A medical image-based parameter acquisition model training method and device

Abstract

This application discloses a method and apparatus for training a parameter acquisition model based on medical images, comprising: firstly, segmenting the template medical image to obtain various template regions based on the number of voxels and voxel intensity values, and determining the category and physical property parameters of the template regions based on the average voxel intensity and standard deviation of voxel intensity; secondly, acquiring various sample medical images and matching each sample medical image with each template region to determine the sample regions and physical property parameters of each category in the sample medical image; finally, using the sample regions of each category as samples and their physical property parameters as labels, training the parameter acquisition model. This solves the technical problem of low efficiency in training data acquisition and low training efficiency of the parameter acquisition model due to the large amount of human intervention required for training, thereby improving the efficiency of training data acquisition and the training efficiency of the parameter acquisition model.

Description

Technical Field

[0001] This application relates to the technical field of medical imaging, specifically to a method and apparatus for training a parameter acquisition model based on medical images. Background Technology

[0002] Currently, medical simulation has gradually evolved from simple visual simulation to physical and physiological simulation. The foundation and core of physical and physiological simulation is biological soft tissue simulation modeling, and the key technology of biological soft tissue simulation modeling lies in the segmentation of medical images and the acquisition of the physical property parameters of each segmented region.

[0003] Existing technologies often employ machine learning to obtain physical property parameters from medical images. For example, a sample medical image can first be manually segmented and identified to obtain various types of segmented regions. Then, the physical property parameters of each type of segmented region are manually determined. Using these data and their corresponding physical property parameters, a parameter acquisition model can be trained. This trained model can then be used to process various medical images, thus solving the problem of obtaining physical property parameters from a large number of medical images.

[0004] However, the above scheme requires a large amount of training data to train the parameter acquisition model, and labeling the segmentation regions and physical property parameters of a large number of medical images requires a lot of human intervention, resulting in low efficiency in acquiring training data, which in turn leads to low training efficiency of the parameter acquisition model. Summary of the Invention

[0005] This application provides a method and apparatus for training a parameter acquisition model based on medical images, which solves the technical problem that training a parameter acquisition model requires a large amount of training data, and annotating the segmentation regions and physical property parameters of a large number of medical images requires a lot of human intervention, resulting in low efficiency in acquiring training data and thus low training efficiency of the parameter acquisition model.

[0006] On the one hand, a method for training a parameter acquisition model based on medical images is provided, the method comprising:

[0007] Obtain template medical images;

[0008] Based on the number of voxels and the voxel intensity value of the template medical image, the template medical image is segmented to obtain each template region;

[0009] For each template region, the category of the template region is determined based on the average voxel intensity and the standard deviation of the voxel intensity of the template region;

[0010] Based on the category of each template region, the physical property parameters of each template region are obtained respectively;

[0011] Acquire each sample medical image, and for each sample medical image, match the sample medical image with each template region in the template medical image to determine the sample regions of each category in the sample medical image, as well as the physical property parameters of the sample regions of each category.

[0012] The parameter acquisition model is trained using the sample regions of each category as samples and the physical property parameters of the sample regions of each category as labels. The trained parameter acquisition model is used to process the various categories of the target medical image to be identified in order to obtain the physical property parameters of the various categories of the regions.

[0013] On another front, a parameter acquisition model training device based on medical images is provided, the device comprising:

[0014] The template medical image acquisition module is used to acquire template medical images;

[0015] The template region acquisition module is used to segment the template medical image based on the number of voxels and the voxel intensity value to obtain each template region.

[0016] The template region category determination module is used to determine the category of each template region based on the average voxel intensity and the standard deviation of voxel intensity of the template region.

[0017] The physical property parameter acquisition module is used to acquire the physical property parameters of each template region according to the category of each template region.

[0018] The sample medical image acquisition module is used to acquire various sample medical images, and for each sample medical image, match the sample medical image with various template regions in the template medical image to determine the sample regions of various categories in the sample medical image, as well as the physical property parameters of the sample regions of various categories.

[0019] The parameter acquisition model training module is used to train the parameter acquisition model using sample regions of each category as samples and physical property parameters of sample regions of each category as annotations. The trained parameter acquisition model is used to process each category region of the target medical image to be identified in order to obtain the physical property parameters of each category region.

[0020] In one possible implementation, the physical property parameters include: density, elasticity, viscosity, surface friction, and damage threshold.

[0021] In one possible implementation, the template region acquisition module is further configured to:

[0022] Obtain the voxel intensity values ​​of the template medical image and calculate the standardized histogram of the voxel intensity values;

[0023] The standardized histogram is divided into multiple first-level segments at equal intervals;

[0024] Obtain the number of voxels in each of the first-level segments, and divide the first-level segments into multiple second-level segments based on the number of voxels;

[0025] Based on the threshold range of the multiple secondary segments, the template medical image is subjected to threshold segmentation to obtain the threshold segmentation result;

[0026] Each connected component in the threshold segmentation result is expanded, and each template region is obtained based on the expanded connected components.

[0027] In one possible implementation, the template region acquisition module is further configured to:

[0028] Obtain the seed point set of each tissue and organ in the template medical image;

[0029] Obtain the spatial division range of the seed point set, and calculate the average voxel intensity and standard deviation of each point within the spatial division range;

[0030] Based on the average voxel intensity and standard deviation of voxel intensity corresponding to each point, the template medical image is segmented by region growing within the template voxel intensity distribution range.

[0031] Obtain each template region based on the region growth segmentation results.

[0032] In one possible implementation, the template region category determination module is further configured to:

[0033] For each template region, when the voxel distribution of the template region is detected to meet the classification conditions, the category of the template region is determined based on the voxel distribution of the template region.

[0034] When the voxel distribution of the template region does not meet the classification conditions, the preliminary classification of the template region is determined based on the voxel intensity value of the template region. Based on the preliminary classification and the voxel intensity of the template region, clustering is performed on the template region and each template region in the template medical image to obtain the category of the template region. The preliminary classification includes at least one of bone, soft tissue, muscle, and fat.

[0035] In one possible implementation, the sample medical image acquisition module is further configured to:

[0036] The sample medical image is coarsely segmented using each template region in the template medical image to obtain each sample region;

[0037] Based on the contour and voxel intensity distribution of each sample region, registration is performed with each template region to obtain the category of each sample region.

[0038] In one possible implementation, the parameter acquisition model training module is further configured to:

[0039] For each sample region, obtain the mean voxel intensity of the sample region, the category of the sample region, and the standard deviation of the voxel intensity of the sample region;

[0040] The mean voxel intensity of the sample region, the category of the sample region, and the standard deviation of the voxel intensity of the sample region are processed by a parameter acquisition model to obtain prediction parameters.

[0041] The physical property parameters corresponding to the category of the sample region, as well as the prediction parameters, are processed by a loss function to obtain a loss function value, so as to update the parameters in the parameter acquisition model using the loss function value.

