present invention relates to a CT image generation method using a quantum algorithm, and more particularly, to an excellent CT image generation method capable of reconstructing and obtaining a CT image in a simple and highly reliable manner by using a projection transformation technique that quantizes a CT image using a quantum algorithm and matches it with an actual projection image.
[0002] Computed tomography (CT) is known as a very useful medical device that can be used to check the condition of the human body by obtaining cross-sectional images of the body by having the subject enter a large, circular machine with an X-ray generator and taking pictures.
[0003] Compared to simple X-rays, these CT images provide images that allow for the inference of actual shapes through dense cross-sections without overlapping structures, offering the advantage of at least allowing for a clearer view of structures and lesions. Therefore, these CT images are widely used for precise examinations of most organs and diseases.
[0004] However, because CT images do not provide clear images, CT image contrast is a crucial factor in accurately diagnosing lesions. Therefore, with the advancement of CT techniques, efforts are continuing to obtain high-contrast CT images.
[0005] Currently, a method of injecting contrast agents that enhance the X-ray attenuation of blood vessels has been utilized to obtain improved images due to the difference in attenuation between the lesion and surrounding tissue. However, this method of contrast agent injection is not only uncomfortable for patients, but also raises concerns about potential side effects, leading to its minimization.
[0006] In this way, computed tomography (CT) is a powerful tool for non-destructively analyzing the internal structure of an object and is a very useful examination method in that it allows for precise examination without damaging the object. However, it is difficult to obtain clear CT images yet, making it difficult to accurately diagnose lesion sites.
[0007] As a technology studied by the inventor of the present invention, WO 2017-188559 A proposes a CT image reconstruction method, which comprises determining at least one fixed point of an object for reconstructing an image in computed tomography, setting the fixed point as a virtual rotation axis so that the fixed point corresponds to a straight line crossing the center of a sinogram, calculating the relative position of a point other than the fixed point of the object with respect to the fixed point within the sinogram, and reconstructing the image based on the calculation of the relative position of the point.
[0008] However, in this case, despite significant progress in CT image reconstruction methods, there were some aspects of the technology that were not sufficiently complete.
[0009] The present invention aims to solve the problems of the prior art by generating very clear CT images using a quantum-based algorithm.
[0010] Therefore, the purpose of the present invention is to provide a method for generating a clear CT image using a quantum algorithm.
[0011] In addition, another object of the present invention is to provide a new QUBO (quadratic unconstrained binary optimization) model that can be applied as a quantum-based algorithm to quantize the CT image as described above and match it with an actual projection image.
[0012] In addition, another object of the present invention is to provide a CT image with excellent resolution by optimizing a superimposed sinogram and an experimentally obtained sinogram by applying the above-described quantum-based algorithm.
[0013] In order to solve the above-described problem of the present invention, the present invention provides a method for generating a segmented CT image from projection data by minimizing the discrepancy between a quantized sinogram derived from a CT image using an experimentally obtained sinogram and a projection used in the experiment.
[0014] According to a preferred embodiment of the present invention, the segmented CT image can be obtained through a process including an optimization step of an overlapping sinogram and an experimentally obtained sinogram.
[0015] According to a preferred embodiment of the present invention, the segmented CT image
[0016] A step of obtaining a known mass decay for a sample;
[0017] A step of multiplying the mass decay and the qubit for each pixel of the CT image;
[0018] A step of obtaining superimposed projection data by applying the same mathematical algorithm as the beam shape obtained from the projection image to the CT image in the superimposed state with the mass attenuation coefficient;
[0019] A step of defining the difference between the overlapping projection image and the actual projection image for each pixel in the form of an optimization problem; and
[0020] Steps to calculate using the QUBO model
[0021] A segmented CT image can be generated by reconstructing the CT image using an algorithm including .
[0022] According to a preferred embodiment of the present invention, the QUBO model can be formulated by summing the QUBO forms for each pixel of all projected images.
[0023] According to a preferred embodiment of the present invention, the global minimum energy of the QUBO model can be calculated using the hybrid solver of the D-Wave system.
