CT dose adaptive optimization system and method

Abstract

The application discloses a CT dose self-adaptive optimization system and method thereof, and belongs to the technical field of medical images, which comprises a CT scanner, an image data processing workstation, a high-risk population analysis device, a dose parameter self-adaptive optimization device, a dose parameter automatic adjustment device and an intelligent learning unit, and based on the theory of information geometry, the dose parameter optimization problem is promoted from the Euclidean space to the Riemannian manifold space, the dose parameter manifold and the image quality manifold are constructed, the geodesic line is calculated on the manifold through the Fisher information matrix as the Riemannian metric, and the optimal dose path is found. The system adopts a double-manifold collaborative optimization mechanism, establishes a mapping relationship between the two manifolds, and realizes accurate optimization of parameters in combination with a geometric-guided genetic algorithm, and through pre-scanning, high-risk population analysis, information geometry optimization and automatic parameter adjustment, individualized dose control is realized, the radiation dose is additionally reduced by 30-45% compared with a traditional method, the image quality is maintained or improved, and the safety of CT examination is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, specifically to a CT dose adaptive optimization system and method based on information geometry and artificial intelligence, used to reduce CT scan radiation dose while ensuring image quality. Background Technology

[0002] Computed tomography (CT) scans play an irreplaceable role in clinical applications as a crucial tool in modern medical diagnosis. However, the X-ray radiation generated during CT scans poses potential health risks, especially to patients requiring repeated CT examinations, children, and radiation-sensitive populations. Therefore, effectively reducing CT radiation dose while ensuring image quality meets diagnostic requirements has become an important research topic in the field of medical imaging.

[0003] Existing CT dose optimization methods mainly include body shape-based automatic tube current adjustment technology, iterative reconstruction algorithms, and low-dose scanning protocols. While these methods reduce radiation dose to some extent, they still have the following limitations: First, dose parameter adjustments are usually based on simple linear relationships or lookup tables, which cannot accurately adapt to individual differences among patients; second, existing methods typically only consider patient body shape or scan sites, ignoring the influence of individual genetic risk factors and pathological characteristics; and third, the parameter optimization process lacks theoretical guidance and relies mainly on empirical adjustments, resulting in limited and unstable optimization effects.

[0004] The development of information geometry and artificial intelligence technologies has provided new technical approaches for CT dose optimization. Information geometry maps probability distributions onto Riemannian manifolds, accurately characterizing the intrinsic geometric structure of the parameter space, while artificial intelligence can learn optimization strategies from massive amounts of data. However, currently, there is a lack of CT dose optimization systems that organically combine information geometry and artificial intelligence technologies, failing to fully leverage the advantages of both technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a CT dose adaptive optimization system and method based on information geometry and artificial intelligence. By constructing an information geometric manifold, calculating the Fisher information matrix, and applying a geometry-guided genetic algorithm, the system achieves precise adaptive optimization of CT dose parameters, significantly reducing radiation dose while ensuring image quality.

[0006] This invention proposes a CT dose adaptive optimization system, comprising: A CT scanner is used to perform CT scans based on scanning parameters and generate CT image data. An image data processing workstation, connected to the CT scanner, is used to process the CT image data; A high-risk population analysis device is connected to the image data processing workstation and is used to receive image data of patients to be diagnosed sent by the image data processing workstation, perform high-risk population analysis based on the image data of patients to be diagnosed, generate high-risk population stratification numbers, and send the high-risk population stratification numbers to the image data processing workstation. A dose parameter adaptive optimization device is connected to the high-risk population analysis device and the image data processing workstation. It is used to receive the high-risk population stratification number, receive pre-scanned image data, construct an information geometric manifold based on the high-risk population stratification number and the pre-scanned image data, calculate the optimal dose parameter on the information geometric manifold, and output the optimal dose parameter. The information geometric manifold includes a dose parameter manifold and an image quality manifold. An automatic dose parameter adjustment device, connected to the dose parameter adaptive optimization device and the CT scanner, is used to receive the optimal dose parameter and automatically adjust the scanning parameters of the CT scanner according to the optimal dose parameter; and The intelligent learning unit, connected to the dose parameter adaptive optimization device and the image data processing workstation, is used to update the information geometric manifold based on historical optimization results.

[0007] Preferably, the dose parameter adaptive optimization device includes: The manifold construction module is used to map the high-risk population stratification number to an n-dimensional parameter space and construct an n-dimensional information geometric manifold, where n is the high-risk population stratification number. The Fisher information matrix calculation module is used to calculate the Fisher information matrix on the information geometric manifold, and the Fisher information matrix is ​​used as a Riemannian metric. A geodesic calculation module is used to calculate the optimal path for dose parameters on the information geometric manifold; The information entropy evaluation module is used to calculate the image quality information entropy and compare it with a preset entropy threshold to determine whether the image quality is acceptable.