[0042] In another aspect, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, the at least one instruction being loaded and executed by the processor to implement the above-described method for training a parameter acquisition model based on medical images.

[0043] In another aspect, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement a parameter acquisition model training method based on medical images as described above.

[0044] The technical solution provided in this application may include the following beneficial effects:

[0045] First, based on the number of voxels and voxel intensity values ​​of the template medical image, the template medical image is segmented to obtain each template region. The category of each template region is determined based on its average voxel intensity and standard deviation. Then, the physical property parameters of each template region are obtained according to its category. Next, sample medical images are acquired, and for each sample medical image, it is matched with each template region in the template medical image to determine the sample regions of each category and their physical property parameters. Finally, using the sample regions of each category as samples and their physical property parameters as annotations, the parameter acquisition model is trained. The trained parameter acquisition model is used to process the various categories of regions in the target medical image to obtain their physical property parameters. This application first segments and classifies a single template medical image, and then uses the segmentation and classification results of that template medical image as a template to segment and classify a large number of sample medical images, thereby training a parameter acquisition model to solve the problem of acquiring physical property parameters of a large number of medical images. This does not require a lot of human intervention, and improves the efficiency of acquiring training data and the training efficiency of the parameter acquisition model. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram illustrating the structure of a parameter acquisition model training system based on medical images, according to an exemplary embodiment.

[0048] Figure 2 This is a flowchart illustrating a method for training a parameter acquisition model based on medical images, according to an exemplary embodiment.

[0049] Figure 3 This is a flowchart illustrating a method for training a parameter acquisition model based on medical images, according to an exemplary embodiment.

[0050] Figure 4 This is a flowchart illustrating an application method for a parameter acquisition model training method based on medical images, according to an exemplary embodiment.

[0051] Figure 5This is a structural block diagram of a parameter acquisition model training device based on medical images, according to an exemplary embodiment.

[0052] Figure 6 A structural block diagram of a computer device illustrated in an exemplary embodiment of this application is shown. Detailed Implementation

[0053] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0054] It should be understood that the term "instruction" mentioned in the embodiments of this application can be a direct instruction, an indirect instruction, or an indication of a relationship. For example, A instructing B can mean that A directly instructs B, such as B being able to obtain information through A; it can also mean that A indirectly instructs B, such as A instructing C, so B can obtain information through C; or it can mean that there is a relationship between A and B.

[0055] In the description of the embodiments of this application, the term "correspondence" may indicate that there is a direct or indirect correspondence between two things, or that there is an association between two things, or that there is a relationship of instruction and being instructed, configuration and being configured, etc.

[0056] Figure 1 This is a schematic diagram illustrating the structure of a parameter acquisition model training system based on medical images, according to an exemplary embodiment. The segmentation system includes a server 110 and a medical image acquisition device 120.

[0057] Optionally, the medical image acquisition device 120 can be a computed tomography (CT) scanner for acquiring three-dimensional medical images.

[0058] Optionally, the medical image acquisition device 120 can communicate with the server 110 through a transmission network (such as a wireless communication network). The medical image acquisition device 120 can upload the acquired medical images to the server 110 through the wireless communication network so that the server 110 can process, train, and model the acquired medical images.

[0059] Optionally, the server 110 can also wirelessly connect to the medical image acquisition device 120 via a wireless communication network. The server 110 can be a server cluster or a distributed system composed of multiple physical servers, or it can be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platform technologies and computing services.

[0060] Optionally, the system may also include a management device for managing the system (such as managing the connection status between the various modules in the medical image acquisition device 120 and the server), the management device being connected to the server via a communication network. Optionally, the communication network may be a wired network or a wireless network.

[0061] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any other network, including but not limited to any combination of local area networks (LANs), metropolitan area networks (MANs), wide area networks (WANs), mobile, wired or wireless networks, private networks, or virtual private networks (VPNs). In some embodiments, technologies and / or formats including Hypertext Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Furthermore, conventional encryption technologies such as Secure Sockets Layer (SSL), Transport Layer Security (TLS), VPNs, and Internet Protocol (IP) security can be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0062] Figure 2 This is a flowchart illustrating a method for training a parameter acquisition model based on medical images, according to an exemplary embodiment. The method is executed by a computer device, which may be, for example... Figure 1 Server 110 is shown in the image. Figure 2 As shown, the method may include the following steps:

[0063] Step S201: Obtain template medical images.

[0064] In one possible implementation, a template medical image is first acquired, which can be a three-dimensional medical image such as a computed tomography (CT) scan. This template medical image can be a whole-body computed tomography (CT) scan of a healthy individual, or a computed tomography (CT) scan of a target area of ​​a healthy individual; for example, the target area can be the thoracic cavity.

[0065] Optionally, the template medical image can be any medical image that requires parameter acquisition model training.

[0066] Step S202: Based on the number of voxels and voxel intensity values ​​of the template medical image, segment the template medical image to obtain each template region.

[0067] In one possible implementation, after acquiring the template medical image, the number of voxels and the voxel intensity value of the template medical image are counted. Since in actual situations, such as when the template medical image is a human chest cavity image, the appearance of various tissues and organs inside the chest cavity in the CT image is different (for example, the voxel distribution in the bone region and the heart region is obviously quite different), the CT image can be segmented into different template regions based on the voxel intensity value and the number of voxels shown in the CT image.

[0068] For example, based on the number of voxels and voxel intensity values ​​of the template medical image, threshold segmentation and region growing segmentation methods can be used to segment the template medical image and obtain each template region.

[0069] Step S203: For each template region, determine the category of the template region based on the average voxel intensity and the standard deviation of the voxel intensity.

[0070] In one possible implementation, after segmenting each template region, the average voxel intensity and standard deviation of voxel intensity in each template region are statistically analyzed to obtain the voxel intensity value coverage range. The category (anatomical structure type of tissue or organ) corresponding to the template region is determined based on the voxel intensity value coverage range.

[0071] The categories of each template region correspond to the various tissues and organs in the medical image of that template. For example, the tissues and organs can be the sternum, ribs, spine, lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, mediastinal pleura, etc., contained in the thoracic cavity.

[0072] Step S204: According to the category of each template region, obtain the physical property parameters of each template region.

[0073] In one possible implementation, after obtaining the category of each template region, since the category of each template region represents the tissue or organ corresponding to that template region, the computer device can obtain the physical property parameters corresponding to each template region by looking up a table according to the category of each template region. For example, the physical property parameters may include: density, elasticity, viscosity, surface friction, and damage threshold.

[0074] Step S205: Acquire each sample medical image, and for each sample medical image, match the sample medical image with each template region in the template medical image to determine the sample regions of each category in the sample medical image, as well as the physical property parameters of the sample regions of each category.

