A cloud storage system for medical image archives

Abstract

This invention relates to the field of cloud storage technology and discloses a cloud storage system for medical image archiving, comprising: a data segmentation unit for segmenting the body data to be archived to extract a metal mask; a first data projection unit for performing projection transformation on the metal mask to obtain a metal trace mask; a second data projection unit for performing projection transformation on the reference body data and the body data to be archived respectively to generate corresponding projection data maps; a data processing unit for calculating angular Fisher information within the projection angle range covered by the metal trace mask and forming an angle sequence; an unstable angle extraction unit for summarizing projection angles that meet the conditions based on feature value thresholds to obtain an unstable angle set; and a cloud storage unit for constructing angular channel incremental data and spatial channel incremental data based on the unstable angle set and completing cloud storage.

Description

Technical Field

[0001] This invention relates to the field of cloud storage technology, and more specifically, to a cloud storage system for medical image archiving. Background Technology

[0002] In clinical practice, CT scans generate massive amounts of data. Hospitals or imaging centers typically need to store this data long-term and share it across different departments or institutions via networks. To reduce storage and transmission costs, a common approach is incremental archiving, which only stores the differences from previous versions. This method can significantly reduce redundancy and improve archiving efficiency in typical scenarios.

[0003] However, when metallic materials are present in the patient's body, the strong absorption and scattering of X-rays by the metal can produce artifacts such as extensive stripes and dark bands in the image. Different reconstruction strategies (such as iterative reconstruction, model-based correction, or artifact suppression algorithms) suppress or correct these artifacts in different ways, resulting in images that vary significantly in texture, noise distribution, and local density. Even under the same examination, strong inconsistencies may appear between reconstructed versions in local areas.

[0004] These local inconsistencies manifest as large, directional, and irregular variations in the residual image. Since traditional incremental archiving methods mostly assume linear similarity of spatial blocks or pixel values, these variations cannot be effectively modeled or reused, leading to an increase in the statistical entropy of the residual. Increased entropy means greater compression difficulty, and the compression efficiency of the incremental chain rapidly declines. Ultimately, when metal-related regions frequently appear in head CT scans, this efficiency collapse is repeatedly triggered, negating the advantages of incremental storage and resulting in a significant waste of storage resources and network bandwidth. Summary of the Invention

[0005] This invention provides a cloud storage system for medical image archiving, which solves the technical problems mentioned in the background art.

[0006] This invention provides a cloud storage system for medical image archiving, comprising:

[0007] The data acquisition unit reads reference version data and archived version data. The reference version data is historical reconstructed version voxel data, and the archived version data is newly reconstructed version voxel data.

[0008] The data segmentation unit performs image segmentation on the archived version data and extracts the metal mask.

[0009] The first data projection unit performs a projection transformation on the metal mask to obtain the metal trace mask corresponding to the metal mask.

[0010] The second data projection unit performs projection transformations on the reference version body data and the version body data to be archived, respectively, to obtain the reference version projection data map and the version projection data map to be archived.

[0011] The data processing unit calculates the angular Fisher information of the projected data map of the version to be archived at each angle in the order of the projection angle within the projection angle range covered by the metal trace mask, and obtains the angular Fisher information angle sequence.

[0012] The unstable angle extraction unit extracts the feature values ​​of the angle sequence of the angle to Fisher information, sets a threshold based on the feature values, and summarizes the projection angles whose angle to Fisher information values ​​meet the threshold conditions to form an unstable angle set.

[0013] The cloud storage unit constructs incremental data for angle channels and incremental data for spatial channels using unstable angle sets, and then processes the incremental data for cloud storage.

[0014] The beneficial effects of this invention include: by identifying and separating angles and regions affected by metal in the projection domain, incremental data for angular and spatial channels are constructed only for these unstable parts, effectively avoiding the compression efficiency collapse caused by excessively high residual entropy values ​​in traditional methods. Therefore, this invention significantly improves the compression efficiency and storage stability of incremental archiving in metal-related scenarios, reduces the waste of storage resources and network bandwidth, and ensures the economy and reliability of long-term archiving and cross-institutional sharing of clinical images. Attached Figure Description

[0015] Figure 1 This is a block diagram of a cloud storage system for medical image archiving according to the present invention. Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0017] like Figure 1 As shown, a cloud storage system for medical image archiving includes:

[0018] The data acquisition unit reads reference version data and archived version data. The reference version data is historical reconstructed version voxel data, and the archived version data is newly reconstructed version voxel data.

[0019] The data segmentation unit performs image segmentation on the archived version data and extracts the metal mask.

[0020] The first data projection unit performs a projection transformation on the metal mask to obtain the metal trace mask corresponding to the metal mask.

[0021] The second data projection unit performs projection transformations on the reference version body data and the version body data to be archived, respectively, to obtain the reference version projection data map and the version projection data map to be archived.

[0022] The data processing unit calculates the angular Fisher information of the projected data map of the version to be archived at each angle in the order of the projection angle within the projection angle range covered by the metal trace mask, and obtains the angular Fisher information angle sequence.

[0023] The unstable angle extraction unit extracts the feature values ​​of the angle sequence of the angle to Fisher information, sets a threshold based on the feature values, and summarizes the projection angles whose angle to Fisher information values ​​meet the threshold conditions to form an unstable angle set.

[0024] The cloud storage unit constructs incremental data for angle channels and incremental data for spatial channels using unstable angle sets, and then processes the incremental data for cloud storage.

[0025] In detail, the reference version of the voxel data is a collection of voxel data generated and archived after medical imaging examinations based on the original acquired projection data using a historical reconstruction algorithm. The reference version of the voxel data includes three-dimensional spatial location information (spatial coordinates corresponding to the patient's anatomical structures) and voxel values ​​for each spatial location (such as CT values ​​in CT scans, reflecting tissue density).

[0026] In detail, the voxel data to be archived is a set of voxel data generated by a new reconstruction algorithm in the same medical imaging examination (originating from the same original projection data acquisition as the reference version) to meet new clinical needs (such as optimizing artifact suppression and improving the clarity of local structures) and has not yet been archived.

[0027] For example, a patient presents with a head injury and requires a cranial CT scan. The patient also has metal dentures in their mouth (which are prone to metal artifacts). Based on the raw X-ray projection data collected during this examination, the hospital generates two different reconstructed voxel data sets.

[0028] Reference version body data generation, including:

[0029] Within one hour of the patient's CT scan, to quickly provide a preliminary diagnostic report, the raw projection data was processed using the clinically standard and computationally fast Filtered Back Projection (FBP) algorithm to generate the first version of head CT voxel data. This version of the voxel data contains complete information on the anatomical structures of the skull, such as brain tissue, skull, and sinuses, but there are obvious streaky artifacts around the metal dentures (due to the weak suppression ability of FBP to the X-ray beam hardening and scattering effects caused by metal); the voxel spacing is 0.5mm × 0.5mm × 1.0mm (spatial resolution meets the requirements of routine diagnosis), and the CT value ranges from -1000HU (air) to 3000HU (metal dentures); this version of the data has been archived through the hospital's PACS system for emergency room physicians to make a preliminary assessment of whether there is a skull fracture or intracranial hemorrhage.

[0030] The generation of the archived version body data includes:

[0031] The following day, when the radiologist reviewed the preliminary report, they discovered that artifacts from the metal dentures interfered with the clear visualization of the adjacent temporal lobe brain tissue. To rule out the possibility of localized brain contusion, iterative reconstruction (IR) algorithms with stronger metal artifact suppression capabilities were used to reprocess the original projection data from the same examination, generating a second version of head CT voxel data. The spatial coordinates of this version of the voxel data were completely consistent with the reference version (ensuring a perfect correspondence between the spatial positions of the temporal lobe brain tissue and the metal dentures), but the stripe artifacts around the metal dentures were significantly reduced (IR iteratively optimized the projection data to correct the model, reducing metal interference); the voxel spacing remained at 0.5mm × 0.5mm × 1.0mm (consistent with the reference version, ensuring effective comparison), but the CT value distribution in the area adjacent to the metal dentures was closer to that of normal brain tissue (the abnormal low-density / high-density areas caused by artifacts were reduced); this version of the data had not yet been archived and needed to be incrementally archived based on the differences from the reference version using the cloud storage system of this invention, avoiding the resource waste caused by full storage.

