System for evaluation and diagnosis of myocardial fibrosis based on photon ct energy scan

Abstract

The application discloses an evaluation and diagnosis system for myocardial fibrosis based on photon CT energy scanning and belongs to the technical field of myocardial fibrosis evaluation. The evaluation and diagnosis system for myocardial fibrosis based on photon CT energy scanning comprises a scanning data acquisition module, a full-automatic heart segmentation module, a biological feature recognition module, a fibrosis region recognition module and a multi-parameter quantitative analysis module. The application solves the problem that the prior art has weak boundary distinguishing ability for myocardium, blood vessels and valves, and is prone to cause segment division deviation. The application realizes automatic and accurate segmentation of left ventricular myocardium, ensures the accuracy of standard 17-segment model division, greatly improves the speed and efficiency of fibrosis region recognition, effectively reduces manual operation intervention, reduces human error, improves the efficiency and accuracy of diagnosis, can comprehensively quantitatively analyze each myocardial segment and output rich key indicators.

Description

Technical Field

[0001] This invention relates to the field of myocardial fibrosis assessment technology, specifically to a system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning. Background Technology

[0002] Myocardial fibrosis is a common pathological change in many cardiovascular diseases at a certain stage. Accurately assessing the degree and extent of myocardial fibrosis is of great significance for disease diagnosis, treatment planning and prognosis.

[0003] Chinese Patent Publication No. CN120473143A discloses a method for determining a myocardial fibrosis assessment model, comprising: acquiring ultrasound images of a target rabbit during echocardiography; extracting features from the ultrasound images using multiple preset feature extraction models to form multiple feature sets; filtering representative features from each feature set using multiple preset feature screening models to form multiple feature representative sets; classifying each feature representative set using multiple preset classifiers to form multiple classification results; and determining a myocardial fibrosis assessment model based on the classification results corresponding to each feature set. This method promotes robust assessment of the severity of myocardial fibrosis in a population, enabling early detection and longitudinal monitoring.

[0004] The aforementioned patents rely solely on ultrasound single-modal data in practical use, lacking energy information and exhibiting weak ability to distinguish the boundaries between myocardium, blood vessels, and valves. This can easily lead to deviations in segment division, directly affecting the accuracy of myocardial fibrosis assessment. Therefore, they do not meet existing needs. In response, we propose an assessment and diagnostic system for myocardial fibrosis based on photon CT energy scanning. Summary of the Invention

[0005] The purpose of this invention is to provide an assessment and diagnostic system for myocardial fibrosis based on photon CT energy scanning. This system achieves energy resolution, effectively improving the information content and accuracy of the data. It enables automatic and precise segmentation of the left ventricular myocardium, ensuring the accuracy of the standard 17-segment model division. It significantly improves the speed and efficiency of fibrotic region identification, effectively reduces manual intervention, lowers human error, and enhances diagnostic efficiency and accuracy. It can perform comprehensive quantitative analysis of each myocardial segment, outputting rich key indicators. It not only provides basic information such as the location, area, and volume of the fibrotic region but also reflects the degree and nature of fibrosis through energy characteristic parameters, providing clinicians with more comprehensive and detailed diagnostic evidence and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning, comprising:

[0007] The scanning data acquisition module is used to collect multi-dimensional energy information datasets of myocardial tissue and perform preprocessing to obtain multi-dimensional energy information data.

[0008] The fully automatic heart segmentation module is used to receive the preprocessed multi-dimensional dataset and use the heart segmentation model to complete the fully automatic heart segmentation to obtain the left ventricular myocardial segmentation mask image.

[0009] The biometric recognition module is used to receive the left ventricular myocardial segmentation mask image and the corresponding multi-dimensional energy information data. Combining the anatomical features and energy metabolism features of myocardial tissue, it divides the standard 17-segment model of the left ventricular myocardium to obtain 17-segment labeled myocardial image data.

[0010] The fibrosis region identification module is used to identify myocardial fibrosis regions using 17-segment labeled myocardial imaging data and multi-parameter energy image data, obtain the boundaries of the fibrosis regions, and record the location and area information of the fibrosis regions.

[0011] The multi-parameter quantitative analysis module is used to perform quantitative analysis of the boundaries of fibrous regions using a smart chip with a heterogeneous computing architecture, and outputs the results of the quantitative analysis.

[0012] Preferably, the scanning data acquisition module specifically includes:

[0013] Personalized scanning plans are generated based on the patient's age, weight, and heart size. The tube voltage, tube current, scanning slice thickness, and scanning speed of the photon CT equipment are dynamically adjusted based on the personalized scanning plan.

[0014] By directly detecting a single X-ray photon and measuring its energy, energy resolution capability is achieved, thereby collecting a dataset of multi-dimensional energy information of the myocardium generated by photon CT scanning.

[0015] The myocardial multidimensional energy information dataset was subjected to denoising, normalization and data format conversion. The denoising process adopted an adaptive wavelet threshold denoising algorithm.

[0016] Based on the data noise characteristics of different energy channels, the threshold parameters are dynamically adjusted to effectively remove electronic noise and photon statistical noise generated during the scanning process;

[0017] Image data at different energy levels are mapped to a unified data range to eliminate data amplitude deviation caused by energy differences;

[0018] The original acquired proprietary format data is converted to DICOM format, which conforms to medical imaging standards, while retaining the original energy information annotations.

[0019] Preferably, the fully automated heart segmentation module specifically includes:

[0020] Construct a heart segmentation model and input a multi-dimensional energy dataset into the heart segmentation model;

[0021] The heart segmentation model performs feature fusion on a multi-dimensional dataset, combining image features from different energy channels with photon energy distribution features to construct a multimodal feature map;

[0022] The multimodal feature map is downsampled to gradually extract high-level semantic features, then upsampled to restore the resolution of the multimodal feature map, and finally outputs a segmented mask image of the left ventricular myocardium.

[0023] Preferably, the construction of the heart segmentation model specifically includes:

[0024] Based on a U-shaped convolutional neural network architecture, an attention mechanism module and a residual connection structure are introduced to construct an initial heart segmentation model based on deep learning;

[0025] Acquire labeled myocardial CT image data, perform data augmentation processing on the acquired myocardial CT image data, and divide the data into training set and validation set after processing;

[0026] Using a multi-dimensional energy dataset as input and myocardial segmentation results as output, the initial myocardial segmentation model is trained on the training set using the gradient descent algorithm.

[0027] The initial heart segmentation model is trained until the segmentation Dice coefficient on the validation set is no less than 0.92 and the fibrotic region recognition accuracy is no less than 0.90, thus obtaining the final heart segmentation model;

[0028] Preferably, the biometric identification module specifically includes:

[0029] Anatomical features of different regions of the myocardium were extracted, including myocardial wall thickness, ventricular cavity morphology, and myocardial texture structure, while energy features of each region were also extracted.

