A remote digitized image analysis collaboration system

Abstract

This invention belongs to the field of image processing technology and discloses a remote digital image analysis collaborative system. It includes: a virtual diagnosis module: acquiring continuous pathological slide sequences and performing spatial registration to obtain registered slides; a model building module: performing three-dimensional voxel reconstruction on the registered slides to obtain a three-dimensional voxel reconstruction model, followed by semantic enhancement to obtain a three-dimensional pathological voxel model; a collaborative interaction module: generating G virtual avatars of experts based on VR technology, annotating the three-dimensional pathological voxel model, and obtaining spatial anchor point annotation data for the experts; a graph-driven module: constructing a structured semantic tree based on the spatial anchor point annotation data, detecting annotation conflicts, and obtaining annotation conflict detection results; and a conflict resolution module: resolving conflicts in the annotation conflict detection results to obtain consensus suggestions. This invention improves the collaborative efficiency and reliability of conclusions in remote diagnosis, providing a systematic solution for the accurate diagnosis of complex pathological cases.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a remote digital image analysis collaborative system. Background Technology

[0002] With the rapid development of telemedicine and digital pathology, pathological diagnosis, as the gold standard for disease diagnosis, increasingly demands multi-expert collaboration. In traditional pathological diagnosis models, experts often rely on independent annotation based on local slide observation.

[0003] Chinese patent application CN116363314A discloses a labeling method, apparatus, electronic device, and computer-readable storage medium: Preprocessing medical image data to obtain preprocessed data corresponding to the medical image data, the preprocessed data including 3D model data and multiple segmented regions; performing 3D reconstruction based on the 3D model data to obtain a 3D model corresponding to the medical image data; displaying the 3D model; and, in response to receiving labeling information, performing visual processing on the 3D model based on the segmented regions corresponding to the labeling information; and storing the labeling information and the visually processed 3D model. This invention performs visual processing on the 3D model corresponding to medical image data in response to labeling information, helping users to easily and intuitively label 3D models without training, reducing the time cost for users to label 3D models, improving the user experience of products using the labeling method provided by this invention, reducing product development and maintenance costs, and enhancing product competitiveness.

[0004] While the above methods can meet the needs of most scenarios, research and practical application of these methods and existing technologies have revealed at least the following shortcomings:

[0005] The above method can only be applied to independent annotation behavior on a single user terminal and cannot resolve conflicts in the interpretation of lesion points by different experts on multiple user terminals.

[0006] In view of this, the present invention proposes a remote digital image analysis collaborative system to solve the above problems. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a remote digital image analysis collaborative system, comprising:

[0008] Virtual diagnostic module: Acquires continuous pathological slide sequences and performs spatial registration to obtain registered slides;

[0009] Model building module: Perform three-dimensional voxel reconstruction on the registered slices to obtain a three-dimensional voxel reconstruction model, and then perform semantic enhancement to obtain a three-dimensional pathological voxel model;

[0010] Collaborative Interaction Module: Based on VR technology, generate virtual avatars of G experts, annotate the three-dimensional pathological voxel model, and obtain spatial anchor point annotation data of the experts;

[0011] Graph-driven module: Constructs a structured semantic tree based on spatial anchor point annotation data, detects annotation conflicts, and obtains annotation conflict detection results;

[0012] Conflict resolution module: Resolves conflicts in the labeled conflict detection results and obtains consensus suggestions.

[0013] Furthermore, methods for obtaining registration slices include:

[0014] Gaussian denoising and grayscale normalization were performed on each pathological slide to obtain a standardized slide sequence;

[0015] The first slice of the standardized slice sequence Using the reference coordinate system, the origin is... The top-left pixel is defined by the x-axis along the width direction, the y-axis along the height direction, and the z-axis along the slice sequence direction; global coordinate system parameters are obtained, including the origin, axis directions, and the initial z-coordinate of each slice.

[0016] For each slice, the anatomical structure region is extracted by threshold segmentation. The binarized image after threshold segmentation is traversed, and the arithmetic mean of the coordinates of all pixels in the binarized image is calculated to obtain the feature points.

[0017] For the i-th and (i+1)-th slices, for each feature point, extract the neighborhood of C×C pixels and count the gray-level pairs within the neighborhood. The probability of its occurrence, and All values ​​are grayscale values. A grayscale matrix is ​​generated. The elements of the grayscale matrix are expanded into a one-dimensional vector by rows / columns to obtain the local descriptors of the feature points. The Euclidean distance between the local descriptors of the i-th slice and the (i+1)-th slice is calculated. Similar feature points with a distance less than the distance threshold are filtered to obtain a set of matching point pairs between adjacent slices.

[0018] Let the point on the (i+1)th slice of the matching point pair set correspond to the point on the ith slice after transformation; minimize the transformation error of the matching point pair using the least squares method; obtain the rigid transformation parameters of adjacent slices, and perform rigid transformation in combination with the rigid transformation parameters, that is, perform rigid transformation in combination with the rigid transformation parameters and transformation formula to obtain coarsely aligned slices.

[0019] Construct an m×n control point grid on the coarse-aligned slice plane, with control point coordinates as follows: , The grid spacing is m; m and n are the number of grid rows and columns, respectively. and The number of grid rows and columns corresponding to the control points; the deformation displacement of any point is obtained by interpolating the control point displacement through the B-spline basis function, and the elastic transformation parameters are obtained; among them, the control point displacement is optimized by using the gradient descent method to obtain the minimum target value with the goal of maximizing gray-level similarity and deformation smoothness, and the elastically registered slice is obtained.

[0020] Based on the rigid transformation parameters and elastic transformation parameters, calculate the total transformation of the i-th slice relative to the first slice; obtain the globally optimized set of transformation parameters; and apply this transformation to any non-adjacent slice. and The deviation between the direct matching transform and the cumulative transform is calculated, and the total global deviation is minimized by adjusting the transform parameters of the intermediate slices; thus, the registration slice sequence is obtained.

[0021] Furthermore, methods for obtaining three-dimensional voxel reconstruction models include:

[0022] The boundaries of the 3D voxel model are determined based on the pixel size d and the slice thickness t. The x-axis boundary is [0, W×d], the y-axis boundary is [0, H×d], and the z-axis boundary is [0, (N-1)×t]; where W is the number of pixels in the slice width; H is the number of pixels in the slice height; and N is the number of registered slices in the registered slice sequence.