[0032] In one embodiment of the present invention, image segmentation of the archived version data and extraction of the metal mask include:

[0033] Read the volume data and voxel spacing of the version to be archived, and use the voxel grid of the volume data to be archived as the processing grid. Perform strength truncation with the lower strength limit and the upper strength limit to obtain truncated volume data. Perform binarization with the metal strength threshold to obtain the initial binary volume data.

[0034] The initial binary volume data is subjected to a three-dimensional 26-neighbor connected component screening, and only connected regions with a voxel count not less than the minimum voxel count threshold are retained; wherein, the minimum voxel count threshold is determined by converting the minimum physical volume parameter and the voxel spacing.

[0035] A three-dimensional spherical structuring element is used to perform the closing operation; the radius of the structuring element for the closing operation is determined by converting the physical radius of the closing operation with the voxel spacing.

[0036] Expansion is performed using a three-dimensional spherical structural element; the radius of the expanding structural element is determined by converting the safe expansion physical radius and the voxel spacing.

[0037] The metal mask is obtained based on the closing operation and expansion.

[0038] In detail, a voxel grid is a three-dimensional spatial structure framework for the voxel data to be archived. Each node of the grid corresponds to the spatial location of a voxel, and the grid arrangement (such as the number of rows and columns, and the spacing between adjacent nodes) perfectly matches the spatial distribution of the voxel data to be archived.

[0039] In detail, since medical images often contain irrelevant intensity interference (such as extremely low intensity values ​​in air regions in CT scans or abnormally high intensity values ​​caused by equipment noise), intensity truncation can limit the effective intensity range of volumetric data and eliminate irrelevant interference data. Therefore, based on the clinical characteristics and equipment parameters of medical images, preset lower and upper intensity limits are established. Each voxel in the volumetric data to be archived is iterated, and voxel intensity values ​​below the lower limit are adjusted to the lower limit, while those above the upper limit are adjusted to the upper limit. Voxels with intensity values ​​between the two limits are left unchanged. The volumetric data obtained after this processing is the truncated volumetric data.

[0040] In detail, based on the characteristic strength of metallic materials in medical images (e.g., the CT value of metal in CT scans is usually significantly higher than that of human tissue, generally above 1000 HU), a metal strength threshold is preset; each voxel in the truncated volume data is traversed, and voxels with a strength value greater than or equal to the metal strength threshold are marked as the first value (1), and voxels with a strength value less than the metal strength threshold are marked as the second value (0). The volume data obtained after this processing, which contains only two types of values, is the initial binary volume data. Through binarization, the metallic and non-metallic regions in the volume data to be archived are initially separated, where the region with the first value initially corresponds to the metallic region, and the region with the second value initially corresponds to the non-metallic region.

[0041] In detail, within the 3D voxel mesh of the initial binary data, the 26-neighborhood of each voxel refers to the set of all spatially adjacent voxels. Specifically, this includes: 6 voxels directly adjacent to the voxel along the x-axis (left-right), y-axis (front-back), and z-axis (up-down); 12 voxels diagonally adjacent along the xy, xz, and yz axes; and 8 voxels diagonally adjacent along the xyz axes, totaling 26 voxels. The neighborhood definition is a spatial criterion for determining the connectivity between voxels; that is, if two voxels marked as 1 belong to the same 26-neighborhood (connected directly or indirectly through 26-neighborhood voxels), they are considered to be in the same connected region.

[0042] In detail, all voxels marked as 1 in the initial binary data (preliminary metal candidate voxels) are traversed. Based on the aforementioned three-dimensional 26-neighborhood rule, interconnected clusters of 1-value voxels (directly or indirectly connected through 26-neighborhoods) are divided into multiple independent connected regions. Each connected region represents a continuous metal candidate region in the initial binary data. For each identified connected region, the total number of 1-value voxels it contains is counted to obtain the voxel count for each connected region. In the initial binary data, voxels marked as 1 may contain noise (such as isolated small voxel clusters caused by device interference) or minor artifacts (high-grayscale areas that are not real metal). By filtering connected regions, these small volume interference regions with no clinical significance can be removed, ensuring that the remaining connected regions are voxel clusters corresponding to real metal structures (such as dentures, fixation plates).

[0043] In detail, the minimum physical volume parameter is a preset minimum physical volume of a real metal structure based on actual clinical scenarios (e.g., set according to the minimum size of common dental dentures and cranial fixation plates, in cubic millimeters). This parameter is the physical basis for distinguishing real metal from small interfering areas. The voxel spacing is the three-dimensional spatial spacing (x-axis spacing, y-axis spacing, z-axis spacing, all in millimeters) of the voxel data to be archived, reflecting the actual physical space size corresponding to a single voxel. The physical volume of a single voxel is calculated as follows: Physical volume of a single voxel = x-axis spacing × y-axis spacing × z-axis spacing (in cubic millimeters). The minimum voxel number threshold is calculated as follows: Minimum voxel number threshold = Minimum physical volume parameter ÷ Physical volume of a single voxel. If the calculated result is not an integer, it needs to be rounded up (to ensure the threshold is not less than the theoretically calculated value, avoiding omission of small volumes of real metal). If the minimum physical volume parameter is 8 cubic millimeters and the voxel spacing is 0.5mm×0.5mm×0.5mm, then the physical volume of a single voxel is 0.125 cubic millimeters, and the minimum voxel number threshold is 8÷0.125=64, meaning that only connected regions with a voxel number ≥64 are retained.

[0044] It's important to note that the closing operation fills in tiny holes within the metal region, ensuring the integrity of its internal structure. The dilation operation expands the boundaries of the metal region, covering any missed tiny metal parts and artifact-affected areas. The volume data processed by the closing operation is used as input for the dilation operation. After dilation, the resulting volume data shows regions with a voxel value of 1 completely covering the real metal and its surrounding areas of interest, while regions with a voxel value of 0 represent non-metallic regions. This volume data is the final metal mask. The closing operation's physical radius is a preset physical space size used to fill in holes within the metal region. It is set based on the minimum hole size of common clinical metal structures (such as dentures and fixation plates), in millimeters, ensuring coverage of any tiny holes that may exist within the metal region. The voxel spacing is the three-dimensional spatial spacing (x-axis, y-axis, z-axis spacing) of the volume data to be archived, in millimeters, reflecting the physical space size corresponding to a single voxel. The calculation method for the radius (number of voxels) of the closing operation structuring element is as follows: divide the physical radius of the closing operation by the voxel spacing (usually taking the minimum value among the three-dimensional voxel spacings to ensure that the structuring element covers enough voxels in all directions). If the calculation result is not an integer, it needs to be rounded up to ensure that the spherical region of the structuring element is not smaller than the preset physical radius of the closing operation in physical space, thereby effectively filling the holes inside the metal region. The safe expansion physical radius is a preset real physical space size used to expand the boundary of the metal region. It is set according to the influence range of metal artifacts and is in millimeters. It ensures that the expanded metal region can cover the metal and its surrounding potential areas affected by artifacts. The calculation method for the radius (number of voxels) of the expanding structuring element is the same as that of the closing operation: divide the safe expansion physical radius by the voxel spacing (taking the minimum value among the three-dimensional voxel spacings). If the calculation result is not an integer, it is rounded up to ensure that the structuring element is not smaller than the preset safe expansion physical radius in physical space, thereby achieving effective expansion of the metal region boundary.

[0045] In one embodiment of the present invention, a projection transformation is performed on the metal mask to obtain a metal trace mask corresponding to the metal mask, including:

[0046] Define the projection geometry of the parallel beam projection model. The projection geometry includes: angle set, angle step size, detector coordinate set, and detector step size.

[0047] Length-weighted projection is performed on the metal mask to calculate the total geometric length of each ray passing through the metal mask voxels, resulting in a length sine curve. Each ray corresponds to an angle in the projection geometry and a detector coordinate. The total geometric length is the sum of the geometric passage lengths within each metal voxel through which the ray passes.

[0048] Calculate the minimum intersection length threshold, which is the safety factor multiplied by the minimum voxel spacing. The minimum voxel spacing is the minimum of the three voxel spacings, and the safety factor is a dimensionless constant not less than one. Binarize the length sine curve with the minimum intersection length threshold, mark the positions in the length sine curve that are not less than the threshold as one, and mark the rest as zero, to obtain the original metal trace mask.

[0049] The original metal trace mask is subjected to a first expansion along the detector direction. The first expansion adopts a one-dimensional symmetric structure element. The number of sampling points of the one-dimensional symmetric structure element is obtained by dividing the physical radius of the detector direction safe expansion by the detector step size and rounding up, thus obtaining a detector direction coherent mask.