[0030] The extracted biological feature vectors are matched with the feature templates of a pre-defined standard 17-segment myocardial model;

[0031] A dynamic weight allocation algorithm is used during the matching process to automatically adjust the weight ratio of anatomical features and energy features in the matching based on the myocardial anatomy variations of different patients.

[0032] After the division is completed, the boundaries of each segment are verified to check whether there is overlap or gap between adjacent segments. If there is overlap or gap, the boundary position is automatically adjusted.

[0033] After adjusting to ensure there is no overlap or gap between adjacent segments, output left ventricular myocardial imaging data with 17 segment labels.

[0034] Preferably, the fibrous region identification module specifically includes:

[0035] Construct a fibrous region identification model, use the fibrous region identification model to learn the energy characteristic differences of different types of fibrosis, and establish a feature matching library for fibrous regions;

[0036] Receive 17-segment labeled myocardial image data and multi-parameter energy image data, and spatially register the 17-segment labeled myocardial image data and multi-parameter energy image data;

[0037] The multi-parameter energy image data of each segment after registration is processed in parallel, and the processed data is input into the fibrosis region recognition model to extract the energy features of each segment.

[0038] Set a feature matching threshold and compare the extracted segment energy features with the fibrosis energy features in the fibrosis region feature matching library;

[0039] When the feature matching degree exceeds the preset threshold, it is determined that there is fibrosis in the region, and the boundary of the fibrosis region is automatically drawn, while the location and area information of the fibrosis region are recorded.

[0040] Preferably, the construction of the fibrous region identification model specifically includes:

[0041] An initial model was built based on the ResNet-152 architecture, and labeled multi-parameter energy image data containing myocardial fibrosis regions and energy feature data of fibrosis were obtained. The obtained data were divided into training set and dataset.

[0042] The initial model is trained using the training set, and after multiple iterations of training, the initial model is tested on the test set.

[0043] When the fiber identification sensitivity reaches 0.91 and the specificity reaches 0.93, the final fiber region identification model is obtained.

[0044] Preferably, the multi-parameter quantitative analysis module specifically includes:

[0045] The intelligent chip adopts a heterogeneous computing architecture and realizes multi-task parallel computing through the OpenCL parallel programming framework;

[0046] Based on the boundary of the fibrotic region, the number of pixels in the fibrotic region of each myocardial segment and the total number of pixels in the segment are calculated using a smart chip to obtain the proportion of fibrotic area. At the same time, the maximum diameter, minimum diameter and morphological parameters of the fibrotic region are calculated.

[0047] CT values ​​of the fibrotic region were extracted at 10 energy levels. The average energy value and standard deviation of each energy level were calculated using a smart chip to generate energy and CT value variation curves.

[0048] At each energy level, the difference in CT values ​​between the fibrotic region and the normal myocardial region in the same segment is calculated using a smart chip to obtain an energy difference distribution histogram.

[0049] Based on the attenuation coefficient data of myocardial tissue at various energies, the average attenuation coefficient of the fibrotic region is calculated using a smart chip.

[0050] The severity of fibrosis is classified by combining the percentage of fibrotic area, average attenuation coefficient, energy and CT value change curves, and energy difference distribution histogram. After the analysis is completed, a visualization chart of quantitative indicators and a quality assessment report are automatically generated.

[0051] Preferably, the initial model is trained using the training set, including:

[0052] Preprocessing and multimodal alignment are performed on the multi-parameter energy image data and fiber energy feature data in the training set to obtain the aligned multi-parameter energy image stack and segment ROI mask;

[0053] A dual-branch feature extraction network is constructed. The dual-branch feature extraction network includes a local energy feature branch and a global segmental structure branch. The local energy feature branch is based on 3DResNet-50. It adds modal attention blocks and energy feature enhancement blocks through four downsampling stages and restores spatial resolution through a 3D transposed convolution upsampling path. It then extracts features from the aligned multi-parameter energy image stack to generate a local energy feature map.

[0054] The global segmental structure branch multiplies the segment ROI mask with the local energy feature map to obtain the local features of 17 segments, which are then clipped by 3D ROI pooling. Segment embedding vectors are generated through global average pooling and segment anatomy prior encoding. The 17 segment vectors are expanded into a pseudo-mesh, and the SwinTransformer is used to capture the inter-segment correlations to output the global segmental features. The global segmental features are copied to the corresponding segment ROI region through segment-spatial broadcasting, and then upsampled by 3D transposed convolution to obtain the global structure feature map.

[0055] A cross-branch attention fusion layer is constructed to calculate the segment-pixel correlation between the local energy feature map and the global structural feature map; based on the correlation, fusion weights are generated to perform weighted fusion of the local energy features and the global structural features, and a cross-modal fusion feature map is output.

[0056] Constructing a segmental fiberization classification head: Apply segmental ROI masks to the cross-modal fused feature map and perform regional average pooling, then output the fiberization classification probability of each of the 17 segments through a fully connected layer;

[0057] Constructing a segmental energy feature regression head: Based on segmental ROI region pooling, the predicted 8-dimensional energy features of 17 segments are output through a fully connected layer;

[0058] Constructing the ROI prediction head: From the 3DResNet-50 stage 3 feature map branch, through 4 layers of 3D convolution and softmax, output the segment ROI prediction mask;

[0059] We construct segmental fiberization classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal cooperation loss, and adversarial robustness loss, respectively.

[0060] The total loss function is obtained by weighted summing of the segmental fibrosis classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal collaboration loss, and adversarial robustness loss. The AdamW optimizer is used to minimize the total loss function, and the initial fibrosis region recognition model is obtained when the training results meet the requirements.

[0061] Preferably, the severity of fibrosis is classified by combining the percentage of fibrotic area, average attenuation coefficient, energy and CT value change curves, and energy difference distribution histogram, including:

[0062] A comprehensive index of fibrosis characteristics is calculated based on the fibrosis area ratio, average attenuation coefficient, energy and CT value variation curves, and energy difference distribution histogram.

[0063] ;

[0064] in, Indicates a comprehensive index of fibrosis characteristics; Indicates the percentage of fibrous area; This represents the average attenuation coefficient of the target area; This represents the average attenuation coefficient of normal tissue. This represents the average photon energy of the target region; This represents the average photon energy of normal tissue; This represents the slope of the CT value change curve; The slope of the curve representing the change in CT values ​​of normal tissue; This indicates the peak position of the histogram of energy difference distribution; This indicates the peak position of the histogram showing the distribution of energy differences in normal tissues. , , , , This represents the weighting coefficients, which are greater than 0 and have a sum of 1.

[0065] The severity assessment value of fibrosis is calculated based on the comprehensive index of fibrosis characteristics.

