Method and device for repairing a tubular structure and computer equipment

Abstract

The application relates to a tubular structure fracture completion method, device and computer equipment, the method comprising: acquiring a medical image of a target object, segmenting a tubular structure in the medical image to obtain an initial tubular structure mask, inputting the initial tubular structure mask into a tubular structure completion model to complete a missing part of the tubular structure in the initial tubular structure mask, obtaining a completed tubular structure mask, inputting the completed tubular structure mask and the medical image into an error learning model to obtain error information, and obtaining a target tubular structure mask based on the completed tubular structure mask and the error information. The above method can realize tubular structure fracture completion processing through a neural network model to obtain a target tubular structure mask. The process does not require manual intervention, avoids completion errors caused by manual intervention, and thus improves the accuracy of tubular structure fracture completion results.

Description

Technical Field

[0001] This application relates to the field of medical technology, and in particular to a method, apparatus and computer equipment for repairing fractures in tubular structures. Background Technology

[0002] Tubular structures (such as blood vessels, trachea, pancreatic duct, and ureter) play a vital role in maintaining human physiological functions. Accurate determination of the location relationship between tubular structures and lesions is typically necessary before clinical surgical planning to improve surgical outcomes. However, due to the complexity of tubular structures and the limited resolution of medical images, tubular structures extracted from medical images often exhibit fragmentation, thus affecting practical clinical applications.

[0003] In related technologies, manual annotation is used to repair fractures in tubular structures, resulting in complete tubular structures. However, these technologies require manual intervention, which can lead to inaccurate results in repairing fractures in tubular structures. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and computer equipment for repairing fractures in tubular structures to address the aforementioned technical problems. This would avoid manual intervention in repairing fractures in tubular structures and make the repair results more accurate.

[0005] In a first aspect, this application provides a method for repairing fractures in tubular structures, the method comprising:

[0006] Acquire medical images of the target object, including tubular structures;

[0007] The tubular structures in the medical images are segmented to obtain an initial tubular structure mask;

[0008] The initial tubular structure mask is input into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, thus obtaining the completed tubular structure mask.

[0009] The completed tubular structure mask and the medical image are input into the error learning model to obtain error information. Based on the completed tubular structure mask and the error information, the target tubular structure mask is obtained. The error information is used to compare the tubular structure in the completed tubular structure mask with the tubular structure in the gold standard of the completed tubular structure mask.

[0010] In one embodiment, the tubular structure completion model includes: a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder; an initial tubular structure mask is input into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, resulting in a completed tubular structure mask, including:

[0011] The initial tubular structure mask is input into the first downsampling encoder to obtain the first feature; the first feature is the result obtained by the first downsampling encoder through downsampling the initial tubular structure mask;

[0012] The first feature is input into the latent distribution space learning layer to obtain the mean and standard deviation distribution of the first feature. Then, the mean and standard deviation distribution are sampled to obtain the second feature.

[0013] The second feature is input into the first upsampling decoder to obtain the completed tubular structure mask; the completed tubular structure mask is the result of the first upsampling decoder upsampling the second feature.

[0014] In one embodiment, an initial tubular structure mask is input to a first downsampling encoder to obtain a first feature, including:

[0015] The initial tubular structure mask is divided into blocks to obtain multiple image blocks;

[0016] Multiple image blocks are input into the first downsampling encoder to obtain multiple first features corresponding to the multiple image blocks.

[0017] In one embodiment, the first feature is input into the latent distribution space learning layer to obtain the second feature, including:

[0018] Multiple primary features are input into the latent distribution space learning layer;

[0019] The mean and variance vectors are obtained by performing mean and variance calculations on each first feature through the latent distribution space learning layer. The initial distribution is obtained based on the mean and variance vectors, and the second feature is obtained by sampling the initial distribution.

[0020] In one embodiment, the training process of the tubular structure completion model includes:

[0021] Obtain the first training sample set, which includes tubular structure mask samples and corresponding gold standards for completing tubular structure masks;

[0022] The tubular mask sample is sequentially input into the first downsampling encoder and the latent distribution space learning layer to obtain the sample features after sampling the predicted initial distribution.

[0023] The sample features are input into the first upsampling decoder to obtain the predicted and completed tubular structure mask;

[0024] Based on the first difference between the initial predicted distribution and the corresponding ideal distribution, and the second difference between the predicted completed tubular structure mask and the gold standard of the completed tubular structure mask, the network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder are adjusted until convergence is achieved, thus obtaining the tubular structure completion model.

[0025] In one embodiment, the error learning model includes: a second downsampling encoder and a second upsampling decoder; the completed tubular structure mask and the medical image are input into the error learning network model to obtain error information, including...

[0026] The completed tubular structure mask and the medical image are input into the second downsampling encoder to obtain the third feature; the third feature is the result obtained by the second downsampling encoder through downsampling processing of the completed tubular structure mask and the medical image;

[0027] The third feature is input into the second upsampling decoder to obtain error information; the error information is the result obtained by the second upsampling decoder after upsampling the third feature.

[0028] In one embodiment, the training process of the error learning model includes:

[0029] Obtain the second training sample set; the second training sample set includes medical image samples, incomplete tubular structure mask samples, and the corresponding incomplete tubular structure mask gold standard;

[0030] The medical image sample and the completed tubular structure mask sample are sequentially input into the second downsampling encoder and the second upsampling decoder to obtain the prediction error information.

[0031] Based on the difference between the predicted error information and the ideal error information, the network parameters of the second downsampling encoder and the second upsampling decoder are adjusted until convergence, thus obtaining the error learning model; the ideal error information is used to characterize the difference between the completed tubular structure mask sample and the completed tubular structure mask gold standard.

[0032] In one embodiment, obtaining the target tubular structure mask based on the completed tubular structure mask and error information includes:

[0033] The target tubular structure mask is obtained by superimposing the completed tubular structure mask with the error information.

[0034] Secondly, this application provides a tubular structure fracture repair device, the device comprising:

[0035] The medical image acquisition module is used to acquire medical images of the target object, including tubular structures.

[0036] The segmentation processing module is used to segment tubular structures in medical images to obtain an initial tubular mask;

[0037] The completion processing module is used to input the initial tubular mask into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular mask, so as to obtain the completed tubular structure mask.

[0038] The completion result repair module is used to input the completed tubular structure mask and medical image into the error learning model to obtain error information, and to obtain the target tubular structure mask based on the completed tubular structure mask and the error information. The error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

[0039] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any of the embodiments in the first aspect described above.