[0050] A second expansion is performed on the detector directional coherent mask along the angular direction. The second expansion uses a one-dimensional symmetric structural element. The number of sampling points of the one-dimensional symmetric structural element is obtained by dividing the angular expansion width by the angular step size and rounding up, thus obtaining the metal trace mask.

[0051] In detail, the parallel beam projection model is the basic geometric framework for transforming a three-dimensional metal mask into a two-dimensional metal trace mask in this invention. Its characteristic is that the projected rays are parallel beams, all rays propagate along the same preset angle and are received by the detector. By setting uniform projection geometric parameters, the projection process of the metal mask is ensured to be quantifiable and reproducible.

[0052] In detail, the angle set is a set of discrete angle values ​​representing the emission direction of the projection ray in three-dimensional space, reflecting all directions of rotation of the projection ray around the imaging center. In clinical scenarios, the angle set typically covers the complete rotation range (0 degrees to 360 degrees) and is discretized into several specific angle values ​​at fixed intervals. For example, the angle set can be represented as 0 degrees, 0.2 degrees, 0.4 degrees...359.8 degrees, containing 1800 discrete angles. The angle set determines the coverage direction of the projection ray on the metal mask, ensuring that the projection information of the metal mask is captured from multiple different angles, avoiding the omission of metal traces due to missing angles.

[0053] In detail, the angle step size is the difference between two adjacent discrete angles in the angle set. It is calculated by dividing the total rotation range of the angle set by the total number of discrete angles (if the total rotation range is 360 degrees and the number of discrete angles is 1800, then the angle step size is 360 degrees ÷ 1800 = 0.2 degrees). The size of the angle step size is determined by the imaging accuracy requirements; the higher the accuracy requirement, the smaller the angle step size. The angle step size controls the sampling density of the projection rays in the angular direction. The smaller the step size, the smaller the difference in projection information between adjacent angles, and the stronger the continuity of the metal traces in the angular direction. This avoids breaks in the metal traces due to excessively large angular intervals, ensuring that the metal trace mask accurately reflects the spatial shape of the metal mask.

[0054] In detail, the detector coordinate set is the set of discrete position values ​​of the detector elements receiving the projected rays on the detector plane. Detectors are typically arranged in a straight line, and the detector coordinate set corresponds to the center position coordinates of these elements. For example, for detectors arranged horizontally, the coordinate set can be represented as -500 mm, -499.5 mm...499.5 mm, 500 mm, containing 2001 discrete coordinate values ​​(the origin is usually aligned with the imaging center). The detector coordinate set determines the receiving position of the projected rays on the detector; each coordinate corresponds to one detector element, converting the energy of projected rays from different directions into specific positional data.

[0055] In detail, the detector step size is the difference between two adjacent discrete coordinates in the detector coordinate set, that is, the physical distance between the centers of two adjacent detector elements. For example, if the detector coordinate set ranges from -500 mm to 500 mm and contains 2001 coordinate values, then the detector step size is (500 mm - (-500 mm)) ÷ (2001 - 1) = 0.5 mm. The detector step size controls the sampling density of the projection rays in the detector direction. The smaller the step size, the higher the spatial resolution of the detector element, which can capture the fine structure of the metal mask projection, avoid the loss of metal trace details due to excessively large detector element spacing, and ensure that the metal trace mask can reflect the projected outline of the metal mask.

[0056] In detail, length-weighted projection is the key operation for converting a 3D metal mask (volume data) into 2D projection data, based on the previously established parallel beam projection model. Its logic is to map the spatial distribution information of the 3D metal region onto a 2D projection plane by calculating the total geometric length of each projection ray passing through the metal mask. Each ray corresponds to a unique combination of an angle and a detector coordinate in the parallel beam projection model. Angle dimension: The emission direction of the ray is determined by the set of angles in the projection geometry (e.g., discrete angles such as 0 degrees, 0.2 degrees, etc.); Detector dimension: The receiving position of the ray is determined by the set of detector coordinates in the projection geometry (e.g., discrete coordinates such as -500 mm, -499.5 mm, etc.). All rays are parallel beams and cover the entire range of angles and detector coordinates set by the projection geometry. Calculation of the geometric passage length of a single voxel: For a single ray, it is determined whether it intersects with a voxel (voxel value 1) in the metal mask. If intersecting, calculate the actual physical distance (in millimeters) the ray travels within the voxel based on its spatial direction and the voxel's 3D boundary. This distance is the geometrical length of the voxel (not the number of voxels, but the actual spatial distance). Accumulation of total geometrical lengths: Iterate through all metal mask voxels traversed by a single ray, summing the geometrical lengths of each voxel to obtain the total geometrical length corresponding to the ray. The total geometrical length reflects the degree of intersection between the ray and the metal region; a larger total geometrical length indicates a thicker metal region traversed by the ray. Iterate through all rays in the parallel beam projection model (i.e., all angle + detector coordinate combinations), using the total geometrical length of each ray as the pixel value of the corresponding angle-detector coordinate position, constructing a 2D data matrix. This matrix is ​​the length sine curve: the row dimension of the length sine curve corresponds to the set of angles in the projection geometry (each row index corresponds to a discrete angle); the column dimension corresponds to the set of detector coordinates in the projection geometry (each column index corresponds to a discrete detector coordinate); the value of each element in the matrix is ​​the total geometrical length of the corresponding ray, and the value range is determined by the actual thickness of the metal mask (the thicker the metal region, the larger the value).

[0057] In detail, voxel spacing is the physical distance between the centers of adjacent voxels in three-dimensional space (x-axis, y-axis, z-axis) of the volumetric data to be archived, and there are three values. The minimum voxel spacing is the minimum of these three voxel spacings, which reflects the highest precision of the volumetric data spatial resolution.

[0058] In detail, the safety factor is a preset dimensionless constant with a value not less than one. Its function is to compensate for the error in the length calculation when the ray intersects the edge of the metal voxel, to avoid misjudging it as a valid metal projection when the ray only slightly touches the edge of the voxel (the geometric passage length is extremely small), and to ensure that the threshold can filter out rays that actually pass through the metal region.

[0059] In detail, the minimum intersection length threshold is calculated as the product of a safety factor and a minimum voxel spacing. This calculation is directly related to the spatial accuracy and error compensation requirements of the volume data. The unit of the resulting threshold is consistent with the voxel spacing (e.g., millimeters), representing the minimum physical length that a ray must effectively pass through a metal voxel. That is, only when the total geometric length of the ray passing through the metal mask reaches or exceeds this threshold is it considered a valid intersection with the metal region. For example, if the minimum voxel spacing is 0.5 millimeters and the safety factor is set to 1.2, then the minimum intersection length threshold is 0.5 × 1.2 = 0.6 millimeters. This means that the total geometric length of the ray passing through the metal mask must be ≥ 0.6 millimeters to be considered a valid metal projection.

[0060] In detail, the binarization operation targets the length sine curve (a two-dimensional matrix, where each element represents the total geometric length of each ray passing through the metal mask). The operation involves traversing each position of the length sine curve (each position corresponds to a ray with an angle and a detector coordinate in the projection geometry): if the total geometric length at that position is not less than the minimum intersection length threshold, the value at that position is marked as one; if the total geometric length at that position is less than the minimum intersection length threshold, the value at that position is marked as zero. Thus, through numerical comparison and binary marking, effective metal projection rays and ineffective rays (such as those that only touch the metal edge or do not pass through the metal) are separated in the length sine curve. Positions marked as one correspond to the projection of the real metal region, while positions marked as zero correspond to non-metallic regions or noise interference.

[0061] In detail, after binarization, the length sine curve is transformed into a two-dimensional matrix containing only zeros and ones, which is the original metal trace mask.

[0062] In detail, a one-dimensional symmetrical structural element is a linear template extending along the detector direction. The two ends of the template are symmetrical with respect to the center, ensuring that the expansion operation expands uniformly on both sides (or in the positive and negative coordinate directions) in the detector direction, avoiding trace offset caused by the asymmetry of the structural element, and ensuring the accuracy of the spatial position of the metal trace in the detector direction.

[0063] In detail, the physical radius of the detector direction safety expansion is a preset actual physical length (in millimeters) used to cover the metal trace breakage in the detector direction. It is set according to the possible offset of the metal projection in the detector direction or the breakage width to ensure that it can fill the tiny breakage.

