Biological tissue model structure optimization model, method, device and medium

Abstract

The invention discloses a biological tissue model structure optimization model, method and device and a medium, and the optimization model is constructed through the following steps: obtaining a learning sample set, the learning sample set comprises a plurality of learning samples, and each learning sample comprises the space coordinates of all points extracted from a biological tissue model sample; a basic model is constructed in advance, the basic model comprises a deformation network and a template network, the deformation network is configured to transform a learning sample into a preset three-dimensional space to obtain a predicted biological tissue model, and the template network is configured to determine an SDF value of the predicted biological tissue model; inputting the learning sample set into the basic model, and performing optimization training on the basic model by using a preset total loss function to obtain a biological tissue model structure optimization model; the total loss function is associated with predicting SDF loss and / or geometric gradient loss of the biological tissue model. According to the method, defects and deformation in the biological tissue model structure can be predicted and repaired, and the precision of the model structure is improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of medicine and computer simulation, and particularly to an optimized model, method, device and medium for the structure of a biological tissue model. Background Art

[0002] Interventional surgery has been widely used in the treatment of various diseases due to its minimally invasive nature, and is particularly important in the treatment of vascular diseases (such as aneurysms, arteriosclerosis, cancer-related vascular reconstruction, etc.). Precise vascular structure reconstruction is the key to ensuring surgical safety and improving the treatment effect. However, existing vascular structure optimization technologies still face many challenges, such as the reduction of vascular diameter caused by segmentation algorithms, the decrease of image resolution, and the accuracy problem of discrete voxel processing. In clinical applications, these problems may lead to large errors and discontinuities in the obtained vascular models, thus affecting the accuracy of interventional treatment and increasing the operation burden on doctors.

[0003] Existing methods usually complete the modeling of vascular structures by processing image stacks layer by layer, so the shape is inevitably affected by the resolution of discrete voxels. In addition, as the vascular diameter gradually decreases, the clarity of CT images also significantly decreases. These factors result in a rougher vascular structure being established, which is vulnerable to spiky noise interference. Although the Laplacian Smoothing method can reduce some noise, it cannot guarantee the high fidelity of the vascular surface while eliminating spiky noise interference.

[0004] With the development of Convolutional Neural Networks (CNN), automated vascular structure optimization methods have been widely used to reduce the workload of manual labor. These methods can infer intraoperative data of the same patient using preoperative data, thus reducing the burden on doctors. However, methods based on CNN usually still model vascular structures in the discrete voxel space, so they still face the above problems. In addition, these methods are prone to breakage problems during the inference process, resulting in discontinuities in the vascular tree, which in turn interferes with preoperative path planning algorithms and may even lead interventional doctors to draw incorrect conclusions.

[0005] The disclosure of the above background art content is only for assisting in understanding the inventive concept and technical solution of the present invention, and it does not necessarily belong to the prior art of this application, nor does it necessarily provide technical guidance; without clear evidence showing that the above content was publicly available before the filing date of this application, the above background art should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0006] The objective of the present invention is to provide an optimized model, method, device and medium for a biological tissue model structure. By introducing the joint modeling of an implicit template and a spatial transformation function, it can more accurately capture the complex geometric features of the biological tissue structure, predict and repair the defects and deformations in the biological tissue model structure, and thus improve the accuracy of the surface reconstruction of the biological tissue model structure.

[0007] To achieve the above objective, the technical solution adopted by the present invention is as follows:

[0008] An optimized model for a biological tissue model structure is constructed in the following manner:

[0009] Obtain a learning sample set, where the learning sample set includes multiple learning samples, and each learning sample includes the spatial coordinates of all points extracted from a biological tissue model sample; the points on the biological tissue model sample are usually the points on the surface of the biological tissue model, also known as vertices or surface vertices, and each of the biological tissue models includes multiple points / vertices;

[0010] Pre-construct a neural network basic model, where the basic model includes a deformation network and a template network. The deformation network is configured to transform the learning samples into a preset three-dimensional space to obtain a predicted biological tissue model, and the template network is configured to determine the SDF value (signed distance function value) of the predicted biological tissue model;

[0011] Input the learning sample set into the basic model, and use a preset total loss function to optimize and train the basic model to obtain an optimized model for the biological tissue model structure; the total loss function is related to the SDF loss and / or geometric gradient loss of the predicted biological tissue model.

[0012] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the expression of the total loss function is:

[0013] L sdf = k1L sdf-val + k2L sdf-geo

[0014] Wherein, L sdf represents the total loss function, L sdf-wal represents the SDF loss of the predicted biological tissue model, L sdf-geo represents the geometric gradient loss of the predicted biological tissue model, k1 is the first coefficient, k1 > 0, and k2 is the second coefficient, k2 > 0.