[0008] Preferably, the dose parameter adaptive optimization device further includes: A dual-manifold mapping module is used to establish a mapping relationship between the dose parameter manifold and the image quality manifold; The mapping learning module is used to train the mapping relationship based on historical mapping data; The covariance structure optimization module is used to calculate the covariance matrix of parameter changes and quality changes, and adjust the optimization path based on the covariance matrix.

[0009] Preferably, the dose parameter adaptive optimization device further includes: A genetic algorithm optimization module is used to perform geometry-based genetic algorithm optimization on the information geometric manifold; The fitness function construction module is used to construct fitness functions based on geodesic distance; The geometric genetic operators module is used to perform geodesic crossover, tangent space mutation, and geodesic selection operations. The population dynamics management module is used to dynamically adjust the population size and distribution based on the local curvature of the manifold.

[0010] Preferably, the high-risk population analysis device includes: A deep learning network is used to analyze the image data of the patient to be diagnosed. The risk stratification module is used to stratify patients based on the analysis results of the deep learning network and generate the number of stratifications for the high-risk population. A risk database is used to store risk stratification criteria and historical risk assessment data; The pre-scan parameter generation module is used to generate pre-scan parameters based on the number of stratifications of the high-risk population and to control the CT scanner to generate the pre-scan image data.

[0011] Preferably, the deep learning network of the high-risk population analysis device is a generative adversarial network, which is used to obtain sample images of the patients to be diagnosed after training for high-risk population analysis.

[0012] Preferably, the expression for the image quality information entropy on the image quality manifold is: H(y)=-∑p(y)log(p(y)) Where y is the image quality quantization representation value, p(y) is the probability of the image quality quantization representation value y, and the image quality information entropy threshold is 1.1.

[0013] Preferably, the automatic dose parameter adjustment device includes: A tube current adaptive adjustment module is used to adjust the tube current of the CT scanner according to the optimal dose parameters; A tube voltage adaptive adjustment module is used to adjust the tube voltage of the CT scanner according to the optimal dose parameters; The adaptive scanning mode selection module is used to select one of the following scanning modes based on the number of stratifications of the high-risk population: plain CT scan, low-dose CT scan, or high-resolution CT scan.

[0014] Preferably, the intelligent learning unit includes: The optimization results evaluation module is used to evaluate the effect of dose optimization and image quality; The Fisher information matrix update module is used to update the Fisher information matrix based on the latest scan results. A mapping function optimization module is used to optimize the mapping function between the dose parameter manifold and the image quality manifold; The fitness function adjustment module is used to adjust the weight coefficients of the fitness function based on historical optimization results.

[0015] CT dose adaptive optimization methods include: Receive image data of patients to be diagnosed from the image data processing workstation; Based on the imaging data of the patients to be diagnosed, high-risk population analysis is performed to generate high-risk population stratification numbers; Pre-scanning parameters are generated based on the stratification number of the high-risk population; The CT scanner is controlled to perform a pre-scan based on the pre-scan parameters, generating pre-scan image data. An information geometric manifold is constructed based on the high-risk population stratification number and the pre-scanned image data, the information geometric manifold including a dose parameter manifold and an image quality manifold; Calculate the image quality information entropy on the information geometric manifold to determine whether the image quality is acceptable; The optimal dose parameters are calculated by applying a geometry-based genetic algorithm on the information geometric manifold. The optimal dose parameter is output to the automatic dose parameter adjustment device; The scanning parameters of the CT scanner are automatically adjusted according to the optimal dose parameters; and The information geometric manifold is updated based on historical optimization results.

[0016] The present invention has the following beneficial effects: 1. By constructing an information geometric manifold, the dose parameter optimization problem is elevated from Euclidean space to Riemannian manifold space, which can accurately capture the nonlinear relationship between parameters and improve the optimization accuracy; 2. Using the Fisher information matrix as a Riemann metric can quantify the sensitivity of parameter changes to image quality and guide optimization in the most effective direction; 3. An innovative dual-manifold collaborative mechanism is proposed to establish a mapping relationship between the dose parameter manifold and the image quality manifold, thereby achieving deep coupling optimization of dose and quality; 4. By integrating genetic algorithms with information geometry, a fitness function based on geodesic distance and a geometry-guided genetic operator are designed, which significantly improves search efficiency and accuracy; 5. Precise dose control based on individual patient risk characteristics reduces radiation dose by 30-45% compared to traditional methods, while maintaining or improving image quality. Attached Figure Description

[0017] Figure 1This is a schematic diagram of the overall structure of the CT dose adaptive optimization system of the present invention; Figure 2 This is a schematic diagram of the module structure of the dose parameter adaptive optimization device of the present invention; Figure 3 This is a schematic diagram illustrating the process of constructing the information geometric manifold of this invention; Figure 4 This is a schematic diagram of the geometry-guided genetic algorithm optimization process of the present invention; Figure 5 This is a flowchart illustrating the adaptive CT dose optimization method of the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 like Figure 1 As shown, the CT dose adaptive optimization system provided by the present invention includes a CT scanner 1, an image data processing workstation 2, a high-risk population analysis device 3, a dose parameter adaptive optimization device 4, a dose parameter automatic adjustment device 5, and an intelligent learning unit 6.