[0040] This application provides a method, apparatus, and computer device for completing tubular structure fractures. The method includes: acquiring a medical image of a target object; segmenting the tubular structures in the medical image to obtain an initial tubular structure mask; inputting the initial tubular structure mask into a tubular structure completion model to complete the missing parts of the tubular structures in the initial tubular structure mask, obtaining a completed tubular structure mask; inputting the completed tubular structure mask and the medical image into an error learning model to obtain error information; and obtaining a target tubular structure mask based on the completed tubular structure mask and the error information. This method can... A neural network model is used to complete the fracture repair of tubular structures, obtaining a mask of the target tubular structure. This process does not require manual intervention, avoiding repair errors caused by manual intervention, thus improving the accuracy of the tubular structure fracture repair results. It can also increase the speed of tubular structure fracture repair. At the same time, this method simplifies the fracture repair process, saves manpower, and reduces the cost of tubular structure fracture repair. In addition, based on the accurate tubular structure fracture repair results, the positional relationship between the target object's tubular structure and the lesion area can be accurately determined, further improving the accuracy of surgical planning. Attached Figure Description

[0041] Figure 1 This is an internal structural diagram of a computer device in one embodiment;

[0042] Figure 2 This is a flowchart illustrating a method for repairing fractures in a tubular structure in one embodiment;

[0043] Figure 3 This is a diagram illustrating the process of repairing a tracheal rupture in one embodiment.

[0044] Figure 4 This is a diagram illustrating the process of repairing a severed portal vein in one embodiment.

[0045] Figure 5 This is a diagram illustrating the process of completing a broken tubular structure in a medical image in one embodiment.

[0046] Figure 6 This is a flowchart illustrating the tubular structure fracture repair method in another embodiment;

[0047] Figure 7 This is a flowchart illustrating the tubular structure fracture repair method in another embodiment;

[0048] Figure 8 This is a flowchart illustrating the tubular structure fracture repair method in another embodiment;

[0049] Figure 9 This is a flowchart illustrating the tubular structure fracture repair method in another embodiment;

[0050] Figure 10 This is a flowchart illustrating the tubular structure fracture repair method in another embodiment;

[0051] Figure 11 This is a diagram showing the overall model structure of the tubular structure completion model and the error learning model in one embodiment.

[0052] Figure 12 This is a structural block diagram of a tubular structure fracture repair device in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] In clinical surgical techniques, such as surgeries involving tubular structures, the primary focus is determining the positional relationship between the tubular structure and the lesion area before surgery. Specifically, determining this relationship typically relies on 3D reconstructed medical images, involving manual segmentation and reconstruction of the organ's tubular structure. Then, visualization techniques such as multi-layer reconstruction (MPR) and virtual reality (VR) are used to further confirm the positional relationship between the tubular structure and the lesion area. Accurate tubular structure segmentation improves the accuracy of subsequent tasks, such as hemodynamic calculations, endoscopic route planning, and disease prediction and modeling. However, in practice, due to the complexity of tubular structures (such as curvature and thickness variations) and the limited resolution of medical images, the reconstructed tubular structure often exhibits breaks, affecting subsequent tasks and ultimately impacting clinical application. To address these issues, it is usually necessary to perform break-filling processing on the segmented and reconstructed tubular structure.

[0055] In related technologies, most methods employ manual completion to achieve tubular structure fracture repair. Alternatively, image processing methods are combined, such as assisted active shape model adjustment, region growing, and centerline growing, to achieve tubular structure fracture repair. However, these techniques often result in inaccurate tubular structure fracture repair results.

[0056] Based on this, embodiments of this application provide a method for repairing fractures in tubular structures, which can improve the accuracy of the repair results. Specifically, this method for repairing fractures in tubular structures is applicable to… Figure 1 The computer device shown can be a standalone server or a server cluster consisting of multiple servers, and can also be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The following embodiments will illustrate the implementation process of the tubular structure fracture repair method.

[0057] like Figure 2 The diagram shown is a flowchart illustrating a method for repairing fractures in a tubular structure according to an embodiment of this application. This method may include the following steps:

[0058] S100. Acquire a medical image of the target object. The medical image includes tubular structures.

[0059] Specifically, the aforementioned medical images can be X-ray images, computed tomography (CT) images, magnetic resonance imaging (MRI) images, or ultrasound images, etc.

[0060] The medical image can be medical data obtained by a medical scanning device scanning the whole body or a target area of ​​a target subject in real time, and then sending the medical data to a computer device, which reconstructs the image from the medical data. Alternatively, the medical image can be medical data obtained by a medical scanning device scanning the whole body or a target area of ​​a target subject in advance, and then sending the medical data to a computer device, which reconstructs the image from the medical data. The medical image can be stored in the cloud, cloud storage, hardware, or local storage, and in actual processing, the computer device can retrieve the medical image of the target subject from the cloud, cloud storage, hardware, or local storage.

[0061] S200. The tubular structure in the medical image is segmented to obtain the initial tubular structure mask.

[0062] It should be noted that computer equipment can segment tubular structures in medical images based on their location to obtain an initial tubular structure mask. Alternatively, to improve the accuracy of segmentation and reduce segmentation errors, the medical image of the target object can be input into a pre-trained tubular structure segmentation model to obtain an initial tubular structure mask. It should be noted that the tubular structure segmentation model can be a network model composed of at least one of the following: a fully convolutional neural network model (such as V-Net, U-Net, and VBNet), a recurrent neural network model, or an adversarial neural network model.

[0063] The initial tubular structure mask can be understood as the tubular structure label mask corresponding to the tubular structure segmentation result in the medical image. The initial tubular structure mask has the same size as the medical image, but each pixel within the tubular structure region has a pixel value of 1, while each pixel outside the tubular structure region has a pixel value of 0. Furthermore, the mean of the initial tubular structure mask can be 1, and the standard deviation can be 0.

[0064] In addition, embodiments of this application can perform preprocessing such as denoising, enhancement, and normalization on medical images to obtain preprocessed images. Then, the preprocessed images are segmented to obtain an initial tubular structure mask, thereby improving the accuracy of the segmentation process and further enhancing the accuracy of the tubular structure fracture completion results. It should be noted that normalization can be understood as a process of normalization based on the mean and standard deviation.

[0065] S300. Input the initial tubular structure mask into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, and obtain the completed tubular structure mask.

[0066] The aforementioned tubular structure completion model can also be a network model composed of at least one of convolutional neural network models, recurrent neural network models, and adversarial neural network models. However, the internal connection structure of the tubular structure completion model differs from that of the tubular structure segmentation model. In the embodiments of this application, the tubular structure completion model can also be a network model composed of one or more other neural network models.

