Patient brain CT image intelligent calibration system and method based on big data

Abstract

The invention relates to the technical field of medical image processing, in particular to a patient brain CT image intelligent calibration system and method based on big data. A density gradient analysis unit is subjected to self-adaptive partitioning according to brain anatomical characteristics, and through three-dimensional direction gradient detection and Laplacian differential operator dual-scale fusion, a brain CT image is obtained; calculating the density gradient value of each region; the displacement compensation weight unit depends on an association rule model trained by big data, a continuous weight field is generated through two-stage decision, and a specific compensation coefficient is loaded to a key area; the image calibration execution unit adopts a differentiation algorithm to carry out sub-pixel-level non-rigid correction on a high-density mutation region and carry out rigid correction on a homogeneous region; and a partition correction result is integrated, so that the influence of head micromotion is effectively counteracted, tissue boundary blur is eliminated, the brain CT image definition and structural accuracy are improved, and reliable support is provided for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a system and method for intelligent calibration of patient brain CT images based on big data. Background Technology

[0002] Medical image processing is an important technology, specifically applied to the calibration stage of patient brain CT images before clinical diagnosis. Its core function is to improve image clarity through precise calibration adapted to the characteristics of different brain tissue regions, thus meeting the clinical need for accurate identification of lesion areas. During brain CT scans, patients are prone to unconscious micro-movements of the head. Due to differences in the degree of density variation in different brain regions, these micro-movements cause different types of interference to the images of different regions. Consequently, a single calibration method cannot simultaneously meet the precise correction needs of each region, resulting in blurring of key lesion-related areas in the image and affecting the accuracy of clinical diagnosis. To address this issue, we provide a big data-based intelligent calibration system and method for patient brain CT images. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent calibration system and method for patient brain CT images based on big data, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a big data-based intelligent calibration system for patient brain CT images is provided, including: The density gradient analysis unit receives the input brain CT image, divides the brain CT image into multiple local regions, and calculates the density gradient value of each local region based on the pixel value change rate of the brain CT image. The density gradient value is used to quantify the degree of spatial abrupt change in brain tissue density. The displacement compensation weight dynamic adjustment unit uses a pre-learned correlation model to determine the displacement compensation weight of each local region based on the density gradient value. The correlation model is established by big data analysis of historical CT image datasets. The historical CT image dataset contains pixel displacement records caused by head micro-movements and density gradient information of the corresponding regions. The correlation model learns the positive correlation between density gradient value and displacement compensation requirement. When the density gradient value is lower than the first threshold, a basic weight is assigned. When the density gradient value is between the first threshold and the second threshold, the weight increases linearly. When the density gradient value is higher than the second threshold, the highest weight is assigned. The weight value range is predefined and stored in the weight mapping table for real-time query. The image calibration execution unit applies a weighted displacement compensation operation to each local region based on the displacement compensation weight. For the local region with the highest weight, a displacement correction algorithm is used to offset the intra-layer displacement caused by head micro-movement by adjusting the pixel position. For the local region with the basic weight, a standard rigid displacement correction is used.

[0005] The second objective of this invention is to provide a method for implementing a big data-based intelligent calibration system for patient brain CT images, including any one of the above-described methods, comprising the following steps: S1. After receiving brain CT images, high-resolution grid units are divided according to the anatomical distribution of gray matter nuclei in the basal ganglia region. The grid boundaries are dynamically densified in the gray matter-white matter boundary area. The density gradient value of each local region is calculated by fusing the three-dimensional directional gradient detection operator with the Laplacian differential operator of the downsampling layer. S2. Based on the density gradient value, the basic weight coefficients are output through the pre-trained association law model. Combined with the real-time anatomical location, a spatial index query is performed in the weight mapping table. Nucleus-specific compensation coefficients are loaded for the subthalamic nucleus region. A continuous displacement compensation weight field is generated through spatial interpolation. S3. For calcifications and vascular wall boundary regions with density gradient values ​​higher than the second threshold, activate the sub-pixel level compensation mode and superimpose the real-time gradient change rate increment. Use thin plate spline interpolation to construct a local deformation field to adjust pixel coordinates. For ventricles and white matter tract regions with density gradient values ​​lower than the first threshold, call the six-degree-of-freedom rigid body transformation model to perform overall rigidity correction. S4. Integrate the partition correction results, eliminate tissue boundary ambiguity, and output the calibrated brain CT image.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs an adaptive anatomical partitioning strategy in the density gradient analysis unit to precisely divide regions according to the anatomical characteristics of brain tissue. Combined with dual-scale gradient fusion technology, it accurately quantifies the degree of density mutation in each region. The correlation model is based on big data mining to explore the mapping relationship between density gradient and displacement compensation. Coupled with a two-level weighted decision logic, specific compensation coefficients are applied to key regions such as the subthalamic nucleus to generate a continuous and accurate displacement compensation weight field. The image calibration execution unit uses a differentiated correction algorithm to offset local micro-movements in high-density mutation regions through subpixel-level non-rigid deformation correction, and to efficiently offset overall displacement in homogeneous regions through rigid correction, balancing calibration accuracy and computational efficiency. Finally, the partitioned correction results are integrated to effectively eliminate image blurring caused by head micro-movements, improve the clarity and structural accuracy of key lesion-related regions in brain CT images, and provide reliable imaging support for accurate clinical diagnosis. Attached Figure Description

[0007] Figure 1 This is an overall block diagram of the present invention; Figure 2 This is the overall flowchart of the present invention.

[0008] The meanings of the labels in the diagram are as follows: 1. Density gradient analysis unit; 2. Displacement compensation weight dynamic adjustment unit; 3. Image calibration execution unit; Detailed Implementation

[0009] 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.

[0010] This invention provides an intelligent calibration system for patient brain CT images based on big data. Please refer to [link / reference]. Figure 1 As shown, it includes: The density gradient analysis unit 1 receives the input brain CT image, divides the brain CT image into multiple local regions, and calculates the density gradient value of each local region based on the pixel value change rate of the brain CT image. The density gradient value is used to quantify the degree of spatial abrupt change in brain tissue density. The displacement compensation weight dynamic adjustment unit 2 uses a pre-learned correlation model to determine the displacement compensation weight of each local region based on the density gradient value. The correlation model is established by big data analysis of historical CT image datasets. The historical CT image dataset contains pixel displacement records caused by head micro-movements and density gradient information of the corresponding regions. The correlation model learns the positive correlation between density gradient value and displacement compensation requirement. When the density gradient value is lower than the first threshold, the basic weight is assigned. When the density gradient value is between the first threshold and the second threshold, the weight increases linearly. When the density gradient value is higher than the second threshold, the highest weight is assigned. The weight value range is predefined and stored in the weight mapping table for real-time query. The image calibration execution unit 3 applies a weighted displacement compensation operation to each local region based on the displacement compensation weight. The displacement correction algorithm is used for the local region with the highest weight to offset the intra-layer displacement caused by the micro-movement of the head by adjusting the pixel position, while the standard rigid displacement correction is used for the local region with the basic weight.

[0011] Density gradient analysis unit 1 divides the brain CT image into multiple local regions using an adaptive anatomical partitioning strategy, specifically including: High-resolution mesh units are generated based on the anatomical distribution characteristics of gray matter nuclei in the basal ganglia. A dynamic densification mesh division method is adopted in the tissue boundary region where gray matter and white matter meet. The mesh coverage is expanded for homogeneous regions filled with cerebrospinal fluid. An overlapping buffer is set at the boundary of each mesh unit to eliminate division errors. The density gradient value is calculated in each mesh through multi-directional pixel value difference operation.

[0012] The principal gradient vector is obtained by applying a three-dimensional directional gradient detection operator to the resolution layer of the original brain CT image. At the same time, the region density change features are captured by the Laplacian differential operator in the macroscale layer generated by downsampling. Finally, the dual-scale calculation results are fused by the gradient sensitivity weighting function.