[0064] In detail, detector step size: the physical distance (in millimeters) between adjacent detector coordinates as defined in projective geometry, reflecting the sampling accuracy of the detector orientation.

[0065] In detail, the number of sampling points is the number of discrete units contained in a one-dimensional symmetric structuring element, which directly determines the expansion range. The calculation process consists of three steps:

[0066] Number of sampling points = physical radius of safe expansion in detector direction ÷ detector step size; the calculation result is rounded up to ensure that the physical length of the structural element is not less than the preset physical radius of safe expansion in detector direction, so as to avoid insufficient expansion range due to rounding and failure to completely fill the fracture.

[0067] In detail, a one-dimensional symmetric structuring element is traversed along the detector direction (column dimension) of the original metallic trace mask at each location: when a pixel marked as 1 (effective metallic projection) exists within the area covered by the structuring element, all locations covered by the structuring element are marked as 1. After this operation, the tiny breaks in the detector direction of the original metallic trace mask are filled, resulting in a detector-direction coherent mask, meaning that the metallic traces in the detector direction of this mask are completely continuous without local breaks.

[0068] In detail, the second expansion operation targets the detector orientation coherence mask, and the operation direction is the angular direction (i.e., the row dimension of the detector orientation coherence mask, corresponding to the set of angles in projection geometry). Its purpose is to fill in any minor breaks that may exist in the mask in the angular direction (such as discontinuities in traces caused by local missing parts of the metal projection at different angles), ensure the coherence of the metal traces in the angular direction, and ultimately form a mask that completely covers the metal projection.

[0069] In detail, the one-dimensional symmetrical structural element along the angular direction is a linear template extending along the angular sequence. The two ends of the template are symmetrical with respect to the center, ensuring that the expansion operation expands uniformly on both sides (or in the positive and negative angular directions) in the angular direction, avoiding trace offset caused by the asymmetry of the structural element, and ensuring the accuracy of the spatial position of the metal trace in the angular direction.

[0070] In detail, the safe expansion angle width in the angular direction is a preset actual angle range (in degrees) used to cover the metal trace fracture in the angular direction. It is set according to the possible offset or fracture angle of the metal projection in the angular direction to ensure that it can fill the small fracture.

[0071] In detail, the angle step size is the difference (in degrees) between adjacent discrete angles defined in projective geometry, reflecting the sampling accuracy of the angle direction.

[0072] In detail, the number of sampling points determines the range of expansion in the angular direction, and the calculation process consists of three steps:

[0073] Number of sampling points = safe expansion angle width in the angular direction ÷ angle step size; the calculation result is rounded up to ensure that the angle range of the structural element is not less than the preset safe expansion angle width in the angular direction, so as to avoid insufficient expansion range due to rounding and failure to completely fill the fracture.

[0074] The one-dimensional symmetric structuring element is traversed along the angular direction (row dimension) of the mask in the detector direction. When a pixel marked as 1 (effective metallic projection) exists within the area covered by the structuring element, all positions covered by the structuring element are marked as 1. After this operation, the tiny breaks in the angular direction of the mask in the detector direction are filled, resulting in the final metallic trace mask. That is, the mask remains complete and coherent in both the detector direction and the angular direction, accurately and without omission reflecting the entire projection area of ​​the three-dimensional metallic mask in the two-dimensional projection domain.

[0075] In one embodiment of the present invention, projection transformations are performed on the reference version body data and the version body data to be archived, respectively, to obtain a reference version projection data map and a version projection data map to be archived, including:

[0076] Under the parallel beam projection model, determine all ray paths corresponding to the angle set and the detector coordinate set;

[0077] Determine the geometrical passage length between the ray and the voxel, and use this geometrical passage length as the weight of the voxel in the corresponding ray accumulation;

[0078] Set the unit conversion factor used to convert volumetric data grayscale to projected cumulative measurement units;

[0079] The reference version projection data map is obtained by calculating the length-weighted projection value of the reference version body data.

[0080] The length-weighted projection value of the body data to be archived is calculated to obtain the projected data of the version to be archived.

[0081] In detail, the angle set includes all discrete projection angles in the projection model (such as 0 degrees and 0.2 degrees within the range of 0 degrees to 360 degrees). Each angle corresponds to a set of propagation directions of parallel rays:

[0082] Establish a three-dimensional coordinate system with the imaging center as the origin (usually the x-axis and y-axis are the horizontal and vertical directions in the imaging plane, and the z-axis is the depth direction of the volume data).

[0083] Each projection angle corresponds to the angle between the ray and the x-axis. Once the angle value is determined, the direction vector of the ray is fixed (e.g., when the angle is θ, the ray...). The direction vector in the plane is (The z-axis has no component, ensuring it is parallel to the imaging plane).

[0084] All rays at the same angle propagate along this fixed direction vector, differing only in position, thus achieving parallel beams with a unified direction but different positions.

[0085] In detail, the detector coordinate set contains all discrete receiver coordinates on the detector plane (e.g., -500 mm to -499.5 mm within the range of -500 mm to 500 mm). Each detector coordinate corresponds to the spatial position of a ray:

[0086] The detector plane is perpendicular to the direction of ray propagation (e.g., when the angle is θ, the normal direction of the detector plane is consistent with the ray direction vector), and there is a fixed spatial offset between the detector coordinates and the imaging center (to ensure that the ray can pass through the volume data area).

[0087] Each detector coordinate corresponds to a receiving unit on the detector plane. The spatial position of this unit is mapped to the incident position of a ray. That is, the ray starts from the side away from the volume data along a fixed direction, passes through the volume data, and is precisely incident on the receiving unit corresponding to the detector coordinate.

[0088] The rays corresponding to different detector coordinates are parallel in space and evenly spaced, with the interval equal to the detector step size (the difference between adjacent coordinates in the detector coordinate set).

[0089] In detail, iterate through each angle in the angle set and each coordinate in the detector coordinate set to form all unique combinations of angle-detector coordinates:

[0090] Each combination corresponds to an independent ray: the angle determines the direction of propagation of the ray, and the detector coordinates determine the spatial position of the ray;

[0091] All combinations cover the entire orientation and position range of the projection model: the angle set covers the complete rotation range (ensuring that volume data is projected from multiple directions), and the detector coordinate set covers the full width of the detector (ensuring that the lateral range of the volume data is fully projected).

[0092] All generated ray paths pass through the volume data region to be projected (the spatial range of the volume data in the reference version and the version to be archived).

[0093] In detail, each voxel in the volume data is traversed to determine whether the current ray intersects with that voxel. The determination is based on whether the ray passes through the three-dimensional boundary of the voxel (the voxel boundary is determined by the spatial coordinates of the voxel and the voxel spacing; for example, if the center coordinates of a voxel are x0, y0, z0, and the voxel spacing along the x-axis is dx, then the voxel boundary in the x-direction is from x0-dx / 2 to x0+dx / 2, and the same applies to the y-axis and z-axis). If the ray intersects with any boundary surface, then the ray is determined to intersect with that voxel.

[0094] In detail, for voxels determined to intersect, the coordinates of the intersection points between the ray and each boundary surface of the voxel are calculated. Using a three-dimensional rectangular coordinate system (consistent with the volume data space coordinate system) as the reference, based on the ray's direction vector (determined by the projection angle) and the ray's initial position (determined by the detector coordinates), the intersection points between the ray and the six boundary surfaces (x±, y±, z±) of the voxel are solved through linear equations, and the valid intersection points located inside the voxel (i.e., the two intersection points that pass through the voxel's inlet and outlet surfaces respectively) are selected.

[0095] In detail, based on the coordinates of the effective intersection points, the distance between two points is calculated using the distance formula between two points in space. This distance is the geometric length through which the ray passes the voxel. If the ray passes parallel to a certain axis of the voxel (such as along the x-axis), the length through is equal to the voxel spacing of the corresponding axis; if the ray passes obliquely through the voxel, the length through is greater than the voxel spacing along a single axis. The specific value is determined by the angle between the ray direction and the voxel boundary.

[0096] In detail, ray accumulation is the step in calculating the projection data map (reference version or archived version). The purpose is to convert the intensity information of three-dimensional voxels (such as the CT value in a CT scan, which reflects tissue density) into the cumulative metric value of the two-dimensional projection domain (i.e. the projection value corresponding to the ray), which is in line with the physical principle of medical image projection imaging (such as the attenuation degree of X-rays when passing through human tissue is positively correlated with tissue thickness).