[0067] Based on the initial tubular structure mask obtained in the above steps, it can be input into the tubular structure completion model. The tubular structure completion model then completes the missing parts of the tubular structure in the initial tubular structure mask to obtain the completed tubular structure mask. It should be noted that the completed tubular structure mask and the initial tubular structure mask are the same size, and the pixel values ​​at corresponding positions in the completed and initial tubular structure masks can be the same or different.

[0068] S400. Input the completed tubular structure mask and the medical image into the error learning model to obtain error information, and obtain the target tubular structure mask based on the completed tubular structure mask and the error information; the error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

[0069] It should be noted that when the tubular structure completion model completes the tubular structure mask, there may be processing errors. This could result in the tubular structure in the completed tubular structure mask not being a complete tubular structure completion mask. Therefore, the completed tubular structure mask and medical image obtained in the previous steps can be input into the error learning model to obtain error information, namely the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

[0070] Optionally, the error learning model can also be a network model composed of at least one of the following: fully convolutional neural network models (such as V-Net, U-Net, and VBNet), recurrent neural network models, and adversarial neural network models. However, the internal connection structure of the error learning model differs from that of the tubular structure completion model and the tubular structure segmentation model. In the embodiments of this application, the error learning model can also be a network model composed of one or more other neural network models.

[0071] Furthermore, based on the error information obtained in the previous steps, the incomplete tubular structure mask and the error information can be processed to obtain the target tubular structure mask, that is, the complete tubular structure incomplete mask corresponding to the medical image.

[0072] In this embodiment of the application, the step of obtaining the target tubular structure mask based on the completed tubular structure mask and error information in S400 above may include: superimposing the completed tubular structure mask and the error information to obtain the target tubular structure mask.

[0073] It should be noted that by using the difference, i.e., the error information, between the completed tubular structure mask and the complete tubular structure mask corresponding to the medical image, the broken tubular structure mask can be repaired. In other words, the completed tubular structure mask and the error information are superimposed to obtain the complete tubular structure mask corresponding to the medical image, i.e., the target tubular structure mask.

[0074] In practical applications, the error information is the same size as the mask for completing the tubular structure. This error information can also be represented by an error mask. If all the error information consists of pixels with a pixel value of 0, then it is equivalent to the error information being 0. In this case, the mask for completing the tubular structure can be determined as the target tubular structure mask. That is, the tubular structure breakage completion method does not need to execute the step in S400 of obtaining the target tubular structure mask based on the mask for completing the tubular structure and the error information.

[0075] For example, the process of repairing a broken tubular structure is illustrated using the trachea and portal vein as tubular structures. Figure 3 The diagram shown illustrates the process of repairing a ruptured trachea. Figure 3 (a) shows the initial tracheal mask corresponding to the trachea (the circled area shows the broken part in the trachea), and (b) shows the completed tracheal mask obtained after completing the initial tracheal mask in (a) (the circled area shows the broken part in the trachea). Figure 3 (b) shows the tubular structure after fracture repair corresponding to the fractured part; (c) shows the target tracheal mask obtained by repairing the fractured tracheal mask in (b) (circled). Figure 3 (a) The tubular structure after repair and restoration of the fractured portion. Additionally... Figure 4 The diagram shown illustrates the process of repairing a severed portal vein. Figure 4 (a) shows the initial portal vein mask corresponding to the portal vein (circles indicate broken portions of the portal vein), and (b) shows the completed portal vein mask obtained after completing the initial portal vein mask in (a) (circles indicate broken portions of the portal vein). Figure 4 (a) shows the tubular structure after the fracture is repaired, corresponding to the fractured portion in (b). (c) shows the target portal vein mask obtained by repairing the fractured portal vein mask in (b) (circled). Figure 4 (a) The tubular structure after repair and restoration of the fractured portion.

[0076] Another example, Figure 5The image shows the initial tubular structure mask and the target tubular structure mask generated during the process of completing a broken tubular structure in a medical image. The tubular structure can be a blood vessel, trachea, pancreatic duct, ureter, etc. Figure 5 In the image, (a) is the medical image O, (b) is the initial tubular structure image O1 corresponding to (a) (the dashed box shows the fractured portion in the initial tubular structure image O1), and (c) is the target tubular structure image O2 corresponding to (b) (the dashed box shows the repaired tubular structure corresponding to the fractured portion in the initial tubular structure image O1). The initial tubular structure image is the medical image corresponding to the initial tubular structure mask, and the target tubular structure image is the medical image corresponding to the target tubular structure mask.

[0077] The tubular structure fracture completion method in this embodiment can acquire a medical image of the target object, segment the tubular structure in the medical image to obtain an initial tubular structure mask, input the initial tubular structure mask into a tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, and obtain a completed tubular structure mask. The completed tubular structure mask and the medical image are input into an error learning model to obtain error information, and the target tubular structure mask is obtained based on the completed tubular structure mask and the error information. This method can be implemented using a neural network model. The tubular structure fracture repair process yields a target tubular structure mask. This process requires no manual intervention, avoiding repair errors caused by manual intervention and thus improving the accuracy of the tubular structure fracture repair results. It can also increase the speed of tubular structure fracture repair. At the same time, this method simplifies the fracture repair process, saves manpower, and reduces the cost of tubular structure fracture repair. In addition, based on the accurate tubular structure fracture repair results, the positional relationship between the target object's tubular structure and the lesion area can be accurately determined, further improving the accuracy of surgical planning.

[0078] Based on the above embodiments, the following describes the tubular structure completion model and the process of inputting an initial tubular structure mask into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, thereby obtaining a completed tubular structure mask. In one embodiment, the tubular structure completion model includes: a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder; as shown... Figure 6 As shown, the above S300 includes the following steps:

[0079] S310. Input the initial tubular structure mask into the first downsampling encoder to obtain the first feature; the first feature is the result obtained by the first downsampling encoder through downsampling processing of the initial tubular structure mask.

[0080] In this embodiment, the tubular structure completion model includes a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder. The first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder are connected sequentially. The first downsampling encoder is the input of the tubular structure completion model, and the first upsampling decoder is the output of the tubular structure completion model.

[0081] It should be noted that the computer device can input the initial tubular structure mask obtained in the previous steps into the first downsampling encoder in the tubular structure completion model. The first downsampling encoder performs downsampling processing on the initial tubular structure mask to obtain a feature map, i.e., the first feature. Here, the first downsampling encoder is a pre-trained network layer.

[0082] S320. Input the first feature into the latent distribution space learning layer to obtain the mean and standard deviation distribution of the first feature. Then, after sampling the mean and standard deviation distribution, the second feature is obtained.

