Artificial intelligence-based calcaneal fracture four-zone automatic segmentation and measurement system

Abstract

This invention discloses an artificial intelligence-based automatic four-zone segmentation and measurement system for calcaneal fractures, relating to the field of three-dimensional image analysis technology. The method acquires the DICOM sequence of foot and ankle CT scans, converts it into NIfTI three-dimensional volume data, and performs voxel resampling, bone window normalization, and orientation standardization. A first segmentation network extracts a complete three-dimensional mask of the calcaneus, establishes a three-dimensional anatomical coordinate system, and clips the region of interest (ROI) of the calcaneus. A second segmentation network generates four-zone probability maps: the anterior region, the posterior region, the superior articular surface-related region, and the medial support region. The spatial assignment of structural blocks is determined by constraints from the complete calcaneal three-dimensional mask, connected component analysis, and the centroid and anchor point projection of the structural blocks. Finally, based on the complete calcaneal contour and the four-zone label map, geometric angles, surface height differences, the number of effective connected components, and the three-dimensional contact ratio are extracted and encapsulated into structured parameter files and quality control markers for easy three-dimensional display and image retrieval.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional image analysis technology, specifically to an artificial intelligence-based automatic segmentation and measurement system for four zones of calcaneal fractures. Background Technology

[0002] Medical image data processing, computer vision, and 3D image analysis are primarily used in hospital radiology departments, orthopedic trauma centers, preoperative image processing workstations, and medical imaging artificial intelligence software platforms. The data processed typically involves CT images of the foot and ankle, generally including image import, format conversion, voxel standardization, bony structure identification, local region extraction, spatial parameter calculation, and result visualization output. The calcaneus, located at the posterior end of the foot, has an irregular shape and is surrounded by the talus, cuboid, and multiple articular surfaces. While CT images can provide relatively complete 3D tomographic information, differences in raw data slice thickness, reconstruction kernel, scanning direction, and voxel spacing necessitate standardization and structuring before stable integration into the image analysis process.

[0003] Chinese patent document CN108491770A discloses a data processing method based on fracture images. This method imports fracture images into a data processing center for image recognition, measures predetermined parameter values ​​in the recognized images, and processes the measured data. This involves marking bone contour lines, fracture lines, marker lines, or marker points on the bones, measuring corresponding parameters according to different fracture classifications, storing these parameters in pre-set databases, and outputting the results. From this patent document, we can see that this type of method represents a basic "image recognition-parameter measurement-result output" approach in fracture image data processing research, but its implementation primarily focuses on the general recognition and parameter processing of fracture images.

[0004] A non-patented document more closely related to foot and ankle CT scenarios is "Real-time Automatic Segmentation and Classification of Calcaneal Fractures in CT Images" published by Rahmaniar and Wang in 2019. This article proposes a computer-aided method for calcaneal fracture CT images: the goal is to obtain faster and more detailed observation results through automated segmentation and classification. The document processes calcaneal fracture CT images, and its technical workflow includes: image preprocessing, candidate region processing, fracture region segmentation, and classification output. This research reflects the trend of using computer vision algorithms to assist in calcaneal fracture image analysis. Additionally, Farda et al. published a study in 2021 titled "Deep Learning Network + Data Augmentation for Sanders Classification of Calcaneal Fractures in CT Images," indicating that research paths already exist in this field that utilize deep learning to obtain classification results from CT images.

[0005] While the aforementioned existing technologies can achieve fracture image recognition, region segmentation, and parameter measurement or classification output to a certain extent, they still have shortcomings in processing the calcaneal structure of foot and ankle CT.

[0006] First, the calcaneus is not a regular long bone; its shape is irregular and blocky. It is also relatively close to the talus, cuboid, and surrounding bony structures. If relying solely on standard fracture image recognition procedures, local boundaries are easily affected by adjacent cortical bone, tomographic artifacts, and fragmented areas. Second, because CT data originates from different devices and scanning protocols, slice thickness, voxel spacing, and orientation matrices are not uniform. The same structure can exhibit different spatial scales and orientations in different cases. Without a unified spatial benchmark, subsequent region segmentation and parameter measurement are prone to coordinate drift. Third, existing classification schemes often focus on the final classification result, rarely providing unified encapsulation of segmentation boundaries, structural block attribution, parameter sources, and abnormal quality states. This makes it difficult to stably track the image region corresponding to each parameter during subsequent 3D display, verification, and retrieval. Therefore, the technical problem to be solved in this field is: to stably obtain image segmentation results and retrievable structural parameters for the calcaneus-related region in foot and ankle CT images with varying scales, orientations, interference from adjacent bones, and discontinuous local structures. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an AI-based automatic four-zone segmentation and measurement system for calcaneal fractures. The system extracts a complete 3D mask of the calcaneus using a first segmentation network, establishes a 3D anatomical coordinate system, and trims the region of interest (ROI) of the calcaneus. A second segmentation network generates four-zone probability maps: the anterior region, the posterior region, the superior articular surface-related region, and the medial support region. The spatial assignment of structural blocks is determined through constraints imposed by the complete 3D mask, connected component analysis, and the centroid and anchor point projection of the structural blocks. Furthermore, based on the complete calcaneal contour and the four-zone label map, geometric angles, surface height differences, the number of effective connected components, and the 3D contact ratio are extracted and encapsulated into a structured parameter file and quality control markers for easy 3D display and image retrieval. This system solves the technical problems described in the background art.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] The AI-based automatic segmentation and measurement system for four zones of calcaneal fracture includes: receiving DICOM sequences and converting them into NIfTI three-dimensional volume data after resampling, bone window normalization, and RAS standardization; inputting the NIfTI three-dimensional volume data into a first segmentation network to obtain a complete calcaneal three-dimensional mask; and establishing a three-dimensional anatomical coordinate system, normalized anatomical coordinates, and calcaneal region of interest based on the complete calcaneal three-dimensional mask.

[0012] The region of interest of the calcaneus is input into the second segmentation network to obtain a four-zone probability map of the anterior region, the posterior region, the upper articular surface related region, and the medial support region. The four-zone probability map is restricted to the complete 3D mask of the calcaneus to generate a four-zone label map. The spatial assignment of the structural block is determined by combining the 3D connected domain, the centroid of the structural block, and the anchor point projection parameters.

[0013] Geometric angles, surface height differences, number of effective connected components, and 3D contact ratio are extracted based on the complete 3D mask of the calcaneus and the four-zone label map.

[0014] The four-zone label map, structural block spatial assignment, geometric angle, surface height difference, number of effective connected domains, and three-dimensional contact ratio are encapsulated into a structured parameter file, and quality control tags are output.

[0015] Furthermore, when the data processing server converts the DICOM sequence into NIfTI 3D volume data, it simultaneously reads the image position, image orientation matrix, pixel spacing, and layer spacing, and generates an affine mapping relationship between the physical coordinate system of the DICOM sequence and the voxel coordinate system of the NIfTI 3D volume data.

[0016] The data processing server establishes a three-dimensional anatomical coordinate system based on the physical coordinates of voxel points in the complete 3D mask of the calcaneus, and maps the voxel points in the complete 3D mask of the calcaneus to normalized anatomical coordinates in the posteroanterior, lateral-medial, and inferior-superior directions, respectively. The region of interest of the calcaneus is obtained by cropping the three-dimensional bounding box of the complete 3D mask of the calcaneus, and the mapping relationship between the region of interest of the calcaneus and the physical coordinate system of the DICOM sequence is preserved.

[0017] Furthermore, the data processing server aligns the four-zone probability map output by the second segmentation network with the complete calcaneal 3D mask voxel by voxel.

[0018] When a voxel is located outside the complete 3D mask of the calcaneus, the voxel is assigned the background value; when a voxel is located inside the complete 3D mask of the calcaneus and its maximum partition probability reaches a preset probability threshold, the voxel is assigned the partition label corresponding to the maximum partition probability; when a voxel is located inside the complete 3D mask of the calcaneus and its maximum partition probability does not reach the preset probability threshold, it is completed based on the support strength of the partition labels already determined in the 26 neighborhoods of the 3D model.

[0019] If the support strength has not yet reached the completion threshold, mark the voxel as a non-critical structural block.

[0020] Furthermore, the data processing server extracts the sagittal projection contour based on the complete calcaneal 3D mask, and determines the geometric angles from the sagittal projection contour.

[0021] The data processing server extracts the surface height difference and the number of effective connected components based on the relevant region of the upper joint surface in the four-zone label map, and extracts the three-dimensional contact ratio based on the rear region and the relevant region of the upper joint surface in the four-zone label map.

[0022] The objects extracted based on geometric angles are independent of those extracted based on surface height difference, number of effective connected regions, and 3D contact ratio, but are all written into the structured parameter file.

[0023] Furthermore, after receiving the DICOM sequence, the data processing server verifies the image position, image orientation matrix, pixel spacing, and layer spacing. When the image position, image orientation matrix, pixel spacing, and layer spacing are all complete, it performs NIfTI 3D volume data conversion, resampling, and orientation normalization. When the layer spacing is missing and the image positions of adjacent slices are continuous, the layer spacing is reconstructed from the image position differences of adjacent slices, and then NIfTI 3D volume data conversion is performed.

[0024] When the image orientation matrix is ​​missing and the orientation cannot be reconstructed from adjacent slice positions, a quality control marker for image orientation verification is generated, and intermediate data that retains the original slice order is output.

[0025] Furthermore, the data processing server performs main connected component ratio verification, hole filling status verification, and boundary coverage status verification on the complete calcaneal 3D mask;

[0026] When the complete 3D mask of the calcaneus passes the main connected component proportion verification, hole filling status verification, and boundary coverage status verification, the region of interest (ROI) of the calcaneus is clipped based on the complete 3D mask of the calcaneus. When the complete 3D mask of the calcaneus fails any of the verifications, the clipping boundary of the ROI of the calcaneus is expanded, and the quality status corresponding to the complete 3D mask of the calcaneus is written into the structured parameter file. When the expanded ROI of the calcaneus still fails the boundary coverage status verification, the inference results of the four-zone probability map are retained, and a quality control mark indicating insufficient complete calcaneus segmentation quality is generated.

[0027] Furthermore, before outputting the structured parameter file, the data processing server performs field integrity checks on the image data number, CT voxel spacing, complete calcaneal mask quality index, four-zone segmentation results, structural block centroid, geometric angle, surface height difference, number of effective connected components, three-dimensional contact ratio, and quality control markers.

[0028] When all fields exist and their format conforms to the field template, output a structured parameter file and a parameter visualization image; when at least one field is missing, output a structured parameter file containing the reason for the missing field and stop generating the parameter visualization image; when a field exists but the parameter source conflicts, retain the original field value and write the name of the conflicting field and the triggering reason in the quality control mark.

[0029] Furthermore, the first segmentation network and the second segmentation network are deployed in the same model service, which includes the first output branch and the second output branch;

[0030] The first output branch receives NIfTI 3D volume data and outputs a complete calcaneal 3D mask and a calcaneal region of interest (ROI); the second output branch receives the ROI generated by the first output branch and outputs a four-zone probability map; the same model service records the model version of the first output branch, the model version of the second output branch, the file identifier of the complete calcaneal 3D mask, and the file identifier of the four-zone probability map in a single processing task.

