Photon computed tomography low dose low concentration contrast agent protocol for assessment of coronary stenosis

Abstract

The application discloses a photon CT low-dose low-concentration contrast agent scheme coronary stenosis evaluation system and belongs to the medical image field.The photon CT low-dose low-concentration contrast agent scheme coronary stenosis evaluation system comprises a multi-energy spectrum data acquisition unit, a data preprocessing unit, an image generation unit, an image segmentation unit, a plaque analysis unit and a stenosis detection unit.The application solves the problems of high radiation dose, large amount of contrast agent and low efficiency and strong subjectivity caused by high dependence on the experience of doctors in the existing coronary CT examination, realizes high-quality coronary imaging and accurate stenosis evaluation under the condition of low dose and low-concentration contrast agent by adopting a photon counting CT detector and optimized low-dose scanning parameters, combining a deep learning algorithm and a multi-material decomposition technology, significantly reduces the radiation exposure and contrast agent related risk of patients, and improves the accuracy and reliability of coronary stenosis detection and plaque component analysis.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, specifically to an assessment system for coronary artery stenosis using a low-dose, low-concentration contrast agent regimen in photonic CT. Background Technology

[0002] Coronary artery disease is one of the leading causes of cardiovascular death worldwide, and its early and accurate diagnosis is crucial for clinical treatment and prognosis. Currently, coronary CT angiography is the primary non-invasive imaging method for diagnosing coronary artery stenosis. However, traditional CT scans require high doses of radiation and high concentrations of iodine contrast agents, increasing the potential risk of radiation damage and contrast agent nephropathy in patients. Furthermore, conventional energy integrating detectors suffer from limited energy resolution and susceptibility to noise interference, affecting the accurate identification of small plaque components and the accurate assessment of stenosis severity. In recent years, the emergence of photon counting CT technology has provided a new direction for solving these problems. It can achieve multi-spectral imaging, possessing higher energy resolution and sensitivity, and is expected to significantly reduce radiation dose and contrast agent usage while improving image quality and diagnostic accuracy. However, existing photon counting CT technology heavily relies on manual operation by physicians during coronary artery stenosis detection, which is time-consuming and subject to subjective differences; therefore, it does not meet current needs. To address this, we propose a photon CT low-dose, low-concentration contrast agent scheme for the assessment of coronary artery stenosis. Summary of the Invention

[0003] The purpose of this invention is to provide an evaluation system for coronary artery stenosis using a low-dose, low-concentration contrast agent regimen in photonic CT. By integrating and optimizing a low-dose scanning protocol, a low-concentration contrast agent injection protocol, and a deep learning-based artificial intelligence-based end-to-end analysis, the system solves the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen in photonic CT, comprising:

[0005] The multi-spectral data acquisition unit is configured to use a photon-counting CT detector and scan patients who have been injected with low-concentration contrast agents based on optimized low-dose scanning parameters to acquire multi-spectral image data.

[0006] The data preprocessing unit is configured to perform big data acquisition and preprocessing on the acquired multi-energy spectrum image data, including image enhancement, noise suppression and energy spectrum correction;

[0007] The image generation unit is configured to perform multi-material decomposition on the base material image data to generate a variety of derived images, including virtual single-level images, iodine images, and calcification images. The multi-material decomposition uses a deep learning algorithm to perform high-precision separation of iodine, calcium, and soft tissue components using a pre-trained convolutional neural network model.

[0008] The image segmentation unit is configured to perform image recognition based on the virtual single-level image, and automatically identify and segment the coronary artery tree structure and the vessel centerline.

[0009] The plaque analysis unit is configured to perform qualitative and quantitative analysis of plaque components based on the segmented vascular region and by fusing information from the iodine map and calcification map, including calcification score calculation and differentiation between non-calcified plaques and mixed plaques.

[0010] The stenosis detection unit is configured to automatically detect the location of stenosis on the centerline of the segmented blood vessel and to quantitatively assess the degree of stenosis based on changes in the lumen diameter.

[0011] Furthermore, the multi-energy spectrum data acquisition unit includes:

[0012] The scanning parameter optimization module is configured to collect patients' physiological characteristics and clinical needs and match them with the historical scanning database. Using a pre-trained deep neural network model, with patients' physiological characteristics and clinical needs as input, it analyzes the complex nonlinear relationship between different parameter combinations in historical scanning data and the final image quality, and predicts the optimal combination of scanning parameters that achieves the lowest radiation dose while meeting the requirements of diagnostic image quality. This includes specific values ​​for tube voltage, tube current, and exposure time.

[0013] The edge computing module is configured to be deployed on the CT equipment and relies on the integrated photon counting CT detector's dedicated smart chip to perform real-time preprocessing of the raw photon counting signal at the data acquisition source, initially completing energy spectrum analysis and noise filtering;

[0014] The data calibration module is configured to be integrated into the aforementioned smart chip to perform real-time calibration on the acquired multi-energy spectrum image data, including energy correction, gain correction and dark current correction, in order to eliminate drift and system errors of the photon counting CT detector.

[0015] The data quality control module is configured to perform real-time quality assessment on the acquired multi-energy spectrum image data, check the integrity of the data, automatically identify and mark abnormal points in the data based on the inspection, and perform preliminary screening and correction of the data according to preset standards.

[0016] Furthermore, the data quality control module executes the following process:

[0017] Receive multi-energy spectrum image data, parse the data packets, extract key information, including the data dimension, number of energy channels and scanning parameters, and ensure that the data format meets the preset standards;

[0018] Compare the expected size and actual size of the multi-energy spectral image data to check the integrity of the multi-energy spectral image data.

[0019] Based on the integrity check results, outliers in the data are automatically identified, their locations and types are recorded in detail, and these outliers are marked in the data. The marking information of the outliers will include the outlier type and location information.

[0020] Based on preset quality standards, the multi-energy spectrum image data is initially screened to remove data segments and channels that do not meet the quality requirements.

[0021] The filtered data is then corrected, including signal smoothing and contrast adjustment.

[0022] After the calibration is completed, the calibrated multi-energy spectral image data will be re-evaluated to ensure that the quality of the multi-energy spectral image data meets the preset standards.

[0023] If the data still has quality issues, feedback will be sent to the multi-energy spectrum data acquisition unit, prompting for re-acquisition or adjustment of acquisition parameters.

[0024] Further, the image generation unit includes:

[0025] The multi-material decomposition module is configured to process the base material image data using a pre-trained convolutional neural network model. The convolutional neural network model automatically extracts features through deep learning algorithms to separate iodine, calcium and soft tissue components in the image and generate corresponding material composition maps.

[0026] The derived image generation module is configured to generate virtual single-level images, iodine maps, and calcification maps based on the decomposition results of the multi-material decomposition module.

[0027] The quality assessment module is configured to assess the quality of the generated virtual single-level images, iodine maps, and calcification maps, checking the image contrast, resolution, and noise level. Based on the assessment results, it automatically adjusts the image processing parameters to optimize image quality.

[0028] Furthermore, the image segmentation unit includes:

[0029] The image recognition module is configured to detect the contours and branches of blood vessels based on the generated virtual single-level image, using machine learning algorithms combined with edge detection algorithms, thereby identifying the coronary vascular structure, including the starting point, branching point and termination point of the coronary vascular tree;

[0030] The vessel segmentation module is configured to segment the boundaries of the coronary vessels and extract the vessel regions based on the identified coronary vessel structures using a region growing method, while simultaneously extracting the vessel centerline based on a distance transformation algorithm.

[0031] The segmentation optimization module is configured to use a multi-scale optimization algorithm to optimize the segmented coronary artery tree structure, including smoothing, artifact and noise removal, and topology optimization of the vascular tree in combination with a vascular anatomy model.