[0075] In one possible implementation, after obtaining the physical property parameters corresponding to each template region, multiple sample medical images are first obtained, and based on the segmentation and classification results of the template medical images, the sample medical images are matched with each template region in the template medical images, thereby determining the sample regions existing in the sample medical images and the categories of each sample region.

[0076] Similarly, since the categories corresponding to each sample region in the medical image have been determined, that is, the tissues and organs corresponding to each sample region in the medical image have been determined, the physical property parameters of each category of sample region can be obtained by looking up a table, or by obtaining the physical property parameters of the template region of the same category.

[0077] Step S206: Using the sample regions of each category as samples and the physical property parameters of the sample regions of each category as labels, train the parameter acquisition model; the trained parameter acquisition model is used to process the various categories of the target medical image to be identified in order to obtain the physical property parameters of each category of the region.

[0078] In one possible implementation, after determining the sample regions of each category in the sample medical image, and the physical property parameters of each category's sample regions, the parameter acquisition model is trained using the sample regions of each category as samples (i.e., the indices of the input parameter acquisition model) and the physical property parameters of each category's sample regions as labels (i.e., the output results of the parameter acquisition model), resulting in a trained parameter acquisition model. This trained parameter acquisition model can then be used to input and identify the various category regions of the target medical image to obtain the physical property parameters of each category's region.

[0079] In other words, through the above scheme, the computer device first generates a template medical image through segmentation and reclassification. Then, the template medical image is matched with other sample medical images to determine the category and physical property parameters of each sample region in the other sample medical images. Therefore, before training the medical image classification model, the acquisition process of each sample medical image does not require manual segmentation and annotation. Moreover, compared with commonly used automatic segmentation algorithms, the scheme in this embodiment only needs to perform the segmentation and reclassification process once to generate the template medical image. Subsequently, it is only necessary to match the template medical image with other sample medical images, without needing to perform segmentation and reclassification algorithms on each sample medical image, thereby greatly improving the generation efficiency of sample medical images (i.e., training data).

[0080] In summary, this application first segments the template medical image based on the number of voxels and voxel intensity values ​​to obtain various template regions, and then determines the category of the template region based on the average voxel intensity and standard deviation of the voxel intensity. Next, it obtains the physical property parameters of each template region based on its category. Then, it acquires various sample medical images and matches each sample medical image with the various template regions in the template medical image to determine the sample regions of each category and their physical property parameters. Finally, it trains the parameter acquisition model using the sample regions of each category as samples and their physical property parameters as annotations. The trained parameter acquisition model is used to process the various category regions of the target medical image to obtain the physical property parameters of each category region. This application first segments and classifies a single template medical image, and then uses the segmentation and classification results of that template medical image as a template to segment and classify a large number of sample medical images, thereby training a parameter acquisition model to solve the problem of acquiring physical property parameters of a large number of medical images. This does not require a lot of human intervention, and improves the efficiency of acquiring training data and the training efficiency of the parameter acquisition model.

[0081] Figure 3 This is a flowchart illustrating a method for training a parameter acquisition model based on medical images, according to an exemplary embodiment. The method is executed by a computer device, which may be, for example... Figure 1 Server 110 is shown in the image. Figure 3 As shown, the method may include the following steps:

[0082] Step S301: Obtain template medical images.

[0083] Step S302: Based on the number of voxels and voxel intensity values ​​of the template medical image, segment the template medical image to obtain each template region.

[0084] In one possible implementation, the voxel intensity values ​​of the template medical image are obtained, and a standardized histogram of the voxel intensity values ​​is plotted.

[0085] The standardized histogram is divided into multiple first-level segments at equal intervals.

[0086] Obtain the number of voxels for each primary segment, and based on the number of voxels, divide the primary segment into multiple secondary segments;

[0087] Based on the threshold range of the multiple secondary segments, the template medical image is subjected to threshold segmentation to obtain the threshold segmentation result;

[0088] The connected components in the threshold segmentation result are expanded, and the template regions are obtained based on the expanded connected components.

[0089] Furthermore, after acquiring the template medical image, the number of voxels and voxel intensity values ​​of the template medical image are first counted, and the range of voxel intensity values ​​is standardized. Then, a standardized histogram of the voxel intensity values ​​is obtained. After obtaining the standardized histogram, the standardized histogram is divided into multiple first-level segments at equal intervals. For each first-level segment, the intensity values ​​or intensity value ranges where the number of voxels is less than the target number (this target number can be set according to the actual situation) are detected. These are used as cutoff points to further divide the first-level segment into second-level segments.

[0090] After obtaining multiple secondary segments, a threshold segmentation method is applied to each secondary segment, using the upper and lower bounds of the secondary segment as the threshold range to perform threshold segmentation on the template medical image, thereby obtaining the threshold segmentation result. After obtaining the threshold segmentation result, a connected component statistics method is used to count the number of voxels in each connected component in the threshold segmentation result, removing connected components with values ​​less than a specific threshold, and then using a spherical dilation operator of a specific diameter to dilate the remaining connected components. The dilated region is taken as the segmentation result (i.e., each template region).

[0091] Each template region corresponds to a specific tissue or organ in the corresponding medical image. For example, these tissues or organs can be the sternum, ribs, spine, lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, mediastinal pleura, etc., contained within the thoracic cavity. However, the threshold segmentation method and connected component statistics method described above can only capture tissues and organs with skeletal support, such as the sternum, ribs, and spine. For tissues and organs without skeletal support, such as the lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, and mediastinal pleura, a seed point set method can be used for segmentation.

[0092] In one possible implementation, the seed point set of each tissue and organ in the template medical image is obtained;

[0093] Obtain the spatial partition range of the seed point set, and calculate the average voxel intensity and standard deviation of each point within the spatial partition range;

[0094] Based on the average voxel intensity and standard deviation of voxel intensity corresponding to each point, the template medical image is segmented by region growing within the range of template voxel intensity distribution.

[0095] Obtain each template region based on the region growth segmentation results.

[0096] Furthermore, seed point sets are obtained for each tissue and organ (tissues and organs without skeletal support, such as lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast and mediastinal pleura) in the template medical image. The spatial division range of the seed point set is manually defined, and the average voxel intensity and standard deviation of the corresponding voxel intensity are calculated for the points within the range. Within the target voxel intensity distribution range, region growing segmentation is performed to obtain the segmentation results of each tissue and organ. Within the template voxel intensity distribution range, region growing segmentation is performed on the template medical image to obtain the segmentation results of each tissue and organ (each template region).

[0097] The intensity distribution range of the template voxel is defined by the range of [mean - A*standard deviation, mean + A*standard deviation] as the upper and lower limits, with the optimal value of A being 1.92.

[0098] After acquiring the template medical images, the voxel intensity value range is standardized to ensure that the voxel intensity distribution range is consistent during subsequent processing.

[0099] Step S303: For each template region, determine the category of the template region based on the average voxel intensity and the standard deviation of the voxel intensity.