[0083] Based on the first feature obtained in the previous steps, it can be input into the latent distribution space learning layer in the tubular structure completion model. The latent distribution space learning layer performs calculations and sampling on the first feature to obtain a feature map, i.e., the second feature. It should be noted that the calculation and sampling on the first feature can be a process of performing calculations on the first feature to obtain the mean and standard deviation distribution, and then sampling the mean and standard deviation distribution to obtain the second feature.

[0084] The latent distribution space learning layer is a pre-trained network layer. Optionally, the above operations can be addition, subtraction, multiplication, and / or division, etc.

[0085] S330. Input the second feature into the first upsampling decoder to obtain the completed tubular structure mask; the completed tubular structure mask is the result obtained by the first upsampling decoder by upsampling the second feature.

[0086] After obtaining the second feature, it can be input into the first upsampling decoder in the tubular structure completion model. The first upsampling decoder upsamples the second feature to obtain a feature map, which is the completed tubular structure mask. The first upsampling encoder is a pre-trained network layer.

[0087] In the embodiments of this application, the above-mentioned complete tubular structure mask can be understood as a complete mask for the preliminary tubular structure.

[0088] The tubular structure fracture completion method in this embodiment can process the initial tubular structure mask sequentially through a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder to obtain a completed tubular structure mask. This method can realize the tubular structure fracture completion process through a neural network model to obtain a completed tubular structure mask. This process does not require manual intervention, avoiding completion errors caused by manual intervention, thereby improving the accuracy of the tubular structure fracture completion result, and can also improve the speed of tubular structure fracture completion.

[0089] The process of inputting the initial tubular structure mask into the first downsampling encoder to obtain the first feature will be described below. In one embodiment, as follows: Figure 7 As shown, the above S310 may include the following steps:

[0090] S311. The initial tubular structure mask is divided into blocks to obtain multiple image blocks.

[0091] Specifically, if the initial tubular structure mask is large, to improve the processing speed and efficiency of the tubular structure completion model, the initial tubular structure mask can be first divided into blocks to obtain multiple image blocks of the same size. The tubular structure completion model then processes the block-based processing results. In this embodiment, the size of the specific image block can be determined based on the amount of data that the tubular structure completion model can process during optimal processing. Simultaneously, the resolution of the initial tubular structure mask can be set to maximize the resolution of the acquired initial tubular structure mask, thereby improving the accuracy of the tubular structure break completion result.

[0092] For example, taking the trachea as an example, the preferred size of the image block is 128*128*128, and the resolution of the initial tubular structure mask is selected according to the actual situation. For example, the preferred resolution of the initial tubular structure mask corresponding to the trachea is 0.7mm. These data are not limited in this application and can also be adjusted according to the actual situation.

[0093] S312. Input multiple image blocks into the first downsampling encoder to obtain multiple first features corresponding to the multiple image blocks.

[0094] Based on the multiple image blocks corresponding to the initial tubular structure mask obtained through the above steps, these multiple image blocks can be input into the first downsampling encoder to obtain multiple first features corresponding to the multiple image blocks, that is, each image block obtains a corresponding first feature. The size of the first feature is the same as the size of the corresponding image block.

[0095] Furthermore, based on the multiple first features obtained in the preceding steps, the process of inputting the first features into the latent distribution space learning layer to obtain the second features is described below. In one embodiment, S320 may include the following steps: inputting multiple first features into the latent distribution space learning layer, performing mean and variance calculations on each first feature through the latent distribution space learning layer to obtain a mean vector and a variance vector, obtaining an initial distribution based on the mean vector and variance vector, and performing sampling processing on the initial distribution to obtain the second features.

[0096] It should be noted that, based on the first features of multiple image patches corresponding to the obtained initial tubular structure mask, these first features can be directly input into the latent distribution space learning layer. Here, it needs to be explained that the latent distribution space learning layer calculates the mean of each first feature by averaging the data, and obtains a mean vector based on the mean of each first feature. Simultaneously, it calculates the variance of each first feature by averaging the data, and obtains a variance vector based on the variance of each first feature. Further, based on the obtained mean vector and variance vector, the initial distribution that the mean vector and variance vector follow is calculated. Then, the initial distribution is sampled to obtain the second feature.

[0097] Optionally, the initial distribution that the mean vector and variance vector follow can be a normal distribution, a Poisson distribution, an exponential distribution, an F distribution, or a 0-1 distribution, etc. For example, to calculate the initial distribution that the mean vector and variance vector follow, the mean vector and variance vector can be input into a pre-trained algorithm model, which outputs the initial distribution that the mean vector and variance vector follow.

[0098] The tubular structure fracture completion method in this embodiment can divide the initial tubular structure mask into blocks to obtain multiple image blocks. The multiple image blocks are input into a first downsampling encoder to obtain multiple first features corresponding to the multiple image blocks. The tubular structure mask is then completed by further processing the multiple first features. This method can divide big data into small data for processing, thereby reducing the complexity of data calculation, avoiding the probability of calculation errors, and improving the accuracy of tubular structure fracture completion results.

[0099] The following describes the process of obtaining the tubular structure completion model in the above steps. The tubular structure completion model is obtained by training a tubular structure completion network. Based on this, in one embodiment, as follows... Figure 8 As shown, the training process of the above tubular structure completion model includes:

[0100] S340. Obtain the first training sample set, which includes tubular structure mask samples and corresponding gold standards for completing tubular structure masks.

[0101] Specifically, the same type of tubular structure within different objects is different, and even different types of tubular structures within the same object are different. Therefore, during the training of the tubular structure completion network, in order to improve the generalization ability of the network and the robustness of tubular structure break completion, the first training sample set may include multiple tubular structure masks corresponding to different types of tubular structures within different objects, i.e., tubular structure mask samples, and the gold standard for completing tubular structures corresponding to the tubular structure masks. The gold standard for completing tubular structures can be understood as the ideal complete tubular structure mask.

[0102] Before training the tubular structure completion network, medical scanning equipment can pre-scan different subjects' whole-body or target areas to obtain medical data, and send the medical data to a computer. The computer reconstructs the medical data to obtain medical images, and then segments the tubular structures in the medical images to obtain tubular structure mask samples. These tubular structure mask samples and the corresponding completion tubular structure mask gold standards can be stored in the cloud, cloud storage, hardware, or local storage. When training the tubular structure completion network, the computer can retrieve the pre-stored tubular structure mask samples and the corresponding completion tubular structure mask gold standards from the cloud, cloud storage, hardware, or local storage.

[0103] In this embodiment of the application, the first training sample set contains a large amount of data to train the optimal tubular structure completion model, thereby enabling the tubular structure completion model to guide the consistency of the tubular structure topology when performing fracture completion processing.

