A deep learning-based bone defect repair material adaptability prediction method

Abstract

The application discloses a kind of bone defect repair material adaptability prediction methods based on deep learning, the bone defect region and host bone region in three-dimensional CT image data are overlapped and evenly divided, and the three-dimensional visual Transformer model of multi-scale position coding is used to carry out position coding and feature extraction to multiple sub-blocks, to obtain anatomical and functional characteristics;Pathophysiological characteristics are constructed after calculating bone metabolic activity score and systemic inflammation score based on blood test data;Anatomical and functional characteristics, pathophysiological characteristics, parameter data and multi-direction multi-head cross attention layer are interacted in multiple rounds, in each round, select a modal feature as a query vector, select any modal feature in the remaining modal feature as a key vector and a value vector;The interactive features of each round are fused, and the adaptability score and risk level of bone repair material are determined by combining the attention weight in the multi-direction multi-head cross attention layer and the bone metabolic activity score.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and relates to, but is not limited to, a method for predicting the compatibility of bone defect repair materials based on deep learning. Background Technology

[0002] Bone defect repair is a common clinical need in orthopedics, widely used in scenarios involving the repair of bone tissue defects caused by trauma, tumor resection, and infection. If bone defects are not repaired in a timely manner, they can subsequently impact a patient's health and even their quality of life. The repair of bone defects heavily relies on implant materials. However, significant differences exist among patients in terms of defect location, size, blood supply, infection status, and overall health, leading to varying outcomes with the same material in different individuals. Therefore, it is crucial to scientifically predict the compatibility of bone repair materials with individual patient conditions and their potential risks.

[0003] In related technologies, on the one hand, judgment is made manually combined with clinical experience. However, this method suffers from problems such as one-sided assessment of suitability and insufficient consideration of individual differences, leading to complications such as immune rejection, mechanical mismatch, and asynchronous degradation, thus reducing postoperative bone integration rates. On the other hand, traditional machine learning models are used to analyze bone defects. However, this method has limited ability to characterize the global spatial dependencies and complex topological structures of the defect area, resulting in insufficient quantitative accuracy of anatomical features. At the same time, most assessment systems are based on only single-modal data, making it difficult to effectively integrate the high-dimensional nonlinear correlations between the patient's individual anatomical environment, physiological state, and repair material properties, resulting in one-sided predictive capabilities.

[0004] Therefore, how to efficiently and accurately predict the compatibility of bone defect repair materials has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method for predicting the compatibility of bone defect repair materials based on deep learning, which at least solves the problem that related technologies cannot efficiently and accurately predict the compatibility of bone defect repair materials.

[0006] According to a first aspect of the present invention, a deep learning-based method for predicting the adaptability of bone defect repair materials is provided, comprising:

[0007] Acquire the patient's 3D CT image data, blood test data, and parameter data of bone repair materials; and divide the target area in the preprocessed 3D CT image data into multiple sub-blocks by overlapping and uniformly dividing the target area into multiple sub-blocks; the target area includes the bone defect area and the host bone area.

[0008] The positional encoding and feature extraction of multiple sub-blocks are performed using a 3D visual Transformer model with multi-scale positional encoding to obtain the first anatomical and functional features;

[0009] Bone metabolism activity score and systemic inflammation score were calculated based on blood test data; and the first pathophysiological feature was determined based on the systemic inflammation score and bone metabolism activity score.

[0010] Based on the first anatomical and functional features, the first pathophysiological features, parameter data, and the multi-directional multi-head cross-attention layer of the multimodal Transformer model, multiple rounds of interaction are performed. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector to obtain the interaction features after each round.

[0011] Interactive features are fused to obtain fused features; and based on the fused features, the attention weights in the multi-directional multi-head cross-attention layer, and the bone metabolic activity score, the suitability score and risk level of bone repair materials are determined respectively.