[0024] According to a preferred embodiment of the present invention, the QUBO model can be obtained by the following mathematical expression 1.
[0025] [Mathematical Formula 1]
[0026]
[0027] Here are each The pixel positions and ratios of the CT image that affect the . Is Location of CT images affecting It means the average mass attenuation coefficient for . Here, The value of is unknown, and to formulate the QUBO model, is a qubit variable is expressed as the sum of .
[0028] According to a preferred embodiment of the present invention, the overlapping projection image can be obtained by a mathematical projection method used in CT images including projection transformation.
[0029] According to a preferred embodiment of the present invention, the X-ray can be selected from among a balanced beam, a cone beam, and a fan beam, and can also be used for electrons or other transmission forms.
[0030] In addition, the present invention is utilized as an image reconstruction means applied to quantum computing for reconstruction of CT images, and includes a QUBO model characterized by being expressed by the mathematical expression 1 or an equation that produces the same result.
[0031] The present invention also comprises a step of obtaining a known mass attenuation for a sample;
[0032] A step of multiplying the mass decay and the qubit for each pixel of the acquired projection image;
[0033] A step of obtaining superimposed projection data by applying the same mathematical algorithm as the beam shape obtained from the projection image to the image image in the superimposed state;
[0034] For each pixel, the difference between the overlapping projection image and the projection image is implemented as a step in the QUBO model in the form of a least squares problem.
[0035] Includes a QUBO model design method including .
[0036] As described above, the method for generating a CT image according to the present invention has the effect of generating a CT image with excellent resolution in a simple manner.
[0037] Figure 1 illustrates an example of a quantum annealer using the D-Wave system used in the present invention.
[0038] Figure 2 illustrates a random transformation according to the present invention.
[0039] Figure 3 shows an example of a state in which each pixel of a CT image to be implemented is made into a quantum superposition state and a superimposed sinogram is created in accordance with a light source.
[0040] FIG. 4 is a schematic diagram illustrating the process of using a superposed sinogram and an experimental sinogram to calculate a QUBO model according to the present invention, and then using a quantum annealer to create a CT image with qubits that satisfy the global minimum energy for the QUBO model.
[0041] Figures 5 and 6 are conceptual examples showing Radon transform in equilibrium light, respectively, according to the present invention.
[0042] Figure 7 shows each pixel of the superposed sinogram to calculate the QUBO model according to the present invention. and each pixel of the sinogram This is a diagram illustrating the process of using .
[0043] Figure 8 is an image showing acquisition and transformation of a projection image, where (a) is an original projection image with a resolution of 2560 x 2160 pixels, (b) is a binned projection image with a resolution of 50 x 45 pixels, and (c) is a projection image with the projection area of the space set to 0.
[0044] Figure 9 shows the segmentation of a conventional CT image, where (a) is the axial level used for CT image reconstruction, which is indicated in red on the X-ray projection image, (b) is a sinogram obtained at axial level 16, and (c) to (e) are CT images and segmented images reconstructed using a classical algorithm.
[0045] Figure 10 is a photograph showing CT image segmentation by a quantum optimization algorithm, where (a) the axial level used for CT image reconstruction is indicated in red on the X-ray projection image, (b) is a sinogram obtained at axial level 16, which is made proportional to the mass attenuation coefficient, (c) is an image segmented by a quantum optimization algorithm, and (d) compares two segmented images obtained by a classical and a quantum algorithm.
[0046] Hereinafter, the present invention will be described in more detail as an embodiment.
[0047] The present invention relates to a method for implementing a CT image with excellent resolution that can more accurately read a computed tomography (CT) image.
[0048] CT is an important imaging technique used for medical analysis of the internal structures of the human body, and is therefore widely used in the medical field.
[0049] Typically, to obtain a segmented CT image, an image segmentation method is required after obtaining a reconstructed CT image, which causes errors in both the reconstruction and segmentation algorithms, resulting in low resolution or unclear interpretation of the CT image, making it difficult.
