A bone structure automatic segmentation method, system, electronic device and storage medium

Abstract

This invention discloses an automatic bone structure segmentation method, system, electronic device, and storage medium. The method utilizes a pre-trained multi-level, multi-task model group for subdividing the bone structure of a single vertebra, achieving automatic segmentation and extraction of bone structures. This reduces manual intervention and solves the problems of low efficiency and unreliability associated with manual annotation and segmentation of bone structures in existing technologies, laying a solid data foundation for subsequent clinical surgical planning. Furthermore, based on clinical indications and planning requirements, the method uses pre-trained primary and secondary cutting models to further subdivide the bone structure of a single vertebra. The multi-level, multi-task fine segmentation model group strategy, segmenting substructures from coarse to fine, can significantly improve segmentation accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an automatic bone structure segmentation method, system, electronic device, and storage medium. Background Technology

[0002] Currently, image-guided stereotactic navigation robots are rapidly developing in surgical subspecialties such as orthopedics, neurosurgery, and dentistry, and are gradually being adopted clinically. In orthopedics, these navigation robots are mainly used for assisted positioning and orientation of bone tissue or bony passages, including key steps such as drilling, grinding, and cutting. Spinal surgery, such as nerve decompression and internal fixation, often involves procedures around sensitive tissues like nerves and blood vessels, posing significant surgical risks. Robot-assisted spinal surgery helps improve the precision of these procedures and reduce complications. For spinal surgical navigation robots that rely on intraoperative image guidance, precise and efficient surgical planning is a crucial step.

[0003] Current surgical planning is typically performed manually by surgeons, a cumbersome and time-consuming process that relies heavily on the surgeon's experience and subjective judgment, making it difficult to guarantee planning quality and efficiency. Therefore, the development of automated planning systems is urgently needed.

[0004] Taking automated pedicle screw planning as an example, early studies mainly employed anatomical morphology optimization, statistical shape models, or template matching. These methods largely relied on manual intervention to set initial parameters, resulting in low template matching accuracy and poor robustness to vertebral variations. Recent research has largely utilized machine learning or deep learning techniques. Due to increased computing power, these methods significantly outperform traditional non-machine learning methods in terms of performance and runtime. In particular, Ronneberger et al.'s 2015 proposal of the U-Net network structure, using fewer training samples for end-to-end training, significantly improved the performance of deep learning networks and produced more accurate results. In such automated pedicle screw planning using machine learning methods, manual annotation is mainly based on setting feature points such as screw origin and insertion points in a single or multiple vertebrae. Nevertheless, mature pedicle screw planning technology still lacks a universally accepted gold standard. Even among clinically considered perfectly acceptable planning methods, significant differences exist. Therefore, training and learning based on manual planning or individual clinical experience may ultimately require fine-tuning of the automated planning results to achieve clinical acceptance. Such clinical controversies are also quite common in other spinal surgeries.

[0005] The spine, as a complex irregular bone, is further subdivided into intricate structures such as the vertebral body, pedicles, laminae, superior and inferior articular processes, and spinous processes. These substructures are adjacent to each other and each has its own unique morphological characteristics, with different shapes and sizes. These morphological characteristics are important references for planning channel screws or decompression internal fixation methods. Therefore, the automatic segmentation and extraction of these intricate vertebral structures is a more objective and reliable method.

[0006] The detailed segmentation images used in this technology are entirely based on manual delineation of the various substructures of the spine, which is inefficient and cannot guarantee accuracy. Furthermore, the boundaries in this technology are highly subjective, and its clinical significance is unclear. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide an automatic bone structure segmentation method, system, electronic device and storage medium to solve the technical problems of low efficiency and unreliability of manual annotation and segmentation of bone structures in related technologies.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0009] According to a first aspect of the present invention, an automatic bone structure segmentation method is provided, comprising:

[0010] Obtain an original image containing the complete vertebral structure, and continuously cut the vertebral structure in the original image to obtain multiple original images of single vertebral structures.

[0011] The original image of each vertebral bone structure is subdivided into substructures, and a mask is added to the subdivided substructures to obtain a substructure mask image.

[0012] For any single vertebral bone structure, the original image and substructure mask of the single vertebral bone structure are input into a pre-trained single vertebral multi-level multi-task model group to obtain the subdivided substructures of the single vertebral bone structure.

[0013] Preferably, the single-spine multi-level multi-task model group includes:

[0014] A first-level cutting model is used to perform first-level cutting on the single vertebral bone structure in the original image based on the substructure mask map, to obtain a coarse-divided image containing the first-level single vertebral bone structure.