[0012] Optionally, the multi-scale position-encoded 3D vision Transformer model includes a block embedding layer, multiple encoder layers, and a fully connected layer;

[0013] The three-dimensional visual Transformer model, which utilizes multi-scale positional encoding, performs positional encoding and feature extraction on multiple sub-blocks to obtain the first anatomical and functional features, including:

[0014] Multiple sub-blocks are input into the block embedding layer of the 3D vision Transformer model, and then divided by a 3D convolutional layer containing a 3D convolutional kernel of a preset size to obtain multiple local blocks. The multiple local blocks are then mapped to construct an initial feature matrix.

[0015] The three-dimensional absolute position code and the local position code are added element-wise to the initial feature matrix to obtain the first feature matrix; the three-dimensional absolute position code is the absolute spatial position of each sub-block in the target region; the local position code is the absolute spatial position of each local block in the sub-block;

[0016] Learnable classification tokens are added to the first feature matrix, and the first feature matrix with the added tokens is input into the first encoder layer to obtain the second feature matrix with fused context information; the second feature matrix is ​​then input into the second encoder layer, and so on, until the last encoder layer outputs the target feature matrix.

[0017] By using a fully connected layer to perform dimensional mapping on the feature vectors of classifiable tokens in the target feature matrix, the first anatomical and functional features are obtained.

[0018] Optionally, the blood test data includes bone metabolism indicators, inflammatory indicators, and preset routine biochemical indicators;

[0019] The bone metabolism vitality score and systemic inflammation score are calculated based on blood test data, including:

[0020] C-reactive protein and tumor necrosis factor were normalized inflammatory markers. Substituting interleukin-6 into the systemic inflammation scoring formula yields the systemic inflammation score;

[0021] Normalized bone metabolism indicators include alkaline phosphatase, osteocalcin, and... Type I collagen is substituted into the bone metabolism vitality scoring formula to obtain the bone metabolism vitality score;

[0022] The formula for the systemic inflammation scoring system is as follows:

[0023] ;

[0024] In the above formula, S For systemic inflammation scores, It is C-reactive protein. Tumor necrosis factor , It is interleukin-6;

[0025] The formula for bone metabolism vitality scoring is as follows:

[0026] ;

[0027] In the above formula, B Bone metabolic activity score, It is an alkaline phosphatase. For osteocalcin, for Type II collagen;

[0028] The determination of the first pathophysiological feature based on systemic inflammation score and bone metabolic activity score includes:

[0029] The first pathophysiological characteristic was obtained by combining systemic inflammation score, bone metabolism activity score, normalized bone metabolism index, normalized inflammation index, and normalized preset routine biochemical index.

[0030] Optionally, the multi-directional, multi-head cross-attention layer based on the first anatomical and functional features, the first pathophysiological features, parameter data, and the multimodal Transformer model performs multiple rounds of interaction. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector, resulting in the interaction features after each round, including:

[0031] The parameter data were normalized, uniquely thermally encoded, and key features were extracted to construct the first material features; and the first anatomical and functional features, the first pathophysiological features, and the first material features were unified in dimensionality to obtain the second anatomical and functional features, the second pathophysiological features, and the second material features.

[0032] The second anatomical and functional features, the second pathophysiological features, and the second material features are input into the self-attention layer of the multimodal Transformer model to obtain the third anatomical and functional features, the third pathophysiological features, and the third material features, which respectively integrate the information from the upper and lower parts.

[0033] The multimodal features, consisting of the third anatomical and functional features, the third pathophysiological features, and the third material features, are input into a multi-directional, multi-head cross-attention layer for multi-round interaction. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector, until the interaction features after each round are obtained.

[0034] Optionally, the determination of the compatibility score and risk level of bone repair materials based on fusion features, attention weights in the multi-directional multi-head cross-attention layer, and bone metabolic activity score includes:

[0035] The compatibility score of the predicted bone repair material is determined based on the fusion features; and in the second round of interaction in the multi-directional multi-head cross attention layer, the weights between inflammatory indicators and surface chemistry in the material features are extracted from the attention weight matrix; the pathophysiological features in the second round are query vectors, and the material features are key vectors and value vectors.

