Human tomographic anatomical image three-dimensional visualization reconstruction method and system

Abstract

The present application relates to the technical field of medical image processing and three-dimensional visualization, and discloses a human body tomographic anatomic image three-dimensional visualization reconstruction method and system, which comprises five steps of interlayer registration and gray scale normalization, multi-organ semantic segmentation based on anatomic position attention module, surface-volume hybrid three-dimensional reconstruction, interactive virtual dissection layer-by-layer peeling visualization and anatomic knowledge graph driven labeling, and a quality feedback signal is sent from a segmentation module to a registration module to form a closed-loop cooperation, thereby providing reusable digital three-dimensional anatomic resources for anatomic teaching of medical colleges.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and three-dimensional visualization technology, and in particular to a method and system for three-dimensional visualization reconstruction of human sectional anatomical images. Background Technology

[0002] In medical school anatomy teaching practice, traditional methods primarily rely on anatomical manipulation and structural observation using human cadaver specimens. However, due to constraints such as the increasing scarcity of cadaver specimens, high preservation and maintenance costs, and biosafety risks, the number of anatomical specimens available for students to manipulate is extremely limited. Furthermore, a core difficulty students face when learning sectional anatomy lies in their inability to establish a spatial correspondence between two-dimensional continuous sectional images and the overall three-dimensional anatomical structure. This cognitive barrier severely impacts the quality and effectiveness of anatomy teaching. Therefore, utilizing digital three-dimensional reconstruction technology to provide reusable virtual anatomical resources for anatomy teaching has become a research hotspot in this field.

[0003] Chinese patent CN110570515B discloses a method for 3D modeling of human skeleton using CT images. The technical solution is as follows: First, a DICOM format image of the human skeleton is obtained using a medical CT scan. Then, a slicing gradient interpolation method based on gray-level gradient characteristics is used to increase the resolution of the image sequence along the scanning direction. Next, based on the probability distribution of each gray level in the image, a maximum entropy segmentation method is selected to determine the threshold parameters. Binary segmentation is then performed on the image to separate the bones from the soft tissue. Finally, tetrahedral voxels of the isosurface are defined using the spacing between three-dimensional structural feature points. A set of tetrahedral voxels is generated using a volume fitting method to construct a 3D model of the human skeleton. This method provides a feasible technical path for skeleton modeling based on CT images.

[0004] However, analysis revealed the following shortcomings in the aforementioned technical solution: First, this method only segments and models the single type of anatomical structure, bones, and cannot simultaneously segment and collaboratively model multiple anatomical structures such as bones, muscles, blood vessels, nerves, and internal organs. Therefore, it is difficult to meet the needs of comprehensive display of the overall human body structure in anatomy teaching. Second, this method uses a traditional threshold segmentation method based on maximum entropy. This method relies on the statistical characteristics of gray-level probability distribution and has limited ability to identify boundaries between soft tissue organs with low gray-level contrast. When faced with complex multi-organ scenes in the human body, the segmentation accuracy is difficult to guarantee.

[0005] Furthermore, this 3D reconstruction scheme employs a tetrahedral voxel stitching approximation method, which is essentially a surface reconstruction strategy based on voxel fitting. While this method can obtain the basic 3D morphology of the skeleton, it has the following technical limitations: Firstly, the accuracy of tetrahedral voxel fitting is limited by the extraction quality and distribution density of feature points. When the interlayer spacing is large or the feature point distribution is uneven, the reconstruction result is prone to surface roughness and loss of detail. Secondly, this method lacks volume rendering capabilities, cannot provide volumetric rendering effects, and cannot support perspective observation of internal structures and arbitrary cross-sectional display.

[0006] Furthermore, the aforementioned technical solutions completely lack interactive visualization and anatomical annotation capabilities. In actual anatomy teaching scenarios, students need to be able to perform interactive operations such as rotating, scaling, and layer-by-layer dissection on the 3D model to observe the spatial relationships of anatomical structures from different angles and depths; they also need to obtain annotation information such as the names, blood supply sources, and nerve innervations of each anatomical structure to aid learning. The methods described above only output static 3D skeletal models and lack the interactive visualization and knowledge annotation capabilities necessary for such teaching, thus making them unsuitable for direct application in the construction of digital anatomy teaching resources.

[0007] In summary, there is an urgent need for a complete technical solution that can automatically segment multiple anatomical structures in continuous tomographic images of the human body, perform high-quality 3D reconstruction, interactive visualization, and intelligent anatomical annotation. This solution would overcome the shortcomings of existing methods in terms of multi-organ coverage, segmentation accuracy, reconstruction quality, interactivity, and annotation functions, and provide high-quality digital 3D anatomical resources for anatomy teaching in medical colleges. Summary of the Invention

[0008] To address at least one of the technical problems existing in the prior art, this invention provides a method for three-dimensional visualization reconstruction of human sectional anatomical images. This method achieves end-to-end reconstruction from continuous sectional images to an interactive three-dimensional anatomical model through five deeply coupled steps: inter-layer registration and grayscale normalization, multi-organ semantic segmentation based on deep learning, surface-volume hybrid three-dimensional reconstruction, interactive virtual anatomical visualization, and automatic annotation driven by anatomical knowledge graph.

[0009] The present invention provides a three-dimensional visualization reconstruction method for human sectional anatomical images, comprising the following steps: Step S1, sectional image inter-layer registration and grayscale normalization step, acquiring a continuous sequence of human sectional images, performing rigid registration based on mutual information on adjacent layers to correct inter-layer displacement and rotation deviations to less than a preset pixel threshold, and performing global grayscale normalization, dynamically adjusting the registration parameters of subsequent layers using inter-layer consistency feedback signals; Step S2, multi-organ semantic segmentation step, inputting the normalized image into a three-dimensional semantic segmentation network based on an encoder-decoder architecture, the encoder extracting multi-scale features layer by layer, the decoder introducing an anatomical position attention module to perform weighted fusion using anatomical prior probability maps, outputting multi-class segmentation annotation results, and obtaining a segmentation boundary continuity evaluation value by calculating the Dice coefficient of the segmentation boundary of adjacent layers, generating a feedback signal to trigger local fine-tuning registration when the evaluation value is lower than a preset quality threshold; Step S3, three-dimensional surface-volume hybrid reconstruction step, applying an improved marching method to the segmentation results. The Cubes algorithm extracts isosurface triangular meshes to construct a surface model, while a ray projection volume rendering method based on organ category opacity transfer function is used to construct a volumetric rendering model. The models are then fused and adaptive mesh subdivision is performed in a unified coordinate system. Step S4 is the interactive virtual anatomy visualization step, which constructs a real-time interactive scene based on the fused 3D model, supporting single organ extraction, transparent overlay of multiple structures, virtual layer-by-layer peeling, and arbitrary cross-sectional display. Step S5 is the anatomical knowledge graph-driven annotation step, which automatically generates structural names, blood supply, nerve innervation, and adjacency annotations based on the pre-constructed anatomical knowledge graph and the segmentation and reconstruction results.

[0010] In the above method, the segmentation quality feedback signal in step S2 forms a closed-loop feedback to the registration process in step S1. When discontinuous regions appear at the segmentation boundary, local fine registration of the corresponding layer is automatically triggered, thereby establishing a collaborative optimization mechanism between segmentation and registration. This enables the overall system to achieve performance improvements in both segmentation accuracy and reconstruction quality that surpass those achieved when each step is executed individually.