[0015] Multiple secondary cutting models are used to perform secondary cutting on the primary single vertebral bone structure in the coarse image to obtain a subdivided image containing the secondary single vertebral bone structure.

[0016] Each single vertebral bone structure corresponds to a primary cutting model, each primary cutting model corresponds to multiple secondary cutting models, and each secondary cutting model corresponds to a type of clinical indication and planning requirement.

[0017] Preferably, the method further includes:

[0018] For the disconnected regions of the coarse segmented image, a connected component analysis algorithm is used to connect them with the surrounding connected regions, and the connected coarse segmented image is then input into the secondary segmentation model.

[0019] For the disconnected regions of the subdivided image, a connected component analysis algorithm is used to connect them with the surrounding connected regions, and the connected subdivided image is used as the output result.

[0020] Preferably, the method further includes:

[0021] Extract the morphological features of secondary single vertebral bone structures from the subdivided images;

[0022] Based on the morphological characteristics, multiple single vertebral bone structures are fused to generate a bony channel formed between the multiple single vertebral bone structures; the bony channel includes at least the spinal canal, intervertebral foramen, and intervertebral disc region.

[0023] Preferably, the first-level cutting model is trained based on a 3D U-net network model and includes:

[0024] Obtain training samples: Obtain an original image containing the complete vertebral structure; continuously segment the vertebral structure in the original image containing the complete vertebral structure to obtain an original image containing multiple individual vertebral structures; select the original image of the single vertebral structure to be trained from the multiple segmented original images of individual vertebral structures; add a mask to the first-level subdivision structure on the original image of the single vertebral structure to be trained to obtain the GT mask; use the original image of the single vertebral structure to be trained and the GT mask as training samples;

[0025] Model architecture construction: Define the network parameters of the 3D U-net network model, including the number of convolutional layers, the number of channels, the loss function, and the number of optimization iterations of the model; define two input channels, which are used to receive the original image of the single vertebral structure to be trained and the GT mask, respectively; define one output channel, which is used to output the coarse-segmented image;

[0026] Model training: Input the training samples into the constructed 3D U-net network model, perform iterative optimization, calculate the loss function between the first-level substructure segmentation mask predicted by the model and the GT mask, and adjust the network parameters according to the loss function until the number of optimization iterations reaches the preset number, and the value of the loss function is within the preset threshold range and tends to be stable. The model is then determined to have converged, and the model at this time is marked as a first-level segmentation model.

[0027] Model prediction: The newly acquired training samples are input into the first-level segmentation model to obtain a coarse-segmented image of the single vertebral bone structure to be trained.

[0028] Preferably, the training samples of the secondary cutting model are the coarse-segmented images output by the primary cutting model, and the coarse-segmented images carry mask information of the secondary single vertebral bone structures to be cut.

[0029] The network model structure of the secondary cutting model is trained using the same training method as the primary cutting model.

[0030] Preferably, before acquiring training samples, the method further includes: image preprocessing, the image preprocessing comprising:

[0031] The pixel spacing in the XYZ scanning directions of the original image of each single vertebral bone structure is adjusted to be consistent, so as to obtain the isotropic pixel volume and mask;

[0032] The image grayscale is divided into different levels, and the original image is then subjected to grayscale normalization processing.

[0033] Construct three-dimensional samples based on single vertebral structures. The size of the samples should at least cover the complete image of the vertebral structure to be segmented, so that the 3D U-net network model can learn the overall structural morphological features.

[0034] Preferably, the step of continuously cutting the vertebral structures in the original image to obtain multiple original images of individual vertebral structures specifically involves:

[0035] Obtain a raw 3D image containing the complete vertebral structure;

[0036] The complete vertebral structure in the original 3D image is continuously cut using a surface component, and the area between two adjacent target surfaces is determined as a single vertebral structure.

[0037] According to a second aspect of the present invention, an automatic bone structure segmentation system is provided, comprising:

[0038] The acquisition module is used to acquire an original image containing a complete vertebral structure, and to continuously cut the vertebral structure in the original image to obtain multiple original images of single vertebral structures.

[0039] The masking module is used to subdivide the single vertebral structure in the original image of each single vertebral structure and add a mask to the subdivided substructure to obtain a substructure mask image.

[0040] The segmentation module is used to input the original image and substructure mask of any single vertebral bone structure into a pre-trained single vertebral multi-level multi-task model group to obtain the subdivided structure of the single vertebral bone structure.

