Calcaneal fracture AI reduction method based on multi-modal data

Abstract

The application provides a calcaneus fracture AI reduction method based on multi-modal data, and relates to the technical field of medical image processing. The method obtains CT voxel data, separates and extracts the sustentaculum tali center as a static space registration reference origin point; maps the healthy side topological atlas based on the origin point and calculates a three-dimensional space transformation matrix of the fracture fragments; performs point cloud reconstruction and orthogonal projection dimension reduction, extracts key peak and valley landmark points to generate a visual measurement parameter vector group; encapsulates the vector group and the space matrix as a feature set, and inputs the comprehensive evaluation module of a medical rule knowledge base; when the comparison deviation is out of limit, a space displacement compensation parameter constrained by tension and interference boundary is generated and reverse injection is performed to execute quantitative correction. The method effectively overcomes the limitations of traditional registration divergence drift and lack of mechanical verification closed loop, and realizes a reduction with both anatomical angle accuracy and internal load stability.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an AI-based reduction method for calcaneal fractures based on multimodal data. Background Technology

[0002] As computer-aided decision-making systems continue to evolve towards high-precision three-dimensional medical image processing, virtual surgical planning for orthopedic trauma has moved from basic two-dimensional geometric measurements to the spatial calculation stage of complex multi-dimensional topological structures. In the field of calcaneal comminuted fractures, which involve high anatomical complexity and biomechanical load-bearing requirements, the ability to reconstruct the absolute spatial position of anatomical structures within a digital twin environment using machine vision and spatial registration algorithms has become a key technological pivot for determining postoperative stress states and joint function recovery.

[0003] When dealing with complex free fragments in calcaneal fractures, traditional image registration algorithms, lacking an absolutely static reference point in the physical world, are prone to systematic global coordinate system divergence during iterative optimization between fragments after irregular three-dimensional displacement and spatial topological deconstruction of the fracture fragments. This leads to macroscopic spatial drift in the post-reduction 3D model. Existing algorithms heavily rely on single-modal geometric contour surface stitching and lack automated visual measurement mechanisms to extract key parameters representing anatomical knowledge after reduction. Furthermore, they lack mechanisms for multi-dimensional joint evaluation and closed-loop fine-tuning of "macroscopic geometric angles" and "internal microscopic load-bearing bone fit" by combining expert knowledge bases. For example, Chinese patent application CN121685877A proposes a fracture intelligent reduction navigation evaluation and analysis method and system based on multimodal imaging, which constrains the reduction solution through bone fragment fissure fiber bundle modeling and multiple convolution calculations. Although the scheme provides some support for the calculation of non-penetration and normal continuity of local fracture ends, it still faces the challenge of spatial coordinate divergence of free bone blocks in the absence of global absolute static anatomical anchor points due to its reliance on local topological constraint mechanism of fracture end pairing. Furthermore, it urgently needs to be supplemented with an automated closed-loop correction capability based on biomechanical boundaries (such as ligament tensile tension) and clinical weight-bearing contact rate. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an AI-based reduction method for calcaneal fractures based on multimodal data. Inspired by the physical constraint mechanism of the calcaneal-talar interosseous ligament, the method extracts the center of the suspensory plexus as the origin of the global static spatial registration reference, constructing an absolutely stable anti-drift anchoring reference. A multi-source heterogeneous feature alignment closed loop is introduced, deeply parametrically fusing the visual measurement parameters (geometric angles) derived from 3D reconstruction projection with the proportion of high-density cortical contact at the fracture ends (load-bearing mechanics). This data is then fed into a pre-built medical rule knowledge base for multi-dimensional rule comparison and fine-tuning compensation. This addresses the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An AI-based reduction method for calcaneal fractures based on multimodal data, executed by a medical image processing system, includes the following steps: Step S1: Obtain multimodal CT voxel data with gray values ​​from the affected and healthy sides of the patient; the image processing module calls the multi-label image segmentation model to separate the calcaneal anatomical partition identifier set containing multiple anatomical structures, and extracts the three-dimensional spatial center position of the suspensory process anatomical label that is relatively stable due to ligament constraint from the calcaneal anatomical partition identifier set, and uses it as the static spatial registration reference origin for registration. Step S2: Based on the static spatial registration reference origin, generate a mirror topology map of the healthy side, and execute an iterative nearest point spatial alignment model to extract the three-dimensional spatial transformation matrix of each free fracture fragment aligned to the mirror topology map; Three-dimensional point cloud reconstruction is performed based on the three-dimensional spatial transformation matrix, and projection dimension reduction is performed on the standard anatomical sagittal projection system to extract the transitional two-dimensional contour feature vector. Step S3: Based on the two-dimensional contour feature vector, trigger the image contour feature extraction module, use the edge contour search algorithm to automatically detect key peak and valley markers, and generate a set of visual measurement parameter vectors containing anatomical angles and size values; The visual measurement parameter vector set and the three-dimensional spatial transformation matrix are uniformly encapsulated into a reset state feature set, which is then input into a comprehensive evaluation module with a pre-configured medical rule knowledge base for multi-dimensional data rule comparison. Step S4: If the deviation calculated by the comprehensive evaluation module exceeds a preset confidence threshold, then receive the fine-tuning strategy containing the spatial attitude fine-tuning intention returned by the comprehensive evaluation module based on the rule comparison result, and parse the fine-tuning strategy to generate spatial displacement compensation parameters; and, The spatial displacement compensation parameters are injected in reverse into the preceding spatial registration node. Quantitative correction is performed on the transformation matrix corresponding to the current calculation frame in the direction of the target alignment deviation until the comparison deviation meets the safety threshold. Finally, a finalized three-dimensional fracture reduction spatial model is output.

[0006] Preferably, the step of projecting and reducing the dimension of the transition to the standard anatomical sagittal projection system to extract the transitional two-dimensional contour feature vector specifically includes: The 3D mesh depth mapping module is invoked to orthogonally project the reconstructed 3D point cloud onto the standard anatomical sagittal projection system; During the projection process, the depth difference data of the three-dimensional point cloud surface relative to the projection reference plane is extracted simultaneously to generate terrain depth array data containing three-dimensional spatial undulation information. The edge contour search algorithm is used to search for depth extreme points in the terrain depth array data to lock the key peak and valley marker points and generate the visual measurement parameter vector set, thereby eliminating the projection deformation error caused by the spatial tilt of the fracture fragments.

[0007] Preferably, the step of uniformly encapsulating the visual measurement parameter vector set and the three-dimensional spatial transformation matrix into a reset state feature set further includes: The gray values ​​representing bone density in the multimodal CT voxel data are pre-statically bound to the corresponding mesh vertices of the three-dimensional point cloud; After performing three-dimensional point cloud reconstruction based on the three-dimensional spatial transformation matrix, the gray values ​​on the splicing section of the free fracture fragments are extracted, filtered by the set bone cortex gray value filtering threshold, and the high-density cortical bone contact ratio score characterizing the load-bearing mechanical stability is calculated. The high-density cortical bone contact ratio score, along with the visual measurement parameter vector set, is encapsulated into the reset state feature set and input into the comprehensive evaluation module. This allows the comprehensive evaluation module to jointly determine the fine-tuning strategy based on both anatomical angle constraints and internal load-bearing contact area.

[0008] Preferably, the step of inputting data into a comprehensive evaluation module based on a medical rule knowledge base for multidimensional data rule comparison specifically includes: A multi-dimensional feature mapping mechanism is constructed within the comprehensive evaluation module, setting the anatomical angle as the geometric appearance optimization target and setting the high-density cortical bone contact ratio score as the load-bearing contact benchmark threshold. The comprehensive evaluation module performs conditional comparison and screening on the reset state feature set. When the anatomical angle deviates from the preset anatomical angle baseline and triggers the stretching compensation rule, and at the same time, it is monitored that the high-density cortical bone contact ratio score is lower than the load-bearing contact benchmark threshold, a cutoff strategy is executed to intercept the correction command for the unidirectional stretching geometry angle. Trigger a sliding rotation search logic toward the Z-axis orthogonal plane to generate a suboptimal fine-tuning strategy that satisfies the load-bearing contact reference threshold.

[0009] Preferably, the step of parsing the returned fine-tuning strategy to generate spatial displacement compensation parameters specifically includes: Establish a three-dimensional spatial interference boundary for the three-dimensional point cloud envelope of free fracture fragments; Extract the spatial attitude fine-tuning intent from the fine-tuning strategy returned by the comprehensive evaluation module, and calculate it as a virtual displacement correction vector applied to the dissecting node; In the spatial interference detection module, the spatial equilibrium state of the free fracture fragment under the action of the virtual displacement correction vector is solved based on the three-dimensional spatial interference boundary. When envelope penetration interference is detected, the translation and rotation increments used to counteract the penetration interference are calculated, and the translation and rotation increments are directly configured as the spatial displacement compensation parameters.

[0010] Preferably, the step of injecting the spatial displacement compensation parameters in reverse into the previous spatial registration node, and performing quantitative correction on the transformation matrix corresponding to the current calculation frame in the target alignment deviation direction, specifically includes: Using the Kalman filter state update matrix equation, the spatial displacement compensation parameter is used as the observation gain and fused with the three-dimensional spatial transformation matrix corresponding to the current calculation frame in the preceding spatial registration node; In the direction of target alignment deviation, compare with the pre-stored soft tissue tension constraint baseline, strip the abnormal translation components in the matrix that exceed the physical tensile limit of the free fracture fragment ligament, generate a smooth correction transformation matrix that has been smoothed based on displacement extrema, and use the smooth correction transformation matrix to cover and update the current transformation matrix.

[0011] Preferably, the method further includes the step of synchronously triggering a surgical difficulty warning process based on spatial registration nodes: During the cyclic execution of quantitative correction, the spatial registration time and alignment deviation sequence for generating the three-dimensional spatial transformation matrix from the preceding spatial registration node are simultaneously extracted. Based on the spatial registration time and alignment deviation sequence, a resetting and alignment difficulty index, which characterizes the resetting and alignment difficulty, is extracted. When the state of the repositioning and alignment difficulty index exceeding the preset alignment failure threshold is detected to meet the preset maximum number of consecutive iterations parameter, the system abnormal interception strategy is triggered, automatically blocking the link to request the default geometric repositioning calculation from the comprehensive evaluation module, and directly generating an early warning signal for prompting alternative surgical planning based on the current free fracture fragment distribution state, and sending the alternative surgical planning early warning signal to the display and control terminal to trigger the interactive rendering closed loop for prompting alternative intervention strategies.

[0012] Preferably, the step of triggering a sliding rotation search logic in the Z-axis orthogonal plane to generate a suboptimal fine-tuning strategy that satisfies the load-bearing contact reference threshold specifically includes: On the Z-axis orthogonal plane, with the centroid of the point cloud of the current free fracture fragment as the spatial reference origin, a two-dimensional sliding search grid covering the preset search range is constructed according to the preset sampling step size parameter; Traverse each discrete grid node in the two-dimensional sliding search grid and call the preset two-dimensional steady-state weighted aggregation rule to calculate the reset steady-state evaluation value corresponding to each discrete grid node; The execution logic of the two-dimensional steady-state weighted aggregation rule is as follows: extract the normalized high-density cortical bone contact ratio score of the location of the discrete grid node, and perform a subtraction operation between the score and the anatomical angle deviation value carrying a preset convergence penalty weight to generate the reset steady-state evaluation value. The target grid node whose reset steady-state evaluation value satisfies the preset steady-state convergence extreme value condition is extracted. The relative offset coordinates of the target grid node with respect to the spatial reference origin are analyzed into translation and rotation mapping vectors. The translation and rotation mapping vectors are used as the feedback of the suboptimal fine-tuning strategy to the preceding spatial registration node to perform subsequent quantitative correction steps.

[0013] Preferably, the step of comparing the abnormal translational components in the stripping matrix that exceed the physical tensile limit of the ligaments in the free fracture fragments with a pre-stored soft tissue tension constraint baseline specifically includes: Based on the three-dimensional spatial transformation matrix corresponding to the current calculation frame in the preceding spatial registration node, the three-dimensional relative displacement vector of the free fracture fragment relative to the static spatial registration reference origin is extracted. The three-dimensional non-uniform tensile limit envelope surface bound to the corresponding calcaneal anatomical partition is retrieved as the soft tissue tension constraint baseline, and the three-dimensional spatial interference penetration amount between the end coordinate point of the three-dimensional relative displacement vector and the three-dimensional non-uniform tensile limit envelope surface is calculated using three-dimensional mesh collision detection logic. If it is determined that the end coordinate point breaks through the three-dimensional non-uniform stretch limit envelope, then the spatial distance from the end coordinate point back to the envelope along the shortest normal path is calculated in three-dimensional space, and a boundary projection convergence vector is generated based on the spatial distance and the direction of the shortest normal path. The boundary projection convergence vector is used to perform inverse numerical compensation on the current three-dimensional relative displacement vector to quantize and remove the abnormal translation components.