[0097] In detail, for each ray, the intensity values ​​of all the voxels it intersects are multiplied by the geometric passage length corresponding to that voxel to obtain the weighted contribution value of that voxel to the current ray projection value. The weight here is the geometric passage length, which means that the degree of contribution of the voxel to the ray projection is positively correlated with the actual thickness of the voxel through which the ray passes. That is, the larger the passage length, the higher the proportion of the voxel in the ray propagation path, and the stronger the influence on the projection value.

[0098] In detail, the weighted contribution values ​​of all voxels intersecting a given ray are summed to obtain the projection value corresponding to that ray. This projection value is the numerical value of the corresponding angle-detector coordinate position in the projection data map. Through this weighted summation, it is ensured that the projection value accurately reflects the true attenuation or density distribution of the ray passing through the volume data area, avoiding projection value deviations caused by ignoring the differences in voxel thickness.

[0099] In detail, the unit conversion factor is a key parameter connecting volumetric data grayscale and projective cumulative measurement. Its function is to realize the physical unit conversion from volumetric data grayscale values ​​to projective cumulative measurement values. Volumetric data grayscale (such as the HU value in CT scans) reflects the density characteristics of tissues, while projective cumulative measurement (such as the attenuation coefficient integral of X-rays passing through tissues) reflects the physical cumulative effect of the interaction between rays and tissues. The two have different physical meanings, and a quantitative correlation needs to be established through this factor to ensure that the projective data conforms to the actual imaging physical process.

[0100] In detail, the unit conversion factor values ​​are determined based on the imaging principles and equipment parameters of medical imaging:

[0101] The relationship between grayscale and physical quantities: grayscale values ​​in volumetric data usually correspond to specific physical quantities (such as the HU value in CT, which is directly related to the linear attenuation coefficient of tissue). It is necessary to first clarify the conversion relationship between grayscale values ​​and this physical quantity (such as HU value = 1000 × (tissue linear attenuation coefficient - water linear attenuation coefficient) / water linear attenuation coefficient).

[0102] Definition of projection cumulative metric: The projection cumulative metric must match the imaging physical process (e.g., the cumulative attenuation of X-ray projection, unit: ...). (i.e., the dimensionless attenuation integral), combined with the above relationship between grayscale and physical quantities, the unit conversion factor is derived.

[0103] Equipment parameter calibration: The coefficients need to be fine-tuned in conjunction with the specific parameters of the imaging equipment (such as X-ray energy and detector sensitivity) to ensure that the converted cumulative measurement value can accurately reflect the real physical effect of the X-ray passing through the body data and avoid projection deviation caused by equipment differences.

[0104] Detailed calculation of single-ray projection values:

[0105] For each ray in the reference version body data, calculate the projection value of that ray according to the following steps:

[0106] Screening intersecting voxels: Identifies all reference version voxels that the current ray passes through;

[0107] Calculate the weighted contribution of voxels: For each intersecting voxel, calculate the voxel gray value × geometric passage length × unit conversion factor to obtain the weighted contribution of the voxel to the current ray projection value (i.e., the contribution of the voxel in the projection cumulative metric).

[0108] The cumulative projection value of a ray is obtained by summing the weighted contributions of all intersecting voxels of the current ray.

[0109] In detail, the projection values ​​of all rays are arranged according to the angle-detector coordinate dimension:

[0110] Row dimension: The set of angles corresponding to the parallel beam projection model, with each row index matching a discrete projection angle;

[0111] Column dimension: corresponds to the set of detector coordinates, with each column index matching a discrete detector coordinate;

[0112] Matrix filling: The projection value of each ray is filled into the corresponding angle-detector coordinate matrix position to form a two-dimensional data matrix, which is the reference version projection data map.

[0113] In detail, the length-weighted projection value is calculated for the archived version body data:

[0114] Calculation of projection value for a single ray: Traverse each ray, filter the voxels of the version to be archived that it passes through, calculate the gray value of each voxel × geometric passage length × unit conversion factor and sum them up to obtain the projection value of the ray.

[0115] Projection data generation for the archived version: Arrange the projection values ​​of all rays into a two-dimensional matrix according to the angle-detector coordinates to form the projection data for the archived version (the matrix size and units are completely consistent with the reference version projection data map).

[0116] In one embodiment of the present invention, within the projection angle range covered by the metal trace mask, the angular Fisher information of the projected data map of the version to be archived is calculated at each angle in the order of projection angles to obtain the angular Fisher information angle sequence, including:

[0117] Determine the effective projection angle set; where the effective projection angle set is the set of angles marked as one on the projection angle of the metal trace mask;

[0118] Based on the effective projection angle, extract the projection samples corresponding to the detector coordinates marked as one on the metal trace mask from the projection data map of the version to be archived, and calculate the sample mean and sample variance for each angle.

[0119] The rate of change of the sample mean and sample variance with respect to the angle is estimated according to the central difference rule in the order of projection angles, thus obtaining the rate of change of the sample mean and the rate of change of the sample variance; wherein, the first and last order of projection angles adopt the one-sided difference rule.

[0120] In an approximate form where the sample mean is used as a location parameter and the sample variance is used as a scale parameter, the rate of change of the sample mean and the rate of change of the sample variance are combined to form the angular Fisher information value.

[0121] Output the angle sequence of Fisher information according to the projection angle order.

[0122] In detail, the metallic trace mask is a two-dimensional matrix generated in the previous step, and its dimensions perfectly match the angle set and detector coordinate set of the parallel beam projection model:

[0123] The row dimension of the matrix corresponds to the set of projection angles in projection geometry, and each row uniquely corresponds to a discrete projection angle (such as 0 degrees, 0.2 degrees, etc.).

[0124] The column dimensions of the matrix correspond to the set of detector coordinates in projective geometry, and each column uniquely corresponds to a discrete detector coordinate (such as -500 mm, -499.5 mm, etc.).

[0125] Each element in the matrix takes the value of 0 or 1: when the element is 1, it means that the ray corresponding to the projection angle-detector coordinates passes through the metal region (belonging to the metal trace); when the element is 0, it means that the ray does not pass through the metal region (non-metal trace).

[0126] In detail, the effective projection angle set is a subset of the projection angle set. The determination of whether a row corresponding to a certain projection angle in the metal trace mask contains an element equal to 1 is performed as follows:

[0127] Traversing projection angles: Following the order of the projection angle set in projection geometry, check the rows (e.g., angles) of the corresponding angles in the metal trace mask one by one. Corresponding to the first row of the mask, angle (corresponding to line 2, etc.)

[0128] Row element check: For each row corresponding to an angle, determine whether the row contains at least one element with a value of 1. If it does, it means that at least one ray passes through the metal area at that angle (i.e., there is a metal trace at that angle); if it does not (all elements in the row are 0), it means that none of the rays at that angle pass through the metal area (no metal trace).

[0129] Set filtering: The projection angles corresponding to rows with at least one 1 element are summarized to form a set of valid projection angles.

[0130] In detail, the extraction of projected samples includes:

[0131] The rows corresponding to the valid angles are selected one by one in the order of the set of valid projection angles. For the current valid angle, the row corresponding to the angle is found in the projection data map of the version to be archived and the metal trace mask (since the two maps have the same dimension, the same angle corresponds to the same row index).

[0132] The detector coordinates (columns) corresponding to the metal traces are filtered by iterating through all columns (detector coordinates) in the row corresponding to the current effective angle of the metal trace mask and selecting the column indices with a value of 1. The detector coordinates corresponding to these column indices are the detector positions where the ray passes through the metal region at the current angle, which are the target positions for sample extraction.

[0133] Based on the column indices selected above, the projection values ​​corresponding to these column indices are extracted from the rows corresponding to the current valid angle in the projection data map of the version to be archived. These projection values ​​together constitute the projection sample for the current valid angle, meaning the sample only contains projection data of the area covered by the metal trace, which can accurately reflect the metal-related projection characteristics at that angle.