[0015] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the SDF loss L of the predicted biological tissue model is calculated by the following formula sdf-wal :

[0016]

[0017] Among them, Ω i represents the entire space of the i-th predicted biological tissue model, S i represents the surface of the i-th predicted biological tissue model, p represents the spatial coordinates of a point, α i represents the feature vector of the i-th biological tissue model sample, β represents the feature vector of a class of biological tissue model samples, s' represents the true SDF value, F(p; α i , β) represents the predicted SDF value for p, φ(F(p; α i , β)) = exp(-δ·||F(p; α i , β)||) is the penalty for predicting the surface points of non-biological tissue models, aiming to make its SDF value close to zero, where δ >> 1, ω s represents the weight of the SDF error value in L sdf-wal , ω φ represents the weight of the SDF spatial distribution in L sdf-wal .

[0018] Furthermore, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the geometric gradient loss L of the predicted biological tissue model is calculated by the following formula sdf-geo :

[0019]

[0020] Among them, Ω i represents the entire space of the i-th predicted biological tissue model, S i represents the surface of the i-th predicted biological tissue model, p represents the spatial coordinates of a point, n' represents the normal vector of the plane where the surface p point of the biological tissue model sample is located, S cos represents the cosine similarity, which is used to measure the consistency between the predicted gradient and the true normal vector, represents the gradient direction of the surface p point of the biological tissue model sample, represents that the step size of the gradient is consistent with the modulus of the unit normal vector, ω n represents the coefficient of the cosine similarity, ω Eik represents the coefficient of the amplitude.

[0021] Furthermore, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, k1 = 1, k2 = 1.

[0022] Furthermore, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the optimization training of the basic model is further included by the following method:

[0023] Configure a latent encoding for the learning sample. The deformation network is configured to transform the learning sample to a preset three-dimensional space based on the latent encoding to obtain a predicted biological tissue model. The template network is configured to determine the SDF value of the predicted biological tissue model based on the latent encoding, and during each round of training, optimize the latent encoding. The latent encoding is related to the shape features of the learning sample. Specifically, the latent encoding is a high-dimensional vector representing the shape features of the learning sample. At the initial stage of training, the latent encoding is a randomly initialized high-dimensional vector; during each round of training, the latent encoding is optimized. After training is completed, the latent encoding is a high-dimensional vector that accurately represents the learning sample. By introducing the latent encoding, the trained deformation network and template network can accurately learn the overall structural features of the learning sample. Furthermore, in practical applications, they can also more efficiently, quickly, and accurately determine the overall structural features of the 3D biological tissue model to be optimized with defects.

[0024] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, using the deformation network to transform the learning sample to a preset three-dimensional space and using the template network to obtain the SDF value of the predicted biological tissue model according to the following formula:

[0025] F(p∈R 3 ,c)=T(D(p,c))

[0026] where F represents the signed distance function, and its output is the SDF value; p represents the spatial coordinates of a point on the biological tissue model sample; c represents the latent encoding of the learning sample; R 3 represents the three-dimensional space of the learning sample; the function D is used to map the spatial coordinates of the point p to the preset three-dimensional space to obtain the corresponding three-dimensional coordinates; T is an implicit function, which is used to return the SDF value of a point in the preset three-dimensional space.

[0027] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the deformation network includes multiple fully connected layers and non-linear activation functions. Its input is the point coordinates and latent encoding in the three-dimensional space, and its output is a three-dimensional displacement vector. The latent encoding is related to the shape features of the biological tissue model.

[0028] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the biological tissue model structure optimization model satisfies: taking a part of the learning sample as the input, its output includes not only the input but also the spatial coordinates of the points that are not included in the input but are included in the learning sample; and / or,

[0029] The preset three-dimensional space is a template space, which is a space representing an implicit template of the biological tissue model structure. Based on the template space, the deformation network and the template network perform unified modeling on different shapes.

[0030] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, the biological tissue model includes one or more of a blood vessel model, a bone model, and an organ model.

[0031] According to another aspect of the present invention, the present invention provides a method for optimizing the structure of a biological tissue model, including the following steps:

[0032] Determine input data, which includes the spatial coordinates of all vertices on the 3D biological tissue model to be optimized and an initial latent code, and the initial latent code is determined by random initialization;

[0033] Input the input data into a pre-constructed biological tissue model structure optimization model to obtain an optimized 3D biological tissue model;

[0034] Wherein, the biological tissue model structure optimization model is constructed in the following manner:

[0035] Obtain a learning sample set, which includes a plurality of learning samples, and each learning sample includes the spatial coordinates of all points extracted from a biological tissue model sample;

[0036] Pre-construct a neural network basic model, which includes a deformation network and a template network. The deformation network is configured to transform the learning sample into a preset three-dimensional space to obtain a predicted biological tissue model, and the template network is configured to determine the SDF value of the predicted biological tissue model;

[0037] Input the learning sample set into the basic model, and use a preset total loss function to optimize and train the basic model to obtain a biological tissue model structure optimization model; the total loss function is related to the SDF loss and / or geometric gradient loss of the predicted biological tissue model.