[0014] Preferably, in the step where the image processing module calls the multi-label image segmentation model to separate the anatomical partition markers of the calcaneus, when encountering loose bone fragments with blurred edges, a projection allocation routing mechanism based on anatomical anchor points is further executed: For isolated fragment point clouds that cannot be directly mapped to a preset dictionary and contain more voxels than the noise filtering threshold, the output of the multi-label image segmentation model is identified and extracted as bone fragments to be classified. Obtain the centroid of the solid point cloud of the bone fragments to be classified, and extract the fixed feature points at both ends of the calcaneal skeleton to construct a three-dimensional spatial anatomical anchor point vector; Based on the spatial inner product operation, the relative projection distance component of the centroid of the entity point cloud on the three-dimensional spatial anatomical anchor point vector is calculated. The relative projection distance component is mapped to the feature anchor point projection normalization coefficient to characterize the relative topological depth position of the free bone fragments on the main calcaneal morphological axis. Based on the quantized numerical distribution of the feature anchor projection normalization coefficient, the spatial nearest neighbor assignment operation is performed in the corresponding orthogonal coordinate system to attach the bone fragments to be classified to the determined calcaneal anatomical partition marker.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention separates multiple anatomical structures through an image processing module and selectively extracts the anatomical label of the suspensory process, which remains relatively stable due to strong ligament constraints, and establishes it as the origin of the global static spatial registration reference. This avoids the coordinate system divergence distortion problem caused by mutual reference of free fragments in traditional iterative nearest-point algorithms, ensuring that each free fracture fragment has an absolute spatial constraint keyhole when aligned to the healthy side's mirror topological map, thus improving the stability of the three-dimensional spatial transformation matrix calculation.

[0016] This invention projects a 3D point cloud onto a standard anatomical sagittal plane to reduce its dimensionality, extracting peak and valley landmarks to generate visual anatomical angles. Simultaneously, it extracts the cross-section of the fracture ends to calculate the contact ratio of high-density cortical bone. Both are then encapsulated in a comprehensive evaluation module for rule-based comparison. This mechanism overcomes the limitations of existing technologies that rely solely on surface-level single-modal geometric contour stitching, achieving gains in both anatomical appearance restoration (Boehler angle reconstruction) and structural stability (preventing internal bone suspension). This effectively avoids the risk of postoperative mechanical collapse caused by simply pursuing aesthetic fit.

[0017] This invention, while receiving displacement compensation parameters generated by a fine-tuning strategy, simultaneously introduces three-dimensional spatial interference boundary detection and performs Kalman filtering smoothing and stripping on abnormal translational components based on a pre-stored soft tissue tension constraint baseline. This effectively blocks excessive stretching or spatial penetration distortion commands, ensuring that the output reset trajectory conforms to the impenetrable characteristics of rigid bone and the physical yield limit of ligaments, thus improving the feasibility and clinical safety of the virtual reset model in surgical procedures. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the physical space execution architecture and technical route of the present invention; Figure 2 This is a schematic diagram of the results of a numerical verification experiment of the present invention for monitoring the reset alignment difficulty index and abnormal circuit breaking mechanism based on timing consumption and alignment deviation calculation. Figure 3 This is a core principle diagram of the present invention for projection dimension reduction onto a standard anatomical sagittal projection system and generating a set of visual measurement parameter vectors. Figure 4This diagram illustrates the computational mechanism of the present invention, which uses a spatial interferometry detection module to perform three-dimensional physical anti-penetration and soft tissue tension constraint verification. Figure 5 This is a schematic diagram of the technical route of the AI ​​reduction method for calcaneal fractures based on multimodal data according to the present invention. Detailed Implementation

[0019] The technical solutions in 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Please see Figures 1 to 5 The present invention provides a technical solution: The AI-based reduction method for calcaneal fractures based on multimodal data proposed in this invention includes the following: The first step is to acquire multimodal CT voxel data with grayscale values ​​from the affected and healthy sides of the patient; the image processing module calls a multi-label image segmentation model to separate a set of calcaneal anatomical region identifiers containing multiple anatomical structures, and extracts the three-dimensional spatial center position of the suspensory process anatomical label, which is relatively stable due to ligament constraint, from the calcaneal anatomical region identifier set, using it as the static spatial registration reference origin. The specific data processing process is as follows: The system retrieves raw multimodal CT voxel sequences from both the affected and healthy sides of the patient in real time via a trusted standard Medical Digital Imaging and Communication (DICOM) network data interface. In this embodiment, a privacy-based desensitization process is performed on the raw multimodal CT voxel sequences within a pre-defined desensitization module, thereby truncating the patient identification information flow and outputting compliant, anonymized multimodal CT voxel data. This further triggers 3D reconstruction and normalization preprocessing of the image data; specifically, the desensitized voxel sequences are converted into a 3D volumetric data structure with standardized dimensions, and a 3D interpolation algorithm is invoked to perform voxel resampling. During the execution of the 3D interpolation algorithm, the isovoxel resampling scale parameter is simultaneously introduced and read, denoted as... It characterizes the consistency accuracy of the physical spacing of the three-dimensional spatial matrix in the three orthogonal dimensions of X, Y, and Z, aiming to align the non-isotropic voxels output by heterogeneous CT devices (e.g., datasets where the slice thickness spacing is greater than the pixel spacing) to a unified geometric metric benchmark. In the specific configuration of this embodiment, this voxel resampling scale parameter... The value of is limited to a closed interval of [0.5 mm, 1.2 mm], with 1 mm being the preferred value.

[0021] Further, a grayscale mapping range for the bone window is introduced, denoted as... Its characterization can highlight the cutoff range of X-ray physical attenuation values ​​for dense bone cortex. In this embodiment, the preferred range is set to a window level of 400 HU and a window width of 1800 HU. Based on the grayscale mapping range of the bone window, linear filtering of grayscale values ​​and standardization of spatial coordinate system orientation (aligned to the standard RAS system) are performed on global voxels to eliminate dimensional deviations caused by differences in the underlying sensor matrix of heterogeneous CT acquisition devices.

[0022] Before implementing online anatomical separation, this embodiment constructs and calls a multi-label image segmentation model based on a dual-model cascade architecture for offline pre-training and calibration. Specifically, a large image database containing historical reduction control samples is acquired, and the fracture side and healthy side control annotation mapping rules are configured during the model training initialization phase. Specifically, voxel sequences of traumatic fracture states on the affected side of historical patients are extracted from the large image database as the input feature space, and voxel sequences of the healthy side morphology after spatial mirror inversion processing of the same historical patient are extracted as expert-level real anatomical mask labels. The input feature space and expert-level real anatomical mask labels are aligned in a global physical space coordinate system, establishing a one-to-one correspondence mapping relationship between the two in the voxel grid dimension. This constructs supervised training sample pairs for the multi-label image segmentation model to perform error backpropagation and weight iteration calculation in the offline phase. Furthermore, when constructing the offline training mechanism of the core segmentation network, sample data containing desensitized historical fracture CT sequences are extracted. A dual-constraint loss fusion formula is constructed in the error backpropagation stage to overcome network inference chaos caused by the topological similarity within the trabecular bone after trauma. After each forward feature inference, the pixel-level difference between the output mask and the expert-level true label is extracted and fed into the basic pixel classification pipeline to generate the cross-entropy loss benchmark. Simultaneously, it is input into the geometry pipeline to generate the anatomical overlap loss baseline value. Introducing shape edge focusing weights. This represents the gradient descent allocation percentage used in the global network convergence game to specifically penalize physical deviations such as anatomical edge overlap (e.g., sharp-angle deformation of broken bone fragment tips). In this embodiment, this morphological edge focusing weight... The value is set to a constant within the range [0.65, 0.85] (preferably set to 0.75). Furthermore, hematoma artifacts exist in the fracture area; if the morphological edge focusing weight... If the value is below 0.65, the network weight update will be based on the cross-entropy loss benchmark. The dominant factor leads to stable model performance when predicting large main bone blocks, but it directly produces overly smoothed blunting for small, free-floating fragments at the edges. If the morphological edge focusing weight exceeds the upper limit of the interval, the model will fall into a local geometry-sensitive state, easily overfitting to metal artifacts or high-frequency speckle noise in CT imaging, causing gradient explosion or network non-convergence. The global loss is finally extracted through the above game-theoretic calibration. The optimal gradient solution is obtained and fixed into the network weight dictionary to complete the basic pre-setup for online inference.

[0023] In this embodiment, during online execution, the pre-defined dual-model cascaded architecture is invoked to perform progressive inference. Specifically, the normalized voxel data is input into the first-level foot and ankle bone segmentation network, and forward feature inference across the entire anatomical domain is performed to extract the complete binary mask of the calcaneus outer contour, which is then recorded as the complete calcaneus mask parameter. It represents the overall peripheral geometric topological boundary of the calcaneus unaffected by the spatial gaps between free fragments from the fracture. Based on this complete calcaneal mask parameter... A 3D bounding box is calculated and extended with a safety margin to clip the region of interest. This region is then input into a second-level fracture segmentation network. This second-level network performs fine-grained feature inference specifically on the microstructure topology of the trabecular bone within the calcaneus, ultimately outputting four categories of calcaneal anatomical region identifiers based on the microstructure topology of the trabecular bone (specifically covering the anterior bone block, posterior tubercle, posterior articular surface, and suspensory process). Specifically, this is represented as follows: Label-1, Anterior Bone Block: Located anterior to the transverse main fracture line, including the anterior process of the calcaneus and the calcaneocuboid articular surface area; Label-2, posterior tubercle: the largest fracture fragment located lateral to the main fracture line in the sagittal plane and posterior to the main transverse fracture line, which is the attachment area of ​​the Achilles tendon; Label-3, Posterior Articular Surface: Contains fragments of the posterior subtalar articular surface, located below the intersection of the two main fracture lines, and is the most critical fragment affecting the functional recovery of the subtalar joint. Label-4, suspensory process: Located medial to the main fracture line in the sagittal plane, it remains relatively stable during fracture due to the constraint of strong ligaments such as the deltoid ligament and the calcaneal interosseous ligament, and serves as a reference point for fracture reduction.

[0024] Furthermore, this embodiment addresses free bone fragments with blurred edges and excessively small size that cannot be directly classified by the network, triggering a projection-based routing mechanism using anatomical anchor points. This is achieved by invoking a spatial extremum search module to automatically extract anchor points, specifically based on the complete calcaneal mask parameter. A three-dimensional orthogonal bounding box conforming to the outer contour of the mask is constructed in a standard anatomical coordinate system. An intersection vertex search is performed along the negative Y-axis (posterior), positive X-axis (outer), and negative Z-axis (lower) directions of this bounding box to extract unique three-dimensional spatial coordinates as fixed anatomical points in the posterolateral inferior region of the calcaneus; an intersection vertex search is performed along the positive Y-axis (anterior), negative X-axis (inner), and positive Z-axis (upper) directions to extract fixed anatomical points in the anterior medial superior region. Vector subtraction is performed on the two extracted physical extreme coordinates to construct a three-dimensional spatial anatomical anchor point vector penetrating the principal axis of the calcaneus; then, the centroid of the solid point cloud of the bone fragments to be classified is extracted. After performing forward hierarchical feature inference of the multi-label image segmentation model, this embodiment executes an anomaly detection mechanism based on voxel aggregation degree for fragment information present in the output tensor. A pre-constructed calcaneal principal anatomical dictionary is used; the data configuration of this dictionary is based on offline spatial dimension statistical measurements of standard healthy calcaneal 3D solid images; the pre-constructed calcaneal principal anatomical dictionary establishes a static constraint correspondence between calcaneal anatomical partition identifiers and standard 3D envelope coordinate references. For free small-volume point cloud entities whose spatial location exceeds the range of the standard 3D envelope coordinate references constrained within the dictionary, and whose number of independent voxels after connected component aggregation exceeds a pre-constructed isolated noise filtering threshold, the data cleaning module extracts them separately and marks them as valid bone fragments to be classified. The pre-constructed isolated noise filtering threshold is preferably configured as 50 consecutive voxels; further, the 3D mesh model of the bone fragment to be classified is retrieved, and the centroid parameter of the solid point cloud representing its absolute physical center (denoted as ) is calculated and generated. ); Read the pre-constructed three-dimensional spatial anatomical anchor point vector (denoted as ) from the posterolateral inferior fixation point of the calcaneus to the anteromedial superior fixation point. This leads to the introduction of spatial dot product dimensionality reduction mapping logic: It calculates the relative projection distance component of the centroid of the solid point cloud onto the anatomical anchor point vector (main anatomical axis) in three-dimensional space through an inner product operation mechanism. Then, it performs a scaling division operation using the square of the vector magnitude to generate the feature anchor point projection normalization coefficient, which characterizes the relative topological depth position of the free bone fragments on the main calcaneal morphological axis. In the above calculation logic, the numerator part... This represents the dot product operation performed by the processor in 3D Euclidean space between the centroid vector of the entity point cloud and the anatomical anchor point vector in 3D space. Its function is to calculate the absolute projected length of the centroid along the principal axis of the anchor point; the denominator... The squared magnitude of the 3D spatial anatomical anchor point vector represents the actual physical reference length of the principal anatomical axis in the current patient entity model. The original, unconstrained relative projection scale is extracted by performing a division operation on these two values. (Outer nested symbols) This represents the upper limit gating filtering logic configured in the data link, used to hard truncate out-of-bounds values ​​greater than 1 to 1; (symbol) The lower limit gating filtering logic is configured to forcibly truncate reverse detachment values ​​less than 0 to 0. This is to exclude classification anomalies caused by free bone fragments detaching from the effective physical domain of the calcaneus due to splashing. Finally, the three-dimensional topological space coordinates are reduced to one-dimensional normalized depth coordinate parameters confined to the closed interval [0,1], thereby excluding clustering divergence distortions caused by abnormal splashed bone fragments during three-dimensional Euclidean distance comparison. A one-dimensional depth topological mapping table is further pre-defined; the segmented interval configuration of this mapping table is based on offline statistical sampling measurements of the anterior and posterior anatomical morphology and physiological proportions of the calcaneus in historical healthy subjects; the one-dimensional depth topological mapping table establishes a binding relationship between the normalized depth distribution value intervals and the calcaneus anatomical partition identifiers. As a specific application example, the preferred corresponding association relationships in the one-dimensional depth topology mapping table are as follows: [0,0.3] closed interval mapping is associated with the posterior tubercle identifier, (0.3,0.6] semi-open and semi-closed interval mapping is associated with the posterior articular surface identifier, and (0.6,1] semi-open and semi-closed interval mapping is associated with the anterior bone fragment identifier. By extracting the feature anchor point projection normalization coefficient and calling the one-dimensional depth topology mapping table to perform conditional interval comparison, the specific segment interval where the coefficient falls is confirmed to obtain the initial assignment label; then, combined with the normalized spatial position parameter of the bone fragment to be classified on the three-dimensional orthogonal coordinate axis, the nearest neighbor in Euclidean space is used. The distance measurement algorithm binds the spatial subordination state of the bone fragments to be classified to the nearest known calcaneal anatomical region identifier node. In this embodiment, after completing the anatomical identifier binding, a global anti-drift reference establishment action is performed. Specifically, the suspensory process anatomical label, which maintains absolute three-dimensional spatial stability and does not shift due to the physical constraint of the calcaneal-talar interosseous ligament, is extracted from the calcaneal anatomical region identifiers. The three-dimensional spatial center position of the voxel set to which this label belongs is extracted, and the corresponding three-dimensional coordinate parameters are statically locked in the system storage space. This establishes the absolute reference center for all subsequent transformations and registration calculations, represented as the static spatial registration reference origin.