[0066] ;

[0067] in, Indicates the assessment value for the severity of fibrosis; This represents the standard deviation of the histogram representing the distribution of energy differences. A comprehensive index representing the fibrotic characteristics of normal tissue; The standard deviation of the histogram representing the distribution of energy differences in normal tissues;

[0068] Based on the severity assessment value of fibrosis, the preset assessment value-severity level table is queried to determine the target severity level of fibrosis.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] This invention achieves energy resolution by directly detecting individual X-ray photons and measuring their energy. This enables the acquisition of richer and more accurate multi-dimensional energy information datasets of the myocardium, providing a high-quality data foundation for subsequent myocardial segmentation, fibrosis region identification, and quantitative analysis. It effectively improves the information content and accuracy of the data. By integrating biometric recognition and smart chip technology, a fully automated myocardial fibrosis assessment and diagnosis process is constructed, achieving automatic and precise segmentation of the left ventricular myocardium. This ensures the accuracy of the standard 17-segment model segmentation and significantly improves the speed and efficiency of fibrosis region identification. The synergistic operation of these three technologies effectively reduces manual intervention, lowers human error, and improves diagnostic efficiency and accuracy. It allows for comprehensive quantitative analysis of each myocardial segment, outputting rich key indicators. It not only provides basic information such as the location, area, and volume of the fibrosis region but also reflects the degree and nature of fibrosis through energy characteristic parameters, providing clinicians with more comprehensive and detailed diagnostic evidence. Attached Figure Description

[0071] Figure 1 This is a block diagram of the photon CT energy scanning-based assessment and diagnosis system for myocardial fibrosis of the present invention;

[0072] Figure 2 This is a flowchart of the photon CT energy scanning-based assessment and diagnosis system for myocardial fibrosis of the present invention;

[0073] Figure 3 This is a flowchart of the multi-parameter quantitative analysis module of the present invention. Detailed Implementation

[0074] 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.

[0075] To address the shortcomings of existing technologies that rely solely on single-modal ultrasound data, lack energy information, have weak ability to distinguish the boundaries between myocardium, blood vessels, and valves, and are prone to segmentation errors, thus affecting the accuracy of direct assessment of myocardial fibrosis, please refer to [link to relevant documentation]. Figures 1-3 This embodiment provides the following technical solution:

[0076] A photon CT energy scanning-based system for assessing and diagnosing myocardial fibrosis includes:

[0077] The scanning data acquisition module is used to acquire and preprocess multi-dimensional energy information datasets of the myocardium to obtain multi-dimensional energy information data. Using a photon CT scanning device, it directly detects individual X-ray photons and accurately measures the energy value of each photon. It has high-resolution energy resolution and can capture continuous energy spectrum data from low to high energy ranges. The energy measurement accuracy is controlled within ±5keV. The multi-dimensional energy information dataset of the myocardium includes myocardial tissue image data, photon energy distribution data, and tissue density-related data at different energy levels. During the acquisition process, the photon detection accuracy is calibrated in real time to ensure that the energy measurement error of each detected X-ray photon does not exceed a preset threshold. At the same time, a dynamic scanning parameter adjustment mechanism is adopted to automatically optimize the scanning current, voltage, and scanning slice thickness according to individual differences such as the patient's body fat percentage and heart position to obtain high-quality multi-dimensional datasets.

[0078] The fully automatic heart segmentation module is used to receive the preprocessed multi-dimensional dataset and use the heart segmentation model to complete the fully automatic heart segmentation to obtain the left ventricular myocardial segmentation mask image.

[0079] The biometric recognition module is used to receive the left ventricular myocardial segmentation mask image and the corresponding multi-dimensional energy information data. Combining the anatomical features and energy metabolism features of myocardial tissue, it divides the standard 17-segment model of the left ventricular myocardium to obtain 17-segment labeled myocardial image data.

[0080] The fibrosis region identification module is used to identify myocardial fibrosis regions using 17-segment labeled myocardial imaging data and multi-parameter energy image data, obtain the boundaries of the fibrosis regions, and record the location and area information of the fibrosis regions.

[0081] The multi-parameter quantitative analysis module is used to perform quantitative analysis of the boundaries of fibrous regions using a smart chip with a heterogeneous computing architecture, and outputs the results of the quantitative analysis.

[0082] The system collects multi-dimensional energy information. The input is a multi-dimensional dataset generated by photon CT scans. It performs fully automated heart segmentation, automatically and accurately segmenting the left ventricular myocardium and further dividing it into a standard 17-segment model. On the multi-parameter energy images, it automatically identifies and delineates fibrotic regions. By learning from a large amount of labeled data, the model grasps the energy characteristics of fibrosis, performs comprehensive quantitative analysis of each myocardial segment, and outputs key indicators. By directly detecting individual X-ray photons and measuring their energy, it achieves energy resolution capability.

[0083] The multi-dimensional energy information acquisition module specifically includes:

[0084] Personalized scanning plans are generated based on the patient's age, weight, and heart size. The tube voltage, tube current, scanning slice thickness, and scanning speed of the photon CT equipment are dynamically adjusted according to the personalized scanning plan to ensure that high-resolution datasets are obtained while reducing radiation dose.

[0085] By directly detecting a single X-ray photon and measuring its energy, energy resolution capability is achieved, thereby collecting a dataset of multi-dimensional energy information of the myocardium generated by photon CT scanning.

[0086] The myocardial multidimensional energy information dataset was subjected to denoising, normalization and data format conversion. The denoising process adopted an adaptive wavelet threshold denoising algorithm.

[0087] Based on the data noise characteristics of different energy channels, the threshold parameters are dynamically adjusted to effectively remove electronic noise and photon statistical noise generated during the scanning process;

[0088] Image data at different energy levels are mapped to a unified data range to eliminate data amplitude deviation caused by energy differences;

[0089] The original acquired proprietary format data was converted to DICOM format, which conforms to medical imaging standards, while retaining the original energy information annotations.

[0090] The fully automated heart segmentation module specifically includes:

[0091] Construct a heart segmentation model and input a multi-dimensional energy dataset into the heart segmentation model;

[0092] The heart segmentation model performs feature fusion on a multi-dimensional dataset, combining image features from different energy channels with photon energy distribution features to construct a multimodal feature map;

[0093] The multimodal feature map is downsampled to gradually extract high-level semantic features, then upsampled to restore the resolution of the multimodal feature map, and finally outputs a segmented mask image of the left ventricular myocardium.

[0094] Constructing a heart segmentation model specifically includes:

[0095] Based on a U-shaped convolutional neural network architecture, an attention mechanism module and a residual connection structure are introduced to construct an initial heart segmentation model based on deep learning, which enhances the ability to extract features of the myocardial boundary region.