[0038] Further, based on any one of the foregoing technical solutions or a combination of multiple technical solutions, using the biological tissue model structure optimization model to optimize the input data to obtain an optimized 3D biological tissue model includes:

[0039] Based on the input data, use the trained deformation network to transform the 3D biological tissue model into a preset three-dimensional space to obtain a current 3D biological tissue optimization model, and use the trained template network to determine the SDF value of the current 3D biological tissue optimization model;

[0040] Determine whether the preset total loss function converges according to the 3D biological tissue optimization model and the SDF value. If the total loss function converges, determine the current 3D biological tissue optimization model as the optimized 3D biological tissue model;

[0041] If the total loss function does not converge, optimize the neural network parameters and the current latent encoding, and perform the next round of optimization of the 3D biological tissue model according to the optimized neural network parameters and the current latent encoding until the total loss function converges;

[0042] Among them, the current latent encoding corresponding to the first round of optimization is obtained by random initialization, and the current latent encoding corresponding to each subsequent round of optimization is determined according to the optimization result of the previous round.

[0043] Specifically, obtain the latent encoding of the 3D biological tissue model to be optimized, that is, the latent encoding of the model to be repaired. Adopt the maximum a posteriori probability estimation method in statistics, that is, perform N different random SDF space samplings on the model to be repaired, which can be used as observation values. The latent encoding that satisfies the minimum average error or the highest similarity of N observations is the latent encoding of the repaired model.

[0044] According to another aspect of the present invention, the present invention provides a biological tissue model structure optimization device. The biological tissue model structure optimization model described in any one of the above embodiments is provided in the biological tissue model structure optimization device, and the biological tissue model structure optimization device is configured to optimize and reconstruct the input 3D biological tissue model to be optimized.

[0045] According to another aspect of the present invention, the present invention provides a computer-readable storage medium for storing program instructions, and the program instructions are configured to be called by a processor to execute the steps of the method described in any one of the above technical solutions or a combination of multiple technical solutions.

[0046] The beneficial effects brought by the technical solutions provided by the present invention are as follows:

[0047] a. By introducing the joint modeling of implicit templates and spatial transformation functions, the present invention can more accurately capture the complex geometric features of biological tissue structures. When dealing with complex shapes, the traditional DeepSDF method often has difficulty accurately reflecting the subtle changes in shapes. However, in this method, by relating the geometric changes of the shape to the differences in the general template SDF, based on the deep implicit neural field and geometric gradient constraints, the broken regions in the original data are jointly modeled by the implicit template and the spatial transformation function. Through the optimization of geometric gradient constraints, the SDF values at the breaks can be accurately predicted and complemented;

[0048] b. The present invention conditionally models the implicit template and the spatial transformation function through latent encoding. The shape features extracted from the input data by the latent encoding can represent the geometric differences of different biological tissue model instances. During the training process, the parameters of the implicit template and the spatial transformation function and the latent encoding are jointly optimized to capture the global and local geometric features of the shape together, ensuring the consistency between the implicit template and the spatial transformation function and avoiding the problem of mismatch between global and local features that may occur in traditional methods, effectively improving the accuracy of surface reconstruction. Description of the Drawings

[0049] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings described below are only some embodiments recorded in the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 Flowchart for constructing a vascular structure optimization model provided for an exemplary embodiment of the present invention;

[0051] Figure 2 Schematic diagram of the architecture of a vascular structure optimization model provided for an exemplary embodiment of the present invention;

[0052] Figure 3 Vascular model diagram before repair provided for an exemplary embodiment of the present invention;

[0053] Figure 4 Vascular model diagram after repair provided for an exemplary embodiment of the present invention;

[0054] Figure 5 Flowchart of a vascular structure optimization method provided for an exemplary embodiment of the present invention;

[0055] Figure 6 Flowchart of a vascular structure optimization process provided for an exemplary embodiment of the present invention. Detailed Description of the Embodiments

[0056] In order to enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0057] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or equipment comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or equipment.

[0058] In order to overcome the problems existing in the prior art, the present application models the vascular structure by introducing an implicit function, and combines the deep implicit neural field with geometric gradient constraints to accurately optimize the surface of the vascular structure, which can improve the reconstruction accuracy of the vascular tree and provide more reliable preoperative planning data for doctors. Based on this, the present application proposes an optimization model, method, device and medium for the biological tissue model structure based on the deep implicit neural field and geometric gradient constraints, aiming to provide more efficient and accurate support for clinical interventional treatment, and has important clinical application value especially in the processing of complex vascular anatomical structures and the refined reconstruction of small blood vessels.

[0059] An implicit neural field is a method of implicitly representing complex data through a neural network. It encodes the geometric and appearance information of a scene, shape or signal into the parameters of the neural network, rather than explicitly storing it in traditional forms such as a mesh, point cloud or polygon. In other words, an implicit neural field implicitly defines the structure and appearance of data through the function mapping of a neural network.