[0025] The second step further involves generating a mirror topology map of the healthy side based on the static spatial registration reference origin mapping, and then executing an iterative nearest-point spatial alignment model to extract the three-dimensional spatial transformation matrix for aligning each free fracture fragment to the mirror topology map. The implementation logic is as follows: Anonymous multimodal healthy side CT voxel data paired with the patient's affected side are extracted. The static spatial registration reference origin is used as the global three-dimensional coordinate zero point. Values ​​of all voxel coordinates are extracted along the orthogonal normal direction of the anatomical sagittal plane and a negative inversion mapping operation is performed to generate a healthy side mirror topology map to characterize the patient's original perfect anatomical surface.

[0026] Point cloud surface topology sets of each free fracture fragment, including the anterior bone block, posterior tubercle, and posterior articular surface, were extracted separately. The surfaces of the corresponding anatomical regions of the contralateral mirror topology map were set as rigid target point base maps for matching and alignment. Iterative nearest-point (ICP) spatial alignment calculations were performed using the static spatial registration reference origin as a unified spatial constraint keyhole.

[0027] This embodiment pre-calls the voxel grid downsampling algorithm before proceeding with the iterative approximation calculation. Specifically, a voxel bounding box structure covering the free fracture fragment is constructed in a three-dimensional coordinate system, and a voxel sampling resolution parameter is configured for this structure, denoted as . This parameter characterizes the physical side length of a single discrete sampling voxel in the data mapping, directly determining the feature point cloud density participating in the iterative alignment calculation. In this embodiment, the voxel sampling resolution parameter is determined and limited to the range of [0 mm, 5 mm], with a preferred value of 3 mm. In this embodiment, after downsampling based on a grid scale preferably of 3 mm, the local point cloud centroid within each retained voxel is extracted as a representative discrete sampling point. During subsequent iterative optimization, the spatial normal Euclidean distance from the discrete sampling points on the surface of the dimensionality-reduced free fracture fragment to the target base map surface is continuously extracted. During the point cloud iterative approximation calculation, the spatial convergence residual threshold is injected into and continuously read through the state monitoring module. This parameter represents the physical limit of the spatial alignment calculation model's termination of the cyclic approximation action and the release of the final transformation matrix. In this embodiment, the spatial convergence residual threshold is locked in the judgment interval of [0.05 mm, 0.15 mm], preferably set to 0.1 mm. The reason for this setting is that the physical gap tolerance of conventional medical-grade fracture internal fixation is about 1 mm. If the value is set higher than 0.15 mm, the registration algorithm will stop prematurely, resulting in the extracted three-dimensional spatial transformation matrix still containing alignment deviations sufficient to cause traumatic arthritis. When the overall root mean square error of the current iteration frame output decreases and meets the spatial convergence residual threshold of 0.1 mm, the iteration termination command will be triggered, and the three-dimensional spatial transformation parameter currently in dynamic equilibrium will be output. Further, the three-dimensional translation vector parameter and the three-dimensional rotation mapping parameter accumulated during this cyclic evolution process for eliminating rigid body displacement errors are extracted, and the two are structurally encapsulated and finally extracted and output as a three-dimensional spatial transformation matrix representing the free fragment returning from the distorted position to the normal physiological pose.

[0028] Furthermore, after completing the calculation of the aforementioned spatial transformation matrix, this embodiment performs a third step: reconstructing the three-dimensional point cloud based on the three-dimensional spatial transformation matrix, and performing projection dimension reduction onto the standard anatomical sagittal projection system to extract the transitional two-dimensional contour feature vector. The specific derivation is as follows: The three-dimensional spatial transformation matrix of each free fracture fragment output from the previous step is extracted and applied to the initial three-dimensional spatial coordinate system of the corresponding bone fragment to perform spatial reconstruction and repositioning of the three-dimensional point cloud. During the point cloud reconstruction process, a closed-loop mechanism for extraction and mapping of bone density is introduced. Specifically, grayscale values ​​(HU values) representing the physical density of bone tissue are read from the source multimodal CT voxel data and pre-defined as static attribute labels, which are bound one-to-one to the bottom-level mesh vertices of the corresponding three-dimensional point cloud. After the free fracture fragment is moved and reconstructed to the repositioned state based on the three-dimensional spatial transformation matrix, the set of grayscale values ​​on the cross-section of the spliced ​​fracture ends where the bone fragments are joined together is extracted. This set is denoted as... The cortical bone grayscale filtering threshold is used to characterize the lower limit of X-ray physical attenuation of dense cortical bone that can provide rigid body support. In this preferred embodiment, the cortical bone grayscale filtering threshold is set to be greater than or equal to 350 HU, which removes low-density hematoma tissue or crushed and lost cancellous bone cavity areas at the fracture end section, leaving only rigid contact points that can achieve mechanical self-locking.

[0029] Further, the envelope of the three-dimensional point cloud that has completed the initial spatial reconstruction is retrieved, and the three-dimensional collision region is extracted using the voxel intersection judgment logic to generate an initial intersection patch array. This array is then used to trigger the boundary stripping logic based on the healthy side mirror topology map to remove the natural bone outer surfaces that are pressed together due to bone block misalignment. Specifically, the spatial nearest neighbor search algorithm is called to construct a KD tree spatial index structure based on the bottom mesh vertices of the healthy side mirror topology map. For each discrete point element to be tested in the initial intersection patch array, the reference mapping reference point with the smallest Euclidean distance is retrieved on the surface of the healthy side mirror topology map based on this spatial index structure. The physical normal vector carried by the reference mapping reference point on the topology map surface is extracted, and the spatial position deviation vector between the discrete point element to be tested and the reference mapping reference point is mathematically projected onto the physical normal vector. This is used to quantify and extract the three-dimensional normal relative depth of each discrete point element relative to the healthy side mirror topology map (which represents the absolute spatial boundary of the smooth outer surface of the natural joint).

[0030] Discrete points with depth errors within a preset cortical thickness tolerance threshold (e.g., 0.5 mm) are identified as invalid natural cortical compression surfaces and physically filtered out, thereby extracting the actual exposed cancellous bone fracture splice section and simultaneously reading the CT grayscale scalar data bound to the valid discrete points on the section.

[0031] This embodiment transforms the density scalar into a physical stability measure by constructing an area integral and mask filtering aggregation step, specifically as follows: It traverses a total of K micro-mesh patches (or point cloud neighborhoods) on the splicing section. Variable i is the local patch discrete index number assigned by the underlying calculation module when traversing the splicing section of the free fracture fragment ends; variable K represents the total physical number of valid micro-mesh patches or point cloud neighborhoods contained on the current splicing section. This number is dynamically constrained by the accuracy of the preceding 3D mesh generation, and its optimal normal distribution is around 10. 3 Up to 10 5 Within the order of magnitude range. In each execution cycle, the grayscale value bound to the current face is determined. Is it greater than or equal to the cortical bone grayscale filtering threshold? The Mask switch module, which triggers a binary logic state, outputs the following: when the condition is met, it outputs a state value of 1, indicating that the patch is dense cortical bone with real rigid support capability; when the condition is not met, it outputs a state value of 0. The above state output values ​​are then compared with the corresponding microscopic physical areas. Multiply the results and perform global summation, then divide by the total global area. The final mapped output is the high-density cortical-bone contact ratio score, denoted as... .

[0032] This embodiment further invokes the 3D mesh depth mapping module to generate 2D morphological parameters for clinical visual assessment and large model verification. The reconstructed 3D point cloud data is then orthogonally projected onto the standard anatomical sagittal projection system (reference plane) for dimensionality reduction. During the rendering pipeline of the orthogonal projection, the depth difference data (physical undulation distance in the vertical space of the Z-axis) of each entity vertex on the 3D point cloud surface relative to the projection reference plane is extracted simultaneously to avoid the contour deformation error caused by the 2D black and white projection when the fracture fragments are tilted in three dimensions.

[0033] Further, the depth information extraction module based on the graphics processing unit obtains the depth space coordinate data during the projection process. The depth mapping logic within the calculation frame for orthogonal projection of the 3D point cloud data is as follows: It uses the established static spatial registration reference origin as the absolute anchor point in three-dimensional space and calls the bounding box extremum search logic to perform a three-dimensional spatial traversal of the healthy side mirror topology map. Specifically, it extracts the positive and negative boundary grid vertices farthest from the absolute anchor point in three-dimensional space along the coronal axis (the physical directions of the inner and outer sides on the anatomical cross-section), and registers them as the natural inner and outer extremum points that have not been damaged by trauma. It extracts the absolute three-dimensional spatial coordinates of the above two extremum points in the global static registration reference system and performs vector subtraction to construct the line connecting the natural inner and outer extremum points. The spatial direction of this line is directly configured as the normal constraint condition, thereby rigidly aligning and locking a unique standard anatomical sagittal projection system that is absolutely orthogonal to the normal constraint condition in the topological environment of the distorted side. Then, it extracts and obtains the normal vector of the projection reference plane from the locked projection system. Offset distance from the origin, which characterizes its absolute position A multi-dimensional coordinate mapping pipeline is constructed through the graphics rendering processing module. The parameter (u,v) represents the two-dimensional discrete pixel grid coordinates within the constructed standard anatomical sagittal projection plane, used to map the specific row and column positions of the final output digital image; the parameter (x,y,z) represents the original three-dimensional physical space coordinates of the projected entity in the global static registration reference frame. This pipeline is applied to the three-dimensional mesh vertices of the entity projected onto any target pixel (u,v). Through its normal vector to the projection reference plane Perform vector dot product and subtract the distance offset from the origin. This is used to solve for the elevation difference components along the normal direction at that point. The outermost bounding absolute value operator... Configured as an undirected distance extraction rule, this means that regardless of whether the currently reconstructed fracture fragment protrudes in front of the projection reference plane or is recessed behind it, the positive and negative signs of the spatial direction are forcibly stripped away, and only the absolute spatial distance of its physical undulations perpendicular to the projected two-dimensional grid is extracted. This embodiment establishes a transformation from high-dimensional (x,y,z) coordinates to a two-dimensional grid (u,v) with added depth distance attributes. This achieves data dimensionality reduction while seamlessly encapsulating the three-dimensional elevation changes caused by spatial misalignment within a single-layer depth array, solving the problem that fully transmissive X-ray two-dimensional silhouettes cannot distinguish spatial overlap.

[0034] Furthermore, in the extraction link of the aforementioned matrix dot product and elevation mapping, a depth buffer anti-overlay mechanism is triggered. Specifically, when it is detected that multiple entity 3D mesh vertices with different spatial elevations are simultaneously mapped to the same 2D discrete pixel raster coordinate (u,v) due to overlapping 3D views, an extreme value aggregation instruction is forcibly scheduled. Specifically, all vertical Euclidean distances mapped to the same coordinate position are compared in a numerical array, and only the vertical Euclidean distance with the largest value is selected and retained as the only valid assignment for that pixel coordinate. The maximum vertical Euclidean distance after selection is directly assigned as the spatial feature of the projected terrain depth. It represents the elevation of the outermost surface three-dimensional physical undulation of fracture fragments superimposed on each pixel coordinate in a two-dimensional plane. Furthermore, the flattened and easily overlapping black-and-white X-ray two-dimensional silhouette is upgraded to a two-dimensional contour feature vector with resistance to three-dimensional tilt distortion of fragments, and represented as terrain depth array data.

[0035] The fourth step further involves triggering the image contour feature extraction module based on the two-dimensional contour feature vector. This module uses an edge contour search algorithm to automatically detect key peak and valley markers, generating a set of visual measurement parameter vectors containing anatomical angles and dimensional values. The physical extraction logic is as follows: Based on the generated two-dimensional contour feature vector (representing terrain depth array data) with elevation and depth dimensions, the retrieval logic of the image contour feature extraction module is triggered. In this embodiment, Gaussian smoothing filtering is first applied to the contour features of the depth array data before performing extreme value search to eliminate local spiculation interference at bone edges caused by free micro-bone fragments or CT reconstruction artifacts.