[0041] According to a third aspect of the present invention, an electronic device is provided, comprising:

[0042] A processor, and a memory connected to the processor;

[0043] The memory is used to store computer programs;

[0044] The processor is used to call and execute the computer program in the memory to perform the above-described method.

[0045] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, characterized in that the computer instructions are used to cause a computer to perform the above-described method.

[0046] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0047] By using a pre-trained multi-level, multi-task model group for single vertebral bone structure subdivision, automatic segmentation and extraction of bone structure is achieved, reducing manual intervention and solving the problems of low efficiency and unreliability of manual annotation and segmentation of bone structure in existing technologies. This lays a solid data foundation for subsequent clinical surgical planning.

[0048] Furthermore, based on clinical indications and planning requirements, pre-trained primary and secondary cutting models are used to further subdivide the single vertebral bone structure into substructures. Based on a multi-level, multi-task fine segmentation model group strategy, the substructures are segmented from coarse to fine, which can better improve the segmentation accuracy. At the same time, depending on the different clinical indications and planning requirements, secondary cutting models are selectively trained to complete different tasks, which reduces the complexity and mutual interference of fine substructure segmentation and can improve the accuracy of subsequent planning.

[0049] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating an automatic bone structure segmentation method according to an exemplary embodiment;

[0051] Figures 2A-2B This is a schematic diagram illustrating cutting using a curved surface component according to an exemplary embodiment;

[0052] Figures 3A-3H This is a schematic diagram of the structure of each individual vertebra according to an exemplary embodiment;

[0053] Figures 4A-4C This is a subdivision map obtained after secondary cutting, as shown in an exemplary embodiment;

[0054] Figure 5 This is a coarse segmentation image obtained after a first-level cut, as shown in an exemplary embodiment.

[0055] Figures 6A-6E This is a schematic diagram illustrating the morphological features of the lumbar spine structure according to an exemplary embodiment;

[0056] Figure 7 This is a schematic diagram of the structure of a 3D U-net network model according to an exemplary embodiment;

[0057] Figure 8 This is an output diagram of a 3D U-net network model illustrated according to an exemplary embodiment;

[0058] Figure 9 This is a flowchart illustrating an automatic lumbar spine segmentation method according to an exemplary embodiment;

[0059] Figure 10 This is a schematic block diagram illustrating an automatic bone structure segmentation system according to an exemplary embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] As mentioned in the background section, the related technologies suffer from the problems of low efficiency and unreliability in manually annotating and segmenting bone structures.

[0062] In order to effectively solve the problems in related technologies, the present invention provides an automatic bone structure segmentation method, system, electronic device and storage medium, which are described in detail below.

[0063] Example 1

[0064] Figure 1 This is a flowchart illustrating an automatic bone structure segmentation method according to an exemplary embodiment, such as... Figure 1 As shown, the method includes:

[0065] Step S11: Obtain an original image containing the complete vertebral structure, and continuously cut the vertebral structure in the original image to obtain multiple original images of single vertebral structures.

[0066] Step S12: Subdivide the single vertebral structure in the original image of each single vertebral structure into substructures, and add a mask to the subdivided substructures to obtain a substructure mask image.

[0067] Step S13: For any single vertebral bone structure, input the original image and substructure mask of the single vertebral bone structure into a pre-trained single vertebral multi-level multi-task model group to obtain the subdivided substructures of the single vertebral bone structure.

[0068] It should be noted that, in practice, the technical solution provided in this embodiment runs in the controller of the medical device, or is loaded into an electronic device connected to the controller. The controller of the medical device executes the corresponding method by calling the program stored in the electronic device.

[0069] The technical solution provided in this embodiment can be integrated into the navigation and planning system of medical robots to quickly and automatically generate morphology-based screw paths. Furthermore, future integration with mechanical analysis such as bone mineral density can expand its applicability and robustness. In addition, detailed segmentation and visualization of the structure can be used for endoscopic navigation, reducing the probability of getting lost in the surgical field.

[0070] In practice, step S11 involves continuously cutting the vertebral structures in the original image to obtain multiple original images of individual vertebral structures, specifically as follows:

[0071] Obtain a raw 3D image containing the complete vertebral structure;

[0072] The complete vertebral structure in the original 3D image is continuously cut using a surface component, and the area between two adjacent target surfaces is determined as a single vertebral structure.