[0023] Obtain voxel resolution of d in both the x-axis and y-axis directions, and voxel resolution of t / k in the z-axis direction, where k is the B-spline basis function interpolation factor;

[0024] The registered slice sequence is normalized to obtain the corrected slice sequence;

[0025] Map the pixels of each calibration slice to three-dimensional voxel coordinates;

[0026] Using the local regions containing Q×Q pixel blocks from two adjacent correction slices as input to the depth completion model, the local regions are obtained in... and The k-1 layer interpolated voxel values ​​between; the mapped voxels and interpolated voxel values ​​of all correction slices are stitched together in z-axis order to form a continuous three-dimensional voxel reconstruction model and a completed slice; among which, and The first Zhang Hedi The z-axis coordinate of the correction slice in the three-dimensional voxel coordinate system.

[0027] Furthermore, methods for obtaining three-dimensional pathological voxel models include:

[0028] Obtain expert annotation data and pixel coordinate markers, and map the pixel coordinates of the annotation area on the completed slice to three-dimensional voxel coordinates to form a set of annotation points in three-dimensional space;

[0029] Voxel space is divided according to resolution for tumor regions in two-dimensional slices. and First, combine the seed point set converted to voxel space by resolution. and traversal and If the voxel gray value is within the gray range of the tumor region and is connected to the existing seed point set in three-dimensional space, the corresponding voxel will be included in the three-dimensional tumor region to obtain preliminary three-dimensional semantic annotation.

[0030] Extract all voxels in the 3D neighborhood of each target voxel in the voxel space, calculate the voxel mean, variance, and gradients in the x, y, and z directions of the voxels in the neighborhood, then calculate the gradient magnitude, take the maximum gradient magnitude to obtain the gradient feature, and concatenate the mean, variance, and gradient feature to obtain the grayscale feature; calculate the shortest Euclidean distance from the target voxel to the known semantic region to obtain the morphological feature.

[0031] By using preliminary 3D semantic annotation, grayscale features, and morphological features as inputs to the label prediction model, a 3D pathological voxel model with semantic labels is obtained.

[0032] Furthermore, methods for obtaining spatial anchor point annotation data from experts include:

[0033] Virtual avatars of various experts are generated based on VR technology, and the position coordinates and viewing direction of the virtual avatars are synchronized in real time based on the spatial coordinate system of the three-dimensional pathological voxel model.

[0034] Experts annotate the three-dimensional pathological voxel model, and the system automatically generates spatial anchor points containing anchor point attributes, including annotation coordinates, annotation type, annotation expert, and annotation timestamp.

[0035] Using dynamic rendering of the view frustum, only the voxel data within the current expert's field of view is rendered to obtain the expert's spatial anchor point annotation data.

[0036] Furthermore, methods for obtaining the expert's spatial anchor point annotation data by employing dynamic frustum rendering, rendering only voxel data within the current expert's field of view, include:

[0037] Obtain the position coordinates, view direction vector, view angle, near clipping plane distance, and far clipping plane distance, and construct the view frustum;

[0038] The coordinate range of the view cone in the three-dimensional pathological voxel model is obtained by taking the intersection of the coordinate range of the view cone and the coordinate range of the three-dimensional pathological voxel model.

[0039] Traverse the voxels in the 3D pathological voxel model whose coordinates fall within the cone coordinate range. For each voxel that has been initially screened, determine whether it is located inside the cone: calculate the distance from the voxel center to the 6 boundary planes of the cone; if the distance from the voxel to all planes is within the plane, it is determined to be a voxel inside the cone; obtain the set of voxels inside the cone by statistics.

[0040] Calculate the Euclidean distance from each voxel in the voxel set to the expert coordinates, and set the N-level rendering detail according to the preset rendering rules and the Euclidean distance; obtain the voxels inside the view frustum with detail level markers;

[0041] Voxels are grouped and processed according to their level of detail. The voxels are projected onto the screen coordinate system. When the expert's viewpoint changes, a new view frustum is obtained and the level of detail is reassigned based on the new view frustum.

[0042] Furthermore, the viewing cone includes a near clipping plane, a far clipping plane, a left side, a right side, an upper side, and a lower side; wherein, the near clipping plane and the far clipping plane are both perpendicular to the line of sight, and are respectively a near clipping distance and a far clipping distance from the observation point; the angle between the left side and the right side and the line of sight is half of the horizontal field of vision angle, defining the horizontal visible range; the angle between the upper side and the lower side and the line of sight is half of the vertical field of vision angle, defining the vertical visible range.

[0043] Furthermore, methods for detecting annotation conflicts include:

[0044] Obtain spatial anchor point annotation data for each expert. Each anchor point includes three-dimensional coordinates, annotation text, expert ID, and annotation time.

[0045] Calculate the spatial distance between anchor points. If the spatial distance between two anchor points is not greater than U voxels, they are determined to be labeled in the same region; otherwise, they are determined to be labeled in different regions.

[0046] Obtain the labeled groups for the same region, and decompose each labeled text into hierarchical terms in the preset ontology library;

[0047] Using the case ID as the root node, a tree structure is constructed according to the ontology hierarchy to obtain a structured semantic tree;

[0048] Define conflict types, including term hierarchy conflict, attribute contradiction conflict, and logical inclusion conflict. Term hierarchy conflict means that terms labeled in the same area are on mutually exclusive branches of the structured semantic tree. Attribute contradiction conflict means that the core attributes of terms are opposite. Logical inclusion conflict means that one term completely includes another term but the labeling is inconsistent.

[0049] Based on the relationships between nodes in a structured semantic tree, conflict detection rules are defined.

[0050] Conflict detection is performed based on structured semantic trees to obtain labeled conflict detection results.

[0051] Furthermore, the tree structure includes first to fourth level nodes, where the first level node is the organ, the second level node is the lesion type, the third level node is the subtype, and the fourth level node is the grade / feature; each node is associated with the corresponding anchor coordinates and expert ID.

[0052] Furthermore, methods for obtaining consensus recommendations include:

[0053] The conflict detection results are broken down into conflict terminology pairs, pathological morphological parameters of conflict areas, expert subspecialties, and historical diagnostic accuracy.

[0054] The binarized image after threshold segmentation is traversed to extract the lesion contour. The length of the long side of the minimum bounding rectangle of the contour is calculated as the lesion diameter. HE-stained sections are used as input to the U-Net segmentation model to obtain the mitotic count. Ki-67 index and ER / PR expression status are obtained. The lesion diameter, mitotic count, Ki-67 index and ER / PR expression status are spliced ​​together to obtain pathological features.

[0055] The expert weight is obtained by weighting the expert’s historical diagnostic accuracy rate, the proportion of the expert’s diagnostic frequency in its subspecialty, and the expert’s historical consensus rate.