[0036] In this embodiment, the contour anti-aliasing smoothing coefficient is retrieved and initialized before calling the edge extremum search algorithm. This parameter, serving as the standard deviation configuration factor for the two-dimensional Gaussian filter kernel, characterizes the pixel window coverage span used to remove high-frequency edge spatial noise (such as isolated protrusions caused by tiny free bone fragments) when processing terrain depth array data. This contour anti-aliasing smoothing coefficient... The boundary is controlled within the numerical range of [2 pixels, 5 pixels], with a preferred calibration of 3 pixels. If the contour anti-aliasing smoothing coefficient is determined to be lower than the set lower limit of 2, the contour lines of the terrain depth array will retain too many CT reconstruction artifacts (such as burrs produced by metal beam hardening), causing the subsequent local zero-crossing detection algorithm to be deceived by false high-frequency extreme points, misjudging non-anatomically significant acute angles as Boehler angle vertices.

[0037] This embodiment, after smoothing and filtering, utilizes an edge contour search algorithm to perform a directed sliding scan along the skeletal topological contour from back to front, directly searching for spatial extreme points in the physical depth dimension within the terrain depth array data. By performing zero-crossing detection on the depth gradient change rate of local regions, multiple local elevation extreme coordinates on the one-dimensional scan trajectory are extracted. Based on the prior physical arrangement sequence of calcaneal anatomy, topological binding mapping rules are executed according to the back-to-foreign scanning order: specifically, the first depth maximum point captured by the scan is mapped and registered as the highest point of the posterior tubercle, the second depth maximum point is mapped and registered as the highest point of the posterior articular surface, and the third depth maximum point is mapped and registered as the highest point of the anterior protrusion, thereby locking these three key peak and valley landmarks using absolute spatial scan order constraints.

[0038] Based on the two-dimensional physical coordinates of the three key peak-valley landmarks extracted above—the highest point of the posterior tubercle, the highest point of the posterior articular surface, and the highest point of the anterior protrusion—in the anatomical sagittal projection system, the feature vector construction and angle mapping module is triggered. This embodiment takes the calculation of the Boehler angle, which characterizes the arch height, as an example, using the highest point of the posterior articular surface as a common geometric vertex (denoted as...). ), and the highest point of the posterior nodule (denoted as ) and the highest point of the forward protrusion (denoted as These are respectively used as the topological ends of the rays, thereby constructing the first anatomical feature vector in the two-dimensional plane. With the second anatomical feature vector Further, the geometric space dot product operation logic is invoked to perform numerical dimensionality reduction: It calculates the inner product of the first and second anatomical feature vectors, divides it by the product of their Euclidean moduli, and then extracts the inverse cosine value of the result. This directly reduces the dimensionality of the dispersed spatial point coordinates into visual anatomical morphological angle parameters, denoted as […]. (Specifically, this refers to the Boehler angle; the Gissane angle is extracted using a multi-point vector dot product operation based on the same source.) This angle parameter characterizes the geometric recovery state of the overall calcaneal arch morphology in a macroscopic dimension. It eliminates the error of subjective line drawing on X-ray films, achieving absolutely automated quantitative measurement based on pixel grid extreme points. In this embodiment, based on conventional medical standards, the normal reference recovery baseline for the Boehler angle is set to 25 to 40 degrees, and the normal reference recovery baseline for the Gissane angle is set to 120 to 145 degrees. Finally, the extracted angle values ​​and anatomical dimension values ​​are uniformly structured to generate a set of visual measurement parameter vectors with clinical evaluative significance.

[0039] The fifth step further involves encapsulating the visual measurement parameter vector set and the three-dimensional spatial transformation matrix into a unified reset state feature set, which is then input into a comprehensive evaluation module pre-configured with a medical rule knowledge base for multi-dimensional data rule comparison. The specific execution mechanism is as follows: The system extracts a 3D spatial transformation matrix (representing the spatial displacement of the microscopic rigid body), a high-density cortical bone contact ratio score (representing the stability of the internal load-bearing physical structure), and a set of visual measurement parameter vectors (containing angles and dimensions with clinical morphological guidance significance). This further triggers an automatic classification preprocessing link based on 3D mesh topology to extract high-confidence prior fracture classification labels. Specifically, the 3D topology feature extraction module is invoked to perform a spatial connected component traversal algorithm based on the 3D 26-neighborhood rule for each independently segmented posterior articular surface partition (the aforementioned Label-3). After forming multiple closed mesh sets through traversal search, an isolated noise volume filtering threshold is introduced. Fragments with a volume smaller than this threshold (e.g., less than 1% of the total volume) are judged as bone debris or free artifacts without structural support significance and are removed. The count values ​​of the remaining independent closed mesh sets are counted in real time to generate and introduce a parameter for the number of effective connected components. This parameter characterizes the absolute discrete number of independent main load-bearing fragments that are physically split into the rear joint surface due to high-energy vertical impact.

[0040] Further based on the parameter of the number of effective connected components The numerical landing point triggers the branch mapping logic to deduce the basic Sanders fracture classification labels. During the offline initialization phase, a mapping rule library, instantiated as a key-value hash structure, is pre-configured based on clinical trauma orthopedic classification standards. The specific construction method of this mapping rule library is as follows: a pure numeric constant representing the number of independent fractured bone fragments on the posterior articular surface is used as the index key field, and standardized character labels representing the medical classification intent are used as the mapping value field. The clinical engineering basis of its mapping rules is that the Sanders classification depends on the number of fracture lines crossing the articular surface on coronal CT (number of bone fragments = number of fracture lines + 1). Based on this, the following defined mapping correspondence is established in the mapping rule library: When determining the number of valid connected components as input When the value of is equal to 2, the 'Sanders II' label will be output through hash addressing mapping; when determining the number of valid connected components, the parameter will be used. When the value of is equal to 3, the mapping outputs the label 'Sanders III'; when determining the number of valid connected components, the parameter... When the value is greater than or equal to 4, it is determined that the physical topology of the joint surface has undergone shattering, and the 'SandersIV type' label is directly mapped and output. Furthermore, an abnormal state fallback isolation route is configured in this mapping rule base, specifically by adding static interception key values ​​of 1 and 0: when the number of valid connected components is passed... When the value is equal to 1 (indicating that the joint surface has not been fractured and is an indication for conservative treatment) or equal to 0 (indicating that the pre-segmentation algorithm has failed or the three-dimensional reconstruction source data has been lost), the abnormal interception protection mechanism will be triggered, blocking the transmission of invalid reset feature data to the post-comprehensive evaluation module, and simultaneously transmitting and rendering a system-level warning status code indicating abnormal / non-fractured working conditions of the basic data to the external display and control terminal.

[0041] In this embodiment, to clarify the spatial constraint criterion for subtype division, a spatial principal component characteristic axis is introduced, denoted as... Specifically, the global entity point cloud coordinates of the previously generated healthy side mirror topology map are extracted, and its data covariance matrix in three-dimensional space is calculated. Eigenvalue decomposition is performed on this covariance matrix to extract three orthogonal eigenvectors corresponding to the magnitudes of the eigenvalues. Based on the biomechanical morphology of the calcaneus, its longest anatomical principal axis is the anterior-posterior direction (corresponding to the first principal component of the largest eigenvalue), while the second widest anatomical axis is the medial-lateral direction. The eigenvector corresponding to the second largest eigenvalue is extracted and recorded as the spatial principal component eigenaxis. This is used to characterize the medial and lateral physical axes of the calcaneus anatomical entity. Furthermore, the centroid coordinates of each connected domain are aligned along axes perpendicular to the characteristic axes of the principal spatial components. The direction is mathematically projected, and the absolute length scalar obtained by projection is divided by the maximum width of the calcaneus on both the inner and outer sides, thereby normalizing it to a pure numerical closed interval of [0,1]. In this embodiment, based on the specific location of the normalized projection point in the outer segment of [0,0.33], the central segment of (0.33,0.66], or the inner segment of (0.66,1], the fine subtypes such as Sanders IIA, IIB, or IIC are further automatically determined.

[0042] Further topological adjacency checks are performed to detect whether there are shared 3D boundary voxels between the 3D envelope meshes of the posterior nodule (Label-2) and the posterior articular surface (Label-3). Simultaneously with the topological adjacency check, the angle collapse warning difference is introduced and read, denoted as... This is used to characterize the physical collapse reduction limit of the Boehler angle currently measured on the affected side of the patient compared to its original perfect anatomical reference (or the mirror reference of the healthy side). In this embodiment, the boundary of the collapse warning difference of this angle is set within a closed interval of [10 degrees, 20 degrees], with a preferred extreme value judgment boundary set at 15 degrees. If the set value exceeds the upper limit of the interval of 20 degrees, the warning triggering condition may become sluggish, leading to the easy misjudgment of early moderate articular surface collapse as a common non-displaced fracture, thus providing the comprehensive assessment engine with an incorrect mild prior label; if the set value is lower than the set lower limit of 10 degrees, the judgment logic may easily trigger false collapse warnings for normal physiological skeletal variations or slight foot tilt in patients on CT scans.

[0043] Further utilizing standard medical data serialization formats, the aforementioned deduced classification prior label data is packaged together with the three physical / geometric heterogeneous parameters to generate a standardized set of reset state features. This set serves as the data payload to trigger the downstream evaluation engine, thereby triggering the comprehensive evaluation module. The core of this comprehensive evaluation module is pre-configured with a medical rule knowledge base calibrated based on offline pre-training and instruction fine-tuning techniques. Specifically, the following offline calibration is performed: a multimodal instruction fine-tuning dataset is constructed by accessing a historical electronic medical record (EMR) database. The input of the dataset is formatted as a serialized structure containing anatomical angle deviation values, high-density cortical bone contact ratio scores, and prior fracture classification labels; the output of the dataset is precisely aligned to the operational intervention sequence text issued by a senior orthopedic expert for this specific situation (e.g., 'slide towards the Z-axis orthogonal plane to find the suboptimal chiseling point' or 'trigger abnormal interception to perform subtalar joint fusion').

[0044] In the network weight update stage, a pre-trained general-purpose large model is selected as the base, and a parameter-efficient fine-tuning algorithm (Low-Rank Adaptive Mapping Algorithm LoRA) is used to inject the fine-tuning matrix via a bypass. After calculating the probability of the model output text during forward propagation, it is compared with the expert intervention sequence at the word level, and the error is quantified using the autoregressive cross-entropy loss function. Gradient backpropagation is performed through the Adaptive Moment Estimator Optimizer (AdamW) to iteratively update the weights of the bypass matrix. When the instruction prediction accuracy on the validation set remains at its extreme convergence for 10 consecutive iterations (no further loss decline occurs), the model weights are automatically solidified and output, and packaged into a medical rule knowledge base capable of conditional comparison and intent delivery for complex heterogeneous feature sets.

[0045] This embodiment further introduces a model quantization compression ratio identifier, denoted as . Its characterization represents the hardware adaptation dimensionality reduction coefficient that lowers the precision of floating-point operations in deep learning. In this embodiment, the large language model preferably adopts a large model with hundreds of billions of parameters pre-trained for the medical vertical domain (such as MedGemma27B), and the compression ratio is based on a model quantization set to 4 bits. Low-bit-width quantization is performed to ensure that the inference model can be successfully loaded into the memory space of the graphics processing unit.

[0046] During online comparison, the deployed comprehensive assessment module receives a set of reduction status features fused with fracture classification labels and invokes an internally constructed multidimensional feature function tree mapping mechanism to perform multidimensional data rule comparison. The multidimensional feature function tree mapping mechanism is pre-constructed as a directed acyclic graph (DAG) structured conditional routing decision tree; the root node of this function tree is configured as the parsing entry point for receiving the reduction status feature set; its internal intermediate layer nodes are defined as feature threshold comparison operators, respectively mapped and bound to the anatomical angle feature dimension and the high-density cortical bone contact ratio score dimension, and utilize a set weight-bearing contact benchmark threshold. and preset anatomical angle baseline The physical extreme value triggering condition serves as the branch routing; the terminal leaf node of this functional tree is preloaded with a structured text template. After the reset state feature set completes the binary logic judgment of being greater than or less than the threshold at each intermediate hierarchical node, the data flow will fall into a specific leaf node along the decision link, triggering that node to concatenate the quantized values ​​in the feature set into the corresponding text template, thereby automatically translating the discrete multidimensional heterogeneous parameters into a structured natural language evaluation array that can be natively parsed by the large language model.

[0047] In this embodiment, the anatomical angle in the visual measurement parameter vector group is set as a soft appearance optimization target for evaluating the degree of restoration of apparent geometry; at the same time, the high-density cortical bone contact ratio score, which represents the actual internal support force, is set as a hard physical load-bearing contact constraint baseline.

[0048] The sixth step is further defined as follows: If the deviation calculated by the comprehensive evaluation module exceeds the preset confidence threshold, the module receives a fine-tuning strategy containing the intention to fine-tune spatial attitude based on the rule-based comparison results. This strategy is then parsed to generate spatial displacement compensation parameters, which are then injected back into the preceding spatial registration node. Quantitative correction is performed on the transformation matrix corresponding to the current calculation frame in the direction of the target alignment deviation until the comparison deviation meets the safety threshold. Finally, a finalized three-dimensional fracture reduction spatial model is output. The specific process is as follows: When the aforementioned feature set enters the comprehensive evaluation module for game theory simulation, in response to the state determination command, the load-bearing contact benchmark threshold is simultaneously issued and activated. It is used to characterize the minimum cortical bone occlusal contact ratio that ensures the patient can withstand gravity without secondary articular surface collapse when removing the cast and bearing weight post-surgery. This load-bearing contact benchmark threshold is used in actual clinical engineering parameter configurations. The value of is converged and mapped to the numerical range of [55%, 75%], preferably 60% of the total area.