[0032] Furthermore, the patch analysis unit performs the following process:

[0033] Receive the blood vessel region segmentation results output by the image segmentation unit, and spatially register the blood vessel region segmentation results with the iodine map and calcification map;

[0034] The registered iodine map and calcification map are analyzed at the pixel level within the vascular region. By using a preset threshold, the distribution areas of calcified plaques, non-calcified plaques, and mixed plaques are identified.

[0035] For calcified patches, the calcification integral value is obtained by counting the number of pixels in the calcified area and combining it with the corresponding weighting factor.

[0036] For non-calcified and mixed plaques, the distribution of iodine content on the iodine map is analyzed, and combined with the calcification information in the calcification map, a qualitative description of the plaque components is achieved, including lipid and fibrous components. Quantitative analysis is then performed by measuring the area and volume parameters of the plaque region.

[0037] Furthermore, the narrow detection unit performs the following process:

[0038] The image segmentation unit outputs the coronary artery centerline and vascular tree structure, which serve as the basic framework for stenosis detection.

[0039] Along the centerline of the blood vessel, a multiplanar reconstruction technique was used to extract a sequence of cross-sectional images of the blood vessel from multi-energy spectral image data;

[0040] On each cross-sectional image, the boundary of the blood vessel lumen is identified by an edge detection algorithm, and then the diameter of the blood vessel lumen is calculated.

[0041] Then, a continuous analysis of the lumen diameter data of multiple cross-sectional images along the blood vessel centerline is performed. By calculating the rate of change of lumen diameter between adjacent cross-sections, the location of stenosis is identified.

[0042] When the rate of change of the lumen diameter exceeds a preset threshold, the location is determined to be a narrow point;

[0043] A quantitative index of the degree of stenosis is calculated by comparing the lumen diameter at the narrowing point with the lumen diameter of the normal reference segment.

[0044] Stenosis degree = [(normal lumen diameter - lumen diameter at stenosis) / normal lumen diameter] × 100%.

[0045] Furthermore, the data preprocessing unit includes:

[0046] The data acquisition module is configured to deploy multiple distributed data acquisition nodes on the photon counting CT detector array to form a distributed sensor network architecture. It is used to synchronously receive the raw photon counting signal and to aggregate the signal of a specific detector area for each node, so as to realize the acquisition of multi-spectral image data.

[0047] The data processing module is configured to perform image enhancement, noise suppression, and energy spectrum correction on the acquired multi-energy spectrum image data.

[0048] Furthermore, the vascular region segmentation results are spatially registered with the iodine map and calcification map, including:

[0049] Obtain the segmentation results, iodine map, and calcification map of the vascular region;

[0050] Spatially align the vascular region segmentation results, iodine map, and calcification map;

[0051] The spatially aligned vascular region segmentation results are then subjected to vascular wall layer segmentation to obtain segmentation results of the intima, media and adventitia layers.

[0052] The spatially aligned calcification map is used to perform stratified localization and hardness grading of calcifications. Multi-scale Frangi filtering is used to enhance the features of calcifications, mark the interaction area between calcifications and the vessel wall, and generate pathological anchor points with hardness grade and the layer parameters of the vessel wall in which they are located.

[0053] An iodine concentration-blood flow velocity correlation model is established on the spatially aligned iodine map. Gradient direction adaptive filtering is used to enhance the edges of the iodine map, which is then transformed into a blood flow velocity distribution and marked in areas of low blood perfusion. Functional anchor points with mean blood flow velocity and iodine concentration parameters are generated.

[0054] Based on the aforementioned layered segmentation results, pathological anchor points, and functional anchor points, dynamic structural anchor points are generated, including intimal inflection points with calcified attachment sites and key hemodynamic parameters.

[0055] Construct cross-modal anchor point morphological association rules;

[0056] Key points with attached spatial-mechanical-functional feature vectors are extracted from the layered segmentation results, enhanced iodine map, and enhanced calcification map, respectively. The feature vectors include three-dimensional coordinates, the vascular layer, and corresponding modality-specific parameters.

[0057] Define the cross-modal keypoint distance;

[0058] Based on cross-modal key point distance combined with morphological association rule constraints, corresponding point pairs are selected by combining the nearest neighbor ratio method with RANSAC denoising.

[0059] Based on the corresponding point pairs, the local bending energy of the blood vessel is calculated, and the geometric shape of the blood vessel is corrected by the deformation compensation model.

[0060] Singular value decomposition (SVD) is used to solve the rigid transformation matrix. Based on the rigid transformation matrix, the iodine map and calcification map are first registered. The optimization objective is to minimize the mean square error between the segmentation result and the transformed image.

[0061] Construct a layered elastic displacement field model of the blood vessel wall;

[0062] Based on the layered elastic displacement field model of the blood vessel wall, the iodine map and calcification map after the first registration are re-registered to obtain the iodine map and calcification map after the second registration.

[0063] Furthermore, the quality of the generated virtual single-level images, iodine maps, and calcification maps is evaluated, including:

[0064] Virtual single-level images, iodine maps, and calcification maps are used as multimodal images;

[0065] Obtain the contrast, resolution, and noise level of each modality in a multimodal image;

[0066] The multimodal image quality co-factor is calculated based on the contrast, resolution, and noise level of each modality.

[0067] ;

[0068] in, Represents the multimodal image quality co-operation coefficient; Represents a virtual single-level image; Iodine diagram; Represents a calcification diagram; Represents modal weights; This represents the contrast of a single-modal image after normalization. This represents the resolution of the single-modal image after normalization. The normalized representation shows the noise level of a single-modal image. Indicates the interaction coefficient of the indicators; Represents the minimum constant;

[0069] Calculate the comprehensive multimodal image quality index based on the aforementioned multimodal image quality synergy coefficient;

[0070] ;

[0071] in, This represents the comprehensive quality index of multimodal images; Indicates the noise suppression correction coefficient; Represents the modal fit coefficient; This represents the average resolution of the three types of images after normalization. This represents the average noise level of the three types of images after normalization. , , These represent the maximum, minimum, and mean contrast values ​​of the three types of images after normalization. , Represents the minimum constant; Indicates the contrast uniformity coefficient; This represents a monotonic calibration function. , , These are calibration coefficients;

[0072] The multimodal image quality comprehensive index is compared with a preset comprehensive index threshold. If the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold, the multimodal image is deemed qualified. If the multimodal image quality comprehensive index is less than the preset comprehensive index threshold, the multimodal image parameters are adjusted until the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold.

[0073] Compared with the prior art, the beneficial effects of the present invention are:

[0074] This invention introduces a deep neural network-based scanning parameter optimization module into the multi-energy spectrum data acquisition unit. Based on the patient's individual physiological characteristics and clinical needs, it dynamically predicts and employs optimal scanning parameters with the lowest radiation dose and lowest contrast agent concentration. This significantly reduces the patient's radiation exposure risk and the possibility of contrast agent nephropathy while ensuring diagnostic image quality, thereby improving examination safety. Furthermore, by deeply integrating deep learning artificial intelligence algorithms into various stages of image generation and analysis, it achieves full automation from automatic coronary artery tree segmentation and precise qualitative and quantitative analysis of plaque components to automatic detection and quantification of stenosis location and severity. This reduces reliance on manual operation by physicians, significantly improving diagnostic efficiency, shortening report generation time, and effectively avoiding assessment biases caused by differences in operator subjective experience, thus enhancing the objectivity, accuracy, and repeatability of diagnostic results. Attached Figure Description

[0075] Figure 1 This is a structural diagram of the photonic CT low-dose, low-concentration contrast agent scheme of the present invention for evaluating coronary artery stenosis. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] To address the issues of high radiation doses, large contrast agent usage, and low efficiency and subjectivity due to the high reliance on physician experience in current coronary CT examinations, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution:

[0078] A photon CT low-dose, low-concentration contrast agent regimen for assessing coronary artery stenosis includes:

[0079] The multi-spectral data acquisition unit is configured to use a photon-counting CT detector and scan patients who have been injected with low-concentration contrast agents based on optimized low-dose scanning parameters to acquire multi-spectral image data.

