Image processing method and device, medical image processing method and device

Abstract

Embodiments of the present specification provide image processing methods and devices, medical image processing methods and devices, wherein the medical image processing method comprises: acquiring a medical image containing vertebral bodies, and performing image feature extraction on the medical image; determining an initial center point, an adjustment vector and an initial category of each vertebral body based on the image features; adjusting the initial center point of each vertebral body according to the adjustment vector to obtain a target center point of each vertebral body and a connection relationship of the target center points; and adjusting the initial category of each vertebral body based on the connection relationship of the target center points to obtain a target category and a target label of each vertebral body.

Description

TECHNICAL FIELD

[0001] Embodiments of the present specification relate to the technical field of image processing, in particular to an image processing method. One or more embodiments of the present specification also relate to an image processing apparatus, a medical image processing method, a medical image processing apparatus, a computing device, and a computer-readable storage medium. BACKGROUND

[0002] In the context of computer-aided spine surgery, vertebra localization, segmentation and segmentation are important components of spine automated analysis. Due to difficulties such as similar appearance, limited field of view, pathological lesions, etc., previous methods usually solve the localization, segmentation and segmentation tasks separately to achieve high accuracy; but in this way, the entire analysis process is particularly complex, and existing spine automated analysis models mainly rely on spine prior information (such as statistical shape, appearance and results) to segment vertebrae, which is extremely dependent on prior information, limiting the robustness of the model.

[0003] Therefore, there is an urgent need to provide a medical image processing method that can unify localization, segmentation and segmentation into one network to simply and quickly identify each vertebra of the spine. SUMMARY

[0004] Therefore, the embodiments of the present specification provide an image processing method. One or more embodiments of the present specification also relate to an image processing apparatus, a medical image processing method, a medical image processing apparatus, a computing device, and a computer-readable storage medium to solve the technical defects existing in the prior art.

[0005] According to a first aspect of the embodiments of the present specification, an image processing method is provided, comprising:

[0006] An image containing a predetermined shape target object is obtained, and image features of the image are extracted;

[0007] Based on the image features, an initial center point, an adjustment vector and an initial class of each predetermined shape target object are determined;

[0008] According to the adjustment vector, the initial center point of each predetermined shape target object is adjusted to obtain a target center point of each predetermined shape target object and a connection relationship of the target center point;

[0009] Based on the connection relationship of the target center point, the initial class of each predetermined shape target object is adjusted to obtain a target class and a target label of each predetermined shape target object.

[0010] According to a second aspect of the embodiments of the present specification, a medical image processing method is provided, comprising:

[0011] obtaining a medical image containing vertebrae, and performing image feature extraction on the medical image;

[0012] determining an initial center point, an adjustment vector, and an initial category of each vertebra based on the image feature;

[0013] adjusting the initial center point of each vertebra according to the adjustment vector to obtain a target center point of each vertebra and a connection relationship of the target center point;

[0014] adjusting the initial category of each vertebra based on the connection relationship of the target center point to obtain a target category and a target label of each vertebra.

[0015] According to a third aspect of the embodiments of the present specification, an image processing apparatus is provided, comprising:

[0016] a first image obtaining module configured to obtain an image containing a predetermined shape target object, and perform image feature extraction on the image;

[0017] a first determining module configured to determine an initial center point, an adjustment vector, and an initial category of each predetermined shape target object based on the image feature;

[0018] a first adjusting module configured to adjust the initial center point of each predetermined shape target object according to the adjustment vector to obtain a target center point of each predetermined shape target object and a connection relationship of the target center point;

[0019] a first obtaining module configured to adjust the initial category of each predetermined shape target object based on the connection relationship of the target center point to obtain a target category and a target label of each predetermined shape target object.

[0020] According to a fourth aspect of the embodiments of the present specification, a medical image processing apparatus is provided, comprising:

[0021] a second image obtaining module configured to obtain a medical image containing vertebrae, and perform image feature extraction on the medical image;

[0022] a second determining module configured to determine an initial center point, an adjustment vector, and an initial category of each vertebra based on the image feature;

[0023] a second adjusting module configured to adjust the initial center point of each vertebra according to the adjustment vector to obtain a target center point of each vertebra and a connection relationship of the target center point;

[0024] The second obtaining module is configured to adjust the initial category of each vertebral body based on the connection relationship of the target center points, to obtain a target category and a target label of each vertebral body.

[0025] According to a fifth aspect of an embodiment of the present specification, a medical image processing method is provided, comprising:

[0026] an interface display module configured to display an image input interface for a user based on a calling request of the user;

[0027] an image receiving module configured to receive a medical image containing vertebral bodies input by the user based on the image input interface, and perform image feature extraction on the medical image;

[0028] a third determining module configured to determine an initial center point, an adjustment vector, and an initial category of each vertebral body based on the image features;

[0029] a third adjusting module configured to adjust the initial center point of each vertebral body according to the adjustment vector, to obtain a target center point of each vertebral body and a connection relationship of the target center points;

[0030] a second adjusting module configured to adjust the initial category of each vertebral body based on the connection relationship of the target center points, to obtain a target category and a target label of each vertebral body;

[0031] a second determining module configured to determine a binarization result of each vertebral body based on the image features, and perform segmentation on each vertebral body based on the binarization result of each vertebral body, the target category, the target label, and the target center point, and return the segmented medical image to the user.

[0032] According to a sixth aspect of an embodiment of the present specification, a medical image processing device is provided, comprising:

[0033] an interface display module configured to display an image input interface for a user based on a calling request of the user;

[0034] an image receiving module configured to receive a medical image containing vertebral bodies input by the user based on the image input interface, and perform image feature extraction on the medical image;

[0035] a third determining module configured to determine an initial center point, an adjustment vector, and an initial category of each vertebral body based on the image features;

[0036] a third adjusting module configured to adjust the initial center point of each vertebral body according to the adjustment vector, to obtain a target center point of each vertebral body and a connection relationship of the target center points;

[0037] a third obtaining module configured to adjust the initial class of each vertebral body based on the connection relationship of the target center point to obtain a target class and a target label of each vertebral body;

[0038] a first segmentation module configured to determine a binarization result of each vertebral body according to the image features, and segment each vertebral body based on the binarization result of each vertebral body, the target class, the target label and the target center point, and return the segmented medical image to the user.

[0039] According to a seventh aspect of an embodiment of the present specification, a medical image processing method is provided, comprising:

[0040] receiving a calling request sent by a user and carrying a medical image containing a vertebral body, and performing image feature extraction on the medical image;

[0041] determining an initial center point, an adjustment vector and an initial class of each vertebral body based on the image features;

[0042] adjusting the initial center point of each vertebral body according to the adjustment vector to obtain a target center point of each vertebral body and a connection relationship of the target center point;

[0043] adjusting the initial class of each vertebral body based on the connection relationship of the target center point to obtain a target class and a target label of each vertebral body;

[0044] determining a binarization result of each vertebral body according to the image features, and segmenting each vertebral body based on the binarization result of each vertebral body, the target class, the target label and the target center point, and returning the segmented medical image to the user.

[0045] According to an eighth aspect of an embodiment of the present specification, a medical image processing device is provided, comprising:

[0046] a request receiving module configured to receive a calling request sent by a user and carrying a medical image containing a vertebral body, and perform image feature extraction on the medical image;

[0047] a fourth determining module configured to determine an initial center point, an adjustment vector and an initial class of each vertebral body based on the image features;

[0048] a fourth adjusting module configured to adjust the initial center point of each vertebral body according to the adjustment vector to obtain a target center point of each vertebral body and a connection relationship of the target center point;

[0049] a fourth obtaining module, configured to adjust the initial class of each vertebral body based on the connection relationship of the target center point, to obtain a target class and a target label of each vertebral body;

[0050] a second segmentation module, configured to determine a binarization result of each vertebral body according to the image feature, and segment each vertebral body based on the binarization result of each vertebral body, the target class, the target label and the target center point, and return the segmented medical image to the user.

[0051] According to a ninth aspect of an embodiment of the present specification, a computing device is provided, comprising:

[0052] a memory and a processor;

[0053] The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, and when the processor executes the computer executable instructions, the steps of the image processing method or the medical image processing method are implemented.

[0054] According to a tenth aspect of an embodiment of the present specification, a computer readable storage medium is provided, which stores computer executable instructions, and when the instructions are executed by a processor, the steps of the image processing method or the medical image processing method are implemented.

[0055] An embodiment of the present specification implements an image processing method and device, and a medical image processing method and device. The medical image processing method comprises: acquiring a medical image containing vertebral bodies, and performing image feature extraction on the medical image; determining an initial center point, an adjustment vector and an initial class of each vertebral body based on the image feature; adjusting the initial center point of each vertebral body according to the adjustment vector, to obtain a target center point of each vertebral body and a connection relationship of the target center point; and adjusting the initial class of each vertebral body based on the connection relationship of the target center point, to obtain a target class and a target label of each vertebral body.

[0056] The medical image processing method acquires the target center point, the adjustment vector and the initial class of the vertebral body through the extracted image feature, then performs fine adjustment of the target center point position of the vertebral body and reconstruction of the class relationship of the vertebral body through the adjustment vector, to obtain accurate positioning of the target center of the vertebral body, and then obtains the segmentation result of the vertebral body by using the reconstructed class relationship of the vertebral body, and further obtains the target class and the target label of the vertebral body, and finally greatly improves the classification accuracy of the vertebral body through the target class and the target label of the vertebral body. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1is a specific schematic diagram of a medical image processing method applied to vertebra recognition provided by one embodiment of the present specification;

[0058] Figure 2 is a flowchart of a medical image processing method provided by one embodiment of the present specification.

[0059] Figure 3 is a flowchart of a first medical image processing method provided by one embodiment of the present specification;

[0060] Figure 4 is a chain structure schematic diagram of a target center point of a vertebra in a medical image processing method provided by one embodiment of the present specification;

[0061] Figure 5 is a processing process schematic diagram of a medical image processing method provided by one embodiment of the present specification;

[0062] Figure 6 is a specific vertebra segmentation schematic diagram in a medical image processing method provided by one embodiment of the present specification;

[0063] Figure 7 is a structure schematic diagram of a medical image processing device provided by one embodiment of the present specification.