[0031] Furthermore, when establishing a three-dimensional anatomical coordinate system, the data processing server selects the coordinate system generation strategy according to the strategy order in the configuration file; when the main connected component of the complete calcaneal three-dimensional mask satisfies the morphological integrity condition, the three-dimensional anatomical coordinate system is generated using the principal axis direction of the voxel points; when the main connected component does not satisfy the morphological integrity condition and the set of endpoints satisfies the separation condition, the three-dimensional anatomical coordinate system is generated using the posterior endpoint, anterior endpoint, upper endpoint, and lower endpoint.

[0032] When the main connected component does not meet the morphological integrity condition and the endpoint set does not meet the separation condition, a three-dimensional anatomical coordinate system is generated by rigid registration between the preset calcaneal template and the complete calcaneal 3D mask, and the selected strategy identifier is written in the quality control mark.

[0033] Furthermore, the structured parameter file includes a JSON field file, NIfTI label extension fields, and a 3D mesh model index file; the JSON field file records image data number, CT voxel spacing, structural block spatial assignment, geometric angle, surface height difference, number of effective connected components, 3D contact ratio, and quality control markers; the NIfTI label extension fields record the label values ​​of the four-zone label map in voxel space;

[0034] The 3D mesh model index file records the file path, partition name, block number, and centroid coordinates of each structural block's 3D mesh model; the JSON field file, NIfTI tag extended fields, and the 3D mesh model index file share the same image data number.

[0035] Furthermore, the data processing server generates a parameter visualization interface, which displays the four-zone label diagram, structural block number, structural block centroid, geometric angle, surface height difference, number of effective connected domains, three-dimensional contact ratio, and quality control markers. The parameter visualization interface displays the front region, rear region, upper joint surface related region, inner support region, and non-critical structural blocks with fixed field names. When quality control markers exist, the parameter visualization interface displays the triggering reason for the quality control markers. When no quality control markers exist, the parameter visualization interface displays the status information of the generated structured parameter file.

[0036] Furthermore, the data processing server provides a parameter export interface, which receives export requests containing image data number, file type field, and request time field.

[0037] When the file type field is a structured parameter file, the parameter export interface returns the file address, field summary, and quality control flag of the structured parameter file; when the file type field is a four-zone segmentation label graph, the parameter export interface returns the file address, voxel spacing, and label name list of the four-zone label graph; when the file type field does not belong to the preset file type, the parameter export interface returns a response message without the parameter file address and writes the unsupported file type flag in the response message.

[0038] Furthermore, the data processing server generates a processing log for each processing task. The processing log records the image data number, DICOM sequence reception time, NIfTI 3D volume data generation status, complete calcaneal 3D mask generation status, calcaneal region of interest generation status, four-zone probability map generation status, four-zone label map generation status, structural block spatial assignment generation status, structured parameter file generation status, and quality control mark generation status.

[0039] When any generation status is a failure, the processing log records the corresponding failure stage name and input file identifier;

[0040] When all generation statuses are successful, the processing log records the address of the structured parameter file, the address of the four-zone segmentation label map, and the address of the parameter visualization image.

[0041] (III) Beneficial Effects

[0042] This invention provides an artificial intelligence-based automatic four-zone segmentation and measurement system for calcaneal fractures, which has the following beneficial effects:

[0043] By converting foot and ankle CT images into three-dimensional volumetric data and performing operations such as voxel resampling, bone window processing, orientation standardization, and physical coordinate mapping, data generated by different devices, slice thicknesses, and positioning can enter the same spatial reference, and subsequent segmentation, parameter extraction, and write-back all have the same coordinate source.

[0044] First, the first segmentation network obtains the complete 3D mask of the calcaneus. Then, the second segmentation network generates labels for the anterior region, posterior region, superior articular surface related region, and medial support region within the region of interest of the calcaneus. This not only completely determines the global localization but also delineates the local partitions, avoiding the boundary confusion caused by a single model only handling foot and ankle localization and calcaneal internal partitioning. The complete 3D mask of the calcaneus is used to constrain the probability map of the four regions. Combined with 3D connected domains, structural block centroids, anchor point projection parameters, and normalized anatomical coordinates, the spatial assignment of structural blocks is determined. Low-confidence voxels, isolated small connected domains, and boundary structural blocks all have a processing destination, making the four-region label map more stable.

[0045] By extracting geometric angles from the complete calcaneal contour, and then extracting information such as surface height difference, number of effective connected domains, and three-dimensional contact ratio with the posterior region from the relevant area of ​​the upper articular surface, the overall contour parameters are decoupled from the local connected structure parameters, reducing the impact of local discontinuities on contour measurement.

[0046] The four-zone label map, structural block spatial assignment, geometric parameters, connectivity parameters, and quality control markers are encapsulated into a set of structured data for consistency verification before output. This ensures that the source of required parameters, the cause of anomalies, and the visualization results share the same data output, making it easy to implement 3D display and image analysis calls. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall process of foot and ankle image processing according to the present invention;

[0048] Figure 2 This is a schematic diagram of the standardization and spatial mapping process of foot and ankle tomographic images according to the present invention;

[0049] Figure 3 This is a schematic diagram of the complete calcaneal 3D mask, 3D anatomical coordinate system, and calcaneal region of interest generation of the present invention.

[0050] Figure 4 This is a schematic diagram of the fusion of the probability map of the four regions of the calcaneus and the complete calcaneus mask constraint according to the present invention;

[0051] Figure 5 This is a schematic diagram illustrating the arrangement of connected domains and anchor point projections of the structural blocks in this invention.

[0052] Figure 6 This is a schematic diagram illustrating the measurement of the sagittal projection contour and geometric angle parameters of the calcaneus according to the present invention.

[0053] Figure 7 This is a schematic diagram of local parameter measurements in the relevant area of ​​the upper articular surface of the present invention;

[0054] Figure 8 This is a schematic diagram of the structured parameter package, rule quality verification, and output of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figures 1-8 This invention provides an artificial intelligence-based automatic four-zone segmentation and measurement system for calcaneal fractures, comprising: Step 1, converting foot and ankle CT image data from different sources, with varying orientations and voxel spacings, into a standardized data environment with a unified spatial reference, a unified grayscale window, and a unified calcaneal representation object. This provides a stable calcaneal region of interest for Step 2, four-zone semantic segmentation. With complete calcaneal 3D mask and three-dimensional anatomical coordinate system .

[0057] Foot and ankle CT image data to be processed is usually stored in DICOM sequence format. Different scanning devices, reconstruction kernels, slice thicknesses, scanning directions, and foot placement methods can cause inconsistencies in the voxel scale, grayscale distribution, and spatial orientation of the same calcaneus. If this type of data is directly input into the subsequent four-region segmentation network, the network needs to simultaneously handle anatomical target recognition, orientation adaptation, and scale correction, which can easily lead to interference from unstructured differences in the segmentation boundaries. Therefore, step one first extracts the pixel spacing, slice thickness, image orientation matrix, and physical coordinate information from the DICOM sequence, and converts the DICOM sequence into NIfTI three-dimensional volume data, so that each voxel point has a traceable spatial position.

[0058] In this process, the discrete slices formed at the image acquisition end are stacked and transformed into unified data that can be used by 3D convolutional networks, 3D connected component algorithms, and spatial projection algorithms.

[0059] For example, a set of foot and ankle CT scan folders is imported into the workstation. The data processing server first sorts the slices by DICOM instance number and image location, then reads the pixel spacing and slice thickness of each slice, assembles the grayscale matrix into NIfTI 3D volume data, and if the image orientation matrix is ​​found to be inconsistent with the RAS coordinate system, it is corrected to a unified orientation through orientation rearrangement and voxel resampling. After this operation is completed, the original slice files are retained, and the standardized CT 3D volume data becomes the input of the subsequent first deep learning segmentation network.

[0060] To ensure that the original CT physical space and the normalized voxel space can be mapped to each other, this step saves the affine transformation matrix from the DICOM physical coordinate system to the NIfTI voxel coordinate system. And save its inverse transformation matrix. For any original physical coordinate point Its corresponding coordinates in the NIfTI voxel coordinate system Determine using the following formula:

[0061]

[0062] Among them, the original physical coordinate points : Represents the three-dimensional homogeneous coordinates in the DICOM physical space, with values ​​determined by the image position, image orientation, and pixel spacing within the DICOM sequence; NIfTI voxel coordinates. : Represents the homogeneous 3D coordinates in the standardized NIfTI 3D volume data, with values ​​corresponding to the voxel index positions after resampling; affine transformation matrix : Represents a fourth-order matrix mapping from the DICOM physical coordinate system to the NIfTI voxel coordinate system; inverse transformation matrix Affine transformation matrix The inverse matrix is ​​used to write back the subsequent segmentation results and spatial parameters to the original CT physical space.

[0063] The complete 3D mask of the calcaneus referenced in subsequent steps calcaneal region of interest Three-dimensional anatomical coordinate system and normalized anatomical coordinates , , All are based on the same spatial reference. Therefore, the region of interest of the calcaneus can be directly utilized. Perform four-region semantic segmentation to reduce the interference of non-anatomical factors on the partition probability map.

[0064] In one implementation, the DICOM sequence of the 1st... Layer, First line, number Voxels, first from image position vectors unit vector in the row direction Column-direction unit vector Interlayer direction unit vector , row pixel spacing Column pixel spacing and interlayer spacing Calculate DICOM physical coordinates :

[0065]

[0066] Among them, DICOM physical coordinate points Represents the three-dimensional coordinates in the original DICOM physical space; image position vector Represents the physical coordinates of the top-left voxel of the DICOM first-layer image; unit vector in the row direction. Represents the row direction and column direction of a DICOM image, respectively, as unit vectors. Indicates the column direction of a DICOM image; interlayer direction unit vector. Depend on and The cross product determines the row pixel spacing; Column pixel spacing and interlayer spacing All are positive numbers, in millimeters; row index Column index and layer index All are non-negative integers.

[0067] DICOM physical coordinates After writing it in homogeneous coordinates, it is transformed by an affine transformation matrix. Mapped to NIfTI voxel coordinates :

[0068]

[0069] Among them, NIfTI voxel coordinate points Represents coordinates in the normalized NIfTI space; affine transformation matrix Represents the rotation, translation, scaling, and orientation rearrangement relationships from the DICOM physical coordinate system to the NIfTI voxel coordinate system; inverse mapping matrix. This refers to the matrix written back from the NIfTI voxel coordinate system to the DICOM physical coordinate system.

[0070] In one implementation, the initial mask of the calcaneus is analyzed using a three-dimensional 26-neighbor connected component analysis, with two voxels. and The conditions for being considered adjacent are:

[0071]

[0072] Among them, voxel coordinates Voxel representation Three index coordinates in NIfTI voxel space; voxel coordinates Voxel representation The three index coordinates are used. The connected component with the largest number of voxels is retained as the main calcaneal connected component. Hole filling is performed using a 3D closing operation or background connected component inverse filling method; the small voxel noise removal threshold can be set to... The volume of the complete calcaneus for The product of the number of intrinsic primes and the physical volume of the monomer primes.