[0020] The CT scanner 1 is used to perform CT scans and generate CT image data according to the scanning parameters. In one embodiment of the present invention, the CT scanner 1 adopts a third-generation spiral CT device, whose main scanning parameters include tube current (unit: mA), tube voltage (unit: kV), exposure time (unit: s), pitch, slice thickness (unit: mm), etc.

[0021] The image data processing workstation 2 is connected to the CT scanner 1 and is used to process CT image data. Preferably, the image data processing workstation 2 uses a high-performance computer and is equipped with medical image processing software, which can perform operations such as reconstruction, display, and analysis on CT image data.

[0022] The high-risk population analysis device 3 is connected to the image data processing workstation 2. It receives image data of patients to be diagnosed from the workstation, performs high-risk population analysis based on the image data, generates a high-risk population stratification number, and sends this stratification number back to the workstation 2. In this invention, the high-risk population stratification number represents the patient's risk level, typically taking the value of an integer from 1 to 9, with higher values ​​indicating higher risk.

[0023] The dose parameter adaptive optimization device 4 is connected to the high-risk population analysis device 3 and the image data processing workstation 2. It receives the high-risk population stratification number and pre-scanned image data, constructs an information geometric manifold based on the high-risk population stratification number and pre-scanned image data, calculates the optimal dose parameters on the information geometric manifold, and outputs the optimal dose parameters. Here, the information geometric manifold includes the dose parameter manifold and the image quality manifold. The information geometric manifold is one of the core innovations of this invention and will be described in detail in subsequent embodiments.

[0024] The automatic dose parameter adjustment device 5 is connected to the dose parameter adaptive optimization device 4 and the CT scanner 1. It receives the optimal dose parameters and automatically adjusts the scanning parameters of the CT scanner 1 based on the optimal dose parameters. In this way, the system can automatically adjust the CT scanning parameters according to the optimization results without manual intervention, thus improving work efficiency.

[0025] The intelligent learning unit 6 is connected to the dose parameter adaptive optimization device 4 and the image data processing workstation 2, and is used to update the information geometric manifold based on historical optimization results. This design enables the system to have self-learning capabilities, continuously optimizing the dose parameter calculation model and improving optimization accuracy.

[0026] Example 2 like Figure 2 As shown, the dose parameter adaptive optimization device 4 of the present invention includes a manifold construction module 41, a Fisher information matrix calculation module 42, a geodesic calculation module 43, and an information entropy evaluation module 44.

[0027] The manifold construction module 41 is used to map the number of high-risk population strata to an n-dimensional parameter space, constructing an n-dimensional information geometric manifold, where n is the number of high-risk population strata. In a preferred embodiment of the invention, when the number of high-risk population strata is 5, the constructed information geometric manifold is a 5-dimensional manifold, and the parameter vector... It includes five parameters: tube current, tube voltage, exposure time, pitch, and layer thickness.

[0028] The manifold construction process is as follows Figure 3 As shown, the main steps include: First, establishing an n-dimensional parameter vector. This includes dose-related parameters such as tube current and tube voltage; then, a family of probability distributions is constructed. , where y is the image quality quantization representation value; then, the parameter space Mapping to the family of probability distributions P forms an n-dimensional manifold M; finally, a local coordinate system is constructed on the manifold M, and a tangent space is established for each point p∈M. And define the inner product structure on the tangent space to form a Riemannian manifold.

[0029] The Fisher information matrix calculation module 42 is used to calculate the Fisher information matrix on the information geometric manifold, as a Riemannian metric. The Fisher information matrix is ​​a metric tensor on the Riemannian manifold, representing the local geometric structure of information in the parameter space. In this invention, the formula for calculating the n×n dimensional Fisher information matrix G(θ) is: , in, Let be the element in the i-th row and j-th column of the Fisher information matrix G(θ), E denotes the mathematical expectation operator, and p(y|θ) represent the conditional probability distribution of image quality y under parameter θ. This represents the log-likelihood function with respect to the parameter. The partial derivatives of , where i and j are parameter indices, with values ​​ranging from 1 to n.

[0030] In practical applications, since directly calculating the Fisher information matrix is ​​quite complex, this invention uses an empirical Fisher information matrix for approximate calculation: , Where N is the number of samples (generally ranging from 100 to 500, determined based on the amount of available data). This represents the image quality quantization value of the k-th sample, where k is the sample index, ranging from 1 to N. (Summarization symbol) This indicates that the summation operation is performed on all N samples.