[0064] Figure 8 is a structure schematic diagram of a first medical image processing device provided by one embodiment of the present specification;

[0065] Figure 9 is a flowchart of a second medical image processing method provided by one embodiment of the present specification;

[0066] Figure 10 is a structure schematic diagram of a second medical image processing device provided by one embodiment of the present specification;

[0067] Figure 11 is a flowchart of a third medical image processing method provided by one embodiment of the present specification;

[0068] Figure 12 is a structure schematic diagram of a third medical image processing device provided by one embodiment of the present specification;

[0069] Figure 13 is a structure block diagram of a computing device provided by one embodiment of the present specification. DETAILED DESCRIPTION

[0070] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. However, the present description can be practiced without the specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present description.

[0071] The terminology used in this description is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the present description. As used in this description and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0072] It will be understood that, although the terms first, second, etc. can be used herein to describe various information, these terms are not intended to denote a temporal or chronological order. Rather, these terms are used solely to distinguish one from another only. For example, without departing from the scope of one or more embodiments, first can be termed second, and similarly, second can be termed first. The term "if can be construed to mean "when" or "upon" or "in response to determining" depending on the context, as used herein.

[0073] First, the noun terms related to one or more embodiments of the present description are explained.

[0074] Spine: composed of a plurality of sequentially connected vertebral bodies.

[0075] Positioning: requires positioning the center point of each vertebral body.

[0076] Segmentation: vertebral bodies are classified from top to bottom as C1-C7 (cervical vertebrae), T1-T12 (thoracic vertebrae), L1-L5 (lumbar vertebrae), S1-S2 (sacral vertebrae) for a total of 26, and segmentation requires giving the classification corresponding to each vertebral body.

[0077] Segmentation: each vertebral body is separated in a 3D CT (Computed Tomography) image.

[0078] Vnet: virtual network, a three-dimensional image segmentation network based on volume, based on FCN (Fully Convolutional Networks).

[0079] In the present specification, an image processing method is provided. One or more embodiments of the present specification also relate to an image processing apparatus, a medical image processing method, a medical image processing apparatus, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments.

[0080] Referring to Figure 1 , Figure 1 A specific schematic diagram of a medical image processing method for identifying a vertebral body is shown.

[0081] Figure 1 The application scenario includes an image acquisition terminal 102 and a server 104. Specifically, the image acquisition terminal 102 acquires a 3D CT image containing a vertebral body, and then sends the CT image to the server 104. After receiving the CT image, the server 104 first inputs the CT image into a Vnet to obtain the basic image features of the CT image, and then inputs the basic image features into a multi-layer convolution layer for convolution to obtain the convolutional basic image features. Then, the convolutional basic image features are respectively input into a first feature extractor, a second feature extractor, a third feature extractor, a fourth feature extractor, and a fifth feature extractor for target image feature extraction to obtain a first target image feature, a second target image feature, a third target image feature, a fourth target image feature, and a fifth target image feature. The first feature extractor, the second feature extractor, the third feature extractor, the fourth feature extractor, and the fifth feature extractor are different in their corresponding tasks, and the features extracted by them are also different. However, in actual applications, since the output results of the heat map network, the relationship vector field network, the offset vector field network, and the vertebral body recognition network are all related to the center point of the vertebral body, the heat map network, the relationship vector field network, the offset vector field network, and the vertebral body recognition network can share the features of one branch, i.e., the first target image feature, the second target image feature, the third target image feature, or the fourth target image feature.

[0082] input the first target image feature into the heat map network to obtain an initial center point of the vertebral body in the CT image; input the second target image feature into the relationship vector field network to obtain a relationship vector of the vertebral body in the CT image; input the third target image feature into the offset vector field network to obtain an offset vector of the vertebral body in the CT image; input the fourth target image feature into the vertebral body recognition network to obtain an initial category and an initial segmentation result of the vertebral body in the CT image; and input the fifth target image feature into the binary segmentation network to obtain a binary image of the vertebral body in the CT image. The network structures of the heat map network, the relationship vector field network, the offset vector field network, the vertebral body recognition network, and the binary segmentation network can be the same as that of the Vnet, but due to fewer network layers and smaller volume, in actual application, they can also be referred to as heads, for example, the heat map network is a heat map head, the relationship vector field network is a relationship vector field head, the offset vector field network is an offset vector field head, the vertebral body recognition network is a vertebral body recognition head, and the binary segmentation network is a binary segmentation head.

[0083] After obtaining the initial center point, the relationship vector, the offset vector, the initial category, and the binary image of the vertebral body in the CT image, the position of the initial center point of the vertebral body in the CT image is fine-tuned and the category relationship is reconstructed through the relationship vector and the offset vector, to obtain a precise positioning of the vertebral body in the CT image, i.e., a target center point, then a segmentation result of the target center point of each vertebral body in the CT image is further obtained by using a maximum likelihood algorithm, and finally, the foreground pixel points of the vertebral body in the binary image are assigned to the target center point of a specific vertebral body according to the offset vector, thereby generating a segmentation map of each vertebral body in the CT image. Through this mode of unifying the positioning, recognition, and segmentation of the vertebral body into one network, the segmentation of each vertebral body in the CT image can be simply and clearly realized.

[0084] In actual application, the image processing method provided by the embodiments of the present specification can not only be applied to the medical image field of identifying and / or segmenting the vertebral body, but also be applied to the process of identifying and / or segmenting other organs in the medical image in the medical field, such as identifying blood vessels, identifying hand bones and foot bones, and identifying thoracic cavities, oral cavities, and head cavities, etc. In addition, the image processing method can also be applied to any non-medical field, for example, it can be applied to some pipeline fields, and based on the image processing method, the identification of a section of pipeline can be realized.

[0085] Referring to Figure 2 , Figure 2 A flowchart of a medical image processing method according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0086] Step 202: An image containing a target object of a predetermined shape is obtained, and image feature extraction is performed on the image.

[0087] The predetermined shape target object includes, but is not limited to, a target object of a specific shape in any field, for example, a pipe of a special shape in the pipe field, or an irregular skull or hand bone in the medical field.

[0088] In the implementation, after obtaining the image containing the predetermined shape target object, the image feature extraction is performed on the image, and the subsequent image processing is performed based on the image features of the image, and the specific implementation manner is as follows:

[0089] The image feature extraction on the image includes:

[0090] The initial image feature extraction is performed on the image based on the feature extraction model, and the initial image feature is obtained.

[0091] The initial image feature is subjected to the convolution processing through the multi-layer convolution layer, and the image feature after the convolution processing is obtained.

[0092] The image feature after the convolution processing is subjected to the re-extraction according to a preset task, and the first image feature and the second image feature are obtained, and the preset task includes a first task, a second task and a third task.

[0093] Step 204: determining the initial center point, the adjustment vector and the initial category of each predetermined shape target object based on the image feature.

[0094] Step 206: adjusting the initial center point of each predetermined shape target object according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center point.

[0095] Step 208: adjusting the initial category of each predetermined shape target object based on the connection relationship of the target center point to obtain the target category and the target label of each predetermined shape target object.

[0096] Optionally, after the target category and the target label of each predetermined shape target object are obtained, the method further includes:

[0097] determining the binarization result of each predetermined shape target object according to the image feature, and segmenting each predetermined shape target object based on the binarization result, the target category, the target label and the target center point of each predetermined shape target object.

[0098] Optionally, the first task is implemented by a first network and a second network, and the second task is implemented by a third network.

[0099] Correspondingly, the determining the initial center point, the adjustment vector and the initial category of each predetermined shape target object based on the image features comprises:

[0100] inputting the first image features into the first network to obtain the initial center point of each predetermined shape target object;

[0101] inputting the first image features into the second network to obtain the adjustment vector of each predetermined shape target object; and

[0102] inputting the first image features into the third network to obtain the initial category of each predetermined shape target object.

[0103] Optionally, the inputting the first image features into the first network to obtain the initial center point of each predetermined shape target object comprises:

[0104] inputting the first image features into the first network to obtain the probability of each pixel in each predetermined shape target object being the initial center point, and determining the initial center point of each predetermined shape target object according to the probability.

[0105] Optionally, after the determining the initial center point of each predetermined shape target object according to the probability, the method further comprises:

[0106] determining the initial center point region of each predetermined shape target object based on each pixel, the initial center point and a preset radius in each predetermined shape target object.

[0107] Optionally, the inputting the first image features into the second network to obtain the adjustment vector of each predetermined shape target object comprises:

[0108] inputting the first image features into a relationship vector network to obtain the relationship vector of the initial center points of two adjacent predetermined shape target objects; and

[0109] inputting the first image features into an offset vector network to obtain the offset vector of each pixel in the initial center point region of each predetermined shape target object corresponding to the initial center point of each predetermined shape target object.

[0110] Optionally, the adjusting the initial center point of each predetermined shape target object according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center points comprises:

[0111] inputting the value of each pixel in the initial center region of each predetermined shape target object and the offset vector into a pre-designed calculation formula to obtain a probability of each pixel in the initial center region of each predetermined shape target object as a center point, and taking the pixel with the maximum probability as a candidate center point of each predetermined shape target object;

[0112] obtaining a target center point of each predetermined shape target object and a connection relationship of the target center point based on the candidate center point of each predetermined shape target object and a relationship vector of initial center points of two adjacent predetermined shape target objects.

[0113] Optionally, the adjusting the initial category of each predetermined shape target object based on the connection relationship of the target center point comprises:

[0114] adjusting the initial category of each predetermined shape target object based on the connection relationship of the target center point by using a maximum likelihood algorithm to obtain a target category and a target label of each predetermined shape target object.

[0115] Optionally, the third task is implemented by a fourth network.

[0116] Correspondingly, the determining the binarization result of each predetermined shape target object according to the image feature comprises:

[0117] inputting the second image feature into the fourth network to obtain the binarization result of each predetermined shape target object.

[0118] Optionally, the segmenting each predetermined shape target object based on the binarization result, the target category, the target label and the target center point of each predetermined shape target object comprises:

[0119] obtaining the coordinates of each pixel of each predetermined shape target object after binarization, and determining the nearest target center point of each pixel based on the coordinates of each pixel, the offset vector and the target center point of each predetermined shape target object;

[0120] determining the target category and the target label of each pixel based on the target category and the target label of the nearest target center point of each pixel, and segmenting each predetermined shape target object according to the target category and the target label of each pixel.