[0100] In one possible implementation, for each template region, when the voxel distribution of the template region is detected to meet the classification conditions, the category of the template region is determined based on the voxel distribution of the template region.

[0101] When the voxel distribution of the template region does not meet the classification conditions, the preliminary classification of the template region is determined based on the voxel intensity value of the template region. Based on the preliminary classification and the voxel intensity of the template region, clustering is performed on the template region and each template region in the template medical image to obtain the category of the template region. The preliminary classification includes at least one of bone, soft tissue, muscle, and fat.

[0102] In one possible implementation, after segmenting each template region, for known anatomical structures (corresponding to those that meet the above classification conditions, such as the definite anatomical structure of the template region being known and having data available on its physical properties, for example: lungs, bones, heart (atria, ventricles, etc., whose physical properties are known), their corresponding physical property parameters such as density, elasticity, viscosity, surface friction, and damage threshold can be obtained by consulting the data, and a set of parameter vectors can be formed.

[0103] For structures lacking clearly defined physical properties (corresponding to those not meeting the classification criteria, such as fat; for fat, only differentiating locations is possible, but whether specific density differences exist is unknown), a preliminary classification is performed based on their physical type (e.g., bone, soft tissue, muscle, fat) and their voxel intensity distribution. This preliminary classification is artificial; after identifying the broad category (bone, soft tissue, muscle, fat, cavity, fluid) based on density values, a classification is obtained by combining this with voxel intensity distribution, such as average voxel intensity, standard deviation of voxel intensity, maximum and minimum voxel intensity values. This classification can be predefined or based on specific segmentation criteria, without a general level range, and gradually supplemented during the actual classification process.

[0104] For each segmented template region, the absolute value of the intensity difference between the average voxel intensities and the overlap of the intensity ranges of [average voxel intensity - standard voxel intensity, average voxel intensity + standard voxel intensity] are first calculated. Then, according to the Gaussian distribution model, the weighted product of these two values ​​is used as the distance measure of the segmented region for clustering, thereby classifying the template region and obtaining its category. The specific classification is obtained through hierarchical clustering, which is more accurate and automatic than the initial manual classification mentioned above. For example, in the thoracic cavity, all cavities can be classified into one category, and all fluid regions can be classified into another, even if these regions are not connected.

[0105] Furthermore, the overlap can be obtained using the following formula:

[0106]

[0107] Where Distance represents the degree of overlap, μa represents the voxel intensity of template region A, μb represents the voxel intensity of template region B, σa represents the standard deviation of the voxel intensity of template region A, and σb represents the standard deviation of the voxel intensity of template region B.

[0108] Step S304: According to the category of each template region, obtain the physical property parameters of each template region.

[0109] In one possible implementation, the physical property parameters include: density, elasticity, viscosity, surface friction, and failure threshold.

[0110] For example, a physical property parameter table can be pre-stored in a computer device. This table records the distribution of physical property parameters of different types of tissues and organs. Once the computer device determines the category of each template region, it can query the physical property parameter table based on the category of the template region to obtain the distribution of physical properties of different types of tissues and organs.

[0111] Optionally, this table of physical property parameters can be obtained statistically based on the results of anatomical experiments.

[0112] Step S305: Acquire medical images of each sample, and for each medical image of the sample, perform coarse segmentation of the medical image of the sample through each template region in the template medical image to obtain each sample region.

[0113] In one possible implementation, after obtaining the physical property parameters corresponding to each template region, sample medical images of multiple healthy individuals are first obtained (for example, sample medical images of 1000 healthy individuals are obtained), and the segmentation and classification results of the template medical images are used as templates to perform coarse segmentation and classification on each sample medical image to obtain each sample region.

[0114] Furthermore, since human anatomy is generally quite standard, template medical images specify the locations of elements such as the heart and lungs. When performing coarse segmentation using these template images, the edges of the template image are first aligned with the edges of the sample image. This can be achieved through the sampling location information in the sample image file. Taking the heart as an example, after coarse segmentation, the actual location of the heart in the sample image will be similar to its location in the template image.

[0115] Step S306: Based on the contour and voxel intensity distribution of each sample region, register them with each template region to obtain the category of each sample region and the physical property parameters of the sample region of each category.

[0116] In one possible implementation, after coarse segmentation, based on the contour and voxel intensity distribution of each sample region, the ICP rigid registration algorithm (Iterative closest point) based on the contour point set and the PCA-SIFT matching algorithm (Scale-invariant feature transform) based on SIFT features are used to automatically match each sample region with the closest region in the template medical image, thereby obtaining a physical parameter relationship mapping table with each sample region type (i.e., each part type or each tissue / organ type), the average voxel intensity, and the standard deviation of voxel intensity as indexes, and density, elasticity, viscosity, surface friction, and damage threshold as parameters.

[0117] Step S307: Using the sample regions of each category as samples and the physical property parameters of the sample regions of each category as labels, train the parameter acquisition model; the trained parameter acquisition model is used to process the various categories of the target medical image to be identified in order to obtain the physical property parameters of each category of the region.

[0118] In one possible implementation, for each sample region, the mean voxel intensity of the sample region, the category of the sample region, and the standard deviation of the voxel intensity of the sample region are obtained;

[0119] The mean voxel intensity of the sample region, the category of the sample region, and the standard deviation of the voxel intensity of the sample region are processed by the parameter acquisition model to obtain the prediction parameters.

[0120] The physical property parameters corresponding to the category of the sample region, as well as the prediction parameters, are processed through a loss function to obtain the loss function value, so that the parameters in the model can be updated by using the loss function value.

[0121] In one possible implementation, after obtaining the physical parameter relationship mapping table, a support vector machine classifier is used. The {region type, average voxel intensity, standard deviation of voxel intensity}, {density, elasticity, viscosity, surface friction, and damage threshold} of each sample medical image data are used as training inputs to train the classifier (i.e., the parameter acquisition model). This allows the input of region type, average voxel intensity, and standard deviation of voxel intensity to provide physical parameters such as density, elasticity, viscosity, surface friction, and damage threshold.

[0122] In summary, this application first segments the template medical image based on the number of voxels and voxel intensity values ​​to obtain various template regions, and then determines the category of the template region based on the average voxel intensity and standard deviation of the voxel intensity. Next, it obtains the physical property parameters of each template region based on its category. Then, it acquires various sample medical images and matches each sample medical image with the various template regions in the template medical image to determine the sample regions of each category and their physical property parameters. Finally, it trains the parameter acquisition model using the sample regions of each category as samples and their physical property parameters as annotations. The trained parameter acquisition model is used to process the various category regions of the target medical image to obtain the physical property parameters of each category region. This application first segments and classifies a single template medical image, and then uses the segmentation and classification results of that template medical image as a template to segment and classify a large number of sample medical images, thereby training a parameter acquisition model to solve the problem of acquiring physical property parameters of a large number of medical images. This does not require a lot of human intervention, and improves the efficiency of acquiring training data and the training efficiency of the parameter acquisition model.