[0073] See Figure 2A and Figure 2B When a curved surface component is initially placed at a location within the complete vertebral structure, it defaults to a flat surface. However, if the bone surface at that location has an irregular shape, it will be manually adjusted to suit the application. Figure 2BThe curved surface shown offers more precise cutting, but is more time-consuming. Both planes and curved surfaces can be manually sized to achieve precise segmentation of bone structures. The curvature of curved surfaces can be set to better wrap around the structure.

[0074] See Figures 3A-3H After continuously cutting the complete vertebral structure, original images of multiple individual vertebral structures were obtained, including: upper cervical vertebrae (C1), upper cervical vertebrae (C2), lower cervical vertebrae (C3-C7), upper thoracic vertebrae (T1-T4), middle thoracic vertebrae (T3-T8), lower thoracic vertebrae (T9-T12), lumbar vertebrae (L1-L5), and sacrum.

[0075] In step S12, a mask is added to the subdivided substructure. This can be done using existing automatic segmentation methods, or it can be done manually using automatic or semi-automatic methods.

[0076] For step S13, in specific practice, the single-spine multi-level multi-task model group includes:

[0077] A first-level cutting model is used to perform first-level cutting on the single vertebral bone structure in the original image based on the substructure mask map, to obtain a coarse-divided image containing the first-level single vertebral bone structure.

[0078] Multiple secondary cutting models are used to perform secondary cutting on the primary single vertebral bone structure in the coarse image to obtain a subdivided image containing the secondary single vertebral bone structure.

[0079] Each single vertebral bone structure corresponds to a primary cutting model, each primary cutting model corresponds to multiple secondary cutting models, and each secondary cutting model corresponds to a type of clinical indication and planning requirement.

[0080] Assuming the single vertebral structure to be segmented is the lumbar vertebra, see [link / reference]. Figures 4A-4C A single lumbar vertebra, by traditional definition, is mainly divided into the vertebral body, pedicles, facet joints (including the superior and inferior articular processes), laminae, transverse processes, and spinous processes. According to pre-defined rules (which can be referenced in spinal development, spinal biomechanics, and clinical applications), see [link to relevant documentation]. Figure 5 The vertebral body can be classified as anterior structure 101, while the facet joints, lamina, transverse processes, and spinous processes can be classified as posterior structures 102.

[0081] It should be noted that since the area near the pedicle is frequently needed as the surgical entry point in subsequent navigation planning, this embodiment proposes the concept of a pedicle entry point area to better handle subsequent entry point planning. The pedicle entry point area can be understood as being composed of parts of the transverse process, superior articular process, and pedicle, and its location intersects with all three, hence it is defined as a boundary area. The advantage of this definition is that it better determines the entry point by identifying the pedicle screw entry area, thus reducing the chance of bone slippage.

[0082] Therefore, considering the importance of the entry point area in robot-assisted pedicle screw fixation, this embodiment, based on the traditional classification method, will... Figure 4A In the left pedicle 2, left transverse process 6, and left superior articular process 8, the junction area of ​​these three is defined as the left pedicle entry point region 4; the junction area of ​​the right pedicle 3, right transverse process 7, and right superior articular process 9 is defined as the right pedicle entry point region 5. In the first-level cutting model, Figure 4A The left pedicle entry point region 4 and the right pedicle entry point region 5 are defined as follows: Figure 5 The rear structure 102 in the middle.

[0083] See Figure 5 Assuming the single vertebral bone structure to be segmented is the lumbar vertebra, the substructures after the first-order segmentation of the lumbar vertebra include: vertebral body 101, posterior structure 102, and pedicle 103.

[0084] See Figure 9 The lumbar spine corresponds to one primary cutting model and multiple secondary cutting models. The purpose of the primary cutting model is to obtain... Figure 5 The coarse-segmented image shown, the purpose of the two-stage cutting model is to obtain Figures 4A-4C The image shown is a detailed breakdown. Each secondary cutting model is trained to suit different clinical indications and planning needs. Therefore, for the lumbar spine, there can be multiple secondary cutting models, each designed to perform a different task.

[0085] See Figure 4A The substructures after secondary lumbar vertebral resection include: vertebral body 1, left pedicle 2, right pedicle 3, left pedicle entry point region 4, right pedicle entry point region 5, left transverse process 6, right transverse process 7, spinous process 14, left superior articular process 8, left inferior articular process 12, right superior articular process 9, right inferior articular process 13, left lamina 10, and right lamina 11, totaling 14 secondary substructures, of which 12 are symmetrical structures and 2 are independent structures (including vertebral body 1 and spinous process 14).