[0080] The data preprocessing unit is configured to perform big data acquisition and preprocessing on the acquired multi-energy spectrum image data, including image enhancement, noise suppression and energy spectrum correction;

[0081] The image generation unit is configured to perform multi-material decomposition on the base material image data to generate a variety of derived images, including virtual single-level images, iodine images, and calcification images. The multi-material decomposition uses a deep learning algorithm to perform high-precision separation of iodine, calcium, and soft tissue components using a pre-trained convolutional neural network model.

[0082] The image segmentation unit is configured to perform image recognition based on the virtual single-level image, and automatically identify and segment the coronary artery tree structure and the vessel centerline.

[0083] The plaque analysis unit is configured to perform qualitative and quantitative analysis of plaque components based on the segmented vascular region and by fusing information from the iodine map and calcification map, including calcification score calculation and differentiation between non-calcified plaques and mixed plaques.

[0084] The stenosis detection unit is configured to automatically detect the location of stenosis on the centerline of the segmented blood vessel and to quantitatively assess the degree of stenosis based on changes in the lumen diameter.

[0085] The results fusion unit is configured to fuse plaque component analysis results, stenosis quantification results, and virtual single-level images, and generate a comprehensive evaluation report, specifically:

[0086] It receives qualitative and quantitative data on plaque composition from the plaque analysis unit and quantitative data on the location and extent of stenosis from the stenosis detection unit;

[0087] Using a three-dimensional spatial registration algorithm, the results of multimodal analysis are mapped and superimposed onto a high-resolution virtual single-level image to generate a comprehensive visualization model that integrates anatomical structure and functional information.

[0088] In the comprehensive visualization model, patches of different components are rendered and distinguished with different colors or transparency, while narrow areas are highlighted with a high-brightness outline.

[0089] Based on a predefined clinical diagnostic rule base and knowledge graph, comprehensive logical reasoning and correlation analysis are performed on all the fused data to automatically generate structured diagnostic conclusions and risk assessments. The conclusions include assessment of the hemodynamic significance of stenosis, evaluation of plaque vulnerability, and recommended subsequent treatment pathways.

[0090] Finally, the comprehensive visualization model, key parameter list, and automated diagnostic conclusions are integrated to output a comprehensive evaluation report that supports interactive browsing and one-click export.

[0091] The technical effects of the above-mentioned solution are as follows: By employing an advanced photon-counting CT detector and optimized low-dose scanning parameters, it can significantly reduce the risk of radiation exposure and contrast agent-induced kidney injury in patients. Simultaneously, it utilizes deep learning-based multi-material decomposition technology to extract iodine, calcium, and soft tissue component information from low-dose, low-concentration data with high precision. This generates high-quality virtual single-level images, iodine maps, calcification maps, and other multi-parameter derived images. Furthermore, through fully automated coronary artery segmentation, qualitative and quantitative analysis of plaque components, and precise quantification of stenosis, an intelligent assessment system integrating data acquisition, processing, and analysis is constructed. This eliminates the subjective differences and time consumption caused by the high reliance on manual operation by physicians in traditional methods. Thus, while ensuring the accuracy of coronary artery stenosis diagnosis and the ability to identify plaque components, it achieves standardization, efficiency, and intelligence in the examination process, providing strong technical support for the early and accurate diagnosis and risk assessment of coronary artery disease.

[0092] The multi-energy spectrum data acquisition unit includes:

[0093] The scanning parameter optimization module is configured to collect patients' physiological characteristics (such as weight, age, and heart rate) and clinical needs and match them with the historical scanning database. Using a pre-trained deep neural network model, with patients' physiological characteristics and clinical needs as input, it analyzes the complex nonlinear relationship between different parameter combinations in historical scanning data and the final image quality to predict the optimal combination of scanning parameters that achieves the lowest radiation dose while meeting diagnostic image quality requirements. This includes specific tube voltage, tube current, and exposure time values. After scanning starts, it continuously monitors the raw data signals during the scanning process and fine-tunes the initial parameters using the same deep neural network model through real-time feedback loops to dynamically respond to possible physiological changes in the patient (such as heart rate fluctuations) and ensure that the entire acquisition process is always in an optimal state.

[0094] The edge computing module is configured to be deployed on the CT equipment and relies on the integrated photon counting CT detector's dedicated smart chip to perform real-time preprocessing of the raw photon counting signal at the data acquisition source, initially completing energy spectrum analysis and noise filtering;

[0095] The data calibration module, configured to be integrated into the aforementioned smart chip, performs real-time calibration on the acquired multi-energy spectrum image data, including energy correction, gain correction, and dark current correction, to eliminate drift and systematic errors in the photon-counting CT detector. Specifically:

[0096] Energy correction is applied in real time to the raw signal sequence output by each pixel unit of the photon counting CT detector. By looking up a pre-calibrated energy-channel address correspondence table, the signal is accurately mapped to the standard energy value.

[0097] Gain correction is performed by monitoring the reference signal in real time and applying a gain compensation coefficient to ensure that the responses of each detector unit are consistent.

[0098] The dark current signal is continuously collected and subtracted to eliminate the noise generated by the detector itself when there is no X-ray irradiation;

[0099] The above series of calibration operations work together to eliminate detector energy response drift, inter-unit gain differences and system inherent noise in real time, and ensure the accuracy and consistency of subsequently acquired multi-energy spectrum image data in both the energy spectrum and spatial dimensions.

[0100] The data quality control module is configured to perform real-time quality assessment on the acquired multi-energy spectrum image data, check the integrity of the data, automatically identify and mark abnormal points in the data based on the inspection, and perform preliminary screening and correction of the data according to preset standards.

[0101] The technical effects of the above-mentioned solution are as follows: The scanning parameter optimization module based on deep neural networks can accurately predict and dynamically adjust the lowest radiation dose scheme while ensuring the quality of diagnostic images, based on the individual physiological characteristics and clinical needs of the patient. This achieves personalized and minimized dose. The edge computing module, combined with a dedicated smart chip, can complete efficient real-time preprocessing at the data acquisition source, thereby significantly improving signal quality and reducing the computational burden on the back-end system. At the same time, the data calibration module integrated into the chip can eliminate the inherent drift and response inconsistency of the photon counting CT detector through real-time calibration, thereby ensuring the high fidelity of multi-spectral data in both energy spectrum and spatial dimensions, thus laying a reliable foundation for subsequent precise material separation. The real-time evaluation and intervention of the data quality control module can immediately identify and mark anomalies, thereby ensuring that the data quality input into the processing flow is always under control and at a high standard.

[0102] In summary, the collaborative work of the scanning parameter optimization module, data calibration module, and data quality control module constitutes an intelligent and robust data acquisition front-end, which not only significantly reduces the patient's radiation risk and contrast agent burden, but also ensures the accuracy, reliability, and consistency of all subsequent analysis and diagnostic results.

[0103] The data quality control module executes the following process:

[0104] Receive multi-energy spectrum image data, parse the data packets, extract key information, including the data dimension, number of energy channels and scanning parameters, and ensure that the data format meets the preset standards;

[0105] Compare the expected size and actual size of the multi-energy spectral image data to check the integrity of the multi-energy spectral image data.

[0106] Based on the integrity check results, outliers in the data are automatically identified, their locations and types are recorded in detail, and these outliers are marked in the data. The marking information of the outliers will include the anomaly type (such as excessive noise, missing data, etc.) and location information.

[0107] Based on preset quality standards, the multi-energy spectrum image data is initially screened to remove data segments and channels that do not meet the quality requirements.

[0108] The filtered data is then corrected, including signal smoothing (such as median filtering) and contrast adjustment.