[0121] It should be noted that the part of the image processing method provided by the embodiments of the present specification corresponding to the embodiments of the first medical image processing method of the following embodiments can be referred to the detailed description in the embodiments of the first medical image processing method, which will not be repeated here.

[0122] The image processing method provided by the embodiments of the present specification extracts the image features to obtain the target center point, the adjustment vector, the initial category and the binarization result of the predetermined shape target object, and then performs fine adjustment of the target center point position of the predetermined shape target object and reconstruction of the category relationship through the adjustment vector, so as to obtain accurate positioning of the target center of the predetermined shape target object. Then, the segmentation result of the predetermined shape target object is obtained by using the reconstructed category relationship of the vertebrae. Finally, the pixel points in the binarization result of the predetermined shape target object are assigned to the target center point of the corresponding predetermined shape target object according to the adjustment vector, so as to generate the segmentation map of each predetermined shape target object. The positioning, recognition and segmentation of the predetermined shape target object are completed in one network, which is simple and efficient to realize the segmentation of the predetermined shape target object in the image.

[0123] Referring to Figure 3 , Figure 3 A flowchart of the first medical image processing method according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0124] Step 302: Obtain a medical image containing a vertebra, and perform image feature extraction on the medical image.

[0125] The medical image containing a vertebra includes but is not limited to a 3D CT image containing a vertebra. In the embodiments of the present specification, the medical image containing a vertebra can be obtained from a medical image library, or a medical image scanned by a CT machine in real time can be directly obtained.

[0126] In specific implementation, after obtaining the medical image containing a vertebra, image feature extraction is needed to be performed on the medical image, and the image features of the medical image are used for subsequent image processing. The specific implementation manner is as follows:

[0127] The image feature extraction on the medical image includes:

[0128] performing initial image feature extraction on the medical image based on a feature extraction model to obtain the initial image features;

[0129] performing convolution processing on the initial image features through a plurality of convolution layers to obtain image features after convolution processing;

[0130] The image features after the convolution processing are extracted again according to preset tasks, to obtain first image features and second image features, and the preset tasks include a first task, a second task and a third task.

[0131] The feature extraction model can be any one of preset feature extraction models that can realize image feature extraction of the medical image of the vertebral body, such as a Vnet model.

[0132] Taking the Vnet model as the feature extraction model and the medical image as a 3D spine volume image as an example, first, the Vnet model is used to perform initial image feature extraction on the 3D spine volume image, to obtain initial image features of the 3D spine volume image, then the image features of the 3D spine volume image are convolved through the multi-layer convolution layer, to obtain image features after the convolution processing, and finally, the image features after the convolution processing are extracted again based on different preset tasks, to realize obtaining the first image features and the second image features.

[0133] In actual application, the Vnet model usually outputs initial image features of the medical image, and the application range of these initial image features is relatively wide, for example, the initial image features can be applied to some image recognition and detection, but the recognition efficiency and detection effect are not very good. In this case, the initial image features are convolved, and the feature extraction is performed according to the preset tasks. The preset task can be understood as an item that needs to be completed by using the image features of the medical image in subsequent specific processing, such as a positioning task, a segmentation task, etc. It can be understood that different preset tasks require different image features, and the image features extracted again based on the preset tasks can be more suitable for subsequent processing tasks, so that the subsequent task processing is more accurate. Since the medical image processing method of the present specification is realized by the entire prediction network, the learning effect and optimization effect of the prediction network are better when the prediction network is trained by using the image features of the sample image extracted again in the specific training stage of the prediction network of the present specification.

[0134] In specific implementation, the multi-layer convolution layer can be understood as three layers or more than three layers. The initial image features are input to the feature extractor again after being processed by the multi-layer convolution, and further feature extraction is performed on the image features after the convolution according to the preset task, to obtain the first image features and the second image features after the feature extraction again. The preset task includes but is not limited to the first task, the second task and the third task. In specific application, the first task can be understood as a center point acquisition task of the vertebral body, the second task can be understood as a segmentation task of the vertebral body, and the third task can be understood as a segmentation task of the vertebral body.

[0135] In a specific application, in order to capture multi-scale features, the feature extraction model of the embodiment of the present specification selects V-Net [9] as a feature extractor. V-Net [9] can simultaneously maintain high-level semantics and low-level image details to ensure the completeness of image features. In addition, V-Net adopts a coding and decoding mode, and the feature map output by the decoder has the same resolution as the original vertebra, which is important for subsequent accurate 3D segmentation of the vertebra. Specifically, the extracted initial image features are further processed through three cascaded convolutions to obtain first image features and second image features.

[0136] In the embodiment of the present specification, the initial image features extracted by the basic feature extraction model are further processed through convolution, and then image features are extracted again based on a preset task, so as to establish a correlation between the image features and the subsequent task to be implemented, and to enable the subsequent task to be implemented to be processed more accurately and efficiently based on the image features.

[0137] Step 304: determining the initial center point, adjustment vector and initial category of each vertebra based on the image features.

[0138] After obtaining the first image features and the second image features of the medical image, it is necessary to obtain the initial center point, adjustment vector and initial category of each vertebra based on the first image features. The specific implementation is as follows:

[0139] The first task is implemented by a first network and a second network, and the second task is implemented by a third network;

[0140] Correspondingly, the determination of the initial center point, adjustment vector and initial category of each vertebra based on the image features comprises:

[0141] inputting the first image features into the first network to obtain the initial center point of each vertebra;

[0142] inputting the first image features into the second network to obtain the adjustment vector of each vertebra; and

[0143] inputting the first image features into the third network to obtain the initial category of each vertebra.

[0144] Wherein, the first network, the second network and the third network are part of the prediction network of the embodiment of the present specification. In actual application, each network can also be understood as a head. The first network can be a heat map head, the second network can be a vector field head, and the third network can be a vertebra recognition head.

[0145] Specifically, the first task is implemented by the first network and the second network, and the second task is implemented by the third network. Still taking the above example, if the first task is a task of obtaining the center point of the vertebral body and the second task is a task of segmenting the vertebral body, then the task of obtaining the center point of the vertebral body is implemented by the heat map head and the vector field head, and the task of segmenting the vertebral body is implemented by the vertebral body recognition head.

[0146] In actual application, since the first task and the second task are both related to the center point of the vertebral body, the first task and the second task can share a feature, that is, the first image feature.

[0147] In specific implementation, the first image feature is input into the heat map head to obtain the initial center point of each vertebral body, the first image feature is input into the vector field head to obtain the adjustment vector of each vertebral body, and the first image feature is input into the vertebral body recognition head to obtain the initial category and the initial label of each vertebral body.

[0148] In the embodiments of the present specification, different heads are used to implement the preset tasks to obtain the initial center point, the adjustment vector, the initial category and the initial label of each vertebral body, so that the target center point and the segmentation result of each vertebral body can be accurately obtained through the initial center point, the adjustment vector, the initial category and the initial label of each vertebral body.

[0149] In another embodiment of the present specification, the inputting of the first image feature into the first network to obtain the initial center point of each vertebral body includes:

[0150] The first image feature is input into the first network to obtain the probability that each pixel in each vertebral body is an initial center point, and the initial center point of each vertebral body is determined according to the probability.

[0151] In actual application, the positioning of the vertebral body is implemented according to the target center point of each vertebral body. In order to effectively implement this point, the heat map head is used to predict the probability that each pixel of each vertebral body corresponds to the center point of the vertebral body, and a binary heat map can be used instead of a traditional Gaussian heat map as a supervised image. In specific application, the real heat map image h(x, x*) of the vertebral body is obtained by formula 1:

[0152]

[0153] wherein x* represents the initial center point of the vertebral body, x represents a pixel in the 3D spine volume image, and R represents a radius selected in the experiment and can be set to 6.

[0154] The heat map head predicts the probability that a pixel is located within a sphere with a radius R centered on the corresponding vertebral body center point, and then takes the pixel with the highest probability as the initial center point of the corresponding vertebral body. The target activation is equal to 1 inside the sphere and 0 outside the sphere, that is, the pixels inside the sphere are represented by 1 and the pixels outside the sphere are represented by 0; therefore, the center point detection of the vertebral body can be regarded as a binary classification task, and the target center point of the vertebral body can also be determined by filtering the local maximum value of each region in the predicted heat map.

[0155] Specifically, the first image feature is input into the heat map head, and the initial center point of each vertebral body can be predicted by the heat map head. Based on the initial center point of each vertebral body, the pixels in the 3D spine volume image, and a preset radius, the heat map of each vertebral body can be obtained by using the above formula 1, and the heat map of each vertebral body can be understood as the initial center point region of each vertebral body in the following embodiments.

[0156] In the embodiments of the present specification, the initial center region of each vertebral body is effectively determined by the heat map head, and the heat map of each vertebral body is obtained based on the initial center region of each vertebral body, so that the initial center point of each vertebral body can be accurately adjusted by the pixels in the heat map of each vertebral body subsequently.

[0157] In specific implementation, after determining the initial center point of each vertebral body according to the probability, the method further includes:

[0158] The initial center point region of each vertebral body is determined based on each pixel in the vertebral body, the initial center point, and a preset radius.

[0159] Specifically, based on each pixel in each vertebral body, the initial center point, and a preset radius, the initial center point region of each vertebral body, which can also be referred to as the heat map of each vertebral body, is calculated by using the above formula 1. In the case that the initial center point region of each vertebral body is accurately determined, the target center point of each vertebral body can be accurately adjusted based on the pixels in each initial center point region subsequently, so as to improve the positioning accuracy of the vertebral body.

[0160] In another embodiment of the present specification, the method of inputting the first image feature into the second network to obtain the adjustment vector of each vertebral body includes:

[0161] The first image feature is input into a relationship vector network to obtain a relationship vector of the initial center points of two adjacent vertebral bodies; and

[0162] The first image feature is input into an offset vector network to obtain an offset vector of each pixel in the initial center point region of each vertebral body corresponding to the initial center point of each vertebral body.