[0073] After obtaining standardized CT 3D volumetric data, this step inputs the data into the first deep learning segmentation network. The input to the first deep learning segmentation network is single-channel CT 3D volumetric data, and the output is the multi-label segmentation result of the bony structures in the foot and ankle region. The calcaneal label is extracted into a calcaneal 3D mask. The entire encoder-decoder architecture exemplifies the use of a 3D convolutional structure. The encoder extracts cortical bone edges, cancellous bone texture, and gaps between adjacent bones; the decoder recovers spatial details through upsampling and skip connections. The output layer uses a softmax function to provide the probability that each voxel belongs to each bony structure label. The softmax function is a normalized classification function readily implementable in this field, used here only for multi-label bony structure segmentation. Its input is the network's output label response values, and its output is the probability for each label.

[0074] A DICOM sequence is a set of medical image files acquired and output by a CT scanner, arranged in spatial order. Each file typically contains one slice, or some scanners may store multiple slices in multi-frame format. In addition to the CT grayscale image itself, the sequence includes metadata such as pixel spacing, slice thickness, image orientation matrix, and physical coordinates. These metadata are used to subsequently convert the raw foot and ankle CT images into NIfTI 3D volumetric data, thereby establishing a unified spatial coordinate relationship.

[0075] To achieve a complete 3D mask of the calcaneus To prevent scattered noise or localized holes from affecting subsequent coordinate system establishment, this step performs maximum connectivity filtering, hole filling, and small voxel noise removal on the calcaneal labels. Maximum connectivity is used to preserve the spatially continuous main calcaneal region; holes are used to fill in internal voxels that are surrounded by the bone cortex but have not been assigned values ​​in the network; small voxel noise is used to remove isolated predicted points that are not connected to the main calcaneal region.

[0076] During implementation, the data processing server reads the multi-label NIfTI file output by the first deep learning segmentation network into memory, selects the calcaneal label channel to form a binary mask, and then performs a three-dimensional 26-neighbor connected component traversal on the binary mask, retains the connected component with the largest number of voxels, and performs morphological closure processing on its internal holes.

[0077] Complete 3D mask of the calcaneus The formation process should also retain the index relationship with the standardized CT three-dimensional volume data.

[0078] In one specific implementation, a complete 3D mask of the calcaneus All voxel physical coordinates Calculate the centroid and scatter matrix :

[0079]

[0080] Among them, the centroid Represents a complete 3D mask of the calcaneus Mean physical coordinates; number of voxels express Total number of intrinsic voxels; scatter matrix Represents the spatial distribution matrix of a point set; transpose symbol This represents the transpose of a vector.

[0081] For the scatter matrix Perform eigenvalue decomposition, and take the unit eigenvectors corresponding to the eigenvalues ​​from largest to smallest as... The direction sign correction adopts the following deterministic rule: a complete calcaneal 3D mask is used. exist The sets of the first 10 endpoints and the last 10 endpoints of the projection are denoted as follows: and ,like If the corresponding area is closer to the anterior part of the calcaneus in the physical space of a CT scan, then... Point in that direction, otherwise take ; Complete 3D mask of the calcaneus exist Compare the upper 10 sets and the lower 10 sets of the upper projection, and let From the lower area to the upper area; Right-handed relationship Alternatively, it can be corrected by candidate points in the inner support area. In case of conflict between the two, the DICOM foot side information shall be given priority.

[0082] For example, if a voxel is identified as a calcaneal voxel in standardized CT 3D volumetric data, then it is assigned to the complete calcaneal 3D mask. If the value is within the range, then assign a value of 0; otherwise, assign a value of 0. Assuming the first deep learning segmentation network exhibits a small number of missegments at locations close to each other, maximum connected component filtering and the gray-level boundary between two adjacent bone gaps can both reduce this error, thus ensuring a complete 3D mask of the calcaneus. It closely resembles the overall outer contour of the calcaneus.

[0083] Furthermore, a complete 3D mask of the calcaneus. It not only carries the integrity and spatial continuity of the contour, but it can also serve as the spatial constraint boundary for the four-zone segmentation in step two, and as the contour source for the geometric parameter measurement in step three. Because the four-zone labels in step two can only be applied to the complete calcaneal 3D mask... It is generated internally, thus avoiding the inclusion of bony structures or background areas outside the calcaneus within the four-zone division.

[0084] Furthermore, regarding the standardization of CT 3D volumetric data, specific aspects include: voxel resampling, grayscale cropping, bone window width and level adjustment, and RAS direction standardization. For voxel resampling, linear interpolation is preferred for processing CT grayscale volumetric data, or nearest neighbor interpolation can be used for label volumetric data. If the hardware environment or image reconstruction method differs, cubic B-spline interpolation can be used for grayscale volumetric data; however, nearest neighbor interpolation is still required for label volumes to avoid non-integer categories at label boundaries. A voxel spacing of 1.0 mm for isotropic voxels is preferred. When the original slice thickness is large, isotropic voxels with an original slice thickness between 0.8 mm and 1.2 mm can be selected as the primary processing targets to balance detail and memory usage.

[0085] Grayscale processing employs a bone window level and width locking scheme, with an optimal window level of 400 HU and a window width of 1800 HU, corresponding to a grayscale window range of -500 HU to 1300 HU. CT values ​​outside the window are truncated to their boundary values, while CT values ​​within the window are linearly normalized to the range required by the network input. This ensures a stable grayscale contrast between the cortical bone, cancellous bone, and surrounding soft tissues within the network input. The execution is performed by a data processing server, with input being NIfTI 3D volumetric data and output being standardized CT 3D volumetric data.

[0086] Furthermore, window normalization can be implemented using the following formula:

[0087]

[0088] Among them, standardized gray values : Represents the normalized grayscale value of the voxel after bone window cropping, with a value range of 0 to 1; original HU grayscale value : Represents the CT grayscale retained after standardization of direction and voxel spacing, the value of which is determined by the original CT acquisition results;

[0089] Grayscale lower limit : Indicates the lower boundary of the bone window clipping, preferably -500 HU; upper limit of grayscale. : Represents the upper boundary of the bone window clipping, preferably 1300 huan; function sum function Perform lower and upper bound truncation respectively. It is a deterministic numerical operator. The input and output are both gray values ​​of a single voxel.

[0090] In one implementation, to avoid window level and window width being merely described by parameters, the grayscale normalization formula is as follows:

[0091]

[0092] Among them, standardized gray values This represents the normalized voxel gray level input to the first deep learning segmentation network, with a value range of [value range missing]. Original HU grayscale value The HU value represents the CT voxel; the lower limit of grayscale. The preferred value is half the window width minus the window position. HU; Grayscale upper limit Add half the window width to the window position, preferably. HU; function Used for lower bound truncation, function For upper limit truncation. Grayscale volume data preferably uses linear interpolation or cubic B-spline interpolation, while label volume data uses nearest neighbor interpolation to avoid label categories being interpolated to non-integer values.

[0093] The process involves the operator importing the DICOM sequence into the data processing server. The server then reads the scan metadata and generates NIfTI 3D volumetric data. Subsequently, all slices are rearranged according to the RAS direction, each voxel is resampled to a uniform voxel spacing, and bone window normalization is performed on the CT values. Finally, the screen displays only the axial, coronal, and sagittal views of the standardized CT 3D volumetric data, along with the corresponding voxel spacing and orientation.

[0094] Furthermore, standardized CT 3D volumetric data can be uniformly represented in terms of spatial scale, orientation, and grayscale range, eliminating the need for the first deep learning segmentation network to adjust the input scale for different scanning devices. This is due to the affine transformation matrix... and inverse transformation matrix The reason for saving this simultaneously is to obtain a complete 3D mask of the calcaneus later. and the calcaneus region of interest Boundary locations are traced back to the original CT physical space for easier 3D display and parameter encapsulation.

[0095] Furthermore, the calcaneus was located using the context of the entire foot and ankle bony structures, and then a complete 3D mask of the calcaneus was extracted. Compared to directly locating the calcaneus region in a foot and ankle CT scan, this method is more flexible and convenient. The first deep learning segmentation network can observe the positional relationships of the talus, cuboid, distal tibia and fibula, and other bony structures of the foot, and can also identify calcaneus labels with only grayscale thresholds. Its operating environment is a data processing server with a graphics processing unit; if installed on a regular workstation, it can also use a central processing unit for inference, requiring only the implementation of 3D convolution and the output of multiple probability maps.

[0096] After the first deep learning segmentation network outputs a multi-label probability map, this step extracts the calcaneal probability map according to the calcaneal label channel and generates an initial calcaneal binary mask using a fixed probability threshold or the maximum probability assignment method.

[0097] In one implementation, the first deep learning segmentation network employs a 3D encoder-decoder structure, with standardized CT 3D volumetric data as input. The output is a multi-label probability body. , where the category index It should include at least the background, calcaneus, and adjacent bony structures of the foot and ankle. The output layer uses softmax normalization.

[0098]

[0099] Among them, the probability of the first network category Voxel representation Belongs to the The probability of osteoid structures ranges from 1 to 1. First network response value This indicates that the first deep learning segmentation network is in voxel For the Unnormalized response of the class; total number of classes The number of output channels for the first network; exponential function Used to convert response values ​​to positive values. During inference, the calcaneus category is selected based on its probability being the highest and not lower than a threshold. The voxel is assigned the initial mask of the calcaneus, and the threshold is... The optimal value is 0.5, and a complete calcaneal 3D mask is formed by filtering for maximum connected components, filling holes, and removing small voxel noise. During training, manually labeled multi-label 3D masks of the foot and ankle are used as supervision labels, and the loss function is a weighted sum of Dice loss and cross-entropy loss. If other 3D segmentation networks are used, the input, output, labels, loss function, and inference threshold should be kept consistent.

[0100] Preferably, when a voxel has a calcaneal probability greater than 0.5 and the highest probability among all bony structure labels, that voxel is included in the initial calcaneal binary mask; when the probability of a local cortical bone edge is low but it is surrounded by the main calcaneal region, it is filled into the complete calcaneal 3D mask by filling in the holes. In this embodiment, after receiving standardized CT 3D volumetric data, the server automatically calls the model checkpoints and outputs a multi-label NIfTI file. Subsequently, it completes calcaneal label extraction and mask cleanup in the background, so that the operator can only see the overall outline of the calcaneus displayed separately.

[0101] To avoid a complete 3D mask of the calcaneus Adhesive to adjacent bony structures, this step calculates the boundary contact state between the main connected region and adjacent labels after filtering for the largest connected region. If narrow bridge-like connections appear, they can be cut off based on the local gray-level gradient and morphological opening operation. The local gray-level gradient can be obtained based on the gray-level difference between adjacent voxels, while the morphological opening operation element uses a spherical structural element, preferably with a radius of 1 to 2 voxels for calculation.