[0049] Further extract the preset anatomical angle baseline based on common anatomical knowledge. Its ideal reference range represents the perfect arch shape. The comprehensive evaluation module performs conditional comparison and screening on the reset state feature set. Specifically, before determining whether the reset state is usable, a numerical deviation discrimination benchmark is constructed based on the reset state feature set: Extract the first normalized deviation of the current anatomical angle relative to the preset anatomical angle baseline, and simultaneously extract the second normalized deviation of the current high-density cortical bone contact ratio score relative to the load-bearing contact benchmark threshold. For the anatomical angle dimension, the absolute difference scalar between the current anatomical angle and the preset anatomical angle baseline is extracted; at the same time, the preset geometric limit collapse tolerance is awakened from the data persistence layer (this parameter is based on the maximum physical angle threshold for irreversible damage to the foot arch structure in biomechanical characterization); the absolute difference scalar is divided by the preset geometric limit collapse tolerance, and the output result is hard-gated and truncated within the pure numerical closed interval of [0,1], thereby reducing the dimension to the dimensionless first normalized deviation.

[0050] For the internal load-bearing dimension, an isomorphic dimension reduction mapping is performed: specifically, the contact loss difference between the load-bearing contact benchmark threshold and the current high-density cortical bone contact ratio score is extracted; and simultaneously, a preset maximum allowable contact loss boundary is introduced, the contact loss difference is divided by this boundary parameter, and extreme value truncation filtering is performed to generate a second normalized deviation locked in the closed interval [0,1].

[0051] Two normalized deviations are assigned preset geometric appearance diagnostic weights and mechanical load-bearing diagnostic weights, respectively. In this embodiment, both weights are set to 0.5. The first normalized deviation is multiplied by the geometric appearance diagnostic weight to generate a morphological penalty term representing the cost of appearance distortion. The second normalized deviation is multiplied by the mechanical load-bearing diagnostic weight to generate a mechanical penalty term representing the cost of internal support loss. A global addition aggregation operation is performed on the morphological penalty term and the mechanical penalty term to synthesize a comprehensive comparison deviation scalar. In this embodiment, a preset reliability threshold is statically loaded and read from the configuration library in advance. This parameter is based on statistical measurements of historical trauma reduction clinical follow-up samples during the offline phase. Specifically, based on control samples from historical electronic medical records clearly marked as "postoperative complications of severe traumatic arthritis or mechanical collapse," the intraoperative repositioning final state images are substituted into the pre-calculated weighted summation calculation logic to extract historical comprehensive comparison deviation clusters. A Gaussian distribution is then used to fit and extract the confidence lower bound of the probability distribution of these deviation clusters, which is then quantified and mapped to generate a preset reliability threshold. In this embodiment, the preset reliability threshold is preferably configured as 0.75. The comprehensive comparison deviation scalar output of the current calculation frame is then compared with this preset reliability threshold. If the comprehensive comparison deviation scalar is determined to be greater than the preset reliability threshold (indicating that the current repositioning state has exceeded the clinically unacceptable comprehensive error tolerance threshold), a subsequent abnormal fine-tuning mechanism is triggered. This mechanism receives a fine-tuning strategy, including spatial pose fine-tuning intentions, returned by the comprehensive evaluation module based on the rule-based comparison results.

[0052] When performing condition screening and strategy adjudication, if the current anatomical angle is detected to deviate from the preset anatomical angle baseline... (This indicates that the instruction engine needs to output a further unidirectional stretching fine-tuning strategy to elevate the collapsed joint surface), but the currently calculated high-density cortical bone contact ratio score is already lower than or equal to the load-bearing contact benchmark threshold. .

[0053] Further triggering of abnormal interception logic for edge cases: If the model is allowed to continue stretching, the rigid fracture ends will misalign and detach, causing the free bone fragments to be suspended in the low-density hematoma cavity area, violating the biomechanical principles of the real physical world. Based on this, a truncation strategy is executed. Specifically, it blocks and intercepts the error correction instructions issued by the large language model for unidirectional stretching geometry. In this embodiment, the intent parsing module is invoked to perform operations on the text sequence based on a pre-configured and persistently stored structured semantic mapping dictionary. The dictionary is constructed as follows: unstructured natural language action words commonly used by the large language model in the medical repositioning domain are used as the mapping key field, and the corresponding encapsulated deterministic control function signaling in the 3D spatial calculation engine is used as the mapping value field. The engineering configuration of the mapping relationship is based on the one-to-one association binding between clinical operation intent and machine-level executable function execution path established by human experts in the offline stage. As a specific application example, the dictionary establishes a correspondence between the identifier text mapping of 'sliding', 'interlocking', and other characterizing attitude fine-tuning intents and the underlying Z-axis translation search trigger function signaling. When regular expression matching extracts a target that matches the core action identifier of the mapping key, the intent parsing module obtains and directly outputs the associated geometric control trigger signal through hash addressing. In response to this control trigger signal, preset parameters are automatically loaded in the 3D coordinate space, triggering a sliding rotation search logic towards the Z-axis orthogonal plane.

[0054] A two-dimensional sliding search grid covering a preset search range is constructed on the Z-axis orthogonal plane, with the centroid of the current free fracture fragment's point cloud as the spatial reference origin. A two-dimensional discrete sampling step size parameter is introduced. It characterizes the translation and rotation resolution of spatial grid nodes in the physical coordinate system. In this embodiment, its preferred values ​​are set to a translation step size of 0.5 mm and a rotation step size of 1 degree, and then according to the two-dimensional discrete sampling step size parameter. A two-dimensional sliding search grid is constructed. Each discrete grid node in this grid is traversed, and the embedded two-dimensional steady-state weighted aggregation logic is invoked to perform iterative calculations. It extracts the position of the discrete grid node pointed to by the current loop pointer and calls the normalized high-density cortical bone contact ratio score in that state. The mapping interval is [0,1]. The absolute deviation amplitude between the current anatomical angle and the ideal perfect baseline at this node pose is extracted synchronously. A preset anatomical limit tolerance is introduced and read; this parameter characterizes the absolute angle collapse red line sufficient to declare the loss of arch structure (flatfoot limit) in macroscopic diagnosis of foot and ankle surgery. In this embodiment, the anatomical limit tolerance value is determined and fixed at 20 degrees based on clinical biomechanical boundaries. If the configured value exceeds the upper limit of 40 degrees, the denominator of the normalization calculation will be too large, causing the small angle deviation features to be numerically compressed, resulting in the subsequent optimization algorithm losing its gradient sensitivity to slight distortions; by dividing the absolute deviation amplitude by this anatomical limit tolerance, a normalized anatomical angle deviation value with a scale limited to the closed interval [0,1] is generated. (When the deviation is greater than or equal to the anatomical limit tolerance, it is hard-truncated to 1). Further, multiplicative logic is used to apply the preset convergence penalty weights. The deviation value is applied as a deduction item. A logical difference operation is performed to remove the morphological penalty deduction item from the load-bearing benefit value, outputting a reset steady-state evaluation value with a partial order relationship, denoted as... .

[0055] In this embodiment, a preset steady-state convergence penalty weight is read before performing aggregation calculation. This dimensionless constant characterizes the tolerance gradient and attenuation penalty for sacrificing external anatomical morphology to achieve internal dense load-bearing support during physical interference optimization. It is calibrated based on offline extreme condition replays from a historical clinical rehabilitation tracking database: through offline analysis of negative surgical evolution samples where the pursuit of perfect anatomical angles led to the failure of internal supporting bone blocks to suspend, the critical point of support loss is statistically extracted. Steady-state convergence penalty weight. The quantization boundary was determined and fixed within the numerical range of [0.1, 0.3], with a preferred tuning value of 0.15. The extracted high-density cortical bone contact ratio score was then compared with a weight carrying a steady-state convergence penalty. The deviation value of the anatomical angle is aggregated by performing a subtraction logic, thereby reducing the dimensionless multidimensional game contradiction between geometry and mechanics to a unidirectional scalar that can be stably sorted and compared.

[0056] Furthermore, this embodiment utilizes a real-time streaming CT voxel sequence of Sanders III type calcaneal posterior articular surface collapse fracture from the hospital's medical imaging center PACS system. The background spatial calculation nodes construct a two-dimensional sliding search grid on the Z-axis orthogonal plane. When the target free fragment encounters physical fracture end suspension (high-density cortical bone contact ratio less than 60%) during unidirectional tensile compensation command execution, the solution is applied. Under the interception condition of the hard red line, according to the preset two-dimensional discrete sampling step size parameter (Translate 0.5 mm, rotate 1 degree) Extract the topological features of each discrete grid node, and continuously call the two-dimensional steady-state weighted aggregation rule to perform iterative calculations.

[0057] Table 1: Parameter mapping array of two-dimensional steady-state weighted aggregation routing rules under two-dimensional sliding search grid nodes As shown in Table 1, when scanning to a discrete grid node with a relative offset coordinate of (0.5, 2), the highest global steady-state convergence extreme value that has not triggered any peripheral anatomical red lines is extracted. The control module captures the translation and rotation mapping vector corresponding to this node, bypasses the frozen default geometric stretching link, and feeds it back to the preceding spatial registration node as the suboptimal fine-tuning strategy with the highest execution priority. This ensures that the fragment fits tightly into the internal dense cancellous bone bed while sacrificing some apparent anatomical angles.

[0058] Furthermore, the target mesh node whose reset steady-state evaluation value satisfies the preset steady-state convergence extreme value condition is extracted. The relative offset coordinates of the target mesh node relative to the spatial reference origin are resolved into translation and rotation mapping vectors. Specifically, the relative offset raster count values ​​of the target mesh node relative to the spatial reference origin in each orthogonal axis are extracted, and the two-dimensional discrete sampling step size parameter configured when constructing the search mesh in the previous step (representing the translation step size reference and the rotation step size reference) is obtained simultaneously. The relative offset raster count values ​​and their corresponding two-dimensional discrete sampling step size parameters are subjected to physical scale restoration calculation (which represents a step-size-based product mapping operation), thereby translating the discrete topological mesh coordinate mapping into translation and rotation mapping vectors with engineering measures (such as physical millimeter distance and spatial deflection degrees). This vector is then distributed as a suboptimal fine-tuning strategy to ensure a firm fit of the bone fracture ends. Upon receiving this suboptimal fine-tuning strategy, it is resolved from the strategy intent into specific spatial displacement compensation parameters. To prevent the generated strategy from causing unrealizable errors such as interpenetration of fracture fragments in the physical world, this embodiment triggers a defense mechanism based on rigid body mechanics for spatial interference detection during the analysis process. Specifically, for the extracted topological set of the point cloud surface of each free fracture fragment, the spatial computing module executes a 3D convex hull generation algorithm (such as the QuickHull algorithm). This algorithm extracts all the outer protruding vertices of the fracture fragments and removes the internal concave points, thereby generating a closed polygonal mesh topological surface, which is then represented as a rigid body spatial interference boundary parameter. This parameter constitutes a virtual mechanical collision sensor that dynamically moves with the bone fragments. Its function is to eliminate the lag in judgment based on manual visual observation of overlap, transforming the interference problem between bone fragments into a geometric envelope intersection calculation. The interference determination mechanism is configured as follows: if, within a specific time frame (representing the current iteration / rendering cycle in the loop calculation process), the rigid space interference boundary parameters of two independent bone fragments... If any triangular facet on the envelope surface intersects with spatial coordinates, it is determined that an unrealizable spatial clipping error has occurred in the physical world.

[0059] The spatial attitude fine-tuning intention implied in the suboptimal fine-tuning strategy is solved along the spatial topology as a virtual displacement correction vector applied to anatomical nodes (such as the centroid of a bone block). This is equivalent to a directional elastic traction force applied to the bone block in a 3D physics engine. Rigid body spatial interference boundary parameters based on global fragments. Set, solve for the virtual displacement correction vector of the free fracture fragment. Three-dimensional spatial equilibrium state when moving under the action of traction force: Specifically, when penetrating interference is detected between the envelopes of two bone fragments (with a spatial mesh overlap rate greater than zero), the physics simulation engine immediately freezes the current displacement and calculates the translational and rotational increments to offset the penetrating interference after being constrained by collision resistance, thereby locking all translational and rotational states of the current fracture fragment. Vertex depth traversal logic is triggered for the mesh intersection region where interference is detected. Specifically, all free mesh vertices that have penetrated into the envelope of another fracture fragment are extracted, and the vertical Euclidean distance from these penetrating mesh vertices to the nearest triangular facet on the surface of the penetrated rigid body is calculated. The set of vertical Euclidean distances is compared, and the distance scalar with the largest value is represented as the maximum penetration depth scalar. At the same time, the set of penetrated triangular facets associated with the maximum penetration depth scalar is extracted, the normal vectors of the penetrated triangular facets are extracted, their arithmetic mean on the X, Y, and Z coordinate axes is calculated, and this is set as the physical bounce repulsion normal.