[0031] The geodesic calculation module 43 is used to calculate the optimal path for dose parameters on the information geometric manifold. A geodesic is the shortest path between two points on a Riemannian manifold; in this invention, it represents the optimal path from the current dose parameter to the optimal dose parameter. The geodesic satisfies the following differential equation: , in, Let represent the i-th component of the parameter vector, and t be the parameter variable on the parameter path (its value range is usually normalized to [0,1]). This is the Christoffel notation (used to describe the connection structure on a Riemannian manifold). and Indicates parameters and The first derivative with respect to t Indicates parameters Second derivative with respect to t. Double summation and notation. This indicates that all combinations of j and k from 1 to n are summed.

[0032] Christoffel symbol Determined by the Fisher information matrix, the calculation formula is as follows: , in, These are the elements of the inverse of the Fisher information matrix. Represents the elements of the Fisher information matrix Regarding parameters The partial derivative of , where l is the parameter index, ranging from 1 to n. Summation symbol. This indicates that the summation is performed on all l from 1 to n.

[0033] In practical applications, this invention employs the fourth-order Runge-Kutta method to solve the geodesic equations. The specific steps are as follows: (1) Set the starting point (Current parameters) and endpoint (Target parameters); (2) Calculate the initial velocity vector ; (3) Solve the differential equation iteratively using the fourth-order Runge-Kutta method; (4) Generate from arrive The optimal parameter path.

[0034] The information entropy evaluation module 44 is used to calculate the image quality information entropy and compare it with a preset entropy threshold to determine whether the image quality is acceptable. In this invention, the image quality information entropy is calculated using the following expression: , Where y is the image quality quantization value, This represents the i-th discrete image quality level. It is the image quality quantization characterization value The probability is given by M, where M is the total number of image quality levels, and i is the quality level index, ranging from 1 to M. (Summarization symbol) This represents the summation of all M quality levels. The logarithmic function log represents the logarithmic operation with the natural logarithm as the base.

[0035] Information entropy characterizes the richness of information in an image; a higher entropy value indicates that the image contains more information. In this invention, the image quality information entropy threshold is set to 1.1. This threshold is determined based on the analysis of a large amount of clinical trial data, ensuring that image quality meets diagnostic requirements while reducing dosage. When the calculated image quality information entropy is greater than or equal to 1.1, the image quality is considered acceptable; when the entropy value is less than 1.1, the image quality is considered unacceptable, and dosage parameters need to be adjusted.

[0036] Example 3 In another embodiment of the present invention, the dose parameter adaptive optimization device 4 further includes a dual manifold mapping module 45, a mapping learning module 46, and a covariance structure optimization module 47.

[0037] The dual-manifold mapping module 45 is used to establish the mapping relationship between the dose parameter manifold S and the image quality manifold Q. The dimension of the dose parameter manifold S corresponds to the number of dose-related parameters (usually 5-7 dimensions), and the core parameters include tube current, tube voltage, exposure time, pitch, layer thickness, etc.; the dimension of the image quality manifold Q corresponds to the image quality evaluation indicators (usually 3-5 dimensions), and the core indicators include signal-to-noise ratio, contrast ratio, spatial resolution, entropy value, etc.

[0038] The dual-manifold mapping module 45 defines a mapping function F: S→Q, which maps points on the dose parameter manifold to the image quality manifold. The mathematical expression of the mapping function F is: , Where s∈S represents a point on the dose parameter manifold (an n-dimensional vector), q∈Q represents a point on the image quality manifold (an m-dimensional vector), and F represents a nonlinear mapping function from S to Q.

[0039] Jacobian matrix of mapping function F Represents a local linear approximation of the mapping: , in, It is an m×n dimensional Jacobian matrix. This represents the derivative matrix of the mapping function F with respect to the parameter s, and its elements are... Let represent the partial derivative of the i-th output component of the mapping function with respect to the j-th input parameter.

[0040] The Jacobian matrix describes how small changes in dose parameters affect image quality and is an important tool for understanding parameter sensitivity.

[0041] The mapping learning module 46 is used to train mapping relationships based on historical mapping data. In this invention, mapping learning employs a nonlinear regression method, utilizing the dose parameter-image quality pair from historical data. Train the mapping function F. Preferably, a radial basis function (RBF) network is used to implement the nonlinear mapping: , Where M is the number of radial basis functions (usually set to 50-200, depending on the model complexity and the amount of data). These are the weighting coefficients (determined through training). The center point is selected (usually from the training data using clustering methods). Radial basis functions (usually Gaussian functions are chosen) ,in This is a shape parameter, typically ranging from 0.1 to 10. Indicates s and The Euclidean distance between them, where i is the basis function index, ranging from 1 to M. Summation symbol. This indicates that the contributions to all M basis functions are summed.

[0042] The covariance structure optimization module 47 is used to calculate the covariance matrix of parameter changes and quality changes, and adjust the optimization path based on the covariance matrix. The formula for calculating the covariance matrix C is: , Where K is the sample size. To expand the sample vector (containing dose parameters and corresponding image quality, with dimensions n+m), The sample mean vector , Representing vectors The transpose of , where i is the sample index, ranging from 1 to K. Summation symbol. This indicates that the summation is performed on all K samples.