[0036] The risk level of immune rejection is determined by summing the weights and the risk threshold; the risk level of mechanical mismatch is determined by the cosine similarity between the elastic modulus of the parameter data and the bone density distribution of the patient host; and the risk level of degradation synchronicity is determined by the decomposition rate and bone metabolic activity score in the parameter data.

[0037] According to a second aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform an operation corresponding to the method described in the first aspect.

[0038] According to the scheme provided in the embodiments of the present invention, the patient's three-dimensional CT image data, blood test data, and parameter data of bone repair materials are acquired; the target region in the preprocessed three-dimensional CT image data is uniformly divided into overlapping blocks to obtain multiple sub-blocks; the target region includes bone defect regions and host bone regions; the position encoding and feature extraction of multiple sub-blocks are performed using a three-dimensional visual Transformer model with multi-scale position encoding to obtain the first anatomical and functional features; the bone metabolic activity score and systemic inflammation score are calculated based on the blood test data; and the first pathophysiological feature is determined based on the systemic inflammation score and bone metabolic activity score; multiple rounds of interaction are performed based on the first anatomical and functional features, the first pathophysiological feature, parameter data, and the multi-directional multi-head cross-attention layer of the multi-modal Transformer model. In each round, one modal feature is selected as the query vector, and any modal feature among the remaining modal features is selected as the key vector and value vector to obtain the interaction features after each round; the interaction features are fused to obtain fused features; and the suitability score and risk level of the bone repair materials are determined based on the fused features, the attention weights in the multi-directional multi-head cross-attention layer, and the bone metabolic activity score. In this process, the collected 3D CT image data, blood test data, and parameter data of bone repair materials are combined to build a comprehensive predictive foundation. The target region in the collected 3D CT image data is segmented to effectively capture the local microstructural features of the bone defect area and the host bone. A multi-scale positional encoding 3D visual Transformer model is used to positionally encode multiple sub-blocks, preserving both the global anatomical context and capturing local microstructural details. This further extracts anatomical and functional features, enabling the characterization of the bone defect's structural morphology and bioactivity, providing a comprehensive and accurate local microenvironmental representation for bone defect repair material compatibility. Processing the three modal features through a multi-directional, multi-head cross-attention layer allows for the learning of cross-modal associations between them. This integrates multi-source information while providing interpretable risk attribution, significantly improving the accuracy and clinical reliability of bone repair material compatibility prediction. Finally, based on fusion features, attention weights in the multi-directional, multi-head cross-attention layer, and bone metabolic vitality scores, the suitability score and suitability risk of bone repair materials are determined. This not only allows for more accurate prediction of suitability results but also quantifies the correlation strength between different factors, intuitively revealing the sources of risk. In summary, this method can efficiently and accurately predict the suitability of bone defect repair materials. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0040] Figure 1 A flowchart illustrating a deep learning-based method for predicting the compatibility of bone defect repair materials, provided in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0044] It should be noted that the terms "first, second, and third" used in the embodiments of the present invention are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0045] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which these embodiments of the invention pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0046] Figure 1This is a flowchart illustrating a deep learning-based method for predicting the compatibility of bone defect repair materials according to an embodiment of the present invention. This deep learning-based method for predicting the compatibility of bone defect repair materials according to an embodiment of the present invention can be executed by an electronic device, such as a computer or server.

[0047] like Figure 1 As shown, a deep learning-based method for predicting the compatibility of bone defect repair materials includes:

[0048] S101. Acquire the patient's three-dimensional CT image data, blood test data, and parameter data of bone repair materials; and divide the target area in the preprocessed three-dimensional CT image data into overlapping and uniform blocks to obtain multiple sub-blocks; the target area includes the bone defect area and the host bone area.