[0013] The correlation pattern model is built through a multimodal data-driven architecture, which specifically includes: The data layer integrates a surgical navigation system to record head micro-movement trajectories, brain CT image sequences, and ground truth values ​​of tissue boundaries annotated by high-resolution MRI. The analysis layer uses a gradient boosting decision tree algorithm to mine the three-dimensional mapping relationship between density gradient values ​​and displacement compensation requirements. The optimization layer introduces an adversarial sample generator to synthesize calcification foci and blood vessel wall boundary samples in the basal ganglia region.

[0014] The displacement compensation weight dynamic adjustment unit 2 adopts a two-level weight decision logic, specifically including: The first level outputs basic weight coefficients through a correlation model. The second level combines real-time anatomical location information to perform spatial index lookup in the weight mapping table, loads nucleus-specific compensation coefficients for the subthalamic nucleus region, and finally generates a continuously distributed displacement compensation weight field through a spatial interpolation algorithm.

[0015] When the density gradient value is below a first threshold, the assigned base weights undergo standardized rigid correction, specifically including: For the central ventricle and homogeneous white matter tract regions, a six-degree-of-freedom rigid body transformation model is invoked for overall displacement compensation. At the same time, the local deformation correction module is disabled to reduce the computational load. The basic weight value is fixed as the minimum compensation intensity defined by the weight mapping table.

[0016] The linearly increasing weights with density gradient values ​​between the first and second thresholds are implemented using a piecewise ramp function, specifically including: Gradient-dependent weight growth curves are constructed in the transition region between the shell and the globus pallidus. The weight growth curves adopt an exponential smooth transition in the steep section at the tissue boundary and switch to a linear growth mode in the homogeneous tissue region. The weight increment is directly proportional to the rate of change of the density gradient.

[0017] The highest weights assigned when the density gradient value is above the second threshold are generated through a dynamic boundary reinforcement mechanism, specifically including: In the basal ganglia region, calcifications and the boundary region of blood vessel walls are activated to form a subpixel-level compensation mode. Preset peak weight coefficients are extracted from the weight mapping table and superimposed with adaptive increments based on real-time gradient change rate to form an overcompensation effect for micro-displacement.

[0018] Image calibration execution unit 3 employs a non-rigid deformation correction algorithm for the local region with the highest assigned weight, specifically including: A local deformation field is constructed using thin-plate spline interpolation to adjust the spatial coordinates of the boundary pixels of calcification foci. For regions with assigned base weights, rigid correction based on rotation and translation matrices is used to preserve their original topological structure without deformation.

[0019] Further explanation is needed: after the density gradient analysis unit 1 receives the input brain CT image, in order to accurately quantify the degree of density variation in different brain tissue regions, it is necessary to use an adaptive anatomical partitioning strategy based on the specificity of the brain's anatomical structure to divide local regions, and then calculate the density gradient value specifically. The specific implementation method is as follows: Density gradient analysis unit 1 divides brain CT images into multiple local regions using an adaptive anatomical partitioning strategy. This strategy dynamically adjusts the partitioning method based on the anatomical distribution characteristics and density uniformity differences of different brain tissues. The core principle is to adapt the size and density of the grid cells to the tissue characteristics, avoiding detail loss or computational redundancy caused by uniform partitioning. Brain CT images are two-dimensional tomographic images of the brain obtained through X-ray computed tomography (CT) scans. Pixel values ​​directly correspond to the density of brain tissue, and there are significant differences in pixel values ​​between different tissues. Specifically, high-resolution grid cells are first generated based on the anatomical distribution characteristics of the gray matter nuclei in the basal ganglia. The gray matter nuclei in the basal ganglia are a cluster of key neural nuclei located deep in the brain, including the putamen, globus pallidus, and subthalamic nucleus. Their anatomical distribution has clear spatial boundaries, and the density variations in this region are complex, making it a key area of ​​focus for clinical diagnosis. High-resolution grid cells refer to small grids with a side length of 1 mm, which can accurately capture the subtle structures and density differences of the gray matter nuclei in the basal ganglia, avoiding the omission of key features due to excessively large grids. During the generation process, the gray matter nuclei in the basal ganglia are first retrieved from a standard brain anatomical atlas. The spatial coordinate range is mapped onto the input brain CT image, and then a regular grid is generated within this range with a side length of 1 mm. This ensures that each gray matter nucleus is completely covered by the grid unit, and that the grid boundary fits as closely as possible to the anatomical boundary of the nucleus. This provides a precise regional division basis for subsequent density gradient calculations. A dynamically refined grid division method is used in the tissue boundary region where gray matter and white matter meet. The tissue boundary region where gray matter and white matter meet refers to the transition region between gray matter (the area where neuronal cell bodies are concentrated and has a higher density) and white matter (the area where nerve fibers are concentrated and has a lower density). The area with drastic density changes is a high-incidence area for image blurring caused by micro-movements of the head, requiring more refined partitioning. The dynamic densification grid partitioning method refers to reducing the side length of the grid unit in this area to 0.5 mm, further densifying it compared to the high-resolution grid in the basal ganglia region. By increasing the number of grids, the accuracy of capturing density abrupt changes is improved. In specific implementation, the boundary between gray matter and white matter (the transition zone from high to low pixel value) is first identified by the pixel value distribution of the CT image. Then, the dynamic densification area is defined by extending 3 grid units to each side of the boundary line. Within this area, the density is increased by 0.A 5mm side-length grid is used to ensure detailed coverage of the density gradient transition, providing support for accurate subsequent density gradient calculations. The grid coverage is expanded for homogeneous regions filled with cerebrospinal fluid (CSF). These homogeneous regions refer to areas such as the ventricles and sulci filled with CSF, where the CSF density is uniform and significantly lower than that of gray and white matter. Density variations in these regions are minimal, eliminating the need for fine-grained partitioning. Expanding the grid coverage involves increasing the side-length of the grid cells to 2mm to reduce the number of grid cells, thus reducing computational load while maintaining reasonable partitioning. Specifically, CSF regions are selected using a pixel value threshold (the pixel values ​​of CSF in CT images are typically below a specific threshold, calibrated using standard anatomical data). Within these regions, a sparse grid is generated with 2mm side-length cells. The coverage of these grid cells is significantly larger than that of the basal ganglia and tissue boundary regions, avoiding redundant calculations while fully covering the homogeneous regions and ensuring comprehensive density gradient calculations. An overlap buffer is set at the boundary of each grid cell to eliminate partitioning errors. The overlap buffer is a 0.1mm overlap between the edges of adjacent grid cells, meaning the starting edge of the next grid cell overlaps with the edge of the previous one. The terminating edges of each grid cell partially overlap. Grid division error refers to issues that may occur during grid division, such as tissue boundaries being truncated by grid edges or omissions at the junctions of adjacent grids. Specifically, after generating grid cells for all regions, an overlap range of 0.1 mm is automatically added to the four edges of each grid. For example, if the horizontal range of a grid is 1-2 mm, the overlap will actually cover 0.9-2.1 mm. Adjacent grids will start covering from 2.0-3.1 mm. This partial overlap ensures that the brain tissue boundary, regardless of which grid edge it falls on, is completely contained by at least one grid, completely eliminating junction errors during division and guaranteeing the continuity of subsequent density gradient calculations. Within each grid, the density gradient value is calculated using multi-directional pixel value difference operations. Multi-directional pixel value difference operations calculate the pixel value difference between adjacent pixels along four directions: horizontal, vertical, 45-degree diagonal, and 135-degree diagonal, thereby obtaining the density change rate in different directions. The density gradient value is an indicator that quantifies the degree of spatial abrupt change in brain tissue density within the grid. The larger the difference, the more drastic the density change, and the higher the gradient value. The specific calculation process is as follows: For all pixels within each grid cell, the pixel value difference (subtracting the previous pixel value from the next pixel value) of adjacent pixels is calculated sequentially in four directions to obtain the difference sequence in each direction. Then, the absolute value of the difference sequence in each direction is taken and the average value is calculated to obtain the density change rate in that direction. Finally, the density change rates in the four directions are weighted and summed, with the horizontal and vertical directions each accounting for 0.3 and the two diagonal directions each accounting for 0.2, to obtain the final density gradient value of the grid cell. This value accurately reflects the spatial abrupt change characteristics of brain tissue density within the grid, providing a core basis for subsequent displacement compensation weight allocation.