[0011] This invention also provides a three-dimensional visualization and reconstruction system for human anatomical tomographic images, comprising: an inter-slice registration and grayscale normalization module, a multi-organ semantic segmentation module, a three-dimensional hybrid reconstruction module, an interactive visualization module, and an anatomical annotation module. Each module corresponds one-to-one with the steps described above. The output of the inter-slice registration and grayscale normalization module serves as the input to the multi-organ semantic segmentation module, the output of the multi-organ semantic segmentation module serves as the input to both the three-dimensional hybrid reconstruction module and the anatomical annotation module, and the output of the three-dimensional hybrid reconstruction module serves as the input to both the interactive visualization module and the anatomical annotation module. The multi-organ semantic segmentation module also sends segmentation quality feedback signals to the inter-slice registration and grayscale normalization module to achieve closed-loop collaboration between modules.

[0012] Compared with existing technologies, this invention has the following beneficial effects: First, this invention achieves automated segmentation and collaborative 3D reconstruction of various anatomical structures, including bones, muscles, blood vessels, nerves, and internal organs, breaking through the limitation of existing technologies that can only handle single skeletal structures; Second, this invention adopts a hybrid strategy combining surface reconstruction and volume rendering, which can provide high-precision surface geometry models and support volume rendering and arbitrary cross-sectional display, enriching the means of 3D visualization; Third, the virtual layer-by-layer peeling and anatomical knowledge graph-driven annotation functions provided by this invention offer highly interactive and information-rich digital teaching resources for anatomy teaching; Fourth, this invention establishes a closed-loop quality feedback mechanism from the segmentation module to the registration module, enabling a collaborative optimization relationship between the segmentation and registration steps. The overall system performance surpasses the level achievable by simply chaining and superimposing the steps, demonstrating a significant synergistic effect. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the three-dimensional visualization reconstruction method for human cross-sectional anatomical images provided in this embodiment of the invention.

[0014] Figure 2 This is a schematic diagram of the architecture of the three-dimensional visualization and reconstruction system for human cross-sectional anatomical images provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the scope of protection of the present invention. Non-substantial improvements and adjustments made to the present invention by those skilled in the art without exceeding the scope of the technical solutions of the present invention are still within the scope of protection of the present invention.

[0016] See Figure 1 This embodiment provides a method for three-dimensional visualization and reconstruction of human sectional anatomical images, using a frozen section sectional image sequence obtained from the China Visualized Human Body Data Set as an example. In this embodiment, the sectional image sequence contains approximately 3600 consecutive transverse section images, with a slice thickness of approximately 0.5 mm, a planar resolution of 3072 × 2048 pixels, and TIFF format RGB color images. It should be noted that the method described in this invention is also applicable to clinical CT scan images or MRI scan images. When the input is a CT image, the slice thickness range is typically 0.5 mm to 2 mm, the planar resolution is not less than 512 × 512 pixels, and the image format is the DICOM standard format.

[0017] Step S1: Inter-slice registration and grayscale normalization of tomographic images. In this step, inter-slice registration processing is first performed on the acquired continuous tomographic image sequence to correct spatial displacement and rotation deviations between adjacent layers caused by factors such as specimen fixed offset, uneven slice thickness, and slight changes in scanning parameters. In one embodiment of the present invention, registration is achieved using a rigid registration method based on mutual information. This method transforms the registration problem into an optimization problem that maximizes the mutual information between two images.

[0018] Specifically, the definition of the first Layer image is , No. Layer image is The goal of the registration process is to solve for the optimal rigid transformation parameters. The transformed image Compared with reference image Maximizing mutual information between them. Among them, and respectively along shaft and Translation along the axis, in pixels; The rotation angle is around the image center, in degrees. Preferably, in this embodiment, the Powell optimization algorithm is used to solve the above mutual information maximization problem, and the initial search range is set to the translation direction. Pixels, rotation direction The convergence accuracy was optimized to 0.1 pixels in the translation direction and 0.01 degrees in the rotation direction.

[0019] After completing the initial mutual information registration, this invention further introduces an inter-layer consistency feedback mechanism to dynamically optimize the registration quality. Specifically, the first layer is defined as follows: Image after layer registration Relative to reference layer image Gray-scale distribution consistency measure for: ,in: For the first Normalized probability distribution of grayscale histogram of the image after layer registration; The gray-level histogram normalized probability distribution of the reference layer image; To be from the distribution To distribution The Kullback-Leibler divergence, used to measure the degree of difference between two probability distributions, is calculated as follows: ,in The number of gray levels is taken in this embodiment. ; The maximum value of the KL divergence across the entire image sequence is used to normalize the consistency metric to... Within the interval, A value closer to 1 indicates better consistency in grayscale distribution. Preferably, in this embodiment, the middle layer of the image sequence (i.e., near the 1800th layer) is used as the reference layer. The reason for choosing the intermediate layer as the reference layer is that it is located at the geometric center of the sequence, which makes the accumulation of registration error at both ends of the sequence most balanced.

[0020] When the consistency metric of a certain layer Below the preset threshold When this occurs, the system automatically expands the registration search range of the current layer and its neighboring layers to 1.5 times the original range, and re-performs registration with a finer search step size. In this embodiment, a preset threshold is used. The value is 0.85, and the selection of this threshold is based on experimental verification. When the subsequent segmentation steps can achieve the required segmentation accuracy, then... At that time, differences in grayscale distribution will cause obvious boundary discontinuities in the output of the segmentation network at the corresponding layer. This feedback mechanism enables the registration process to adaptively refine the adjustment for layers with poor quality, rather than applying the same fixed registration strategy to all layers, thereby effectively controlling computational costs while ensuring overall registration quality.

[0021] After registration, global grayscale normalization is performed on the registered tomographic image sequence. In a preferred embodiment of the present invention, grayscale normalization is achieved using a histogram matching method, with the reference layer as the reference. The grayscale histogram is used as the target distribution, and nonlinear mapping transformations are performed on the grayscale values ​​of the remaining image layers. For color tomographic images (such as RGB images in human visualization datasets), the image is first converted from the RGB color space to the Lab color space before performing grayscale normalization. Histogram matching normalization is performed only on the luminance channel L, while keeping the chrominance channels a and b unchanged. After normalization, the image is converted back to the RGB color space. The technical advantages of this approach are: firstly, it unifies the luminance distribution of each image layer to facilitate consistent input features for subsequent segmentation networks; secondly, it preserves the color information of the original image to support the realistic rendering requirements of subsequent visualization displays.

[0022] In the verification test of this embodiment, after performing the above-mentioned inter-layer registration and grayscale normalization processing on the image sequence of the visualized human body dataset containing 3600 layers, the inter-layer translational deviation decreased from the original average of 2.3 pixels to 0.6 pixels, and the inter-layer rotational deviation decreased from the average of 0.8 degrees to 0.2 degrees. The registration accuracy meets the design requirements of translational deviation less than 1 pixel and rotational deviation less than 0.5 degrees. At the same time, the inter-layer standard deviation of grayscale distribution of each layer decreased from 18.7 before normalization to 3.2 after normalization, and the grayscale consistency was improved by approximately 83%. The registered and normalized tomographic image sequence output in step S1 will be used as the input data for multi-organ semantic segmentation in step S2.