[0109] After the calibration is completed, the calibrated multi-energy spectral image data will be re-evaluated to ensure that the quality of the multi-energy spectral image data meets the preset standards.

[0110] If the data still has quality issues, feedback will be sent to the multi-energy spectrum data acquisition unit, prompting for re-acquisition or adjustment of acquisition parameters.

[0111] The technical effects of the above solution are as follows: By parsing and checking the integrity of the data, the basic compliance of the data can be guaranteed. With the help of automated anomaly detection and fine marking, specific problems such as excessive noise and missing data can be accurately located, thus providing clear targets for subsequent screening and correction. Based on preset quality standards, preliminary screening and targeted correction processing such as signal smoothing and contrast adjustment can effectively improve the signal-to-noise ratio and consistency of the data. Finally, through a closed-loop feedback mechanism, the corrected data is iteratively evaluated, and when the quality does not meet the standards, re-acquisition or parameter adjustment is actively triggered, ensuring that the quality of multi-energy spectrum image data is controllable and reliable throughout the entire process from reception to output.

[0112] Image generation unit, including:

[0113] The multi-material decomposition module is configured to process the base material image data using a pre-trained convolutional neural network model. The convolutional neural network model automatically extracts features through deep learning algorithms to separate iodine, calcium and soft tissue components in the image and generate corresponding material composition maps.

[0114] The convolutional neural network model adopts the U-Net architecture, which includes an encoder-decoder structure. It learns a large amount of labeled data through pre-training. The model training process is based on transfer learning on the training set and performance is evaluated through the test set and validation set. It supports self-training and self-learning to adapt to different patient data.

[0115] The derived image generation module is configured to generate virtual single-level images, iodine maps, and calcification maps based on the decomposition results of the multi-material decomposition module, specifically:

[0116] Iodine composition information was extracted from the multi-material decomposition results, and an iodine map was generated;

[0117] Extract calcium composition information from multi-material decomposition results and generate calcification maps;

[0118] Iodine distribution map and calcification distribution map were extracted from the base material image, respectively;

[0119] The quality assessment module is configured to evaluate the quality of the generated virtual single-level images, iodine maps, and calcification maps, checking the image contrast, resolution, and noise level. Based on the evaluation results, it automatically adjusts image processing parameters to optimize image quality. Specifically:

[0120] For each generated derived image (including virtual single-level images, iodine maps, and calcification maps), a set of quantitative quality indicators are extracted simultaneously, including:

[0121] Calculate the signal-to-noise ratio and contrast-to-noise ratio of the image to objectively assess the noise level;

[0122] Measure the signal intensity value of the region of interest within a standard area of ​​known anatomical structure to evaluate image contrast;

[0123] The spatial resolution of an image is evaluated by analyzing the sharpness of its edges or by using a modulation transfer function.

[0124] The above quantitative indicators are compared with the preset diagnostic quality thresholds. Based on the comparison results, a comprehensive quality score is automatically generated for each image, and images that do not meet the standards and their specific non-compliance items are marked (such as insufficient contrast-to-noise ratio of iodine images and artifacts in calcification images). The thresholds are set differently according to different image types and clinical assessment goals.

[0125] Based on the quality assessment results, the image processing parameters are automatically adjusted and fed back to the data preprocessing unit, triggering the reprocessing of the original data.

[0126] After one or more rounds of parameter optimization and image reprocessing, the final version of the image set is subjected to a final quality assessment. Only when all key quality indicators meet the preset standards will the image data be approved for output and transmitted to the downstream image segmentation unit.

[0127] The technical effects of the above solution are as follows: By introducing a deep learning model based on the U-Net architecture for multi-material decomposition, iodine, calcium, and soft tissue components can be separated from the matrix material data acquired by photon-counting CT with high precision and automation. This overcomes the problem of decreased accuracy caused by noise interference and spectral overlap under low-dose and low-contrast conditions in traditional decomposition methods, thereby improving the accuracy and robustness of the material composition map. On this basis, the derived image generation module uses the decomposition results to generate virtual single-level images, iodine maps, and calcification maps, thus providing rich imaging evidence for subsequent diagnosis. The quality assessment module quantitatively evaluates key indicators such as image contrast, resolution, and noise level, and automatically adjusts processing parameters for iterative optimization, ensuring that all the final output derived images always meet strict diagnostic quality requirements, thereby guaranteeing the accuracy and reliability of subsequent coronary artery stenosis and plaque analysis results.

[0128] Image segmentation unit, including:

[0129] The image recognition module is configured to detect the contours and branches of blood vessels based on the generated virtual single-level image, using machine learning algorithms combined with edge detection algorithms, thereby identifying the coronary vascular structure, including the starting point, branching point and termination point of the coronary vascular tree;

[0130] The vessel segmentation module is configured to segment the boundaries of the coronary vessels and extract the vessel regions based on the identified coronary vessel structures using a region growing method, while simultaneously extracting the vessel centerline based on a distance transformation algorithm.

[0131] The segmentation optimization module is configured to use a multi-scale optimization algorithm to optimize the segmented coronary artery tree structure, including smoothing, artifact and noise removal, and topology optimization of the vascular tree in combination with a vascular anatomy model.

[0132] In this embodiment, the starting point and branch point of the blood vessel determined by the image recognition module are used as seed points. The three-dimensional segmentation of the blood vessel lumen is performed using an adaptive threshold region growing algorithm. This algorithm dynamically adjusts the growth threshold according to the local image grayscale statistical characteristics to ensure that the blood vessel boundary can be accurately defined even in low-contrast areas and small branches, thereby completely extracting the lumen region of the entire coronary vascular tree. After obtaining the three-dimensional mask of the blood vessel region, a three-dimensional Euclidean distance transformation is performed to calculate the distance from each voxel in the mask to the nearest blood vessel boundary. Then, by tracking the voxel point with the local maximum distance value in each blood vessel segment and connecting them to form a continuous central path, the blood vessel centerline representing the geometric central axis of the blood vessel is accurately extracted.

[0133] In this embodiment, surface fitting technology is used to smooth the initial vessel boundaries obtained by region growing, in order to eliminate jagged artifacts caused by noise while preserving the true anatomical structure. For the vessel centerline, moving average filtering or curve smoothing algorithm based on the minimum energy principle is applied to ensure the continuity and smoothness of the centerline and avoid unreasonable sharp turns. The segmentation results are compared with a preset coronary artery anatomical statistical shape model, and graph theory algorithms are used to analyze the topological connectivity of the centerline branches. False branches or missing branches caused by image noise or segmentation errors are automatically detected and corrected to ensure that the topological structure of the vessel tree is consistent with the real anatomical structure.

[0134] The technical effects of the above solution are as follows: The image recognition module integrates machine learning and edge detection algorithms, which can intelligently identify the starting point, branch point, and complete tree structure of the coronary artery, thus laying an accurate topological foundation for subsequent segmentation. The vessel segmentation module combines region growing and distance transformation algorithms, which not only accurately delineates the inner and outer boundaries of the vessel to extract the complete vessel region, but also simultaneously generates an accurate vessel centerline, thus providing a key geometric framework for subsequent stenosis detection. The segmentation optimization module, through multi-scale optimization and the introduction of anatomical prior knowledge, can eliminate local segmentation errors caused by image noise, artifacts, or partial volume effects, and make reasonable corrections to the topological structure of the vessel tree, thereby ensuring a high degree of consistency between geometric accuracy and anatomical realism in the final segmentation result, and thus providing a reliable and complete anatomical basis for subsequent accurate plaque analysis and quantitative assessment of stenosis.

[0135] The plaque analysis unit executes the following process:

[0136] The system receives the blood vessel region segmentation results output by the image segmentation unit, and performs spatial registration between the blood vessel region segmentation results and the iodine map and calcification map to ensure that the three are accurately aligned in anatomical position.