[0096] Acquire labeled myocardial CT image data, including images of normal myocardial tissue and images of patients with different degrees of myocardial fibrosis. The labeled content includes the outline of the left ventricular myocardium and the boundary information between the myocardium and surrounding tissues. Perform data augmentation processing on the acquired myocardial CT image data, including rotation, scaling, noise addition and energy dimension data augmentation. After processing, the data is divided into training set and validation set.

[0097] Using a multi-dimensional energy dataset as input and myocardial segmentation results as output, the initial myocardial segmentation model is trained on the training set using the gradient descent algorithm.

[0098] The initial heart segmentation model was trained until it achieved a Dice coefficient of at least 0.92 and an accuracy of at least 0.90 in identifying fibrotic regions on the validation set. This resulted in the final heart segmentation model. Through iterative training, the network parameters were continuously optimized, enabling the model to automatically and accurately segment the left ventricular myocardial region on a preprocessed multi-dimensional dataset.

[0099] The biometric identification module specifically includes:

[0100] Anatomical features of different regions of the myocardium were extracted, including myocardial wall thickness, ventricular cavity morphology, and myocardial texture structure. At the same time, energy features of each region were extracted, such as the absorption coefficient of X-rays of different energies and the peak value of photon energy distribution. These features together constitute the bio-feature vector of the myocardial region.

[0101] The extracted biofeature vectors are matched with the feature templates of a pre-defined standard 17-segment myocardial model. The feature templates of the standard 17-segment myocardial model are constructed from a large amount of normal human myocardial anatomy data and energy feature data, and include the typical feature range of each segment.

[0102] A dynamic weight allocation algorithm is used during the matching process. Based on the myocardial anatomy variations of different patients, the weight ratio of anatomical features and energy features in the matching is automatically adjusted to ensure the accuracy of segment division.

[0103] After the division is completed, the boundaries of each segment are verified to check whether there is overlap or gap between adjacent segments. If there is overlap or gap, the boundary position is automatically adjusted.

[0104] After adjusting to ensure there is no overlap or gap between adjacent segments, output left ventricular myocardial imaging data with 17 segment labels.

[0105] The fibrous region identification module specifically includes:

[0106] Construct a fibrous region identification model, use the fibrous region identification model to learn the energy characteristic differences of different types of fibrosis, and establish a feature matching library for fibrous regions;

[0107] It receives 17-segment labeled myocardial image data and multi-parameter energy image data, and performs spatial registration between the 17-segment labeled myocardial image data and the multi-parameter energy image data to ensure accurate matching of the energy image data corresponding to each myocardial segment.

[0108] The multi-parameter energy image data of each segment after registration is processed in parallel, and the processed data is input into the fibrosis region recognition model to extract the energy features of each segment.

[0109] Set a feature matching threshold and compare the extracted segment energy features with the fibrosis energy features in the fibrosis region feature matching library;

[0110] When the feature matching degree exceeds the preset threshold, it is determined that there is fibrosis in the region, and the boundary of the fibrosis region is automatically drawn, while the location and area information of the fibrosis region are recorded.

[0111] Constructing a fibrous region identification model specifically includes:

[0112] An initial model was built based on the ResNet-152 architecture, and labeled multi-parameter energy image data containing myocardial fibrosis regions and energy feature data of fibrosis were obtained. The obtained data were divided into training set and dataset. The energy features of fibrosis are specifically manifested as the unique energy absorption pattern, photon energy distribution characteristics and energy difference characteristics between the fibrosis region and normal myocardial tissue under different energy X-ray irradiation.

[0113] The initial model is trained using the training set, and after multiple iterations of training, the initial model is tested on the test set.

[0114] When the fiber identification sensitivity reaches 0.91 and the specificity reaches 0.93, the final fiber region identification model is obtained.

[0115] The multi-parameter quantitative analysis module specifically includes:

[0116] The intelligent chip adopts a heterogeneous computing architecture (integrating a 4-core ARM Cortex-A78 CPU, Mali-G710 GPU and 1024 dedicated neural network computing units, with a peak computing power of 20 TOPS), and realizes multi-task parallel computing through the OpenCL parallel programming framework;

[0117] Based on the boundary of the fibrotic region, the number of pixels in the fibrotic region of each myocardial segment and the total number of pixels in the segment are calculated using a smart chip to obtain the proportion of fibrotic area. At the same time, the maximum diameter, minimum diameter and morphological parameters of the fibrotic region are calculated.

[0118] CT values ​​of the fibrotic region were extracted at 10 energy levels. The average energy value and standard deviation of each energy level were calculated using a smart chip to generate energy and CT value variation curves.

[0119] At each energy level, the difference in CT values ​​between the fibrotic region and the normal myocardial region in the same segment is calculated using a smart chip to obtain an energy difference distribution histogram.

[0120] Based on the attenuation coefficient data of myocardial tissue at various energies, the average attenuation coefficient of the fibrotic region is calculated using a smart chip.

[0121] The severity of fibrosis is classified by combining the proportion of fibrotic area, average attenuation coefficient, energy and CT value change curves, and energy difference distribution histogram. After the analysis is completed, a visualization chart of quantitative indicators and a quality assessment report are automatically generated, and the total time for quantitative analysis is controlled within 30 seconds. At the same time, an outlier detection algorithm is used to identify and correct extreme data in the calculation process, such as abnormal fluctuations in CT values ​​caused by image noise, to ensure the accuracy of key indicators. After the analysis is completed, a quality assessment report of quantitative indicators is automatically generated, including data reliability score, error source analysis, etc.

[0122] The initial model is trained using the training set, including:

[0123] Preprocessing and multimodal alignment are performed on the multi-parameter energy image data and fiber energy feature data in the training set to obtain the aligned multi-parameter energy image stack and segment ROI mask;

[0124] A dual-branch feature extraction network is constructed. The dual-branch feature extraction network includes a local energy feature branch and a global segmental structure branch. The local energy feature branch is based on 3DResNet-50. It adds modal attention blocks and energy feature enhancement blocks through four downsampling stages and restores spatial resolution through a 3D transposed convolution upsampling path. It then extracts features from the aligned multi-parameter energy image stack to generate a local energy feature map.

[0125] The global segmental structure branch multiplies the segment ROI mask with the local energy feature map to obtain the local features of 17 segments, which are then clipped by 3D ROI pooling. Segment embedding vectors are generated through global average pooling and segment anatomy prior encoding. The 17 segment vectors are expanded into a pseudo-mesh, and the SwinTransformer is used to capture the inter-segment correlations to output the global segmental features. The global segmental features are copied to the corresponding segment ROI region through segment-spatial broadcasting, and then upsampled by 3D transposed convolution to obtain the global structure feature map.