[0023] Step S2: Multi-organ semantic segmentation. This step aims to automatically perform semantic segmentation on various human anatomical structures in the normalized tomographic image sequence output from Step S1. In one embodiment of the present invention, the segmentation targets cover the following anatomical structure categories: bones (including cortical and cancellous bone), muscles (including skeletal and smooth muscles), blood vessels (including arteries and veins), nerves, and internal organs (including liver, spleen, kidney, lung, heart, stomach, intestines, etc.), totaling no fewer than 15 subcategories. The segmentation task is implemented using a three-dimensional semantic segmentation network based on an encoder-decoder architecture. This network takes three-dimensional voxel blocks as input and outputs multi-category pixel-level labeled images of the same size as the input.

[0024] Specifically, the encoder path of the 3D semantic segmentation network used in this embodiment consists of 5 stages, each stage containing 2 residual convolutional blocks and 1 downsampling module. The input voxel block size of the first stage is... The output feature map has 32 channels; the second to fifth stages sequentially downsample the spatial resolution to the original size. , , and The number of feature map channels increases accordingly to 64, 128, 256, and 512. Each residual convolutional block contains two sets of... The sequence consists of convolutional-batch normalization-ReLU activation and one residual skip connection. The downsampling module uses a stride of [missing value]. of Convolution is implemented to extract higher-level semantic features while reducing spatial resolution.

[0025] The decoder path also consists of five stages, and its structure is symmetrical to that of the encoder path. Each decoding stage first passes through... The transposed convolution performs upsampling to restore spatial resolution. The upsampled feature map is then concatenated along the channel dimension with the feature maps of the corresponding stages in the encoder path, followed by feature fusion and refinement through two residual convolutional blocks. The innovation of this invention lies in introducing an anatomical location attention module at each stage of the decoder path. The core idea of ​​this module is to use a pre-constructed prior spatial probability map of human anatomy to perform spatially location-aware weighted fusion of the decoded features.

[0026] The specific implementation of the anatomical location attention module is as follows: First, a priori spatial probability map of human anatomy, aligned with the spatial space of the input tomographic image sequence, is pre-constructed. ,in , and These represent the number of voxels in the width, height, and depth directions of the map, respectively. Number of anatomical structure categories (in this embodiment) The position of each voxel in the atlas place dimensional vector This represents the prior probability value of the spatial location belonging to each anatomical structure category, with the probability value ranging from [value missing]. Furthermore, the sum of the probabilities of each category is 1. This atlas is calculated based on the category frequency of each voxel location by statistically analyzing a large-scale labeled anatomical dataset. In this embodiment, it is constructed based on a dataset of 50 manually labeled human visualizations.

[0027] In the decoder path In this stage, let the decoded feature map be... , probability graph Downsampling to the same level using trilinear interpolation The same space size Then through Convolution will The number of channels from Mapped to Obtain the attention weight map Weighted feature map Calculated by element-wise multiplication:

[0028] ,in: This represents the element-wise multiplication operation (Hadamard product). This represents the Sigmoid activation function, which maps attention weights to... This allows for soft-selective enhancement of decoded features—when the prior probability of an anatomy corresponding to a certain spatial location is high, the feature response at that location is enhanced; conversely, it is suppressed. Subsequently, the weighted feature map... Compared with the original decoded feature map After splicing along the channel dimension Convolutional layer dimensionality reduction and fusion The final anatomical location-aware feature map is obtained by using multiple channels. The technical effect of this design is that the anatomical location attention module uses prior knowledge to guide the segmentation network to focus on the correct organ type in the correct spatial location, effectively reducing the probability of missegmentation in areas far from the target organ. The improvement in segmentation accuracy is particularly significant for structures with small volume and low grayscale contrast (such as nerves and small blood vessels).

[0029] During network training, a composite loss function is used. Optimize network parameters:

[0030] ,in: The weighted cross-entropy loss is used to address the imbalance of sample sizes among organ categories, with the weight of each category inversely proportional to its number of voxels. The Dice coefficient loss is used to directly optimize the overlap of segmented regions; The boundary distance loss is used to penalize boundary offset by calculating the average surface distance between the predicted boundary and the true boundary. Its purpose is to improve the accuracy and smoothness of the segmentation boundary. , and Here are the weighting coefficients for each loss term; in this embodiment, they are respectively set to values ​​of [value]. , and The selection criteria for this weight configuration are as follows: and Maintaining equal weight as the primary loss item As an auxiliary loss term, it is given a smaller weight to avoid over-constraining the boundaries and affecting the overall segmentation performance. Training uses the Adam optimizer, with an initial learning rate of [value missing]. The cosine annealing strategy is used to gradually decay the temperature to... The total number of training rounds is 200, and the batch size is 4.

[0031] Another key innovation of this step lies in the introduction of a segmentation quality feedback mechanism to achieve closed-loop collaboration between steps S2 and S1. Specifically, a segmentation boundary continuity evaluation value is calculated for each layer of the segmentation result output by the network. : ,in: The number of categories to be divided; and The first Layer and first The first in the layer The set of pixels representing the segmentation boundaries of organoids; This represents the symmetric difference operation; This represents the number of pixels in the set. The range of values ​​is , The closer the value is to 1, the smoother and more continuous the boundary change between adjacent layers. The closer to 0, the more drastic the change in the segmentation boundary. Below the preset quality threshold (In this embodiment) When the Dice coefficient is not less than 0.85, the system automatically sends a feedback signal to step S1, triggering the action of the first step. Five layers of images, including the first layer and two layers before and after it (a total of five layers), undergo local fine-grained registration to correct potential inter-layer registration deviations. Experiments have verified that after introducing this closed-loop feedback mechanism, the average Dice coefficient of the overall segmentation increases from 0.89 before the feedback was introduced to 0.92, and the proportion of discontinuous layers at the boundaries decreases from 4.2% to 0.8%.

[0032] Step S3: 3D Surface-Volume Hybrid Reconstruction. This step receives the multi-organ segmentation and annotation results from Step S2 and the normalized tomographic image sequence from Step S1, respectively constructs a 3D surface model and a volumetric rendering model, and then fuses them in a unified 3D coordinate system. The technical significance of this surface-volume hybrid reconstruction strategy is that the surface model can provide accurate geometric morphological descriptions, suitable for external morphological observation and interactive operation; the volumetric model can preserve internal density information, suitable for perspective observation and arbitrary cross-sectional display. The complementary fusion of the two can simultaneously meet the dual needs of external morphological display and internal structural exploration in anatomy teaching.

[0033] For the construction of 3D surface models, this invention employs an improved Marching Cubes algorithm. The classic Marching Cubes algorithm extracts isosurface triangular meshes from voxel meshes using a fixed grayscale isosurface threshold. However, due to the differentiated grayscale distribution characteristics of different types of anatomical structures in the human body, using a single fixed threshold is insufficient to achieve good surface extraction results for all organs. Therefore, this invention proposes an adaptive isosurface threshold determination method based on organ category. Specifically, for the ... Organoid segmentation mask Within the voxel region, adaptive isosurface thresholding Determined in the following manner: ,in: For the first Organoid segmentation mask The mean of all voxel gray values, in units consistent with the gray dimension of the image (0 to 255 for an 8-bit image). For the first Organoid segmentation mask The mean gray-level gradient magnitude at the boundary voxel is obtained by calculating and averaging the three-dimensional Sobel gradient at the mask boundary. This is the category-related moderating factor, and its value range is... For hard tissues such as bones with high gray values ​​and large boundary gradients, Take a positive value (in this embodiment, the skeleton class takes...) To increase the threshold and thus obtain a more accurate bone cortical surface; for soft tissue organs with low grayscale contrast, such as the liver and spleen, Take negative or zero values ​​(in this embodiment, internal organs are taken as negative values). This adaptive thresholding method reduces the threshold to avoid missing weak gradient regions at organ edges. The technical advantage of this method lies in its differentiated processing of grayscale characteristics for different anatomical structures, reducing the average Hausdorff distance of surface extraction by approximately 25% compared to a fixed threshold method.