[0056] The guide recommendation level is obtained according to the predefined recommendation level mapping rules. For each conflicting term, its text similarity with the preset guide standard is calculated. The matching degree is obtained by combining the guide recommendation level.

[0057] The pathological features, expert weights, and matching degree are used as inputs to a weighted fusion graph neural network to obtain the consensus probability of each term. The term with the highest probability is extracted as the consensus conclusion. In the weighted fusion graph neural network, each node represents a conflicting term, and the node features are the pathological features and matching degree corresponding to the conflicting term. The edge weight between two term nodes is obtained by weighting the difference in expert weights and the similarity of pathological features calculated by cosine similarity. When there are 2 experts, the weight difference between experts is calculated to obtain the expert weight difference. When there are more than 2 experts, the expert weight of each expert is calculated separately. Then, for each conflicting term, the statistical value of the individual weights of all experts supporting the conflicting term is taken as the representative weight of the conflicting term. The absolute value of the difference between the representative weights of the two conflicting terms is calculated to obtain the expert weight difference between the two term nodes.

[0058] The technical effects and advantages of the remote digital image analysis collaborative system of the present invention are as follows:

[0059] This invention utilizes a virtual diagnostic module to perform Gaussian denoising, grayscale normalization, and spatial registration on continuous pathological slide sequences, constructing registered slides in a unified coordinate system. A model building module combines pixel size and slide thickness to complete 3D voxel reconstruction and assign semantic labels, forming a 3D pathological voxel model that provides a precise spatial reference for multi-expert collaboration. A collaborative interaction module generates virtual avatars based on VR technology, enabling simultaneous annotation by multiple experts on the 3D model. Dynamic frustum rendering technology efficiently acquires spatial anchor point data containing spatial coordinates and annotation types, overcoming the limitations of independent annotation by a single user. A map-driven module divides the same region into annotation groups by calculating the spatial distance between anchor points. By mapping annotated text to ontology-level terms and constructing a structured semantic tree, three conflict types—terminology hierarchy, attribute contradiction, and logical inclusion—are defined, along with detection rules. This approach accurately identifies disagreements among multiple experts. Furthermore, by extracting pathological features, calculating expert weights and guideline matching degrees, and utilizing a weighted fusion graph neural network, the solution generates the conclusion with the highest consensus probability, effectively resolving conflicts. The overall solution enables remote, immersive collaborative annotation by multiple experts. Through a unified three-dimensional semantic framework and a structured conflict resolution mechanism, it overcomes the problem of traditional methods being unable to handle disagreements in lesion interpretation among multiple experts, improving the collaborative efficiency and reliability of remote diagnosis. This provides a systematic solution for the accurate diagnosis of complex pathological cases. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the structure of a remote digital image analysis collaborative system according to the present invention;

[0061] Figure 2 This is a schematic diagram of the data flow in this invention;

[0062] Figure 3 This is a schematic diagram of Embodiment 2 of the present invention;

[0063] Figure 4 This is a schematic diagram of the data flow in Embodiment 2 of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1

[0066] Please see Figure 1 , Figure 2 As shown, this embodiment provides a remote digital image analysis collaborative system, including:

[0067] Virtual diagnostic module: Acquires continuous pathological slide sequences and performs spatial registration to obtain registered slides;

[0068] Methods for obtaining registration slices include:

[0069] Gaussian denoising and gray-level normalization are performed on each pathological slide to obtain a standardized slide sequence; for example, Gaussian denoising is performed based on Gaussian filtering with a kernel size of L×L, and then the gray-level values ​​of the slides are linearly stretched to the range of [0,255] to obtain normalized gray-level values.

[0070] The first slice of the standardized slice sequence Using the reference coordinate system, the origin is... The top-left pixel is defined by the x-axis along the width direction, the y-axis along the height direction, and the z-axis along the slice sequence direction; global coordinate system parameters are obtained, including the origin, axis directions, and the initial z-coordinate of each slice.

[0071] For each slice, the anatomical structure region is extracted by threshold segmentation. The threshold T is obtained by optimization calculation using the Otsu algorithm combined with the natural heuristic optimization algorithm. The binarized image after threshold segmentation is traversed, and the arithmetic mean of the coordinates of all pixels in the binarized image is calculated to obtain the feature points.

[0072] For the i-th and (i+1)-th slices, for each feature point, extract the neighborhood of C×C pixels and count the gray-level pairs within the neighborhood. The probability of occurrence is used to generate a grayscale matrix. The elements of the grayscale matrix are expanded into a one-dimensional vector by rows / columns, such as an 8×8 matrix being expanded into a 64-dimensional vector, which serves as the local descriptor of the feature point. The Euclidean distance between the local descriptors of the i-th slice and the (i+1)-th slice is calculated. Similar feature points with a distance less than the distance threshold are filtered out to obtain a set of matching point pairs between adjacent slices.

[0073] Let the points on the (i+1)th slice of the matching point pair set be... The point corresponding to the i-th slice after transformation , such as transformation formula ;in, The rotation angle; and These represent the translations along the x-axis and y-axis, respectively; the transformation error of the matching point pairs is minimized using the least squares method, such as... ,in, For the first The first slice One feature point, for The coordinates; For the first Zhang slice and the first Pairs of matching feature points on a slice; For the first The first slice One feature point, for The coordinates; obtain the rigid transformation parameters of adjacent slices. The rigid transformation is performed by combining the rigid transformation parameters, that is, by combining the rigid transformation parameters with the transformation formula to perform the rigid transformation, and a coarse alignment slice is obtained.

[0074] Construct an m×n control point grid on the coarse-aligned slice plane, with control point coordinates as follows: , The grid spacing is determined by the displacement of the control points. Any point can be obtained by interpolation using B-spline basis functions. Deformation displacement , For the first in the grid line, number Column control points at Displacement in the direction; For the first in the grid line, number Column control points at Displacement in direction; such as , ;in, and The basis functions are cubic B-spline functions, specifically:

[0075] ; ; Obtain elastic transformation parameters Among them, the control point displacement aims to maximize gray-level similarity and deformation smoothness, such as the target... ,in, Mutual information is used to measure the grayscale similarity between coarsely aligned slices and slices after rigid transformation; This is the sum of squares of the second derivatives of the displacements at the control points; The deformation smoothing constraint factor is used; the minimum objective value is obtained by iterative optimization using the gradient descent method, resulting in elastically registered slices. ;

[0076] Based on the rigid transformation parameters and the elastic transformation parameters, calculate the total transformation of the i-th slice relative to the first slice, such as... ,in, For transformation composition; obtain the globally optimized set of transformation parameters; for any non-adjacent slices and Calculate the transformation of its direct match. With cumulative transformation The deviation is minimized by adjusting the transformation parameters of the intermediate slices, such as... ; Obtain the registration slice sequence.