[0126] A cross-branch attention fusion layer is constructed to calculate the segment-pixel correlation between the local energy feature map and the global structural feature map; based on the correlation, fusion weights are generated to perform weighted fusion of the local energy features and the global structural features, and a cross-modal fusion feature map is output.

[0127] Constructing a segmental fiberization classification head: Apply segmental ROI masks to the cross-modal fused feature map and perform regional average pooling, then output the fiberization classification probability of each of the 17 segments through a fully connected layer;

[0128] Constructing a segmental energy feature regression head: Based on segmental ROI region pooling, the predicted 8-dimensional energy features of 17 segments are output through a fully connected layer;

[0129] Constructing the ROI prediction head: From the 3DResNet-50 stage 3 feature map branch, through 4 layers of 3D convolution and softmax, output the segment ROI prediction mask;

[0130] We construct segmental fiberization classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal cooperation loss, and adversarial robustness loss, respectively.

[0131] The total loss function is obtained by weighted summing of the segmental fibrosis classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal collaboration loss, and adversarial robustness loss. The AdamW optimizer is used to minimize the total loss function, and the initial fibrosis region recognition model is obtained when the training results meet the requirements.

[0132] In this embodiment, the multi-parameter energy image data includes raw multi-energy bins data (40-50keV, 50-60keV, 60-70keV, 70-80keV, 80-90keV) directly output by photon CT and material decomposition images (iodine map, water map, effective atomic number map). The fibrosis energy feature data includes segmental fibrosis grading labels and multi-dimensional energy feature labels (8 dimensions: iodine concentration, iodine concentration standard deviation, ECV, CT value energy spectrum slope, effective atomic number, photon counting noise level, extracellular space ratio, and fibrosis region ratio). Adaptive artifact removal is performed on the multi-parameter energy image data, including charge sharing correction, pulse accumulation correction, and K-escape correction specific to photon CT, and each energy channel is standardized to clinical standard ranges (iodine concentration: 0-5.0 mg / mL, ECV: 0.2-0.5, energy spectrum slope: -0). 0.5-0.0 HU / keV, effective atomic number: 7.0-8.5); spatial coarse registration is performed based on cardiac anatomical landmarks (left ventricular apex, mitral valve annulus center), and fine registration is performed through B-spline non-rigid registration with mutual information maximization constraint, aligning all energy channel images to the spatial coordinate system of the 70keV virtual monoenergetic image; based on the AHA17 segmental anatomical atlas, segment ROI masks are generated by automatic segmentation and manual correction using nnU-Net, and the association between the segment ROI masks and fibrosis energy feature labels is established; for samples with missing energy channels, images of missing energy channels are predicted by a lightweight 3DCNN (4-layer convolution, number of channels [16,32,16,5]) based on the inter-channel Pearson correlation coefficient matrix, and the images are completed to obtain the aligned multi-parameter energy image stack (128×128×32×5) and segment ROI mask (128×128×32×17).

[0133] In this embodiment, there are four downsampling stages: Stage 1 outputs 64×64×16×64, Stage 2 outputs 32×32×8×128, Stage 3 outputs 16×16×4×256, and Stage 4 outputs 8×8×2×512.

[0134] In this embodiment, the overall network architecture is designed as follows: Input layer: aligned multi-parameter energy image stack (128×128×32×5) + segment ROI mask (128×128×32×17); Dual-branch structure: local branch outputs local energy feature map (128×128×32×128); global branch outputs global structure feature map (128×128×32×128); fusion layer outputs cross-modal fusion feature map (128×128×32×256); Output layer contains three types of outputs: segment fiberization classification result (17×4-dimensional probability vector), segment energy feature vector (17×8-dimensional feature value), and segment ROI prediction mask (128×128×32×17).

[0135] In this embodiment, the modal attention block calculates the "material-specific discriminant score" for each energy channel. ,in, Energy Channel The material-specific discriminant score quantifies the sensitivity of this channel to fibrosis. Energy Channel The photon count value at pixel i Energy channels for normal myocardium Average photon count, where N is the total number of pixels in the current feature map; channel weights are generated based on the score: ; Indicates energy channel Dynamic weights; The energy channel index is represented; the weights of the five energy channels are dynamically adjusted by a multilayer perceptron; the energy feature enhancement block targets the fibrosis-sensitive channels (40-50keV low-energy channels, iodine map channels), and adds a 1×1×1 convolutional layer to the iodine map channels to enhance the feature response of the high iodine concentration region (iodine concentration > 150% of normal myocardium).

[0136] In this embodiment, the 3DROI pooling is clipped to 16×16×4×128.

[0137] In this embodiment, segment embedding vectors (17×256) are generated by global average pooling and segmental anatomy prior encoding (17-dimensional one-hot vectors concatenated with normalized spatial coordinates). The 17 segment vectors are expanded into a 4×4×2 pseudo-mesh (filled with one zero vector virtual segment), and the SwinTransformer (window size 2×2×2, 3 layers) is used to capture the inter-segment relationships and output global segment features. The global segment features are copied to the corresponding segment ROI region through segment-spatial broadcasting, and then upsampled to 128×128×32 by 3D transposed convolution to obtain the global structural feature map (128×128×32×128).

[0138] In this embodiment, calculating the segment-pixel correlation between the local energy feature map and the global structural feature map includes: ,in, For pixels, for To which segment, For pixels With the segment The characteristic correlation between them; Local energy feature branch at pixel The eigenvector at that location; For global segmental structure branching in segments The eigenvector at that location; The L2 norm of the local eigenvectors; is the L2 norm of the global eigenvectors.

[0139] In this embodiment, the fusion weights are generated based on correlation: ; For pixels The fusion weight; For pixels Index of the corresponding myocardial segment; This is the activation function.

[0140] In this embodiment, local energy features and global structural features are weighted and fused: ; This is the fused feature vector.

[0141] In this embodiment, the fused feature map is subjected to region average pooling using segment ROI masks, and the fiberization grading probabilities of each of the 17 segments are output through a fully connected layer (256→64→4).

[0142] In this embodiment, a segmental energy feature regression head is constructed: also based on segmental ROI region pooling, the predicted 8-dimensional energy features of 17 segments are output through a fully connected layer (256→128→8).

[0143] In this embodiment, the ROI prediction head is constructed by: starting from the 3DResNet-50 stage 3 feature map branch, passing through 4 layers of 3D convolution (number of channels [128,64,32,17]) and softmax output segment ROI prediction mask (128×128×32×17).

[0144] In this embodiment, segmental fiberization classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal cooperation loss, and adversarial robustness loss are constructed respectively, including:

[0145] Constructing segmental fiberization classification loss;

[0146] An energy feature consistency loss is constructed based on the MSE loss of energy feature numerical accuracy and the constraint loss of energy feature physiological rationality.