[0060] Read the pre-configured virtual rigid body penalty coefficient This parameter represents the numerical amplification weight that converts the geometric interference depth into a rigid body repulsion increment in the virtual physical simulation space. In this embodiment, the value of this parameter is limited to the range of [0.5, 0.9], with a preferred value of 0.7. Based on the preferred penalty weight of 0.7, the maximum penetration depth scalar is compared with the virtual rigid body penalty coefficient. The components are multiplied and assigned vector properties along the rebound repulsion normal to generate a three-dimensional translational compensation component to counteract the penetration interference. The specific mesh intersection contact point where the maximum penetration depth scalar occurs is extracted, and a spatial offset vector is constructed from the centroid of the free fracture fragment point cloud to the contact point. The generated three-dimensional translational compensation component is matrix-mapped with the spatial offset vector by a three-dimensional cross-product to calculate the corresponding three-dimensional rotational compensation component generated under the action of this eccentric resistance. Finally, the two compensation components are jointly encapsulated and output.

[0061] In this embodiment, after completing the rigid anti-penetration stripping between the bone fragments, a three-dimensional non-uniform tensile limit envelope surface, pre-statically bound to the anatomical partition markers of each calcaneus, is retrieved and loaded. This envelope serves as a soft tissue tension constraint baseline to prevent the compensation parameter from tearing the externally connected soft tissues. This three-dimensional non-uniform tensile limit envelope surface is a spatial threshold field generated based on the foot and ankle biomechanical database during the offline calibration phase. In the offline phase, the physiological yield limit data of ligaments attached to various bone fragments (such as the calcaneofibular ligament and the intercalcaneal ligament) in different spatial directions are statistically analyzed, and this non-uniform spatial threshold field is constructed using these as boundaries. Taking the anterior calcaneal protrusion as an example, the tolerance measure of its corresponding envelope surface in the Z-axis (vertical stretching) direction is 2 mm, while the tolerance measure in the Y-axis (anterior-posterior sliding) direction can reach 4.5 mm. The limitation is that fracture reduction cannot simply pursue geometric occlusion while ignoring soft tissue tension. If the three-dimensional displacement parameter is detected to break through the curved boundary of any side of the envelope surface, in actual surgery, it means applying a mechanical traction force sufficient to cause ligament tearing in the corresponding direction. This non-uniform offline benchmark mapping enables the virtual space to achieve an equivalent model of the physical boundaries of complex soft tissues in the human body without the need for any external real-time tension sensors.

[0062] Before generating the boundary projection convergence vector, this embodiment also includes a physical entity-based tension calibration synchronous verification mechanism to ensure that the soft tissue tension constraint baseline generated offline based on the historical database can adapt to the actual biomechanical environment of the current specific patient. This embodiment is equipped with a tension verification module in the external hardware access layer. During the surgical reduction process, the physical tension verification module obtains the soft tissue yield resistance characteristics of the anatomical anchor points on the affected side of the patient (such as the attachment point of the calcaneofibular ligament) under passive traction, and uses this characteristic as the physical observation gain to perform dynamic scaling correction of the pre-stored three-dimensional non-uniform tensile limit envelope surface to a safety threshold.

[0063] The physical tension verification module employs a medical-grade miniature force gauge or a piezoelectric force sensor. In practice, it simply needs to be aseptically attached or hooked onto the corresponding tendon or ligament anchor point to convert the actual physical pulling resistance into a standardized electrical signal and feed it back to the system bus. Based on the three-dimensional spatial transformation matrix of the current loop frame in the previous spatial registration node, the three-dimensional relative displacement vector of the free fracture fragment relative to the static spatial registration reference origin is extracted.

[0064] A static data mapping table is pre-established in the configuration library, and the human standard biomechanical mean parameter matching each calcaneal anatomical region identifier is loaded. The configuration library establishes a relationship between the calcaneal anatomical region identifier as the retrieval index key and the expected tensile resistance value of a specific ligament as the corresponding mapping value; this relationship is configured based on the statistical mean of physical tensile force at the yield critical point of a specific ligament in healthy individuals of the same age and sex from an offline medical statistical database. As a specific application example, for the calcaneal anatomical region identifier corresponding to the calcaneofibular ligament attachment point, the human standard biomechanical mean parameter mapped and bound in the configuration library is preferably configured to 150 Newtons.

[0065] In response to the physical force-induced state transmitted back from the data bus, the analog-to-digital conversion and calibration mapping module is invoked to extract the voltage amplitude of the standardized electrical signal. Linear scaling and dimensional conversion are then performed based on the pre-written factory physical sensitivity coefficient of the sensor to calculate the soft tissue yield resistance characteristic with true Newtonian dimensions. The value of the true soft tissue yield resistance characteristic is then divided by the human body's standard mechanical mean parameter to generate the original stiffness scaling factor. Upper and lower limit safety truncation operations are performed on the original stiffness scaling factor. As a specific application example, the preferred safety truncation interval for the stiffness scaling factor is [0.6, 1.5]. The dimensionless stiffness scaling factor output after the safety truncation operation is extracted to characterize the actual deviation ratio. Based on this generated stiffness scaling factor, the three-dimensional vertex coordinates of the pre-bound three-dimensional non-uniform tensile limit envelope surface are subjected to proportional matrix multiplication scaling calculations with the static spatial registration reference origin as the radiation center.

[0066] The modified envelope surface is retrieved as the soft tissue tension constraint baseline, and the three-dimensional spatial interference penetration between the end coordinates of the three-dimensional relative displacement vector and the three-dimensional non-uniform tensile limit envelope surface is calculated using three-dimensional mesh collision detection logic. In this embodiment, the three-dimensional mesh collision detection logic specifically uses the end coordinates of the three-dimensional relative displacement vector as a virtual physical probe and extracts the three-dimensional non-uniform tensile limit envelope surface after the aforementioned dynamic scaling correction. This envelope surface is composed of a closed manifold mesh constructed from a series of triangular facets.

[0067] A detection ray is projected outward along the direction of the three-dimensional relative displacement vector, using the static registration reference origin as the emission source. All triangular faces of the manifold mesh are traversed, and the coordinates of the penetration intersection point between the ray and the mesh are calculated using the Möller-Trumbore ray-triangle intersection algorithm. If the endpoint coordinate point is determined to have exceeded the penetration intersection point in the ray direction, the Euclidean distance between these two coordinate points is extracted to calculate the initial three-dimensional spatial interference penetration. After determining that the envelope surface has been breached, a geometric ray projection calculation is initiated from the endpoint coordinate point as the origin towards the envelope surface. A shortest normal retreat path perpendicular to the local tangent plane of the envelope surface is solved, and the Euclidean distance scalar of this path in the three-dimensional coordinate system is measured. The Euclidean distance scalar is numerically multiplied by its corresponding inverse normal vector to synthesize a boundary projection convergence vector with dual attributes of directional constraint and physical distance measure. By using vector addition, the boundary projection convergence vector is used to perform reverse numerical compensation calculation on the current three-dimensional relative displacement vector. This is done to quantify and remove abnormal translational components that exceed the physical tensile limit of the ligaments of the free fracture fragments, ensuring that the coordinates of the fracture fragments are constrained within the safe envelope plane.

[0068] If it is determined that the endpoint of the 3D relative displacement vector has penetrated and broken through the 3D non-uniform tensile limit envelope, a geometric ray projection is initiated towards the envelope with the endpoint as the origin. The shortest normal retreat path perpendicular to the envelope tangent is calculated, and the Euclidean distance scalar of this path in the 3D coordinate system is measured, representing the penetration depth. The Euclidean distance scalar is then multiplied numerically with its corresponding inverse normal vector to synthesize a boundary projection convergence vector possessing both directional constraint and distance measurement attributes. The boundary projection converges to a vector using the vector addition rule. Perform a compensation calculation against the original three-dimensional relative displacement vector that caused the clipping. Ensure the bone block is "pulled back" to the edge state of the safe envelope.

[0069] In this embodiment, before performing the above compensation calculation and outputting the final displacement increment, the highest priority arbitration mechanism of the multidimensional physical constraint defense line is triggered: specifically, before extracting the pre-compensated displacement coordinates, a pose feedforward trial calculation is performed; the three-dimensional spatial transformation matrix corresponding to the current loop frame in the preceding spatial registration node is extracted, and a matrix weighted fusion operation is performed on it with the spatial displacement compensation parameters generated and solved by the preceding logic to generate a pre-compensated spatial transformation matrix to represent the retreat intention; the global vertex array of the three-dimensional point cloud envelope of the current free fracture fragment is extracted, and an affine transformation mapping calculation is performed on the global vertex array using the pre-compensated spatial transformation matrix to generate and extract the pre-compensated displacement coordinates; the pre-compensated displacement coordinates are used as a physical probe, and the spatial interference detection module is called a second time to verify whether it has secondary interference with the rigid body spatial interference boundary parameters of other surrounding bone fragments.

[0070] If secondary interference is detected, it will be confirmed that the current bone fragment is in a conflict situation where "yielding will result in bone fragment protrusion, while pushing away will result in ligament tearing." Under this condition, the rigid body physics establishes impenetrability as the red line, and immediately locks all translational and rotational states of the fracture fragment; Simultaneously, the data feed to the Kalman filter state update equation is intercepted and directly transitioned to the system anomaly interception strategy, generating a severe crushing warning signal indicating that the current joint surface is beyond repair. This severe crushing warning signal is then sent to the display and control terminal to trigger an interactive rendering action that prompts alternative intervention strategies. Upon receiving the signal, the display and control terminal highlights the joint surface topology region that requires fusion due to severe crushing in its operation window.

[0071] If no secondary interference occurs, the subsequent calculations continue, further triggering the spatial trajectory smoothing filter module. This module utilizes the Kalman filter state update matrix equation to eliminate pose abrupt jitter caused by discrete control commands. Specifically, the three-dimensional spatial transformation matrix calculated in the current loop frame is constructed as a priori state estimation vector (denoted as...). The spatial displacement compensation parameters after stripping the abnormal stretching component of soft tissue are encapsulated into an independent observation matrix for the current period (denoted as...). The Kalman gain formula and the posterior state update formula are as follows: in The standard identity mapping matrix is ​​used to map the state space to the observation space. To predict the error covariance matrix, an observation noise covariance matrix R is introduced and configured. This matrix R characterizes the confidence uncertainty of the data fluctuation noise carried by the current input environmental interferometric feedback. During initialization, static values ​​are assigned to the diagonal elements of the observation noise covariance matrix R. The variance values ​​of the diagonal elements of the observation noise covariance matrix R are preferably configured to be limited to the order of magnitude [0.01, 0.05]. The prior state estimation vector... With the posterior state vector All are defined as 6×1 column vector data structures for matching the pose of rigid bodies in space (which internally encapsulate the three-dimensional translational components along the X / Y / Z axes and the three-dimensional rotational components along the roll / pitch / yaw dimensions in sequence); the independent observation matrix of the input external displacement compensation parameters. The same configuration is used for 6×1 observation compensation data frames; R corresponds to an instantiation of a 6×6 covariance matrix. The superscript T control character represents a transpose read / write operation performed on the array rows and columns, while the superscript -1 represents the matrix inversion operation logic performed based on Gaussian-Jordan elimination or a matrix decomposition library.

[0072] The background computing nodes construct composite matrix data packets capable of being reverse-engineered. Specifically, they extract the prediction error covariance matrix, which contains the uncertainty of its own state transition. And the observation noise covariance matrix R, which characterizes the dynamic state fluctuations of the aforementioned entity under force. The Kalman gain-based solution formula... The standard identity mapping matrix H, which maps the state space to the observation space, is extracted using the matrix addressing module, and its rows and columns are transposed to obtain the result. The predicted error covariance matrix From the left and right sides respectively, they are mapped to the standard unit matrix H and Perform matrix multiplication to calculate the projection matrix of the state prior variance in the current observation space; then perform element-wise matrix addition on this projection matrix and the observation noise covariance matrix R to construct a composite matrix data packet. .

[0073] Perform matrix inversion operation logic based on Gaussian-Jordan elimination on the composite matrix data packet, and multiply by... Factors are then used to extract and output a 6×6 Kalman gain weight matrix for the current iteration calculation period. .

[0074] Extract the independent observation matrix for the current period The observation residual vector is obtained by subtracting the projection of the prior state estimation vector into the observation space; the Kalman gain weight matrix is ​​then used. Perform matrix multiplication weighting on the observed residual vector and then superimpose the weighted results onto the prior state estimation vector. This generates the posterior state vector. For posterior state vectors with a 6×1 column vector data structure Perform reverse data unpacking in the spatial dimension, extracting the internal 3D translation and rotation components, and then calling a homogeneous coordinate transformation algorithm to repackage these components into a 4×4 smoothed transformation matrix. This smoothed transformation matrix is ​​then used to directly overwrite and update the 3D spatial transformation matrix corresponding to the current calculation frame in the previous spatial registration node.

[0075] Loading the covariance update formula: Where I represents the initialized 6×6 standard unit diagonal matrix; the Kalman gain weight matrix is ​​subtracted from this standard unit diagonal matrix I. The product of the standard unit mapping matrix H and the result is then compared with the prediction error covariance matrix of the current frame. Perform matrix multiplication to calculate and output the posterior error covariance matrix for the current period. This matrix physically represents the objective attenuation of the system's state uncertainty after absorbing the physical observation gain of the current frame.

[0076] Furthermore, in this embodiment, when dealing with physical interference scenarios where free fracture fragments on the lateral wall overlap and compress with the morphological axis of the main calcaneus, the soft tissue yield resistance characteristics transmitted back by the peripheral entity tension verification module via a piezoelectric force sensor are read in real time. The 3D mesh collision detection logic extracts penetration interference features and, based on the calculated maximum penetration depth and physical rebound repulsion normal, continuously generates 3D displacement compensation parameters with a rigid pushing intention. High-frequency noise reduction and smoothing fusion are performed on the displacement command stream constrained by the 3D non-uniform tensile limit envelope surface using the Kalman filter state update matrix equation.