[0077] The virtual diagnostic module eliminates data interference by performing Gaussian denoising and grayscale normalization on continuous pathological slide sequences, establishes a unified global coordinate system to ensure spatial consistency, extracts anatomical feature points through threshold segmentation and calculates local descriptors to achieve precise matching of adjacent slides, and combines rigid and elastic transformations to eliminate displacement, rotation, and nonlinear deformation between slides. Finally, it obtains the registered slide sequence and constructs a three-dimensional pathological voxel model. This process provides a unified and accurate three-dimensional spatial reference for annotation by different experts on multiple user terminals, enabling each expert's annotation of lesion points to be accurately located and associated in the same coordinate system. This avoids false discrepancies caused by spatial misalignment of slides and ensures that expert interpretation discrepancies are based on differences in real pathological features. This provides a reliable spatial benchmark and data foundation for subsequent conflict detection and conflict resolution for multi-expert annotation discrepancies, effectively solving the technical problem that existing methods can only support independent annotation on a single user terminal and cannot handle discrepancies in lesion point interpretations by multiple experts.

[0078] Model building module: Perform three-dimensional voxel reconstruction on the registered slices to obtain a three-dimensional voxel reconstruction model, and then perform semantic enhancement to obtain a three-dimensional pathological voxel model;

[0079] Methods for obtaining three-dimensional voxel reconstruction models include:

[0080] The boundaries of the 3D voxel model are determined based on the pixel size d and the slice thickness t. The x-axis boundary is [0, W×d], the y-axis boundary is [0, H×d], and the z-axis boundary is [0, (N-1)×t]; where W is the number of pixels in the slice width; H is the number of pixels in the slice height; and N is the number of registered slices in the registered slice sequence.

[0081] Obtain voxel resolution of d in both the x-axis and y-axis directions, and voxel resolution of t / k in the z-axis direction, where k is the B-spline basis function interpolation factor;

[0082] The registered slice sequence is normalized to obtain the corrected slice sequence;

[0083] Pixels of each correction slice Mapping to 3D voxel coordinates ,like ;in, The physical dimension of the voxel in the x-direction; The physical dimension of the voxel in the y-direction;

[0084] Using the local regions containing Q×Q pixel blocks from two adjacent correction slices as input to the depth completion model, the local regions are obtained in... and The k-1 layer interpolation voxel values ​​between; the mapping voxels and interpolation voxel values ​​of all correction slices are spliced ​​together in z-axis order to form a continuous three-dimensional voxel model and a complete slice.

[0085] The method for obtaining a 3D voxel reconstruction model determines the boundary and resolution of the 3D voxel model by determining the pixel size and slice thickness. It normalizes the registered slice sequence and maps pixels to 3D voxel coordinates. Then, it uses a depth completion model to interpolate and complete the blank areas between adjacent slices, ultimately forming a continuous 3D voxel model and completed slices. This process constructs a complete and continuous 3D spatial structure, providing a unified and coherent lesion observation benchmark for different experts on multiple user terminals. This allows experts to interpret lesion points based on the same 3D spatial context, avoiding misunderstandings caused by information breaks between slices. It ensures that divergent interpretations are directed at lesion features in the same 3D space, thus laying a coherent spatial data foundation for accurate identification and resolution of annotation conflicts among multiple experts. This effectively supports conflict resolution for multi-expert divergent interpretations and solves the technical problem that existing methods can only support independent annotation by a single user and cannot handle multi-expert disagreements.

[0086] Methods for obtaining three-dimensional pathological voxel models include:

[0087] Obtain expert annotation data and pixel coordinate markers, and map the pixel coordinates of the annotation area on the completed slice to three-dimensional voxel coordinates to form a set of annotation points in three-dimensional space;

[0088] Tumor region in two-dimensional slices and First, convert it into a seed point set in voxel space. and traversal and If the voxel gray value is within the gray range of the tumor region and is connected to the existing seed point set in three-dimensional space, the corresponding voxel will be included in the three-dimensional tumor region to obtain preliminary three-dimensional semantic annotation.

[0089] Extract all voxels in the 3D neighborhood of each target voxel, calculate the voxel mean, variance, and gradients in the x, y, and z directions of the voxels in the neighborhood, then calculate the gradient magnitude, take the maximum gradient magnitude to obtain the gradient feature, and concatenate the mean, variance, and gradient feature to obtain the grayscale feature; calculate the shortest Euclidean distance from the target voxel to the known semantic region to obtain the morphological feature.

[0090] By using preliminary 3D semantic annotation, grayscale features, and morphological features as inputs to the label prediction model, a 3D pathological voxel model with semantic labels is obtained.

[0091] The method for obtaining a three-dimensional pathological voxel model maps the two-dimensional pixel coordinates annotated by experts to three-dimensional voxel coordinates, constructs preliminary semantic annotations by combining the grayscale range of the tumor region and three-dimensional connectivity, and extracts the grayscale and morphological features of the voxels. After processing by a label prediction model, a three-dimensional pathological voxel model with semantic labels is obtained. This process provides a unified three-dimensional semantic framework for multi-expert annotation, enabling different experts' annotations of lesions to not only correspond accurately in space, but also to achieve structured association through semantic labels. This avoids misunderstandings caused by differences in semantic descriptions and ensures that disagreements among multiple experts on lesion interpretation can be accurately transformed into semantic conflicts. This provides a structured semantic data foundation for the subsequent detection and resolution of these conflicts, effectively solving the technical problem that existing methods only support independent annotation by a single user and cannot handle disagreements among multiple experts on lesion interpretation.

[0092] Collaborative Interaction Module: Based on VR technology, generate virtual avatars of G experts, annotate the three-dimensional pathological voxel model, and obtain spatial anchor point annotation data of the experts;

[0093] Methods for obtaining spatial anchor point annotation data from experts include:

[0094] Virtual avatars of various experts are generated based on VR technology, and the position coordinates and viewing direction of the virtual avatars are synchronized in real time based on the spatial coordinate system of the three-dimensional pathological voxel model.

[0095] Experts annotate the three-dimensional pathological voxel model, and the system automatically generates spatial anchor points containing anchor point attributes, including annotation coordinates, annotation type, annotation expert, and annotation timestamp.

[0096] Using the dynamic rendering technology of view frustum, only the voxel data within the current expert's field of view is rendered to obtain the expert's spatial anchor point annotation data.