[0104] S350. The tubular structure mask sample is sequentially input into the first downsampling encoder and the latent distribution space learning layer to obtain the sample features after sampling processing of the predicted initial distribution.

[0105] The tubular structure completion network can be trained using tubular structure mask samples and the corresponding gold standard for completing the tubular structure mask, resulting in a tubular structure completion model. During the training process, multiple tubular structure mask samples from the first training sample set are sequentially input into the first downsampling encoder and the latent distribution space learning layer of the tubular structure completion network. This yields the sample features output by the latent distribution space learning layer after sampling the predicted initial distribution. Specifically, the latent distribution space learning layer first obtains the predicted initial distribution, and then samples the predicted initial distribution to output sample features.

[0106] It should be noted that during the training of the tubular structure completion network, multiple tubular structure mask samples are of the same size. If the sizes of multiple tubular structure mask samples in the first training sample set are different, the multiple tubular structure mask samples can be cropped into images of the same size before being sequentially input into the first downsampling encoder and the latent distribution space learning layer. Furthermore, the aforementioned initial prediction distribution can be the initial distribution that follows between the training mean vector and the training variance vector corresponding to the tubular structure mask samples. This initial prediction distribution can also be a normal distribution, Poisson distribution, exponential distribution, F distribution, or 0-1 distribution, etc.

[0107] Meanwhile, in the embodiments of this application, the processing procedure performed by the latent distribution space learning layer can be designed as a processing procedure based on the variational autoencoder network model.

[0108] S360. Input the sample features into the first upsampling decoder to obtain the predicted and completed tubular structure mask.

[0109] After obtaining the sample features, these features can be further input into the first upsampling decoder in the tubular structure completion network to obtain the predicted completed tubular structure mask. The predicted completed tubular structure mask has the same size as the sample tubular structure mask.

[0110] S370. Based on the first difference between the initial predicted distribution and the corresponding ideal distribution, and the second difference between the predicted completed tubular structure mask and the gold standard of the completed tubular structure mask, adjust the network parameters of the first downsampling encoder, the latent distribution space learning layer and the first upsampling decoder until convergence, and obtain the tubular structure completion model.

[0111] In the embodiments of this application, the type of the initial predicted distribution and the corresponding ideal distribution, such as normal distribution, Poisson distribution, exponential distribution, F distribution or 0-1 distribution, may be the same or different, but the type of the ideal distribution is fixed.

[0112] Based on the initial predicted distribution obtained above, a first difference between the initial predicted distribution and the corresponding ideal distribution can be calculated using a first loss function, and a second difference between the predicted completed tubular structure mask and the gold standard of the completed tubular structure mask can be calculated using a second loss function. The network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder in the tubular structure completion network are further adjusted according to the first and second differences. Then, steps S350-S370 are continued until the tubular structure completion network converges, resulting in a trained tubular structure completion model, i.e., a trained first downsampling encoder, latent distribution space learning layer, and first upsampling decoder. The first and second loss functions can be L1 norm loss functions, cross-entropy loss functions, mean squared error loss functions, binary cross-entropy loss functions, etc.

[0113] It should be noted that when training the tubular structure completion network, the network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder in the tubular structure completion network can be obtained through initialization. The convergence condition of the above-mentioned tubular structure completion network can be that both the first difference and the second difference are less than a first preset difference threshold, or that the number of iterations of S350-S370 is greater than or equal to a first preset iteration number threshold. Here, the first preset difference threshold can be a positive number close to 0, and the first preset iteration number threshold can be a value greater than 0.

[0114] The tubular structure fracture completion method in this application embodiment can train the tubular structure completion network to obtain the optimal tubular structure completion model, thereby improving the accuracy of the completed tubular structure mask obtained through the optimal tubular structure completion model.

[0115] Based on the above embodiments, the process of inputting the completed tubular structure mask and medical image into the error learning network model to obtain error information will be described below. In one embodiment, the error learning model includes: a second downsampling encoder and a second upsampling decoder; as shown... Figure 9 As shown, the above S400 includes the following steps:

[0116] S410. The completed tubular structure mask and the medical image are input into the second downsampling encoder to obtain the third feature. The third feature is the result of downsampling the completed tubular structure mask and the medical image by the second downsampling encoder.

[0117] In this embodiment, the error learning model includes a second downsampling encoder and a second upsampling decoder. The second downsampling encoder and the second upsampling decoder are connected sequentially, with the second downsampling encoder serving as the input to the error learning model and the second upsampling decoder serving as its output.

[0118] It should be noted that the computer device can input the completed tubular structure mask and medical image obtained in the previous steps into the second downsampling encoder in the error learning model. The second downsampling encoder performs downsampling processing on the completed tubular structure mask and medical image to obtain a feature map, i.e., the third feature. The second downsampling encoder is pre-trained.

[0119] S420. The third feature is input to the second upsampling decoder to obtain error information. The error information is the result of the second upsampling decoder performing upsampling processing on the third feature.

[0120] After obtaining the third feature, it can be input into the second upsampling decoder in the error learning model. The second upsampling decoder upsamples the third feature to obtain error information. The second upsampling encoder is pre-trained.

[0121] The tubular structure fracture completion method in this embodiment can sequentially process the completed tubular structure mask and the medical image using a second downsampling encoder and a second upsampling decoder to obtain error information. This method can process the completed tubular structure mask using a neural network model to obtain the incomplete tubular structure mask, i.e., the error information, in the completed tubular structure mask. This process does not require manual intervention, avoiding errors caused by human intervention, thereby improving the accuracy of the obtained incomplete tubular structure mask and providing a basis for further obtaining an accurate target tubular structure mask.

[0122] The process of obtaining the error learning model in the above steps is described below. The error learning model is obtained by training an error learning network. Based on this, in one embodiment, as follows... Figure 10 As shown, the training process of the above error learning model includes:

[0123] S430. Obtain the second training sample set; the second training sample set includes medical image samples, completed tubular structure mask samples, and the corresponding completed tubular structure mask gold standard.

[0124] Specifically, the same type of tubular structure within different objects is different, and even different types of tubular structures within the same object are different. Therefore, during the training of the error learning network, in order to improve the generalization ability and robustness of the error learning network, the second training sample set may include medical images corresponding to different types of tubular structures of different objects (i.e., medical image samples), complete tubular structure mask samples corresponding to the medical image samples, and complete tubular structure mask gold standards corresponding to the complete tubular structure mask samples. The complete tubular structure mask gold standard in this step is the same as the complete tubular structure mask gold standard in S340 above.