[0043] The covariance matrix C is a (n+m)×(n+m) dimensional symmetric matrix whose elements This represents the covariance between the i-th and j-th variables. The covariance matrix reflects the correlation between parameter changes and quality changes, providing guidance for the optimization path. In practical applications, this invention employs a covariance-guided path adjustment strategy: reducing the optimization step size in highly sensitive regions (regions with large covariance) to avoid unstable regions and ensure the stability and reliability of the optimization process.

[0044] Example 4 In another embodiment of the present invention, the dose parameter adaptive optimization device 4 further includes a genetic algorithm optimization module 48, a fitness function construction module 49, a geometric genetic operator module 50, and a population dynamics management module 51.

[0045] The genetic algorithm optimization module 48 is used to perform geometry-based genetic algorithm optimization on the information geometric manifold. For example... Figure 4 As shown, the geometry-based genetic algorithm optimization process includes steps such as initial population generation, fitness evaluation, selection operation, crossover operation, mutation operation, and termination judgment. Unlike traditional genetic algorithms, the genetic algorithm of this invention operates on a Riemannian manifold, taking into account the geometric structure of the parameter space, thus improving search efficiency and accuracy.

[0046] The fitness function construction module 49 is used to construct a fitness function based on geodesic distance. In this invention, the fitness function... The definition of is: , Where D is the measured value of image quality (usually a dimensionless indicator such as signal-to-noise ratio or contrast-to-noise ratio, typically ranging from 0 to 10). For the current parameter To ideal parameters The geodesic distance (distance calculated on a Riemannian manifold, units dependent on parameters), λ1 and The weighting coefficients are dimensionless, ranging from 0 to 1, and satisfy the following conditions: ).

[0047] During the optimization process, the weighting coefficients λ1 and λ2 are dynamically adjusted: in the initial stage, the focus is on optimizing image quality (λ1>λ2, typical values ​​λ1=0.7, λ2=0.3); in the intermediate stage, the balance between quality and dose is achieved (λ1≈λ2, typical values ​​λ1=0.5, λ2=0.5); and in the later stage, the focus is on dose optimization (λ1<λ2, typical values ​​λ1=0.3, λ2=0.7).

[0048] The geometric genetic operator module 50 is used to perform geodesic crossover, tangent space mutation, and geodesic selection operations. These geometrically guided genetic operators are another innovation of this invention, organically combining genetic algorithms with information geometry to improve optimization efficiency.

[0049] The specific steps of the geodesic crossover operation are as follows: Select two parent parameter points p1 and p2; calculate the geodesic line L connecting p1 and p2; select the crossover point c = α·p1 + (1-α)·p2 on the geodesic line L, where α is the geodesic line parameter and its value range is [0,1]; generate two child parameters, which are located on different segments of the geodesic line; verify the validity of the child parameter and make corrections if necessary.

[0050] The specific steps of the tangent space mutation operation are as follows: in the tangent space of parameter point p... Generate a random vector v; map v to a new point p' on the manifold using an exponential mapping; control the mutation magnitude to be inversely proportional to the local curvature of the manifold; perform boundary checks to ensure that the parameters are within the effective range; calculate the fitness value of the mutation point.

[0051] The specific steps of the geodesic selection strategy are as follows: calculate the selection probability of each individual based on the fitness value; apply geodesic distance weighted correction to the selection probability; use roulette or tournament methods to select individuals; retain elite individuals to ensure that the optimal solution is not lost; and dynamically adjust the selection pressure to balance exploration and development.

[0052] The population dynamics management module 51 is used to dynamically adjust the population size and distribution based on the local curvature of the manifold. In this invention, the basic population size N is set to 100-200 individuals, but it is dynamically adjusted according to the manifold complexity: the local population density is increased in high curvature regions (regions where the manifold is more curved) to improve search accuracy; and the population density is reduced in low curvature regions (regions where the manifold is relatively flat) to improve search efficiency.

[0053] In addition, the population dynamics management module 51 also implements population differentiation and merging strategies to avoid premature convergence: when the population diversity drops below a threshold (usually 30% of the initial diversity), a population differentiation operation is triggered to introduce new random individuals; when multiple subpopulations exist and their optimal solution similarity is higher than a threshold (usually 90%), a population merging operation is triggered to integrate search resources.

[0054] Example 5 like Figure 1 As shown, the high-risk population analysis device 3 of the present invention includes a deep learning network 31, a risk stratification module 32, a risk database 33, and a pre-scanning parameter generation module 34.