[0163] Specifically, the second network includes a relationship vector network and an offset vector network. In the case of the vector field head, the relationship vector network can be understood as a relationship vector field head, and the offset vector network can be understood as an offset vector field head.

[0164] In actual application, while the first image feature is input into the first network to obtain the initial center point of each vertebral body, the first image feature is also input into the relationship vector field head to obtain the relationship vector of the initial center points of two adjacent vertebral bodies, and the first image feature is input into the offset vector field head to obtain the offset vector of each pixel in the initial center point region of each vertebral body corresponding to the initial center point of each vertebral body.

[0165] In specific implementation, in order to obtain accurate positioning and identification of the vertebral bodies, the vector field head is used to predict two types of vectors: one three-dimensional local offset vector for determining the accurate key point coordinates of each vertebral body, and one three-dimensional relationship vector for obtaining the correct topological relationship between the vertebral bodies. The three-dimensional local offset vector is obtained through the offset vector field head, which is used to determine the accurate target center point coordinates of each vertebral body. The three-dimensional relationship vector is obtained through the relationship vector field head, which is used to obtain the correct topological relationship between the vertebral bodies. The obtained offset vector constitutes an offset vector map, which has three channels for representing the offset vector of each pixel in the spherical region of the vertebral body in the three-dimensional space corresponding to the center point of each vertebral body as defined in formula 1. The obtained relationship vector constitutes a relationship vector map, which has 2*3 channels for representing the vector pointing to the center points of two adjacent vertebral bodies, representing the chain structure of the spine, and the relationship vectors of the top and bottom vertebral bodies only point to their own centers because the top and bottom vertebral bodies are the starting point and end point of the chain. When the vector field head is pre-trained, the L1 loss function can be used to punish the error on the vector field, and only the inside of the disc is back-propagated.

[0166] In the embodiments of the present application, the relationship vector and the offset vector of the vertebral body are obtained through the relationship vector field head and the offset vector field head, so that the initial center point of the vertebral body can be adjusted based on the relationship vector and the offset vector to obtain accurate positioning and identification of the initial center point of each vertebral body.

[0167] Step 306: adjusting the initial center point of each vertebral body according to the adjustment vector to obtain the target center point of each vertebral body and the connection relationship of the target center point.

[0168] The adjustment vector includes a relationship vector and an offset vector, and the initial center point of each vertebra is adjusted according to the relationship vector and the offset vector to obtain a target center point of each vertebra and a connection relationship of the target center point, and the specific implementation is as follows:

[0169] The adjustment vector includes a relationship vector and an offset vector, and the initial center point of each vertebra is adjusted according to the relationship vector and the offset vector to obtain a target center point of each vertebra and a connection relationship of the target center point, and the specific implementation is as follows:

[0170] The value of each pixel in the initial center region of each vertebra and the offset vector are input into a pre-designed calculation formula to obtain the probability of each pixel in the initial center region of each vertebra as a center point, and the pixel with the maximum probability is taken as the candidate center point of each vertebra.

[0171] Based on the relationship vector of the candidate center point of each vertebra and the initial center point of two adjacent vertebrae, the target center point of each vertebra and the connection relationship of the target center point are obtained.

[0172] Specifically, the offset vector field can be subjected to Hough-voting, and the corresponding activation value in the heat map is used as a voting weight to generate a score map, and the probability of each pixel point in the spherical region of each vertebra in the heat map as a center point can be determined through the score map.

[0173] In practical applications, the pre-designed calculation formula can be formula 2, and through formula 2, the voting result of each pixel point in the spherical region of each vertebra in the heat map can be obtained.

[0174]

[0175] Wherein, F(x+V(x)-y) represents a trilinear interpolation kernel, x represents a pixel in a 3D spine volume image, H(x) represents the coordinates of the pixel, V(x) represents the offset vector of the pixel, and y represents all values that satisfy the condition ||x+V(x)-y||<1, that is, it can be understood as a voting result.

[0176] Specifically, the voting result of each pixel in the spherical region of each vertebra can be calculated through formula 2, then the voting result is converted into a probability value, the probability value is reflected on each pixel in the spherical region of each vertebra in the form of a score map, and finally the position corresponding to the local maximum value of the pixel in the spherical region of each vertebra in the score map is determined as the candidate center point.

[0177] After obtaining the candidate center points, the candidate center points are formed into a spine chain structure and the candidate center points are formed into a queue. Specifically, the highest-scored point in the candidate center points is selected as the initial seed for the breadth-first search, and then the dequeued points and their relationship vectors point to the potential positions of their adjacent vertebrae. When the minimum distance is greater than a threshold (for example, the threshold is set to 50 pixels) or the predicted adjacent point is itself, the search is terminated, and the remaining candidate center points in the candidate center point queue are discarded as false positive points. Finally, an accurate spine chain structure is formed based on the remaining candidate center points, which represents the connection relationship of all the target center points of the vertebrae.

[0178] Referring to Figure 4 , Figure 4 is a schematic diagram of the chain structure of the target center points of the vertebrae in the medical image processing method of the embodiments of the present disclosure.

[0179] Specifically, Figure 4 is a flowchart of the final target center points of the vertebrae generated according to the above processing method of obtaining the target center points of the vertebrae, wherein Figure 4 (a) represents a heat map of the initial center points of the vertebrae, Figure 4 (b) represents a candidate center point map after refining the positions of the initial center points of the vertebrae using the offset vector field, Figure 4 (c) represents a schematic diagram of reconstructing the chain structure of the spine using the relationship vector field.

[0180] In the embodiments of the present disclosure, the offset vector can obtain more accurate positions of the center points of each vertebra, and the relationship vector can obtain a more reliable spatial chain structure of the spine, which can effectively filter out false positive points, so that subsequent segmentation of the vertebrae can be based on the reconstructed spine chain structure to achieve accurate segmentation of the vertebrae.

[0181] Step 308: adjusting the initial class of each vertebra based on the connection relationship of the target center points to obtain the target class and target label of each vertebra.

[0182] In practical applications, it is difficult to directly identify the category and label of a single vertebra even artificially due to the similarity of adjacent vertebrae. In the prior art, experienced doctors generally first locate a vertebra with obvious bone characteristics, and then infer the remaining vertebrae according to their relative positions. Therefore, in the implementation of the present specification, the spine is divided into five independent categories, such as C2, T1, T11, L5, and S1, according to the anatomical characteristics of the spine. Then, the third network, i.e., the vertebra recognition head, is trained based on the above five independent categories, so that the vertebra recognition head becomes a classification subnetwork. In the embodiment of the present specification, the five categories of the above-mentioned spine vertebrae can be predicted, and a probability map with the same resolution as the original volume is generated. In the specific training process, the cross-entropy loss is calculated and back-propagated only within a sphere with a radius R centered on the vertebra.

[0183] In a specific implementation, the vertebra recognition head described above is used, and the category of the target center point of a given vertebra can be determined by the channel with the maximum response. However, this approach completely ignores the constraints of adjacent vertebrae and may result in unrealistic conflicts in many cases. Therefore, considering the connection relationship of the target center point of each vertebra, i.e., the chain structure, can be constructed by the relationship vector in formula 3. In practical applications, the maximum likelihood algorithm is selected to accurately adjust the initial category of each vertebra identified by the vertebra recognition head. The specific implementation is as follows:

[0184] The adjustment of the initial category of each vertebra based on the connection relationship of the target center point to obtain the target category of each vertebra includes:

[0185] Adjusting the initial category of each vertebra based on the connection relationship of the target center point using a maximum likelihood algorithm to obtain the target category and target label of each vertebra.

[0186] Wherein, the maximum likelihood algorithm is represented by formula 4:

[0187]

[0188] Wherein, M(j) represents the mapping from a class to a corresponding channel, p represents the probability, and k represents the vertebra.

[0189] Specifically, the five types of vertebral bodies are identified by the head, and then the initial class of each vertebral body is calculated based on the maximum likelihood algorithm, and finally the initial class of each vertebral body is adjusted based on the chain structure diagram of the target center point to obtain the target class and target label of each vertebral body; for example, the initial class of vertebral body 1 is L1, and the initial class of vertebral body 2 is L5, but it can be clearly seen from the chain structure of the vertebral body that vertebral body 1 and vertebral body 1 are two adjacent vertebral bodies, so the initial class of vertebral body 1 or vertebral body 2 can be adjusted based on the chain structure of the vertebral body to obtain the accurate target class of vertebral body 1 and vertebral body 2, and then the accurate target label is obtained based on the target class, for example, which vertebral body is vertebral body 1 and which vertebral body is vertebral body 2.

[0190] The medical image processing method provided by the embodiments of the present specification obtains the target center point, the adjustment vector and the initial class of the vertebral body by extracting the image features, and then performs fine adjustment of the target center point position of the vertebral body and reconstruction of the class relationship of the vertebral body by the adjustment vector, so as to obtain accurate positioning of the target center of the vertebral body, and then obtain the segmentation result of the vertebral body by using the reconstructed class relationship of the vertebral body, and further obtain the target class and target label of the vertebral body, and finally greatly improve the classification accuracy of the vertebral body by the target class and target label of the vertebral body.

[0191] In another embodiment of the present specification, after obtaining the target class and target label of each predetermined shape target object, the method further comprises:

[0192] The binary result of each vertebral body is determined according to the image features, and each vertebral body is segmented based on the binary result, the target class, the target label and the target center point of each vertebral body.

[0193] Specifically, the third task is implemented by a fourth network;

[0194] Correspondingly, the binary result of each vertebral body is determined according to the image features, and each vertebral body is segmented based on the binary result, the target class, the target label and the target center point of each vertebral body.

[0195] The second image features are input into the fourth network to obtain the binary result of each vertebral body.

[0196] Wherein, the fourth network can be understood as a binary head, the second image features are input into the binary head, and the binary result of each vertebral body, i.e. the binary foreground pixels, can be obtained, and finally each vertebral body is segmented based on the binary result, the target class, the target label and the target center point of each vertebral body.

[0197] The challenging vertebra segmentation problem in the embodiments of the present specification is decomposed into a binary segmentation problem and an instance assignment problem. First, the pixel points of the region of the entire spine are segmented by binary segmentation, and then the class and label of each pixel point are obtained based on the target label and target class of each vertebra obtained above, to realize accurate contour determination of each vertebra.