[0102] Furthermore, the complete 3D mask of the calcaneus Independent of the overall environment of the foot and ankle, this avoids the problem of interosseous adhesion caused by segmentation based solely on thresholds, and also avoids the loss caused by directly cutting out the calcaneal region.

[0103] Then, using the complete 3D mask of the calcaneus Establish a corresponding three-dimensional anatomical coordinate system for all voxel points. .

[0104] Specifically, first, create a complete 3D mask of the calcaneus. A voxel with a median value of 1 is converted into a point set in physical space, and then the centroid of the point set is calculated. Principal component analysis was performed on the decentralized point set to obtain the anterior-posterior axis of the calcaneus. , medial and lateral axes of the calcaneus and the upper and lower axes of the calcaneus At this point, to prevent sign flipping on the left or right foot sides or in the scanning direction, the anteroposterior axis of the calcaneus is first corrected based on the DICOM image orientation information, the length distribution of the anterior and posterior ends of the calcaneus, and the position of adjacent bony structures. and the upper and lower axes of the calcaneus The positive direction of the calcaneus refers to the direction from the posterior region of the calcaneus to the anterior region, and the positive direction of the superior-inferior axis of the calcaneus refers to the direction from the inferior region of the calcaneus to the superior region. The medial-lateral axis of the calcaneus... The positive direction is the same as the medial-lateral direction of the calcaneus.

[0105] Normalized anatomical coordinates are determined according to the following formula:

[0106]

[0107] Among them, voxel points : Represents a complete 3D mask of the calcaneus Inner Physical coordinates of an individual element; center of mass : Represents a complete 3D mask of the calcaneus Average physical coordinates of all voxel points; anteroposterior axis of the calcaneus , medial and lateral axes of the calcaneus and the upper and lower axes of the calcaneus Three-dimensional anatomical coordinate system Three unit direction vectors; lower bounds of front and rear projections Upper limit of front and rear projection These represent all voxel points along the anterior and posterior axes of the calcaneus. The minimum and maximum projection values ​​on the surface;

[0108] Lower limit of internal and external projection Upper limit of internal and external projection These represent all voxel points along the medial and lateral axes of the calcaneus. Minimum and maximum projection values; lower and upper projection limits Upper and lower projection limits These represent all voxel points on the upper and lower axes of the calcaneus, respectively. Minimum and maximum projection values ​​on the plane; normalized coordinates in the back-forward direction. Normalized coordinates of the outer and inner directions Normalized coordinates in the up and down directions All values ​​are limited to the range of 0 to 1 and are used to transmit comparable spatial location information to step two.

[0109] Establishing a three-dimensional anatomical coordinate system Then, step one is based on the complete calcaneus 3D mask. The 3D bounding box is calculated from the voxel boundaries, and safety margins are added in the front-back, inside-out, and top-bottom directions. The preferred safety margin is 20 to 30 voxels; when the resampling voxel spacing is 1.0 mm, this corresponds to a spatial boundary of approximately 20 to 30 mm. The expanded bounding box is then used to trim the calcaneal region of interest from the standardized CT 3D volumetric data. Simultaneously, the complete 3D mask of the calcaneus corresponding to it is cropped out. Local version and affine mapping information. If the edge of the calcaneus is close to the original scan boundary, the portion of the bounding box that extends beyond the boundary is filled with background grayscale without changing the region of interest of the calcaneus. The actual voxel value inside.

[0110] An equivalent implementation includes: principal component analysis can be replaced by orientation estimation based on calcaneal endpoint search, with the endpoint search using a complete 3D calcaneal mask. Establish coordinate axes for the front and rear extreme points, upper and lower extreme points, and inner and outer extreme points; the bounding box can use an axis-aligned bounding box or a three-dimensional anatomical coordinate system. The above alternatives all maintain the same output object, namely the complete calcaneal 3D mask. calcaneal region of interest Three-dimensional anatomical coordinate system and normalized anatomical coordinates , , .

[0111] Furthermore, the region of interest of the calcaneus It was not obtained through a fixed cutting size, but through a complete 3D mask of the calcaneus. and three-dimensional anatomical coordinates It was determined that irrelevant background voxels could be reduced while preserving the entire contour of the calcaneus. Normalized anatomical coordinates were then used. , , After synchronous output, step two can directly use these coordinates to determine the spatial attribution of the front region, rear region, upper joint surface related region and inner support region, forming a continuous data chain from image standardization to four-zone segmentation input.

[0112] Step 2: Receive the complete 3D calcaneal mask output from Step 1. calcaneal region of interest Three-dimensional anatomical coordinate system and normalized anatomical coordinates , , The system divides the calcaneus into four types of computable spatial units: the anterior region, the posterior region, the superior articular surface-related region, and the medial support region. It outputs the volume, centroid, attribution result, and confidence level of the structural block.

[0113] Step one has already separated the calcaneal target from the whole foot space in the foot and ankle CT images and used a complete 3D calcaneal mask. Locked voxel boundaries; but complete 3D mask of the calcaneus It can only indicate whether a voxel belongs to the calcaneus, but it cannot indicate which spatial region the voxel belongs to within the calcaneus. If subsequent geometric parameter extraction still directly deals with the overall mask, the system can only obtain a single overall outline, making it difficult to form volume, centroid, contact relationships, and local distribution sequences at the structural block level.

[0114] Therefore, step two, without changing the spatial reference of step one, involves defining the region of interest of the calcaneus. The data is fed into a second deep learning segmentation network, which enables the network to specifically identify the internal region of the calcaneus within a smaller spatial range, rather than undertaking the task of locating the bony structures of the entire foot again.

[0115] The input to the second deep learning segmentation network is the region of interest of the calcaneus. Its output is not the direct final label, but rather a probability map of the front region. Probability map of the rear region Probability map of the relevant area of ​​the upper joint surface Probability diagram of inner support area The above four types of probability maps and the complete calcaneus 3D mask. Simultaneously entering the spatial constraint fusion process, only those located within the complete calcaneal 3D mask... Only internal voxels are allowed to generate four-zone labels.

[0116] Furthermore, the semantic information obtained by the network recognition is restricted to the defined geometric boundaries of the calcaneus, so that the semantic segmentation of the four regions is simultaneously constrained by image texture, local morphology and overall spatial boundary.

[0117] For example, after completing step one, the data processing server will select the region of interest of the calcaneus. As a single-channel 3D grayscale block input, it serves as the second deep learning segmentation network. The network outputs four [intercepts / parameters]. The server then reads the complete calcaneal 3D mask, which was synchronously cropped in step one, from the 3D response map of consistent size. Partial version, All other responding voxels were set to the background; subsequently in The four probabilities are compared on a voxel-by-voxel basis, and the voxels are initially assigned to the regions with the highest probabilities.

[0118] Specifically, the processing logic of step two mainly consists of three consecutive steps. The first is probability allocation, where the deep learning segmentation network gives the probability that each voxel belongs to one of the four spatial regions; the second part is organizing the candidate connected components of the structural block. Voxels that are spatially adjacent in the same partition are merged as candidate connected components of the structural block, and isolated noise with too small a volume is removed; the third part is spatial attribution, using the normalized anatomical coordinates obtained in step one. , , And the anchor point projection parameters calculated in this step Stable classification of candidate connected components of structural blocks.

[0119] The four-class partition probabilities are formed using the softmax function, which is an explicit normalization operator in this step. Its input is the four-region response values ​​output by the second deep learning segmentation network, and the output is the non-negative probability of the same voxel in each of the four regions. The sum of the four probabilities is 1. Specifically:

[0120]

[0121] Among them, the probability of the fourth zone : Represents structural elements Belongs to the The probability of a class space region, with a range of values, is: ; structural elements : Indicates the region of interest for the calcaneus Inner One undetermined voxel; four-zone response values : indicates that the second deep learning segmentation network is in the structure voxels The above is the first The unnormalized response value output by the class space region takes the value of a real number;

[0122] Category Index Values ​​1, 2, 3, and 4 correspond to the anterior region, posterior region, superior articular surface region, and medial support region, respectively; summation index. Values ​​range from 1 to 4; exponential function It is a deterministic exponential mapping that converts the response values ​​of the four zones into positive values ​​and then performs proportional normalization.

[0123] In one implementation, the input to the second deep learning segmentation network is the calcaneal region of interest. Corresponding standardized grayscale data and optional The local mask channel outputs a four-channel response image. Category Index These correspond to the anterior region, posterior region, upper joint surface related region, and medial support region, respectively.

[0124] The probability map for the four zones is calculated using the following formula:

[0125]

[0126] Among them, the probability of the fourth zone Voxel representation Belongs to the Probability of partitioning; response value of four zones The unnormalized response output by the second deep learning segmentation network; category index. Used to traverse four partitions. Training labels are... The aligned four-zone label map uses a loss function that is a weighted sum of class-weighted Dice loss and cross-entropy loss. Higher class weights can be assigned to the inner support region and the upper joint surface-related region. During inference, only when... When the probability is less than the threshold, four zones are assigned labels; Perform neighborhood completion or mark as a non-critical structural block at the time. threshold The preferred value is 0.45 to 0.60.

[0127] After the probability allocation is completed, the server... Voxels within a voxel but whose probabilities in all four categories are less than a preset threshold are subjected to neighborhood completion. The neighborhood voting method for selecting this voxel is a three-dimensional 26-neighbor voting, which reads the determined labels of its 26 surrounding voxels and selects the labels that appear most frequently and have a high average probability as completion labels; if there are few labels in the neighborhood or the average probability is still lower than the threshold, it is no longer recognized as a non-critical structural block. Candidate voxels.

[0128] First, the probability map of the front region is read. Probability map of the rear region Probability map of the relevant area of ​​the upper joint surface Probability diagram of inner support area The four types of probability maps were then compared with the complete 3D mask of the calcaneus. exist Alignment is performed voxel-by-voxel in the coordinate system. If a certain structural voxel... Complete 3D mask of the calcaneus The value of the middle is The voxel is directly set as the background; if the value is... Then, its four probabilities are retained and used in label determination. Since step one has already saved... The mapping relationship between the coordinate system and the original CT physical coordinate system. After the label determination is completed, this step can write the four-zone label map back to the original CT physical space.

[0129] In one embodiment, a second deep learning segmentation network is deployed on a data processing server equipped with a graphics processing unit. The network structure adopts a 3D encoder-decoder form, with the input block being the region of interest of the calcaneus after cropping in step one, and the number of output channels fixed at four. Forward inference is first run to obtain the four-zone response map, and then softmax normalization is performed to obtain the four-zone probability map. Finally, the complete 3D calcaneus mask is loaded. Partial version, full 3D mask of the calcaneus The external voxels are written with background labels, and the internal voxels of the complete calcaneal 3D mask are written with the initial four-zone labels with the highest probability. This can also be implemented in a central processing unit inference environment with equivalent functionality, without changing the processing flow and output data structure.