[0097] Methods for obtaining the expert's spatial anchor point annotation data by employing dynamic frustum rendering technology, which renders only voxel data within the current expert's field of view, include:

[0098] The system acquires the position coordinates, gaze direction vector, field of view angle, near clipping plane distance, and far clipping plane distance. The gaze direction vector is provided by the head tracking device. The horizontal field of view angle defaults to 60°, and the vertical field of view angle defaults to 45°. The near clipping plane distance is the closest visible distance from the observation point, typically 0.1m, and the far clipping plane distance is the farthest visible distance from the observation point, typically 10m. A visual cone is constructed, including the near clipping plane, far clipping plane, left side, right side, top side, and bottom side. The near and far clipping planes are perpendicular to the gaze direction and are the near and far clipping distances from the observation point, respectively. The angles between the left and right sides and the gaze direction are half the horizontal field of view angle, limiting the horizontal visible range. The angles between the top and bottom sides and the gaze direction are half the vertical field of view angle, limiting the vertical visible range.

[0099] The coordinate range of the view cone in the three-dimensional pathological voxel model is obtained by taking the intersection of the coordinate range of the view cone and the coordinate range of the three-dimensional pathological voxel model.

[0100] Traverse the voxels in the 3D pathological voxel model whose coordinates fall within the cone coordinate range. For each initially selected voxel, determine whether it is located inside the cone: calculate the distance from the voxel center to the 6 boundary planes of the cone; if the distance from the voxel to all planes is within the plane (e.g., in front of the near clipping plane, behind the far clipping plane, between the left and right sides, etc.), then it is determined to be a voxel inside the cone; obtain the set of voxels inside the cone by statistics.

[0101] Calculate the Euclidean distance from each voxel in the voxel set to the expert coordinates, and set an N-level rendering detail based on the preset rendering rules and the Euclidean distance; obtain the voxels inside the view frustum with detail level markers; if N=3, this includes:

[0102] Near-range voxels (e.g., distance ≤ 2 meters): Preserve complete details, render each voxel separately, and retain original attributes (e.g., fine grayscale, specific semantic tags).

[0103] Mid-range voxels (e.g., 2-5 meters): Merge adjacent voxels (e.g., 2×2×2 voxels are merged into 1 rendering unit) and take the average attribute value;

[0104] For distant voxels (e.g., >5 meters): significantly merge voxels (e.g., merge 4×4×4 voxels into 1 rendering unit) and simplify attributes (e.g., retain only major category tags such as "organization" and "background").

[0105] Voxels are grouped and processed according to their level of detail. For example, near voxels use their original attributes directly, while mid / far voxels are generated into simplified rendering units according to merging rules and assigned corresponding colors.

[0106] Voxels are projected onto the screen coordinate system using a graphics rendering pipeline (such as WebGL);

[0107] When the expert's perspective changes, such as turning the head or moving the position, the view frustum is rebuilt, the new view frustum is calculated, voxels are selected and level of detail is assigned; only newly visible voxels caused by the change in perspective and voxels that have left the field of view are updated, while the rendering results of unchanged areas are retained to reduce redundant calculations.

[0108] The collaborative interaction module generates expert virtual avatars based on VR technology. By synchronizing positions and perspectives in real time, it ensures that multiple experts annotate pathological voxel models in a unified 3D coordinate system. The system generates spatial anchor point data containing coordinates, type, expert, and timestamp. Simultaneously, using dynamic cone rendering technology, it constructs cones based on expert perspective parameters and filters internal voxels. Multi-level rendering details are set according to distance to efficiently present the visible area. When the perspective changes, only the changing voxels are updated. This process provides an immersive and synchronized collaborative annotation environment for multiple experts, enabling annotation operations by different experts to be based on the same 3D spatial reference. The annotation data contains clear spatial coordinates and expert information, ensuring both the accuracy and efficiency of annotation. It also provides structured and correlated raw data for subsequent comparison of different experts' interpretations of lesion points and identification of discrepancies. This lays the interactive and data foundation for solving the technical problem that existing methods only support independent annotation by a single user and cannot handle discrepancies in interpretations by multiple experts.

[0109] Graph-driven module: Constructs a structured semantic tree based on spatial anchor point annotation data, detects annotation conflicts, and obtains annotation conflict detection results;

[0110] Methods for detecting annotation conflicts include:

[0111] Obtain spatial anchor point annotation data for each expert. Each anchor point includes three-dimensional coordinates, annotation text, expert ID, and annotation time.

[0112] Calculate the spatial distance between anchor points. If the spatial distance between two anchor points is not greater than U voxels, they are determined to be labeled in the same region; otherwise, they are determined to be labeled in different regions (only focusing on label conflicts in the same region).

[0113] Obtain labeled groups for the same area;

[0114] Each labeled text is broken down into hierarchical terms in a predefined ontology library; for example, the label “invasive ductal carcinoma grade 3” is mapped to: breast (organ), carcinoma (lesion type), invasive ductal carcinoma (subtype), grade 3 (grading); or the label “intraductal carcinoma with microinvasive” is mapped to: breast, carcinoma, intraductal carcinoma (subtype), with microinvasive (feature).

[0115] Using the case ID as the root node, a tree structure is constructed according to the ontology hierarchy to obtain a structured semantic tree. The structured semantic tree includes nodes at levels one to four. Among them, the first-level node is the organ, the second-level node is the lesion type, the third-level node is the subtype, and the fourth-level node is the grade / feature. Each node is associated with the corresponding anchor coordinates and expert ID.

[0116] Define conflict types, including term hierarchy conflict, attribute contradiction conflict, and logical inclusion conflict. Term hierarchy conflict means that terms labeled in the same area are on mutually exclusive branches of the structured semantic tree. Attribute contradiction conflict means that the core attributes of terms are opposite. Logical inclusion conflict means that one term completely includes another term but the labeling is inconsistent.

[0117] Based on the relationships between nodes in a structured semantic tree, conflict detection rules are defined.

[0118] If two terms are sibling nodes in the structured semantic tree and are mutually exclusive, then they are determined to be hierarchical conflicts.

[0119] If the attribute values ​​of two terms are opposite in the ontology, then it is determined to be an attribute contradiction;

[0120] If two terms conflict in terms of the categories of a parent node and a child node, it is determined to be a logical inclusion conflict.

[0121] Conflict detection is performed based on structured semantic trees to obtain labeled conflict detection results:

[0122] Extract the node paths of the labeled Gi for each group of the same region in the structured semantic tree;

[0123] Check if there are mutually exclusive branches in the path of the two labeled nodes;

[0124] Compare whether the attributes of sibling attribute nodes are the same;

[0125] Check if there is an inclusion relationship between nodes and if there is a category contradiction.