[0020] After dividing the local region through adaptive anatomical partitioning and performing multi-directional pixel value difference operations within the mesh, in order to balance the precision of density gradient calculation (capturing small local density abrupt changes) and global consistency (reflecting the density distribution trend over a large area), it is necessary to further optimize the results through dual-scale detection and fusion. The specific implementation method is as follows: First, a three-dimensional gradient detection operator is applied to the original brain CT image resolution layer to obtain the principal gradient vector. The original brain CT image resolution layer refers to the original pixel resolution layer when the CT image is acquired, without any scaling or downsampling processing, which can completely preserve the fine structural information at the pixel level. The three-dimensional gradient detection operator is a mathematical operator that can simultaneously detect the rate of change of pixel values ​​in three spatial directions: x (horizontal direction, left and right of the image), y (vertical direction, up and down of the image), and z (thickness direction, front and back of the image). Compared with two-dimensional operators, it is more in line with the density distribution characteristics of the three-dimensional anatomical structure of the brain. The principal gradient vector is the vector with the largest gradient magnitude in the three directions, which includes both the rate of change of density (magnitude) in that direction and the direction of change (positive or negative sign). It can accurately locate the spatial direction and intensity of the most intense density change in a local area. The specific implementation process is as follows: For each voxel within a local region's grid cell, a 3D directional gradient detection operator is applied one by one to calculate the pixel value difference (subtracting the previous voxel value from the subsequent voxel value) between adjacent voxels in the x, y, and z directions, respectively, yielding gradient components in the three directions. The amplitude of each voxel's three gradient components is calculated (combining the intensity of change in the three directions), and the gradient component with the largest amplitude is selected as the principal gradient vector for that voxel. The average of the principal gradient vectors of all voxels within the grid cell is taken to obtain the local gradient features of that grid cell at the original resolution layer, ensuring the capture of minute local density abrupt changes. Simultaneously, at the macroscale layer generated by downsampling, the Laplacian differential operator is used to capture regional density change features. Downsampling refers to scaling down the CT image at the original resolution layer. The image resolution is reduced by merging the pixel values ​​of adjacent voxels to generate a macroscale layer. The purpose is to filter local noise and highlight the overall density distribution trend. This macroscale layer, a low-resolution layer formed after downsampling, reflects the overall density distribution pattern of brain tissue, avoiding the omission of global trends due to over-focusing on local areas. The Laplacian differential operator, a second-order differential operator, excels at detecting the second-order rate of change of pixel values, i.e., the curvature of density change. It can effectively identify transition regions where density increases and decreases (or decreases) from low to high (or vice versa), accurately capturing the density distribution characteristics of a large area. Regional density change characteristics refer to the relative density change between a local area and its surrounding large area at the macroscale, reflecting the overall density distribution pattern. The specific implementation process is as follows: The original resolution brain CT image is downsampled by 2x. A mean sampling method (averaging the pixel values ​​of 2×2×2 adjacent voxels) is used to generate a macroscale layer image. The local region division of the macroscale layer is kept consistent with the original resolution layer (grid cells are scaled up proportionally to ensure corresponding coverage). The Laplacian differential operator is applied to each grid cell to calculate the second derivative of all voxels within that cell (reflecting the curvature of density changes). The absolute values ​​of the second derivatives are taken and averaged to obtain the global gradient feature of that grid cell at the macroscale layer. This feature effectively filters the original image. Local noise in the layer is eliminated to highlight the density distribution trend over a large area. Finally, the results of the dual-scale calculation are fused using a gradient sensitivity weighting function. This function is a fusion function that dynamically allocates weights based on the reliability of the dual-scale gradient features. The core logic is: the gradient features of the original resolution layer are more reliable in terms of local fine structure, while the gradient features of the macro-scale layer are more reliable in terms of global trend. Weight allocation allows the two to complement each other's advantages. The dual-scale calculation result is the mean of the principal gradient vector of the original resolution layer (local fine gradient) and the mean of the Laplacian derivative of the macro-scale layer (global trend gradient). The specific fusion process is as follows: First, the gradient results at both scales are normalized, mapping the values ​​to the 0-1 range to eliminate the difference in magnitude caused by resolution differences. Then, the weights of each scale are calculated using a gradient sensitivity weighting function. If the variance of the principal gradient vector magnitude of the original resolution layer is small (indicating stable local gradients and low noise), a higher weight (e.g., 0.6) is assigned; if the variance is large (high local noise), the weight is reduced (e.g., 0.4). Correspondingly, if the Laplacian derivative result of the macroscale layer has high consistency with multiple surrounding grid cells (clear global trend), a higher weight (e.g., 0.4-0.6) is assigned; if the consistency is low, the weight is reduced. Finally, the normalized gradient results at both scales are multiplied by their respective weights and summed to obtain the final density gradient value of the grid cell. This fusion method retains the accurate capture of small local density abrupt changes by the original resolution layer and integrates the grasp of global density trends by the macroscale layer, effectively improving the accuracy and stability of the density gradient value and providing a more reliable basis for the accurate allocation of subsequent displacement compensation weights.