[0060] An implicit neural field usually maps the input space (such as the position coordinates in three-dimensional space) to the output space (such as color, density or surface normal, etc.). For example, for the three-dimensional reconstruction of a scene, a function f: R 3 →R 4 can be defined, where the input is the three-dimensional coordinate P = (x, y, z) and the output is the color and density (r, g, b, σ). This function is implemented by a neural network and learns the geometry and appearance of the scene through training data.

[0061] An implicit function is a mathematical concept used to implicitly represent geometric shapes or data structures. It typically maps an input space to a scalar value or vector value, and defines geometric boundaries through the zeros (or specific values) of the function. For example, the signed distance function (SDF) is an implicit function that maps points in space to the distance value from that point to the nearest surface, where the surface is defined by the zeros. In recent years, implicit functions have shown great potential in modeling complex geometric shapes, especially when dealing with complex geometric shapes, the application of implicit functions has gradually shown great potential.

[0062] A deep implicit neural field is an extension of the implicit neural field, which combines the powerful capabilities of deep learning and uses a deep neural network to implicitly represent complex data structures. A deep implicit neural field can be regarded as a method of using a deep neural network to implement an implicit function.

[0063] A deformation network is a neural network architecture for processing geometric deformations. Its core idea is to enable the network to better adapt to the shape and position changes of the target by introducing a deformable sampling grid or convolutional operation. For example, the Deformable Convolutional Networks introduce learnable offsets in the convolutional operation, enabling the convolutional kernel to deform adaptively, thus better capturing the geometric changes of the target. The Spatial Transformer Networks (STN) enhance the network's adaptability to geometric transformations by explicitly performing spatial transformations (such as translation, rotation, scaling, etc.) on the input data.

[0064] In an embodiment of the present invention, a biological tissue model structure optimization model is provided. Taking the optimization of the vascular model structure as an example, see Figure 1 and Figure 2 , the model is constructed in the following manner:

[0065] Obtain a learning sample set, where the learning sample set includes multiple learning samples, and each learning sample includes the spatial coordinates of all points extracted from a vascular model sample. The vascular model sample is a standard vascular model. Specifically, a vascular model that conforms to the vascular structure of the human body (or other animals) and is intact can be regarded as the vascular model sample;

[0066] A basic model is pre - constructed. The basic model includes a deformation network and a template network. The deformation network is configured to transform the learning sample into a preset three - dimensional space to obtain a predicted blood vessel model. The template network is configured to determine the SDF value of the predicted blood vessel model. Among them, the preset three - dimensional space is the template space, that is, a normalized three - dimensional space, which is used to represent the geometric structure of a general shape. In this space, the geometric changes of all shapes are unified under a common reference frame (a fixed and unified coordinate system). The core idea of the template space is to represent the signed distance function (SDF) of the shape through an implicit function (i.e., the template network), so as to realize the unified modeling of different shapes.

[0067] Input the learning sample set into the basic model, and use a preset total loss function to optimize and train the basic model to obtain a blood vessel structure optimization model. The total loss function is related to the SDF loss and / or geometric gradient loss of the predicted blood vessel model.

[0068] Based on the total loss function, train the basic model so that the obtained blood vessel structure optimization model satisfies: taking a part of the learning sample as input, its output not only includes the input but also the spatial coordinates of the points that are not included in the input but are included in the learning sample. In addition, it can also correct and repair the points on the blood vessel model that deviate from the standard.

[0069] In an embodiment of the present invention, during the process of optimizing and training the basic model, a latent encoding, that is, an initial latent encoding, is pre - configured for each of the learning samples. It is a high - dimensional vector, and the initial latent encodings of different learning samples can be the same or different. In each subsequent round of training, the deformation network transforms the learning sample into a normalized three - dimensional space based on the latent encoding to obtain a predicted blood vessel model. The template network determines the SDF value of the predicted blood vessel model based on the latent encoding, and in each round of training, during the optimization process of the network parameters, the latent encoding is also optimized. For the trained model, each learning sample corresponds to a latent encoding related to its shape. This latent encoding can reveal its global shape characteristics in advance and is also the ID of each blood vessel model.

[0070] Latent Encoding is a representation method that compresses data into a low-dimensional space and is commonly used in generative models or autoencoders. Through latent encoding, complex input data (such as images, point clouds, etc.) can be mapped to a low-dimensional latent space and operations can be performed in this space. The deformation network in this application includes multiple fully connected layers and non-linear activation functions. Its input is the point coordinates in three-dimensional space and the latent encoding, and its output is a three-dimensional displacement vector. In this application, the latent encoding is related to the shape characteristics of the blood vessel model. This application uses latent encoding to represent the global features of the blood vessel model. Using the latent encoding as the input of the deformation network can help the deformation network better understand the global structure of the input data and process complex geometric models more efficiently and with higher quality.

[0071] In the classical DeepSDF method, the change of shape is directly represented by the change of the signed distance (Signed Distance Functions, SDF) value itself, that is, the signed distance function. For any point p in space, the SDF value represents the distance from this point to the nearest surface. If the point is inside the object, the value is negative; if the point is outside the object, the value is positive; if the point is exactly on the surface, the value is zero. Different from this, the inventive concept of this application is that for a class of shapes represented by SDF, their geometric changes can be reflected by the differences relative to a general template SDF.