[0125] Among them, the medical image samples, the completed tubular structure mask samples, and the corresponding completed tubular structure mask gold standards are all obtained in advance before training the error learning network. They can be stored in advance in the cloud, cloud storage, hardware, or local location. When training the error learning network, the computer device can obtain the pre-stored medical image samples, completed tubular structure mask samples, and the corresponding completed tubular structure mask gold standards from the cloud, cloud storage, hardware, or local location.

[0126] S440. The medical image sample and the completed tubular structure mask sample are sequentially input into the second downsampling encoder and the second upsampling decoder to obtain the prediction error information.

[0127] Specifically, the error learning network can be trained using medical image samples, completed tubular structure mask samples, and the corresponding gold standard of completed tubular structure masks to obtain the error learning model. During the training process, the medical image samples and completed tubular structure mask samples from the second training sample set can be sequentially input into the second downsampling encoder and the second upsampling decoder in the error learning network to obtain the prediction error information output by the second upsampling decoder.

[0128] It should be noted that the medical image samples and the completed tubular structure mask samples are the same size. If the medical image samples and the completed tubular structure mask samples in the second training sample set are different sizes, the medical image samples and the completed tubular structure mask samples can be cropped into images of the same size before being input into the second downsampling encoder and the second upsampling decoder in sequence.

[0129] S450. Based on the difference between the prediction error information and the error information gold standard, adjust the network parameters of the second downsampling encoder and the second upsampling decoder until convergence is achieved, thus obtaining the error learning model. The error information gold standard is used to characterize the difference between the completed tubular structure mask sample and the completed tubular structure mask gold standard.

[0130] Based on the prediction error information obtained in the above steps, the difference between the prediction error information and the gold standard error information can be calculated using the third loss function. The network parameters of the second downsampling encoder and the second upsampling decoder are then adjusted according to this difference. The steps S440-S450 are then executed again until the error learning network converges, resulting in a trained error learning model, i.e., a trained second downsampling encoder and a trained second upsampling decoder. The third loss function can be the same as or different from the second and first loss functions; this embodiment does not limit this. In this embodiment, the third loss function is the mean squared loss function.

[0131] It should be noted that before training the error learning network, the network parameters of the second downsampling encoder and the second upsampling decoder in the error learning network can be obtained through initialization. The convergence condition of the above error learning network can be that the difference between the predicted error information and the gold standard error information is less than a second preset difference threshold, or that the number of iterations of S440-S450 is greater than or equal to a second preset iteration number threshold.

[0132] Wherein, the second preset difference threshold can be a positive number close to 0, and the second preset iteration number threshold can be a value greater than 0. In the embodiments of this application, the second preset difference threshold and the first preset difference threshold can be equal or unequal, and the second preset iteration number threshold and the first preset iteration number threshold can be equal or unequal.

[0133] For example, the following uses the tubular structure completion network and the error learning network as examples from the VBNet network model to illustrate the two network structures, such as... Figure 11 The diagram shows the overall network structure of the tubular structure completion network and the error learning network. During training, tubular structure mask samples M can be input into the tubular structure completion network. This network includes a first downsampling encoder D1 (4 downsampling layers and 1 convolutional layer), a latent distribution space learning layer R, and a first upsampling decoder U1 (4 upsampling layers and 1 convolutional layer). The error learning network includes a second downsampling encoder D2 (4 downsampling layers and 1 convolutional layer) and a second upsampling decoder U2 (4 upsampling layers and 1 convolutional layer). Figure 11 In this diagram, BD1 represents the initial predicted distribution, BD2 represents the ideal distribution, L1 represents the first loss function, L2 represents the second loss function, L3 represents the third loss function, E' represents the prediction error information, E represents the gold standard for error information, M' represents the predicted completion of the tubular structure mask, Imag is the medical image sample, G is the gold standard for the completion of the tubular structure mask, and M” represents the target tubular structure mask. Figure 10 In this context, Mean represents the process of calculating the average, and Std represents the process of calculating the variance.

[0134] Figure 11 The first downsampling encoder D1 in the middle includes five rectangles, each rectangle representing a network layer. The first layer on the left represents a convolutional layer, and the four layers on the right represent four downsampling layers. Figure 11 The first upsampling encoder U1 consists of five rectangles, each representing a network layer. The four layers on the left represent upsampling layers, and the one layer on the right represents a convolutional layer. Correspondingly, Figure 11 The positions of the sampling layers and convolutional layers in the second downsampling encoder D2 and the second upsampling decoder U2 are also of the same type, which will not be elaborated further. Additionally, Figure 11 In this context, R, P, L, S, and I all represent the directions of different positions on the tubular structure based on the sagittal plane of the target object. R represents right, P represents back, L represents left, S represents up, and I represents down.

[0135] In this example, the ideal distribution can be set as a standard normal distribution. During training, the first loss function is set as the KL divergence loss function as a constraint between the predicted initial distribution and the ideal distribution. By adjusting the network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder, the predicted initial distribution is forced to approach the ideal distribution. However, in actual model training, in order to obtain better network parameters, it is preferable to set the first loss function as the KL divergence plus cross-entropy loss function during training, and set the minimization of this loss function as the optimization objective, α as the adjustment parameter, and use the Adam parameter optimization method to achieve network parameter optimization, that is, iteratively adjust the network parameters.

[0136] The tubular structure fracture repair method in this embodiment can train an error learning network to obtain an optimal error learning model. This allows the method to further detect fracture repair errors on the tubular structure mask using the optimal error learning model. Based on the error detection results, i.e., error information, the tubular structure mask can be repaired. This method performs fracture repair on the tubular structure through a dual repair process, which can greatly improve the accuracy and completeness of the tubular structure fracture repair results.

[0137] In one embodiment, this application also provides a method for repairing a fractured tubular structure, the method comprising the following steps:

[0138] (1) Obtain medical images of the target object, including tubular structures.

[0139] (2) The tubular structures in the medical image are segmented to obtain the initial tubular structure mask.

[0140] (3) The initial tubular structure mask is divided into blocks to obtain multiple image blocks.

[0141] (4) Input multiple image blocks into the first downsampling encoder in the tubular structure completion model to obtain multiple first features corresponding to multiple image blocks; the first feature is the result obtained by the first downsampling encoder performing downsampling processing on the initial tubular structure mask.

[0142] (5) Input multiple first features into the learning layer of the latent distribution space.

[0143] (6) The mean vector and variance vector are obtained by first calculating the mean and variance of each first feature through the latent distribution space learning layer. Then, the initial distribution is obtained based on the mean vector and variance vector. Finally, the initial distribution is sampled to obtain the second feature. The second feature is the result of the latent distribution space learning layer performing operation and sampling on the first feature.