[0021] After the density gradient analysis unit 1 completes the accurate calculation of the density gradient values ​​of each local region, in order to establish a reliable correlation between the density gradient values ​​and the displacement compensation requirements, and thus provide a scientific basis for the allocation of displacement compensation weights, the correlation model needs to be constructed through a multimodal data-driven architecture, integrating multi-source data and mining deep mapping relationships. The specific implementation method is as follows: The correlation model is constructed using a multimodal data-driven architecture. This architecture integrates various types of data from different sources, using data as the core to drive model learning. It overcomes the limitations of single-data sets, comprehensively capturing the complex relationships between head micro-movements, brain tissue anatomy, and image displacement. The correlation model is the core model used to output the basic weights for displacement compensation. Its core objective is to learn the correspondence between density gradient values ​​and displacement compensation requirements, matching appropriate compensation intensities to regions with different density gradients. Specifically, this involves first building a data layer, integrating surgical navigation system recording head micro-movement trajectories, brain CT image sequences, and high-resolution MRI-annotated tissue boundary ground truth values. The data layer is the foundation of model construction, responsible for collecting and integrating... The process of processing and standardizing multi-source data provides high-quality input for subsequent analysis. The head micro-motion trajectory recorded by the surgical navigation system refers to the real-time capture of minute head displacement data during CT scans, including the direction, amplitude, and timestamp of the displacement. This accurately reflects the actual micro-motions that cause image blurring. Brain CT image sequences are multiple consecutive CT images acquired during the same patient's scan, containing pixel displacement differences caused by head micro-motions. This is the core image data for model learning. The ground truth of tissue boundaries annotated by high-resolution MRI refers to the precisely annotated anatomical boundaries of brain tissue using high-resolution MRI images (soft tissue resolution is superior to CT). This serves as a standard reference for judging displacement deviations in CT images, ensuring the accuracy of model learning. During the integration process, the three types of data were first standardized. The displacement data of the head micro-motion trajectory was converted into coordinate offsets corresponding to the pixels in the CT images. The brain CT image sequence was sorted by timestamp. The ground truth values ​​of tissue boundaries annotated by MRI were mapped to the same coordinate system of the CT images using image registration technology. Then, the three types of data were aligned by timestamp to ensure that each frame of CT image corresponds to the head micro-motion trajectory and the matching ground truth value of the tissue boundary at the same time. Finally, a three-dimensional data set of CT image-micro-motion trajectory-tissue boundary ground truth value was formed, and a historical CT image dataset of no less than 10,000 cases was constructed to provide sufficient samples for model training. Then, an analysis layer was built, and the gradient boosting decision tree algorithm was used to mine the three-dimensional relationship between density gradient values ​​and displacement compensation requirements. The mapping relationship, or analysis layer, is the core computational layer of the model, responsible for mining key correlation patterns from integrated multi-source data. The gradient boosting decision tree algorithm is a machine learning algorithm based on the ensemble of multiple decision trees. By iteratively generating weak decision trees and continuously correcting errors, it can effectively handle nonlinear data, capture complex feature correlations, and adapt to the nonlinear relationship between density gradient values ​​and displacement compensation requirements. The three-dimensional mapping relationship refers to the correspondence between density gradient values, brain tissue anatomical location, and displacement compensation requirements. Displacement compensation requirements refer to the pixel adjustment amplitude and method needed to offset micro-movements of the head, which needs to be comprehensively determined by combining the density gradient (reflecting the degree of density abrupt changes at tissue boundaries) and anatomical location (reflecting the importance of the tissue and its susceptibility to micro-movements). During the mining process...First, feature engineering is performed on the three-dimensional data set. Density gradient values ​​of each local region are extracted from the CT image, representing the calculated dual-scale fused gradient. Anatomical location labels of local regions are extracted from the ground truth of tissue boundaries. Pixel displacement amounts of corresponding regions are extracted from the head micro-motion trajectory, representing the actual displacement compensation requirements. Density gradient values ​​and anatomical location labels are used as input features to the model, and pixel displacement amounts are used as output labels. 80% of the data is used as the training set, and 20% as the validation set. Then, a gradient boosting decision tree model is initialized, setting hyperparameters such as the number of decision trees, tree depth, and learning rate. Experimental calibration shows that the number of decision trees is set to 100, the tree depth to 6, and the learning rate to 0.1. Iterative training is performed using the training set, generating... A decision tree is used to calculate the error between the model's predicted value and the actual displacement compensation requirement. The next decision tree is trained with the goal of minimizing this error, progressively optimizing the model's performance. During training, the model's prediction accuracy is monitored in real time using a validation set. Training stops when the accuracy improvement falls below a preset threshold for 10 consecutive validation rounds. Ultimately, the model learns the three-dimensional mapping relationship between density gradient value, anatomical location, and displacement compensation requirement. Based on the input density gradient value and anatomical location, it can output the corresponding predicted value of displacement compensation requirement, which serves as the basis for subsequent weight allocation. Finally, an optimization layer is built, introducing an adversarial example generator to synthesize calcification foci and blood vessel wall boundary samples in the basal ganglia region. The optimization layer is an auxiliary layer used to supplement scarce samples and improve the model's generalization ability. The adversarial sample generator is a sample generation module built on a generative adversarial network. Through the adversarial interaction between the generator and the discriminator, it can generate synthetic samples that are highly similar to real sample features, solving the model bias problem caused by insufficient scarce samples. Basal ganglion calcifications and vessel wall boundaries refer to CT image samples containing basal ganglion calcifications and vessel wall boundaries. These samples account for a small proportion of real data but have a significant impact on model accuracy and need to be supplemented through synthesis. In specific implementation, the adversarial sample generator consists of two parts: a generator and a discriminator. The generator uses a convolutional neural network structure, taking a random noise vector and feature parameters of real basal ganglion calcifications and vessel wall boundaries as input, and generating simulated sample images through convolution operations. The discriminator... Using the same convolutional neural network, the model takes samples synthesized by the generator and real samples as input. By learning the features of real samples, it determines the authenticity of the input samples. During training, the generator continuously optimizes its parameters to generate realistic samples that can fool the discriminator, while the discriminator continuously optimizes its parameters to accurately distinguish between real and fake samples. The two work together and compete against each other until the probability of the generator's output being misclassified as a real sample by the discriminator stabilizes at 45%-55%, meaning that the generated samples and real samples have no significant difference in features. The 5000 generated synthetic samples are then proportionally added to the original training set to retrain the gradient boosting decision tree model. By supplementing the scarce samples, the model learns the mapping relationships of high-density abrupt change regions more fully, avoiding prediction bias caused by insufficient samples.To further improve the accuracy and generalization ability of the model in predicting displacement compensation needs in complex anatomical structures, the correlation model, through the collaborative construction of data, analysis, and optimization layers, can comprehensively integrate multi-source data features and accurately uncover the deep correlation between density gradient values ​​and displacement compensation needs, providing reliable model support for the subsequent dynamic adjustment of displacement compensation weights.

[0022] After the correlation model completes the learning of the three-dimensional mapping relationship between density gradient values ​​and displacement compensation requirements, in order to ensure that the displacement compensation weights not only conform to the characteristics of density abrupt changes but also adapt to the functional differences in brain anatomical structures, especially the high-precision calibration requirements of key nuclei, the displacement compensation weight dynamic adjustment unit 2 adopts a two-level weight decision logic. Through the synergistic optimization of basic weights and specific compensation, it generates a weight distribution that accurately adapts to the whole brain region. The specific implementation method is as follows: The displacement compensation weight dynamic adjustment unit 2 adopts a two-level weight decision logic. This two-level decision logic refers to a decision mechanism that completes weight calculation in two steps. The first level obtains a universal basic weight based on the density gradient value. The second level supplements specific compensation weights based on the anatomical location. This ensures both the correlation between weights and density abrupt changes and takes into account the special calibration needs of key areas. The displacement compensation weight quantifies the intensity coefficient of displacement correction required for each local area. The higher the weight, the greater the correction strength, and the better it can counteract image blurring caused by head micro-movements. Specifically, the first-level decision is executed, outputting the basic weight coefficient through a correlation model. This correlation model is a pre-trained model built on a multimodal data-driven architecture, which has learned the positive correlation between density gradient values ​​and displacement compensation requirements. The basic weight coefficient is the initial compensation weight obtained solely based on the density gradient value, reflecting the basic correction intensity required for density abrupt changes in the area. Its value range is 0.1-0.9, where 0.1 is the minimum compensation intensity and 0.9 is the maximum basic intensity. The specific implementation process is as follows: The displacement compensation weight dynamic adjustment unit 2 receives the density gradient value of each grid cell output by the density gradient analysis unit 1, and inputs it along with the real-time anatomical location label into the correlation model. Based on the pre-learned 3D mapping relationship, the model outputs a basic weight coefficient of 0.1 for regions with density gradient values ​​below the first threshold (calibrated to 0.3), corresponding to basic compensation requirements. For regions with density gradient values ​​between the first and second thresholds (calibrated to 0.7), it outputs a linearly increasing basic weight coefficient of 0.2-0.7. For regions with density gradient values ​​above the second threshold, it outputs a basic weight coefficient of 0.8-0.9, corresponding to high-intensity compensation requirements, ensuring that the basic weights accurately match the displacement compensation requirements caused by density abrupt changes. Then, the second-level decision is executed, combining real-time anatomical location information to perform a spatial index lookup in the weight mapping table. For the subthalamic nucleus region, a nucleus-specific compensation coefficient is loaded. Real-time anatomical location information refers to the coordinate mapping of each grid cell... The brain anatomical region identifiers obtained from the imaging are generated by registering the system's built-in standard brain anatomical atlas with the input CT image, enabling precise location of the anatomical structure to which the grid unit belongs. The weight mapping table is a coefficient lookup table pre-stored in the system, containing specific compensation coefficients corresponding to different anatomical regions. The coefficients for key nuclei regions are optimized with clinical data, while the coefficients for ordinary regions are set to 0. Spatial index query is a query method that quickly matches the corresponding coefficient in the weight mapping table based on the anatomical location identifier of the grid unit, ensuring real-time response. The subthalamic nucleus region is a key neural nucleus in the basal ganglia, involved in motor control and cognitive function. Its anatomical structure is delicate and sensitive to displacement, and the calibration accuracy directly affects the clinical diagnosis results, requiring additional specific compensation. The nucleus-specific compensation coefficient is an additional compensation coefficient specifically set for the subthalamic nucleus region, with a value of 0.2-0.3, used to strengthen the calibration intensity of this region and compensate for possible insufficient accuracy of the basic weights. The specific implementation process is as follows: The displacement compensation weight dynamic adjustment unit 2 analyzes the real-time anatomical location information of each grid cell. If the grid cell is identified as belonging to the subthalamic nucleus region, it is confirmed by coordinate range matching. Then, the weight mapping table is queried through spatial index, and a nucleus-specific compensation coefficient of 0.2-0.3 is loaded. If it belongs to other general anatomical regions, a specific compensation coefficient of 0 is loaded to ensure that only key nuclei receive additional compensation and avoid overcalibration. Finally, a continuously distributed displacement compensation weight field is generated through spatial interpolation algorithm. Spatial interpolation algorithm is a mathematical algorithm used to fill the weight gaps between discrete grid cells. Here, bilinear interpolation is used, which can calculate the weight at the grid gap through the weight values ​​of adjacent grids, ensuring that the weight distribution is continuous and smooth. The displacement compensation weight field is a continuous weight distribution covering the entire brain CT image. Each pixel has a unique corresponding displacement compensation weight, rather than only grid cells having weights, avoiding weight abrupt changes caused by grid division and ensuring the continuity of calibration operations. The specific implementation process is as follows: First, the basic weight coefficients of the first-level output are added to the nucleus-specific compensation coefficients loaded in the second level to obtain the final discrete weight value of each grid unit. For example, for the grid unit in the subthalamic nucleus region, the basic weight coefficient is 0.9 + the specific compensation coefficient is 0.3 = 1.2. When it exceeds the basic range, 1.0 is taken as the upper limit. Then, the bilinear interpolation algorithm is applied to the weight values ​​of all discrete grid units. Based on the weight of each grid unit, the weight values ​​of all pixels at the grid boundary and inside are calculated to make the weight change of adjacent pixels smooth and without obvious discontinuities. Finally, a displacement compensation weight field covering the entire brain CT image is generated. The weight value of each pixel is between 0.1 and 1.0. This not only preserves the basic compensation intensity corresponding to the density gradient, but also enhances the calibration accuracy of key areas such as the subthalamic nucleus. This provides a continuous and accurate weight basis for the partition weighted compensation of the subsequent image calibration execution unit 3.