[0123] Figure 4 This is a flowchart illustrating an exemplary embodiment of a method for training a parameter acquisition model based on medical images. In practical application, this method uses a trained parameter acquisition model to process various three-dimensional voxel models of a target medical image to obtain the physical property parameters of each voxel model. Based on these voxel models and their physical property parameters, a physical model of the target medical image with high consistency in physical performance with that of the real human body is established. This method is executed by a computer device, which can be, for example,... Figure 1 Server 110 is shown in the image. Figure 4 As shown, the method may include the following steps:

[0124] Step S401: Obtain template medical images.

[0125] Step S402: Based on the number of voxels and voxel intensity values ​​of the template medical image, segment the template medical image to obtain each template region.

[0126] Step S403: For each template region, determine the category of the template region based on the average voxel intensity and the standard deviation of the voxel intensity of the template region.

[0127] Step S404: According to the category of each template region, obtain the physical property parameters of each template region.

[0128] Step S405: Acquire each sample medical image, and for each sample medical image, perform coarse segmentation of the sample medical image through each template region in the template medical image to obtain each sample region.

[0129] Step S406: Based on the contour and voxel intensity distribution of each sample region, register them with each template region to obtain the category of each sample region and the physical property parameters of the sample region of each category.

[0130] Step S407: Using the sample regions of each category as samples and the physical property parameters of the sample regions of each category as labels, train the parameter acquisition model; the trained parameter acquisition model is used to process the various categories of the target medical image to be identified in order to obtain the physical property parameters of each category of the region.

[0131] The specific implementation methods of steps S401 to S407 are as follows: Figure 3 The specific implementation methods of steps S301 to S307 in the illustrated embodiment are similar, and will not be repeated here.

[0132] Step S408: Acquire the target medical image.

[0133] In one possible implementation, the target medical image can be a three-dimensional medical image such as a computed tomography (CT) scan. The target medical image can be a segmented whole-body image of a healthy individual, a single-shot whole-body image of a healthy individual, a computed tomography (CT) scan of a target area of ​​a healthy individual, or a single-shot whole-body image. For example, the target area could be the chest cavity.

[0134] The target medical image is any medical image that requires physical modeling.

[0135] Step S409: Based on the number of voxels and voxel intensity values ​​of the target medical image, segment the target medical image to obtain each three-dimensional voxel model.

[0136] In one possible implementation, the voxel intensity values ​​of the target medical image are obtained, and a standardized histogram of the voxel intensity values ​​is plotted.

[0137] The standardized histogram is divided into multiple first-level segments at equal intervals.

[0138] Obtain the number of voxels for each primary segment, and based on the number of voxels, divide the primary segment into multiple secondary segments;

[0139] Based on the threshold range of these multiple secondary segments, the target medical image is subjected to threshold segmentation to obtain the threshold segmentation result;

[0140] The connected components in the threshold segmentation result are dilated, and the three-dimensional voxel models are obtained based on the dilated connected components.

[0141] Furthermore, after acquiring the target medical image, the number of voxels and voxel intensity values ​​of the target medical image are first counted, and the range of voxel intensity values ​​is standardized. Then, a standardized histogram of the voxel intensity values ​​is obtained. After obtaining the standardized histogram, the standardized histogram is divided into multiple first-level segments at equal intervals. For each first-level segment, the intensity values ​​or intensity ranges where the number of voxels is less than the target number (the target number can be set according to the actual situation) are detected. These are used as cutoff points to further divide the first-level segment into second-level segments.

[0142] After obtaining multiple secondary segments, a threshold segmentation method is applied to each segment, using the upper and lower bounds of the secondary segment as the threshold range to perform threshold segmentation on the target medical image, thus obtaining the threshold segmentation result. After obtaining the threshold segmentation result, a connected component statistics method is used to count the number of voxels in each connected component in the threshold segmentation result. Connected components with a value less than a specific threshold are removed, and the remaining connected components are dilated using a spherical dilation operator with a specific diameter. The dilated region is then taken as the segmentation result (i.e., each three-dimensional voxel model).

[0143] Each three-dimensional voxel model corresponds to a specific tissue or organ in the target medical image. For example, this tissue or organ could be the sternum, ribs, spine, lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, mediastinal pleura, etc., contained within the thoracic cavity. However, the threshold segmentation method and connected component statistics method described above can only capture tissues and organs with skeletal support, such as the sternum, ribs, and spine. For tissues and organs without skeletal support, such as the lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, and mediastinal pleura, a seed point set method can be used for segmentation.

[0144] In one possible implementation, a seed point set of each tissue and organ in the target medical image is obtained;

[0145] Obtain the spatial partition range of the seed point set, and calculate the average voxel intensity and standard deviation of each point within the spatial partition range;

[0146] Based on the average voxel intensity and standard deviation of voxel intensity corresponding to each point, the target medical image is segmented by region growing within the target voxel intensity distribution range.

[0147] Each 3D voxel model is obtained based on the region growth segmentation results.

[0148] Furthermore, seed point sets are obtained for each tissue and organ (tissues and organs without skeletal support, such as lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast and mediastinal pleura) in the target medical image. The spatial division range of the seed point set is manually defined, and the average voxel intensity and standard deviation of the voxel intensity of the points within the range are statistically analyzed. Within the target voxel intensity distribution range, region growing segmentation is performed to obtain the segmentation results of each tissue and organ. Within the target voxel intensity distribution range, region growing segmentation is performed on the target medical image to obtain the segmentation results of each tissue and organ (each three-dimensional voxel model).

[0149] The target voxel intensity distribution range is defined by upper and lower limits of [mean - A*standard deviation, mean + A*standard deviation], with the optimal value of A being 1.92.

[0150] After acquiring the target medical image, the voxel intensity value range is standardized to ensure that the voxel intensity distribution range is consistent during subsequent processing.

[0151] Step S410: For each three-dimensional voxel model, obtain the category of the three-dimensional voxel model based on the average voxel intensity and standard deviation of voxel intensity in the three-dimensional voxel model.

[0152] In one possible implementation, after segmenting the individual three-dimensional voxel models, for known anatomical structures (where the specific anatomical structure of the three-dimensional voxel model is known and its physical properties are documented, such as lungs, bones, and the heart (atria and ventricles), the corresponding physical property parameters such as density, elasticity, viscosity, surface friction, and damage threshold can be obtained by consulting relevant literature, forming a set of parameter vectors. For structures without clearly defined physical property parameters (e.g., fat; for fat, only the location can be distinguished, but whether its specific density differs), a preliminary classification is performed based on its physical type, such as bone, soft tissue, muscle, or fat, combined with its voxel intensity distribution.