[0134] Detailed calculations of the sample mean and sample variance for each angle, including:

[0135] The sample mean is calculated by taking the arithmetic mean of all projected samples at the current valid angle. The sample mean reflects the central tendency of the projected values ​​of the metal trace region at the current angle and is a fundamental indicator for subsequent analysis of the overall level of the projected data at that angle. The formula is: Sample mean = (Sum of all projected sample values ​​at the current angle) ÷ (Number of projected samples at the current angle)

[0136] The sample variance is calculated by first determining the difference between each projected value and the sample mean for all projected samples at the current valid angle, then squaring the differences, and finally calculating the arithmetic mean of these squared values. This arithmetic mean is the sample variance for that angle. The sample variance reflects the dispersion of the projected values ​​of the metal trace region at the current angle and is a key indicator for subsequent analysis of the fluctuation characteristics of the projected data at that angle. The calculation formula is: Sample variance = [Sum of squares of (each projected sample value - sample mean)] ÷ (Number of projected samples at the current angle);

[0137] The calculation constraints are performed separately for each valid angle, ensuring that the sample mean and variance of different angles do not interfere with each other. Because the coverage area (sample size) and projection value distribution of the metal traces differ at different angles, separate calculations ensure that the statistical indicators for each angle can truly reflect its own characteristics.

[0138] Detailed central difference rules include:

[0139] Let the angle corresponding to the i-th position (2≤i≤N-1) in the projection angle order be . The sample mean is The sample variance is Its preceding rank (the first rank) The angle corresponding to the position is The sample mean is The sample variance is The next one (the first) The angle corresponding to the position is The sample mean is The sample variance is .

[0140] The rate of change of the sample mean reflects how quickly the sample mean changes with the angle at the i-th position. The formula is: Rate of change of the sample mean = Where the denominator For the first Rank and The physical difference of the position angle (in degrees), the numerator is the difference of the mean of the corresponding position sample, and the rate of change is in (projection value unit / degree).

[0141] Calculation of the rate of change of sample variance: The rate of change of sample variance reflects the first... At each position angle, the rate of change of the sample variance with respect to the angle is calculated using the formula: Rate of change of sample variance = The calculation logic is consistent with the rate of change of the sample mean, and the unit of the rate of change is (projected value unit² / degree).

[0142] Detailed one-sided difference rules include:

[0143] Let the angle corresponding to the first and second be . The sample mean is The sample variance is The angle corresponding to its adjacent next position (the second position) is The sample mean is The sample variance is .

[0144] Rate of change of sample mean =

[0145] Rate of change of sample variance =

[0146] Let the angle corresponding to the last position (Nth position) be... The sample mean is The sample variance is Its adjacent preceding position (the first) The angle corresponding to the position is The sample mean is The sample variance is .

[0147] Rate of change of sample mean = ;

[0148] Rate of change of sample variance = .

[0149] In detail, in the statistical analysis of projection data in this invention, the sample mean and sample variance play the roles of location parameter and scale parameter, respectively:

[0150] The sample mean reflects the central tendency of the projection values ​​of the metal trace area at this angle. The corresponding position parameter determines the central position of the projection data distribution, and its change reflects the overall horizontal angular shift of the projection value.

[0151] The sample variance (the sample variance of each effective angle projection calculated in the previous steps) reflects the dispersion of the projected values ​​of the metal trace region at that angle. The corresponding scale parameter determines the diffusion range of the projected data distribution, and its change reflects the angular difference in the amplitude of the projected value fluctuation.

[0152] In detail, the approximate form adopts the Gaussian distribution as the approximate statistical distribution model for the projection data. The reason is that the values ​​of the projection data of medical images in the local area of ​​the metal trace (excluding extreme noise) usually show continuous and approximately symmetrical distribution characteristics, which matches the statistical characteristics of the Gaussian distribution. Furthermore, the Fisher information of the Gaussian distribution can be directly calculated through the rate of change of the position parameter (mean) and the scale parameter (variance). This approximation can simplify the calculation and quantify the intensity of the abrupt change in the distribution characteristics of the projection data with the angle, which meets the system's requirements for the analysis of metal-related angular information.

[0153] The angular Fisher information value represents the contribution of the superimposed position and scale parameters to the angular changes. That is, the greater the rate of change of the sample mean (rate of change of the position parameter), the more significant the angular abrupt change in the center of the projected value; the greater the rate of change of the sample variance (rate of change of the scale parameter), the more significant the angular abrupt change in the dispersion of the projected value. Together, they constitute a quantitative index of the intensity of the angular information abrupt change at that angle, namely the angular Fisher information value.

[0154] In detail, for each effective projection angle (processed sequentially according to the projection angle), the rate of change of the sample mean based on that angle (denoted as ) ), the rate of change of sample variance (denoted as ), sample variance (denoted as ), calculate the angular Fisher information value using the following formula (denoted as ), ):

[0155] Location parameter contribution: This reflects the contribution of the rate of change of the sample mean to the information. It is calculated by dividing the square of the rate of change of the sample mean by the sample variance. ;

[0156] The reason for using the sample variance in the denominator is that the smaller the sample variance, the more concentrated the projected values ​​are, and the stronger the influence of sudden changes in the central position corresponding to the same rate of change of the mean. Therefore, dividing by the variance can amplify the weight of this influence.

[0157] Scale parameter contribution: Reflects the contribution of the rate of change of sample variance to the information. It is calculated by dividing the square of the rate of change of sample variance by the square of the sample variance, and then multiplying by a coefficient of 1 / 2. ;

[0158] The origin of the coefficient 1 / 2: In the Gaussian distribution, the Fisher information inherent coefficient corresponding to the scale parameter (variance) is 1 / 2, which is an inherent result of the statistical characteristics of this distribution; the reason for using the square of the variance in the denominator: the smaller the sample variance, the stronger the impact of the abrupt change in the degree of dispersion corresponding to the same rate of change of variance, so dividing by the square of the variance can amplify the weight of this impact.

[0159] Total angle Fisher information value: Add the above two items to obtain the angle Fisher information value for that angle. The formula is: .

[0160] In detail, the output order of the angular Fisher information angle sequence must be completely consistent with the order of the projection angles, that is, arranged according to the physical order of the effective projection angles (such as the increasing order of angles from 0 degrees to 360 degrees in the parallel beam projection model). The angular Fisher information value of each effective angle corresponds to its position in the order of projection angles.

[0161] In one embodiment of the present invention, feature values ​​of the angle sequence of angular Fisher information are extracted, a threshold is set based on the feature values, and projection angles whose angular Fisher information values ​​satisfy the threshold condition are summarized to form an unstable angle set, including:

[0162] The median of the angle sequence of the angle Fisher information is used as the threshold baseline;

[0163] In detail, the median is a robust statistic, insensitive to extreme values ​​in the sequence (such as the Fisher information value at a few very high angles), and stably reflects the average level of the sequence. Using it as a threshold baseline can avoid baseline shifts caused by extreme values, ensuring the objectivity and reliability of subsequent threshold settings, and laying the foundation for screening angles that are significantly higher than the average mutation level.

[0164] The absolute median difference of the angle sequence of the angle Fisher information is calculated as the eigenvalue;

[0165] In detail, the absolute median is also a robust statistic used to quantify the dispersion of an angular Fisher information sequence, that is, the range of fluctuation of values ​​in the sequence around the median. Compared to the standard deviation (which is susceptible to extreme values), the absolute median more accurately reflects the normal fluctuation level of the sequence. Using it as a feature value can ensure that subsequent thresholds can accurately match the actual fluctuation characteristics of the sequence, avoiding over-amplification or over-narrowing of the screening range.

[0166] Set threshold control parameters;

[0167] In detail, the threshold control parameter is a preset dimensionless constant, typically ranging from 1.0 to 2.0 (which can be adjusted according to the common influence range of metal artifacts in clinical imaging). It has no fixed unit and is only used to adjust the strictness of the threshold. By adjusting the threshold control parameter, the screening range of unstable angles can be flexibly controlled: the larger the parameter value, the higher the final angular threshold, the fewer unstable angles are screened, and only angles with extremely strong angular information mutations are retained; the smaller the parameter value, the lower the angular threshold, the more unstable angles are screened, and more angles with moderate mutation intensity can be covered, ensuring that the screening needs can be adapted to the actual metal artifact situation.

[0168] The angular threshold is obtained by multiplying the median and threshold control parameter by the sum of the eigenvalues.

[0169] In detail, the formula for calculating the angular threshold is: Angular Threshold = Median + Threshold Control Parameter × Absolute Median Difference. Here, the threshold control parameter × absolute median difference represents the allowable upper limit increment of fluctuation—representing the additional fluctuation range that can be judged as a significant mutation based on the average fluctuation level of the sequence; added to the median, it forms the critical value that distinguishes between normal mutation angles and significant mutation angles. The angular threshold is essentially a criterion for judging the intensity of angular information mutations: if the angular Fisher information value of a certain angle is not less than this threshold, it indicates that its mutation intensity is significantly higher than the average level and normal fluctuation range of the sequence, belonging to a high-risk mutation angle that requires close attention.