[0130] Furthermore, the second deep learning segmentation network only generates one region label map within the four boundaries, and cannot contain adjacent bony structures, soft tissues, or background voxels. That is, the four-region probability map and the complete calcaneus 3D mask are not identical. After constraint fusion, the network output is transformed from pixel-level probability into a spatial segmentation result constrained by geometric boundaries.

[0131] In one implementation, for low-confidence voxels Take its three-dimensional 26-neighbor set Calculate each label neighborhood support value :

[0132]

[0133] Among them, neighborhood support value Indicates low confidence voxels Completed to the first Class tag support strength; neighborhood set express 26 neighboring voxels; indicator function In neighboring voxels The value has been assigned to the first The value is 1 if the class label is selected, otherwise it is 0; the probability of the four zones is... Representing neighborhood voxels Belongs to the The probability of class partitioning. If the maximum If the completion threshold is exceeded, assign the corresponding label; otherwise, mark it as... .

[0134] The initial four-zone label map is processed for four types of labels. The server performs a 3D 26-neighborhood connected component analysis on each of the front, rear, upper joint surface-related, and inner support regions, grouping similar voxels with face, edge, or corner contacts into candidate connected components for the same structural block. The input to the connected component analysis is the four-zone label map, and the output is a connected component number map of the structural block, the number of voxels, volume, centroid, and boundary voxel set for each structural block. This process transforms scattered voxel labels into 3D structural units with volume and position.

[0135] Furthermore, to filter out isolated small regions caused by network boundary fluctuations, this step employs a minimum retention volume. As a noise filtering boundary:

[0136]

[0137] Among them, minimum retention volume : Represents the lowest volume boundary where the candidate connected components of the structural block are preserved, and takes a non-negative volume value; complete calcaneal volume : Represents a complete 3D mask of the calcaneus The physical volume obtained by converting all voxels is calculated by multiplying the number of voxels by the physical volume of a single voxel; function : Represents a deterministic operator that takes the larger value, used to select filtering conditions between fixed volume boundaries and overall volume ratio boundaries; volume unit : represents cubic millimeters. This formula combines the lower limit of absolute volume with the proportion of the complete calcaneus volume, giving isolated noise a uniform filtering boundary.

[0138] For example, when processing the region related to the upper joint surface, the server will also use 5 of the total volume of that region as an additional filtering boundary; if a small connected component is lower than the minimum retention volume... If the boundary of the region's volume ratio is specified, it is added to the noise list without undergoing the structural block assignment process. If a small connected region has a continuous boundary contact with a major structural block, it will be merged into the nearest similar structural block according to the length of the contact boundary between them, in order to balance noise suppression and integrity.

[0139] Furthermore, the four-zone label map is organized into a numberable, measurable, and rewritable structural block connected domain, so that isolated noise will not directly affect the subsequent centroid and projection parameters; the volume threshold adopts both absolute boundary and proportional boundary, so that small volume pseudo-connected domains have clear processing rules.

[0140] Furthermore, the candidate connected components of the structural blocks are mapped to the three-dimensional anatomical coordinate system established in step one. In the middle, the server first calculates the centroid of the candidate connected components of each structural block. Then read the normalized coordinates of the back-forward direction corresponding to the centroid. Normalized coordinates of the outer and inner directions Normalized coordinates in the up and down directions To avoid unstable boundary region attribution due to relying solely on a single directional coordinate, this step defines the outer lower anchor point. and the anterior inner upper anchor point Rear outer anchor point To meet , , Complete 3D mask of the calcaneus voxel set centroid; anterior medial superior anchor point To meet , , Complete 3D mask of the calcaneus Center of mass of a voxel set.

[0141] For the Candidate connected components of a structural block, anchor point projection parameters Determine using the following formula:

[0142]

[0143] Among them, anchor point projection parameters : indicates the first The candidate connected components of the structural block are along the outer lower anchor point. To the anterior inner upper anchor point The normalized projected position of the direction, with a value range of . ;Center of mass of structural block : indicates the first The average point of all voxel physical coordinates within the candidate connected domain of a structural block; the outer bottom anchor point. : Represents the centroid of the posterolateral region obtained by filtering normalized anatomical coordinates;

[0144] Front inner upper anchor point : Represents the centroid of the anterior superior medial region obtained by filtering normalized anatomical coordinates; vector inner product symbol : Represents the dot product of three-dimensional vectors; vector magnitude : Represents the Euclidean length of the anchor vector; clipping function : Indicates when the input value Output 0 when the input value is less than 0, and output 0 when the input value is less than 0. Output 1 if the value is greater than 1, otherwise output [missing value]. The function form is already defined.

[0145] Spatial attribution determination employs a joint rule of probability and location. If the candidate connected components of a structural block satisfy... , , And the probability of the inner support area If the probability is the highest or second highest, it belongs to the inner support area; if it satisfies , And the probability of the related area of ​​the upper joint surface If the probability is the highest, it is assigned to the region related to the upper joint surface; if the condition is met... or anchor point projection parameters If it satisfies the following conditions, it belongs to the front region; or anchor point projection parameters If a candidate connected region of a structural block does not meet the priority criteria of the inner support region and the upper joint surface related region, it is assigned to the posterior region. If a candidate connected region of the same structural block meets multiple conditions simultaneously, it is assigned in the order of inner support region, upper joint surface related region, anterior region, and posterior region; if it still cannot be assigned, it is marked as a non-critical structural block. It retains its volume, centroid, normalized coordinates, nearest critical partition distance, and original spatial coordinates.

[0146] In terms of on-site actions, after the server completes the connected component numbering, it generates a centroid point on each structural block and projects this centroid point onto the 3D anatomical coordinate system from step one. The interface can display four region colors, structural block numbers, and centroid positions; when a structural block is marked as a non-critical structural block... At that time, the system only records its spatial parameters in the structure list and does not use the structure block for the calculation of the key structural parameters in step three.

[0147] In an equivalent implementation, the anchor point and anchor point It can be determined by a fixed normalized window centroid or by a complete calcaneal 3D mask. The endpoint search results are determined; as long as the same three-dimensional anatomical coordinate system is still used. The output results of the same four-zone terminology system are parallel implementations that belong to the same processing principle.

[0148] Furthermore, the structural block attribution is achieved by simultaneously utilizing the structural block centroid, normalized anatomical coordinates, and anchor point projection parameters. The priority rule ensures that the inner support region and the related region of the upper articular surface obtain stable assignment when their boundaries overlap.

[0149] After step two, the system outputs four labeled regions: the anterior region, the posterior region, the relevant region of the superior articular surface, and the medial support region. It also outputs the structural block's connected domain number, structural block volume, structural block centroid, normalized anatomical coordinates, and anchor point projection parameters. Spatial partitioning results, volume percentage of each partition, and segmentation confidence.

[0150] Furthermore, a complete 3D mask of the calcaneus. Boundary constraints are applied to the four-zone probability map to prevent the four-zone labels from detaching from the calcaneal space; connected component sorting and volume threshold filtering transform the voxel-level segmentation results into structure block-level objects, which facilitates subsequent measurement and encapsulation.

[0151] Step 3: Receive the complete 3D calcaneal mask output from Step 1. Three-dimensional anatomical coordinate system and the rear region output in step two Related areas of the superior articular surface Without outputting disease type names, clinical classification names, health status judgments, or treatment suggestions, the parameterized extraction of calcaneal contour geometry, connectivity structure of related areas of the superior articular surface, surface height difference, and spatial contact relationship is completed.

[0152] The following steps are all performed by the data processing server. Step 3 uses a complete 3D mask of the calcaneus. The contour measurement object is used instead of the four-zone labels from step two as the angle calculation object; at the same time, the relevant area of ​​the upper joint surface is used. As the object of connected structure analysis, the rear region Area related to the superior articular surface Spatial adjacency relationships are used as the object of contact parameter calculation. This setup is because the four-zone labels are used to represent local partitions, which may exhibit discontinuities, gaps, or low-confidence areas at the boundaries of structural blocks; complete calcaneal 3D mask. Preserving the overall outer contour of the calcaneus is suitable as a source of geometric angle parameters. Step three only outputs the first geometric angle parameters. Second geometric angle parameters Surface height difference parameters Number of effective connected components 3D contact ratio Connected component centroid sorting sequence and quality control markers .

[0153] Step two has already identified the region of interest for the calcaneus. The voxels within are divided into anterior regions. Rear area Related areas of the superior articular surface and inner support area It outputs the centroid, volume, and normalized anatomical coordinates of each structural block. If the relevant area of ​​the upper articular surface is directly used... When extracting angles using other single-region labels, the angle values ​​will be affected by local label boundaries; if only the complete calcaneal 3D mask is used... However, it is impossible to obtain the number of locally connected components and the contact ratio.

[0154] In terms of on-site operation, after the operator imports the foot and ankle CT image data, the server caches the outputs of steps one and two; upon entering step three, the server reads the complete calcaneal 3D mask. Three-dimensional anatomical coordinate system Rear area Area related to the superior articular surface The interface only displays the outline points, connected component numbers, geometric segments, and parameter list.

[0155] Furthermore, the angle calculation is not affected by the discontinuity of the four-zone labels, and the connectivity structure parameters are not affected by the selection of two-dimensional slices. The structured output can be encapsulated in step four.

[0156] Step 3: Along the inner and outer axes Complete 3D mask of the calcaneus Projected onto the front and rear axles and upper and lower shafts Extract the upper edge contour line from the sagittal projection plane, and find the first marker point within the three normalized windows. Second marker and the third marker The first geometric angle parameters are obtained. Then the server at the second marker point. With the third marker Find the indentation point between The second geometric angle parameter can be obtained by piecewise fitting. .

[0157] After that, the server switched to the area related to the upper joint surfaces. The number of effective connected components was obtained through gap correction, three-dimensional 26-neighborhood connected component analysis, and reference plane fitting. With surface height difference parameter Finally, for the rear area and the area related to the superior articular surface Perform expansion with the same radius and calculate the three-dimensional contact ratio. .

[0158] The Gaussian smoothing, contour top edge extraction, 3D 26-neighborhood connected component analysis, line fitting, reference plane fitting, and morphological dilation mentioned above are all directly implementable image processing operators. Gaussian smoothing uses a one-dimensional discrete Gaussian kernel, with the kernel radius preferably... Three-dimensional 26-neighborhood connected domain analysis classifies similar voxels of surface contact, edge contact, and corner contact into the same connected domain; morphological expansion uses spherical structural elements with a radius of 1 voxel.

[0159] The server will use the entire 3D mask of the calcaneus. Each voxel point within is projected onto the front and rear axes. With upper and lower axes Construct a plane. If multiple voxels fall in the same two-dimensional grid, the grid is recorded as the foreground. Then, extract the highest foreground point on each grid in the front and back directions to obtain its upper edge outline, and use Gaussian smoothing to weaken the voxel step boundaries.

[0160] During marker point localization, the server normalizes the coordinates along the back-forward direction. Divide the search into three windows: the rear window is... The upper middle window is The front window is The server selects the highest point of the outline as the first marker point in each of the three windows. Second marker and the third marker If candidate points have the same height, the candidate point with the smaller distance from the window center is selected. The interface displays three search bands and an upper outline; parameter values ​​are written to the output file by the server.