[0050] The present invention is characterized by applying a new method that essentially uses an advanced quantum optimization algorithm called QUBO (Quadratic Unconstrained Binary Optimization) to solve this problem.
[0051] In the present invention, a CT image may be defined as a tomography image that includes an X-ray image, but essentially uses a series of tomography techniques to visualize invisible information. Therefore, it includes images generated by electrons or synchroton, in addition to X-rays.
[0052] In the present invention, what is expressed as a projection transformation is defined identically to include several projection transformations such as Radon transformation, and includes, for example, the use of cone beam projection transformation in the medical field.
[0053] In the present invention, the mass attenuation coefficient refers to a physical value that appears when a projected beam passes through the interior of an object, and is defined to include concepts such as a scattering coefficient and an absorption coefficient. One of the features of the present invention is that it can represent a CT image generated by overlapping these mass attenuation coefficients, and thus, for example, can represent a CT image in a superimposed state that can complement X-ray coherence.
[0054] The present invention includes a method that can utilize a quantum optimization algorithm designed using quantum computing techniques.
[0055] According to a preferred embodiment of the present invention, a method of utilizing a newly designed QUBO (quadratic unconstrained binary optimization) model specifically designed for such an algorithm is proposed.
[0056] The QUBO model in the present invention can be mathematically designed in various ways. However, the QUBO model exemplified in the present invention is designed to be calculated using the simplest formula. Therefore, it should be understood that the present invention includes QUBO models with the same purpose, including those designed using various mathematical design methods. Accordingly, the QUBO model expressed as a mathematical formula in the present invention can be included as an equation produced for the same purpose.
[0057] In addition, in the present invention, the QUBO model is a model applied to quantize a CT image and match it with an actual projection image, and the difference between the overlapping projection image and the actual projection image for each pixel is defined in the form of an optimization equation, and can be defined as an application technology of an important algorithm used to generate a more improved image.
[0058] According to a preferred embodiment of the present invention, the QUBO model is well suited for use with gate-model-based quantum computers, particularly when combined with the Quantum Approximate Optimization Algorithm (QAOA). To maximize performance when working with the QUBO model, it is preferable to use a quantum annealer instead of a gate-model quantum computer. This is based on the advantages of a quantum annealer, which provides superior performance over a gate-model quantum computer. D-Wave's Advantage quantum annealer currently boasts impressive specifications, including over 5,000 qubits and 35,000 couplers. This allows for the use of 180 logical qubits with interconnects. Notably, D-Wave also offers a hybrid model that supports up to 1 million variables and 200 million biases. Therefore, the present invention utilizes a newly designed QUBO model that applies these technologies.
[0059] According to a preferred embodiment of the present invention, the QUBO algorithm can obtain a segmented CT image from X-ray projection data by minimizing the discrepancy between an experimentally obtained sinogram and a quantized sinogram derived from a quantized segmented CT image using Radon transform.
[0060] According to a preferred embodiment of the present invention, D-Wave's hybrid solver system can be utilized for verification on actual X-ray data.
[0061] Computed tomography (CT), which is intended to improve usability and image interpretation information in the present invention, is essentially a powerful tool for non-destructively analyzing the internal structure of an object, and has the characteristic of allowing precise inspection without damaging the object.
[0062] Typically, obtaining segmented CT images from projection data requires two steps. First, a backprojection algorithm is applied to the projection data to reconstruct the CT image. A segmentation algorithm is then applied to the reconstructed CT image to obtain a segmented image. The CT image obtained in this manner is then used to reconstruct the internal structure of the object based on projection images taken from different angles using a backprojection algorithm. Backprojection utilizes various algorithms, including iterative methods, fast Fourier transforms, artificial intelligence, and optimization. If the optimization algorithm utilized here utilizes a complete sinogram pattern, it can generate a more accurate image that closely resembles the actual internal structure.