[0126] The method for detecting annotation conflicts acquires a set of expert spatial anchor points containing information such as 3D coordinates and annotation text. It calculates the spatial distance between anchor points to group annotations within the same region. The annotation text is then decomposed into hierarchical terms from an ontology library, and a structured semantic tree containing node paths and associated information is constructed. Three conflict types are defined: term hierarchy, attribute contradiction, and logical inclusion, along with corresponding detection rules. The method then checks the mutual exclusion of node paths, consistency of attributes at the same level, and the rationality of inclusion relationships within the annotations of the same region to obtain conflict detection results. This process accurately identifies interpretation disagreements among multiple experts in the same lesion region annotation due to mutual exclusion of term hierarchy, opposite core attributes, or logical inclusion contradictions. It transforms scattered annotation information into structured conflict data that can be systematically analyzed, providing a clear object and basis for subsequent conflict resolution. This effectively solves the technical problem that existing methods only support independent annotation by a single user and cannot identify and handle disagreements among multiple experts regarding lesion point interpretations.

[0127] Conflict resolution module: Resolves conflicts in the labeled conflict detection results and obtains consensus suggestions.

[0128] Methods for obtaining consensus recommendations include:

[0129] The conflict detection results are broken down into conflict terminology pairs, pathological morphological parameters of conflict areas, expert subspecialties, and historical diagnostic accuracy.

[0130] The binarized image after threshold segmentation is traversed to extract the lesion contour. The length of the long side of the minimum bounding rectangle of the contour is calculated as the lesion diameter. HE-stained sections are used as input to the U-Net segmentation model to obtain the mitotic count. Ki-67 index and ER / PR expression status are obtained. The lesion diameter, mitotic count, Ki-67 index and ER / PR expression status are spliced ​​together to obtain pathological features.

[0131] The expert weight is obtained by weighting the expert's historical diagnostic accuracy rate, the proportion of the expert's diagnostic frequency in its sub-specialty, and the expert's historical consensus rate. The weighting parameters can be obtained by optimization through a natural heuristic optimization algorithm. The expert's historical consensus rate is obtained by calculating the ratio of the number of times the expert "supported consensus" in the historical rounds to the total number of effective rounds in which the expert participated. The effective rounds are the rounds in which the final consensus was reached.

[0132] According to the predefined recommendation level mapping rules, as shown in Table 1, the rules can be modified according to actual needs to obtain the guideline recommendation level. For each conflicting term, the text similarity between it and the preset guideline standard is calculated. Combined with the guideline recommendation level, the matching degree is calculated. The preset guideline standard can be an authoritative and industry-recognized medical pathology diagnosis guideline / standard that is pre-embedded in the system.

[0133] Table 1. Predefined Recommendation Level Mapping Rules

[0134]

[0135] The weighted fusion graph neural network (GFNN) uses pathological features, expert weights, and matching degrees as inputs to obtain the consensus probability of each term. The term with the highest probability is extracted as the consensus conclusion. In the GFNN, each node represents a conflicting term, and the node features are the corresponding pathological features and matching degrees. The edge weights between two term nodes are obtained by weighting the difference in expert weights and the pathological feature similarity calculated using cosine similarity. The weighting parameters can be optimized using a natural heuristic optimization algorithm. When there are two experts, the weight difference between experts is calculated. When there are more than two experts, the expert weights of each expert are calculated separately. For each conflicting term, the statistical value of the individual weights of all experts supporting the conflicting term is taken as the representative weight of the conflicting term. The absolute value of the difference between the representative weights of the two conflicting terms is calculated to obtain the expert weight difference between the two term nodes.

[0136] The method for obtaining consensus recommendations decomposes conflicting information into conflicting terminology pairs, pathological morphological parameters, and expert backgrounds. Pathological features such as lesion diameter and mitotic count are extracted, and expert weights are calculated by combining historical diagnostic accuracy and subspecialty frequency proportions. Matching degree is obtained based on guideline recommendation level and text similarity. This information is then input into a weighted fusion graph neural network, which outputs the consensus probability of each term and takes the highest one as the conclusion. This process transforms the disagreements among multiple experts on lesion interpretation into quantifiable and fusionable multi-dimensional data. By integrating objective pathological features, expert credibility, and clinical guideline basis, highly credible consensus recommendations are systematically generated. This effectively solves the technical problem of existing methods that only support independent annotation by a single user and cannot form a unified conclusion on disagreements among multiple experts, providing a reliable decision-making reference for multi-expert collaborative diagnosis.

[0137] Example 2

[0138] Please see Figure 3 , Figure 4 This embodiment provides a deviation correction module applied to Embodiment 1, comprising:

[0139] If a conflict is detected, the historical annotation data of experts is structured and parsed to extract terminology bias features, spatial bias features, attribute bias features, and logical bias features. One-hot encoding is used to convert the above four types of bias features into binary features, the severity of each type of bias is numerically quantified, and each bias feature is assigned a corresponding weight through an attention mechanism. The vector of expert bias is then concatenated. Among them, the frequency of conflict between expert annotations and the preset gold standard at four levels (organ, lesion type, subtype, grade / feature) is statistically analyzed to obtain terminology bias features.

[0140] The Euclidean distance distribution between the expert-annotated anchor points and the gold standard three-dimensional coordinates is calculated to identify significant offset areas and obtain spatial deviation characteristics.

[0141] Based on the reverse attribute definition of the ontology library, such as ER positive and ER negative, the probability of attribute contradictions in statistical expert annotations is obtained to obtain attribute bias characteristics.

[0142] By identifying parent-child node conflicts between detected terms, experts can obtain logical deviation characteristics.

[0143] During conflict resolution, the initial expert weights are calculated in real time based on the current conflict context, the expert bias vector, and the experts' historical diagnostic accuracy.

[0144] The system analyzes pathological, molecular, and imaging features, dynamically assigns feature weights through an attention mechanism, and performs a dot product operation with the expert bias vector to correct the expert weights, thus obtaining corrected expert weights.

[0145] Methods for obtaining molecular characteristics include:

[0146] DNA is extracted and purified, and fragmented using methods such as ultrasonic disruption or enzyme digestion to obtain DNA fragments of predetermined lengths. End repair, A-tailing, and adapter ligation are then performed on these fragments to construct sequencing libraries. The library fragments are enriched by PCR amplification, and high-throughput sequencing is then performed. After sequencing, the reads are aligned to the human reference genome, and variant detection and annotation are performed to identify gene mutation types.