[0144] (7) Input the second feature into the first upsampling decoder in the tubular structure completion model to obtain the completed tubular structure mask; the completed tubular structure mask is the result of the first upsampling decoder upsampling the second feature.

[0145] The training process for the tubular structure completion model includes:

[0146] (8) Obtain the first training sample set, which includes tubular structure mask samples and corresponding complete tubular structure mask gold standards.

[0147] (9) Input the tubular mask samples sequentially into the first downsampling encoder and the latent distribution space learning layer in the tubular completion network to obtain the sample features after sampling the predicted initial distribution.

[0148] (10) Input the sample features into the first upsampling decoder to obtain the predicted complete tubular structure mask.

[0149] (11) Based on the first difference between the initial predicted distribution and the corresponding ideal distribution, and the second difference between the predicted complete tubular structure mask and the gold standard of the complete tubular structure mask, adjust the network parameters of the first downsampling encoder, the latent distribution space learning layer and the first upsampling decoder until convergence, and obtain the tubular structure completion model.

[0150] (12) Input the completed tubular structure mask and the medical image into the second downsampling encoder in the error learning model to obtain the third feature; the third feature is the result obtained by the second downsampling encoder through downsampling processing of the completed tubular structure mask and the medical image.

[0151] (13) Input the third feature into the second upsampling decoder in the error learning model to obtain error information; the error information is the result of the second upsampling decoder upsampling the third feature; the error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

[0152] The training process of the error learning model includes:

[0153] (14) Obtain the second training sample set; the second training sample set includes medical image samples, complete tubular structure mask samples and corresponding complete tubular structure mask gold standards.

[0154] (15) Input the medical image samples and the completed tubular structure mask samples into the second downsampling encoder and the second upsampling decoder in the error learning network in sequence to obtain the prediction error information.

[0155] (16) Based on the difference between the prediction error information and the error information gold standard, adjust the network parameters of the second downsampling encoder and the second upsampling decoder until convergence is achieved to obtain the error learning model; the error information gold standard is used to characterize the difference between the completed tubular structure mask sample and the completed tubular structure mask gold standard.

[0156] (17) The completed tubular structure mask is superimposed with the error information to obtain the target tubular structure mask.

[0157] The specific execution process of (1) to (17) above can be found in the description of the above embodiments. The implementation principle and technical effect are similar, and will not be repeated here.

[0158] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0159] Based on the same inventive concept, this application also provides a tubular structure fracture repair device for implementing the tubular structure fracture repair method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more tubular structure fracture repair device embodiments provided below can be found in the limitations of the tubular structure fracture repair method described above, and will not be repeated here.

[0160] In one embodiment, Figure 12 This is a schematic diagram of a tubular structure fracture repair device according to one embodiment of this application. The tubular structure fracture repair device provided in this embodiment can be applied to computer equipment. Figure 12 As shown, the tubular structure fracture repair device of this application embodiment may include: a medical image acquisition module 11, a segmentation processing module 12, a repair processing module 13, and a repair result repair module 14.

[0161] Among them, the medical image acquisition module 11 is used to acquire medical images of the target object, including tubular structures;

[0162] The segmentation processing module 12 is used to segment tubular structures in medical images to obtain an initial tubular mask;

[0163] The completion processing module 13 is used to input the initial tubular mask into the structural completion model to complete the missing parts of the tubular structure in the initial tubular mask and obtain the completed tubular structure mask.

[0164] The completion result repair module 14 is used to input the completed tubular structure mask and medical image into the error learning model to obtain error information, and to obtain the target tubular structure mask based on the completed tubular structure mask and the error information; the error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

[0165] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0166] In one embodiment, the tubular structure completion model includes: a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder; the completion processing module 13 includes: a first downsampling processing unit, an arithmetic processing unit, and a first upsampling processing unit.

[0167] The first downsampling processing unit is used to input the initial tubular structure mask into the first downsampling encoder to obtain the first feature; the first feature is the result obtained by the first downsampling encoder through downsampling processing of the initial tubular structure mask;

[0168] The computational processing unit is used to input the first feature into the latent distribution space learning layer, first obtain the mean and standard deviation distribution of the first feature, and then sample the mean and standard deviation distribution to obtain the second feature;

[0169] The first upsampling processing unit is used to input the second feature into the first upsampling decoder to obtain a completed tubular structure mask; the completed tubular structure mask is the result obtained by the first upsampling decoder through upsampling processing of the second feature.

[0170] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0171] In one embodiment, the first downsampling processing unit is specifically used for:

[0172] The initial tubular structure mask is divided into blocks to obtain multiple image blocks;

[0173] Multiple image blocks are input into the first downsampling encoder to obtain multiple first features corresponding to the multiple image blocks.

[0174] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0175] In one embodiment, the first upsampling processing unit is specifically used for:

[0176] Multiple primary features are input into the latent distribution space learning layer;

[0177] The mean and variance vectors are obtained by performing mean and variance calculations on each first feature through the latent distribution space learning layer. The initial distribution is obtained based on the mean and variance vectors, and the second feature is obtained by sampling the initial distribution.

[0178] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0179] In one embodiment, the tubular structure fracture completion device further includes a completion network training module.

[0180] Specifically, the network completion training module is used for:

[0181] Obtain the first training sample set, which includes tubular structure mask samples and corresponding gold standards for completing tubular structure masks;

[0182] The tubular mask sample is sequentially input into the first downsampling encoder and the latent distribution space learning layer to obtain the sample features after sampling the predicted initial distribution.

[0183] The sample features are input into the first upsampling decoder to obtain the predicted and completed tubular structure mask;

[0184] Based on the first difference between the initial predicted distribution and the corresponding ideal distribution, and the second difference between the predicted completed tubular structure mask and the gold standard of the completed tubular structure mask, the network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder are adjusted until convergence is achieved, thus obtaining the tubular structure completion model.

[0185] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0186] In one embodiment, the error learning model includes: a second downsampling encoder and a second upsampling decoder; the completion result repair module 14 includes: a second downsampling processing unit and a second upsampling processing unit.

[0187] The second downsampling processing unit is used to input the completed tubular structure mask and the medical image into the second downsampling encoder to obtain the third feature; the third feature is the result obtained by the second downsampling encoder through downsampling processing of the completed tubular structure mask and the medical image;

[0188] The second upsampling processing unit is used to input the third feature into the second upsampling decoder to obtain error information; the error information is the result obtained by the second upsampling decoder after upsampling the third feature.

[0189] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0190] In one embodiment, the tubular structure fracture repair device further includes an error learning network training module.