[0063] According to a preferred embodiment of the present invention, the present invention designs an algorithm that minimizes the disparity between a reconstructed CT image and an actual sinogram by repeatedly updating a sinogram through Radon transform for a randomly generated CT image based on such a technical basis.
[0064] However, computational results involving conventional computers often compromise the quality of acquired CT images. To address this issue, the present invention proposes a novel QUBO model.
[0065] According to a preferred embodiment of the present invention, the QUBO modeling method comprises comparing each pixel of a superimposed sinogram with each pixel of an experimental sinogram. However, the present invention essentially comprises optimizing the superimposed sinogram and the experimental sinogram.
[0066] The QUBO model proposed in the present invention can be implemented on a quantum computer, leveraging its computational efficiency and similarity to optimization algorithms. Nevertheless, improving CT image quality remains a challenge within the QUBO model due to the limited number of logical qubits and the low probability of finding the minimum energy state in a quantum computer.
[0067] According to a preferred embodiment of the present invention, when calculating a QUBO model for CT image reconstruction according to the method of the present invention using D-Wave's hybrid solver, it is expected that very good results can be obtained even using about 10,000 logical qubits.
[0068] In the present invention, based on this theoretical basis, the following new QUBO model for quantum segmentation capable of performing both CT image reconstruction and CT image segmentation is proposed.
[0069] As an embodiment of the present invention, it is assumed that the X-ray mass attenuation of a sample is already known. The new algorithm multiplies the X-ray mass attenuation by a qubit for each pixel of a CT image. Each pixel is a quantum superposition state that can represent all values of the sample and space. The new algorithm then applies the same mathematical algorithm as the X-ray beam shape obtained from the X-ray projection image to the CT image in the superposition state to produce superposition projection data. For each pixel, the difference between the superposition projection image and the X-ray projection image is calculated in the form of a QUBO in the form of a least squares problem. Finally, the QUBO model can be formulated as follows: Equation 1 by summing the QUBO forms for each pixel of all projected images.
[0070] [Mathematical Formula 1]
[0071]
[0072] Here, each of the above elements is as defined above.
[0073] According to the present invention, each term of the QUBO model derived through this formalization process can be expressed by the following mathematical expression 2.
[0074] [Equation 2]
[0075]
[0076] In the above formula, each element is as defined above.
[0077] According to a preferred embodiment of the present invention, a new algorithm can be verified using a parallel beam X-ray source, and the size of the image used in the experiment is, for example, 50 x 50, and a total of 2,500 logical qubits can be used. The mathematical projection method used for the superimposed CT images is the Radon transform. The global minimum energy of the QUBO model can be calculated using the hybrid solver of the D-Wave system.
[0078] While one embodiment of the present invention has been verified using a parallel beam X-ray source, the new algorithm can be used with any type of X-ray source. For example, the X-ray source can be selected from among a parallel beam, a cone beam, and a fan beam. A cone beam is preferably used for medical CT. Additionally, it can also be used in a synchrotron radiation accelerator or a cryogenic electron microscope.
[0079] Meanwhile, the basic technology for projection transformation, such as Radon transformation, utilized in the present invention is described.
[0080] The Radon transform mathematically describes how to obtain projection data for a parallel beam type X-ray source.
[0081] According to a preferred embodiment of the present invention, the Radon transform can be calculated as follows.
[0082]
[0083] Here, p represents the slope of the line, τ represents the intercept, and δ(x) represents the delta function.
[0084] According to a preferred embodiment of the present invention, a sinogram is generated by accumulating Radon transforms based on angles. For example, when using identical samples at different locations, the projected positions due to the Radon transform change. The shape of the projected object in a sinogram free of motion artifacts is called an ideal sinogram pattern, which is an important factor in reconstructing CT images.
[0085] In relation to CT image reconstruction through this Radon transform, an example of an optimization algorithm can be constructed in the following way.
[0086] That is, as an alignment method for correcting motion artifacts in the entire projection image during scanning, unlike previous approaches, it satisfies the Helgason-Ludwig consistency condition by calculating points within the sample in virtual space. Therefore, high-quality CT images can be acquired even when the rotation center of the space changes.