[0072] Therefore, in this application, based on the latent encoding of each learning sample, the deformation network is used to transform the learning sample to a preset three-dimensional space, and the SDF value of the predicted blood vessel model is obtained by using the template network, which is expressed as follows:

[0073] F(p∈R 3 ,c)=T(D(p,c))

[0074] where F represents the signed distance function, and its output is the SDF value. p represents the spatial coordinates of the points on the blood vessel model sample. c represents the latent encoding of the learning sample. R 3 represents the three-dimensional space of the learning sample. The spatial transformation function D: R 3 ×X→R 3 is used to map the spatial coordinates of point p to a preset three-dimensional space to obtain the corresponding three-dimensional coordinates. The implicit function T: R 3 →R is an implicit function, which is used to return the SDF value of the points on the predicted blood vessel model in the preset three-dimensional space.

[0075] In this application, D is a conditional space transformation function that performs a space transformation on the input points based on the latent code c. Specifically, the space transformation function D is a non-linear deformation field that learns a complex mapping function through a deep neural network and can capture local geometric changes. Its mathematical form is: D(p, c) = p + Δp. The space transformation function D maps the input point p from the original space to the implicit template space. And T is an implicit function that is used to represent the general SDF (i.e., the implicit template) independent of the latent code c. Therefore, for a set of shapes, the function T() represents their general template SDF, which is subsequently referred to as the implicit template. In a specific query process, if it is necessary to calculate the signed distance value of a point p of a specific object, first, the point p is mapped to the canonical position in the implicit template through the space transformation function D, and then the canonical position is queried through the implicit template T to obtain the corresponding SDF value. In other words, the implicit template is deformed according to the latent code c to model different SDFs. The latent code is related to the shape features of the learning samples. Specifically, the latent code is a high-dimensional vector representing the shape features of the learning samples. At the initial stage of training, the latent code is a randomly initialized high-dimensional vector; the latent code is optimized in each round of training process, and after the training is completed, the latent code is a high-dimensional vector that accurately represents the learning samples shown. By introducing the latent code, the trained deformation network and template network can accurately learn the overall structural features of the learning samples, and thus, in practical applications, it is also possible to more efficiently, quickly, and accurately determine the overall structural features of the 3D biological tissue model with defects to be optimized.

[0076] In one embodiment of the present invention, the vascular structure is optimized based on the following total loss function, and the representation of the total loss function is:

[0077] L sdf = L sdf-val + L sdf-geo

[0078] where L sdf represents the total loss function, L sdf-wal represents the SDF loss of the predicted vascular model, and L sdf-geo represents the geometric gradient loss of the predicted vascular model.

[0079] In this embodiment, the SDF loss L of the predicted vascular model is calculated by the following formula sdf-wal :

[0080]

[0081] where Ω i represents the entire space of the i-th predicted vascular model, and S iDenote the surface of the $i$-th predicted vascular model, and $p$ represents the spatial coordinates of a point; $\alpha$ i Denote the feature vector of the $i$-th vascular model sample, which is randomly initialized at the beginning of training and subsequently updated and optimized by the deformation network; $\beta$ represents the feature vector of a class of vascular model samples, which is also randomly initialized at the beginning of training and subsequently updated and optimized by the template network; $s'$ represents the true SDF value, and $F(p;\alpha$ i , $\beta)$ represents the predicted SDF value for $p$; $\varphi(F(p;\alpha$ i , $\beta)) = \exp(-\delta\cdot||F(p;\alpha$ i , $\beta)||)$ is the penalty for predicting non-vascular model surface points, aiming to make its SDF value close to zero, so that the predicted vascular model pays more attention to the vicinity of the shape surface. When the predicted SDF value is larger, it means the point is farther from the surface and the loss function is smaller. When the predicted SDF value is closer to 0, it means the point is closer to the surface and the loss function is larger; where $\delta\gg1$, $\omega$ S Denote the weight for predicting the SDF of a point in $L$ sdf-wal , and $\omega$ φ Denote the weight for predicting the normal vector of a point in $L$ sdf-wal .

[0082] The SDF loss of the predicted vascular model directly supervises the SDF regression result of the predicted vascular model obtained in each round of training, and realizes more accurate surface modeling by penalizing the predicted values of non-surface points.