[0153] For each segmented 3D voxel model, first calculate the absolute value of the intensity difference between the average voxel intensities in the 3D voxel model and the overlap of the intensity ranges of [average voxel intensity - standard voxel intensity, average voxel intensity + standard voxel intensity]. Then, according to the Gaussian distribution model, use the weighted product of the two as the distance measure of the segmented region for clustering, thereby realizing the specific classification of the 3D voxel model and obtaining the category of the 3D voxel model.

[0154] Step S411: Based on the corresponding category of each three-dimensional voxel model, obtain the physical property parameters corresponding to each three-dimensional voxel model.

[0155] Furthermore, after obtaining the category of each three-dimensional voxel model, the category, average voxel strength, and standard deviation of voxel strength of the three-dimensional voxel model are used as indexes to obtain the physical property parameters corresponding to each three-dimensional voxel model from the pre-trained classification model.

[0156] Step S412: For each three-dimensional voxel model, based on the voxel intensity value distribution of the three-dimensional voxel model, perform surface reconstruction on each three-dimensional voxel model to obtain the surface model.

[0157] In one possible implementation, for each three-dimensional voxel model, the three-dimensional voxel model is divided into sub-regions based on the voxel intensity value distribution of the three-dimensional voxel model;

[0158] Based on each sub-region, the surface of each 3D voxel model is reconstructed to obtain the surface model.

[0159] In one possible implementation, the distribution of voxel intensity values, the average voxel intensity, and the standard deviation of voxel intensity within the three-dimensional voxel model are statistically analyzed.

[0160] Based on the statistical results, within the target voxel intensity distribution range, the voxel intensity value distribution of the three-dimensional voxel model is averaged and segmented.

[0161] Based on the upper and lower bounds of the average segmentation results, the interior of the three-dimensional voxel model is segmented into multiple segments using thresholding to obtain multiple sub-regions within the three-dimensional voxel model.

[0162] In one possible implementation, the subregions are divided starting from the boundaries of adjacent subregions, and the space between all subregions is allocated to different subregions so that the union of each subregion is equal to the three-dimensional voxel model.

[0163] Furthermore, after obtaining each three-dimensional voxel model, the distribution of voxel intensity values, the average voxel intensity, and the standard deviation of voxel intensity within each three-dimensional voxel model are statistically analyzed. Based on the statistical results, within the target voxel intensity distribution range, the voxel intensity value distribution of the three-dimensional voxel model is averaged and segmented to obtain the average segmentation result (for example, the average segmentation can divide the voxel intensity value distribution into 5 segments).

[0164] After obtaining the average segmentation result, the interior of the 3D voxel model is first segmented into multiple sub-regions based on the upper and lower bounds of the average segmentation result. Then, using the boundaries of adjacent sub-regions as starting points, a graph cut algorithm is used to segment the model, allocating the space between all sub-regions to different sub-regions so that the union of all sub-regions equals the 3D voxel model.

[0165] After allocating the space between all sub-regions to different sub-regions, based on the physical property parameters corresponding to the three-dimensional voxel model, the viscosity relationship between adjacent parts of each sub-region is set according to the average voxel strength of the two, thereby realizing the surface reconstruction of each three-dimensional voxel model and obtaining the surface model.

[0166] The target voxel intensity distribution range is defined by upper and lower limits of [mean - A*standard deviation, mean + A*standard deviation], with the optimal value of A being 1.92.

[0167] Step S413: Based on the surface model, each three-dimensional voxel model, and the physical property parameters of each three-dimensional voxel model, establish the physical model corresponding to the target medical image.

[0168] In one possible implementation, after acquiring the surface model, based on the individual three-dimensional voxel models and their physical property parameters, a corresponding physical model is established in a three-dimensional visualization system equipped with a physics engine. That is, a physical model is established in which the target medical image has a high degree of consistency with the real human body in terms of physical performance.

[0169] The following simple example illustrates the content disclosed in the above embodiments:

[0170] Taking medical imaging of the human thoracic cavity as an example, the tissues and organs within the thoracic cavity include the sternum, ribs, spine, lung parenchyma, trachea and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, and mediastinal pleura. The training of a parameter acquisition model and the construction of a physical model for medical images of the human thoracic cavity include the following steps:

[0171] 1) Obtain CT medical images of the chest cavity of any healthy person.

[0172] 2) The voxel intensity distribution of the CT medical image is standardized to [0, 1024].

[0173] 3) Obtain the standardized histogram of the CT medical image and divide the standardized histogram of the CT medical image into 20 equal parts to obtain the first-level segment. The voxel intensity range of each segment is approximately 50.

[0174] 4) Detect points in each first-level segment where the number of voxels is less than 30 and use them as cutoff points. When there are multiple consecutive cutoff points, a cutoff range is formed. Using these cutoff points and cutoff ranges as intervals, the first-level segment is further divided into second-level segments.

[0175] 5) Threshold segmentation is performed based on the threshold range of each secondary segment, and the volume of each connected component in the segmentation result is calculated. Connected components with a volume less than 200 are removed, and the remaining connected components are dilated using a spherical operator with a diameter of 5. The dilated regions are then segmented to identify the sternum, ribs, and spine. Seed point sets for organs and bronchi, heart, blood vessels, diaphragm, pectoralis major muscle, breast, mediastinal pleura, and cavities are obtained. The spatial range of the seed point set is manually defined, and the mean and standard deviation of the corresponding voxel intensity values ​​of the points within the range are calculated. Using [mean - 1.92 * standard deviation, mean + 1.92 * standard deviation] as the upper and lower limits, region growing segmentation is performed to obtain the segmentation results for each tissue and organ.

[0176] 6) For the segmentation results of various tissues and organs, for known anatomical structures, their corresponding physical properties such as density, elasticity, viscosity, surface friction, and damage threshold can be obtained by consulting relevant literature, forming a set of parameter vectors. For structures without clearly defined physical property parameters, preliminary classification can be performed based on their physical type, such as bone, soft tissue, muscle, and fat, combined with their voxel intensity distribution.

[0177] 7) Based on the segmentation results of each tissue and organ, clustering is performed according to the statistical voxel intensity average value and the Gaussian distribution model. The weighted product of the absolute value of the intensity difference between the voxel intensity average values ​​and the overlap of the intensity ranges of [mean value - standard deviation, mean value + standard deviation] is used as the distance measure of the segmented region to obtain the specific category of each segmented region.

[0178] 8) Using the segmentation and classification results established in the above steps as templates, segment CT images of 1000 healthy individuals. Then, use the ICP rigid registration algorithm based on contour point sets and the PCA-SIFT matching algorithm based on SIFT features to automatically match each segmented region obtained from the 1000 data with the closest region in the template. This will result in a physical parameter mapping table with each part type, average voxel intensity, and standard deviation of voxel intensity as indexes, and density, elasticity, viscosity, surface friction, and damage threshold as parameters.