[0077] Table 2: Time-series state tracking array of Kalman filter fusion anti-drilling displacement compensation parameters The values ​​in Table 2 above are in mm. During periods N to N+1, fast backoff stripping is performed using a Kalman gain weight maintained at a high level. Entering period N+2, when the calculated ray tip touches the tension constraint baseline, triggering a stretching anomaly, geometric ray projection is invoked to generate a reverse convergence vector, and the independent observation matrix is... The radical stripping component in the process is directly reduced to zero, and the Kalman gain of the linkage experiences a cliff-like decline. Smooth truncation may lead to a mutation instruction that causes bone fragments to fly out.

[0078] Furthermore, extract the independent observation matrix for the current period. The newly acquired collision retreat observation residual is calculated by subtracting the mapping component of the prior state estimation vector in the observation space using an arithmetic logic unit; and by using a Kalman gain weight matrix. Numerical weighted multiplication filtering is performed on the collision retreat observation residuals, and the product is superimposed on the historical smooth pose (prior state estimation vector). The above describes the update of the a priori state. The pre-introduced observation noise covariance matrix R, in physical engineering, characterizes the confidence uncertainty of the data fluctuation noise carried by the current input environmental interference feedback itself; as a specific application example, the diagonal elements of the observation noise covariance matrix R are preferably configured to be limited to the interval [0.01, 0.05]. The Kalman gain weight matrix is ​​based on this optimal gain coefficient. The observed spatial displacement compensation parameters are used as the observation gain and fused with the 3D spatial transformation matrix of the current loop frame for weighted calculation. Quantitative correction is performed on the transformation matrix corresponding to the current calculation frame in the target alignment deviation direction, generating a posterior state vector in a 6×1 column vector data structure. Inverse data unpacking of the spatial dimension is performed on the posterior state vector, extracting its internal 3D translation and rotation components. Simultaneously, the pre-stored soft tissue tension constraint baseline is acquired, and displacement extremum verification is performed on the 3D translation components. Abnormal coordinate offset values ​​exceeding the ligament's physical tensile limit are forcibly stripped and truncated. A homogeneous coordinate transformation algorithm is called to re-encapsulate the verified 3D translation and rotation components, thereby generating a 4×4 dimension smoothed correction transformation matrix based on displacement extremum smoothing. This smoothed correction transformation matrix is ​​then used to update the current transformation matrix corresponding to the current calculation frame via memory addressing.

[0079] This embodiment cyclically executes the above-mentioned closed-loop deduction of "comparison-game-anti-penetration analysis-quantitative correction", and extracts the comprehensive comparison deviation scalar of the current frame in real time within each calculation cycle; The status monitoring module continuously compares the decreasing gradient of the comprehensive comparison deviation scalar until it is monitored that the comprehensive comparison deviation scalar has stabilized and fallen back to the safe envelope of the preset reliability threshold (indicating that the three-dimensional pose of the current bone block after fine-tuning has once again met the clinically acceptable comprehensive error tolerance boundary), and immediately triggers the loop termination command. At this point, the transformation matrix parameters, which are in a convergent state globally, are solidified. The initial 3D surface mesh vertex set of each free fracture fragment is extracted from the data persistence layer. The 3D coordinates (x, y, z) of each discrete vertex in the vertex set are extracted, and the dimension filling logic is called to add a constant scaling dimension component (statically assigned a constant value of 1) to each 3D coordinate, thus completing and constructing a 4×1 homogeneous coordinate column vector (x, y, z, 1) matching the matrix operation dimension. The solidified 4×4 dimension transformation matrix parameters are used to perform matrix multiplication mapping operations on this homogeneous coordinate column vector to solve for the transformed homogeneous coordinate results. The scaling dimension component at the last position is removed to calculate the 3D absolute spatial coordinate array of each fracture fragment in the final reduction state. The absolute spatial coordinate array and the corresponding mesh patch topology are structurally encapsulated to generate the final digital 3D fracture reduction spatial model. This model is then sent to the external display and control terminal via a communication interface to guide clinicians in performing the actual reduction surgery. During the background loop execution of quantitative correction and ICP iteration to find the optimal solution, the spatial registration time and alignment residual deviation sequence when the preceding spatial registration node generates the three-dimensional spatial transformation matrix are extracted simultaneously.

[0080] This embodiment extracts the resetting difficulty index based on the time-series fluctuation data stream, denoted as... Its determining logic is as follows: By extracting the spatial registration time of the current multi-frame sequence (e.g., single-frame ICP solution time), using the maximum-minimum normalization function. Mapped to a uniform scale; alignment deviation sequence of the most recent few frames extracted synchronously. The time series variance parameter Variance is calculated to characterize the severity of error oscillations. This refers to the historical alignment residual temporal variance parameter; this embodiment sets a first-in-first-out queue data structure with a temporal sliding data window. The fixed length parameter of this queue is matched to the total number of independent computation frames (20 consecutive frames) in the most recent preset computation cycle. Two data dimensions are defined within the temporal sliding data window: the first dimension is used to synchronously push the spatial registration time record when the preceding spatial registration node generates a single-frame 3D spatial transformation matrix; the second dimension is used to synchronously push the ICP alignment residual scalar value corresponding to the current loop frame. During the cyclic execution of quantitative correction, whenever a new computation frame data is generated and pushed to the tail of the queue, a shift register action is triggered, and the oldest historical data frame at the head of the queue is automatically popped and destroyed, thereby continuously maintaining a dynamically updated but constant-length recently reset state evolution observation section.

[0081] This embodiment addresses the time consumption of spatial registration across multiple frames. The normalization process based on data boundary mapping involves: continuously scanning and locking the maximum and minimum extreme points of latency consumption within a time-series sliding data window; subtracting the latency data corresponding to the minimum extreme point from the actual latency features captured in the current calculation frame, and dividing this difference by the physical difference range between the maximum and minimum extreme points, thereby mapping it to a dimensionless scaling parameter confined to the closed interval [0,1]. This is done simultaneously with the intercepted alignment deviation sequence. Extract the independent registration residual distance scalars of several consecutive frames within the time-series sliding data window. First, perform a traversal summation operation to obtain the arithmetic mean residual value. Then, extract the residual data of each frame within the window and perform a subtraction operation with the arithmetic mean residual value. Perform a square operation on all the resulting differences and then perform a global accumulation and aggregation. Finally, divide the accumulated total value by the total number of buffered frames in the data window to deduce and deconstruct the variance characteristic quantity used to characterize the dispersion of physical interference oscillations.

[0082] In this embodiment, both are multiplied by a preset system balance weight. and Then, weighted aggregation is performed to output a resetting engagement difficulty index that characterizes the physical fit resistance. Among them, the balancing weight Characterized by the penalty coefficient for computational resource consumption, balancing weights The penalty coefficient characterizes the oscillation of physical alignment error (persistent interference collisions between bone blocks), and both satisfy the following conditions: In this embodiment, based on the clinical bottom line of orthopedic reduction where precision must be guaranteed even at the cost of longer reduction times, The quantization value is limited to the range [0.1, 0.3], preferably 0.2. Limited to the range of [0.7, 0.9], preferably 0.8.

[0083] Pre-initialize the upper limit parameter for the number of consecutive iterations In this embodiment, the The preferred setting is 20 consecutive calculation frames. When the reset difficulty index is monitored... If the preset alignment failure threshold is exceeded and the number of frames maintained in this abnormal state reaches 20, the geometric optimization calculation will be terminated and the system's abnormal interception strategy will be directly triggered. The preset alignment failure threshold was obtained through offline analysis of a historical database of severely comminuted control group samples who ultimately underwent subtalar articular fusion surgery due to excessive articular surface damage. The system also extracts the repositioning alignment difficulty index for all similar historical failure scenarios during virtual assembly. The peak value is fitted with a Gaussian probability distribution, and the lower bound of the confidence boundary where the cumulative probability reaches 95% is extracted from the distribution model. This lower bound is then mapped and quantized as a preset merging failure threshold. In this embodiment, the preset merging failure threshold is preferably calibrated to 0.85. In applications that are compatible with different surgical environments and equipment computing resource tolerances, the preset merging failure threshold is configured within the numerical range of [0.80, 0.95].

[0084] Furthermore, based on real-time access to medical imaging interfaces, source data of severe cases involving complete comminuted destruction of articular surfaces are collected. A time-series sliding data window continuously captures and extracts resource consumption data of spatial registration nodes and residual oscillation scalars of small spatially isolated fragments over multiple consecutive computation cycles. These features are then synchronously pushed into a first-in, first-out (FIFO) queue mechanism to drive the anomaly monitoring link to perform circuit breaker-level quantitative tracking.

[0085] Table 3: Reset Alignment Difficulty Index Monitoring and Early Warning Scheduling Array Based on Time Consumption and Alignment Deviation Timing Evolution When the simulation reaches frame K+8, due to the topological chaos of the free bone paste and bone fragments, the difficulty index of resetting and merging exceeds the 0.85 threshold, triggering an abnormal state that continues to accumulate. At frame K+27, the counter hits the pre-configured maximum number of consecutive iterations over 20 frames. Immediately cut off the tensor data feed to the large model and ICP iteration engine, directly extract the three-dimensional envelope feature set of the currently damaged joint surface, call the independent rendering pipeline to overlay a red warning mask for the comminuted area in the surgeon's interactive display window, and simultaneously issue a structured early warning message to abandon the splicing and switch to the subtalar joint fusion alternative intervention strategy.

[0086] Further triggering of abnormal interception: Based on the current distribution status of free fracture fragments, generate an early warning signal to suggest alternative surgical plans (such as recommending subtalar joint fusion), and send the alternative surgical plan early warning signal to the display and control terminal to trigger the interactive rendering closed loop to suggest alternative intervention strategies (e.g., forcibly highlight the subtalar joint surface topology area that needs to be fused due to comminution in the display and control window).

[0087] Introducing a reserved interface for individualized internal fixation intervention, denoted as It represents a standardized communication gateway that sends the post-reduction spatial pose topology to the external 3D printing engine or surgical navigation equipment. The standardized communication gateway defines a data structure payload format, which includes a sequence of vertex coordinates generated from the analysis of the current reduced bone block's 3D surface mesh, the topological relationships of the facet normals, and the finalized feature set of the 3D fracture reduction spatial model. Activation The parameters of the peripheral point cloud of the free fracture fragments in the current reset state are encapsulated in a structured data structure payload format and then sent to external hardware devices in a targeted manner.

[0088] Synchronously configure the model's continuous evolution feedback interface, denoted as... Its representation is used for the cloud uplink to transmit real postoperative validation data, ensuring closed-loop adaptive iteration of the large language model and image weights. The interface internally encapsulates a standardized reverse feedback data message format. This message format includes a standardized set of reset status features for the patient's current reset cycle (recorded as inference input), and "final intervention success / failure indicator" or "subjective manual adjustment parameter increment" entered by clinicians post-surgery based on actual weight-bearing results (recorded as expert-level annotation output). After completing the full validation process, it is activated... The aforementioned message data is uploaded to a large image database in the cloud and used as a new positive and negative feedback sample pair. This data is then directly input into the preceding instruction fine-tuning dataset based on the low-rank adaptive mapping algorithm (LoRA), thereby triggering the periodic adaptive iteration of the model weights.

[0089] Figure 1This case demonstrates the system-level closed loop driving multimodal data flow within a digital orthopedic workstation. The left side of the screen displays the 3D point cloud topology of the calcaneus reconstructed from the patient's multimodal CT voxel data. The "Anchoring Registration Origin and 3D Matrix Solving" node in the roadmap extracts the relatively stable suspensory process center position from the calcaneus anatomical partition identifier set, locking it as the static spatial reference origin for registration. An iterative nearest-point model is then used to extract the 3D spatial transformation matrix mapping each fracture fragment to the healthy side's mirror topology, eliminating spatial drift and divergence caused by the lack of absolute reference. The data link then enters the "Multi-source Heterogeneous Feature Dimensionality Reduction Extraction" and "Multi-dimensional Medical Rule Joint Evaluation" stages. At this point, the local detail window in the image intuitively reflects the extraction of 2D contour extreme points to generate visually measurable anatomical parameters, and the calculation of the high-density cortical bone contact ratio score extracted from the fluoroscopic fracture section. These heterogeneous features are then uniformly encapsulated and sent to the medical rule knowledge base for multi-dimensional rule comparison. When the reduction deviation is determined to exceed the confidence threshold, the "physical boundary anti-penetration closed-loop correction" logic is immediately activated. The polygonal interference boundary surrounding the outside of the bone block in the image represents the rigid body mechanics and soft tissue tension constraint baseline injected during the analytical fine-tuning strategy. By injecting the calculated spatial displacement compensation parameters back into the preceding node to perform quantitative correction, fine-tuning evolution is achieved while ensuring the physical interference baseline. Finally, a three-dimensional fracture reduction spatial model with structural stability that combines anatomical shape and load-bearing robustness is output.