[0034] After extracting the isosurface triangular mesh, a Taubin filter is applied to the original mesh to eliminate the staircase effect caused by voxel discretization. The Taubin filter is a volume-preserving mesh smoothing method that alternates between forward and reverse Laplacian smoothing, each determined by a smoothing factor. and reverse factor The iteration count is set to 50. Compared to simple Laplacian smoothing, Taubin filtering can smooth the surface while avoiding continuous shrinkage of the mesh volume, thus preserving the overall geometric features of the anatomical structure without distortion. Preferably, adaptive mesh subdivision based on surface curvature is performed after Taubin filtering: for regions with high curvature (such as articular surfaces and vascular branches), Loop subdivision is performed to increase the density of triangular facets, thereby preserving detailed morphological features; for flat regions with low curvature, the original mesh density is kept unchanged to control the total number of facets within a reasonable range. In this embodiment, the curvature threshold is set to... mm The threshold is determined based on the fact that areas with curvature higher than this value usually correspond to morphological turning points of anatomical structures, and the preservation of details in these areas is of great significance for teaching demonstrations.

[0035] For constructing the volumetric rendering model, this invention employs a ray casting volume rendering method. The ray casting method starts from the viewpoint, emits rays into the 3D volumetric data space, samples voxel data along the ray path at fixed step sizes, and accumulates the color and opacity values ​​of each sampled point according to a front-to-back compositing rule to finally obtain the color value of the screen pixel. The innovation of this invention lies in designing an opacity transfer function based on organ category to achieve a volumetric rendering effect where multiple organs are visible in layers.

[0036] Specifically, for grayscale values The sampling points, if their organ category is (Determined by the segmentation and labeling results in step S2), then the opacity of this sampling point... Determined by the following transfer function: ,in: For the first The basic opacity coefficient of organoids ranges from [value range missing]. In this embodiment, the skeleton class is taken as Muscle type Vascular retrieval Internal organs A higher base opacity indicates that the organ type is less transparent in the rendering. For the first The central value of the grayscale distribution of organoids is taken as the median of the grayscale values ​​within the segmented region of that organoid. For the first The standard deviation of the grayscale distribution of organoids controls the rate at which the opacity decreases as the grayscale deviates from the center value; This is a user-adjustable organ visibility weighting factor, with a value range of [value range missing]. The default value is 1.0, which users can adjust in real time through the interactive interface. This technology dynamically changes the transparency of each organ. The technical effect of this transfer function is that the opacity mapping curves for each organ category are independent, ensuring that bones, as high-opaque structures, are always visible, while muscles and internal organs have lower opacity, thus achieving a layered perspective effect. Users can adjust the opacity... Selectively hide or show specific organs.

[0037] Finally, the 3D surface model and the volumetric rendering model are fused in a unified 3D world coordinate system. During the fusion process, the two models share the same spatial transformation matrix and viewpoint parameters. During rendering, volumetric rendering is first performed using volumetric ray casting to generate the volumetric rendering image. Then, the depth buffer of the surface model is compared with the depth information of the volumetric rendering. For areas visible from the surface model, the surface shading result is used to overlay the volumetric rendering result; for areas occluded by the surface, the perspective effect of the volumetric rendering is preserved. The technical advantage of this fusion rendering strategy is that users can observe both the clear surface outline of organs and see the internal structure through semi-transparent areas.

[0038] Step S4: Interactive Virtual Anatomy Visualization. This step constructs a visualization scene supporting real-time interactive operations based on the fused 3D model output from Step S3. In one embodiment of the invention, the interactive visualization module is implemented based on the OpenGL graphics rendering pipeline, with a rendering frame rate of no less than 30 FPS to ensure smooth user interaction. This step supports the following four core interactive operation modes.

[0039] The first operating mode is single organ extraction and display. Users can select the target organ name by clicking with the mouse or choosing it from the organ list. The system extracts the corresponding 3D surface model and volumetric rendering data based on the segmentation and annotation results from step S2, rendering only the 3D model of the selected organ and hiding other structures. In this mode, users can perform 360-degree rotation observation of the selected organ. Supported rotation operations include free rotation around the three axes of the world coordinate system and local rotation around the organ's own geometric center, with a rotation accuracy of 0.1 degrees. Zooming is also supported, ranging from 0.1 times to 10 times the original size, to facilitate observation of the organ's fine structural features.

[0040] The second operating mode is a multi-structure combination transparent overlay display. Users can simultaneously select multiple organ structures for combined display. The opacity of each structure is controlled by the transfer function in step S3, and users can adjust the visibility weight of each structure in real time using a slider. Preferably, the system provides several preset combination display schemes, such as skeletal-muscle combined display, cardiovascular system display, digestive system display, etc. Each scheme has preset opacity configuration parameters for relevant organs, and teachers can directly call the preset schemes to quickly display the anatomical structure combination of a specific system.

[0041] The third operating mode is virtual anatomical layer-by-layer dissection display, which is the core innovative function of this invention in terms of interactive visualization. The virtual anatomical layer-by-layer dissection operation simulates the process of gradually dissecting tissue from superficial to deep layers in real anatomical operations. This invention defines five anatomical levels from the outside in: Level 1 is the skin surface layer, corresponding to the skin and subcutaneous fat structures in the segmentation category; Level 2 is the superficial fascia layer, corresponding to the superficial fascia and superficial vascular and nerve structures; Level 3 is the muscle layer, corresponding to skeletal muscle groups and intermuscular septa structures; Level 4 is the deep fascia layer, corresponding to the deep fascia and organ capsule structures; Level 5 is the deep structural layer, corresponding to the bone, deep blood vessels, nerve trunks, and internal organ structures.

[0042] The virtual stripping operation is implemented as follows: when the user selects to strip the first... When assigning levels, the system will assign levels 1 to 2. Visibility weights of all organ structures corresponding to the level Set it to 0 (i.e., completely transparent), and at the same time set the number of... Organ structures up to level 5 remain visible or are displayed at a preset opacity. The dissection process supports two modes: progressive dissection (dissecting one level at a time) and skip-dissection (directly jumping to a specified level). In a preferred embodiment of the invention, the dissection process is also accompanied by a transition animation effect. When dissecting a certain level, the opacity of the corresponding structure at that level linearly changes from its current value to 0 within 0.5 seconds, providing a more intuitive visual experience of layered dissection. The technical advantage of this virtual layer-by-layer dissection function is that students can reveal the internal structures of the human body layer by layer, just as in a real anatomical operation, establishing a spatial relationship recognition from shallow to deep layers. At the same time, because it is a digital operation, the dissected levels can be restored at any time for repeated observation, overcoming the irreversible limitation of traditional cadaver dissection.