[0086] In summary, taking the lumbar spine as an example, the substructures after primary resection mainly include: vertebral body, pedicle, and posterior structures; the substructures after secondary resection are mainly a subdivision of the posterior structures, including: transverse process, pedicle entry point area, superior articular process, inferior articular process, lamina, and spinous process.

[0087] It is understood that the technical solution provided in this embodiment, targeting clinical indications and planning needs, adopts a multi-level, multi-task fine segmentation model group strategy. This decomposes the complex task of fine spinal segmentation into multiple sub-models to fulfill indication-based planning requirements, reducing the complexity of each model and improving its responsiveness. Users can select different secondary segmentation models according to different tasks, making it widely applicable and highly scalable.

[0088] In practice, the method also includes:

[0089] For the disconnected regions of the coarse segmented image, a connected component analysis algorithm is used to connect them with the surrounding connected regions, and the connected coarse segmented image is then input into the secondary segmentation model.

[0090] For the disconnected regions of the subdivided image, a connected component analysis algorithm is used to connect them with the surrounding connected regions, and the connected subdivided image is used as the output result.

[0091] In practice, the method also includes:

[0092] Extract the morphological features of secondary single vertebral bone structures from the subdivided images;

[0093] Based on the morphological characteristics, multiple single vertebral bone structures are fused to generate a bony channel formed between the multiple single vertebral bone structures; the bony channel includes at least the spinal canal, intervertebral foramen, and intervertebral disc region.

[0094] See Figures 6A-6E From a clinical application perspective, bony channels are of great significance. Therefore, after the sub-structures are further subdivided, bony channels need to be generated separately. Bony channels include, but are not limited to: the superior endplate of the vertebral body, the inferior endplate of the vertebral body, the spinal canal, the intervertebral foramen, the facet joints, and the lateral recesses.

[0095] For example, the spinal canal is a bony passage that naturally forms between multiple lumbar vertebrae when they are combined vertically.

[0096] To understand the specific location of the bony canals, please refer to [link / reference]. Figures 6A-6E , Figure 6A The diagram shows the superior endplate (the area within the upper dashed line in the diagram) and the inferior endplate (the area within the lower dashed line in the diagram) of the vertebral body. Figure 6B This represents the spinal canal (the area within the dashed line in the diagram contains the spinal cord). Figure 6CThis indicates the intervertebral foramen (within the area indicated by the dotted line in the diagram, where nerve roots enter and exit). Figure 6D This indicates the facet joints and articular surfaces (within the area indicated by the dashed lines in the diagram). Figure 6E This indicates the lateral recess (within the area indicated by the dotted line in the diagram).

[0097] It is understood that the technical solution provided in this embodiment takes into account the bony channels between multiple single vertebral bone structures after cutting. This provides a precise substructure segmentation map for subsequent clinical surgical planning, laying a solid foundation for achieving rapid, reliable and stable initial planning, and ultimately improving surgical efficiency, enhancing intraoperative decision support and safety reminders.

[0098] As mentioned above, the single-spine multi-level multi-task model group includes a first-level cutting model and multiple second-level cutting models, all of which are pre-trained.

[0099] Before model training, image preprocessing is required. This image preprocessing includes:

[0100] 1. Adjust the pixel spacing in the XYZ scanning directions of the original image of each single vertebral bone structure to be consistent, so as to obtain isotropic pixel volume and mask;

[0101] 2. Divide the image grayscale into different levels and perform grayscale processing on the original image;

[0102] 3. Construct three-dimensional samples based on single vertebral bone structures. The sample size should at least cover the complete image of the bone structure to be segmented, so that the 3D U-net network model can learn the overall structural morphological features.

[0103] In addition, during the image preprocessing stage, osteophytes need to be manually segmented from the vertebral structure using a mesh component (after the osteophytes are segmented, no mask is added to the segmented osteophytes, so the model training will not process the image of the osteophytes) to avoid the osteophytes affecting the model training results. Due to the varied morphology of osteophytes, their segmentation will not be done automatically for the time being.

[0104] In addition, sample data augmentation during network model training can include adding random noise, rotation around the three axes XYZ within a certain angle range, scaling within a certain range, etc.

[0105] Understandably, image preprocessing can ensure that the data input into the subsequent first-level and second-level cutting models can be accurately identified and effectively processed, thereby improving the reliability and accuracy of model training.