[0023] After the displacement compensation weight dynamic adjustment unit 2 generates a continuous displacement compensation weight field through the two-level weight decision logic, for regions where the density gradient value is lower than the first threshold, since the brain tissue density distribution is uniform and the degree of spatial mutation is low, the micro-movement of the head is mostly an overall displacement rather than a local deformation. Therefore, there is no need for complex local correction. Only standardized rigid correction needs to be performed to balance the calibration effect and computational efficiency. The specific implementation method is as follows: When the density gradient value is below the first threshold, the assigned base weights perform standardized rigid correction. The first threshold is a calibrated critical value for the density gradient; below this value, it means that the rate of change of pixel values ​​within the region is small, the brain tissue density is uniform, and the effect of head micro-movements on the image is mainly overall translation or rotation. The base weight is the minimum compensation intensity defined in the weight mapping table, fixed at 0.1. This value corresponds to the lowest correction strength, adapting to the slight displacement compensation needs of uniform regions. Standardized rigid correction refers to a correction method that does not change the internal topology of the image, but only offsets displacement through rigid transformations such as overall translation and rotation. The core is to keep the relative positions between tissues unchanged and avoid image distortion caused by over-correction. Specifically, this includes first clarifying... The calibration targets the central ventricle and the homogeneous white matter tract region. The central ventricle is the core area of ​​the brain filled with cerebrospinal fluid, which has a uniform density and is significantly lower than the surrounding brain tissue, with an extremely low density gradient. The homogeneous white matter tract region is composed of nerve fiber bundles, with a single tissue composition, uniform density distribution, and no obvious density abrupt change boundaries. The common feature of these two regions is that they do not require local deformation correction; overall rigid displacement is sufficient to meet the calibration requirements. The system accurately delineates the boundaries of the central ventricle and the homogeneous white matter tract region from brain CT images through real-time anatomical location recognition and based on the coordinate range of standard brain anatomical atlases, ensuring that the calibration operation only applies to the target area and does not affect other density abrupt change regions.Next, a six-degree-of-freedom rigid body transformation model is invoked for overall displacement compensation. This model enables translation (positional movement in the x, y, and z directions) and rotation (rotation around the x, y, and z axes) in three-dimensional space. The six degrees of freedom completely cover all rigid displacement forms that may result from head micro-movements, accurately offsetting image deviations caused by overall displacement. Overall displacement compensation refers to applying a uniform rigid transformation to the defined target area, adjusting the position of all pixels by the same translation and rotation angles, while maintaining the relative positions of pixels within the area. The specific implementation process is as follows: First, the displacement parameters of the target area (including translation in the x, y, and z directions and rotation angles around the three axes) are extracted from the output of the correlation model. These parameters are derived from the mapping relationship between historical micro-movement trajectories and density gradient values, accurately reflecting the transformation amount required to offset head micro-movements. The displacement parameters are then input into the six-degree-of-freedom rigid body transformation model, which generates the corresponding transformation matrix based on the parameters. Matrix operations are then performed... The model performs coordinate transformation on each pixel within the target area. For example, if a pixel needs to be translated 2 pixels along the x-axis and rotated 0.5 degrees around the z-axis, the model calculates its new coordinates after the transformation. After the transformation, the image position of the target area is aligned with the ideal position under the state of no micro-motion, thus offsetting the overall offset caused by head micro-motion. At the same time, the local deformation correction module is disabled to reduce the computational load. The local deformation correction module is used to handle local small deformations in areas of density abrupt change, and its computational complexity is much higher than that of the rigid correction module. Since there is no need for local deformation in the central ventricle and homogeneous white matter tract regions, enabling this module is not only meaningless but also increases the system's computational load and prolongs the calibration time. In specific operation, the system automatically sends a disable command to the local deformation correction module while starting the six-degree-of-freedom rigid body transformation model, shutting down its core functions such as convolution operation and deformation field construction, and only retaining the computational resources required for rigid correction. This reduces the system's computational load by more than 40%, ensuring the real-time nature of the calibration process and meeting the needs of rapid clinical diagnosis. Throughout the standardized rigid correction process, the base weight value is always fixed at the minimum compensation intensity (0.1) defined by the weight mapping table. This weight value determines the strength of the transformation matrix, ensuring that the correction intensity is precisely matched with the slight displacement requirements of the uniform region. This prevents image blurring due to insufficient correction and pixel coordinate shift due to overcorrection. Through this series of operations, the overall displacement of the central ventricle and homogeneous white matter tract regions is effectively offset, and the image is restored to a clear state without micro-motion interference. At the same time, the internal topological structure of the tissue remains unchanged, laying the foundation for subsequent integration of zonal correction results and output of a complete calibrated image.