[0055] The deep learning network 31 is used to analyze the image data of patients to be diagnosed. In a preferred embodiment of the present invention, the deep learning network 31 is a generative adversarial network (GAN), which is used to acquire sample images of patients to be diagnosed after training for high-risk population analysis. The GAN consists of a generator network and a discriminator network. Through adversarial training, it learns the data distribution and can generate high-quality sample images, thereby improving the accuracy of analysis.

[0056] The risk stratification module 32 is used to stratify patients based on the analysis results of the deep learning network 31, generating a high-risk population stratification number. In this invention, the risk stratification adopts a multi-level stratification strategy, comprehensively assessing the risk level based on the patient's imaging characteristics, medical history, and genetic factors, and generating an integer risk stratification number ranging from 1 to 9.

[0057] Risk Database 33 is used to store risk stratification criteria and historical risk assessment data. Risk Database 33 contains risk assessment criteria for different disease types (such as lung cancer, breast cancer, colorectal cancer, etc.) and a large amount of historical case data, providing data support for risk stratification.

[0058] The pre-scan parameter generation module 34 generates pre-scan parameters based on the number of risk stratifications for high-risk individuals and controls the CT scanner 1 to generate pre-scan image data. Pre-scan parameters typically use a lower dose than the formal scan parameters to obtain basic image information for subsequent optimization. In this invention, the pre-scan parameters are dynamically adjusted according to the number of risk stratifications: high-risk patients (strata 7-9) use a relatively high pre-scan dose (approximately 30% of the standard dose); medium-risk patients (strata 4-6) use a medium pre-scan dose (approximately 20% of the standard dose); and low-risk patients (strata 1-3) use a lower pre-scan dose (approximately 10% of the standard dose).

[0059] Example 6

[0060] In another embodiment of the present invention, the automatic dose parameter adjustment device 5 includes a tube current adaptive adjustment module 52, a tube voltage adaptive adjustment module 53, and a scanning mode adaptive selection module 54.

[0061] The tube current adaptive adjustment module 52 is used to adjust the tube current of the CT scanner 1 according to the optimal dose parameters. Tube current is a major parameter affecting radiation dose and is directly proportional to it. In this invention, the tube current adjustment is based on manifold optimization results, and different adjustment strategies are adopted according to different risk levels of patients: high-risk patients (segment number 7-9) have a tube current set to 12-15 mA; medium-risk patients (segment number 4-6) have a tube current set to 8-12 mA; and low-risk patients (segment number 1-3) have a tube current set to 5-8 mA.

[0062] The tube voltage adaptive adjustment module 53 is used to adjust the tube voltage of the CT scanner 1 according to the optimal dose parameters. The tube voltage affects the penetration capability of X-rays and image contrast. In this invention, the tube voltage adjustment is also based on manifold optimization results, but also takes into account the patient's body size and the scanning site: for patients of the same risk level, the tube voltage is appropriately increased for larger patients (usually by 10-20 kV), and the tube voltage is appropriately decreased for smaller patients (usually by 10-20 kV).

[0063] The adaptive scanning mode selection module 54 is used to select one of the following scanning modes based on the number of slicing layers in a high-risk patient: plain CT scan, low-dose CT scan, or high-resolution CT scan. In this invention, the scanning mode selection strategy is as follows: high-risk patients (slicing layers 7-9) use plain CT scan, prioritizing image quality; medium-risk patients (slicing layers 4-6) use low-dose CT (LDCT) scan, balancing image quality and radiation dose; and low-risk patients (slicing layers 1-3) use high-resolution CT (HDCT) scan, minimizing radiation dose while ensuring basic image quality.

[0064] Example 7 In another embodiment of the present invention, the intelligent learning unit 6 includes an optimization result evaluation module 61, a Fisher information matrix update module 62, a mapping function optimization module 63, and a fitness function adjustment module 64.

[0065] The optimization result evaluation module 61 is used to evaluate the dose optimization effect and image quality. Evaluation indicators include dose reduction rate, image signal-to-noise ratio, and contrast-to-noise ratio. In this invention, the ideal optimization result should meet the following requirements: dose reduction rate not less than 30%, image signal-to-noise ratio decrease not exceeding 5%, and contrast-to-noise ratio decrease not exceeding 3%.

[0066] Fisher information matrix update module 62 is used to update the Fisher information matrix based on the latest scan results. Due to individual patient differences and changes in equipment status, the Fisher information matrix needs to be updated regularly to maintain its accuracy. In this invention, an incremental update strategy is adopted, which weights and fuses new data with historical data: , in, For the updated Fisher information matrix, This is the original Fisher information matrix. The Fisher information matrix calculated based on the new samples. This is the weighting coefficient (typically between 0.7 and 0.9, representing the proportion of historical data retained). The choice needs to balance the impact of historical experience and new data, with a larger... A smaller value helps maintain model stability. Values ​​help the model adapt to new data quickly.