[0087] Based on this, an algorithm for reconstructing CT images based on sinogram patterns can be introduced. In particular, this algorithm can control the reconstructed location of a sample within a CT image by manipulating the sinogram pattern.
[0088] According to a preferred embodiment of the present invention, by randomly reconstructing CT images within a specific region of interest, the algorithm minimizes the discrepancy between the sinogram generated from the initial CT image via the Radon transform and the actual sinogram. Furthermore, because it utilizes the complete sinogram pattern of the sample, it exhibits robustness against artifacts at specific angles.
[0089] Based on this method, the CT image reconstruction algorithm can be mathematically formulated in the following optimized form.
[0090]
[0091] In the above equation, T is one of the CT images, S is the given sinogram, and R is the Radon transform. Here, the Radon transform calculation uses PyTorch (1.9.0+cu111), and the optimization calculation can use the Adam optimizer of the PyTorch library.
[0092] According to the present invention, the following method is proposed for QUBO formulation for CT image reconstruction.
[0093] Quantum analyzers and gated-model quantum computers provide the QUBO and Ising models as quantum optimization algorithms. While quantum annealing calculates the minimum value of the model, gated-model quantum computers use QAOA to calculate the maximum value. This invention describes the use of QUBO modeling for minimization.
[0094] In general, the pixel values of a projection image are correlated with the intensity of X-rays passing through the thickness of the sample. In the region where the Beer-Lambert law applies, for a particular axial level The value of the i-th position, indicated by , can be expressed using the following equation (4).
[0095] (4)
[0096] Here, S(x,y) is the X-ray mass attenuation coefficient for the sample at position (x,y), and the l-axis is defined as perpendicular to the X-ray direction. This equation can be expressed as a sum in discrete form as shown in Equation (5) below.
[0097] (5)
[0098] Here, and are each The pixel positions and ratios of the CT image that affect the . Is Location of CT images affecting can be assumed to be the average X-ray mass attenuation coefficient for . is for the projection angle θ , and the sinogram S is the accumulation of P along the projection angle. Here, The value of is unknown, and to formulate the QUBO model, is a qubit variable is expressed as the sum of .
[0099] The QUBO model for each pixel in the sinogram is computed as follows:
[0100]
[0101] Therefore, the QUBO model for CT image reconstruction is calculated as follows.
[0102]
[0103] A method for reconstructing a CT image using the QUBO model for manufacturing a quantum CT designed in the above manner can be summarized as follows in one embodiment.
[0104] (1) A quantum annealer using the D-Wave system can use Figure 1.
[0105] (2) Random transformation can be performed as shown in Fig. 2.
[0106] (3) Each pixel of the CT image to be implemented is made into a quantum superposition state, and 2. A superposed sinogram can be created by matching it to the X-ray light source. (Fig. 3)
[0107] (4) To calculate the above equation (9) (QUBO), a superposed sinogram and an X-ray sinogram are used, and then a quantum annealer can create a CT image with qubits that satisfy the global minimum energy for the QUBO model. (Fig. 4)
[0108] (5) Radon transform in equilibrium light can be illustrated in Figs. 5 and 6.
[0109] (6) To calculate the above equation (9), each pixel of the superposed sinogram and each pixel of the X-ray sinogram can be used. ( ) (Fig. 7)
[0110] (7) Run QUBO modeling.
[0111] CT images can be reconstructed using the QUBO model obtained through the above process.
[0112] Hereinafter, a process for obtaining a CT image using the method of the present invention as described above will be described by presenting an example. However, the present invention is not limited to this example.
[0113]
[0114] <Example>
[0115] 1) Data collection
[0116] X-ray microcomputed tomography (μCT) was performed at beamline 6C Biomedical Imaging of Pohang Gwangwon-II16.
[0117] Figure 8 is an image showing the acquisition and conversion of an X-ray projection image, where (a) is an original X-ray projection image with a resolution of 2560 x 2160 pixels, (b) is a binned X-ray projection image with a resolution of 50 x 45 pixels, and (c) is an X-ray projection image with the projection area of space set to 0.