[0170] Summarize the projected angles in the angle sequence of Fisher information that are not less than the angle threshold, and arrange them in order of projection angle to form an unstable angle set.

[0171] In detail, the unstable angle set marks the projection angles that cause significant abrupt changes in information, i.e., these unstable angles are the regions that cause the efficiency of traditional incremental storage to collapse.

[0172] In one embodiment of the present invention, constructing incremental angle channel data using an unstable angle set includes:

[0173] The difference projection data map is obtained by subtracting the reference version projection data map from the projection data map of the version to be archived.

[0174] In detail, the process iterates through all angle-detector coordinate positions of the two versions of the projection data map. The projection value of the corresponding position in the version to be archived is subtracted from the projection value of the corresponding position in the reference version of the projection data map, yielding the projection difference for each position. All projection differences are then arranged according to the original angle-detector coordinate dimension to form a difference projection data map. The purpose of this operation is to quantify the differences between the two versions of projection data into differences. The difference projection data map centrally reflects the projection change of the version to be archived relative to the reference version, serving as the basis for subsequent incremental data extraction. In other words, the essence of incremental storage is to preserve version differences, not the entire dataset.

[0175] Filter the difference projection data map according to the unstable angle set in the projection angle dimension: set the position of the projection angle that is not in the unstable angle set to zero;

[0176] In detail, traverse all rows of the difference projection data plot in order of projection angle:

[0177] If the projection angle corresponding to the current row belongs to the set of unstable angles, then the projection difference of all detector coordinate positions in that row is kept unchanged.

[0178] If the projection angle corresponding to the current row does not belong to the unstable angle set, then the projection difference of all detector coordinate positions in that row is set to zero. The purpose of this operation is to extract the differences corresponding to unstable angles from the difference projection data map, retain only the angular abrupt differences with high compression difficulty, and set the low-difficulty differences corresponding to stable angles (which will be processed by the spatial channel later) to zero, thereby achieving difference focusing in the angular dimension.

[0179] The filtered results of the difference projection data map are cropped according to the metal trace mask in the detector coordinate dimension: the positions not marked by the metal trace mask are set to zero; thus, the incremental data of the angle channel is obtained.

[0180] In detail, iterate through all angles of the difference projection data map after filtering by the above angle dimensions to determine the detector coordinates:

[0181] If the detector coordinates corresponding to the current position are marked as one by the metal trace mask (belonging to the metal trace region), then the projection difference of that position is retained unchanged.

[0182] If the detector coordinates corresponding to the current position are not marked by the metal trace mask (non-metal trace region), the projection difference at that position is set to zero. After this clipping operation, the final two-dimensional data matrix is ​​the angle channel incremental data, which only contains the projection difference of unstable angles and metal trace regions. It focuses entirely on differences that are difficult to compress and prone to efficiency collapse. It can be stored in a dedicated angle channel to avoid the compression efficiency decrease caused by mixing with low-difficulty differences.

[0183] In detail, the angle channel incremental data is the difference carrier for incremental storage. Its data volume is only concentrated in the local area of ​​unstable angle + metal trace (far smaller than the full difference projection data map), and it is specifically stored for high-difficulty differences.

[0184] In one embodiment of the present invention, spatial channel incremental data is constructed using an unstable angle set, including:

[0185] An angle gating function is generated in the projection angle dimension based on the unstable angle set, and the projection angles belonging to the unstable angle set are marked as one, while the rest are marked as zero.

[0186] The spatial channel projection weights are constructed based on the angle gating function and the metal trace mask. When the projection angle is marked as one and the metal trace mask is marked as one, the weight is zero; the combined position of the other angle and detector coordinates is set to one.

[0187] The spatial channel projection data map is obtained by cropping the difference projection data map point by point using the spatial channel projection weight; the voxel center coordinates are determined by the voxel spacing and the voxel center is mapped to the detector coordinates by calculating the cosine and sine.

[0188] Within the detector step length range, a neighbor interpolation kernel is used to sample the spatial channel projection data map. The neighbor interpolation kernel is a function that takes a value of one within the detector step length half-width range and a value of zero outside the detector step length half-width range. The sampled values ​​of the spatial channel projection data map are multiplied by the detector step length and the angle step length according to the projection angle set and the detector coordinate set, and then discretely accumulated to obtain the spatial channel incremental data.

[0189] In detail, the angle gating function is a one-dimensional marking function with the same dimension as the projection angle set. Its generation depends entirely on the unstable angle set: it iterates through each angle in the projection angle set; if the angle belongs to the unstable angle set, the corresponding position of the function is marked as one; if the angle does not belong to the unstable angle set, it is marked as zero. For example, the projection angle set is... , , , The set of unstable angles is and Then the angle gating function is , respectively corresponding to The labeling results.

[0190] In detail, the angle gating function is one-dimensional (angle dimension only), while the metal trace mask is two-dimensional (angle × detector coordinates). The angle dimensions of the two are perfectly matched (both correspond to the set of projected angles). Two-dimensional spatial channel projection weights (dimension and difference projection data) can be constructed through dimension expansion and point-by-point logical operations. Figure 1 (All values ​​are angle × detector coordinates).

[0191] In detail, for each angle-detector coordinate combination position, the weight value is determined according to the following rules:

[0192] If the angle gating function at this position is marked as one (belonging to an unstable angle) and the metal trace mask is marked as one (belonging to a metal trace region), then the weight value is zero;

[0193] Except for the cases mentioned above, the weight value of all other angle-detector coordinate combination positions is taken as one.

[0194] In detail, the core function of the spatial channel projection weight is to delineate the processing boundaries of the two channels: the position with a weight of zero corresponds to the unstable angle + metal trace region. The difference in this region is handled by the angle channel, and the spatial channel needs to actively avoid it to avoid duplicate storage; the position with a weight of one corresponds to the stable angle region or the unstable angle but non-metal trace region. The difference compression in these regions is easy and they are the processing targets of the spatial channel.

[0195] In detail, the process iterates through all angle-detector coordinate positions of the difference projection data map (the result of subtracting the reference version from the archived version projection data map). For each position, the projection difference is multiplied point-by-point by the spatial channel projection weight: positions with a weight of one retain the original projection difference; positions with a weight of zero have their projection difference set to zero. The resulting two-dimensional data matrix is ​​the spatial channel projection data map, containing only the stable region differences that the spatial channel needs to process.

[0196] In detail, the voxel center coordinates are mapped to the detector coordinates:

[0197] Voxel center coordinates determination: Based on the voxel spacing (x-axis, y-axis, z-axis spacing) of the volume data to be archived, calculate the three-dimensional coordinates of the center of each voxel (e.g., the starting coordinate of a voxel on the x-axis is...). The voxel spacing is Then the center x-coordinate is (The same applies to the y-axis and z-axis).

[0198] Coordinate mapping logic: Through cosine and sine operations, the three-dimensional coordinates of the voxel center (especially the planar coordinates perpendicular to the ray propagation direction) are converted into detector coordinates in the parallel beam projection model. This mapping is based on the angle parameters of the projection geometry (ray propagation direction) to ensure that each voxel center can accurately correspond to a unique coordinate on the detector plane.

[0199] Detailed neighbor interpolation kernel sampling:

[0200] Definition of nearest neighbor interpolation kernel: The interpolation kernel is a function that matches the detector step size. It takes a value of one within the half-width range of the detector step size (e.g., the detector step size is 0.5 mm and the half-width is 0.25 mm), and a value of zero outside the range.

[0201] Sampling operation: Based on the detector coordinates of the voxel center mapping, within the detector step size range, the projection sampling value of the corresponding position is extracted from the spatial channel projection data map using the nearest neighbor interpolation kernel. That is, if the mapped coordinates fall within a certain detector step size interval, the projection value of that interval is directly taken as the sampling value to ensure that the sampling value can accurately reflect the difference of the corresponding projection area of ​​the voxel.

[0202] In detail, the process iterates through all voxels, and multiplies the projected sample value of each voxel by the detector step size (sampling interval in the detector direction) and the angle step size (sampling interval in the angle direction) according to the range of the projection angle set and the detector coordinate set, to obtain the weighted contribution value of that voxel. Then, the weighted contribution values ​​of all voxels are arranged and accumulated according to spatial coordinates (x-axis, y-axis, z-axis) to finally form three-dimensional spatial channel incremental data. That is, the channel incremental data is the difference data in the image domain, covering all differences in stable angular regions or unstable angular regions but non-metallic trace regions. It can be stored by conventional image domain incremental compression, complementing the angle channel incremental data to achieve complete incremental archiving.