[0106] In practice, the first-level cutting model is trained based on a 3D U-net network model and includes:

[0107] Obtain training samples: Obtain the original image containing the complete vertebral structure; continuously segment the vertebral structure in the original image containing the complete vertebral structure to obtain the original image containing multiple single vertebral structures; select the original image of the single vertebral structure to be trained from the multiple segmented original images of single vertebral structures; add a mask to the first-level subdivision structure on the original image of the single vertebral structure to be trained to obtain the GT mask (Ground Truth, GT ground truth); use the original image of the single vertebral structure to be trained and the GT mask as training samples.

[0108] Model architecture construction: Define the network parameters of the 3D U-net network model, including the number of convolutional layers, the number of channels, the loss function, and the number of optimization iterations of the model; define two input channels, which are used to receive the original image of the single vertebral structure to be trained and the GT mask, respectively; define one output channel, which is used to output the coarse-segmented image;

[0109] Model training: Input the training samples into the constructed 3D U-net network model, perform iterative optimization, calculate the loss function between the first-level substructure segmentation mask predicted by the model and the GT mask, and adjust the network parameters according to the loss function until the number of optimization iterations reaches the preset number, and the value of the loss function is within the preset threshold range and tends to be stable. The model is then determined to have converged, and the model at this time is marked as a first-level segmentation model.

[0110] Model prediction: The newly acquired training samples are input into the first-level segmentation model to obtain a coarse-segmented image of the single vertebral bone structure to be trained.

[0111] Preferably, the training samples of the secondary cutting model are the coarse-segmented images output by the primary cutting model, and the coarse-segmented images carry mask information of the secondary single vertebral bone structures to be cut.

[0112] The network model structure and training method of the second-level cutting model are the same as those of the first-level cutting model, except that the network model parameters are different (the network model parameters include, but are not limited to: convolution type, number of convolution layers, weight distribution in the loss function, number of channels, and the number of optimization iterations of the model are slightly adjusted).

[0113] Taking the lumbar spine as an example, see Figure 7 The input to the 3D U-net network model is the original image of a single lumbar vertebra and its substructure mask. The output of the first-level segmentation model is a coarse-segmented image including the background, vertebral body, pedicle, and posterior structures. See the output image for reference. Figure 8 As shown.

[0114] The loss function can be various methods such as weighted cross-entropy or Dice loss function, or a combination thereof. In practice, given that there may be significant differences in volume between categories during substructure segmentation, this embodiment preferably uses weighted cross-entropy as the loss function. For example, the weights of [background, vertebral body, pedicle, posterior structure] are set to [1, 1, 6, 1] respectively, to reduce the problem of volume imbalance between categories and increase the weight of the smaller pedicle in the loss function.

[0115] It is understood that the technical solution provided in this embodiment can be applied to various application scenarios, including but not limited to processing the complex and varied substructure segmentation from the cervical spine to the thoracic and lumbar spine, and the substructure segmentation of any segment of the lumbar and thoracic spine, which only requires different classification methods to be used for sub-substructure subdivision.

[0116] In addition, this embodiment also proposes a multi-level, multi-task-based fine segmentation model group strategy. The segmentation substructure from coarse to fine can better improve the segmentation accuracy. At the same time, depending on the different clinical indications and planning needs, secondary segmentation models for different tasks are selectively trained, which reduces the complexity and mutual interference of the model's fine substructure segmentation and can improve the accuracy of subsequent planning.

[0117] Understandably, the introduction of multi-level models reduces the difficulty of training network models. For a network model, the more categories it needs to segment, the more complex the network structure itself may be, the larger the amount of training data required, and the easier it is for categories to be confused. This is especially true for delicate structures like the spine, where each substructure is relatively small, there are no obvious grayscale differences between substructures, and too many classifications make post-processing more difficult to distinguish. The technical solution provided in this implementation selectively trains secondary segmentation models to complete different tasks according to different clinical indications and planning needs, reducing the complexity and mutual interference of fine substructure segmentation in the model, and improving the accuracy of subsequent planning.

[0118] Example 2

[0119] Figure 10 This is a schematic block diagram illustrating an automatic bone structure segmentation system 100 according to an exemplary embodiment, such as... Figure 10 As shown, the system 100 includes:

[0120] The acquisition module 110 is used to acquire an original image containing a complete vertebral structure, and to continuously cut the vertebral structure in the original image to obtain multiple original images of single vertebral structures.

[0121] The masking module 120 is used to subdivide the single vertebral structure in the original image of each single vertebral structure and add a mask to the subdivided substructure to obtain a substructure mask image.