[0024] After allocating the basic weights for density gradient values ​​below the first threshold, for regions where density gradient values ​​fall between the first threshold (calibrated to 0.3) and the second threshold (calibrated to 0.7), the degree of density abrupt change is moderate, and the density change patterns of different sub-regions (such as tissue transition zones and homogeneous zones) differ. A single linear growth model is insufficient to accurately match the displacement compensation needs of each sub-region. Therefore, the linear increase of weights is implemented using a piecewise ramp function, ensuring both an overall increasing trend and adapting to local tissue characteristics. The specific implementation method is as follows: The linear increase in weights for density gradient values ​​between the first and second thresholds is achieved using a piecewise ramp function. This piecewise ramp function calculates weights by using different slopes for different segments based on the actual patterns of density changes within the region. This adapts to the gradual increase in density gradient within the intermediate range, avoiding mismatches between weights and compensation requirements caused by a single growth mode. The linear increase means that the weights steadily increase with the density gradient value, ensuring a positive correlation between weights and the degree of density abrupt change, consistent with the mapping logic learned by the correlation model. Specifically, this involves first focusing on the transition region between the putamen and globus pallidus. This transition region is an important component of the gray matter nuclei in the basal ganglia and belongs to the core region of the intermediate density gradient range. This region is partly a transition zone between gray and white matter, and partly a relatively homogeneous tissue. Its complex density distribution necessitates the segmented construction of weight growth curves to accurately match the displacement compensation requirements of different sub-regions. Gradient-dependent weight growth curves are constructed within this transition zone. These curves are determined by the rate of change of the density gradient in the region; regions with a high rate of change have steeper curves and faster weight growth, while regions with a low rate of change have gentler curves and relatively gentler weight growth. This ensures a precise match between the weights and the characteristics of local density abrupt changes. During the construction process, the transition zone between the shell / core and the pale sphere is first divided into sub-regions based on the pixel value change rate. The transition is then calculated... The pixel value change rate of each grid cell within the region is used to define grid cells with a change rate higher than 0.5 as the steep tissue boundary segment, and grid cells with a change rate lower than 0.5 as the homogeneous tissue region. The steep tissue boundary segment refers to the region where the shell-core meets the globus pallidus or the surrounding white matter. In this region, the density changes rapidly from low to high (or vice versa), and the density gradient value fluctuates significantly, requiring a more flexible weighting pattern to avoid abrupt changes. The homogeneous tissue region refers to the relatively simple structure and uniform density distribution within the transition zone, where the density gradient value changes smoothly, making a simple linear growth pattern suitable. An exponential smooth transition is used for the steep tissue boundary segment. This exponential smooth transition involves introducing a smoothing coefficient to make the weights increase slowly in this segment. To avoid a sharp increase in weight due to rapid changes in density gradient, a smooth transition in compensation intensity is achieved, preventing over-correction or under-correction. The smoothing coefficient was experimentally calibrated to 0.2. This coefficient adjusts the growth rate of the exponential function, allowing the weight to rise slowly from the starting point of the corresponding density gradient value. Even if the density gradient value increases rapidly, the weight will not jump, ensuring that the displacement compensation intensity is precisely matched with the slight deformation requirements of the tissue boundary. In specific implementation, the basic weight corresponding to the density gradient value at the beginning of the segment is used as the starting point. The weight increment corresponding to each density gradient value is calculated through the exponential function. The increment gradually increases with the increase of the density gradient value, but the growth rate is limited by the smoothing coefficient, ultimately achieving a slow and gradual transition effect without abrupt changes.In homogeneous tissue regions, the model is converted to a linear growth mode. The linear growth mode refers to the relationship between the weight and the density gradient value, which is a fixed growth relationship. That is, a fixed slope is set, and the weight increment increases uniformly with the increase of the density gradient value. This mode is suitable for the characteristics of the stable change of the density gradient in homogeneous tissue regions. It is simple to calculate and can ensure the linear matching of compensation requirements. The slope of this mode is calibrated to 1.2, that is, for every 0.1 increase in the density gradient value, the weight increases by 0.12 accordingly, ensuring that the overall linear increase requirement is met. In practice, the weight value at the starting position of the homogeneous tissue region is used as the benchmark. The weight increment is obtained by multiplying the difference between the current density gradient value and the initial density gradient value by a fixed slope. This increment is then added to the benchmark weight to obtain the final weight corresponding to that density gradient value. For example, if the density gradient value at a point in the homogeneous tissue region is 0.5, the initial benchmark weight is 0.4, the weight increment is (0.5-0.3)×1.2=0.24, and the final weight is 0.4+0.24=0.64. This maintains linear growth and adapts to the stable density changes in the region. Throughout the segmented construction process, the weight increment is directly proportional to the density gradient change rate. The weight increment refers to the increase in weight for each density gradient value compared to the previous gradient value, while the density gradient change rate refers to the increase in weight between adjacent density gradients. The ratio of the difference in values ​​to the corresponding spatial distance reflects the rate of increase in density gradient. A positive proportional relationship between the two means that the higher the rate of change of density gradient, the larger the weight increment, and the lower the rate of change of density gradient, the smaller the weight increment. This ensures that the weight increase always precisely matches the actual speed of density mutation. Through the above design of the piecewise ramp function, in the region where the density gradient value is between the first and second thresholds, the weight achieves an overall linear increase from 0.2 to 0.7. Furthermore, through the exponential smooth transition of the steep rise segment of the tissue boundary and the linear growth of the homogeneous tissue region, it adapts to the density change patterns of different sub-regions. This makes the weight allocation conform to the positive correlation of the correlation law model and accurately match the displacement compensation needs of each sub-region, providing more targeted weight support for subsequent image calibration.

[0025] After assigning weights to regions with density gradient values ​​between the first and second thresholds, for regions with density gradient values ​​higher than the second threshold (calibrated to 0.7 by multimodal data), due to the extremely drastic changes in brain tissue density, and the presence of areas with delicate structures crucial for diagnosis such as basal ganglia calcifications and vascular wall boundaries, even minor head movements can lead to severe image blurring. Ordinary weights are insufficient for accurate calibration. Therefore, the highest weight is generated through a dynamic boundary enhancement mechanism to achieve targeted high-intensity compensation. The specific implementation method is as follows: The highest weight assigned when the density gradient value is higher than the second threshold is generated through a dynamic boundary reinforcement mechanism. This dynamic boundary reinforcement mechanism is a weight enhancement strategy designed for high-density abrupt change regions. Its core is to combine regional anatomical characteristics with real-time density changes, and generate a compensation intensity higher than the conventional weight upper limit through a basic peak weight and adaptive increment, ensuring that the displacement of fine structural regions is completely offset. The highest weight is the maximum compensation weight value supported by the system, with a basic range of 1.0-1.2, which can provide the strongest displacement compensation force for high-density abrupt change regions and adapt to their extreme requirements for calibration accuracy. Specifically, this involves first clearly defining the target area, then activating a subpixel-level compensation mode in the basal ganglia calcifications and vascular wall boundaries. Basal ganglia calcifications are highly dense calcified tissues with clear boundaries and significant density abrupt changes, serving as crucial markers for lesion diagnosis in clinical practice. Vascular wall boundaries are the junction between the inner wall of brain blood vessels and surrounding tissues; their structure is intricate, with significant density differences, meaning even minute displacements can blur the boundaries, affecting the diagnosis of vascular lesions. The subpixel-level compensation mode overcomes the limitations of traditional pixel-level adjustments by subdividing pixels into smaller units (e.g., 0.1 pixels) for displacement correction. This mode can precisely compensate for minute displacements smaller than 1 pixel, addressing the problem of insufficient pixel-level compensation accuracy in high-density abrupt change areas. During activation... The system first uses real-time anatomical location recognition and density gradient value dual determination to accurately delineate the range of calcifications and vascular wall boundaries in the basal ganglia region. It then matches areas with density gradient values ​​higher than 0.7 using standard anatomical atlas coordinates. Subsequently, it automatically activates the sub-pixel level compensation module and closes the conventional pixel level correction channel to ensure that the compensation operation focuses on the minute displacements of fine structures. Next, it extracts the preset peak weight coefficients from the weight mapping table, which is a pre-stored coefficient library of the system. The preset peak weight coefficient for high-density mutation regions is fixed at 1.0. This value is the highest basic weight calibrated based on a large amount of historical case data, which can meet the basic compensation needs of most high-density mutation regions, while reserving space for subsequent incremental overlay. During the extraction process, the system quickly locates the corresponding weight mapping entry for the delineated region using spatial indexing, directly retrieves the peak weight coefficient of 1.0 as the base value for the highest weight, ensuring the speed and consistency of weight allocation with the baseline. An adaptive increment based on the real-time gradient change rate is then superimposed. The real-time gradient change rate refers to the rate at which the density gradient value of the current region changes with spatial location. It is obtained by calculating the difference in density gradient values ​​between adjacent grid cells, reflecting the steepness of density abrupt changes. A higher change rate indicates a sharper region boundary and higher sensitivity to displacement. The adaptive increment is an additional weight value dynamically adjusted according to the real-time gradient change rate, ranging from 0.1 to 0.2. Its core logic is that the steeper the density abrupt change, the larger the increment, ensuring precise matching between the weight and the actual displacement compensation requirements of the region. The specific calculation process is as follows: For each grid cell within the defined region, the density gradient difference between it and its four adjacent grid cells is calculated. The average of these differences is taken as the real-time gradient change rate of that grid cell. If the change rate is higher than 0.2 (the steepness threshold calibrated experimentally), an adaptive increment of 0.2 is added. If the change rate is between 0.1 and 0.2, an adaptive increment of 0.1 is added. If the change rate is lower than 0.1, no increment is added (only the peak weight coefficient of 1.0 is retained). The maximum weight value after addition is set to 1.2 to avoid excessive image stretching and distortion caused by excessive weight. This results in an overcompensation effect for micro-motion displacement. The overcompensation effect refers to a calibration effect where the compensation intensity is slightly higher than the actual displacement requirement. Through sub-pixel level correction driven by the highest weight (1.0-1.2), not only can the intra-layer displacement caused by head micro-motion be completely offset, but the density contrast at the region boundary can also be slightly enhanced. This makes calcifications and vessel wall boundaries clearer and sharper in the image. This overcompensation is not indiscriminate enhancement, but rather a precise adaptation based on the real-time gradient change rate. It solves the core problem of easy blurring due to displacement in high-density mutation areas, and avoids the risk of overcorrection through weight upper limit control. It ensures that the calibrated area has no displacement deviation and retains the authenticity of the original anatomical structure. The implementation of the entire dynamic boundary enhancement mechanism, from precise region identification to sub-pixel level mode activation, and then to the generation of peak weight and adaptive incremental weight, forms a closed-loop compensation logic for high-density mutation areas. This ensures that the highest weight has sufficient compensation strength and can be dynamically adjusted according to the actual characteristics of the region. It provides accurate and powerful weight support for the non-rigid deformation correction of the subsequent image calibration execution unit 3, ensuring the image clarity and structural accuracy of key diagnostic areas such as calcifications and vessel wall boundaries in the basal ganglia.