[0198] In specific implementation, the binarization head outputs two channels of results, corresponding to segmenting the entire spine, and each pixel in the spine region obtains its target class and target label from the target center point closest to the predicted target center point of the pixel, which is calculated based on the Euclidean distance: d(x+v, x*i), where x is the coordinate of the pixel, v is the offset vector predicted by the vector field head, and x*i is the center of the vertebra; the specific implementation is as follows:

[0199] The segmentation of each vertebra based on the binary result, target class, target label, and target center point of each vertebra includes:

[0200] The coordinates of each pixel of each vertebra after binarization are obtained, and based on the coordinates of each pixel, the offset vector, and the target center point of each vertebra, the closest target center point of each pixel is determined;

[0201] The target class and target label of each pixel are determined based on the target class and target label of the closest target center point of each pixel, and the segmentation of each vertebra is realized according to the target class and target label of each pixel.

[0202] Specifically, the second image feature is input into the binarization head to obtain the pixels of the entire spine region and the coordinates of each pixel, and then the closest target center point of each pixel is calculated by the above Euclidean distance, and then the target class and target label of the vertebra with the closest target center point are taken as the target class and target label of the pixel, and so on. In this way, the target class and target label of each pixel in the entire spine region can be obtained, and the accurate contour of each vertebra of the spine can be determined based on the target class and target label of each pixel. Finally, based on the accurate contour of each vertebra, accurate and convenient segmentation of the vertebrae of the spine can be realized to obtain the segmentation map of each vertebra in the medical image.

[0203] The medical image processing method provided in the embodiments of the present specification simultaneously implements positioning, segmentation and segmentation tasks by using a unified framework, and the tasks are related to each other. The offset vector field obtained through positioning can adjust the position of the target center point of the vertebral body to be more accurate, and the relationship vector field obtained through positioning obtains the chain structure of each vertebral body of the spine, which can significantly filter out false positive vertebral center points. Finally, the vertebral body is segmented and segmented through the accurate target center point of the vertebral body. The medical image processing method provided in the embodiments of the present specification greatly improves the performance of segmenting the vertebral body in the medical image.

[0204] The following describes the embodiments of the present specification in conjunction with the accompanying Figure 5 The medical image processing method provided in the embodiments of the present specification is further described by taking the application of the medical image processing method to the segmentation of the vertebral body in the spine CT image as an example. Among them, Figure 5 FIG. 1 shows a processing process schematic diagram of a medical image processing method provided in an embodiment of the present specification, which specifically includes the following steps.

[0205] Step one: obtain a spine CT image, and process the spine CT image into multiple 3D spine volume images.

[0206] Step two: input each 3D spine volume image into Vnet for initial feature extraction.

[0207] Step three: perform convolution processing on the 3D spine volume image features after initial feature extraction through three cascaded convolutions, and then perform secondary feature extraction on the 3D spine volume image features after convolution processing based on a preset task.

[0208] In the embodiments of the present specification, the preset task can be regarded as a center point acquisition task of the vertebral body, an offset vector acquisition task of the vertebral body, a relationship vector acquisition task of the vertebral body, a category acquisition task of the vertebral body, and a binary result acquisition task of the vertebral body. In actual application, in addition to the binary result acquisition task of the vertebral body, the other preset tasks are all related to the center point, so the other preset tasks can share the features of one branch, and the binary result acquisition task uses the features of one branch.

[0209] Step four: input the 3D spine volume image features after convolution processing into a heat map head, a relationship vector field head, an offset vector field head, a vertebral body recognition head, and a binary segmentation head to obtain the initial center point, the relationship vector, the offset vector, the initial category and the binary result of each vertebral body in the 3D spine volume image.

[0210] Step five: adjust the initial center point and initial class of each vertebra in the 3D spine volume image through the relationship vector, offset vector of each vertebra in the 3D spine volume image, to obtain the target center point and target class of each vertebra.

[0211] Step six: according to the offset vector, assign each pixel point of the binarization result of the vertebra in the binary segmentation head output to the target center point of the vertebra closest to it, match the corresponding target class for each pixel point based on the target class of the target center point of the vertebra closest to it, and generate a segmentation map of each vertebra in the 3D spine volume image based on the target class of each pixel point.

[0212] Referring to Figure 6 , Figure 6 The medical image processing method provided by the embodiments of the present specification shows the specific vertebra segmentation diagram, i.e. the specific initial center point, target center point and segmentation diagram obtained after processing each vertebra in the 3D spine volume image.

[0213] Specifically, Figure 6 (a) can be regarded as a diagram of the initial center point of the side surface of each vertebra in the 3D spine volume image, Figure 6 (b) can be regarded as a diagram of the initial center point of the front surface of each vertebra in the 3D spine volume image, Figure 6 (c) can be regarded as a diagram of the target center point of the side surface of each vertebra in the 3D spine volume image, Figure 6 (d) can be regarded as a diagram of the target center point of the front surface of each vertebra in the 3D spine volume image, Figure 6 (e) can be regarded as a diagram of the segmentation of each vertebra in the 3D spine volume image.

[0214] The network framework used in the medical image processing method provided in the embodiments of the present specification mainly consists of six parts: (1) a feature extractor; (2) a heat map head; (3) a relationship vector field head; (4) an offset vector field head; (5) a vertebra recognition head; and (6) a binary segmentation head. In actual application, during the training of the framework, different supervision signals are provided to the five heads in a multi-task manner, while in the inference stage, i.e., the actual application stage, the 3D spine volume image is first input into the feature extractor, and the generated features are used by different heads at the same time. In the specific post-processing stage, the center point position of the vertebra is fine-tuned and the category relationship is reconstructed by the offset vector field and the relationship vector field, and the accurate positioning of the vertebra center point is obtained from the heat map. Then, the segmentation result of each vertebra center point is further obtained by using the maximum likelihood algorithm. Finally, according to the offset vector field, the foreground pixel points of the vertebra in the binary segmentation are assigned to the specific vertebra center point (i.e., the nearest vertebra center point of each pixel point), so as to generate the segmentation map of each vertebra. The network framework used in the medical image processing method of the embodiments of the present specification is a fully convolutional and end-to-end manner. The vertebra positioning, recognition and segmentation are unified into one network through the end-to-end manner. In addition, a simpler and more explicit method is provided to embed the anatomical prior information of the spine into the convolutional neural network, which can achieve better positioning and segmentation. Therefore, the medical image processing method provided in the embodiments of the present specification can provide a simple and efficient vertebra segmentation implementation.

[0215] Corresponding to the method embodiments described above, the present specification also provides image processing device embodiments, Figure 7 The structure of an image processing device provided by one embodiment of the present specification is shown in a schematic diagram. As shown in the figure, Figure 7 The device comprises:

[0216] The first image acquisition module 702 is configured to acquire an image containing a predetermined shape target object, and perform image feature extraction on the image;

[0217] The first determination module 704 is configured to determine the initial center point, the adjustment vector and the initial category of each predetermined shape target object based on the image features;

[0218] The first adjustment module 706 is configured to adjust the initial center point of each predetermined shape target object according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center point;

[0219] The first obtaining module 708 is configured to adjust the initial category of each predetermined shape target object based on the connection relationship of the target center point, so as to obtain the target category and target label of each predetermined shape target object.

[0220] Optionally, the apparatus further includes:

[0221] The third segmentation module is configured to determine the binarization result of each predetermined shape target object according to the image feature, and segment each predetermined shape target object based on the binarization result, target category, target label and target center point of each predetermined shape target object.

[0222] Optionally, the first image obtaining module 702 is further configured to:

[0223] perform initial image feature extraction on the image based on a feature extraction model to obtain the initial image feature;

[0224] perform convolution processing on the initial image feature through a plurality of convolution layers to obtain a convolution-processed image feature;

[0225] perform re-extraction on the convolution-processed image feature according to a preset task to obtain a first image feature and a second image feature, the preset task including a first task, a second task and a third task.

[0226] Optionally, the first task is implemented by a first network and a second network, and the second task is implemented by a third network.

[0227] Correspondingly, the first determining module 704 is further configured to:

[0228] input the first image feature into the first network to obtain the initial center point of each predetermined shape target object;

[0229] input the first image feature into the second network to obtain the adjustment vector of each predetermined shape target object; and

[0230] input the first image feature into the third network to obtain the initial category of each predetermined shape target object.

[0231] Optionally, the first determining module 704 is further configured to:

[0232] input the first image feature into the first network to obtain the probability that each pixel in each predetermined shape target object is an initial center point, and determine the initial center point of each predetermined shape target object according to the probability.

[0233] Optionally, the apparatus further comprises:

[0234] The fifth determining module is configured to determine an initial center point region of each of the predetermined shape target objects based on each pixel in each of the predetermined shape target objects, the initial center point, and a preset radius.

[0235] Optionally, the first determining module 704 is further configured to:

[0236] input the first image features into a relationship vector network to obtain a relationship vector of the initial center points of the two adjacent predetermined shape target objects; and

[0237] input the first image features into an offset vector network to obtain an offset vector corresponding to the initial center point of each of the predetermined shape target objects for each pixel in the initial center point region of each of the predetermined shape target objects.

[0238] Optionally, the first adjusting module 706 is further configured to:

[0239] input the value of each pixel in the initial center region of each of the predetermined shape target objects and the offset vector into a preset calculation formula to obtain a probability that each pixel in the initial center region of each of the predetermined shape target objects is a center point, and take the pixel with the maximum probability as a candidate center point of each of the predetermined shape target objects;

[0240] obtain a target center point of each of the predetermined shape target objects and a connection relationship of the target center point based on the candidate center point of each of the predetermined shape target objects and the relationship vector of the initial center points of the two adjacent predetermined shape target objects.

[0241] Optionally, the first obtaining module 708 is further configured to:

[0242] adjust the initial category of each of the predetermined shape target objects based on the connection relationship of the target center point by using a maximum likelihood algorithm to obtain a target category and a target label of each of the predetermined shape target objects.

[0243] Optionally, the third task is implemented by a fourth network.

[0244] Correspondingly, the third segmentation module is further configured to:

[0245] input the second image features into the fourth network to obtain a binarization result of each of the predetermined shape target objects.