[0043] The fourth operating mode is arbitrary spatial section display. Users can set virtual sections at any spatial location on the 3D model by dragging and dropping. The direction and position of the normal vector of the section can be adjusted in real time. After the section passes through the 3D model, the structure on one side of the section is clipped and hidden, and the original tomographic image or the superimposed image of the segmentation and annotation results at the corresponding location is displayed at the section. This section display function allows students to intuitively compare the 3D spatial structure with the corresponding 2D tomographic image, thereby effectively establishing the spatial correspondence between the tomographic image and the 3D structure. Preferably, the system supports a three-orthogonal section mode, that is, simultaneously displaying three orthogonal sections: the horizontal plane (cross section), the sagittal plane, and the coronal plane. It is used in conjunction with the 3D model for four-view linkage display. When the user adjusts the position of the section in any view, the other three views are updated synchronously in real time.

[0044] Step S5: Anatomical Knowledge Graph-Driven Annotation. This step aims to automatically generate annotation information such as the name, blood supply source, nerve supply, and adjacent relationships of each anatomical structure in the 3D model based on the pre-constructed anatomical knowledge graph, the segmentation and annotation results of Step S2, and the spatial location information of the 3D model in Step S3. In anatomy teaching, students not only need to observe the morphology of anatomical structures, but also need to understand the name, function, blood supply, and nerve supply of each structure. Therefore, automated annotation is of great significance for improving the information richness of teaching resources.

[0045] In one embodiment of the present invention, the anatomical knowledge graph is modeled and stored using a property graph data structure. Nodes in the graph represent anatomical structures. In this embodiment, approximately 2800 anatomical structure nodes are included, covering the major anatomical structures of all systems in the body. Each node contains the following attribute information: structure name (including the standard Chinese name and the Latin international anatomical name), the anatomical system to which it belongs (e.g., the musculoskeletal system, digestive system, circulatory system, etc.), blood supply source (names of supplying arteries and their branching relationships), venous return pathway, nerve innervation (names of innervating nerves and the source of spinal cord segments), and a description of the standard anatomical location. Edges in the graph represent relationships between anatomical structures. Relationship types include: adjacency (two structures are spatially adjacent), containment (one structure is enclosed by or located within another structure), and functional association (two structures are functionally closely related, such as the bone to which the origin and insertion points of a muscle are attached).

[0046] The automatic annotation process is implemented as follows: First, based on the segmentation annotation results of step S2, the category label and spatial location information of each segmented region are obtained. The spatial location information includes the three-dimensional centroid coordinates, spatial occupancy range, and spatial topological relationship with adjacent segmented regions of the segmented region. Then, a mapping relationship from the segmented regions to knowledge graph nodes is constructed. In a preferred embodiment of the present invention, this mapping is implemented through a two-level matching strategy: the first level is coarse matching based on category labels, which searches for the corresponding set of structural nodes in the knowledge graph according to the category labels of the segmented regions (such as "liver," "spleen," etc.); the second level is fine matching based on spatial location, which compares the three-dimensional centroid coordinates of the segmented regions with the pre-stored standard spatial location ranges of each structure in the knowledge graph, and selects the target node with the most matching spatial location from the coarse matching results.

[0047] After mapping is completed, the system extracts attribute information from the target node and automatically generates the following annotations: Name annotation, which displays the structure name as a text label at a prominent location on the surface of the corresponding organ in the 3D model; Blood supply annotation, which establishes a visual association between the target organ and its blood supply artery in the form of colored lines, with red representing arterial blood supply and blue representing venous return; Nerve innervation annotation, which establishes a visual association between the target organ and its innervating nerve in the form of yellow lines; Adjacency annotation, which automatically highlights its adjacent structures and displays the spatial adjacency relationship with semi-transparent lines when the user selects a structure. The specific process of generating the colored lines in the above blood supply and nerve innervation annotations is as follows: First, based on the target organ node determined by the above two-level matching strategy, the blood supply source attribute (including the name of the supplying artery) and nerve innervation attribute (including the name of the innervating nerve) of the node are read from the anatomical knowledge graph; then, using the read blood supply artery name or innervating nerve name as the search keyword, the segmentation regions with the category label of blood vessel or nerve and corresponding to the same anatomical structure name after two-level matching are searched in all segmentation regions output in step S2, thereby obtaining the three-dimensional surface model of the supplying artery or innervating nerve constructed in step S3; next, the nearest surface point pair between the three-dimensional surface model of the target organ and the three-dimensional surface model of the supplying artery or innervating nerve is calculated, that is, in the target organ... The system iterates through the triangular mesh vertex set of the organ and the triangular mesh vertex set of the supplying artery or innervating nerve, calculating the Euclidean distance and selecting the pair of vertices with the smallest distance as organ-side anchor points and blood vessel-side anchor points or nerve-side anchor points, respectively. Finally, a cubic Bézier curve is generated using these two anchor points as endpoints to serve as the annotation line. The two intermediate control points of the Bézier curve are set at positions offset from their respective anchor points by a preset distance along the surface normal vector direction. This preset distance is one-third of the Euclidean distance between the two anchor points, ensuring the line curves naturally to avoid penetrating the 3D model of other anatomical structures. For blood supply annotation lines, the rendering color is set to red when the supplying artery type is arterial and blue when it is venous return. For nerve innervation annotation lines, the rendering color is uniformly set to yellow. When the segmentation result of step S2 fails to identify an independent segmentation region corresponding to a certain supplying artery or innervating nerve recorded in the knowledge graph, the system degenerates to labeling the name of the blood vessel or nerve on the surface of the target organ using text tags, without generating an annotation line. Preferably, the annotation information supports hierarchical display of detail: brief mode displays only the structure name, standard mode displays the name and blood supply information, and detailed mode displays all attribute information.

[0048] Furthermore, this step also includes an automatic adjacency discovery mechanism based on spatial relationship reasoning. For two spatially adjacent organ regions in the segmentation results of step S2... and If the minimum Hausdorff distance between their 3D surface models is less than a preset adjacent distance threshold (In this embodiment) The threshold of mm is selected based on the fact that there are usually fascial gaps between major anatomical structures in the human body (the gap width is generally no more than 3 mm). If these gaps are not predefined, a spatial adjacency relationship is determined between the two structures, and a corresponding adjacency edge is established in the knowledge graph. This spatial relationship reasoning mechanism can discover adjacency relationships in the knowledge graph that are not yet predefined, thereby enabling dynamic expansion and adaptive improvement of the labeled content.

[0049] See Figure 2 This embodiment provides a three-dimensional visualization and reconstruction system for human sectional anatomical images. This system corresponds one-to-one with each step in the above method embodiment, including a slice registration and grayscale normalization module 1, a multi-organ semantic segmentation module 2, a three-dimensional hybrid reconstruction module 3, an interactive visualization module 4, and an anatomical annotation module 5. In one embodiment of the present invention, the system is deployed on a high-performance workstation equipped with an NVIDIA RTX 4090 graphics processor (24 GB VRAM), an Intel Core i9-13900K processor (5.8 GHz), and 64 GB of DDR5 memory. The operating system is Ubuntu 22.04, the deep learning framework is PyTorch 2.0, and the three-dimensional rendering framework uses OpenGL 4.6.