[0067] The mapping function optimization module 63 is used to optimize the mapping function between the dose parameter manifold and the image quality manifold. As the sample data increases, the mapping function needs continuous optimization to improve mapping accuracy. In this invention, the mapping function optimization employs an online learning method, updating the parameters of the RBF network in real time. , in, and These are the weight coefficients before and after the update, respectively. The learning rate (usually set to 0.01 initially, gradually decreasing with each iteration) The actual observed image quality (m-dimensional vector). The image quality (m-dimensional vector) predicted by the current mapping function. Let be the value of the i-th basis function at point s, where i is the index of the basis function. This formula represents the adjustment of the weights based on the prediction error (qF(s)); the larger the error, the larger the adjustment.

[0068] The fitness function adjustment module 64 is used to adjust the weight coefficients of the fitness function based on historical optimization results. (Weight coefficients of the fitness function) and The optimization needs to be dynamically adjusted based on the results to balance the goals of image quality and dose reduction. In this invention, the weight adjustment strategy is as follows: if the image quality fails to meet the standard in N consecutive optimization results (typically N=5), then the weight is increased. (Increase by 0.05-0.1); If the dose reduction rate fails to meet the target in N consecutive optimization results, then increase... (Increase by 0.05-0.1). This adaptive adjustment strategy can dynamically balance quality and dosage targets based on actual optimization results, improving system adaptability.

[0069] Example 8 This invention also provides a CT dose adaptive optimization method, such as... Figure 5 As shown, the method includes the following steps: Step 101: Receive the image data of the patient to be diagnosed sent by the image data processing workstation.

[0070] Step 102: Perform high-risk population analysis based on the imaging data of the patients to be diagnosed, and generate high-risk population stratification numbers. Preferably, a generative adversarial network is used for high-risk population analysis to generate integer risk stratification numbers with values ​​ranging from 1 to 9.

[0071] Step 103: Generate pre-scanning parameters based on the stratification number of the high-risk population. The pre-scanning parameters are usually at a lower dose than the formal scan parameters, with the aim of obtaining basic image information for subsequent optimization.

[0072] Step 104: Control the CT scanner to perform a pre-scan based on the pre-scan parameters and generate pre-scan image data.

[0073] Step 105: Construct an information geometric manifold based on the stratification number of high-risk groups and pre-scan image data. The information geometric manifold includes a dose parameter manifold and an image quality manifold. For the specific construction process, please refer to the description in Example 2.

[0074] Step 106: Calculate the image quality information entropy on the information geometric manifold to determine whether the image quality is acceptable. The image quality information entropy is calculated using the following expression: , Where y is the image quality quantization value, This represents the i-th discrete image quality level. It is the image quality quantization characterization value The probability is given by M, where M is the total number of image quality levels (usually between 10 and 20, determined by the precision of the quality grading), and i is the quality level index, ranging from 1 to M. The summation symbol is used. This represents the summation of all M quality levels. The logarithmic function log represents the logarithmic operation with the natural logarithm as the base.

[0075] When the calculated image quality information entropy is greater than or equal to a preset threshold of 1.1, the image quality is considered acceptable; when the entropy value is less than 1.1, the image quality is considered unacceptable, and the dosage parameters need to be adjusted. The threshold of 1.1 was determined through extensive clinical trials and achieves a good balance between reducing dosage and ensuring image quality.

[0076] Step 107: Apply a geometry-based genetic algorithm to the information geometric manifold to calculate the optimal dose parameters. The geometry-based genetic algorithm is one of the core innovations of this invention; see Example 4 for a detailed description of the process.

[0077] Step 108: Output the optimal dose parameters to the automatic dose parameter adjustment device.

[0078] Step 109: Automatically adjust the scanning parameters of the CT scanner according to the optimal dose parameters. See Example 6 for the specific adjustment strategy.

[0079] Step 110: Update the information geometric manifold based on historical optimization results. See the description in Example 7 for the update process.

[0080] Through the above steps, the CT dose adaptive optimization method of the present invention can achieve precise adaptive optimization of CT dose parameters, significantly reduce radiation dose while ensuring image quality, and provide patients with a safer CT examination experience.

[0081] This invention is not limited to the above embodiments. Those skilled in the art can make modifications and variations to this invention without departing from the spirit and scope of this invention, and all such modifications and variations fall within the protection scope of this invention.

Claims

1. A CT dose adaptive optimization system, characterized in that, include: A CT scanner is used to perform CT scans based on scanning parameters and generate CT image data. An image data processing workstation, connected to the CT scanner, is used to process the CT image data; A high-risk population analysis device is connected to the image data processing workstation and is used to receive image data of patients to be diagnosed sent by the image data processing workstation, perform high-risk population analysis based on the image data of patients to be diagnosed, generate high-risk population stratification numbers, and send the high-risk population stratification numbers to the image data processing workstation. A dose parameter adaptive optimization device is connected to the high-risk population analysis device and the image data processing workstation. It is used to receive the high-risk population stratification number, receive pre-scanned image data, construct an information geometric manifold based on the high-risk population stratification number and the pre-scanned image data, calculate the optimal dose parameter on the information geometric manifold, and output the optimal dose parameter. The information geometric manifold includes a dose parameter manifold and an image quality manifold. An automatic dose parameter adjustment device, connected to the dose parameter adaptive optimization device and the CT scanner, is used to receive the optimal dose parameter and automatically adjust the scanning parameters of the CT scanner according to the optimal dose parameter; and The intelligent learning unit, connected to the dose parameter adaptive optimization device and the image data processing workstation, is used to update the information geometric manifold based on historical optimization results.