[0083] In this embodiment, the geometric gradient loss $L$ of the predicted vascular model is calculated by the following formula sdf-geo :

[0084]

[0085] where $\Omega$ i denotes the entire space of the $i$-th predicted vascular model, $S$ i denotes the surface of the $i$-th predicted vascular model, $p$ represents the spatial coordinates of a point, $n'$ represents the normal vector of the plane where the surface point $p$ of the vascular model sample is located, and $S$ cos denotes the cosine similarity, which is used to measure the consistency between the predicted gradient and the true normal vector, denotes the gradient direction of the surface point $p$ of the vascular model sample. When the gradient direction of the predicted vascular model is closer to the direction of the normal vector, the accuracy of the reconstructed predicted vascular model is higher; denotes that the step size of the gradient is consistent with the modulus of the unit normal vector. In other words, the gradient of the predicted model should be consistent with the normal vector in both the direction and magnitude aspects. This constraint enables the implicit template to better preserve geometric details when reconstructing complex vascular structures; $\omega$ n denotes the coefficient of the cosine similarity, $\omega$Eik The coefficient representing the amplitude. The amplitude and cosine similarity are not in the same unit of measurement. Through ω n and ω Eik reasonably balance the two to the same order of magnitude. Specifically, ω n and ω Eik are usually empirical values.

[0086] The proposed vascular optimization model in this application can jointly model the fractured regions in the original data during the reconstruction process by an implicit template and a spatial transformation function. Through the optimization of geometric gradient constraints, the SDF values at the fractures are accurately predicted and completed. Figure 3 is the original data before reconstruction, Figure 4 is the data after reconstruction. It can be seen that not only the geometric structure of the fractured region is restored on the surface of the reconstructed vascular model, but also a smooth transition with the surrounding regions is maintained, fully demonstrating the advantages of this model and method in terms of detail repair and geometric consistency. In addition, by comparing the original data and the reconstruction results, it can be observed that while maintaining the vascular topological structure, the model improves the accuracy and integrity of surface reconstruction.

[0087] It should be noted that the method of this application reconstructs all points on the input, rather than only the damaged regions. Specifically, this method jointly models all input points through an implicit template and a spatial transformation function to perform unified SDF value prediction and surface optimization. During the reconstruction process, the model automatically learns the global geometric features of the input data, including the geometric information of the damaged regions, and completes and repairs the damaged regions through geometric gradient constraints and loss functions, including correcting and repairing the deformed regions.

[0088] The innovation of this application lies in that by introducing the joint modeling of an implicit template and a spatial transformation function, it can more accurately capture the complex geometric features of the vascular structure. Traditional DeepSDF methods often have difficulty accurately reflecting the subtle changes in shape when dealing with complex shapes, while this method relates the geometric changes of the shape to the differences in the general template SDF. Conditional modeling is performed on the implicit template and the spatial transformation function through latent encoding. The latent encoding is the shape feature extracted from the input data and can represent the geometric differences of different vascular instances. During the training process, the parameters of the implicit template and the spatial transformation function are jointly optimized so that the two can work together to jointly capture the global and local geometric features of the shape. Through conditional modeling, the implicit template can adapt to various shape changes, while the spatial transformation function can capture the local geometric deformations of different instances. The joint optimization ensures the consistency between the implicit template and the spatial transformation function, avoiding the problem of global-local feature mismatch that may occur in traditional methods and effectively improving the accuracy of surface reconstruction.

[0089] In addition, the introduction of the geometric gradient can ensure that the geometric properties of the reconstructed surface are consistent with the real surface, and can also better maintain the smoothness and consistency of the vascular surface, thus having broad application prospects in the field of medical image processing and other fields.

[0090] In practical applications, this method can be used for the three-dimensional reconstruction of the vascular tree. Especially in medical image analysis, it can help doctors more accurately identify and diagnose vascular diseases. By optimizing the surface reconstruction of the vascular structure, this method not only improves the reconstruction accuracy, but also enhances the robustness and generalization ability of the model. In addition, the framework of this method has good scalability and can be applied to surface reconstruction tasks of other complex shapes, such as the processing of biological tissue models such as organs and bones or biomedical images.

[0091] To sum up, this method proposes an efficient surface optimization model for vascular structures by introducing the joint modeling of implicit templates and spatial transformation functions and combining geometric gradient constraints. While improving the surface reconstruction accuracy, this model maintains good geometric smoothness and consistency, and has important theoretical significance and practical application value.

[0092] In an embodiment of the present invention, a method for optimizing the structure of a biological tissue model is proposed. Taking the structure optimization of a vascular model as an example, see Figure 5 , the method includes the following steps:

[0093] Determine the input data, where the input data includes the spatial coordinates of all vertices on the 3D vascular model to be optimized and the latent encoding of the 3D vascular model to be optimized, and the latent encoding is related to the shape characteristics of the 3D vascular model to be optimized;

[0094] Input the input data into a pre-constructed vascular structure optimization model to obtain an optimized 3D vascular model;

[0095] Among them, the vascular structure optimization model is constructed in the following way:

[0096] Obtain a learning sample set, where the learning sample set includes multiple learning samples, and each learning sample includes the spatial coordinates of all points extracted from a vascular model sample;

[0097] Pre-construct a basic model, where the basic model includes a deformation network and a template network. The deformation network is configured to transform the learning sample into a preset three-dimensional space to obtain a predicted vascular model, and the template network is configured to determine the SDF value of the predicted vascular model;

[0098] Input the learning sample set into the basic model, and optimize and train the basic model using a preset total loss function to obtain a blood vessel structure optimization model; the total loss function is related to the SDF loss and / or geometric gradient loss of the predicted blood vessel model.