[0191] Specifically, the error learning network training module is used for:

[0192] Obtain the second training sample set; the second training sample set includes medical image samples, incomplete tubular structure mask samples, and the corresponding incomplete tubular structure mask gold standard;

[0193] The medical image sample and the completed tubular structure mask sample are sequentially input into the second downsampling encoder and the second upsampling decoder to obtain the prediction error information.

[0194] Based on the difference between the predicted error information and the ideal error information, the network parameters of the second downsampling encoder and the second upsampling decoder are adjusted until convergence, thus obtaining the error learning model; the ideal error information is used to characterize the difference between the completed tubular structure mask sample and the completed tubular structure mask gold standard.

[0195] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0196] In one embodiment, the completion result repair module 14 is specifically used for:

[0197] The target tubular structure mask is obtained by superimposing the completed tubular structure mask with the error information.

[0198] The tubular structure fracture repair device provided in this application embodiment can be used to execute the technical solution in the above-described tubular structure fracture repair method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0199] Specific limitations regarding the tubular fracture repair device can be found in the limitations of the medical image processing method described above, and will not be repeated here. Each module in the aforementioned tubular fracture repair device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0200] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may continue as follows: Figure 1As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores medical images of the target object. The network interface communicates with external endpoints via a network connection. When executed by the processor, the computer program implements a method for repairing fractures in a tubular structure.

[0201] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0202] In one embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the technical solution in the above-described embodiment of the tubular structure fracture repair method of this application. The implementation principle and technical effect are similar, and will not be repeated here.

[0203] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the technical solution of the tubular structure fracture repair method described above in this application. The implementation principle and technical effect are similar, and will not be repeated here.

[0204] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the technical solution of the tubular structure fracture repair method described above in this application. The implementation principle and technical effect are similar and will not be repeated here.

[0205] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0206] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0207] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for repairing fractures in a tubular structure, characterized in that, The method includes: Acquire a medical image of the target object, the medical image including a tubular structure; The tubular structures in the medical image are segmented to obtain an initial tubular structure mask; The initial tubular structure mask is input into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, thereby obtaining the completed tubular structure mask; the tubular structure completion model includes a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder; The completed tubular structure mask and the medical image are input into the error learning model to obtain error information, and the target tubular structure mask is obtained based on the completed tubular structure mask and the error information; the error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

2. The method according to claim 1, characterized in that, The step of inputting the initial tubular structure mask into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular structure mask, thereby obtaining a completed tubular structure mask, includes: The initial tubular structure mask is input to the first downsampling encoder to obtain a first feature; the first feature is the result obtained by the first downsampling encoder through downsampling the initial tubular structure mask; The first feature is input into the latent distribution space learning layer to obtain the mean and standard deviation distribution of the first feature. Then, the mean and standard deviation distribution are sampled to obtain the second feature. The second feature is input to the first upsampling decoder to obtain the completed tubular structure mask; the completed tubular structure mask is the result of the first upsampling decoder upsampling the second feature.

3. The method according to claim 2, characterized in that, The step of inputting the initial tubular structure mask into the first downsampling encoder to obtain the first feature includes: The initial tubular structure mask is divided into blocks to obtain multiple image blocks; The plurality of image blocks are input into the first downsampling encoder to obtain a plurality of first features corresponding to the plurality of image blocks.

4. The method according to claim 3, characterized in that, The step of inputting the first feature into the latent distribution space learning layer to obtain the second feature includes: Multiple first features are input into the latent distribution space learning layer; The mean vector and variance vector are obtained by performing mean and variance calculation on each of the first features through the latent distribution space learning layer. The initial distribution is obtained based on the mean vector and the variance vector, and the second feature is obtained by sampling the initial distribution.

5. The method according to claim 2, characterized in that, The training process of the tubular structure completion model includes: Obtain a first training sample set, which includes tubular structure mask samples and corresponding gold standards for completing tubular structure masks; The tubular mask sample is sequentially input into the first downsampling encoder and the latent distribution space learning layer to obtain the sample features after sampling the predicted initial distribution. The sample features are input into the first upsampling decoder to obtain the predicted and completed tubular structure mask; Based on the first difference between the initial predicted distribution and the corresponding ideal distribution, and the second difference between the predicted completed tubular structure mask and the gold standard of the completed tubular structure mask, the network parameters of the first downsampling encoder, the latent distribution space learning layer, and the first upsampling decoder are adjusted until convergence is achieved, thus obtaining the tubular structure completion model.

6. The method according to any one of claims 1-5, characterized in that, The error learning model includes: a second downsampling encoder and a second upsampling decoder; the step of inputting the completed tubular structure mask and the medical image into the error learning network model to obtain error information includes: The completed tubular structure mask and the medical image are input into the second downsampling encoder to obtain a third feature; the third feature is the result obtained by the second downsampling encoder through downsampling processing of the completed tubular structure mask and the medical image; The third feature is input to the second upsampling decoder to obtain the error information; the error information is the result obtained by the second upsampling decoder after upsampling the third feature.

7. The method according to claim 6, characterized in that, The training process of the error learning model includes: Obtain a second training sample set; the second training sample set includes medical image samples, incomplete tubular structure mask samples, and corresponding incomplete tubular structure mask gold standards; The medical image sample and the completed tubular structure mask sample are sequentially input into the second downsampling encoder and the second upsampling decoder to obtain prediction error information; Based on the difference between the predicted error information and the ideal error information, the network parameters of the second downsampling encoder and the second upsampling decoder are adjusted until convergence is achieved, thus obtaining the error learning model; the ideal error information is used to characterize the difference between the completed tubular structure mask sample and the completed tubular structure mask gold standard.

8. The method according to any one of claims 1-5, characterized in that, The process of obtaining the target tubular structure mask based on the completed tubular structure mask and the error information includes: The target tubular structure mask is obtained by superimposing the completed tubular structure mask with the error information.

9. A device for repairing fractures in a tubular structure, characterized in that, The device includes: A medical image acquisition module is used to acquire a medical image of a target object, wherein the medical image includes a tubular structure; The segmentation processing module is used to segment the tubular structures in the medical image to obtain an initial tubular mask; The completion processing module is used to input the initial tubular mask into the tubular structure completion model to complete the missing parts of the tubular structure in the initial tubular mask, thereby obtaining a completed tubular structure mask; the tubular structure completion model includes: a first downsampling encoder, a latent distribution space learning layer, and a first upsampling decoder; The completion result repair module is used to input the completed tubular structure mask and the medical image into the error learning model to obtain error information, and to obtain the target tubular structure mask based on the completed tubular structure mask and the error information; the error information is used to characterize the difference between the tubular structure in the completed tubular structure mask and the tubular structure in the gold standard of the completed tubular structure mask.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-8.

Images