[0147] Segment structural constraint loss is constructed based on the IOU loss between the predicted and actual ROI of the segment and the position loss due to the deviation of the segment center coordinates.

[0148] A cross-modal collaborative loss is constructed, and the independent features of each energy channel are extracted in the shallow layer of the local energy feature branch. The mean absolute difference between the predicted channel correlation matrix and the clinical true correlation matrix is ​​calculated.

[0149] An adversarial robust loss is constructed, with the generator G taking a noisy multi-parameter image as input and outputting a pseudo-fiber energy image; the discriminator D distinguishes between real and fake samples based solely on the input image; and a feature consistency constraint is added.

[0150] In this embodiment, a segmental fiberization classification loss is constructed. ,in, Total loss for segmental fibrosis classification; Let be the grading probability of segment s. For genuine rating labels, Represents the cross-entropy loss function. Segment weights ( ).

[0151] In this embodiment, the MSE loss in the accuracy of the energy characteristic numerical value is: , This results in a loss of accuracy in energy characteristic values. Let be the k-th dimension energy eigenvalue predicted by the model for segment s; The true value of the k-th dimension energy feature of segment s; and the constraint loss on the physiological rationality of the energy feature: ,in, Loss due to constraints on the physiological rationality of energy characteristics; The midpoint of the clinically normal range for characteristic k is... The width is within the normal clinical range; Indexed by energy feature dimension; Total energy loss: ; This indicates the loss of consistency in total energy characteristics.

[0152] In this embodiment, the IOU loss between the predicted and actual ROI of the segment is as follows: , This represents the segment ROI segmentation IOU loss; The intersection-union ratio (IU) of the predicted segment sROI and the actual ROI. The segmentation region of segment s predicted by the model. The actual segmentation region of segment s, and the positional loss due to the deviation of the segment center coordinates. ; This is the loss due to the deviation in the center position of the segment; The predicted 3D coordinates of the center point of segment s; The actual 3D coordinates of the center point of segment s; Euclidean norm; Total structural loss: ; This represents the total segmental structural constraint loss.

[0153] In this embodiment, independent features of each energy channel are extracted in the shallow layer of the local energy feature branch (ResNet stage 1 output), and the mean absolute difference between the predicted channel correlation matrix and the clinical true correlation matrix (based on statistics of the complete training set) is calculated: ; For cross-modal cooperative loss; , For energy channel index pairs, < This represents all possible channel pair combinations; After extracting features for the model, energy channels and The predictive correlation coefficient between them; Energy channels in real clinical data and The correlation coefficient between them.

[0154] In this embodiment, the adversarial robust loss, ; For adversarial robust loss; For real samples Expectations; To generate pseudo samples Expectations; For the discriminator to judge real samples The discrimination probability; For the discriminator to detect spurious samples The discrimination probability; The mean square error between real and pseudo samples in the local feature space; Local energy feature branch for real samples The characteristic output; For local energy feature branches of pseudo samples The characteristic output.

[0155] In this embodiment, the total loss function is: ; This is the total loss function for model training.

[0156] In this embodiment, the segmental fibrosis classification head performs region average pooling on the fused feature map for the ROI region of each segment s, and outputs the fibrosis classification probability through a fully connected layer; the energy feature regression head extracts 8-dimensional energy features based on the same ROI region: material decomposition features (iodine concentration, ECV, effective atomic number) are calculated by fitting the photon counting energy spectrum (e.g., ECV = (retardation period iodine concentration - plain scan period iodine concentration) / blood pool iodine concentration), and energy spectrum characteristic features (CT value energy spectrum slope, photon counting noise level, fibrosis region proportion) are derived by analyzing the multi-energy channel response curve; the ROI prediction head generates a prediction mask from the ResNet-50 stage 3 feature map branch and through a convolutional layer.

[0157] The working principle and beneficial effects of the above technical solution are as follows: Preprocessing and multimodal alignment are performed on the multi-parameter energy image data and fiber energy feature data in the training set to obtain the aligned multi-parameter energy image stack and segment ROI mask. This step effectively integrates data from different modalities, allowing the model to acquire data information from multiple perspectives, enriching the data expression form, and providing a more comprehensive and accurate data foundation for subsequent feature extraction and analysis, thus helping to improve the model's ability to identify fiber regions. Based on 3DResNet-50, modal attention blocks and energy feature enhancement blocks are added through four downsampling stages. Spatial resolution is restored through a 3D transposed convolution upsampling path. This design allows for in-depth mining of local energy features in the multi-parameter energy image stack. Modal attention blocks enable the model to focus more on important information in different modalities, energy feature enhancement blocks further strengthen the expression of energy features, and upsampling to restore spatial resolution ensures the integrity of spatial information in the feature map, enabling the model to capture subtle local energy changes and more accurately grasp the local features of fibrous regions. The segment ROI mask is multiplied with the local energy feature map, and a series of operations generate global segment features. This branch can consider the inter-segment correlations from a global perspective. By capturing the inter-segment relationships through the SwinTransformer, the output global structural feature map provides the model with overall structural information. The local energy feature branch and the global segmental structure branch complement each other, enabling the model to focus on both local details and the overall structure, thus improving its comprehensive ability to identify fibrotic regions. A cross-branch attention fusion layer is constructed to calculate the segment-pixel correlation between the local energy feature map and the global structural feature map. Based on this correlation, fusion weights are generated to perform weighted fusion of the local energy features and the global structural features. Finally, a segmental fibrosis classification head, a segmental energy feature regression head, and a ROI prediction head are constructed. The model outputs the fibrillation classification probability, the predicted 8-dimensional energy features of each of the 17 segments, and the segment ROI prediction mask. This multi-task learning approach allows the model to learn and optimize on different tasks, enabling it to analyze and judge fibrillated regions from multiple perspectives, thus improving its generalization ability and accuracy. Segment fibrillation classification loss, energy feature consistency loss, segment structure constraint loss, cross-modal collaboration loss, and adversarial robustness loss are constructed and weighted to obtain the total loss function. Different loss functions constrain different tasks and objectives, allowing for model optimization from multiple aspects.

[0158] The severity of fibrosis is classified into levels based on the percentage of fibrotic area, average attenuation coefficient, energy and CT value variation curves, and energy difference distribution histogram, including:

[0159] A comprehensive index of fibrosis characteristics is calculated based on the fibrosis area ratio, average attenuation coefficient, energy and CT value variation curves, and energy difference distribution histogram.

[0160] ;

[0161] in, Indicates a comprehensive index of fibrosis characteristics; Indicates the percentage of fibrous area; This represents the average attenuation coefficient of the target area; This represents the average attenuation coefficient of normal tissue. This represents the average photon energy of the target region; This represents the average photon energy of normal tissue; This represents the slope of the CT value change curve; The slope of the curve representing the change in CT values ​​of normal tissue; This indicates the peak position of the histogram of energy difference distribution; This indicates the peak position of the histogram showing the distribution of energy differences in normal tissues. , , , , This represents the weighting coefficients, which are greater than 0 and have a sum of 1.