[0246] Optionally, the third segmentation module is further configured to:

[0247] obtain the coordinates of each pixel of the each predetermined shape target object after binarization, and determine the target center point closest to each pixel based on the coordinates of each pixel, the offset vector and the target center point of the each predetermined shape target object;

[0248] determine the target category and target label of each pixel based on the target category and target label of the target center point closest to each pixel, and segment the each predetermined shape target object according to the target category and target label of each pixel.

[0249] The above is a schematic scheme of the image processing device of the embodiment. It should be noted that the technical scheme of the image processing device belongs to the same concept as the technical scheme of the image processing method described above, and the details of the technical scheme of the image processing device which are not described in detail can be referred to the description of the technical scheme of the image processing method.

[0250] Corresponding to the method embodiments described above, the present specification also provides medical image processing device embodiments, Figure 8 A structural schematic diagram of a first medical image processing device provided by an embodiment of the present specification is shown. As shown in the figure, Figure 8 The device comprises:

[0251] The second image acquisition module 802 is configured to acquire a medical image containing a vertebral body and perform image feature extraction on the medical image;

[0252] The second determination module 804 is configured to determine the initial center point, the adjustment vector and the initial category of each vertebral body based on the image features;

[0253] The second adjustment module 806 is configured to adjust the initial center point of each vertebral body according to the adjustment vector to obtain the target center point of each vertebral body and the connection relationship of the target center points;

[0254] The second obtaining module 808 is configured to adjust the initial category of each vertebral body based on the connection relationship of the target center points to obtain the target category and target label of each vertebral body.

[0255] Optionally, the device further comprises:

[0256] The fourth segmentation module is configured to determine the binarization result of each vertebral body according to the image features, and segment each vertebral body based on the binarization result, the target category, the target label and the target center point of each vertebral body.

[0257] Optionally, the second image acquisition module 802 is further configured to:

[0258] perform initial image feature extraction on the medical image based on a feature extraction model to obtain initial image features;

[0259] perform convolution processing on the initial image features through a plurality of convolution layers to obtain image features after convolution processing;

[0260] perform re-extraction on the image features after convolution processing according to preset tasks to obtain first image features and second image features, the preset tasks including a first task, a second task and a third task.

[0261] Optionally, the first task is implemented by a first network and a second network, and the second task is implemented by a third network.

[0262] Correspondingly, the second determination module 804 is further configured to:

[0263] input the first image features into the first network to obtain initial center points of each vertebral body;

[0264] input the first image features into the second network to obtain adjustment vectors of each vertebral body; and

[0265] input the first image features into the third network to obtain initial categories of each vertebral body.

[0266] Optionally, the second determination module 804 is further configured to:

[0267] input the first image features into the first network to obtain a probability that each pixel in each vertebral body is an initial center point, and determine the initial center point of each vertebral body according to the probability.

[0268] Optionally, the apparatus further includes:

[0269] a region determination module configured to determine an initial center point region of each vertebral body based on each pixel in each vertebral body, an initial center point and a preset radius.

[0270] Optionally, the second determination module 804 is further configured to:

[0271] input the first image features into a relationship vector network to obtain a relationship vector of initial center points of two adjacent vertebral bodies; and

[0272] input the first image features into an offset vector network to obtain an offset vector of each pixel in the initial center point region of each vertebral body corresponding to the initial center point of each vertebral body.

[0273] Optionally, the second adjustment module 806 is further configured to:

[0274] inputting the value of each pixel in the initial central region of each vertebral body and the offset vector into a pre-designed calculation formula to obtain a probability of each pixel in the initial central region of each vertebral body as a center point, and taking the pixel with the maximum probability as a candidate center point of each vertebral body;

[0275] obtaining a target center point of each vertebral body and a connection relationship of the target center point based on the candidate center point of each vertebral body and a relationship vector of the initial center points of two adjacent vertebral bodies.

[0276] Optionally, the second obtaining module 808 is further configured to:

[0277] adjusting the initial category of each vertebral body based on the connection relationship of the target center point by using a maximum likelihood algorithm to obtain a target category and a target label of each vertebral body.

[0278] Optionally, the third task is implemented by a fourth network.

[0279] Correspondingly, the fourth segmentation module is further configured to:

[0280] inputting the second image feature into the fourth network to obtain a binarization result of each vertebral body.

[0281] Optionally, the fourth segmentation module is further configured to:

[0282] obtaining the coordinates of each pixel of each vertebral body after binarization, and determining the nearest target center point of each pixel based on the coordinates of each pixel, the offset vector and the target center point of each vertebral body.

[0283] determining the target category and the target label of each pixel based on the target category and the target label of the nearest target center point of each pixel, and implementing segmentation of each vertebral body according to the target category and the target label of each pixel.

[0284] In the implementation of the present specification, the medical image processing apparatus obtains the target center point, the adjustment vector, the initial category and the binarization result of the vertebral body through the extracted image features, then performs fine adjustment of the position of the target center point of the vertebral body and reconstruction of the category relationship of the vertebral body through the adjustment vector, so as to obtain accurate positioning of the target center of the vertebral body, and then obtains the segmentation result of the vertebral body by using the reconstructed category relationship of the vertebral body, finally assigns the pixel points in the binarization result of the vertebral body to the target center point of a specific vertebral body according to the adjustment vector to generate a segmentation map of each vertebral body, and the positioning, recognition and segmentation of the vertebral body are completed in one network, so that the segmentation of the vertebral body in the medical image is simply and efficiently realized.

[0285] The above is a schematic scheme of the medical image processing device of the embodiment. It should be noted that the technical scheme of the medical image processing device belongs to the same concept as the technical scheme of the first medical image processing method described above, and the details of the technical scheme of the medical image processing device that are not described in detail can be referred to the description of the technical scheme of the medical image processing method.

[0286] Referring to Figure 9 , Figure 9 A flowchart of a second medical image processing method according to an embodiment of the present specification is shown, which specifically includes the following steps.

[0287] Step 902: Based on the user's calling request, an image input interface is displayed for the user.

[0288] Specifically, in the case of receiving the user's calling request, the image input interface is determined according to the calling request, and the image input interface is displayed to the user, and the user can input the medical image containing the vertebral body through the image input interface.

[0289] In practical applications, the user includes but is not limited to merchants or individual users, etc.

[0290] Step 904: Receiving the medical image containing the vertebral body input by the user based on the image input interface, and performing image feature extraction on the medical image.

[0291] Step 906: Based on the image features, determine the initial center point, adjustment vector and initial category of each vertebral body.

[0292] Step 908: Adjust the initial center point of each vertebral body according to the adjustment vector to obtain the target center point of each vertebral body and the connection relationship of the target center point.

[0293] Step 910: Adjust the initial category of each vertebral body based on the connection relationship of the target center point to obtain the target category and target label of each vertebral body.

[0294] Step 912: Determine the binarization result of each vertebral body according to the image features, and segment each vertebral body based on the binarization result of each vertebral body, target category, target label and target center point, and return the segmented medical image to the user.

[0295] It should be noted that the part of the second medical image processing method provided by the embodiment of the present specification corresponding to the above-mentioned embodiment of the first medical image processing method can be referred to the detailed description in the above-mentioned embodiment of the first medical image processing method, which will not be repeated here.

[0296] In the implementation of the present specification, the medical image processing method obtains the target center point of the vertebral body, the adjustment vector, the initial category, and the binarization result through the extracted image features, then performs fine adjustment of the target center point position of the vertebral body and reconstruction of the category relationship of the vertebral body through the adjustment vector, so as to obtain accurate positioning of the target center of the vertebral body, and then obtains the segmentation result of the vertebral body by using the reconstructed category relationship of the vertebral body, and finally assigns the pixel points in the binarization result of the vertebral body to the target center point of the specific vertebral body according to the adjustment vector, to generate the segmentation map of each vertebral body, which is completed by unifying the positioning, recognition, and segmentation of the vertebral body in one network, and is simple and efficient in realizing the segmentation of the vertebral body in the medical image.

[0297] Corresponding to the method embodiment, the present specification also provides a medical image processing device embodiment, Figure 10 The structure schematic diagram of the second medical image processing device provided by an embodiment of the present specification is shown. As shown in the figure, Figure 10 The device comprises:

[0298] The interface display module 1002 is configured to display an image input interface for the user based on the user's calling request;

[0299] The image receiving module 1004 is configured to receive the medical image containing the vertebral body input by the user based on the image input interface, and perform image feature extraction on the medical image;

[0300] The third determination module 1006 is configured to determine the initial center point, the adjustment vector, and the initial category of each vertebral body based on the image features;

[0301] The third adjustment module 1008 is configured to adjust the initial center point of each vertebral body according to the adjustment vector, to obtain the target center point of each vertebral body and the connection relationship of the target center point;

[0302] The third obtaining module 1010 is configured to adjust the initial category of each vertebral body based on the connection relationship of the target center point, to obtain the target category and the target label of each vertebral body;

[0303] The first segmentation module 1012 is configured to determine the binarization result of each vertebral body according to the image features, and segment each vertebral body based on the binarization result, the target category, the target label, and the target center point of each vertebral body, and return the segmented medical image to the user.

[0304] In the implementation of the specification, the medical image processing apparatus obtains the target center point of the vertebral body, the adjustment vector, the initial category, and the binarization result through the extracted image features, then performs fine adjustment of the target center point position of the vertebral body and reconstruction of the category relationship of the vertebral body through the adjustment vector, thereby obtaining accurate positioning of the target center of the vertebral body, and then obtains the segmentation result of the vertebral body by using the reconstructed category relationship of the vertebral body, and finally assigns the pixel points in the binarization result of the vertebral body to the target center point of the specific vertebral body according to the adjustment vector, to generate the segmentation map of each vertebral body, which is completed by unifying the positioning, recognition, and segmentation of the vertebral body in one network, and is simple and efficient in realizing the segmentation of the vertebral body in the medical image.

[0305] The above is a schematic scheme of the second medical image processing apparatus of the embodiment. It should be noted that the technical scheme of the medical image processing apparatus belongs to the same concept as the technical scheme of the second medical image processing method described above, and the details of the technical scheme of the medical image processing apparatus that are not described in detail can be referred to the description of the technical scheme of the second medical image processing method.