[0050] The inter-layer registration and grayscale normalization module 1 is configured to acquire a sequence of continuous tomographic images of the human body, perform rigid registration based on mutual information on adjacent layers to correct inter-layer translational and rotational deviations to less than a preset pixel threshold, and perform global grayscale normalization on the registered images. This module receives segmentation quality feedback signals from the multi-organ semantic segmentation module 2. When the feedback signal indicates that the continuity of the segmentation boundary of a certain layer is below a threshold, the module automatically expands the registration search range for the corresponding layer region and performs local fine-tuning registration, as detailed in step S1 of the method embodiment. In this system embodiment, the processing performance of this module is as follows: for a single-layer 512×512 pixel CT image, the single-layer registration processing time is approximately 0.2s; for a 3072×2048 pixel visualized human dataset image, the single-layer registration processing time is approximately 1.5s.

[0051] The multi-organ semantic segmentation module 2 is configured to receive the normalized image sequence output by the inter-layer registration and gray-level normalization module 1, and output multi-class pixel-level segmentation annotation results through the encoder-decoder-based 3D semantic segmentation network detailed in step S2 of the method embodiment. This module internally includes an anatomical location attention module, which uses a prior spatial probability map of human anatomy to weightedly fuse decoded features to improve segmentation accuracy. Simultaneously, this module calculates the continuity evaluation value of the segmentation boundary and sends a segmentation quality feedback signal to the inter-layer registration and gray-level normalization module 1, achieving closed-loop collaborative optimization between modules. In this system embodiment, the processing time for full-sequence segmentation of a complete visualized human dataset containing approximately 3600 layers is approximately 45 minutes, of which network inference time is approximately 35 minutes and post-processing (including connected component analysis, small region filtering, and boundary smoothing) time is approximately 10 minutes.

[0052] The 3D hybrid reconstruction module 3 is configured to receive the segmentation annotation results output by the multi-organ semantic segmentation module 2 and perform surface-volume hybrid 3D reconstruction according to the method detailed in step S3 of the method embodiment. The improved MarchingCubes algorithm submodule of this module is responsible for extracting adaptive isosurface triangular meshes from the segmentation masks of each organ and performing Taubin filtering smoothing and curvature adaptive mesh subdivision; the ray projection volume rendering submodule is responsible for constructing a volumetric rendering model based on the organ category opacity transfer function; and the model fusion submodule is responsible for performing depth fusion rendering of the surface model and the volumetric rendering model in a unified coordinate system. In this system embodiment, taking approximately 15 organ categories of the whole body as an example, the complete surface reconstruction processing time is approximately 8 minutes, and the total number of triangular mesh faces generated is approximately 35 million; the construction of the volumetric rendering model is processed in real time, and the rendering frame rate reaches 45 FPS on the aforementioned hardware platform.

[0053] The interactive visualization module 4 is configured to construct a real-time interactive visualization scene based on the fused 3D model output by the 3D hybrid reconstruction module 3. This module supports the four core interactive operation modes detailed in step S4 of the method embodiment: single organ extraction and display, multi-structure transparent overlay display, virtual anatomical layer-by-layer peeling display, and arbitrary spatial section display. Preferably, this module also provides teaching assistance functions, including operation recording and playback functions (teachers can record the operation process as a script file for students to self-study and replay), a perspective bookmark function (which can save and quickly restore specific perspectives and display configurations), and measurement tools (supporting distance and angle measurement between two points). In this system embodiment, the response latency of the interactive operation is less than 33ms (i.e., the frame rate is not less than 30 FPS), ensuring the smoothness and real-time performance of the user's interactive operation.

[0054] The anatomical annotation module 5 is configured to automatically generate name annotations, blood supply annotations, nerve innervation annotations, and adjacency annotations for each anatomical structure based on a pre-constructed anatomical knowledge graph and the outputs of the multi-organ semantic segmentation module 2 and the 3D hybrid reconstruction module 3, following the two-level matching strategy and spatial relationship reasoning mechanism detailed in step S5 of the method embodiment. This module also supports the export of annotation information, allowing the annotation results to be exported in JSON format for use by other anatomy teaching software. In this system embodiment, the automatic annotation and matching processing time for approximately 2800 anatomical structure nodes throughout the body is approximately 15 seconds, and the annotation accuracy rate (i.e., the ratio of correct matching to corresponding knowledge graph nodes) is approximately 94%.

[0055] The data flow and collaborative relationships between the modules are as follows: the output data of the inter-layer registration and grayscale normalization module 1 is transmitted to the multi-organ semantic segmentation module 2 via the data bus as segmentation input; the segmentation annotation results of the multi-organ semantic segmentation module 2 are simultaneously transmitted to the 3D hybrid reconstruction module 3 and the anatomical annotation module 5; the fused 3D model of the 3D hybrid reconstruction module 3 is transmitted to the interactive visualization module 4 and the anatomical annotation module 5; the annotation results of the anatomical annotation module 5 are transmitted to the interactive visualization module 4 for visualization overlay display. Furthermore, the multi-organ semantic segmentation module 2 sends segmentation quality feedback signals to the inter-layer registration and grayscale normalization module 1 through a feedback channel. This deep coupling and closed-loop feedback architecture between modules enables the overall system performance to surpass the simple overlay effect of each module running independently. It should be noted that the processing power and hardware configuration of the above modules can be flexibly adjusted according to the needs of the actual teaching scenario. For example, in offline batch processing scenarios with lower processing speed requirements, the hardware configuration can be appropriately reduced to lower system deployment costs; in classroom demonstration scenarios with higher real-time interactive performance requirements, the rendering frame rate can be improved by increasing the number of GPUs or using higher-specification GPUs. The system architecture design of this invention has good modularity and scalability. The modules communicate with each other through standardized data interfaces, which makes it easy to replace or upgrade individual modules as needed without affecting the normal operation of other modules.

[0056] To verify the technical effectiveness of the method and system described in this invention, a systematic performance evaluation was conducted under the following test conditions. The test dataset consists of two parts: the first part is three sets of complete human frozen section image sequences from the Chinese Visual Human Body Dataset, each set containing approximately 3600 slices, with a slice thickness of 0.5 mm and a resolution of 3072×2048 pixels; the second part is 20 sets of clinical whole-body CT scan data from the radiology department of a tertiary hospital, with a slice thickness of 1 mm and a resolution of 512×512 pixels, covering the area from the skull to the pelvis. All test data are equipped with gold standard annotations for multi-organ segmentation, manually annotated by three anatomical experts with over 10 years of experience. The test hardware platform is consistent with that described in the system embodiment, using a high-performance workstation equipped with an NVIDIA RTX4090 graphics card.

[0057] Regarding multi-organ segmentation accuracy, the method of this invention achieves an average Dice coefficient of 0.92 on the visualized human dataset, with Dice coefficients of 0.96 for skeletons, 0.91 for muscles, 0.87 for blood vessels, 0.83 for nerves, and 0.93 for internal organs. On the clinical CT dataset, the average Dice coefficient is 0.90, and the segmentation accuracy for each category is basically consistent with the results on the visualized human dataset. Compared to the baseline model without the anatomical location attention module, the overall segmentation Dice coefficient is improved by 0.04 after introducing the attention module, with an improvement of 0.07 for small-volume structures such as nerves and small blood vessels. This improvement indicates that the anatomical prior spatial probability map has a significant guiding effect on the segmentation of small target organs. Compared to the version without the closed-loop feedback mechanism, the overall Dice coefficient is further improved by 0.03 after introducing feedback, and the proportion of discontinuous layers at the segmentation boundary decreases from 4.2% to 0.8%, verifying that the closed-loop collaborative mechanism between steps S2 and S1 can effectively eliminate the segmentation quality degradation caused by registration bias.