[0161] Furthermore, the first geometric angle parameter Calculate using the following formula:

[0162]

[0163] Where: First geometric angle parameter : Represents the included angle of the upper edge contour formed by the three marker points, with a value ranging from 0 to 1. The output can be converted to degrees; arctangent function This is an arctangent function with quadrant determination. The inputs are the vertical and horizontal differences, and the output is a directed angle; the height coordinates of the first marker point. The second marker's height coordinates The third marker's height coordinates These are the normalized coordinates of the three marker points in the upper and lower directions. The values ​​on the digit range from 0 to 1.

[0164] Coordinates before and after the first marker point Coordinates before and after the second marker point Coordinates before and after the third marker point These are the normalized coordinates of the three marker points in the backward-forward direction. The values ​​on the integer part range from 0 to 1.

[0165] If the first geometric angle parameter Greater than The server retrieves the supplementary angle. Furthermore, the contour projection is derived from the complete calcaneus 3D mask. The marker positioning is defined by a normalized window, which together reduce the influence of scan direction and voxel scale differences on angular parameters and provide a geometric basis for the visualization of the parameters in step four.

[0166] The server searches for indentations within the contour segment. Indentation Determined by both local low points and curvature response: First, filter the upper and lower coordinates. Candidate points with a height smaller than the average height of adjacent windows are selected, and then points with a larger absolute value of curvature are chosen. Curvature is calculated using first-order and second-order differences, and the boundary is extended using neighboring points.

[0167] Determine the depression point After that, the server will To the depression point The set of contour points is fitted to the first contour line segment, and the concave points are... To the third marker The set of contour points is fitted to form the second contour line segment. Least squares straight line fitting is preferred; when isolated outliers exist within the contour segment, random sampling consistency straight line fitting is used, and the random seed, interior point distance threshold, and iteration count are recorded. Only the second geometric angle parameters are recorded in the output. and quality control markings .

[0168] Among them, the second geometric angle parameter Calculate using the following formula:

[0169]

[0170] Where: the second geometric angle parameter : Represents the interior angle between two locally fitted contour lines, with a value ranging from 0 to 1. First fitted slope :for to The slope of the straight line of the contour segment, this term takes a real number value; the second fitted slope :for to The slope of the straight line of the contour segment, this term takes a real number value; inverse cosine function. Input value range: 0 to 1. Absolute value sign. Eliminate sign differences in different directions. If the fitting residual is greater than a threshold, the server retains the second geometric angle parameter. And set quality control marks. .

[0171] Furthermore, the depression point The boundary misjudgment is avoided by simply taking the lowest point, as the local height and curvature are jointly determined. Piecewise fitting transforms the jagged profile into calculable line segments, making the second geometric angle parameter... It has a clear set of input points, fitting rules, and anomaly handling boundaries.

[0172] The server is The original CT grayscale values ​​and local grayscale gradients of the interior and its boundary neighborhood are read to screen candidate voxels for low-density gaps; bony thresholding. A fixed value within the range of 150 HU to 300 HU is preferred. If the low-density gap candidate voxel is located between two high-density structural blocks, and both sides are adjacent to... If a foreground voxel is adjacent to another voxel, the server temporarily removes that candidate voxel and then applies the corrected voxel. Perform a 3D 26-neighbor connectivity analysis to obtain the number of effective connected components. The connected component volume, connected component centroid, and outward / inward orientation sorting sequence are calculated. The removed voxels do not alter the original four-region label map saved in step two.

[0173] Subsequently, the server will... For each valid connected component, the upper surface point set is extracted. The upper surface point set refers to the points on the upper and lower axes. There are no similar foreground voxels occluding the boundary in the positive direction, and the reference plane is obtained by fitting the point set of the main surface. It is best to choose to use the least squares plane for fitting. If there is CT image data on the opposite side, first mirror-register the corresponding area on the opposite side to the current three-dimensional anatomical coordinate system, and then use the registered corresponding surface as the reference plane.

[0174] Surface height difference The specific value is calculated according to the following formula.

[0175]

[0176] Where: surface height difference parameter : Indicates the area related to the upper articular surface The maximum height difference between the upper surface point and the reference plane is a non-negative length value; upper surface point : Indicates the area related to the upper articular surface The upper surface point set of The Middle A three-dimensional physical coordinate point; upper surface point set : Represents all points on the upper surface obtained by the boundary extraction rules;

[0177] Reference plane normal vector : Represents the unit normal vector of the fitted reference plane, taken as a three-dimensional unit vector; reference plane reference point : Used to represent the preferred value of a point on the reference plane. The centroid of the point set on the main surface is replaced by the transpose symbol; transpose symbol This refers to the transpose of a vector, and the target value for calculation is the directed distance from a point to a plane. If the reference plane fitting residual exceeds a pre-defined boundary value, the server will implement appropriate control flags. And write the trigger reason in step four.

[0178] Finally, the server separately processed the rear area. and the area related to the superior articular surface A three-dimensional expansion with a radius of 1 voxel is performed to determine the number of voxels overlapping after expansion, and the three-dimensional contact ratio is obtained from this. As shown below:

[0179]

[0180] Where: Three-dimensional contact ratio : Indicates the rear area Area related to the superior articular surface The degree of spatial contact within the dilated neighborhood ranges from 0 to 1; the number of overlapping voxels. : Indicates the rear area Area related to the superior articular surface The number of voxels shared after three-dimensional expansion of the same radius, taking a non-negative integer value; the number of surface voxels in the rear region. : Indicates the rear area The number of boundary surface voxels, taking positive integer values;

[0181] Number of surface voxels in the relevant area of ​​the superior articular surface : Indicates the area related to the upper articular surface The number of boundary surface voxels, taking positive integer values; minimum value function : indicates that the smaller of the two surface voxel counts is used to normalize the contact scale.

[0182] In one implementation, Within and at the boundary of a voxel neighborhood, for voxels Calculate local gradient strength :

[0183]

[0184] Among them, local gradient intensity Voxel representation Degree of abrupt change in surrounding grayscale; CT grayscale Voxel representation HU value; directional offset These represent the offset of a voxel along the three axes of the three-dimensional voxel space.

[0185] like and And in Both exist in the neighborhood of two voxels on opposite sides. of Foreground voxels will be Marked as interstitial candidate voxels. Bone threshold. Preferably 150HU to 300HU; high density difference Preferably 100 HU; gradient threshold Can retrieve the current The upper quartile of the grayscale gradient of the bone window. Labeled interstitial candidate voxels are only temporarily removed in the connected component analysis copy, without altering the original four-zone label map.

[0186] Furthermore, gap correction combines grayscale gaps, local gradients, and 3D adjacency relationships to reduce false connectivity caused by smoothing labels; surface height difference parameters Determined by the upper surface point set and the reference plane, facilitating parameter visualization in step four; 3D contact ratio. The spatial contact between two regions is transformed into a storable numerical value without outputting any clinical judgment.

[0187] In one implementation, the upper surface point set Defined as: for any element point If along the upper and lower axes If no similar foreground voxels are found in one or two consecutive examinations in the forward direction, then... Belongs to the upper surface point set The reference plane is obtained through least-squares plane fitting:

[0188]

[0189] Wherein, the reference plane normal vector Unit vector; reference plane reference point Preferred selection The center of mass; points on the upper surface express Any three-dimensional physical coordinate point in the data. Surface height difference parameter. The calculation is as follows:

[0190]

[0191] Among them, the surface height difference parameter This represents the maximum distance from a point on the upper surface of the relevant region of the articular surface to the reference plane, and is a non-negative length value. If contralateral CT image data exists, the corresponding reference plane should first be obtained through rigid mirror registration; otherwise, it should be used... The upper surface of the largest connected region is fitted to a plane.

[0192] After step three, the server outputs the first geometric angle parameter. Second geometric angle parameters Surface height difference parameters Number of effective connected components Connected domain volume, connected domain centroid, outside-inside orientation sorting sequence, three-dimensional contact ratio and quality control markings Among them, the first geometric angle parameter With the second geometric angle parameter Derived from complete 3D calcaneal mask Overall contour, surface height difference parameters Number of effective connected components Originating from the area related to the superior articular surface , Originating from the rear area Area related to the superior articular surface The neighborhood overlap relationship.

[0193] Furthermore, the overall contour parameters, local connectivity parameters, and region contact parameters are extracted from different data objects, but all are located in the three-dimensional anatomical coordinate system. The output consists entirely of image processing parameters and spatial geometry parameters, which can be directly encapsulated and visualized in step four.

[0194] Step 4: Combine the four-zone label map and structural block spatial assignment results generated in Step 2 with the geometric parameters and connectivity parameters generated in Step 3 into a structured parameter package that can be stored, displayed, and verified.

[0195] The following steps are all performed by the data processing server. Step four inherits the output from the front area of ​​step two. Rear area Related areas of the superior articular surface Inner support area Non-critical structural blocks Connectivity number of each structural block, volume of each structural block, centroid of each structural block, and normalized anatomical coordinates. , , Anchor point projection parameters And the segmentation confidence, and take the first geometric angle parameter output from step three. Second geometric angle parameters Surface height difference parameters Number of effective connected components Connected component centroid sorting sequence, three-dimensional contact ratio The next step involves calculating confidence levels and quality control markers for parameters. This step does not change the values ​​of the aforementioned parameters or re-infer the meaning of the images. Instead, it uses field locking, rule validation, and visualization write-back to create a unified data output for the image processing results.

[0196] Steps one through three generate spatial reference, region labels, and structural parameters, respectively. If these results are output separately, subsequent 3D display, parameter descriptions, and data storage will face multiple coordinate sources, filenames, and confidence levels, potentially leading to a disconnect between structural block numbers and parameter fields. Therefore, step four places all results under the same image data number and uses a complete calcaneal 3D mask. calcaneal region of interest Three-dimensional anatomical coordinate system Using the four-zone label map as the top-level index, each angle, volume, connected component, and contact parameter can be traced back to a specific voxel region.

[0197] For example, after the operator completes the processing of a set of foot and ankle CT image data in the workstation interface, the data processing server generates a four-region segmentation NIfTI label map, a three-dimensional mesh model of the structural blocks, JSON files of geometric parameters and connectivity parameters, and parameter visualization images in the same folder. The interface only displays the region name, parameter name, field source, and quality control markers, and does not display the disease name, clinical classification conclusion, or treatment plan.

[0198] Therefore, the working principle of step four is to transform the image segmentation link into a parameter encapsulation link, binding the computable object, the calculation result, and the display object within the same coordinate system.

[0199] The data processing server first reads the output files from steps two and three, verifying that each file has the same image data number, the same voxel spacing, and the same affine transformation matrix; then, it generates structured JSON data according to the field template. The field template includes at least the image data number, CT voxel spacing, complete calcaneal mask quality index, four-zone segmentation result, volume and volume percentage of each zone, number of connected components in the structural block, centroid of the structural block, and first geometric angle parameters. Second geometric angle parameters Surface height difference parameters Number of effective connected components Connected component centroid sorting sequence, three-dimensional contact ratio The reasons for segmentation confidence, parameter calculation confidence, quality control flags, and triggering quality control flags.