[0147] After obtaining fresh tumor tissue or blood samples, tumor cells are isolated from the tissue or tumor cells from peripheral blood are used directly; cells are incubated with fluorescently labeled anti-PD-L1 antibodies to allow the antibodies to bind to the PD-L1 protein on the cell surface; the fluorescence intensity of the cells is detected by flow cytometry, and the proportion of PD-L1 positive cells and the distribution of fluorescence intensity in the cell population are analyzed to obtain the PD-L1 expression level;

[0148] The gene mutation type and PD-L1 expression level are spliced ​​together to obtain molecular characteristics.

[0149] Methods for obtaining image features include:

[0150] The CT image is normalized to obtain a normalized CT image. The Region of Interest (ROI) is obtained through an automatic segmentation algorithm. Within the ROI, different pixel pair distances and orientations are defined, and the frequency of pixel pairs with specific gray value combinations is counted to construct a gray-level co-occurrence matrix. The texture entropy is calculated based on the gray-level co-occurrence matrix.

[0151] During an MRI examination, the patient is given an intravenous injection of gadolinium contrast agent, and MRI scans are performed according to a specific time sequence to obtain image data at different time points.

[0152] The obtained DCE-MRI images are imported into preset post-processing software for image registration; the tumor or tissue of interest region is manually or semi-automatically segmented on the image. Using the software's built-in pharmacokinetic model, the contrast agent concentration over time within the region of interest is calculated to obtain enhancement parameters.

[0153] The texture entropy and enhancement parameters are concatenated to obtain image features.

[0154] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A remote digital image analysis collaborative system, characterized in that, include: Virtual diagnostic module: Acquires continuous pathological slide sequences and performs spatial registration to obtain registered slides; Model building module: Perform three-dimensional voxel reconstruction on the registered slices to obtain a three-dimensional voxel reconstruction model, and then perform semantic enhancement to obtain a three-dimensional pathological voxel model; Collaborative Interaction Module: Based on VR technology, generate virtual avatars of G experts, annotate the three-dimensional pathological voxel model, and obtain spatial anchor point annotation data of the experts; Graph-driven module: Constructs a structured semantic tree based on spatial anchor point annotation data, detects annotation conflicts, and obtains annotation conflict detection results. Obtain spatial anchor point annotation data from each expert; Calculate the spatial distance between anchor points; Obtain the labeled groups for the same region, and decompose each labeled text into hierarchical terms in the preset ontology library; Using case ID as the root node, a tree structure is constructed according to the ontology hierarchy to obtain a structured semantic tree; the tree structure includes nodes of level one to four, where level one nodes are organs, level two nodes are lesion types, level three nodes are subtypes, and level four nodes are grades / features; Define conflict types, including term hierarchy conflict, attribute contradiction conflict, and logical inclusion conflict; term hierarchy conflict means that terms labeled in the same area are on mutually exclusive branches of the structured semantic tree, attribute contradiction conflict means that the core attributes of terms are opposite, and logical inclusion conflict means that one term completely includes another term but the labeling is inconsistent. Based on the relationships between nodes in a structured semantic tree, conflict detection rules are defined. Conflict detection is performed based on structured semantic trees to obtain labeled conflict detection results; Conflict resolution module: Resolves conflicts in the labeled conflict detection results and obtains consensus suggestions. The conflict detection results are broken down into conflict terminology pairs, pathological morphological parameters of conflict areas, expert subspecialties, and historical diagnostic accuracy. The binarized images after threshold segmentation are traversed, and the lesion diameter, mitotic count, Ki-67 index and ER / PR expression status are extracted and stitched together to obtain pathological features. The expert weight is obtained by weighting the expert’s historical diagnostic accuracy rate, the proportion of the expert’s diagnostic frequency in its subspecialty, and the expert’s historical consensus rate. The guide recommendation level is obtained according to the predefined recommendation level mapping rules. For each conflicting term, its text similarity with the preset guide standard is calculated. The matching degree is obtained by combining the guide recommendation level. The pathological features, expert weights, and matching degree are used as inputs to a weighted fusion graph neural network to obtain the consensus probability of each term, and the term with the highest probability is extracted as the consensus conclusion. In the weighted fusion graph neural network, each node represents a conflicting term, and the node features are the pathological features and matching degree corresponding to the conflicting term; the edge weight between two term nodes is obtained by weighting the difference in expert weights and the similarity of pathological features calculated by cosine similarity.

2. The remote digital image analysis collaborative system according to claim 1, characterized in that, Methods for obtaining registration slices include: Gaussian denoising and grayscale normalization were performed on each pathological slide to obtain a standardized slide sequence; The first slice of the standardized slice sequence Using the reference coordinate system, the origin is... The top-left pixel is defined by the x-axis along the width direction, the y-axis along the height direction, and the z-axis along the slice sequence direction; global coordinate system parameters are obtained, including the origin, axis directions, and the initial z-coordinate of each slice. For each slice, the anatomical structure region is extracted by threshold segmentation. The binarized image after threshold segmentation is traversed, and the arithmetic mean of the coordinates of all pixels in the binarized image is calculated to obtain the feature points. For the i-th and (i+1)-th slices, for each feature point, extract the neighborhood of C×C pixels and count the gray-level pairs within the neighborhood. The probability of its occurrence, and All values ​​are grayscale values. A grayscale matrix is ​​generated. The elements of the grayscale matrix are expanded into a one-dimensional vector by rows / columns to obtain the local descriptors of the feature points. The Euclidean distance between the local descriptors of the i-th slice and the (i+1)-th slice is calculated. Similar feature points with a distance less than the distance threshold are filtered to obtain a set of matching point pairs between adjacent slices. Let the point on the (i+1)th slice of the matching point pair set correspond to the point on the ith slice after transformation; minimize the transformation error of the matching point pair using the least squares method; obtain the rigid transformation parameters of adjacent slices, and perform rigid transformation in combination with the rigid transformation parameters, that is, perform rigid transformation in combination with the rigid transformation parameters and transformation formula to obtain coarsely aligned slices. Construct an m×n control point grid on the coarse-aligned slice plane, with control point coordinates as follows: , is the grid spacing; m and n are the number of grid rows and columns, respectively. and The number of grid rows and columns corresponding to the control points; the deformation displacement of any point is obtained by interpolating the control point displacement through the B-spline basis function, and the elastic transformation parameters are obtained; among them, the control point displacement is optimized by using the gradient descent method to obtain the minimum target value with the goal of maximizing gray-level similarity and deformation smoothness, and the elastically registered slice is obtained. Calculate the total transformation of the i-th slice relative to the first slice according to the rigid transformation parameters and the elastic transformation parameters; obtain the globally optimized set of transformation parameters; for any non-adjacent slices and , where j < k, calculate the deviation between its directly matched transformation and the cumulative transformation, and minimize the total global deviation by adjusting the transformation parameters of the intermediate slices; obtain the registered slice sequence.