[0306] Referring to Figure 11 , Figure 11 A flowchart of a third medical image processing method according to an embodiment of the specification is shown, which specifically includes the following steps.

[0307] Step 1102: receiving a calling request sent by a user and carrying a medical image containing a vertebral body, and performing image feature extraction on the medical image.

[0308] Step 1104: determining the initial center point, the adjustment vector, and the initial category of each vertebral body based on the image features.

[0309] Step 1106: adjusting the initial center point of each vertebral body according to the adjustment vector to obtain the target center point of each vertebral body and the connection relationship of the target center point.

[0310] Step 1108: adjusting the initial category of each vertebral body based on the connection relationship of the target center point to obtain the target category and the target label of each vertebral body.

[0311] Step 1110: determining the binarization result of each vertebral body according to the image features, and segmenting each vertebral body based on the binarization result, the target category, the target label, and the target center point of each vertebral body, and returning the segmented medical image to the user.

[0312] It should be noted that the third medical image processing method provided by the embodiments of the present specification corresponds to the above-mentioned embodiments of the first medical image processing method, and the detailed description can be referred to in the above-mentioned embodiments of the first medical image processing method, and will not be repeated here.

[0313] In practical applications, the medical image processing method is applied to a local server, and an API interface is provided for a user. After receiving an API calling request sent by the user, the medical image processing method is used to quickly and accurately segment the vertebral body in the medical image based on the medical image containing the vertebral body carried in the calling request of the user, so as to improve the user experience.

[0314] Corresponding to the above method embodiments, the present specification also provides medical image processing device embodiments, Figure 12 The structure of the third medical image processing device provided by an embodiment of the present specification is shown. As shown in Figure 12 The device comprises:

[0315] The request receiving module 1202 is configured to receive a calling request sent by a user and carrying a medical image containing a vertebral body, and perform image feature extraction on the medical image;

[0316] The fourth determination module 1204 is configured to determine the initial center point, the adjustment vector and the initial category of each vertebral body based on the image feature;

[0317] The fourth adjustment module 1206 is configured to adjust the initial center point of each vertebral body according to the adjustment vector to obtain the target center point of each vertebral body and the connection relationship of the target center point;

[0318] The fourth obtaining module 1208 is configured to adjust the initial category of each vertebral body based on the connection relationship of the target center point to obtain the target category and the target label of each vertebral body;

[0319] The second segmentation module 1210 is configured to determine the binarization result of each vertebral body according to the image feature, and segment each vertebral body based on the binarization result of each vertebral body, the target category, the target label and the target center point, and return the segmented medical image to the user.

[0320] In the implementation of the specification, the medical image processing apparatus obtains the target center point of the vertebral body, the adjustment vector, the initial category, and the binarization result through the extracted image features, then performs fine adjustment of the target center point position of the vertebral body and reconstruction of the category relationship of the vertebral body through the adjustment vector, so as to obtain accurate positioning of the target center of the vertebral body, and then obtains the segmentation result of the vertebral body by using the reconstructed category relationship of the vertebral body, and finally assigns the pixel points in the binarization result of the vertebral body to the specific target center point of the vertebral body according to the adjustment vector, to generate the segmentation map of each vertebral body, which is completed by unifying the positioning, recognition, and segmentation of the vertebral body in one network, and is simple and efficient to realize the segmentation of the vertebral body in the medical image.

[0321] The above is a schematic scheme of the third medical image processing apparatus of the embodiment. It should be noted that the technical scheme of the medical image processing apparatus belongs to the same concept as the technical scheme of the third medical image processing method described above, and the details of the technical scheme of the medical image processing apparatus that are not described in detail can be referred to the description of the technical scheme of the third medical image processing method.

[0322] Referring to Figure 13 , Figure 13 A structural block diagram of a computing device 1300 according to an embodiment of the specification is shown. The components of the computing device 1300 include, but are not limited to, a memory 1310 and a processor 1320. The processor 1320 is connected to the memory 1310 through a bus 1330, and a database 1350 is used to save data.

[0323] The computing device 1300 also includes an access device 1340, which enables the computing device 1300 to communicate via one or more networks 1360. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 1340 can include one or more of any type of network interface (e.g., network interface card (NIC)) such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a worldwide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near-field communication (NFC) interface, and the like, wired or wireless.

[0324] In an embodiment of the specification, the above-mentioned components of the computing device 1300 and other components not shown in Figure 13 may be connected to each other, for example, through a bus. It should be understood that Figure 13 The structural block diagram of the computing device shown is only for the purpose of example, and is not a limitation on the scope of the specification. Those skilled in the art can add or replace other components as needed.

[0325] The computing device 1300 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or PC. The computing device 1300 can also be a mobile or stationary server.

[0326] The processor 1320 is configured to execute computer-executable instructions to perform the steps of the image processing method and the medical image processing method.

[0327] The above is a schematic scheme of the computing device of the embodiment. It should be noted that the technical scheme of the computing device belongs to the same concept as the technical schemes of the image processing method and the medical image processing method, and the details of the technical scheme of the computing device that are not described in detail can be referred to the description of the technical schemes of the image processing method and the medical image processing method.

[0328] An embodiment of the present specification also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps of the medical image processing method.

[0329] The above is a schematic scheme of the computer-readable storage medium of the embodiment. It should be noted that the technical scheme of the storage medium belongs to the same concept as the technical schemes of the image processing method and the medical image processing method, and the details of the technical scheme of the storage medium that are not described in detail can be referred to the description of the technical schemes of the image processing method and the medical image processing method.

[0330] The above describes specific embodiments of the present specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than the order in which they are recited and still achieve desirable results. In addition, the processes depicted in the figures do not necessarily require the particular order shown, or sequential order, to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0331] The computer readable medium can include any entity or apparatus capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, software distribution medium, etc. It should be noted that the computer readable medium can include appropriate additions or subtractions according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0332] It should be noted that for the foregoing method embodiments, the description is made for the sake of brevity, and therefore, each of the method embodiments can include additional steps as appropriate to the alternate embodiments and vice versa, some of which have been discussed above. It is also noted that while the method embodiments have been described as a single process, the process can be separated into a number of processes for different aspects, each of which, and

[0333] In the above embodiments, the description of each embodiment is focused on different aspects, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0334] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details and limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and changes can be made. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and use the present specification. The present specification is limited by the claims and their entire scope and equivalents.

Claims

1. An image processing method, comprising: Acquire an image containing a target object of a predetermined shape, and extract image features from the image; Based on the image features, the initial center point, adjustment vector, and initial category of each predetermined shape target object are determined, wherein the adjustment vector is determined according to the offset vector and the relation vector, the offset vector is used to determine the precise key point coordinates of each predetermined shape target object, and the relation vector is used to determine the correct topological relationship between each predetermined shape target object; The initial center point of each predetermined shape target object is adjusted according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center point; The step of adjusting the initial center point of each predetermined shape target object according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center point includes: determining a candidate center point according to each pixel value in the initial central region of each predetermined shape target object and the offset vector; and obtaining the target center point of each predetermined shape target object and the connection relationship of the target center point based on the candidate center point of each predetermined shape target object and the relationship vector of the initial center points of two adjacent predetermined shape target objects. The initial category of each predetermined shape target object is adjusted based on the connection relationship of the target center point to obtain the target category and target label of each predetermined shape target object.

2. The image processing method according to claim 1, further comprising, after obtaining the target category and target label of each predetermined shape target object: The binarization result of each predetermined shape target object is determined based on the image features, and each predetermined shape target object is segmented based on the binarization result, target category, target label, and target center point.

3. The image processing method according to claim 2, wherein the image feature extraction of the image comprises: The image is initially extracted based on a feature extraction model to obtain the initial image features; The initial image features are convolved through multiple convolutional layers to obtain the convolved image features. The convolutional image features are extracted again according to a preset task to obtain first image features and second image features. The preset task includes a first task, a second task and a third task.

4. The image processing method according to claim 3, wherein the first task is implemented by a first network and a second network, and the second task is implemented by a third network; Accordingly, determining the initial center point, adjustment vector, and initial category of each predetermined shape target object based on the image features includes: The first image features are input into the first network to obtain the initial center point of each predetermined shape target object; The first image features are input into the second network to obtain the adjustment vector for each predetermined shape target object; as well as The first image features are input into the third network to obtain the initial category of each predetermined shape target object.

5. The image processing method according to claim 4, wherein inputting the first image features into the first network to obtain the initial center point of each predetermined shape target object comprises: The first image features are input into the first network to obtain the probability that each pixel in each predetermined shape target object is the initial center point, and the initial center point of each predetermined shape target object is determined according to the probability.

6. The image processing method according to claim 5, further comprising, after determining the initial center point of each predetermined shape target object according to the probability: The initial center point region of each predetermined shape target object is determined based on each pixel, the initial center point, and the preset radius.

7. The image processing method according to claim 6, wherein inputting the first image features into the second network to obtain the adjustment vector of each predetermined shape target object comprises: The first image features are input into a relation vector network to obtain the relation vector of the initial center points of two adjacent target objects of a predetermined shape. as well as The first image features are input into the offset vector network to obtain the offset vector of each pixel in the initial center point region of each predetermined shape target object, corresponding to the initial center point of each predetermined shape target object.

8. The image processing method according to claim 7, wherein determining the candidate center point based on each pixel value in the initial center region of each predetermined shape target object and the offset vector comprises: The value of each pixel in the initial center region of each predetermined shape target object and the offset vector are input into a preset calculation formula to obtain the probability that each pixel in the initial center region of each predetermined shape target object is the center point, and the pixel with the highest probability is taken as the candidate center point of each predetermined shape target object.

9. The image processing method according to claim 1, wherein adjusting the initial category of each predetermined shape target object based on the connection relationship of the target center point to obtain the target category of each predetermined shape target object includes: Based on the connection relationship of the target center point, the initial category of each predetermined shape target object is adjusted using the maximum likelihood algorithm to obtain the target category and target label of each predetermined shape target object.

10. The image processing method according to claim 6, wherein the third task is implemented by a fourth network; Accordingly, determining the binarization result of each predetermined shape target object based on the image features includes: The second image features are input into the fourth network to obtain the binarization result of each predetermined shape target object.