[0118] The projection image of Fig. 8 (a) was obtained at a resolution of 2560 x 2160 pixels and converted to a resolution of 60 x 45 pixels in Fig. 8 (b) for experimental purposes. 50 × 50 binning was performed based on the center of rotation (binning is a method of calculating the average by summing pixels). Each projection image set was measured every 0.5˚, and a total of 360 projection images were measured per set.
[0119] Here, the center of rotation is calculated and binning is performed for image preprocessing, but only the center of rotation is calculated for convenience. However, according to this embodiment, the center of rotation can be calculated when modeling QUBO without having to calculate it. Depending on the performance of the quantum computer, image preprocessing may not be necessary. Pixel Since it is used as , the projection image for air must be 0. Therefore, in Fig. 8, the average of the air part in the middle image (b) is calculated and subtracted from the whole to obtain the right image (c). If we take (c as a constant), we can calculate it directly from the image on the right (c).
[0120]
[0121] 2) Quantum segmentation algorithm for X-ray data
[0122] Assume that the sample exists in a three-dimensional space. Let α=μ / ρ be the X-ray mass attenuation coefficient and a natural number. If the number of different X-ray mass attenuation coefficients that each grid space can have is m+1, then each pixel of the reconstructed image is expressed as one of the combinations of qubits and binary numbers in the following equation (10), or in a form that can represent all natural numbers (if it is not a natural number, it can be linearly transformed into a natural number).
[0123] (10)
[0124] Here is 0 or 1 is a set for positions (i,j) in two-dimensional space Represents one element of . In 3D space, (i,j,z) is used instead of (i,j), where z is the axis level.
[0125] To apply an optimization algorithm to X-ray projection data in the form of parallel beams, the Radon transform is used for the superimposed CT images. IP is defined as the superimposed sinogram transformed by the CT image I. The s-th position of IP for the projection angle θ is calculated as shown in Equation (11) below.
[0126] (11)
[0127] Here When the CT image is projected It means the pixel that affects Is This is the area that overlaps when projected.
[0128] Applying the least squares equation of the difference between P(θ.s) and IP(θ.s), the QUBO model is calculated as follows (12).
[0129]
[0130] In the above equation (14), the second term is a linear term in the QUBO model and the third term represents a part of the optimized value. The first term without is computed as follows:
[0131]
[0132] To derive equation (16) from equation (15) above, Since is 0 or 1 can be used to transform the square term. In the second term of the above equation (16), i, j and k are and cannot be the same at the same time. The first term of the above equation (14) can be calculated as the sum of the linear term and the quadratic term as in equation (16).
[0133] At this time, since the two sinograms P and IP can be compared, each pixel value of the two sinograms is subtracted and squared to calculate the energy minimization model.
[0134]
[0135] Here, θ is the projection angle, s is the sensor position, and dθ is the change in the projection angle. In addition, F(θ,s) is expressed as a linear term that is a second-order term excluding the constant term. In the QUBO model, the constant term is excluded. The minimum value of the QUBO model is the opposite sign of the sum of the constant terms.
[0136]
[0137]
[0138] 3) Data collection and preprocessing of X-ray projection images
[0139]
[0140]
[0141] In the above equation (1) can take from 0 to the maximum value of the X-ray mass attenuation coefficient of the sample. Therefore, the sinogram of the superposition state obtained by the Radon transform also starts from 0. To express the actual X-ray projection image of the projection image created by the Radon transform, A constant value must be added or subtracted from the X-ray projection image. To express the X-ray coherence effect, More qubits need to be allocated. In this experiment, the spatial average of the X-ray projection image is taken and subtracted from the image (see Figure 1c). Also, the X-ray coherence effect is ignored. The sample is a tooth model made of resin. To minimize the use of logical qubits, the sample was made of a single material. In this case, Is can be expressed as
[0142]
[0143] 4) Conclusion
[0144] In this example, we compare segmented images obtained from classical CT images with segmented images generated using a quantum annealer.