[0122] The segmentation module 130 is used to input the original image and substructure mask of any single vertebral bone structure into a pre-trained single vertebral multi-level multi-task model group to obtain the subdivided structure of the single vertebral bone structure.

[0123] It should be noted that, in practice, the technical solution provided in this embodiment runs in the controller of the medical device, or is loaded into an electronic device connected to the controller. The controller of the medical device executes the corresponding method by calling the program stored in the electronic device.

[0124] The implementation methods and beneficial effects of the above modules can be found in the description of the relevant steps in the above embodiments, and will not be repeated in this embodiment.

[0125] It is understood that the technical solution provided in this embodiment uses a pre-trained multi-level multi-task model group to subdivide the bone structure of a single vertebra into substructures, thereby realizing automatic segmentation and extraction of bone structures, reducing manual intervention, and solving the problems of low efficiency and unreliability of manual annotation and segmentation of bone structures in the prior art, thus laying a solid data foundation for subsequent clinical surgical planning.

[0126] Furthermore, based on clinical indications and planning requirements, pre-trained primary and secondary cutting models are used to further subdivide the single vertebral bone structure into substructures. Based on a multi-level, multi-task fine segmentation model group strategy, the substructures are segmented from coarse to fine, which can better improve the segmentation accuracy. At the same time, depending on the different clinical indications and planning requirements, secondary cutting models are selectively trained to complete different tasks, which reduces the complexity and mutual interference of fine substructure segmentation and can improve the accuracy of subsequent planning.

[0127] Example 3

[0128] An electronic device according to an exemplary embodiment includes:

[0129] A processor, and a memory connected to the processor;

[0130] The memory is used to store computer programs;

[0131] The processor is used to call and execute the computer program in the memory to perform the above-described method.

[0132] It is understood that the technical solution provided in this embodiment uses a pre-trained multi-level multi-task model group to subdivide the bone structure of a single vertebra into substructures, thereby realizing automatic segmentation and extraction of bone structures, reducing manual intervention, and solving the problems of low efficiency and unreliability of manual annotation and segmentation of bone structures in the prior art, thus laying a solid data foundation for subsequent clinical surgical planning.

[0133] Furthermore, based on clinical indications and planning requirements, pre-trained primary and secondary cutting models are used to further subdivide the single vertebral bone structure into substructures. Based on a multi-level, multi-task fine segmentation model group strategy, the substructures are segmented from coarse to fine, which can better improve the segmentation accuracy. At the same time, depending on the different clinical indications and planning requirements, secondary cutting models are selectively trained to complete different tasks, which reduces the complexity and mutual interference of fine substructure segmentation and can improve the accuracy of subsequent planning.

[0134] Example 4

[0135] An exemplary embodiment illustrates a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described above.

[0136] It is understood that the technical solution provided in this embodiment uses a pre-trained multi-level multi-task model group to subdivide the bone structure of a single vertebra into substructures, thereby realizing automatic segmentation and extraction of bone structures, reducing manual intervention, and solving the problems of low efficiency and unreliability of manual annotation and segmentation of bone structures in the prior art, thus laying a solid data foundation for subsequent clinical surgical planning.

[0137] Furthermore, based on clinical indications and planning requirements, pre-trained primary and secondary cutting models are used to further subdivide the single vertebral bone structure into substructures. Based on a multi-level, multi-task fine segmentation model group strategy, the substructures are segmented from coarse to fine, which can better improve the segmentation accuracy. At the same time, depending on the different clinical indications and planning requirements, secondary cutting models are selectively trained to complete different tasks, which reduces the complexity and mutual interference of fine substructure segmentation and can improve the accuracy of subsequent planning.

[0138] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a memory, magnetic disk, optical disk, etc.

[0139] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for automatic segmentation of bone structures, characterized in that, The method comprises the following steps: obtaining an original image containing a complete vertebral structure, continuously cutting the vertebral structure in the original image to obtain a plurality of original images of single vertebral structures; sub-structure subdivision is performed on the single vertebral structure in the original image of each single vertebral structure, and a sub-structure mask is added to the subdivided sub-structure to obtain a sub-structure mask image; for any single vertebral structure, the original image and the sub-structure mask image of the single vertebral structure are input into a pre-trained single vertebra multi-level multi-task model group to obtain the subdivided sub-structure of the single vertebral structure; the single vertebra multi-level multi-task model group comprises: a first cutting model for performing first cutting on the single vertebral structure in the original image according to the sub-structure mask image to obtain a coarse division image containing a first single vertebral structure; a plurality of second cutting models for performing second cutting on the first single vertebral structure in the coarse division image to obtain a subdivided image containing a second single vertebral structure; each single vertebral structure corresponds to a first cutting model, each first cutting model corresponds to a plurality of second cutting models, and each second cutting model corresponds to a type of clinical indication and planning requirement; the method further comprises: for the unconnected region of the coarse division image, the connected region analysis algorithm is used to connect with the surrounding connected region, and the connected coarse division image is input into the second cutting model; for the unconnected region of the subdivided image, the connected region analysis algorithm is used to connect with the surrounding connected region, and the connected subdivided image is taken as an output result.