[0203] In detail, the spatial channel incremental data focuses on stable regional differences with low compression difficulty, forming a non-overlapping and fully covered dual-channel architecture with the angular channel incremental data (focusing on core differences with high difficulty). This avoids the efficiency problems of traditional full-volume incremental data and ensures storage efficiency and data integrity by adapting to conventional compression algorithms through image domain storage. Together, they constitute the incremental data of medical images, realizing the efficient use of cloud storage resources.

[0204] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A cloud storage system for medical image archiving, characterized in that, include: The data acquisition unit reads reference version data and archived version data. The reference version data is historical reconstructed version voxel data, and the archived version data is newly reconstructed version voxel data. The data segmentation unit performs image segmentation on the archived version data and extracts the metal mask. The first data projection unit performs a projection transformation on the metal mask to obtain the metal trace mask corresponding to the metal mask. The second data projection unit performs projection transformations on the reference version body data and the version body data to be archived, respectively, to obtain the reference version projection data map and the version projection data map to be archived. The data processing unit calculates the angular Fisher information of the projected data map of the version to be archived at each angle in the order of the projection angle within the projection angle range covered by the metal trace mask, and obtains the angular Fisher information angle sequence. The unstable angle extraction unit extracts the feature values ​​of the angle sequence of the angle to Fisher information, sets a threshold based on the feature values, and summarizes the projection angles whose angle to Fisher information values ​​meet the threshold conditions to form an unstable angle set. The cloud storage unit constructs incremental data for angle channels and incremental data for spatial channels using unstable angle sets, and then processes the incremental data for cloud storage.

2. The cloud storage system for medical image archiving according to claim 1, characterized in that, Image segmentation is performed on the archived version data to extract the metal mask, including: Read the volume data and voxel spacing of the version to be archived, and use the voxel grid of the volume data to be archived as the processing grid. Perform strength truncation with the lower strength limit and the upper strength limit to obtain truncated volume data. Perform binarization with the metal strength threshold to obtain the initial binary volume data. The initial binary volume data is subjected to a three-dimensional 26-neighbor connected component screening, and only connected regions with a voxel count not less than the minimum voxel count threshold are retained; wherein, the minimum voxel count threshold is determined by converting the minimum physical volume parameter and the voxel spacing. A three-dimensional spherical structuring element is used to perform the closing operation; the radius of the structuring element for the closing operation is determined by converting the physical radius of the closing operation with the voxel spacing. Expansion is performed using a three-dimensional spherical structural element; the radius of the expanding structural element is determined by converting the safe expansion physical radius and the voxel spacing. The metal mask is obtained based on the closing operation and expansion.

3. The cloud storage system for medical image archiving according to claim 2, characterized in that, Performing a projection transformation on the metal mask yields the corresponding metal trace mask, including: Define the projection geometry of the parallel beam projection model. The projection geometry includes: angle set, angle step size, detector coordinate set, and detector step size. Length-weighted projection is performed on the metal mask to calculate the total geometric length of each ray passing through the metal mask voxels, resulting in a length sine curve. Each ray corresponds to an angle in the projection geometry and a detector coordinate. The total geometric length is the sum of the geometric passage lengths within each metal voxel through which the ray passes. Calculate the minimum intersection length threshold, which is the safety factor multiplied by the minimum voxel spacing. The minimum voxel spacing is the minimum of the three voxel spacings, and the safety factor is a dimensionless constant not less than one. Binarize the length sine curve with the minimum intersection length threshold, mark the positions in the length sine curve that are not less than the threshold as one, and mark the rest as zero, to obtain the original metal trace mask. The original metal trace mask is subjected to a first expansion along the detector direction. The first expansion adopts a one-dimensional symmetric structure element. The number of sampling points of the one-dimensional symmetric structure element is obtained by dividing the physical radius of the detector direction safe expansion by the detector step size and rounding up, thus obtaining a detector direction coherent mask. A second expansion is performed on the detector directional coherent mask along the angular direction. The second expansion uses a one-dimensional symmetric structural element. The number of sampling points of the one-dimensional symmetric structural element is obtained by dividing the angular expansion width by the angular step size and rounding up, thus obtaining the metal trace mask.

4. A cloud storage system for medical image archiving according to claim 3, characterized in that, Perform projection transformations on the reference version body data and the version body data to be archived, respectively, to obtain the reference version projection data map and the version projection data map to be archived, including: Under the parallel beam projection model, determine all ray paths corresponding to the angle set and the detector coordinate set; Determine the geometrical passage length between the ray and the voxel, and use this geometrical passage length as the weight of the voxel in the corresponding ray accumulation; Set the unit conversion factor used to convert volumetric data grayscale to projected cumulative measurement units; The reference version projection data map is obtained by calculating the length-weighted projection value of the reference version body data. The length-weighted projection value of the body data to be archived is calculated to obtain the projected data of the version to be archived.

5. A cloud storage system for medical image archiving according to claim 4, characterized in that, Within the projection angle range covered by the metal trace mask, the angular Fisher information of the projected data map of the version to be archived is calculated at each angle according to the projection angle sequence, resulting in the angular Fisher information angle sequence, including: Determine the effective projection angle set; where the effective projection angle set is the set of angles marked as one on the projection angle of the metal trace mask; Based on the effective projection angle, extract the projection samples corresponding to the detector coordinates marked as one on the metal trace mask from the projection data map of the version to be archived, and calculate the sample mean and sample variance for each angle. The rate of change of the sample mean and sample variance with respect to the angle is estimated according to the central difference rule in the order of projection angles, thus obtaining the rate of change of the sample mean and the rate of change of the sample variance; wherein, the first and last order of projection angles adopt the one-sided difference rule. In an approximate form where the sample mean is used as a location parameter and the sample variance is used as a scale parameter, the rate of change of the sample mean and the rate of change of the sample variance are combined to form the angular Fisher information value. Output the angle sequence of Fisher information according to the projection angle order.

6. A cloud storage system for medical image archiving according to claim 5, characterized in that, Feature values ​​are extracted from the angle sequence of angular Fisher information. A threshold is set based on these feature values. Projected angles whose angular Fisher information values ​​satisfy the threshold condition are then summarized to form a set of unstable angles, including: The median of the angle sequence of the angle Fisher information is used as the threshold baseline; The absolute median difference of the angle sequence of the angle Fisher information is calculated as the eigenvalue; Set threshold control parameters; The angular threshold is obtained by multiplying the median and threshold control parameter by the sum of the eigenvalues. Summarize the projected angles in the angle sequence of Fisher information that are not less than the angle threshold, and arrange them in order of projection angle to form an unstable angle set.

7. A cloud storage system for medical image archiving according to claim 6, characterized in that, Incremental angle channel data is constructed using a set of unstable angles, including: The difference projection data map is obtained by subtracting the reference version projection data map from the projection data map of the version to be archived. Filter the difference projection data map according to the unstable angle set in the projection angle dimension: set the position of the projection angle that is not in the unstable angle set to zero; The filtered results of the difference projection data map are cropped according to the metal trace mask in the detector coordinate dimension: the positions not marked by the metal trace mask are set to zero; thus, the incremental data of the angle channel is obtained.

8. A cloud storage system for medical image archiving according to claim 7, characterized in that, Spatial channel incremental data is constructed using an unstable angle set, including: An angle gating function is generated in the projection angle dimension based on the unstable angle set, and the projection angles belonging to the unstable angle set are marked as one, while the rest are marked as zero. The spatial channel projection weights are constructed based on the angle gating function and the metal trace mask. When the projection angle is marked as one and the metal trace mask is marked as one, the weight is zero; the combined position of the other angle and detector coordinates is set to one. The spatial channel projection data map is obtained by cropping the difference projection data map point by point using the spatial channel projection weight; the voxel center coordinates are determined by the voxel spacing and the voxel center is mapped to the detector coordinates by calculating the cosine and sine. Within the detector step length range, a neighbor interpolation kernel is used to sample the spatial channel projection data map. The neighbor interpolation kernel is a function that takes a value of one within the detector step length half-width range and a value of zero outside the detector step length half-width range. The sampled values ​​of the spatial channel projection data map are multiplied by the detector step length and the angle step length according to the projection angle set and the detector coordinate set, and then discretely accumulated to obtain the spatial channel incremental data.

Images