[0137] Pixel-level analysis is performed on the registered iodine and calcification maps within the vascular region. By using preset thresholds, the distribution areas of calcified plaques, non-calcified plaques, and mixed plaques are identified. The threshold setting process analyzes the distribution characteristics of the corresponding CT values ​​of known plaque component regions in the iodine and calcification maps. Using a clustering algorithm, the natural aggregation centers of different components (such as calcification, lipid-rich necrotic cores, and fibrous tissue) in the feature space are identified. The process also takes into account individual patient differences (such as contrast agent injection protocols and different enhancement levels caused by cardiac output) and the characteristics of the scanning equipment for normalization calibration. This results in a highly specific and sensitive, adaptively adjustable threshold range for the identification of calcified, non-calcified, and mixed plaques, rather than a single fixed value.

[0138] For calcified patches, the calcification integral value is obtained by counting the number of pixels in the calcified area and combining it with the corresponding weighting factor.

[0139] For non-calcified and mixed plaques, the distribution of iodine content on the iodine map is analyzed, and combined with the calcification information in the calcification map, a qualitative description of the plaque components is achieved, including lipid and fibrous components. Quantitative analysis is then performed by measuring the area and volume parameters of the plaque region.

[0140] The technical effects of the above-mentioned solution are as follows: by performing pixel-level fusion analysis of precisely segmented vascular anatomy and functional iodine and calcium information, a leap from single morphological assessment to precise qualitative and quantitative analysis of coronary plaques is achieved. This design can automatically and accurately identify the different components and spatial distribution of calcified, non-calcified, and mixed plaques. It not only provides standardized calcification scores, but also enables non-invasive assessment of plaque vulnerability through analysis of iodine content and distribution patterns.

[0141] The narrow detection unit performs the following process:

[0142] The image segmentation unit outputs the coronary artery centerline and vascular tree structure, which serve as the basic framework for stenosis detection.

[0143] Along the centerline of the blood vessel, a multiplanar reconstruction technique was used to extract a sequence of cross-sectional images of the blood vessel from multi-energy spectral image data;

[0144] On each cross-sectional image, the boundary of the blood vessel lumen is identified by an edge detection algorithm, and then the diameter of the blood vessel lumen is calculated.

[0145] Then, a continuous analysis of the lumen diameter data of multiple cross-sectional images along the blood vessel centerline is performed. By calculating the rate of change of lumen diameter between adjacent cross-sections, the location of stenosis is identified.

[0146] When the rate of change of the lumen diameter exceeds a preset threshold, the location is determined to be a stenosis. The preset threshold is dynamically set based on retrospective analysis and machine learning training of a large number of clinically diagnosed coronary artery stenosis cases. By analyzing the distribution characteristics of the rate of change of the lumen diameter before and after the known stenosis location in historical data, and combining the qualitative labeling of the degree of stenosis (e.g., mild, moderate, severe) by expert physicians, a statistical model is used to determine the optimal threshold range that can most effectively distinguish between true stenosis and normal vascular physiological fluctuations or image noise interference. Moreover, this threshold is not a fixed value, but will be adaptively adjusted according to the specific type of the detected blood vessel (e.g., left main coronary artery, left anterior descending artery, circumflex artery, etc.), the physiological curvature of the blood vessel segment, and the overall blood vessel size characteristics of the patient, so as to ensure that the sensitivity and specificity of stenosis location identification maintain the optimal balance between different individuals and different blood vessel segments.

[0147] A quantitative indicator of the degree of stenosis is calculated by comparing the lumen diameter at the stenosis point with the lumen diameter of a normal reference segment (usually a non-stenotic segment proximal or distal to the stenosis point).

[0148] Stenosis degree = [(normal lumen diameter - lumen diameter at stenosis) / normal lumen diameter] × 100%.

[0149] The technical effect of the above solution is as follows: by combining the automatically segmented vascular centerline with multi-planar reconstruction technology, a fully automatic and high-precision quantitative analysis of the coronary vascular tree is achieved, which can sensitively identify the location of focal stenosis and automatically calculate the accurate percentage of stenosis by objectively comparing it with the adjacent normal reference segment, thereby avoiding the assessment bias caused by subjective judgment based on the doctor's visual estimation.

[0150] Spatial registration was performed between the vascular region segmentation results and the iodine and calcification maps, including:

[0151] Obtain the segmentation results, iodine map, and calcification map of the vascular region;

[0152] Spatially align the vascular region segmentation results, iodine map, and calcification map;

[0153] The spatially aligned vascular region segmentation results are then subjected to vascular wall layer segmentation to obtain segmentation results of the intima, media and adventitia layers.

[0154] The spatially aligned calcification map is used to perform stratified localization and hardness grading of calcifications. Multi-scale Frangi filtering is used to enhance the features of calcifications, mark the interaction area between calcifications and the vessel wall, and generate pathological anchor points with hardness grade and the layer parameters of the vessel wall in which they are located.

[0155] An iodine concentration-blood flow velocity correlation model is established on the spatially aligned iodine map. Gradient direction adaptive filtering is used to enhance the edges of the iodine map, which is then transformed into a blood flow velocity distribution and marked in areas of low blood perfusion. Functional anchor points with mean blood flow velocity and iodine concentration parameters are generated.

[0156] Based on the aforementioned layered segmentation results, pathological anchor points, and functional anchor points, dynamic structural anchor points are generated, including intimal inflection points with calcified attachment sites and key hemodynamic parameters.

[0157] Construct cross-modal anchor point morphological association rules;

[0158] Key points with attached spatial-mechanical-functional feature vectors are extracted from the layered segmentation results, enhanced iodine map, and enhanced calcification map, respectively. The feature vectors include three-dimensional coordinates, the vascular layer, and corresponding modality-specific parameters.

[0159] Define the cross-modal keypoint distance;

[0160] Based on cross-modal key point distance combined with morphological association rule constraints, corresponding point pairs are selected by combining the nearest neighbor ratio method with RANSAC denoising.

[0161] Based on the corresponding point pairs, the local bending energy of the blood vessel is calculated, and the geometric shape of the blood vessel is corrected by the deformation compensation model.

[0162] Singular value decomposition (SVD) is used to solve the rigid transformation matrix. Based on the rigid transformation matrix, the iodine map and calcification map are first registered. The optimization objective is to minimize the mean square error between the segmentation result and the transformed image.

[0163] Construct a layered elastic displacement field model of the blood vessel wall;

[0164] Based on the layered elastic displacement field model of the blood vessel wall, the iodine map and calcification map after the first registration are re-registered to obtain the iodine map and calcification map after the second registration.

[0165] In this embodiment, the spatially aligned vascular region segmentation results are subjected to vascular wall layer segmentation, including: the spatially aligned segmentation results are segmented using multiple thresholds + Hessian matrix, the innermost boundary of the vascular layer is extracted based on the CT value threshold (30-60HU) + Hessian matrix tubular features; the media is expanded outward by 1-2mm (referring to the thickness of normal vascular wall) + elastic modulus constraint (E=2.0MPa) to avoid confusion with the adventitia adipose tissue; the adventitia is expanded outward by 0.5-1mm + CT value threshold (-100~-20HU, to distinguish adipose tissue).

[0166] In this embodiment, the spatially aligned calcification map is used for stratified localization and hardness grading of calcification foci. The interaction region between calcification foci and vessel wall is marked, and pathological anchor points with hardness grade and vessel wall layer parameters are generated. The calcification foci features are enhanced by multi-scale Frangi filtering, including: determining the vessel wall layer where the calcification foci are located (intima calcification: CT value > 600 HU, elastic modulus 10.0 MPa; media calcification: CT value 300-600 HU, elastic modulus 4.0 MPa) by combining the layered structure of the segmentation results with the spatially aligned calcification map; marking the pathological anchor point (geometric center of calcification foci) of each calcification foci with parameters: hardness grade, vessel wall layer, and shortest distance to the functional anchor point; using distance transformation + dilation operation, an interaction region of 1 mm around the calcification foci is generated, and the vessel wall thickness change rate (interaction region thickness / normal region thickness) in this region is calculated. Regions with a change rate > 0.3 are marked as high-deformation interaction regions.