[0058] Regarding the quality of 3D reconstruction, the average Hausdorff distance between the surface model constructed by the method of this invention and the gold standard 3D model is 0.8 mm, and the average surface distance is 0.3 mm. Comparison with the tetrahedral voxel fitting method used in CN110570515B in terms of bone reconstruction accuracy shows that the average Hausdorff distance of the bone surface constructed by the method of this invention is 0.5 mm, which is better than the 1.2 mm of the CN110570515B method. It is particularly noteworthy that the CN110570515B method can only handle single bone structures and requires manual setting of the segmentation threshold, while the method of this invention can automatically reconstruct more than 15 anatomical structure types simultaneously. Regarding adaptive isosurface thresholding, compared with the standard Marching Cubes algorithm using a fixed global threshold, the adaptive thresholding method of this invention reduces the average Hausdorff distance of soft tissue organ surface extraction from 1.5 mm to 0.9 mm, a reduction of 40%. Furthermore, the combined use of Taubin filtering and curvature-adaptive mesh subdivision improves the surface detail retention of high-curvature areas such as joint surfaces and blood vessel branch points by approximately 35%, while keeping the total number of mesh faces below 35 million, ensuring the feasibility of subsequent real-time rendering.

[0059] Regarding volumetric rendering quality, the organ-category-based opacity transfer function proposed in this invention, compared to the traditional single transfer function based on grayscale values, can effectively distinguish different organ types with overlapping grayscale values. Experimental results show that between the liver and adjacent organs with similar grayscale values, such as the stomach and right kidney, the category-based transfer function can assign different opacity mappings through segmentation annotation results, making the boundaries of each organ clearly distinguishable in the volumetric rendering view. In contrast, traditional methods cannot achieve effective organ differentiation within the same grayscale range.

[0060] Regarding interactive visualization performance, on the aforementioned hardware platform, the rendering frame rate of interactive scenes containing a 3D model of multiple organs throughout the body (approximately 35 million triangular facets) remained stable between 35 and 50 FPS. The response latency for virtual anatomical layer-by-layer dissection operations was less than 20ms, and the response latency for arbitrary section setting operations was less than 30ms, both meeting the requirements for smooth real-time interaction. In terms of annotation module performance, the automatic annotation matching processing time for approximately 2800 anatomical structure nodes throughout the body was approximately 15 seconds, with an annotation accuracy of approximately 94%, including 97% accuracy for name annotation, 92% for blood supply annotation, 90% for nerve innervation annotation, and 93% for adjacency relationship annotation.

[0061] Regarding its effectiveness in teaching applications, the system of this invention underwent a semester-long trial in the Department of Anatomy at a medical college. The 120 undergraduate medical students participating in the trial were divided into an experimental group (using the system of this invention to assist teaching) and a control group (using only traditional teaching methods), with 60 students in each group. The end-of-semester practical anatomy assessment showed that the experimental group students scored an average of 82.5 points (out of 100) on the sectional anatomy identification questions, significantly higher than the control group's 71.3 points, with a statistically significant difference. Particularly in questions involving the course of blood vessels and nerves and the spatial relationships of multiple organs, the experimental group students achieved an accuracy rate approximately 18 percentage points higher than the control group, indicating that the three-dimensional visualization and virtual dissection functions have a significant auxiliary teaching effect on establishing spatial relationship cognition. Questionnaire results showed that 87% of the experimental group students considered the virtual layer-by-layer dissection function "very helpful" in understanding anatomical hierarchies, 92% of the students believed that the anatomical annotation function significantly reduced the time required to memorize anatomical terms, and 85% of the students expressed a desire to continue using the system as an auxiliary learning tool in subsequent courses.

[0062] In summary, the three-dimensional visualization reconstruction method and system for human anatomical cross-section images provided by this invention have achieved the expected technical goals in terms of multi-organ segmentation accuracy, three-dimensional reconstruction quality, interactive visualization performance, and teaching application effect, providing high-quality, reusable digital three-dimensional anatomical resources for anatomy teaching in medical colleges.

[0063] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for three-dimensional visualization and reconstruction of human sectional anatomical images, characterized in that, Includes the following steps: Step S1: Obtain a sequence of continuous tomographic images of the human body, perform rigid registration based on mutual information on adjacent layer images to correct the inter-layer translational and rotational deviations to less than a preset pixel threshold, and perform global grayscale normalization processing. Use the grayscale distribution deviation between the current layer and the reference layer as a consistency feedback signal to dynamically adjust the registration parameters of subsequent layers. Step S2: Input the normalized image sequence into a 3D semantic segmentation network based on an encoder-decoder architecture. The encoder extracts multi-scale features layer by layer. The decoder introduces an anatomical position attention module and uses the human anatomy prior spatial probability map to weightedly fuse the decoded features to output multi-class pixel-level segmentation annotation results. The anatomical structure categories of the segmentation annotation results include bones, muscles, blood vessels, nerves, and internal organs. The specific implementation of the anatomical position attention module is as follows: multiply the spatial probability distribution map of each organ in the human anatomy prior spatial probability map with the feature map of the corresponding scale in the decoder path channel by channel to obtain a weighted feature map. Then, the weighted feature map and the original decoded feature map are concatenated along the channel dimension and fused by a convolutional layer to output refined segmentation features with anatomical position awareness. The continuity evaluation value of the partition boundary is obtained by calculating the Dice coefficient of the partition boundary between adjacent layers. The formula for calculating the continuity evaluation value of the segmentation boundary is: ， in: The number of categories to be divided; and The first Layer and first The first in the layer The set of pixels representing the segmentation boundaries of organoids; This represents the symmetric difference operation; Represents the number of pixels in the set; when the segmentation boundary continuity evaluation value When the quality falls below a preset quality threshold, a feedback signal is generated, triggering step S1 to process the first... Local fine-grained registration was performed on five layers of images, including the first layer and two layers before and after it; Step S3: The organ segmentation masks output in step S2 are used to extract isosurface triangular meshes to construct a 3D surface model. The improved Marching Cubes algorithm includes adaptively determining the isosurface threshold for different organ types. The formula for calculating the adaptive isosurface threshold is: ， in: For the first Organoid segmentation mask The mean gray value of all voxels within the cell; For the first Organoid segmentation mask Mean value of gray-level gradient magnitude at the boundary voxel; This is the category-related moderating factor, and its value range is... For hard tissues like bones Positive values ​​are used to increase the threshold, especially for soft tissues such as internal organs. Negative or zero values ​​are used to reduce the threshold; at the same time, a volumetric rendering model is constructed using a ray projection volume rendering method based on the organ category opacity transfer function, and adaptive mesh subdivision based on surface curvature is performed and fused in a unified coordinate system. Step S4: Construct a real-time interactive visualization scene based on the fused 3D model, supporting single organ extraction and display, multi-structure transparent overlay display, virtual anatomical layer-by-layer peeling display, and arbitrary spatial section display. The layer depth sequence of the virtual anatomical layer-by-layer peeling display includes five anatomical levels defined from the outside to the inside: skin surface, superficial fascia layer, muscle layer, deep fascia layer, and deep structural layer. Each anatomical level corresponds to at least one set of organ structures of at least one segmentation category in step S2. The layer-by-layer peeling removes the superficial structure in sequence according to the anatomical layer depth sequence to expose the deep structure. Step S5: Based on the pre-constructed anatomical knowledge graph containing anatomical structure names, blood supply sources, nerve innervation, and adjacent relationships, and combining the segmentation and annotation results output in Step S2 with the spatial location information of the three-dimensional model output in Step S3, the matching anatomical attributes are queried, and name annotations, blood supply annotations, nerve innervation annotations, and adjacent relationship annotations are automatically generated. The generation method of the colored connecting lines in the blood supply annotations and nerve innervation annotations is as follows: Based on the target organ node, the name of the blood supply artery and the name of the innervating nerve of the node are read from the anatomical knowledge graph. The read blood supply artery name or innervating nerve name is used as the search keyword. In all segmentation regions output in Step S2, segmentation regions with the category label of blood vessel or nerve and corresponding to the same anatomical structure name are searched to obtain the three-dimensional surface model of the blood supply artery or innervating nerve constructed in Step S3. The nearest surface point pair between the three-dimensional surface model of the target organ and the three-dimensional surface model of the blood supply artery or innervating nerve is calculated as the organ-side anchor point and the blood vessel-side anchor point or the nerve-side anchor point. A cubic Bézier curve is generated with the two anchor points as endpoints as the annotation connecting lines.