[0200] After the fields are generated, the rule validation module performs a consistency check on the structured JSON data. This check is not aimed at medical judgment, but rather at determining whether the input files correspond to each other, whether the parameters fall within the algorithm's allowed range, and whether the visualized image can be written back to the original CT physical space. If the validation passes, the server continues to generate parameter visualization images and standardized interface files; if the validation fails, the server retains the original parameters, but writes the trigger reason in the quality control flag and states in the parameter description text that the field requires manual review of the image processing process.

[0201] Furthermore, the data processing server uses the image data number as the root field and the complete calcaneal 3D mask. The four-region NIfTI label map is the main file field, and the structural block number is the array field, which includes the region, volume, centroid, and normalized anatomical coordinates of each structural block. , , Anchor point projection parameters The nearest critical partition distance and split confidence are written to the same structure block record. For non-critical structure blocks... The server still writes its spatial parameters, but specifies in the field properties that they are only used for integrity recording, 3D reconstruction and subsequent image processing.

[0202] Preferably, the structured JSON data simultaneously records the field source path, generation timestamp, and execution entity name; the field source path points to the corresponding NIfTI tag file, 3D mesh file, or parameter calculation record, and the execution entity name is fixed as the data processing server. This setting ensures that the same parameter references the same source field when displayed on the interface, reconstructed in 3D, and stored in the file, avoiding mismatches caused by manually copying parameters.

[0203] Among them, the field completeness index Determine using the following formula:

[0204]

[0205] Where: Field completeness index : Indicates the population status of necessary fields in structured JSON data, with a value range of 100%. Used to determine whether the parameter package has complete output conditions; field number : indicates the first field in the field template There are 1 required field, with values ​​taking positive integers; the total number of fields... : Indicates the number of required fields in the field template, and takes a positive integer value; field weight : indicates the first The weights of the necessary fields in the parameter package are positive numbers, with higher weights preferred for image data numbering, affine transformation matrix, four-zone label map, and quality control markers; field validity indicator values. : indicates the first If a necessary field exists and is correctly formatted, the value is 1 if the field exists and passes the format check; otherwise, the value is 0.

[0206] In this embodiment, the server first writes the image data number and CT voxel spacing, then writes the complete calcaneal mask quality index and four-zone segmentation results; and writes the anterior region item by item according to the structural block number. Rear area Related areas of the superior articular surface Inner support area and non-critical structural blocks Spatial parameters.

[0207] In equivalent implementations, JSON, XML, or binary medical image extension fields can be used to carry structured data. However, the field names, symbol meanings, and output objects must remain consistent. This ensures that the structural block numbering, partitioning, and geometric parameters are fixed within the same parameter package, avoiding name variations for the same structural block in different output files. This is achieved through field completeness metrics. Determine whether the parameter package can be output so that the display end can directly read the fields.

[0208] The rule validation module first reads the quality index of the complete calcaneal mask and judges all complete calcaneal 3D masks. Check if the proportion of the main connected components is less than a preset threshold; then read the four-region segmentation results and determine if any key segmented regions are below the preset threshold; then read the relevant regions of the upper joint surfaces again. Number of valid connected components Compare the gap detection records from step three with those from step three to determine if there are any contradictions; finally, read the surface height difference parameters. 3D contact ratio The upper surface fitting residual is used to determine whether the sources of surface parameter calculations are consistent.

[0209] Among them, cross-gating value Determine using the following formula:

[0210]

[0211] Where: cross-gating value : Indicates the degree to which structured JSON data passes rule validation; the values ​​are limited to a specific range after trimming. Used to generate quality control markers; mask quality confidence. : Represents a complete 3D mask of the calcaneus The confidence value formed by the main connected component, hole filling, and boundary cover states has a range of values. Partition consistency confidence : Represents the front area of ​​the four-zone label map Rear area Related areas of the superior articular surface and inner support area with normalized anatomical coordinates , , The degree of consistency, with a range of values ​​being: ;

[0212] Parameter source confidence : Represents the first geometric angle parameter Second geometric angle parameters Surface height difference parameters Number of effective connected components and three-dimensional contact ratio Whether a complete source record exists, the range of values ​​is: ; Write-back consistent confidence This indicates whether the coordinates of the four-zone label image and the parametric visualization image are consistent after being written back to the original CT physical space. Its value range is... ; Abnormal trigger quantity The rule validation module finds the number of abnormal entries, which is then normalized. Its value range is... Weighting coefficient , , , and All values ​​are non-negative, and their values ​​are determined by the configuration file to adjust the cross-gating values ​​of each confidence value and anomaly trigger value. The impact of [the event / event].

[0213] In one implementation, for the first Candidate connected components of structural blocks Partition confidence Determine using the following formula:

[0214]

[0215] Among them, partition confidence Represents a structural block Assigned to the target partition The degree of credibility, with a value range of . ; Structural block element set Indicates the first Connected domains; Number of structural block voxels This represents the total number of voxels within the connected component; the probability of the target partition. Voxel representation Belongs to the assigned partition The probability of; the volume of the structural block express Physical volume; minimum retention volume This represents the noise filtering volume threshold. If The structure block is then marked as Or it may trigger a quality control flag.

[0216] In one implementation, the parameter source confidence level It can be determined by the contour effectiveness, plane fitting residuals, and connectivity consistency:

[0217]

[0218] Among them, the confidence level of the parameter source This indicates the overall reliability of the parameter source in step three; the quality of the angle source. This indicates whether the number of contour points within the search window containing the three marker points meets the minimum requirement; surface parameter quality. This represents the quality of the reference plane fitting residual after normalization; contact parameter quality. Indicates the degree of consistency between the inflated contact boundary and the actual boundary voxel; weighting coefficient. All are non-negative numbers and their sum is 1. When At that time, a quality control mark is generated without changing the parameter value.

[0219] In the embodiment, if the complete calcaneal 3D mask Insufficient proportion of the main connected component indicates insufficient quality of complete calcaneal segmentation in the quality control markers; if the relevant area of ​​the superior articular surface... Number of valid connected components Inconsistent with the gap detection record, the server-written connectivity structure parameters are pending verification; if the surface height difference parameter... The reference plane fitting residual is inconsistent with its numerical source; the surface parameter calculation written by the server needs to be verified; if the three-dimensional contact ratio The number of voxels at the actual contact boundary is inconsistent; the spatial contact parameters written by the server need to be verified. The above description pertains to image processing quality and does not represent clinical judgment.

[0220] Furthermore, the rule verification module incorporates mask quality, four-zone segmentation, parameter source, and coordinate write-back into the same gating chain, enabling the source of errors to be located to specific fields; cross-gating values Without changing the parameter values, only control the generation boundaries of quality control markers and parameter description text.

[0221] Furthermore, the data processing server first generates a 3D mesh model of each structural block based on the four-zone NIfTI label map. Preferably, the MarchingCubes isosurface extraction algorithm is used to generate mesh vertices and triangular faces. Then, the affine transformation matrix saved in step one is used to map the mesh vertices back to the original CT physical space. A surface extraction algorithm with the same function can also be used, as long as the output is the 3D mesh model of the structural block, the coordinates of the mesh vertex, and the field of the region to which it belongs.

[0222] Subsequently, the server will send the first geometric angle parameters. Second geometric angle parameters Surface height difference parameters Number of effective connected components 3D contact ratio Overlay the image onto the parametric visualization image and mark the first marker point, the second marker point, the third marker point, the depression point, the reference plane, and the contact boundary.

[0223] Among them, visualized write-back error Determine using the following formula:

[0224]

[0225] Where: Visualized write-back error : Represents the maximum coordinate deviation between the parametric visualization image and the 3D mesh model after forward and reverse coordinate mapping, taking a non-negative length value; visualization point set : Represents the set of marker points, structural block centroids, mesh vertex sampling points, and contact boundary points that participate in the visual write-back check;

[0226] Inverse transformation matrix : Represents the affine transformation matrix mapping from the NIfTI voxel coordinate system back to the DICOM physical coordinate system; affine transformation matrix : Represents the affine transformation matrix mapping from the DICOM physical coordinate system to the NIfTI voxel coordinate system; visualizes checkpoints. : Represents a set of visualized points Any three-dimensional coordinate point in the array; L2 norm : Represents the Euclidean length of the three-dimensional coordinate difference, used to characterize the magnitude of the write-back deviation;

[0227] If the structured parameter description text generation module is retained, its input is limited to field names, field values, field sources, and quality control tags in structured JSON data, and its output is limited to parameter descriptions, data source descriptions, and quality control descriptions. This text generation module does not participate in segmentation, measurement, connected component analysis, or rule validation, and does not modify the first geometric angle parameters. Second geometric angle parameters Surface height difference parameters Number of effective connected components or three-dimensional contact ratio .

[0228] Furthermore, the 3D mesh model, parameter visualization images, and structured parameter description text are all derived from the same structured JSON data, and the output objects share the same field source; visualization of write-back errors. This allows for the calculation of boundaries in the overlay and physical coordinate write-back, facilitating subsequent image analysis, 3D display, and data storage retrieval.

[0229] In one implementation, the input to the structured parameter description text generation module only includes field names, field values, field source paths, quality control flags, and trigger reasons from the structured JSON; the output only includes parameter source descriptions, segmentation region descriptions, geometric parameter descriptions, connectivity structure parameter descriptions, and quality control flag descriptions. After generation, string consistency checks compare the values ​​in the text with the JSON fields. If they do not match, the text is rejected and a fixed template description is provided.

[0230] In this embodiment, all image processing steps are performed by the data processing server. The data processing server reads the image position, image orientation, pixel spacing, and layer spacing from the DICOM sequence, and transforms the DICOM physical coordinates using an affine transformation matrix. Mapped to the NIfTI voxel coordinate system, and through the inverse mapping matrix Write back to the original CT physical space. CT grayscale data are resampled using linear interpolation or cubic B-spline interpolation, and label data are resampled using nearest neighbor interpolation. CT grayscale values ​​are normalized after bone window cropping. scope.

[0231] The first deep learning segmentation network employs a 3D encoder-decoder structure. The input is standardized CT 3D volumetric data, and the output is a multi-label probability volume of foot and ankle bony structures. The output layer uses a softmax function, and the voxels with the highest probability for the calcaneus category, not lower than a threshold, form the initial calcaneus mask. The initial calcaneus mask is then filtered through a 3D 26-neighbor maximum connected component, filled with holes, and small voxel noise removed to obtain the complete 3D calcaneus mask. Subsequently, regarding Principal component analysis was performed on the physical coordinates of all voxels to obtain... The normalized anatomical coordinates are generated by correcting the direction sign based on the endpoint set, DICOM foot side information, and right-handed system relationship. and the calcaneus region of interest .