[0179] 9) Using a support vector machine classifier, the classifier is trained by taking the {type of each part, average voxel strength, standard deviation of voxel strength} and {density, elasticity, viscosity, surface friction, and damage threshold} of 1000 data points as training inputs. After inputting the type, average voxel strength, and standard deviation of voxel strength, the classifier can provide density, elasticity, viscosity, surface friction, and damage threshold (i.e., train the parameter acquisition model).

[0180] 10) For new CT medical images, after voxel intensity standardization, the segmentation results of each tissue and organ are obtained by following the 3-5 steps, that is, multiple three-dimensional voxel models are obtained.

[0181] 11) For the interior of each three-dimensional voxel model, statistically analyze the distribution of voxel intensity values, statistically analyze the average voxel intensity and standard deviation of voxel intensity, obtain the voxel intensity value of the tissue center, and divide the distribution into 5 segments from "mean - 1.92 * standard deviation" to "mean + 1.92 * standard deviation". According to the upper and lower boundaries of the segments, perform multi-segment threshold segmentation on the interior of the tissue organ model to obtain multiple regions inside each three-dimensional voxel model.

[0182] 12) Using the boundaries of adjacent regions as the starting point, the GraphCut algorithm is used for segmentation. The two adjacent regions are used as the starting and ending points respectively, so that the space between all regions is allocated to different regions, and the union of all regions is equal to the model of the tissue organ (i.e., the three-dimensional voxel model).

[0183] 13) Using the established SVM classifier, the physical property parameters of each internal region type, the average voxel intensity, and the standard deviation of voxel intensity are used as inputs. The viscosity relationship between adjacent parts of the internal region is set according to the average voxel intensity of the two.

[0184] 14) Through the above steps, the physical model of a new CT medical image is automatically constructed, which includes physical models of various tissues and organs in the thoracic cavity and the internal structures of these tissues and organs. This model and the corresponding voxel model are then imported into a 3D visualization system with a physics engine to establish a simulation physical model.

[0185] In summary, this application first segments the template medical image based on the number of voxels and voxel intensity values ​​to obtain various template regions, and then determines the category of the template region based on the average voxel intensity and standard deviation of the voxel intensity. Next, it obtains the physical property parameters of each template region based on its category. Then, it acquires various sample medical images and matches each sample medical image with the various template regions in the template medical image to determine the sample regions of each category and their physical property parameters. Finally, it trains the parameter acquisition model using the sample regions of each category as samples and their physical property parameters as annotations. The trained parameter acquisition model is used to process the various category regions of the target medical image to obtain the physical property parameters of each category region. This application first segments and classifies a single template medical image, and then uses the segmentation and classification results of that template medical image as a template to segment and classify a large number of sample medical images, thereby training a parameter acquisition model to solve the problem of acquiring physical property parameters of a large number of medical images. This does not require a lot of human intervention, and improves the efficiency of acquiring training data and the training efficiency of the parameter acquisition model.

[0186] Figure 5 This is a structural block diagram illustrating a parameter acquisition model training device based on medical images, according to an exemplary embodiment. The device includes:

[0187] Template medical image acquisition module 501 is used to acquire template medical images;

[0188] The template region acquisition module 502 is used to segment the template medical image based on the number of voxels and the voxel intensity value to obtain each template region.

[0189] Template region category determination module 503 is used to determine the category of each template region based on the average voxel intensity and the standard deviation of voxel intensity of the template region.

[0190] The physical property parameter acquisition module 504 is used to acquire the physical property parameters of each template region according to the category of each template region;

[0191] The sample medical image acquisition module 505 is used to acquire each sample medical image, and for each sample medical image, match the sample medical image with each template region in the template medical image to determine the sample regions of each category in the sample medical image, as well as the physical property parameters of the sample regions of each category.

[0192] The parameter acquisition model training module 506 is used to train the parameter acquisition model using the sample regions of each category as samples and the physical property parameters of the sample regions of each category as annotations. The trained parameter acquisition model is used to process the various categories of the target medical image to be identified in order to obtain the physical property parameters of each category of the region.

[0193] In one possible implementation, the physical property parameters include: density, elasticity, viscosity, surface friction, and failure threshold.

[0194] In one possible implementation, the template area acquisition module 502 is further configured to:

[0195] Obtain the voxel intensity values ​​of the medical image template and generate a standardized histogram of the voxel intensity values;

[0196] The standardized histogram is divided into multiple first-level segments at equal intervals.

[0197] Obtain the number of voxels for each primary segment, and based on the number of voxels, divide the primary segment into multiple secondary segments;

[0198] Based on the threshold range of the multiple secondary segments, the template medical image is subjected to threshold segmentation to obtain the threshold segmentation result;

[0199] The connected components in the threshold segmentation result are expanded, and the template regions are obtained based on the expanded connected components.

[0200] In one possible implementation, the template area acquisition module 502 is further configured to:

[0201] Obtain the seed point set of each tissue and organ in the medical image template;

[0202] Obtain the spatial partition range of the seed point set, and calculate the average voxel intensity and standard deviation of each point within the spatial partition range;

[0203] Based on the average voxel intensity and standard deviation of voxel intensity corresponding to each point, the template medical image is segmented by region growing within the range of template voxel intensity distribution.

[0204] Obtain each template region based on the region growth segmentation results.

[0205] In one possible implementation, the template region category determination module 503 is further configured to:

[0206] For each template region, when the voxel distribution of the template region is detected to meet the classification conditions, the category of the template region is determined based on the voxel distribution of the template region.

[0207] When the voxel distribution of the template region does not meet the classification conditions, the preliminary classification of the template region is determined based on the voxel intensity value of the template region. Based on the preliminary classification and the voxel intensity of the template region, clustering is performed on the template region and each template region in the template medical image to obtain the category of the template region. The preliminary classification includes at least one of bone, soft tissue, muscle, and fat.

[0208] In one possible implementation, the sample medical image acquisition module 505 is further configured to:

[0209] The sample medical image is coarsely segmented using each template region in the template medical image to obtain each sample region.

[0210] Based on the contour and voxel intensity distribution of each sample region, registration is performed with each template region to obtain the category of each sample region.

[0211] In one possible implementation, the parameter acquisition model training module 506 is also used for:

[0212] For each sample region, obtain the mean voxel intensity of that sample region, the category of that sample region, and the standard deviation of the voxel intensity of that sample region;

[0213] The mean voxel intensity of the sample region, the category of the sample region, and the standard deviation of the voxel intensity of the sample region are processed by the parameter acquisition model to obtain the prediction parameters.

[0214] The physical property parameters corresponding to the category of the sample region, as well as the prediction parameters, are processed through a loss function to obtain the loss function value, so that the parameters in the model can be updated by using the loss function value.