2. The method according to claim 1, characterized in that, In step S1, the preset pixel threshold specifically means that the translation deviation is less than 1 pixel and the rotation deviation is less than 0.5 degrees; the global grayscale normalization process adopts the histogram matching method, taking the intermediate layer image of the continuous tomographic image sequence as the reference layer, and mapping the grayscale values ​​of the remaining layers to the grayscale distribution range consistent with the reference layer.

3. The method according to claim 1, characterized in that, In step S1, the sources of the continuous tomographic image sequence include frozen section images and clinical CT or MRI scan images from the Chinese Visual Human Dataset. The slice thickness of the tomographic images ranges from 0.5 mm to 2 mm, and the planar resolution is not less than 512 × 512 pixels.

4. The method according to claim 1, characterized in that, In step S2, the preset quality threshold is 0.85, that is, when the segmentation boundary continuity evaluation value is... The feedback signal is triggered when the value is below 0.

85.

5. The method according to claim 1, characterized in that, In step S3, Taubin filtering is performed on the isosurface triangular mesh to eliminate the staircase effect while preserving the geometric features of the anatomical structure.

6. The method according to claim 1, characterized in that, In step S3, the ray projection volume rendering method uses an opacity transfer function based on organ category, for grayscale values... The sampling points, if their organ category is The opacity of that sampling point Determined by the following transfer function: ， in: For the first The basic opacity coefficient of organoids has a range of values. ; For the first The central value of the grayscale distribution of organoids is taken as the median of the grayscale values ​​within the segmented region of that organoid. For the first Standard deviation of the grayscale distribution of organoids; This is a user-adjustable organ visibility weighting factor, with a value range of [value range missing]. The opacity transfer function supports user-interactive adjustment to dynamically change the transparency of each organ.

7. The method according to claim 1, characterized in that, In step S5, the two intermediate control points of the Bezier curve are respectively set at a position offset by a preset distance from their respective anchor points along the surface normal vector direction. The preset distance is one-third of the Euclidean distance between the two anchor points, so that the connecting line is naturally curved to avoid penetrating into the three-dimensional model of other anatomical structures. For the blood supply labeling line, the rendering color is set to red when the blood supply artery type is arterial and to blue when it is venous return. For the nerve innervation labeling line, the rendering color is uniformly set to yellow.

8. The method according to claim 1, characterized in that, Step S5 also includes an automatic adjacency discovery mechanism based on spatial relationship reasoning: for two spatially adjacent organ regions in the segmentation results of step S2... and If the minimum Hausdorff distance between their 3D surface models is less than a preset adjacent distance threshold If a spatial adjacency relationship exists between two structures, then a corresponding adjacency edge is established in the anatomical knowledge graph.

9. The method according to claim 1, characterized in that, In step S5, the construction of the anatomical knowledge graph is based on a standardized human anatomical terminology system. The nodes in the anatomical knowledge graph represent anatomical structural entities. The node attributes include the structural name, the system to which it belongs, the blood supply arteries and veins, and the nerve innervation information. The edges between nodes represent adjacent relationships, containment relationships, or functional association relationships.

10. A three-dimensional visualization and reconstruction system for human sectional anatomical images, used to implement the method described in any one of claims 1-9, characterized in that, include: The inter-layer registration and gray-level normalization module is configured to acquire a sequence of continuous tomographic images of the human body, perform rigid registration based on mutual information on adjacent layer images to correct inter-layer translational and rotational deviations, perform global gray-level normalization on the registered images, and dynamically adjust the registration parameters according to the gray-level distribution consistency feedback signal. The multi-organ semantic segmentation module is configured to receive the normalized image sequence output by the inter-layer registration and grayscale normalization module, and output multi-class pixel-level segmentation annotation results through a 3D semantic segmentation network based on an encoder-decoder architecture. The decoder path of the 3D semantic segmentation network includes an anatomical location attention module. The anatomical location attention module outputs refined segmentation features by multiplying the spatial probability distribution map of each organ in the prior spatial probability map of human anatomy with the feature map of the corresponding scale in the decoder path channel by channel to obtain a weighted feature map, and then concatenating it with the original decoded feature map along the channel dimension for dimensionality reduction and fusion. The module also outputs refined segmentation features based on the continuity evaluation value of the segmentation boundary. A segmentation quality feedback signal is sent to the inter-layer registration and grayscale normalization module to trigger local fine registration of the corresponding layer and the two layers before and after it, for a total of five layers. The 3D hybrid reconstruction module is configured to receive the segmentation annotation results output by the multi-organ semantic segmentation module, construct a 3D surface model using an improved Marching Cubes algorithm with adaptive determination of isosurface thresholds for different organ types, and construct a volumetric rendering model using a ray projection volume rendering method based on organ category opacity transfer function, and fuse the two in a unified coordinate system. The interactive visualization module is configured to construct a real-time interactive visualization scene based on the fused 3D model output by the 3D hybrid reconstruction module. It supports virtual anatomical layer-by-layer peeling and arbitrary angle rotation observation operations, which can be achieved by extracting a single organ, transparently superimposing multiple structures, and defining five anatomical levels from the outside to the inside: the skin surface layer, superficial fascia layer, muscle layer, deep fascia layer and deep structural layer. The anatomical annotation module is configured to automatically generate name annotations, blood supply annotations, nerve innervation annotations, and adjacency annotations for each anatomical structure based on a pre-built anatomical knowledge graph and the output results of the multi-organ semantic segmentation module and the three-dimensional hybrid reconstruction module. The lines in the blood supply annotations and nerve innervation annotations are generated using cubic Bézier curves with the nearest surface point pair between the three-dimensional surface model of the target organ and the three-dimensional surface model of the blood supply artery or innervation nerve as the endpoints.

Images