2. The system according to claim 1, characterized in that, The dose parameter adaptive optimization device includes: The manifold construction module is used to map the high-risk population stratification number to an n-dimensional parameter space and construct an n-dimensional information geometric manifold, where n is the high-risk population stratification number. The Fisher information matrix calculation module is used to calculate the Fisher information matrix on the information geometric manifold, and the Fisher information matrix is ​​used as a Riemannian metric. A geodesic calculation module is used to calculate the optimal path for dose parameters on the information geometric manifold; The information entropy evaluation module is used to calculate the image quality information entropy and compare it with a preset entropy threshold to determine whether the image quality is acceptable.

3. The system according to claim 2, characterized in that, The dose parameter adaptive optimization device further includes: A dual-manifold mapping module is used to establish a mapping relationship between the dose parameter manifold and the image quality manifold; The mapping learning module is used to train the mapping relationship based on historical mapping data; The covariance structure optimization module is used to calculate the covariance matrix of parameter changes and quality changes, and adjust the optimization path based on the covariance matrix.

4. The system according to claim 1, characterized in that, The dose parameter adaptive optimization device further includes: A genetic algorithm optimization module is used to perform geometry-based genetic algorithm optimization on the information geometric manifold; The fitness function construction module is used to construct fitness functions based on geodesic distance; The geometric genetic operators module is used to perform geodesic crossover, tangent space mutation, and geodesic selection operations. The population dynamics management module is used to dynamically adjust the population size and distribution based on the local curvature of the manifold.

5. The system according to claim 1, characterized in that, The high-risk population analysis device includes: A deep learning network is used to analyze the image data of the patient to be diagnosed; The risk stratification module is used to stratify patients based on the analysis results of the deep learning network and generate the number of stratifications for the high-risk population. A risk database is used to store risk stratification criteria and historical risk assessment data; The pre-scan parameter generation module is used to generate pre-scan parameters based on the number of stratifications of the high-risk population and to control the CT scanner to generate the pre-scan image data.

6. The system according to claim 5, characterized in that, The deep learning network of the high-risk population analysis device is a generative adversarial network, which is used to obtain sample images of the patients to be diagnosed after training for high-risk population analysis.

7. The system according to claim 1, characterized in that, The expression for the image quality information entropy on the image quality manifold is: H(y)=-∑p(y)log(p(y)) Where y is the image quality quantization representation value, p(y) is the probability of the image quality quantization representation value y, and the image quality information entropy threshold is 1.

1.

8. The system according to claim 1, characterized in that, The automatic dose parameter adjustment device includes: Tube current adaptive adjustment module, used to adjust the tube current of the CT scanner according to the optimal dose parameters; A tube voltage adaptive adjustment module is used to adjust the tube voltage of the CT scanner according to the optimal dose parameters; The adaptive scanning mode selection module is used to select one of the following scanning modes based on the number of stratifications of the high-risk population: plain CT scan, low-dose CT scan, or high-resolution CT scan.

9. The system according to claim 1, characterized in that, The intelligent learning unit includes: The optimization results evaluation module is used to evaluate the effect of dose optimization and image quality; The Fisher information matrix update module is used to update the Fisher information matrix based on the latest scan results. A mapping function optimization module is used to optimize the mapping function between the dose parameter manifold and the image quality manifold; The fitness function adjustment module is used to adjust the weight coefficients of the fitness function based on historical optimization results.

10. A CT dose adaptive optimization method, characterized in that, include: Receive image data of patients to be diagnosed from the image data processing workstation; Based on the imaging data of the patients to be diagnosed, high-risk population analysis is performed to generate high-risk population stratification numbers; Pre-scanning parameters are generated based on the stratification number of the high-risk population; The CT scanner is controlled to perform a pre-scan based on the pre-scan parameters, generating pre-scan image data. An information geometric manifold is constructed based on the high-risk population stratification number and the pre-scanned image data, the information geometric manifold including a dose parameter manifold and an image quality manifold; Calculate the image quality information entropy on the information geometric manifold to determine whether the image quality is acceptable; The optimal dose parameters are calculated by applying a geometry-based genetic algorithm on the information geometric manifold. The optimal dose parameter is output to the automatic dose parameter adjustment device; The scanning parameters of the CT scanner are automatically adjusted according to the optimal dose parameters. as well as The information geometric manifold is updated based on historical optimization results.

Images