[0167] In this embodiment, an iodine concentration-blood flow velocity correlation model is established on the spatially aligned iodine map, which is then transformed into a blood flow velocity distribution and low perfusion areas are marked. Functional anchor points with mean blood flow velocity and iodine concentration parameters are generated, and gradient direction adaptive filtering is used to enhance the edges of the iodine map. This includes: establishing a correlation model between iodine concentration and blood flow velocity on the spatially aligned iodine map. ;in Indicates the calibration coefficient; Indicates blood flow velocity; Indicates iodine concentration; Indicates the radius of the blood vessel; Indicates blood viscosity ; This represents the pressure gradient (estimated by the rate of change in diameter of adjacent vessel segments). ; The diameter at the narrowest point; The normal segment diameter is used. The iodine map is converted into a blood flow velocity distribution using this model. Regions with blood flow velocities <20cm / s are marked as low perfusion areas, with their centroids serving as functional anchor points. Gradient direction adaptive filtering is used: for low perfusion areas, small kernel filtering (3×3) is used along the direction of blood vessel to preserve edges, and large kernel filtering (5×5) is used in the direction perpendicular to the vessel to remove noise; for normal perfusion areas, a uniform 4×4 kernel filter is used.

[0168] In this embodiment, dynamic structural anchors are generated based on the layered segmentation results, pathological anchors, and functional anchors, including intimal inflection points at calcification attachment sites and hemodynamic key point parameters: Anchor type 1: Intimal inflection point at calcification attachment site (calcification causes intimal deformation, where the Hessian matrix eigenvalue abruptly changes), with parameters: calcification volume, degree of intimal deformation (deviation rate of intimal boundary from ideal circle); Anchor type 2: Hemodynamic key points (such as pre-stenotic dilatation segment, branch flow junction), identified by calculating the rate of change of vessel diameter in the segmentation results, with parameters: diameter change rate, shortest distance from calcification.

[0169] In this embodiment, key point extraction is performed as follows: Each key point is accompanied by a spatial-mechanical-functional feature vector: Segmentation result key points: Feature vector = (3D coordinates, blood vessel level, blood vessel diameter, local curvature); Iodine map key points: Feature vector = (3D coordinates, blood flow velocity, iodine concentration, blood vessel segment diameter); Calcification map key points: Feature vector = (3D coordinates, calcification hardness level, blood vessel level, deformation rate of the interaction area); Extraction range: Extraction is performed within the union region of 5mm around the dynamic structural anchor point, 3mm around the functional anchor point, and 2mm around the pathological anchor point.

[0170] In this embodiment, the cross-modal keypoint distance is defined. ; Indicates the distance between key points across modalities; Represents the three-dimensional Euclidean distance. ; This represents the distance between the normalized elastic moduli. , , These are the elastic moduli corresponding to the two key points. This is the normalized baseline value; This represents the distance of the normalized local curvature difference. , , The largest eigenvalue of the Hessian matrix. (This represents the maximum physiological curvature of the coronary arteries). , , This represents the weighting coefficient.

[0171] In this embodiment, the morphological association rules include: geometric continuity constraint that the rate of change of the curvature of the vessel centerline is <0.15 / mm; mechanical association constraint that when the hardness grade of calcification is > soft calcification (CT value >300HU), the intimal inflection point anchor point is preferentially associated and the layer where the calcification is located is consistent with the layer where the structural anchor point is located; and functional-structural coupling constraint that the difference in the rate of change of the vessel segment diameter between the low perfusion area and the structural anchor point is <0.3. When screening corresponding point pairs, only candidate point pairs that meet the constraints of the morphological association rules are retained.

[0172] In this embodiment, the local tortuosity energy of the blood vessel is calculated based on the corresponding point pairs, and the geometric morphology of the blood vessel is corrected through a deformation compensation model, including: calculating the local tortuosity energy of the blood vessel centerline: ,in The coordinates of the blood vessel centerline are: For arc length parameters; The vector represents the curvature of the centerline, and its magnitude represents the local bending intensity; high bending energy regions are identified: these regions are prone to artifacts caused by heartbeats; a deformation compensation field is constructed. ( The damping coefficient is... This indicates the energy at the local tortuosity of the blood vessel. (Represents the deformation compensation displacement field), compensating for displacement along the normal direction; the compensation field is applied to correct the vessel segmentation results, eliminating geometric distortion caused by pulsation artifacts.

[0173] In this embodiment, singular value decomposition (SVD) is used to solve the rigid transformation matrix, and the matrix is ​​applied to complete the first registration, including: solving the rigid transformation matrix using singular value decomposition (SVD) based on the deformity-compensated vessel segmentation results and iodine / calcification maps; the matrix includes: rotation Peaceful relocation , The mean vector of the corresponding points in the segmentation result; Indicates rotation; Indicates translation. The mean vector of corresponding points in the iodine map / calcification map is used. The mean square error between the segmentation result and the iodine map / calcification map is calculated based on the rigid transformation matrix. The first registration is completed when the mean square error is <0.05.

[0174] In this embodiment, a layered elastic displacement field model of the blood vessel wall is constructed: based on the layered structure (intima, media, adventitia) of the segmentation results, an independent elastic displacement field is designed for each layer, and the core formula is as follows: for any layer of the blood vessel wall ( Endometrium : median, Elastic displacement fields were designed for the outer membrane. ,satisfy ,in, Presentation layer The elastic displacement field; Represents the gradient operator; The divergence operator represents the stress equilibrium condition; Presentation layer elastic modulus ( , , ); Optimize weights by region, high deformation interaction area (1mm around calcification): weight Using small step size iteration (step size 0.1mm), normal vascular area: weight A large step size of 0.3 mm is adopted. During the iteration process, the deformation error of each layer (the Euclidean distance after registration of the key points of the layer) is calculated in real time. When the average error of all layers is <0.2 mm, the iteration is stopped, and the layered elastic displacement field model is obtained.

[0175] In this embodiment, the iodine map and calcification map after the first registration are re-registered based on the layered elastic displacement field model of the blood vessel wall to obtain the iodine map and calcification map after the second registration. This includes: applying the intima layer displacement field to the iodine map after the first registration, focusing on the areas of low blood perfusion. Ensure precise alignment between the low-perfusion zone and the intima structure; for the calcification map after the first registration, apply the corresponding layer displacement field according to the layer where the calcification foci are located (intima calcification application). Application of medial calcification This process ensures that the calcifications match the vascular wall layer, resulting in a second-registered iodine map and calcification map.

[0176] The working principle and beneficial effects of the above technical solution are as follows: By combining rigid registration and elastic registration, the spatial misalignment problem of vascular region segmentation results, iodine maps, and calcification maps is solved, allowing for precise correspondence between structural, pathological (calcification), and functional (blood flow) information; the construction of pathological, functional, and dynamic structural anchor points, combined with cross-modal association rules, achieves precise matching of key points in different modalities, and uncovers the intrinsic connection between structure, pathology, and function; through bending energy calculation, deformation compensation model correction, and secondary registration using a layered elastic displacement field model, the registration accuracy is further improved, adapting to the physiological morphology of blood vessels.

[0177] The quality of the generated virtual single-level images, iodine maps, and calcification maps was evaluated, including:

[0178] Virtual single-level images, iodine maps, and calcification maps are used as multimodal images;

[0179] Obtain the contrast, resolution, and noise level of each modality in a multimodal image;

[0180] The multimodal image quality co-factor is calculated based on the contrast, resolution, and noise level of each modality.