[0026] After the displacement compensation weight dynamic adjustment unit 2 generates a continuously distributed displacement compensation weight field, the image calibration execution unit 3 needs to adapt the differentiated correction algorithm according to the weight level of different regions. The highest weight region needs to accurately offset local minor deformations, and the basic weight region needs to efficiently offset the overall displacement, so as to balance calibration accuracy and computational efficiency. The specific implementation method is as follows: Image calibration execution unit 3 employs a non-rigid deformation correction algorithm for the local regions assigned the highest weight. These regions are those with density gradient values ​​exceeding the second threshold and weights (1.0-1.2) generated through a dynamic boundary enhancement mechanism. These are primarily calcifications in the basal ganglia and vessel wall boundaries. These regions have intricate structures and dramatic density abrupt changes; even slight head movements can easily trigger minor local deformations, not overall displacement. Therefore, non-rigid correction is necessary. The non-rigid deformation correction algorithm is a correction method that does not rely on a fixed topology. It can flexibly adjust pixel positions according to local deformation requirements, allowing different regions to have different displacements, accurately offsetting irregular minor deformations. This differs from rigid correction, which maintains the overall structure unchanged. Specifically, this involves first constructing a local deformation field using thin-plate spline interpolation. Thin-plate spline interpolation is an interpolation method based on fitting smooth deformation curves using control points. It can calculate the continuous deformation trend of the entire region using displacement information from a small number of key control points, combining flexibility and smoothness to adapt to the irregular deformation correction needs of calcifications and vessel wall boundaries. The local deformation field is a continuous field describing the displacement direction and amplitude of each pixel within the highest weighted region. Each pixel has its own displacement parameters to ensure accurate cancellation of local deformation. The pixels at the calcification boundary are the pixels where the calcification in the basal ganglia region intersects with the surrounding tissue. The positional accuracy of these points directly affects lesion diagnosis and requires careful adjustment. The construction process is as follows: First, control points are automatically extracted from the highest weighted region. Feature pixels at the boundaries of calcifications and the edges of blood vessel walls are selected and determined through pixel value mutation detection. Each feature point corresponds to a target position, determined by the displacement compensation requirement output by the correlation model, i.e., the ideal pixel coordinates when there is no micro-motion. The original coordinates of these control points and the target coordinates are input into the thin-plate spline interpolation model. The model calculates the displacement vectors of all pixels in the region by fitting a smooth interpolation function, including the displacement in the x, y, and z directions. The displacement vectors of all pixels together constitute the local deformation field, ensuring the continuity of the deformation field. Without breaks, image distortion is avoided after correction. Then, the spatial coordinates of pixels at the calcification boundary are adjusted using a local deformation field. Specifically, the image calibration execution unit 3 calls the displacement vector of each pixel in the local deformation field to adjust the original coordinates of the calcification boundary and surrounding pixels point by point. If the displacement vector of a pixel is +0.3 pixels in the x-direction and -0.2 pixels in the y-direction, then the original x-coordinate of that pixel is increased by 0.3 and the y-coordinate is decreased by 0.2 to obtain the corrected new coordinates. During the adjustment process, the pixel values ​​remain unchanged; only the spatial position is changed to ensure the accuracy of the density information at the calcification boundary. For the boundary pixels of the blood vessel walls, the same logic is used to adjust them point by point, restoring the boundaries from a blurred state to a clear and sharp state, completely offsetting the image deviation caused by local minor deformations. For the regions with assigned base weights, rigid correction based on rotation and translation matrices is used. The regions with base weights refer to regions with density gradient values ​​below the first threshold and assigned weights of 0.1, mainly the central area of ​​the ventricles and the homogeneous white matter tract region. These regions have uniform density, and slight head movements only cause overall displacement without local deformation, so complex non-rigid correction is not required. The rotation and translation matrix is ​​a mathematical matrix that describes the overall rotation and translation in three-dimensional space. The matrix contains translation amounts in the x, y, and z directions and rotation angles around the three axes, which can complete the uniform displacement adjustment of the entire region through a single matrix operation. Rigid correction is a correction method that keeps the internal topology of the region unchanged. All pixels are adjusted by the same rotation angle and translation amount to ensure that the internal relative positions of the ventricles and white matter tracts do not change, avoiding structural distortion caused by overcorrection. The original topology refers to the relative positional relationship between pixels in the base weight region. Rigid correction ensures that the tissue morphology is consistent with the real anatomical structure by maintaining this relationship. The specific implementation process is as follows: First, the overall displacement parameters (translation and rotation angle) of the basic weight region are extracted from the correlation law model and converted into the corresponding rotation and translation matrix. This matrix is ​​applied to the original coordinates of all pixels in the region, and the position adjustment of all pixels is completed at once through matrix operation. For example, if a region needs to be translated 2 pixels along the x-axis and rotated 0.5 degrees around the z-axis, the matrix operation will adjust the coordinates of all pixels according to the same rules. During the correction process, the local deformation correction module is disabled simultaneously, and only the computational resources related to matrix operation are retained, which improves the computational efficiency by more than 50%. After correction, the overall displacement of the basic weight region is completely canceled, the image position is aligned with the state without micro-motion, and the internal topology remains unchanged, which meets the calibration requirements of homogeneous tissue. Through the above differentiated correction strategy, the image calibration execution unit 3 not only accurately handles the local micro-deformation of the highest weight region through non-rigid deformation correction algorithm, but also efficiently handles the overall displacement of the basic weight region through rigid correction, ensuring that the calibration requirements of different types of regions are met. This lays the core foundation for the subsequent integration of partition correction results and output of clear and accurate brain CT images.

[0027] In this invention, the density gradient analysis unit 1 adaptively partitions the brain according to its anatomical characteristics, and calculates the density gradient value of each region through dual-scale fusion of three-dimensional directional gradient detection and Laplacian differential operator; the displacement compensation weight unit relies on a correlation law model trained on big data, and generates a continuous weight field through two-level decision-making, and applies specific compensation coefficients to key regions; the image calibration execution unit 3 adopts a differentiated algorithm to perform sub-pixel-level non-rigid correction on high-density abrupt change regions and rigid correction on homogeneous regions; by integrating the partition correction results, the influence of head micro-movements is effectively offset, tissue boundary blurring is eliminated, and the clarity and structural accuracy of brain CT images are improved, providing reliable support for clinical diagnosis.