3. The remote digital image analysis collaborative system according to claim 1, characterized in that, Methods for obtaining three-dimensional voxel reconstruction models include: The boundaries of the 3D voxel model are determined based on the pixel size d and the slice thickness t. The x-axis boundary is [0, W×d], the y-axis boundary is [0, H×d], and the z-axis boundary is [0, (N-1)×t]; where W is the number of pixels in the slice width; H is the number of pixels in the slice height; and N is the number of registered slices in the registered slice sequence. Obtain voxel resolution of d in both the x-axis and y-axis directions, and voxel resolution of t / k in the z-axis direction, where k is the B-spline basis function interpolation factor; The registered slice sequence is normalized to obtain the corrected slice sequence; Map the pixels of each calibration slice to three-dimensional voxel coordinates; Using the local regions containing Q×Q pixel blocks from two adjacent correction slices as input to the depth completion model, the local regions are obtained in... and The k-1 layer interpolated voxel values ​​between; the mapped voxels and interpolated voxel values ​​of all correction slices are stitched together in z-axis order to form a continuous three-dimensional voxel reconstruction model and a completed slice; among which, and The first Zhang Hedi The z-axis coordinate of the correction slice in the three-dimensional voxel coordinate system.

4. The remote digital image analysis collaborative system according to claim 1, characterized in that, Methods for obtaining three-dimensional pathological voxel models include: Obtain expert annotation data and pixel coordinate markers, and map the pixel coordinates of the annotation area on the completed slice to three-dimensional voxel coordinates to form a set of annotation points in three-dimensional space; Voxel space is divided according to resolution for tumor regions in two-dimensional slices. and First, combine the seed point set converted to voxel space by resolution. and traversal and If the voxel gray value is within the gray range of the tumor region and is connected to the existing seed point set in three-dimensional space, the corresponding voxel will be included in the three-dimensional tumor region to obtain preliminary three-dimensional semantic annotation. Extract all voxels in the 3D neighborhood of each target voxel in the voxel space, calculate the voxel mean, variance, and gradients in the x, y, and z directions of the voxels in the neighborhood, then calculate the gradient magnitude, take the maximum gradient magnitude to obtain the gradient feature, and concatenate the mean, variance, and gradient feature to obtain the grayscale feature; calculate the shortest Euclidean distance from the target voxel to the known semantic region to obtain the morphological feature. By using preliminary 3D semantic annotation, grayscale features, and morphological features as inputs to the label prediction model, a 3D pathological voxel model with semantic labels is obtained.

5. The remote digital image analysis collaborative system according to claim 1, characterized in that, Methods for obtaining spatial anchor point annotation data from experts include: Virtual avatars of various experts are generated based on VR, and the position coordinates and viewing direction of the virtual avatars are synchronized in real time based on the spatial coordinate system of the three-dimensional pathological voxel model. Experts annotate the three-dimensional pathological voxel model, and the system automatically generates spatial anchor points containing anchor point attributes, including annotation coordinates, annotation type, annotation expert, and annotation timestamp. Using dynamic rendering of the view frustum, only the voxel data within the current expert's field of view is rendered to obtain the expert's spatial anchor point annotation data.

6. The remote digital image analysis collaborative system according to claim 5, characterized in that, Methods for obtaining spatial anchor point annotation data of experts by using dynamic frustum rendering, which renders only voxel data within the current expert's field of view, include: Obtain the position coordinates, view direction vector, view angle, near clipping plane distance, and far clipping plane distance, and construct the view frustum; The coordinate range of the view cone in the three-dimensional pathological voxel model is obtained by taking the intersection of the coordinate range of the view cone and the coordinate range of the three-dimensional pathological voxel model. Traverse the voxels in the 3D pathological voxel model whose coordinates fall within the cone coordinate range. For each voxel that has been initially screened, determine whether it is located inside the cone: calculate the distance from the voxel center to the 6 boundary planes of the cone; if the distance from the voxel to all planes is within the plane, it is determined to be a voxel inside the cone; obtain the set of voxels inside the cone by statistics. Calculate the Euclidean distance from each voxel in the voxel set to the expert coordinates, and set the N-level rendering detail according to the preset rendering rules and the Euclidean distance; obtain the voxels inside the view frustum with detail level markers; Voxels are grouped and processed according to their level of detail. The voxels are projected onto the screen coordinate system. When the expert's viewpoint changes, a new view frustum is obtained and the level of detail is reassigned based on the new view frustum.

7. A remote digital image analysis collaborative system according to claim 6, characterized in that, The viewing cone includes a near clipping plane, a far clipping plane, a left side, a right side, a top side, and a bottom side; wherein, the near clipping plane and the far clipping plane are both perpendicular to the line of sight, and are respectively located at the near clipping distance and the far clipping distance from the observation point; the angle between the left side and the right side and the line of sight is half of the horizontal field of vision angle, defining the horizontal visible range; the angle between the top side and the bottom side and the line of sight is half of the vertical field of vision angle, defining the vertical visible range.

8. The remote digital image analysis collaborative system according to claim 1, characterized in that, In the spatial anchor point annotation data of each expert, each anchor point includes three-dimensional coordinates, annotation text, expert ID, and annotation time; If the spatial distance between two anchor points is not greater than U voxels, they are determined to be labeled in the same region; otherwise, they are determined to be labeled in different regions.

9. A remote digital image analysis collaborative system according to claim 1, characterized in that, In the tree structure, each node is associated with its corresponding anchor point coordinates and expert ID.

10. A remote digital image analysis collaborative system according to claim 1, characterized in that, Methods for obtaining pathological features include: The binarized image after threshold segmentation is traversed to extract the lesion contour. The length of the long side of the minimum bounding rectangle of the contour is calculated as the lesion diameter. HE-stained sections are used as input to the U-Net segmentation model to obtain the mitotic count. Ki-67 index and ER / PR expression status are obtained. The lesion diameter, mitotic count, Ki-67 index and ER / PR expression status are spliced ​​together to obtain pathological features. When there are 2 experts, calculate the weight difference between experts to obtain the expert weight difference. When there are more than 2 experts, calculate the expert weight of each expert. Then, for each conflicting term, take the statistical value of the individual weights of all experts who support the conflicting term as the representative weight of the conflicting term. Calculate the absolute value of the difference between the representative weights of the two conflicting terms to obtain the expert weight difference between the two term nodes.

Images