[0215] In summary, this application first segments the template medical image based on the number of voxels and voxel intensity values ​​to obtain various template regions, and then determines the category of the template region based on the average voxel intensity and standard deviation of the voxel intensity. Next, it obtains the physical property parameters of each template region based on its category. Then, it acquires various sample medical images and matches each sample medical image with the various template regions in the template medical image to determine the sample regions of each category and their physical property parameters. Finally, it trains the parameter acquisition model using the sample regions of each category as samples and their physical property parameters as annotations. The trained parameter acquisition model is used to process the various category regions of the target medical image to obtain the physical property parameters of each category region. This application first segments and classifies a single template medical image, and then uses the segmentation and classification results of that template medical image as a template to segment and classify a large number of sample medical images, thereby training a parameter acquisition model to solve the problem of acquiring physical property parameters of a large number of medical images. This does not require a lot of human intervention, and improves the efficiency of acquiring training data and the training efficiency of the parameter acquisition model.

[0216] Figure 6 This illustration shows a structural block diagram of a computer device according to an exemplary embodiment of this application. The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the aforementioned method for training a parameter acquisition model based on medical images.

[0217] The processor can be a central processing unit (CPU). It can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations thereof.

[0218] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of this invention. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the methods described in the above embodiments.

[0219] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0220] One embodiment of this application also provides a computer storage medium for storing a computer program, which, when executed by a processor, implements the above-described method for training a parameter acquisition model based on medical images.

[0221] Those skilled in the art will understand that all or part of the processes in the above-described embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0222] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for training a parameter acquisition model based on medical images, characterized in that, The method comprises: acquiring a template medical image; segmenting the template medical image based on the number of voxels and the voxel intensity values of the template medical image to obtain each template region; for each template region, determining the category of the template region according to the average value of the voxel intensity and the standard deviation of the voxel intensity of the template region; acquiring the physical property parameters of each template region according to the category of each template region; acquiring each sample medical image, and for each sample medical image, matching the sample medical image with each template region in the template medical image to determine the sample region of each category in the sample medical image and the physical property parameters of the sample region of each category; training a parameter acquisition model with the sample region of each category as the sample and the physical property parameters of the sample region of each category as the label; the trained parameter acquisition model is used to process each category region of a target to-be-identified medical image to obtain the physical property parameters of each category region; the physical property parameters include density, elasticity, viscosity, surface friction and damage threshold; wherein the training of the parameter acquisition model with the sample region of each category as the sample and the physical property parameters of the sample region of each category as the label comprises: for each sample region, acquiring the average value of the voxel intensity of the sample region, the category of the sample region and the standard deviation of the voxel intensity of the sample region; processing the average value of the voxel intensity of the sample region, the category of the sample region and the standard deviation of the voxel intensity of the sample region through the parameter acquisition model to obtain predicted parameters; processing the physical property parameters corresponding to the category of the sample region and the predicted parameters through a loss function to obtain a loss function value, so as to update the parameters in the parameter acquisition model through the loss function value; the parameter acquisition model is a support vector machine classifier.

2. The method of claim 1, wherein, The segmentation of the template medical image based on the number of voxels and the voxel intensity values of the template medical image to obtain each template region comprises: acquiring the voxel intensity values of the template medical image and counting the normalized histogram of the voxel intensity values; dividing the normalized histogram into multiple first segments at equal intervals; acquiring the number of voxels of each first segment and dividing the first segment into multiple second segments based on the number of voxels; threshold segmenting the template medical image according to the threshold range of the multiple second segments to obtain a threshold segmentation result; dilating each connected domain in the threshold segmentation result and acquiring each template region according to each dilated connected domain.

3. The method of claim 1, wherein, The segmentation of the template medical image based on the number of voxels and the voxel intensity values of the template medical image to obtain each template region comprises: acquiring a seed point set of each tissue organ in the template medical image; acquire a spatial division range of the seed point set, and count average values and standard deviations of voxel intensities corresponding to each point in the spatial division range; based on the average values and standard deviations of voxel intensities corresponding to each point, perform region growing segmentation on the template medical image in a template voxel intensity distribution range; acquire each template region according to the region growing segmentation result.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: determining a category of each template region according to the average value and the standard deviation of voxel intensities of the template region. When it is detected that the voxel distribution of the template region satisfies a classification condition, the category of the template region is determined according to the voxel distribution of the template region. When it is detected that the voxel distribution of the template region does not satisfy the classification condition, a preliminary classification of the template region is determined according to the voxel intensity value of the template region, and a clustering process is performed on the template region and each template region in the template medical image according to the preliminary classification of the template region and the voxel intensity of the template region, so as to obtain the category of the template region. The preliminary classification includes at least one of bone, soft tissue, muscle and fat.

5. The method according to any one of claims 1 to 3, characterized in that, The method further includes: performing coarse segmentation on the sample medical image through each template region in the template medical image, so as to obtain each sample region; and respectively registering each sample region with each template region based on the contour and the voxel intensity distribution of each sample region, so as to obtain the category of each sample region. The device includes: a template medical image acquisition module configured to acquire a template medical image; 6.A medical image-based parameter acquisition model training apparatus, characterized by comprising: a template region acquisition module configured to segment the template medical image based on a voxel number and a voxel intensity value of the template medical image, and obtain each template region; a template region category determination module configured to determine a category of each template region according to an average value and a standard deviation of voxel intensities of the template region; a physical property parameter acquisition module configured to acquire a physical property parameter of each template region according to the category of each template region; a sample medical image acquisition module configured to acquire each sample medical image, match the sample medical image with each template region in the template medical image for each sample medical image, determine a sample region of each category in the sample medical image, and acquire a physical property parameter of the sample region of each category; a parameter acquisition model training module configured to train a parameter acquisition model by taking the sample region of each category as a sample and taking the physical property parameter of the sample region of each category as a label. The trained parameter acquisition model is configured to process each category region of a target to-be-identified medical image, so as to obtain a physical property parameter of each category region. The physical property parameter includes density, elasticity, viscosity, surface friction and breaking threshold. ​ ​ The parameter acquisition model training module is specifically configured to: For each sample region, obtain a voxel intensity mean value of the sample region, a category of the sample region, and a voxel intensity standard deviation of the sample region; The voxel intensity mean value of the sample region, the category of the sample region, and the voxel intensity standard deviation of the sample region are processed by the parameter acquisition model to obtain a prediction parameter; The physical property parameter corresponding to the category of the sample region and the prediction parameter are processed by a loss function to obtain a loss function value, so as to update the parameters in the parameter acquisition model through the loss function value; The parameter acquisition model is a support vector machine classifier.

7. A computer device, comprising: The computer device includes a processor and a memory, and the memory stores at least one instruction, which is loaded and executed by the processor to implement the parameter acquisition model training method based on medical images according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by the processor to implement the parameter acquisition model training method based on medical images according to any one of claims 1 to 5.

Images