[0028] Please see Figure 2 As shown, a second objective of this invention is to provide a method for implementing an intelligent calibration system for patient brain CT images based on big data, including any of the above-mentioned features, comprising the following steps: S1. After receiving brain CT images, high-resolution grid units are divided according to the anatomical distribution of gray matter nuclei in the basal ganglia region. The grid boundaries are dynamically densified in the gray matter-white matter boundary area. The density gradient value of each local region is calculated by fusing the three-dimensional directional gradient detection operator with the Laplacian differential operator of the downsampling layer. S2. Based on the density gradient value, the basic weight coefficients are output through the pre-trained association law model. Combined with the real-time anatomical location, a spatial index query is performed in the weight mapping table. Nucleus-specific compensation coefficients are loaded for the subthalamic nucleus region. A continuous displacement compensation weight field is generated through spatial interpolation. S3. For calcifications and vascular wall boundary regions with density gradient values ​​higher than the second threshold, activate the sub-pixel level compensation mode and superimpose the real-time gradient change rate increment. Use thin plate spline interpolation to construct a local deformation field to adjust pixel coordinates. For ventricles and white matter tract regions with density gradient values ​​lower than the first threshold, call the six-degree-of-freedom rigid body transformation model to perform overall rigidity correction. S4. Integrate the partition correction results, eliminate tissue boundary ambiguity, and output the calibrated brain CT image.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A big data-based intelligent calibration system for patient brain CT images, characterized in that, include: The density gradient analysis unit (1) receives the input brain CT image, divides the brain CT image into multiple local regions, and calculates the density gradient value of each local region based on the pixel value change rate of the brain CT image, wherein the density gradient value is used to quantify the degree of spatial mutation of brain tissue density. The displacement compensation weight dynamic adjustment unit (2) uses a pre-learned association law model to determine the displacement compensation weight of each local area based on the density gradient value. The association law model is established by big data analysis of historical CT image datasets. The historical CT image dataset contains pixel displacement records caused by head micro-movements and density gradient information of the corresponding areas. The association law model learns the positive correlation between density gradient value and displacement compensation requirement. When the density gradient value is lower than the first threshold, the basic weight is assigned. When the density gradient value is between the first threshold and the second threshold, the weight increases linearly. When the density gradient value is higher than the second threshold, the highest weight is assigned. The weight value range is predefined and stored in the weight mapping table for real-time query. The image calibration execution unit (3) applies a weighted displacement compensation operation to each local region based on the displacement compensation weight. The displacement correction algorithm is used for the local region with the highest weight to offset the intra-layer displacement caused by the micro-movement of the head by adjusting the pixel position. For the local region with the basic weight, standard rigid displacement correction is used.

2. The intelligent calibration system for patient brain CT images based on big data according to claim 1, characterized in that: The density gradient analysis unit (1) divides the brain CT image into multiple local regions using an adaptive anatomical partitioning strategy, specifically including: High-resolution mesh units are generated based on the anatomical distribution characteristics of gray matter nuclei in the basal ganglia. A dynamic densification mesh division method is adopted in the tissue boundary region where gray matter and white matter meet. The mesh coverage is expanded for homogeneous regions filled with cerebrospinal fluid. An overlapping buffer is set at the boundary of each mesh unit to eliminate division errors. The density gradient value is calculated in each mesh through multi-directional pixel value difference operation.

3. The intelligent calibration system for patient brain CT images based on big data according to claim 2, characterized in that, The principal gradient vector is obtained by applying a three-dimensional directional gradient detection operator to the resolution layer of the original brain CT image. At the same time, the region density change features are captured by the Laplacian differential operator in the macroscale layer generated by downsampling. Finally, the dual-scale calculation results are fused by the gradient sensitivity weighting function.

4. The intelligent calibration system for patient brain CT images based on big data according to claim 1, characterized in that: The correlation model is constructed through a multimodal data-driven architecture, specifically including: The data layer integrates a surgical navigation system to record head micro-movement trajectories, brain CT image sequences, and ground truth values ​​of tissue boundaries annotated by high-resolution MRI. The analysis layer uses a gradient boosting decision tree algorithm to mine the three-dimensional mapping relationship between density gradient values ​​and displacement compensation requirements. The optimization layer introduces an adversarial sample generator to synthesize calcification foci and blood vessel wall boundary samples in the basal ganglia region.

5. The intelligent calibration system for patient brain CT images based on big data according to claim 4, characterized in that: The displacement compensation weight dynamic adjustment unit (2) adopts a two-level weight decision logic, specifically including: The first level outputs basic weight coefficients through a correlation model. The second level combines real-time anatomical location information to perform spatial index lookup in the weight mapping table, loads nucleus-specific compensation coefficients for the subthalamic nucleus region, and finally generates a continuously distributed displacement compensation weight field through a spatial interpolation algorithm.

6. The intelligent calibration system for patient brain CT images based on big data according to claim 5, characterized in that: When the density gradient value is below a first threshold, the assigned base weights undergo standardized rigid correction, specifically including: For the central ventricle and homogeneous white matter tract regions, a six-degree-of-freedom rigid body transformation model is invoked for overall displacement compensation. At the same time, the local deformation correction module is disabled to reduce the computational load. The basic weight value is fixed as the minimum compensation intensity defined by the weight mapping table.

7. The intelligent calibration system for patient brain CT images based on big data according to claim 1, characterized in that: The linearly increasing weights with density gradient values ​​between the first and second thresholds are implemented using a piecewise ramp function, specifically including: Gradient-dependent weight growth curves are constructed in the transition region between the shell and the globus pallidus. The weight growth curves adopt an exponential smooth transition in the steep section at the tissue boundary and switch to a linear growth mode in the homogeneous tissue region. The weight increment is directly proportional to the rate of change of the density gradient.

8. The intelligent calibration system for patient brain CT images based on big data according to claim 1, characterized in that: The highest weights assigned when the density gradient value is above the second threshold are generated through a dynamic boundary reinforcement mechanism, specifically including: In the basal ganglia region, calcifications and the boundary region of blood vessel walls are activated to form a subpixel-level compensation mode. Preset peak weight coefficients are extracted from the weight mapping table and superimposed with adaptive increments based on real-time gradient change rate to form an overcompensation effect for micro-displacement.

9. The intelligent calibration system for patient brain CT images based on big data according to claim 1, characterized in that: The image calibration execution unit (3) employs a non-rigid deformation correction algorithm for the local region with the highest assigned weight, specifically including: A local deformation field is constructed using thin-plate spline interpolation to adjust the spatial coordinates of the boundary pixels of calcification foci. For regions with assigned base weights, rigid correction based on rotation and translation matrices is used to preserve their original topological structure without deformation.

10. A method for implementing the intelligent calibration system for patient brain CT images based on big data as described in any one of claims 1-9, characterized in that: Includes the following steps: S1. After receiving brain CT images, high-resolution grid units are divided according to the anatomical distribution of gray matter nuclei in the basal ganglia region. The grid boundaries are dynamically densified in the gray matter-white matter boundary area. The density gradient value of each local region is calculated by fusing the three-dimensional directional gradient detection operator with the Laplacian differential operator of the downsampling layer. S2. Based on the density gradient value, the basic weight coefficients are output through the pre-trained association law model. Combined with the real-time anatomical location, a spatial index query is performed in the weight mapping table. Nucleus-specific compensation coefficients are loaded for the subthalamic nucleus region. A continuous displacement compensation weight field is generated through spatial interpolation. S3. For calcifications and vascular wall boundary regions with density gradient values ​​higher than the second threshold, activate the sub-pixel level compensation mode and superimpose the real-time gradient change rate increment. Use thin plate spline interpolation to construct a local deformation field to adjust pixel coordinates. For ventricles and white matter tract regions with density gradient values ​​lower than the first threshold, call the six-degree-of-freedom rigid body transformation model to perform overall rigidity correction. S4. Integrate the partition correction results, eliminate tissue boundary ambiguity, and output the calibrated brain CT image.

Images