11. The image processing method according to claim 10, wherein segmenting each predetermined shape target object based on the binarization result, target category, target label, and target center point of each predetermined shape target object comprises: Obtain the coordinates of each pixel after binarization of each predetermined shape target object, and determine the nearest target center point of each pixel based on the coordinates of each pixel, the offset vector, and the target center point of each predetermined shape target object; Based on the target category and target label of the nearest target center point of each pixel, the target category and target label of each pixel are determined, and each predetermined shape target object is segmented according to the target category and target label of each pixel.

12. A medical image processing method, comprising: Acquire medical images containing vertebral bodies and extract image features from the medical images; Based on the image features, the initial center point, adjustment vector, and initial category of each vertebra are determined. The adjustment vector is determined according to the offset vector and the relation vector. The offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae. The initial center point of each vertebra is adjusted according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the adjustment of the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point includes: determining a candidate center point based on each pixel value in the initial center region of each vertebra and the offset vector; and obtaining the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The initial category of each vertebra is adjusted based on the connection relationship of the target center point to obtain the target category and target label of each vertebra.

13. The medical image processing method according to claim 12, further comprising, after obtaining the target category and target label of each vertebra, the method includes: The binarization result of each vertebra is determined based on the image features, and each vertebra is segmented based on the binarization result of each vertebra, the target category, the target label, and the target center point.

14. The medical image processing method according to claim 13, wherein the image feature extraction of the medical image comprises: The medical image is initially extracted based on a feature extraction model to obtain the initial image features; The initial image features are convolved through multiple convolutional layers to obtain the convolved image features. The convolutional image features are extracted again according to a preset task to obtain first image features and second image features. The preset task includes a first task, a second task and a third task.

15. The medical image processing method according to claim 14, wherein the first task is implemented by a first network and a second network, and the second task is implemented by a third network; Accordingly, determining the initial center point, adjustment vector, and initial category of each vertebra based on the image features includes: The first image features are input into the first network to obtain the initial center point of each vertebra; The first image features are input into the second network to obtain the adjustment vector for each vertebra; as well as The first image features are input into the third network to obtain the initial category of each vertebra.

16. The medical image processing method according to claim 15, wherein inputting the first image features into the first network to obtain the initial center point of each vertebral body comprises: The first image features are input into the first network to obtain the probability that each pixel in each vertebra is the initial center point, and the initial center point of each vertebra is determined according to the probability.

17. The medical image processing method according to claim 16, further comprising, after determining the initial center point of each vertebra based on the probability: The initial center point region of each vertebra is determined based on each pixel in each vertebra, the initial center point, and the preset radius.

18. The medical image processing method according to claim 17, wherein inputting the first image features into the second network to obtain the adjustment vector for each vertebral body comprises: The first image features are input into a relation vector network to obtain the relation vector of the initial center points of two adjacent vertebrae. as well as The first image features are input into an offset vector network to obtain the offset vector of each pixel in the initial center point region of each vertebra corresponding to the initial center point of each vertebra.

19. The medical image processing method according to claim 18, wherein adjusting the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center points comprises: The value of each pixel in the initial central region of each vertebra and the offset vector are input into a preset calculation formula to obtain the probability that each pixel in the initial central region of each vertebra is the center point, and the pixel with the highest probability is selected as the candidate center point of each vertebra. Based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae, the target center point of each vertebra and the connection relationship of the target center point are obtained.

20. The medical image processing method according to claim 12, wherein adjusting the initial category of each vertebra based on the connectivity of the target center point to obtain the target category of each vertebra includes: Based on the connection relationship of the target center point, the initial category of each vertebra is adjusted using the maximum likelihood algorithm to obtain the target category and target label of each vertebra.

21. The medical image processing method according to claim 17, wherein the third task is implemented by a fourth network; Accordingly, determining the binarization result of each vertebra based on the image features includes: The second image features are input into the fourth network to obtain the binarization result of each vertebra.

22. The medical image processing method according to claim 21, wherein segmenting each vertebra based on the binarization result of each vertebra, the target category, the target label, and the target center point comprises: Obtain the coordinates of each pixel after binarization of each vertebra, and determine the nearest target center point of each pixel based on the coordinates of each pixel, the offset vector, and the target center point of each vertebra; The target category and target label of each pixel are determined based on the target category and target label of the nearest target center point of each pixel, and each cone is segmented according to the target category and target label of each pixel.

23. An image processing apparatus, comprising: The first image acquisition module is configured to acquire an image containing a target object of a predetermined shape and to extract image features from the image. The first determining module is configured to determine the initial center point, adjustment vector, and initial category of each predetermined shape target object based on the image features, wherein the adjustment vector is determined according to the offset vector and the relation vector, the offset vector is used to determine the precise key point coordinates of each predetermined shape target object, and the relation vector is used to determine the correct topological relationship between each predetermined shape target object; The first adjustment module is configured to adjust the initial center point of each predetermined shape target object according to the adjustment vector to obtain the target center point of each predetermined shape target object and the connection relationship of the target center point; the first adjustment module is further configured to determine the candidate center point according to each pixel value in the initial center region of each predetermined shape target object and the offset vector; and obtain the target center point of each predetermined shape target object and the connection relationship of the target center point based on the candidate center point of each predetermined shape target object and the relationship vector of the initial center points of two adjacent predetermined shape target objects. The first obtaining module is configured to adjust the initial category of each predetermined shape target object based on the connection relationship of the target center point, so as to obtain the target category and target label of each predetermined shape target object.

24. A medical image processing apparatus, comprising: The second image acquisition module is configured to acquire medical images containing vertebrae and extract image features from the medical images. The second determining module is configured to determine the initial center point, adjustment vector, and initial category of each vertebra based on the image features, wherein the adjustment vector is determined according to the offset vector and the relation vector, the offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae; The second adjustment module is configured to adjust the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the second adjustment module is further configured to determine the candidate center point according to each pixel value in the initial center region of each vertebra and the offset vector; and obtain the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The second acquisition module is configured to adjust the initial category of each vertebra based on the connection relationship of the target center point to obtain the target category and target label of each vertebra.

25. A medical image processing method, comprising: Display an image input interface to the user based on the user's request; Receive a medical image containing vertebrae input by the user based on the image input interface, and extract image features from the medical image; Based on the image features, the initial center point, adjustment vector, and initial category of each vertebra are determined. The adjustment vector is determined according to the offset vector and the relation vector. The offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae. The initial center point of each vertebra is adjusted according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the adjustment of the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point includes: determining a candidate center point based on each pixel value in the initial center region of each vertebra and the offset vector; and obtaining the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The initial category of each vertebra is adjusted based on the connection relationship of the target center point to obtain the target category and target label of each vertebra; The binarization result of each vertebra is determined based on the image features, and each vertebra is segmented based on the binarization result, target category, target label, and target center point. The segmented medical image is then returned to the user.

26. A medical image processing apparatus, comprising: The interface display module is configured to display an image input interface to the user based on the user's request. The image receiving module is configured to receive a medical image containing a vertebral body input by the user based on the image input interface, and to extract image features from the medical image; The third determining module is configured to determine the initial center point, adjustment vector, and initial category of each vertebra based on the image features, wherein the adjustment vector is determined according to the offset vector and the relation vector, the offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae; The third adjustment module is configured to adjust the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the third adjustment module is further configured to determine the candidate center point according to each pixel value in the initial center region of each vertebra and the offset vector; and obtain the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The third acquisition module is configured to adjust the initial category of each vertebra based on the connection relationship of the target center point to obtain the target category and target label of each vertebra. The first segmentation module is configured to determine the binarization result of each vertebra based on the image features, segment each vertebra based on the binarization result of each vertebra, the target category, the target label, and the target center point, and return the segmented medical image to the user.

27. A medical image processing method, comprising: Receive a request from a user to retrieve a medical image containing a vertebra, and extract image features from the medical image; Based on the image features, the initial center point, adjustment vector, and initial category of each vertebra are determined. The adjustment vector is determined according to the offset vector and the relation vector. The offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae. The initial center point of each vertebra is adjusted according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the adjustment of the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point includes: determining a candidate center point based on each pixel value in the initial center region of each vertebra and the offset vector; and obtaining the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The initial category of each vertebra is adjusted based on the connection relationship of the target center point to obtain the target category and target label of each vertebra; The binarization result of each vertebra is determined based on the image features, and each vertebra is segmented based on the binarization result, target category, target label, and target center point. The segmented medical image is then returned to the user.

28. A medical image processing apparatus, comprising: The request receiving module is configured to receive a call request sent by a user carrying a medical image containing a vertebra, and to extract image features from the medical image; The fourth determining module is configured to determine the initial center point, adjustment vector, and initial category of each vertebra based on the image features, wherein the adjustment vector is determined according to the offset vector and the relation vector, the offset vector is used to determine the precise key point coordinates of each vertebra, and the relation vector is used to determine the correct topological relationship between the vertebrae; The fourth adjustment module is configured to adjust the initial center point of each vertebra according to the adjustment vector to obtain the target center point of each vertebra and the connection relationship of the target center point; the fourth adjustment module is further configured to determine the candidate center point according to each pixel value in the initial center region of each vertebra and the offset vector; and obtain the target center point of each vertebra and the connection relationship of the target center point based on the candidate center point of each vertebra and the relationship vector of the initial center points of two adjacent vertebrae. The fourth obtaining module is configured to adjust the initial category of each vertebra based on the connection relationship of the target center point to obtain the target category and target label of each vertebra. The second segmentation module is configured to determine the binarization result of each vertebra based on the image features, segment each vertebra based on the binarization result of each vertebra, the target category, the target label, and the target center point, and return the segmented medical image to the user.

29. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, wherein when the processor executes the computer-executable instructions, it implements the steps of the image processing method according to any one of claims 1-11 or the medical image processing method according to any one of claims 12-22, 25, and 27.

30. A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the image processing method according to any one of claims 1-11 or the medical image processing method according to any one of claims 12-22, 25, and 27.

31. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the image processing method according to any one of claims 1-11 or the medical image processing method according to any one of claims 12-22, 25, and 27.

Images