[0162] The severity assessment value of fibrosis is calculated based on the comprehensive index of fibrosis characteristics.

[0163] ;

[0164] in, Indicates the assessment value for the severity of fibrosis; This represents the standard deviation of the histogram representing the distribution of energy differences. A comprehensive index representing the fibrotic characteristics of normal tissue; The standard deviation of the histogram representing the distribution of energy differences in normal tissues;

[0165] Based on the severity assessment value of fibrosis, the preset assessment value-severity level table is queried to determine the target severity level of fibrosis.

[0166] The working principle and beneficial effects of the above technical solution are as follows: It integrates information from multiple dimensions, including the proportion of fibrosis area, average attenuation coefficient, energy and CT value change curves, and energy difference distribution histogram. The proportion of fibrosis area directly reflects the coverage of fibrosis in the target area; the average attenuation coefficient reflects the attenuation characteristics of tissue to radiation and is related to the density and structure of fibrotic tissue; the energy and CT value change curves reflect the characteristic changes of tissue under different scanning conditions; and the energy difference distribution histogram shows the characteristics of energy distribution from a statistical perspective. By integrating these multi-dimensional features, the characteristics of fibrotic tissue can be comprehensively and accurately described, avoiding the limitations of single-indicator assessment, thus more accurately assessing the severity of fibrosis. Based on the fibrosis severity assessment value, a preset assessment value-severity level table is queried to determine the target fibrosis severity level. This standardized level classification method makes it comparable between different patients or different test results.

[0167] Working principle: When using the photon CT energy scanning-based assessment and diagnosis system for myocardial fibrosis of this invention, according to... Figure 1 , Figure 2 and Figure 3 This includes the following steps:

[0168] S1: Use the scanning data acquisition module to collect a multi-dimensional energy information dataset of myocardial tissue and perform preprocessing to obtain multi-dimensional energy information data;

[0169] S2: The preprocessed multi-dimensional dataset is received using the fully automatic heart segmentation module, and the heart segmentation model is used to complete the fully automatic heart segmentation to obtain the left ventricular myocardial segmentation mask image.

[0170] S3: Utilize the biometric recognition module to receive the left ventricular myocardial segmentation mask image and the corresponding multi-dimensional energy information data. Combine the anatomical features and energy metabolism features of myocardial tissue to divide the standard 17-segment model of the left ventricular myocardium and obtain 17-segment labeled myocardial image data.

[0171] S4: The fibrosis region identification module uses 17-segment labeled myocardial image data and multi-parameter energy image data to identify myocardial fibrosis regions, obtain the boundaries of the fibrosis regions, and record the location and area information of the fibrosis regions.

[0172] S5: The multi-parameter quantitative analysis module uses a smart chip with a heterogeneous computing architecture to perform quantitative analysis on the boundary of the fibrous region and output the results of the quantitative analysis.

[0173] 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.

[0174] 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 and diagnosing myocardial fibrosis based on photon CT energy scanning, characterized in that, include: The scanning data acquisition module is used to collect multi-dimensional energy information datasets of myocardial tissue and perform preprocessing to obtain multi-dimensional energy information data. The fully automatic heart segmentation module receives the preprocessed multi-dimensional dataset and uses the heart segmentation model to complete the fully automatic heart segmentation, obtaining the left ventricular myocardial segmentation mask image. The biometric recognition module is used to receive the left ventricular myocardial segmentation mask image and the corresponding multi-dimensional energy information data. Combining the anatomical features and energy metabolism features of myocardial tissue, it divides the standard 17-segment model of the left ventricular myocardium to obtain 17-segment labeled myocardial image data. The fibrosis region identification module is used to identify myocardial fibrosis regions using 17-segment labeled myocardial imaging data and multi-parameter energy image data, obtain the boundaries of the fibrosis regions, and record the location and area information of the fibrosis regions. The multi-parameter quantitative analysis module is used to perform quantitative analysis of the boundary of the fibrous region using a smart chip with a heterogeneous computing architecture, and output the results of the quantitative analysis. The multi-parameter quantitative analysis module specifically includes: The intelligent chip adopts a heterogeneous computing architecture and realizes multi-task parallel computing through the OpenCL parallel programming framework; Based on the boundary of the fibrotic region, the number of pixels in the fibrotic region of each myocardial segment and the total number of pixels in the segment are calculated using a smart chip to obtain the proportion of fibrotic area. At the same time, the maximum diameter, minimum diameter and morphological parameters of the fibrotic region are calculated. CT values ​​of the fibrotic region were extracted at 10 energy levels. The average energy value and standard deviation of each energy level were calculated using a smart chip to generate energy and CT value variation curves. At each energy level, the difference in CT values ​​between the fibrotic region and the normal myocardial region in the same segment is calculated using a smart chip to obtain an energy difference distribution histogram. Based on the attenuation coefficient data of myocardial tissue at various energies, the average attenuation coefficient of the fibrotic region is calculated using a smart chip. The severity of fibrosis is classified by combining the proportion of fibrotic area, average attenuation coefficient, energy and CT value change curves, and energy difference distribution histogram. After the analysis is completed, a visualization chart of quantitative indicators and a quality assessment report are automatically generated. The severity of fibrosis is classified by combining the percentage of fibrotic area, average attenuation coefficient, energy and CT value variation curves, and energy difference distribution histograms, including: A comprehensive index of fibrosis characteristics is calculated based on the fibrosis area ratio, average attenuation coefficient, energy and CT value variation curves, and energy difference distribution histogram. ； in, Indicates a comprehensive index of fibrosis characteristics; Indicates the percentage of fibrous area; This represents the average attenuation coefficient of the target area; This represents the average attenuation coefficient of normal tissue. This represents the average photon energy of the target region; This represents the average photon energy of normal tissue; This represents the slope of the CT value change curve; The slope of the curve representing the change in CT values ​​of normal tissue; This indicates the peak position of the histogram of energy difference distribution; This indicates the peak position of the histogram showing the distribution of energy differences in normal tissues. , , , , This represents the weighting coefficients, which are greater than 0 and have a sum of 1. The severity assessment value of fibrosis is calculated based on the comprehensive index of fibrosis characteristics. ； in, Indicates the assessment value for the severity of fibrosis; This represents the standard deviation of the histogram showing the distribution of energy differences. A comprehensive index representing the fibrotic characteristics of normal tissue; The standard deviation of the histogram representing the distribution of energy differences in normal tissues; Based on the severity assessment value of fibrosis, the preset assessment value-severity level table is queried to determine the target severity level of fibrosis.

2. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 1, characterized in that, The scanning data acquisition module specifically includes: Personalized scanning plans are generated based on the patient's age, weight, and heart size. The tube voltage, tube current, scanning slice thickness, and scanning speed of the photon CT equipment are dynamically adjusted based on the personalized scanning plan. By directly detecting a single X-ray photon and measuring its energy, energy resolution capability is achieved, thereby collecting a dataset of multi-dimensional energy information of the myocardium generated by photon CT scanning. The myocardial multidimensional energy information dataset was subjected to denoising, normalization and data format conversion. The denoising process adopted an adaptive wavelet threshold denoising algorithm. Based on the data noise characteristics of different energy channels, the threshold parameters are dynamically adjusted to effectively remove electronic noise and photon statistical noise generated during the scanning process; Image data at different energy levels are mapped to a unified data range to eliminate data amplitude deviation caused by energy differences; The original acquired proprietary format data was converted to DICOM format, which conforms to medical imaging standards, while retaining the original energy information annotations.

3. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 2, characterized in that, The fully automated heart segmentation module specifically includes: Construct a heart segmentation model and input a multi-dimensional energy dataset into the heart segmentation model; The heart segmentation model performs feature fusion on a multi-dimensional dataset, combining image features from different energy channels with photon energy distribution features to construct a multimodal feature map; The multimodal feature map is downsampled to gradually extract high-level semantic features, then upsampled to restore the resolution of the multimodal feature map, and finally outputs a segmented mask image of the left ventricular myocardium.

4. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 3, characterized in that, The construction of the heart segmentation model specifically includes: Based on a U-shaped convolutional neural network architecture, an attention mechanism module and a residual connection structure are introduced to construct an initial heart segmentation model based on deep learning; Acquire labeled myocardial CT image data, perform data augmentation processing on the acquired myocardial CT image data, and divide the data into training set and validation set after processing; Using a multi-dimensional energy dataset as input and myocardial segmentation results as output, the initial myocardial segmentation model is trained on the training set using the gradient descent algorithm. The initial heart segmentation model is trained until the segmentation Dice coefficient on the validation set is no less than 0.92 and the fibrotic region recognition accuracy is no less than 0.90, thus obtaining the final heart segmentation model.

5. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 4, characterized in that, The biometric identification module specifically includes: Anatomical features of different regions of the myocardium were extracted, including myocardial wall thickness, ventricular cavity morphology, and myocardial texture structure, while energy features of each region were also extracted. The extracted biological feature vectors are matched with the feature templates of a pre-defined standard 17-segment myocardial model; A dynamic weight allocation algorithm is used during the matching process to automatically adjust the weight ratio of anatomical features and energy features in the matching based on the myocardial anatomy variations of different patients. After the division is completed, the boundaries of each segment are verified to check whether there is overlap or gap between adjacent segments. If there is overlap or gap, the boundary position is automatically adjusted. After adjusting to ensure there is no overlap or gap between adjacent segments, output left ventricular myocardial imaging data with 17 segment labels.

6. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 5, characterized in that, The fibrous region identification module specifically includes: Construct a fibrous region identification model, use the fibrous region identification model to learn the energy characteristic differences of different types of fibrosis, and establish a feature matching library for fibrous regions; Receive 17-segment labeled myocardial image data and multi-parameter energy image data, and spatially register the 17-segment labeled myocardial image data and multi-parameter energy image data; The multi-parameter energy image data of each segment after registration is processed in parallel, and the processed data is input into the fibrosis region recognition model to extract the energy features of each segment. Set a feature matching threshold and compare the extracted segment energy features with the fibrosis energy features in the fibrosis region feature matching library; When the feature matching degree exceeds the preset threshold, it is determined that there is fibrosis in the region, and the boundary of the fibrosis region is automatically drawn, while the location and area information of the fibrosis region are recorded.

7. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 6, characterized in that, The construction of the fibrous region identification model specifically includes: An initial model was built based on the ResNet-152 architecture, and labeled multi-parameter energy image data containing myocardial fibrosis regions and energy feature data of fibrosis were obtained. The obtained data were divided into training set and test set. The initial model is trained using the training set, and after multiple iterations of training, the initial model is tested on the test set. When the fiber identification sensitivity reaches 0.91 and the specificity reaches 0.93, the final fiber region identification model is obtained.

8. The system for assessing and diagnosing myocardial fibrosis based on photon CT energy scanning according to claim 6, characterized in that, The initial model is trained using the training set, including: Preprocessing and multimodal alignment are performed on the multi-parameter energy image data and fiber energy feature data in the training set to obtain the aligned multi-parameter energy image stack and segment ROI mask; A dual-branch feature extraction network is constructed. The dual-branch feature extraction network includes a local energy feature branch and a global segmental structure branch. The local energy feature branch is based on 3DResNet-50. It adds modal attention blocks and energy feature enhancement blocks through four downsampling stages and restores spatial resolution through a 3D transposed convolution upsampling path. It then extracts features from the aligned multi-parameter energy image stack to generate a local energy feature map. The global segmental structure branch multiplies the segment ROI mask with the local energy feature map to obtain the local features of 17 segments, which are then clipped by 3D ROI pooling. Segment embedding vectors are generated through global average pooling and segment anatomy prior encoding. The 17 segment vectors are expanded into a pseudo-mesh, and the SwinTransformer is used to capture the inter-segment correlations to output the global segmental features. The global segmental features are copied to the corresponding segment ROI region through segment-spatial broadcasting, and then upsampled by 3D transposed convolution to obtain the global structure feature map. A cross-branch attention fusion layer is constructed to calculate the segment-pixel correlation between the local energy feature map and the global structural feature map; based on the correlation, fusion weights are generated to perform weighted fusion of the local energy features and the global structural features, and a cross-modal fusion feature map is output. Constructing a segmental fiberization classification head: Apply segmental ROI masks to the cross-modal fused feature map and perform regional average pooling, then output the fiberization classification probability of each of the 17 segments through a fully connected layer; Constructing a segmental energy feature regression head: Based on segmental ROI region pooling, the predicted 8-dimensional energy features of 17 segments are output through a fully connected layer; Constructing the ROI prediction head: From the 3DResNet-50 stage 3 feature map branch, through 4 layers of 3D convolution and softmax, output the segment ROI prediction mask; We construct segmental fiberization classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal cooperation loss, and adversarial robustness loss, respectively. The total loss function is obtained by weighted summing of the segmental fibrosis classification loss, energy feature consistency loss, segmental structure constraint loss, cross-modal collaboration loss, and adversarial robustness loss. The AdamW optimizer is used to minimize the total loss function, and the initial fibrosis region recognition model is obtained when the training results meet the requirements.

Images