[0181] ;

[0182] in, Represents the multimodal image quality co-operation coefficient; Represents a virtual single-level image; Iodine diagram; Represents a calcification diagram; Represents modal weights; This represents the contrast of a single-modal image after normalization. This represents the resolution of the single-modal image after normalization. The normalized representation shows the noise level of a single-modal image. Indicates the interaction coefficient of the indicators; Represents the minimum constant;

[0183] Calculate the comprehensive multimodal image quality index based on the aforementioned multimodal image quality synergy coefficient;

[0184] ;

[0185] in, This represents the comprehensive quality index of multimodal images; Indicates the noise suppression correction coefficient; Represents the modal fit coefficient; This represents the average resolution of the three types of images after normalization. This represents the average noise level of the three types of images after normalization. , , These represent the maximum, minimum, and mean contrast values ​​of the three types of images after normalization. , Represents the minimum constant; Indicates the contrast uniformity coefficient; This represents a monotonic calibration function. , , These are calibration coefficients;

[0186] The multimodal image quality comprehensive index is compared with a preset comprehensive index threshold. If the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold, the multimodal image is deemed qualified. If the multimodal image quality comprehensive index is less than the preset comprehensive index threshold, the multimodal image parameters are adjusted until the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold.

[0187] The working principle and beneficial effects of the above technical solution are as follows: By constructing a multimodal image quality synergy coefficient, the contrast, resolution, and noise level of virtual single-level images, iodine images, and calcification images are comprehensively considered, and modal interaction terms are introduced, thus overcoming the limitations of single-modal quality assessment; because the three types of images are complementary in clinical practice, synergistic assessment can more realistically reflect the overall effectiveness of multimodal data; by setting a comprehensive index threshold, a closed-loop mechanism of assessment-judgment-adjustment is formed, and when the quality does not meet the standards, the image generation parameters can be adjusted in a targeted manner until the quality requirements are met.

[0188] The data preprocessing unit includes:

[0189] The data acquisition module is configured to deploy multiple distributed data acquisition nodes on the photon counting CT detector array to form a distributed sensor network architecture. It is used to synchronously receive the raw photon counting signal and to aggregate the signal of a specific detector area for each node, so as to realize the acquisition of multi-spectral image data.

[0190] The data processing module is configured to perform image enhancement, noise suppression, and energy spectrum correction on the acquired multi-energy spectrum image data, specifically:

[0191] Image enhancement: Image enhancement processing is performed on multi-energy spectral projection data, including image sharpening, image contrast adjustment, image filtering, image correction and image normalization. Among them, the image enhancement module uses a multi-scale filtering algorithm combined with adaptive histogram equalization to improve image quality.

[0192] Noise suppression: Noise frequency bands are separated and filtered out using a threshold denoising method based on wavelet transform, while retaining key signal features;

[0193] Energy spectrum correction: Using a pre-calibrated photon counting CT detector energy spectrum response function, a polynomial fitting algorithm is used to compensate for the response differences between different energy channels, ensuring the accuracy of energy spectrum data.

[0194] The technical effects of the above solution are as follows: the data acquisition module utilizes a distributed sensor network architecture to ensure the synchronous reception and accurate convergence of the original photon counting signals, thereby improving the efficiency and stability of data acquisition. The data processing module significantly improves image quality through comprehensive processing methods such as image enhancement, noise suppression, and energy spectrum correction. Through the collaborative work of the data acquisition module and the data processing module, efficient acquisition and high-quality preprocessing of multi-energy spectrum image data are achieved.

[0195] Working principle: Patient data is acquired through optimized low-dose scanning parameters and low-concentration contrast agent protocols. Deep learning algorithms are used for multi-material decomposition to generate derived images such as virtual single-level images, iodine maps, and calcification maps. The coronary artery tree is then automatically segmented, and iodine and calcification map information is fused for qualitative and quantitative analysis of plaque components. Changes in lumen diameter are automatically detected along the vessel centerline to accurately quantify the degree of stenosis. Finally, all analysis results are integrated to generate a comprehensive assessment report. Based on this design, while ensuring high diagnostic accuracy, the risk of radiation exposure and contrast agent nephropathy for patients is significantly reduced. It also overcomes the reliance on manual operation by physicians in traditional methods, significantly improving assessment efficiency and result consistency, providing a reliable solution for the early and accurate diagnosis of coronary artery disease.

[0196] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0197] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A system for assessing coronary artery stenosis using a low-dose, low-concentration contrast agent regimen in photon CT, characterized in that: include: The multi-spectral data acquisition unit is configured to use a photon-counting CT detector and scan patients who have been injected with low-concentration contrast agents based on optimized low-dose scanning parameters to acquire multi-spectral image data. The data preprocessing unit is configured to perform big data acquisition and preprocessing on the acquired multi-energy spectrum image data, including image enhancement, noise suppression and energy spectrum correction; The image generation unit is configured to perform multi-material decomposition on the preprocessed multi-energy spectrum image data to generate a variety of derived images, including virtual single-level images, iodine maps, and calcification maps. The multi-material decomposition adopts a deep learning algorithm and uses a pre-trained convolutional neural network model to perform high-precision separation of iodine, calcium, and soft tissue components. The image segmentation unit is configured to perform image recognition based on the virtual single-level image, and automatically identify and segment the coronary artery tree structure and the vessel centerline. The plaque analysis unit is configured to perform qualitative and quantitative analysis of plaque components based on the segmented vascular region and by fusing information from the iodine map and calcification map, including calcification score calculation and differentiation between non-calcified plaques and mixed plaques. The plaque analysis unit receives the vascular region segmentation results output by the image segmentation unit, and performs spatial registration of the vascular region segmentation results with the iodine map and calcification map, including: Obtain the segmentation results, iodine map, and calcification map of the vascular region; Spatially align the vascular region segmentation results, iodine map, and calcification map; The spatially aligned vascular region segmentation results are then subjected to vascular wall layer segmentation to obtain a layered segmentation result including the intima, media, and adventitia. The spatially aligned calcification map is used to perform stratified localization and hardness grading of calcifications. Multi-scale Frangi filtering is used to enhance the features of calcifications, mark the interaction area between calcifications and the vessel wall, and generate pathological anchor points with hardness grade and the layer parameters of the vessel wall in which they are located. An iodine concentration-blood flow velocity correlation model is established on the spatially aligned iodine map. Gradient direction adaptive filtering is used to enhance the edges of the iodine map, which is then transformed into a blood flow velocity distribution and marked in areas of low blood perfusion. Functional anchor points with mean blood flow velocity and iodine concentration parameters are generated. Based on the aforementioned layered segmentation results, pathological anchor points, and functional anchor points, dynamic structural anchor points are generated, including intimal inflection points with calcified attachment sites and key hemodynamic parameters. Construct cross-modal anchor point morphological association rules; The stenosis detection unit is configured to automatically detect the location of stenosis on the centerline of the segmented blood vessel and to quantitatively assess the degree of stenosis based on changes in the lumen diameter.

2. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The multi-energy spectrum data acquisition unit includes: The scanning parameter optimization module is configured to collect patients' physiological characteristics and clinical needs and match them with the historical scanning database. Using a pre-trained deep neural network model, with patients' physiological characteristics and clinical needs as input, it analyzes the complex nonlinear relationship between different parameter combinations in historical scanning data and the final image quality, and predicts the optimal combination of scanning parameters that achieves the lowest radiation dose while meeting the requirements of diagnostic image quality. This includes specific values ​​for tube voltage, tube current, and exposure time. The edge computing module is configured to be deployed on the CT equipment and relies on the integrated photon counting CT detector's dedicated smart chip to perform real-time preprocessing of the raw photon counting signal at the data acquisition source, initially completing energy spectrum analysis and noise filtering; The data calibration module is configured to be integrated into the aforementioned smart chip to perform real-time calibration on the acquired multi-energy spectrum image data, including energy correction, gain correction and dark current correction, in order to eliminate drift and system errors of the photon counting CT detector. The data quality control module is configured to perform real-time quality assessment on the acquired multi-energy spectrum image data, check the integrity of the data, automatically identify and mark abnormal points in the data based on the inspection, and perform preliminary screening and correction of the data according to preset standards.

3. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 2, characterized in that, The data quality control module executes the following process: Receive multi-energy spectrum image data, parse the data packets, extract key information, including the data dimension, number of energy channels and scanning parameters, and ensure that the data format meets the preset standards; Compare the expected size and actual size of the multi-energy spectral image data to check the integrity of the multi-energy spectral image data. Based on the integrity check results, outliers in the data are automatically identified, their locations and types are recorded in detail, and these outliers are marked in the data. The marking information of the outliers will include the outlier type and location information. Based on preset quality standards, the multi-energy spectrum image data is initially screened to remove data segments and channels that do not meet the quality requirements. The filtered data is then corrected, including signal smoothing and contrast adjustment. After the calibration is completed, the calibrated multi-energy spectral image data will be re-evaluated to ensure that the quality of the multi-energy spectral image data meets the preset standards. If the data still has quality issues, feedback will be sent to the multi-energy spectrum data acquisition unit, prompting for re-acquisition or adjustment of acquisition parameters.

4. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The image generation unit includes: The multi-material decomposition module is configured to process the base material image data using a pre-trained convolutional neural network model. The convolutional neural network model automatically extracts features through deep learning algorithms to separate iodine, calcium and soft tissue components in the image and generate corresponding material composition maps. The derived image generation module is configured to generate virtual single-level images, iodine maps, and calcification maps based on the decomposition results of the multi-material decomposition module. The quality assessment module is configured to assess the quality of the generated virtual single-level images, iodine maps, and calcification maps, checking the image contrast, resolution, and noise level. Based on the assessment results, it automatically adjusts the image processing parameters to optimize image quality.

5. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The image segmentation unit includes: The image recognition module is configured to detect the contours and branches of blood vessels based on the generated virtual single-level image, using machine learning algorithms combined with edge detection algorithms, thereby identifying the coronary vascular structure, including the starting point, branching point and termination point of the coronary vascular tree; The vessel segmentation module is configured to segment the boundaries of the coronary vessels and extract the vessel regions based on the identified coronary vessel structures using a region growing method, while simultaneously extracting the vessel centerline based on a distance transformation algorithm. The segmentation optimization module is configured to use a multi-scale optimization algorithm to optimize the segmented coronary artery tree structure, including smoothing, artifact and noise removal, and topology optimization of the vascular tree in combination with a vascular anatomy model.

6. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The patch analysis unit performs the following process: Receive the blood vessel region segmentation results output by the image segmentation unit, and spatially register the blood vessel region segmentation results with the iodine map and calcification map; The registered iodine map and calcification map are analyzed at the pixel level within the vascular region. By using a preset threshold, the distribution areas of calcified plaques, non-calcified plaques, and mixed plaques are identified. For calcified patches, the calcification integral value is obtained by counting the number of pixels in the calcified area and combining it with the corresponding weighting factor. For non-calcified and mixed plaques, the distribution of iodine content on the iodine map is analyzed, and combined with the calcification information in the calcification map, a qualitative description of the plaque components is achieved, including lipid and fibrous components. Quantitative analysis is then performed by measuring the area and volume parameters of the plaque region.

7. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The narrow detection unit performs the following process: The image segmentation unit outputs the coronary artery centerline and vascular tree structure, which serve as the basic framework for stenosis detection. Along the centerline of the blood vessel, a multiplanar reconstruction technique was used to extract a sequence of cross-sectional images of the blood vessel from multi-energy spectral image data; On each cross-sectional image, the boundary of the blood vessel lumen is identified by an edge detection algorithm, and then the diameter of the blood vessel lumen is calculated. Then, a continuous analysis of the lumen diameter data of multiple cross-sectional images along the blood vessel centerline is performed. By calculating the rate of change of lumen diameter between adjacent cross-sections, the location of stenosis is identified. When the rate of change of the lumen diameter exceeds a preset threshold, the location is determined to be a narrow point; A quantitative index of the degree of stenosis is calculated by comparing the lumen diameter at the narrowing point with the lumen diameter of the normal reference segment. Stenosis degree = [(normal lumen diameter - lumen diameter at stenosis) / normal lumen diameter] × 100%.

8. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, The data preprocessing unit includes: The data acquisition module is configured to deploy multiple distributed data acquisition nodes on the photon counting CT detector array to form a distributed sensor network architecture. It is used to synchronously receive the raw photon counting signal and to aggregate the signal of a specific detector area for each node, so as to realize the acquisition of multi-spectral image data. The data processing module is configured to perform image enhancement, noise suppression, and energy spectrum correction on the acquired multi-energy spectrum image data.

9. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 1, characterized in that, Spatial registration of the vascular region segmentation results with iodine and calcification maps also includes: Key points with attached spatial-mechanical-functional feature vectors are extracted from the layered segmentation results, enhanced iodine map, and enhanced calcification map, respectively. The feature vectors include three-dimensional coordinates, the vascular layer, and corresponding modality-specific parameters. Define the cross-modal keypoint distance; Based on cross-modal key point distance combined with morphological association rule constraints, corresponding point pairs are selected by combining the nearest neighbor ratio method with RANSAC denoising. Based on the corresponding point pairs, the local bending energy of the blood vessel is calculated, and the geometric shape of the blood vessel is corrected by the deformation compensation model. Singular value decomposition (SVD) is used to solve the rigid transformation matrix. Based on the rigid transformation matrix, the iodine map and calcification map are first registered. The optimization objective is to minimize the mean square error between the segmentation result and the transformed image. Construct a layered elastic displacement field model of the blood vessel wall; Based on the layered elastic displacement field model of the blood vessel wall, the iodine map and calcification map after the first registration are re-registered to obtain the iodine map and calcification map after the second registration.

10. The system for evaluating coronary artery stenosis using a low-dose, low-concentration contrast agent regimen for photon CT according to claim 4, characterized in that, The quality of the generated virtual single-level images, iodine maps, and calcification maps was evaluated, including: Virtual single-level images, iodine maps, and calcification maps are used as multimodal images; Obtain the contrast, resolution, and noise level of each modality in a multimodal image; The multimodal image quality co-factor is calculated based on the contrast, resolution, and noise level of each modality. ； in, Represents the multimodal image quality co-operation coefficient; Represents a virtual single-level image; Iodine diagram; Represents a calcification diagram; Represents modal weights; This represents the contrast of a single-modal image after normalization. This represents the resolution of the single-modal image after normalization. The normalized representation shows the noise level of a single-modal image. Indicates the interaction coefficient of the indicators; Represents the minimum constant; Calculate the comprehensive multimodal image quality index based on the aforementioned multimodal image quality synergy coefficient; ； in, This represents the comprehensive quality index of multimodal images; Indicates the noise suppression correction coefficient; Represents the modal fit coefficient; This represents the average resolution of the three types of images after normalization. This represents the average noise level of the three types of images after normalization. , , These represent the maximum, minimum, and mean contrast values ​​of the three types of images after normalization. , Represents the minimum constant; Indicates the contrast uniformity coefficient; This represents a monotonic calibration function. , , These are calibration coefficients; The multimodal image quality comprehensive index is compared with a preset comprehensive index threshold. If the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold, the multimodal image is deemed qualified. If the multimodal image quality comprehensive index is less than the preset comprehensive index threshold, the multimodal image parameters are adjusted until the multimodal image quality comprehensive index is greater than or equal to the preset comprehensive index threshold.

Images