[0099] See Figure 6 , in the process of optimizing the input data using the biological tissue model structure optimization model to obtain an optimized 3D biological tissue model. Based on the input data, use the trained deformation network to transform the 3D biological tissue model into a preset three-dimensional space to obtain the current 3D biological tissue optimization model, and use the trained template network to determine the SDF value of the current 3D biological tissue optimization model.

[0100] Determine whether the preset total loss function converges according to the 3D biological tissue optimization model and the SDF value. If the total loss function converges, determine that the current 3D biological tissue optimization model is the optimized 3D biological tissue model.

[0101] If the total loss function does not converge, optimize the neural network parameters and the current latent code, and perform the next round of optimization of the 3D biological tissue model according to the optimized neural network parameters and the current latent code until the total loss function converges.

[0102] Among them, the current latent code corresponding to the first round of optimization is obtained by random initialization, and the current latent code corresponding to each subsequent round of optimization is determined according to the optimization result of the previous round. Specifically, obtain the latent code of the 3D biological tissue model to be optimized, that is, the latent code of the model to be repaired. Use the maximum a posteriori probability estimation method in statistics, that is, perform N different random SDF space samplings on the model to be repaired, which can be used as observation values. The latent code that satisfies the minimum average error or the highest similarity of N observations is the latent code of the repaired model.

[0103] In an embodiment of the present invention, a biological tissue model structure optimization device is provided. The biological tissue model structure optimization device is provided with the biological tissue model structure optimization model as described in any one of the above embodiments or a combination of multiple embodiments. The biological tissue model structure optimization device is configured to optimize and reconstruct the input 3D biological tissue model to be optimized to obtain a more accurate 3D biological tissue model with repaired damage and corrected deformation.

[0104] In an embodiment of the present invention, a computer-readable storage medium is provided for storing program instructions, and the program instructions are configured to be called by a processor to execute the steps of the method as described in any one of the above embodiments.

[0105] It should be noted that the above embodiments of the method for optimizing the biological tissue model structure, the device for optimizing the biological tissue model structure, and the computer-readable storage medium belong to the same inventive concept. By reference, all the content of the embodiments of the algorithm for optimizing the biological tissue model structure is incorporated into the embodiments of the method for optimizing the biological tissue model structure, the device for optimizing the biological tissue model structure, and the computer-readable storage medium.

[0106] It should be noted that in this document, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0107] The above are only specific embodiments of the present application. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A biological tissue model structure optimization model, characterized in that: Constructed by: Acquire a learning sample set, wherein the learning sample set includes a plurality of learning samples, and each learning sample includes the spatial coordinates of all points extracted from a biological tissue model sample; Pre-constructing a neural network basic model, the basic model includes a deformation network and a template network, the deformation network is configured to transform the learning sample into a preset three-dimensional space to obtain a predicted biological tissue model, and the template network is configured to determine the SDF value of the predicted biological tissue model; The learning sample set is input into the basic model, and the basic model is optimized and trained using a preset total loss function to obtain a biological tissue model structure optimization model; the total loss function is related to the SDF loss and / or geometric gradient loss of the predicted biological tissue model.

2. The biological tissue model structure optimization model according to claim 1, characterized in that: The total loss function is expressed as: THE s df=k1L sdf-val +k2L sdf-geo Among them, L sdf Represents the total loss function, L sdf-wal represents the SDF loss of the predicted biological tissue model, L sdf-geo It represents the geometric gradient loss of the predicted biological tissue model, k1 is the first coefficient, k1>0, k2 is the second coefficient, k2>0.

3. The biological tissue model structure optimization model according to claim 2, characterized in that: The SDF loss L of the predicted biological tissue model is calculated by the following formula: sdf-wal : Among them, Ω i represents the entire space of the i-th predicted biological tissue model, S i represents the surface of the i-th predicted biological tissue model, p represents the spatial coordinates of the point, α i represents the characteristic vector of the i-th biological tissue model sample, β represents the characteristic vector of a class of biological tissue model samples, s' represents the true SDF value, F(p; α i , β) represents the SDF value predicted for p, φ(F(p; α i ,β))=exp(-δ·||F(p;α i ,β)||) is a penalty for the prediction of the surface points of the non-biological tissue model, aiming to make its SDF value close to zero, where δ>>1, ω s Indicates that L sdf-wal The weight of the SDF error value in ,ω φ Indicates that L sdf-wal The weight of the spatial distribution of SDF in .