[0145] Figure 9 shows segmentation of a conventional CT image, where (a) the axial levels used for CT image reconstruction are indicated in red on the X-ray projection image. (b) is a sinogram obtained at axial level 16. (c) to (e) are CT images and segmented images reconstructed using a classical algorithm.
[0146] For the classical segmented image, a sinogram corresponding to the axis level indicated by the red line in (a) in Fig. 9 was generated (see (b) in Fig. 9). The CT image shown in (c) in Fig. 9 was reconstructed using the fast Fourier transform algorithm provided by MATLAB. Subsequently, the interior of the tooth was identified by applying the MATLAB binarization algorithm to the CT image, as shown in (d) in Fig. 9.
[0147] Finally, a segmented image of axis level 16 was obtained for the tooth sample (see Fig. 9 (e)).
[0148] We implemented a quantum segmentation algorithm using the D-Wave system's hybrid solver. This novel quantum algorithm enables the computation of segmented images directly from X-ray data in a single step.
[0149] Figure 10 is a photograph showing CT image segmentation by a quantum optimization algorithm. (a) The axial levels used for CT image reconstruction are indicated in red on the X-ray projection image. (b) is a sinogram obtained at axial level 16. This sinogram is made proportional to the X-ray mass attenuation coefficient. (c) is an image segmented by a quantum optimization algorithm. (d) compares two segmented images obtained by a classical and a quantum algorithm.
[0150] As shown in Fig. 10(b), the sinogram can be derived from the projection image proportional to the X-ray mass attenuation coefficient shown in Fig. 10(a). The goal is to express all values through the superimposed CT images by formulating the QUBO model and calculating the pixel values of the sinogram using Equation (17). The theoretically calculated global minimum energy is determined by the negative sum of the square values of each pixel in the sinogram, which is -821370.3333333333 in our dataset. We allocated 3 seconds to obtain three minimum energy values, -817517.07239, -817516.349165, and -817517.950592, using the solver for the QUBO model. Fig. 10(c) shows the segmented image corresponding to the last energy value. In addition, (d) of Fig. 10 is a comparative analysis of the segmented image derived from the classical algorithm of (e) in Fig. 9 and the segmented image obtained through the quantum optimization algorithm of (c) in Fig. 10.
[0151]
[0152] From the above experimental results, the following conclusions were reached.
[0153] Quantum algorithms for CT image reconstruction require a number of qubits equal to the number of bits in each pixel. For example, reconstructing a 50 x 50 CT image with a resolution of 64 bits would require a total of 160,000 logical qubits.
[0154] In contrast, the quantum segmentation algorithm can reconstruct and segment CT images using just 2,500 logical qubits, provided the sample's X-ray mass attenuation coefficient is known. Furthermore, this new algorithm can produce more precisely segmented CT images using fewer logical qubits.
[0155] Fig. 10(b) shows a sinogram derived using MATLAB and later converted to integer values in Python. The sinogram values in Python and the Radon transform applied to the segmented image in Fig. 9(e) were utilized to calculate the X-ray mass attenuation coefficient of the sample.
[0156] Although the present invention used approximate coefficient values, it was confirmed that the method of the present invention can produce more accurate results by obtaining coefficients through experimental means. Furthermore, setting the values of each column in the sinogram to a constant and applying uniform normalization to the entire dataset is expected to yield even better results.
[0157] Furthermore, according to the present invention, it is impossible to achieve the theoretically achievable global minimum energy because the new algorithm uses the X-ray mass attenuation coefficient. This is because the pixels surrounding the shell of the sample are mixed with the sample and the surrounding space, making it impractical to represent the pixels based only on the X-ray mass attenuation coefficient of each pixel. However, during the segmentation process, these pixels are converted to binary values of 0 or 1, and the new algorithm can effectively solve this problem. In the embodiment of the present invention, the results were calculated by running the hybrid solver of the D-Wave system three times for 3 seconds each. The obtained energy was different for each iteration, and providing more annealing time would yield better results.