[0090] Figure 2 The X-axis represents the cyclic calculation frame sequence offset, and the Y-axis represents the scalar value of the mapped reset matching difficulty index and the abnormal state retention count. Specifically, marker I1 actually represents the critical value of 0.85, the preset matching failure threshold used to trigger anomaly detection; while visual state marker F1 accurately reflects the value of 20, the upper limit parameter for the number of consecutive iterations to prevent local deadlock, derived from logic; marker K1 represents the value of 8, the cyclic iteration frame sequence offset when the reset matching difficulty index exceeds the red line; and marker C1 reveals the absolute peak value of the reset matching difficulty index of 0.942 when the computation link is finally blocked and the limit boundary interception is triggered. Figure 2 As can be seen from the mapped temporal data flow mechanism, after receiving the heavily fragmented post-articular voxel sequence, the spatial alignment residual and matrix computation overhead are extracted synchronously with weights. As the computation frames progress, the entity curve representing the reset alignment difficulty index rapidly climbs at frame sequence node K1 and exceeds the diagnostic red line of 0.85. The reset alignment difficulty index maintains high-level fluctuations within a high-frequency oscillation range of 0.902 to 0.942. In this embodiment, a rigidly monotonically increasing abnormal state holding counter is activated. This scheme continuously tracks the state based on a temporal sliding data window until the counter exceeds the F1 safety upper limit red line of 20.

[0091] like Figure 3As shown, a 3D point cloud reconstruction is performed based on a 3D spatial transformation matrix. The 3D mesh depth mapping module is then invoked to orthogonally project the reconstructed 3D point cloud onto a reference plane. During the projection process, the depth difference data of the 3D point cloud surface relative to the projection reference plane is extracted simultaneously, generating terrain depth array data containing 3D spatial undulation information. An edge contour search algorithm is used to search for depth extrema points in the terrain depth array data to identify key peak and valley markers and generate a set of visual measurement parameter vectors containing anatomical angles and dimensions. Through this entity elevation extraction mechanism based on orthogonal projection, the complex 3D physical undulation features are dumped into a single-layer terrain depth array data, eliminating projection deformation errors caused by the spatial tilt of fracture fragments at the execution level.

[0092] like Figure 4 As shown in the node structure, a three-dimensional spatial interference boundary is established for the three-dimensional point cloud envelope of the free fracture fragment. Based on the three-dimensional spatial transformation matrix corresponding to the current calculation frame in the previous spatial registration node, the three-dimensional relative displacement vector of the free fracture fragment relative to the static spatial registration reference origin is extracted. The three-dimensional non-uniform tensile limit envelope surface bound to the corresponding calcaneal anatomical partition is retrieved as the soft tissue tension constraint baseline, and the three-dimensional spatial interference penetration between the end coordinate point of the three-dimensional relative displacement vector and the three-dimensional non-uniform tensile limit envelope surface is calculated using three-dimensional mesh collision detection logic. If it is determined that the end coordinate point breaks through the three-dimensional non-uniform tensile limit envelope surface, the spatial distance of the end coordinate point back to the envelope surface along the shortest normal path is solved in three-dimensional space, and a boundary projection convergence vector is generated based on the spatial distance and the direction of the shortest normal path. The boundary projection convergence vector is used to perform reverse numerical compensation on the current three-dimensional relative displacement vector, quantify the abnormal translation components in the stripping matrix that exceed the physical tensile limit of the ligament of the free fracture fragment, and generate a smooth correction transformation matrix after smoothing processing based on displacement extrema.

[0093] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. An AI-based reduction method for calcaneal fractures based on multimodal data, characterized in that, Performed by the medical image processing system, the specific steps include: Step S1: Obtain multimodal CT voxel data with gray values ​​from the affected and healthy sides of the patient; the image processing module calls the multi-label image segmentation model to separate the calcaneal anatomical partition identifier set containing multiple anatomical structures, and extracts the three-dimensional spatial center position of the suspensory process anatomical label that is relatively stable due to ligament constraint from the calcaneal anatomical partition identifier set, and uses it as the static spatial registration reference origin for registration. Step S2: Based on the static spatial registration reference origin, generate a mirror topology map of the healthy side, and execute an iterative nearest point spatial alignment model to extract the three-dimensional spatial transformation matrix of each free fracture fragment aligned to the mirror topology map; Three-dimensional point cloud reconstruction is performed based on the three-dimensional spatial transformation matrix, and projection dimension reduction is performed on the standard anatomical sagittal projection system to extract the transitional two-dimensional contour feature vector. Step S3: Based on the two-dimensional contour feature vector, trigger the image contour feature extraction module, use the edge contour search algorithm to automatically detect key peak and valley markers, and generate a set of visual measurement parameter vectors containing anatomical angles and size values; The visual measurement parameter vector set and the three-dimensional spatial transformation matrix are uniformly encapsulated into a reset state feature set, which is then input into a comprehensive evaluation module with a pre-configured medical rule knowledge base for multi-dimensional data rule comparison. Step S4: If the deviation calculated by the comprehensive evaluation module exceeds a preset confidence threshold, then receive the fine-tuning strategy containing the spatial attitude fine-tuning intention returned by the comprehensive evaluation module based on the rule comparison result, and parse the fine-tuning strategy to generate spatial displacement compensation parameters; and, The spatial displacement compensation parameters are injected in reverse into the preceding spatial registration node. Quantitative correction is performed on the transformation matrix corresponding to the current calculation frame in the direction of the target alignment deviation until the comparison deviation meets the safety threshold. Finally, a finalized three-dimensional fracture reduction spatial model is output.

2. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 1, characterized in that: The steps of projecting and reducing the dimension of the transition to the standard anatomical sagittal projection system to extract the transitional two-dimensional contour feature vectors specifically include: The 3D mesh depth mapping module is invoked to orthogonally project the reconstructed 3D point cloud onto the standard anatomical sagittal projection system; During the projection process, the depth difference data of the three-dimensional point cloud surface relative to the projection reference plane is extracted simultaneously to generate terrain depth array data containing three-dimensional spatial undulation information. The edge contour search algorithm is used to search for depth extreme points in the terrain depth array data to lock the key peak and valley marker points and generate the visual measurement parameter vector set, thereby eliminating the projection deformation error caused by the spatial tilt of the fracture fragments.

3. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 2, characterized in that: The step of encapsulating the visual measurement parameter vector set and the three-dimensional spatial transformation matrix into a unified reset state feature set further includes: The gray values ​​representing bone density in the multimodal CT voxel data are pre-statically bound to the corresponding mesh vertices of the three-dimensional point cloud; After performing three-dimensional point cloud reconstruction based on the three-dimensional spatial transformation matrix, the gray values ​​on the splicing section of the free fracture fragments are extracted, filtered by the set bone cortex gray value filtering threshold, and the high-density cortical bone contact ratio score characterizing the load-bearing mechanical stability is calculated. The high-density cortical bone contact ratio score, along with the visual measurement parameter vector set, is encapsulated into the reset state feature set and input into the comprehensive evaluation module. This allows the comprehensive evaluation module to jointly determine the fine-tuning strategy based on both anatomical angle constraints and internal load-bearing contact area.

4. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 3, characterized in that: The steps for inputting data into the comprehensive evaluation module based on a medical rule knowledge base for multidimensional data rule comparison specifically include: A multi-dimensional feature mapping mechanism is constructed within the comprehensive evaluation module, setting the anatomical angle as the geometric appearance optimization target and setting the high-density cortical bone contact ratio score as the load-bearing contact benchmark threshold. The comprehensive evaluation module performs conditional comparison and screening on the reset state feature set. When the anatomical angle deviates from the preset anatomical angle baseline and triggers the stretching compensation rule, and at the same time, it is monitored that the high-density cortical bone contact ratio score is lower than the load-bearing contact benchmark threshold, a cutoff strategy is executed to intercept the correction command for the unidirectional stretching geometry angle. Trigger a sliding rotation search logic toward the Z-axis orthogonal plane to generate a suboptimal fine-tuning strategy that satisfies the load-bearing contact reference threshold.

5. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 4, characterized in that: The steps for parsing the returned fine-tuning strategy to generate spatial displacement compensation parameters specifically include: Establish a three-dimensional spatial interference boundary for the three-dimensional point cloud envelope of free fracture fragments; Extract the spatial attitude fine-tuning intent from the fine-tuning strategy returned by the comprehensive evaluation module, and calculate it as a virtual displacement correction vector applied to the dissecting node; In the spatial interference detection module, the spatial equilibrium state of the free fracture fragment under the action of the virtual displacement correction vector is solved based on the three-dimensional spatial interference boundary. When envelope penetration interference is detected, the translation and rotation increments used to counteract the penetration interference are calculated, and the translation and rotation increments are directly configured as the spatial displacement compensation parameters.

6. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 5, characterized in that: The step of injecting the spatial displacement compensation parameters in reverse into the previous spatial registration node, and performing quantitative correction on the transformation matrix corresponding to the current calculation frame in the target alignment deviation direction, specifically includes: Using the Kalman filter state update matrix equation, the spatial displacement compensation parameter is used as the observation gain and fused with the three-dimensional spatial transformation matrix corresponding to the current calculation frame in the preceding spatial registration node; In the direction of target alignment deviation, compare with the pre-stored soft tissue tension constraint baseline, strip the abnormal translation components in the matrix that exceed the physical tensile limit of the free fracture fragment ligament, generate a smooth correction transformation matrix that has been smoothed based on displacement extrema, and use the smooth correction transformation matrix to cover and update the current transformation matrix.

7. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 6, characterized in that: The method also includes a step of synchronously triggering a surgical difficulty warning process based on spatial registration nodes: During the cyclic execution of quantitative correction, the spatial registration time and alignment deviation sequence for generating the three-dimensional spatial transformation matrix from the preceding spatial registration node are simultaneously extracted. Based on the spatial registration time and alignment deviation sequence, a resetting and alignment difficulty index, which characterizes the resetting and alignment difficulty, is extracted. When the state of the repositioning and alignment difficulty index exceeding the preset alignment failure threshold is detected to meet the preset maximum number of consecutive iterations parameter, the system abnormal interception strategy is triggered, automatically blocking the link to request the default geometric repositioning calculation from the comprehensive evaluation module, and directly generating an early warning signal for prompting alternative surgical planning based on the current free fracture fragment distribution state, and sending the alternative surgical planning early warning signal to the display and control terminal to trigger the interactive rendering closed loop for prompting alternative intervention strategies.

8. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 4, characterized in that: The steps of triggering a sliding rotation search logic towards the Z-axis orthogonal plane to generate a suboptimal fine-tuning strategy that satisfies the load-bearing contact reference threshold specifically include: On the Z-axis orthogonal plane, with the centroid of the point cloud of the current free fracture fragment as the spatial reference origin, a two-dimensional sliding search grid covering the preset search range is constructed according to the preset sampling step size parameter; Traverse each discrete grid node in the two-dimensional sliding search grid and call the preset two-dimensional steady-state weighted aggregation rule to calculate the reset steady-state evaluation value corresponding to each discrete grid node; The execution logic of the two-dimensional steady-state weighted aggregation rule is as follows: extract the normalized high-density cortical bone contact ratio score of the location of the discrete grid node, and perform a subtraction operation between the score and the anatomical angle deviation value carrying a preset convergence penalty weight to generate the reset steady-state evaluation value. The target grid node whose reset steady-state evaluation value satisfies the preset steady-state convergence extreme value condition is extracted. The relative offset coordinates of the target grid node with respect to the spatial reference origin are analyzed into translation and rotation mapping vectors. The translation and rotation mapping vectors are used as the feedback of the suboptimal fine-tuning strategy to the preceding spatial registration node to perform subsequent quantitative correction steps.

9. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 6, characterized in that: The steps for comparing the abnormal translational components in the stripping matrix that exceed the physical tensile limit of the ligaments in the free fracture fragments with a pre-stored soft tissue tension constraint baseline include: Based on the three-dimensional spatial transformation matrix corresponding to the current calculation frame in the preceding spatial registration node, the three-dimensional relative displacement vector of the free fracture fragment relative to the static spatial registration reference origin is extracted. The three-dimensional non-uniform tensile limit envelope surface bound to the corresponding calcaneal anatomical partition is retrieved as the soft tissue tension constraint baseline, and the three-dimensional spatial interference penetration amount between the end coordinate point of the three-dimensional relative displacement vector and the three-dimensional non-uniform tensile limit envelope surface is calculated using three-dimensional mesh collision detection logic. If it is determined that the end coordinate point breaks through the three-dimensional non-uniform stretch limit envelope, then the spatial distance from the end coordinate point back to the envelope along the shortest normal path is calculated in three-dimensional space, and a boundary projection convergence vector is generated based on the spatial distance and the direction of the shortest normal path. The boundary projection convergence vector is used to perform inverse numerical compensation on the current three-dimensional relative displacement vector to quantize and remove the abnormal translation components.

10. The AI ​​reduction method for calcaneal fractures based on multimodal data according to claim 1, characterized in that: The step of separating the calcaneal anatomical partition markers by calling the multi-label image segmentation model from the image processing module further executes a projection assignment routing mechanism based on anatomical anchor points when encountering loose bone fragments with blurred edges. For isolated fragment point clouds that cannot be directly mapped to a preset dictionary and contain more voxels than the noise filtering threshold, the output of the multi-label image segmentation model is identified and extracted as bone fragments to be classified. Obtain the centroid of the solid point cloud of the bone fragments to be classified, and extract the fixed feature points at both ends of the calcaneal skeleton to construct a three-dimensional spatial anatomical anchor point vector; Based on the spatial inner product operation, the relative projection distance component of the centroid of the entity point cloud on the three-dimensional spatial anatomical anchor point vector is calculated. The relative projection distance component is mapped to the feature anchor point projection normalization coefficient to characterize the relative topological depth position of the free bone fragments on the main calcaneal morphological axis. Based on the quantized numerical distribution of the feature anchor projection normalization coefficient, the spatial nearest neighbor assignment operation is performed in the corresponding orthogonal coordinate system to attach the bone fragments to be classified to the determined calcaneal anatomical partition marker.

Images