[0232] The second deep learning segmentation network uses As input, the output is a four-zone response map, which is then processed by the softmax function to obtain... The system only allows Internal voxels are assigned four-region labels; low-confidence voxels are completed using three-dimensional 26-neighborhood support values, and voxels that fail to complete are moved to non-critical structural blocks. Candidate set. Three-dimensional 26-neighborhood connected component analysis was performed on each of the four region labels, with a volume less than [missing information]. Isolated connected components are filtered out. (Later, the outer lower anchor point...) and the anterior inner upper anchor point Depend on The centroid of the fixed window is determined, and the centroid of the structural block is determined. Projection parameters are obtained by anchor point projection. , and then combine Spatial assignment is determined by partition probability.

[0233] Geometric parameter measurements with complete calcaneal 3D mask For the object, along The projected image is obtained as a sagittal projection, and then subjected to upper edge contour extraction, Gaussian smoothing, and search window positioning. ,Depend on Calculate the first geometric angle parameters .exist to Within the contour segment, the system determines the depression point by combining local height and discrete curvature. Fitting respectively and Calculate the second geometric angle parameter for two straight lines. For the relevant area of ​​the superior articular surface The system identifies gap candidate voxels by using the HU threshold, local gradient, and bilateral high-density neighborhood. After removing them from the temporary replica, it performs a three-dimensional 26-neighborhood connected component analysis to obtain the number of effective connected components. Then, extract the upper surface point set, fit the reference plane, and calculate the surface height difference parameter. Rear area Area related to the superior articular surface After three-dimensional dilation of the same radius, the number of overlapping voxels is calculated. and three-dimensional contact ratio .

[0234] In the structured output stage, the data processing server will output the four-zone label diagram, structure block number, structure block volume, structure block centroid, and so on. , , , , , , Parameter confidence and quality control tags are written to structured JSON data. The rule validation module calculates the field completeness index. Cross-gating value and visualize write-back error When fields are missing, partition locations conflict with coordinate rules, parameter sources are incomplete, or coordinate write-back errors exceed a threshold, only quality control flags are generated, without changing the original parameters. If the structured parameter description text generation module is retained, it only takes JSON fields as input and outputs parameter descriptions, data source descriptions, and quality control descriptions, without generating diagnostic, therapeutic, or treatment opinions.

[0235] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0236] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0237] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0238] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0239] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-based automatic four-zone segmentation and measurement system for calcaneal fractures, characterized in that: include, Receive DICOM sequences and convert them into NIfTI 3D volume data after resampling, bone window normalization, and RAS normalization; The NIfTI 3D volume data is input into the first segmentation network to obtain a complete 3D mask of the calcaneus, and a 3D anatomical coordinate system, normalized anatomical coordinates, and region of interest of the calcaneus are established based on the complete 3D mask of the calcaneus. The region of interest of the calcaneus is input into the second segmentation network to obtain a four-zone probability map of the anterior region, the posterior region, the upper articular surface related region, and the medial support region. The four-zone probability map is restricted to the complete 3D mask of the calcaneus to generate a four-zone label map. The spatial assignment of the structural block is determined by combining the 3D connected domain, the centroid of the structural block, and the anchor point projection parameters. Geometric angles, surface height differences, number of effective connected components, and 3D contact ratio are extracted based on the complete 3D mask of the calcaneus and the four-zone label map. The four-zone label map, structural block spatial assignment, geometric angle, surface height difference, number of effective connected domains, and three-dimensional contact ratio are encapsulated into a structured parameter file, and quality control tags are output.

2. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: When the data processing server converts the DICOM sequence into NIfTI 3D volume data, it simultaneously reads the image position, image orientation matrix, pixel spacing, and layer spacing, and generates an affine mapping relationship between the physical coordinate system of the DICOM sequence and the voxel coordinate system of the NIfTI 3D volume data. The data processing server establishes a three-dimensional anatomical coordinate system based on the physical coordinates of voxel points in the complete 3D mask of the calcaneus, and maps the voxel points in the complete 3D mask of the calcaneus to normalized anatomical coordinates in the posteroanterior, lateral-medial, and inferior-superior directions, respectively. The region of interest of the calcaneus is obtained by cropping the three-dimensional bounding box of the complete 3D mask of the calcaneus, and the mapping relationship between the region of interest of the calcaneus and the physical coordinate system of the DICOM sequence is preserved.

3. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: The data processing server aligns the four-zone probability map output by the second segmentation network with the voxel-by-voxel complete calcaneal 3D mask; When a voxel is located outside the complete 3D mask of the calcaneus, the voxel is assigned the background value; when a voxel is located inside the complete 3D mask of the calcaneus and its maximum partition probability reaches a preset probability threshold, the voxel is assigned the partition label corresponding to the maximum partition probability; when a voxel is located inside the complete 3D mask of the calcaneus and its maximum partition probability does not reach the preset probability threshold, it is completed based on the support strength of the partition labels already determined in the 26 neighborhoods of the 3D model. If the support strength has not yet reached the completion threshold, mark the voxel as a non-critical structural block.

4. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 3, characterized in that: The data processing server extracts the sagittal projection contour based on the complete calcaneus 3D mask and determines the geometric angles from the sagittal projection contour. The data processing server extracts the surface height difference and the number of effective connected components based on the relevant region of the upper joint surface in the four-zone label map, and extracts the three-dimensional contact ratio based on the rear region and the relevant region of the upper joint surface in the four-zone label map. The objects extracted based on geometric angles are independent of those extracted based on surface height difference, number of effective connected regions, and 3D contact ratio, but are all written into the structured parameter file.

5. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 4, characterized in that: After receiving the DICOM sequence, the data processing server verifies the image position, image orientation matrix, pixel spacing, and layer spacing. When the image position, image orientation matrix, pixel spacing, and layer spacing are all complete, it performs NIfTI 3D volume data conversion, resampling, and orientation normalization. When the layer spacing is missing and the image positions of adjacent slices are continuous, the layer spacing is reconstructed from the image position differences of adjacent slices, and then NIfTI 3D volume data conversion is performed. When the image orientation matrix is ​​missing and the orientation cannot be reconstructed from adjacent slice positions, a quality control marker for image orientation verification is generated, and intermediate data that retains the original slice order is output.

6. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 5, characterized in that: The data processing server performs main connected component ratio verification, hole filling status verification, and boundary coverage status verification on the complete 3D calcaneal mask; When the complete 3D mask of the calcaneus passes the main connected component proportion verification, hole filling status verification, and boundary coverage status verification, the region of interest (ROI) of the calcaneus is clipped based on the complete 3D mask of the calcaneus. When the complete 3D mask of the calcaneus fails any of the verifications, the clipping boundary of the ROI of the calcaneus is expanded, and the quality status corresponding to the complete 3D mask of the calcaneus is written into the structured parameter file. When the expanded ROI of the calcaneus still fails the boundary coverage status verification, the inference results of the four-zone probability map are retained, and a quality control mark indicating insufficient complete calcaneus segmentation quality is generated.

7. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: Before outputting the structured parameter file, the data processing server performs field integrity checks on image data number, CT voxel spacing, complete calcaneal mask quality index, four-zone segmentation results, structural block centroid, geometric angle, surface height difference, number of effective connected components, three-dimensional contact ratio, and quality control markers. When all fields exist and their format conforms to the field template, output a structured parameter file and a parameter visualization image; when at least one field is missing, output a structured parameter file containing the reason for the missing field and stop generating the parameter visualization image; when a field exists but the parameter source conflicts, retain the original field value and write the name of the conflicting field and the triggering reason in the quality control mark.

8. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 7, characterized in that: The first segmentation network and the second segmentation network are deployed in the same model service, which includes the first output branch and the second output branch. The first output branch receives NIfTI 3D volume data and outputs a complete calcaneal 3D mask and a calcaneal region of interest (ROI); the second output branch receives the ROI generated by the first output branch and outputs a four-zone probability map; the same model service records the model version of the first output branch, the model version of the second output branch, the file identifier of the complete calcaneal 3D mask, and the file identifier of the four-zone probability map in a single processing task.

9. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: When establishing a three-dimensional anatomical coordinate system, the data processing server selects the coordinate system generation strategy according to the strategy order in the configuration file. When the main connected component of the complete calcaneal three-dimensional mask meets the morphological integrity condition, the three-dimensional anatomical coordinate system is generated using the principal axis direction of the voxel points. When the main connected component does not meet the morphological integrity condition and the set of endpoints meets the separation condition, the three-dimensional anatomical coordinate system is generated using the posterior endpoint, anterior endpoint, upper endpoint, and lower endpoint. When the main connected component does not meet the morphological integrity condition and the endpoint set does not meet the separation condition, a three-dimensional anatomical coordinate system is generated by rigid registration between the preset calcaneal template and the complete calcaneal 3D mask, and the selected strategy identifier is written in the quality control mark.

10. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: The structured parameter file includes a JSON field file, NIfTI tag extended fields, and a 3D mesh model index file; The JSON field file records the image data number, CT voxel spacing, structural block spatial assignment, geometric angle, surface height difference, number of effective connected components, 3D contact ratio, and quality control markers; the NIfTI label extension field records the label values ​​of the four-zone label map in voxel space; The 3D mesh model index file records the file path, partition name, block number, and centroid coordinates of each structural block's 3D mesh model; the JSON field file, NIfTI tag extended fields, and the 3D mesh model index file share the same image data number.

11. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 1, characterized in that: The data processing server generates a parameter visualization interface, which displays the four-zone label diagram, structural block number, structural block centroid, geometric angle, surface height difference, number of effective connected domains, three-dimensional contact ratio, and quality control markers. The parameter visualization interface displays the front region, rear region, upper joint surface related region, inner support region and non-critical structural blocks with fixed field names; when quality control markers exist, the parameter visualization interface displays the triggering reason of the quality control markers; when no quality control markers exist, the parameter visualization interface displays the status information of the generated structured parameter file.

12. The automatic four-zone segmentation and measurement system for calcaneal fractures according to claim 11, characterized in that: The data processing server provides a parameter export interface, which receives export requests containing image data number, file type field, and request time field. When the file type field is a structured parameter file, the parameter export interface returns the file address, field summary, and quality control flag of the structured parameter file; when the file type field is a four-zone segmentation label graph, the parameter export interface returns the file address, voxel spacing, and label name list of the four-zone label graph; when the file type field does not belong to the preset file type, the parameter export interface returns a response message without the parameter file address and writes the unsupported file type flag in the response message.

13. The automatic segmentation and measurement system for four zones of calcaneal fracture according to claim 12, characterized in that: The data processing server generates a processing log for each processing task. The processing log records the image data number, DICOM sequence reception time, NIfTI 3D volume data generation status, complete calcaneal 3D mask generation status, calcaneal region of interest generation status, four-zone probability map generation status, four-zone label map generation status, structural block spatial assignment generation status, structured parameter file generation status, and quality control mark generation status. When any generation status is a failure, the processing log records the corresponding failure stage name and input file identifier; When all generation statuses are successful, the processing log records the address of the structured parameter file, the address of the four-zone segmentation label map, and the address of the parameter visualization image.

Images