2. The method of claim 1, wherein, Further comprising: extracting the morphological features of the second single vertebral structure in the subdivided image; fusing a plurality of single vertebral structures according to the morphological features to generate a bony channel formed between the plurality of single vertebral structures; the bony channel at least comprises a spinal canal, an intervertebral foramen and an intervertebral disc region.

3. The method of claim 1, wherein, The first cutting model is trained based on a 3D U-net network model and comprises: obtaining a training sample: obtaining an original image containing a complete vertebral structure, continuously cutting the vertebral structure in the original image containing a complete vertebral structure to obtain an original image containing a plurality of single vertebral structures, selecting the original image of the single vertebral structure to be trained from the plurality of single vertebral structures after cutting, adding a mask to the first subdivided structure of the original image of the single vertebral structure to be trained to obtain a GT mask, and taking the original image of the single vertebral structure to be trained and the GT mask as a training sample; constructing a model architecture: defining the network parameters of the 3D U-net network model, the network parameters comprising: the number of convolution layers, the number of channels, the loss function, and the optimization iteration number of the model; defining the input channel as two, which is used to receive the original image of the single vertebral structure to be trained and the GT mask; defining the output channel as one, which is used to output the coarse division image; Model training: input the training sample into the constructed 3D U-net network model, perform iterative optimization, calculate the loss function of the first-level substructure segmentation mask of the model prediction output and the GT mask, and adjust the network parameters according to the loss function until the optimization iteration reaches the preset number, and the loss function value is in the preset threshold range and tends to be stable, determine the model convergence, and mark the model at this time as the first-level segmentation model; Model prediction: input the reacquired training sample into the first-level segmentation model to obtain a coarse segmentation image of the current single vertebra structure to be trained.

4. The method of claim 1, wherein The training sample of the second-level segmentation model is a coarse segmentation image output by the first-level segmentation model, and the coarse segmentation image carries mask information of a second-level single vertebra structure to be segmented; The network model structure of the second-level segmentation model is trained based on the same training method as the first-level segmentation model.

5. The method of claim 3, wherein, Before the training sample is acquired, image preprocessing is further included, and the image preprocessing includes: Adjusting the pixel spacing in the XYZ three scanning directions in the original image of each single vertebra structure to be consistent to obtain isotropic pixel volume and mask; Dividing the image grayscale into different levels, and performing grayscale normalization processing on the original image; Constructing a three-dimensional sample in units of single vertebra structure, and the size of the sample covers at least the complete image of the vertebra structure to be segmented, so that the 3D U-net network model can learn the overall structural feature.

6. The method according to any one of claims 1 to 5, characterized in that, The continuous segmentation of the vertebra structure in the original image to obtain a plurality of original images of single vertebra structures is specifically: Acquiring a 3D original image containing a complete vertebra structure; Using a curved surface component to continuously segment the complete vertebra structure in the 3D original image, and determining the region between adjacent two target curved surfaces as a single vertebra structure.

7. A bone structure automatic segmentation system, characterized by, The system includes the bone structure automatic segmentation method of any one of claims 1 to 6. An acquisition module is configured to acquire an original image containing a complete vertebra structure, and continuously segment a vertebra structure in the original image to obtain a plurality of original images of single vertebra structures; A mask module is configured to subdivide a single vertebra structure in the original image of each single vertebra structure, and add a mask to the subdivided substructure to obtain a substructure mask image; A segmentation module is configured to input the original image and the substructure mask image of any single vertebra structure into a pre-trained single vertebra multi-level multi-task model group to obtain a subdivided structure of the single vertebra structure.

8. An electronic device, comprising: It includes: A processor and a memory connected to the processor; The memory is configured to store a computer program; The processor is configured to call and execute the computer program in the memory to execute the method of any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein, The computer instructions are configured to enable the computer to execute the method of any one of claims 1 to 6.

Images