4. The biological tissue model structure optimization model according to claim 3, characterized in that: The geometric gradient loss L of the predicted biological tissue model is calculated by the following formula: sdf-geo : Among them, Ω i represents the entire space of the i-th predicted biological tissue model, S i represents the surface of the i-th predicted biological tissue model, p represents the spatial coordinates of the point, n' represents the normal vector of the plane where the surface p of the biological tissue model sample is located, S cos Represents cosine similarity, which is used to measure the consistency between the predicted gradient and the true normal vector. represents the gradient direction of the surface point p of the biological tissue model sample, Indicates that the gradient stride is consistent with the modulus of the unit normal vector, ω n The coefficient representing the cosine similarity, ω Eik A coefficient representing the magnitude.

5. The biological tissue model structure optimization model according to claim 2, characterized in that: k1=1, k2=1.

6. The biological tissue model structure optimization model according to claim 1, characterized in that: It also includes optimizing and training the basic model in the following ways: A latent code is configured for the learning sample, the deformation network is configured to transform the learning sample into a preset three-dimensional space based on the latent code to obtain a predicted biological tissue model, the template network is configured to determine the SDF value of the predicted biological tissue model based on the latent code, and the latent code is optimized during each round of training.

7. The biological tissue model structure optimization model according to claim 6, characterized in that: Based on the following formula, the learning sample is transformed into a preset three-dimensional space using the deformation network and the SDF value of the predicted biological tissue model is obtained using the template network: F(p∈R 3 ,c)=T(D(p,c)) Where F represents a signed distance function, whose output is the SDF value, p represents the spatial coordinates of a point on the biological tissue model sample, c represents the potential code of the learning sample, and R 3 represents the three-dimensional space of the learning sample, the function D is used to map the spatial coordinates of point p to a preset three-dimensional space to obtain the corresponding three-dimensional coordinates, and T is an implicit function used to return the SDF value of a point in the preset three-dimensional space.

8. The biological tissue model structure optimization model according to claim 6, characterized in that: The deformation network includes multiple layers of fully connected layers and nonlinear activation functions, whose input is the point coordinates in three-dimensional space and the potential code, and whose output is a three-dimensional displacement vector.

9. The biological tissue model structure optimization model according to claim 1, characterized in that: The biological tissue model structure optimization model satisfies: taking part of the learning sample as input, the output obtained includes not only the input but also the spatial coordinates of points that are not included in the input but included in the learning sample; and / or, The preset three-dimensional space is a template space, which is a space representing an implicit template of a biological tissue model structure. Based on the template space, the deformation network and the template network perform unified modeling of different shapes.

10. The biological tissue model structure optimization model according to claim 1, characterized in that: The biological tissue model includes one or more of a blood vessel model, a bone model and an organ model.

11. A method for optimizing the structure of a biological tissue model, characterized in that: The following steps are involved: Determining input data, wherein the input data includes spatial coordinates of all vertices on the 3D biological tissue model to be optimized and an initial latent code; The input data is input into a pre-constructed biological tissue model structure optimization model to obtain an optimized 3D biological tissue model; wherein the biological tissue model structure optimization model is constructed in the following manner: Acquire a learning sample set, wherein the learning sample set includes a plurality of learning samples, and each learning sample includes the spatial coordinates of all points extracted from a biological tissue model sample; Pre-constructing a neural network basic model, the basic model includes a deformation network and a template network, the deformation network is configured to transform the learning sample into a preset three-dimensional space to obtain a predicted biological tissue model, and the template network is configured to determine the SDF value of the predicted biological tissue model; The learning sample set is input into the basic model, and the basic model is optimized and trained using a preset total loss function to obtain a biological tissue model structure optimization model; the total loss function is related to the SDF loss and / or geometric gradient loss of the predicted biological tissue model.

12. The biological tissue model structure optimization method according to claim 11, characterized in that: Optimizing the input data using the biological tissue model structure optimization model to obtain an optimized 3D biological tissue model, including: Based on the input data, the 3D biological tissue model is transformed into a preset three-dimensional space using the trained deformation network to obtain a current 3D biological tissue optimization model, and the SDF value of the current 3D biological tissue optimization model is determined using the trained template network; Determining whether the preset total loss function converges according to the 3D biological tissue optimization model and the SDF value, and if the total loss function converges, determining that the current 3D biological tissue optimization model is an optimized 3D biological tissue model; If the total loss function does not converge, the neural network parameters and the current potential code are optimized, and the next round of optimization of the 3D biological tissue model is performed according to the optimized neural network parameters and the current potential code until the total loss function converges; The current potential code corresponding to the first round of optimization is obtained by random initialization, and the current potential code corresponding to each subsequent round of optimization is determined according to the optimization result of the previous round.

13. A biological tissue model structure optimization device, characterized in that: The biological tissue model structure optimization device is provided with a biological tissue model structure optimization model as described in any one of claims 1 to 10, and the biological tissue model structure optimization device is configured to optimize and reconstruct an input 3D biological tissue model to be optimized.

14. A computer-readable storage medium for storing program instructions, characterized in that: The program instructions are configured to be called by a processor to execute the steps of the biological tissue model structure optimization method according to any one of claims 11 to 12.

Images