Claims
1. A method for generating a CT image using a quantum algorithm, characterized in that a segmented CT image is obtained by reconstructing from projection data using a quantum algorithm that minimizes the discrepancy between an experimentally obtained sinogram and a quantized sinogram derived from a CT image quantized with a mass attenuation coefficient using a projection transformation.
2. A method for generating a CT image using a quantum algorithm, characterized in that the CT image is generated through a process including an optimization step of a superimposed sinogram and an experimental sinogram in claim 1.
3. In claim 1, the reconstructed CT image A step of obtaining a known mass decay for a sample; A step of multiplying the mass decay and the qubit for each pixel of the CT image; A step of obtaining superimposed projection data by applying the same mathematical algorithm as the beam shape obtained from the projection image to the CT image in the superimposed state; A step of implementing a QUBO (quadratic unconstrained binary optimization) model in the form of an optimization problem of the difference between the overlapping projection image and the projection image for each pixel; and Steps for calculating using the QUBO model; A method for generating a CT image using a quantum algorithm, characterized in that the CT image is reconstructed using an algorithm including .
4. A method for generating a CT image using a quantum algorithm, characterized in that, in claim 3, the QUBO model is formulated by summing the QUBO formats for each pixel of all projected images.
5. A method for generating a CT image using a quantum algorithm, characterized in that, in claim 3, the method is calculated using the global optimal energy of the QUBO model.
6. A method for generating a CT image using a quantum algorithm, characterized in that, in claim 3, the overlapping projection image is obtained by a mathematical projection method used for a CT image including a projection transformation.
7. A method for generating a CT image using a quantum algorithm, characterized in that in any one of claims 3 to 6, the QUBO model is expressed by the following mathematical expression 1 or an equation that produces the same result. [Mathematical Formula 1] Here, are each The pixel positions and ratios of the CT image that affect the . Is Location of CT images affecting means the average mass attenuation coefficient for . Here, of The value is unknown, and to formulate the QUBO model, is a qubit variable is expressed as the sum of .
8. A method for generating a CT image using a quantum algorithm, characterized in that in claim 3, the projection beam is selected from among a balanced beam, a cone beam, and a fan beam, and can be used in a synchrotron accelerator or a cryogenic electron microscope.
9. A QUBO (quadratic unconstrained binary optimization) model, which is used as an image reconstruction means applied to quantum computing for the reconstruction of CT images and is characterized by being expressed by the following mathematical formula 1 or an equation that produces the same result: [Mathematical Formula 1] Here, are each The pixel positions and ratios of the CT image that affect the . Is Location of CT images affecting means the average mass attenuation coefficient for . Here, of The value is unknown, and to formulate the QUBO model, is a qubit variable is expressed as the sum of .
10. Step of obtaining a known X-ray mass attenuation for the sample; A step of multiplying the mass decay and the qubit for each pixel of the acquired projection image; A step of obtaining superimposed projection data by applying the same mathematical algorithm as the beam shape obtained from the projection image to the image image in the superimposed state; For each pixel, the difference between the overlapping projection image and the X-ray projection image is implemented as a step of the QUBO (quadratic unconstrained binary optimization) model in the form of a least squares problem. A method for designing a QUBO model including .
11. A QUBO model design method characterized in that, in claim 10, the design is performed using the following mathematical formula 2: [Equation 2] Here, are each The pixel positions and ratios of the CT image that affect the . Is Location of CT images affecting means the average mass attenuation coefficient for . Here, of The value is unknown, and to formulate the QUBO model, is a qubit variable is expressed as the sum of .
Citation Information
Patent Citations
A CT projection denoising and reconstruction method and apparatus based on noise generation mechanism and data-driven tight framework
CN110136218B
A Low-Dose CT Reconstruction Method Based on Interpolation Convolutional Neural Networks
CN112489156B
Computer tomography apparatus and method for reconstrucing a computer tomography image thereof
KR1020150099384A