[0049] In embodiments of this invention, three-dimensional computed tomography (CT) image data, blood test data, and parameter data of bone repair materials are acquired. The three-dimensional CT image data includes the bone defect area, surrounding host bone, and adjacent anatomical structures. The bone defect area is the region where bone should have been present but is now missing or occupied by non-bone tissue. The surrounding host bone refers to the patient's own original bone tissue that has not been destroyed by the lesion and still possesses activity and structural integrity. Adjacent anatomical structures refer to non-bone tissues or organs spatially close to the bone defect or host bone, such as blood vessels, nerves, and vital organs. Blood test data includes bone metabolism indicators, inflammatory indicators, and preset routine biochemical indicators. The parameter data of the bone repair materials is structured data of key physical, chemical, biological, and clinical performance indicators of various biomaterials used for bone defect repair, including material category and core parameters. Material categories include three systems: bioceramics, metals, and polymers. Core parameters include porosity, specific surface area, and degradation curves.

[0050] Furthermore, Hounsfield Unit (HU) values ​​of pixels are extracted from the 3D CT image data. After air removal and metal artifact correction, the HU values ​​are normalized to the [0, 1] range using linear mapping to obtain preprocessed 3D CT image data. The preprocessed 3D CT image data is then input into a trained U-Net segmentation model. The U-Net segmentation model's encoder uses a 3D ResNet-18 network structure, and the decoding layer consists of deconvolutional layers. Finally, the U-Net segmentation model outputs a binary segmentation mask, where regions with a mask value of 1 represent the target region, including corresponding bone defects and host bone tissue, while regions with a mask value of 0 represent soft tissue and interfering tissues such as air. For the target region, 64×64×64 voxel overlap blocks are formed with a stride of 32 to obtain multiple 3D sub-blocks.

[0051] S102. Using a multi-scale positional encoding 3D vision Transformer model, positional encoding and feature extraction are performed on multiple sub-blocks to obtain the first anatomical and functional features.

[0052] In some embodiments of the present invention, the multi-scale position-encoded 3D vision Transformer model is a model that introduces local position encoding and 3D absolute position encoding into the standard 3D vision Transformer model. That is, after multiple sub-blocks are input into the multi-scale position-encoded 3D vision Transformer model, each sub-block is divided, and then 3D absolute position encoding and local position encoding are added on the basis of the divided local blocks. Finally, the corresponding feature extraction processing is performed through subsequent layers to obtain the first anatomical and functional features.

[0053] S103. Calculate the bone metabolism activity score and systemic inflammation score based on blood test data; and determine the first pathophysiological feature based on the systemic inflammation score and bone metabolism activity score.

[0054] In some embodiments of the present invention, the bone metabolism activity score can assess the dynamic balance between osteogenic and osteoclastic activities, reflecting the positive ability to promote bone repair. The systemic inflammation score is a derived indicator that comprehensively quantifies the patient's systemic chronic or subclinical inflammation level, reflecting the negative microenvironment that inhibits bone regeneration. Blood test data includes bone metabolism indicators and inflammation indicators. These indicators can be substituted into their respective scoring formulas to obtain the bone metabolism activity score and the systemic inflammation score. Furthermore, the first pathophysiological feature can be integrated based on the bone metabolism activity score and the systemic inflammation score.

[0055] S104. Based on the first anatomical and functional features, the first pathophysiological features, parameter data, and the multi-directional multi-head cross-attention layer of the multimodal Transformer model, perform multiple rounds of interaction. In each round, select one modal feature as the query vector, and select any modal feature from the remaining modal features as the key vector and value vector to obtain the interaction features after each round.

[0056] In the embodiments of the present invention, after constructing the features corresponding to the parameter data, the first anatomical and functional features, the first pathophysiological features, and the feature semantics extracted from the parameter data are aligned. Then, after contextual information fusion through the self-attention mechanism in the multimodal Transformer model, the three modal features are input into the multi-directional multi-head cross-attention layer. In each round, one modal feature is selected from the three modal features as the query vector, and one of the remaining two modal features is selected as the key vector and value vector. The query vector is different in each round, and the key vector and value vector are also different. Finally, the interaction features after each round are obtained.

[0057] S105. The interaction features are fused to obtain fused features; and the suitability score and risk level of bone repair materials are determined based on the fused features, the attention weights in the multi-directional multi-head cross-attention layer and the bone metabolism vitality score.

[0058] In an embodiment of the present invention, the interaction features after three rounds of interaction are fused to obtain fused features. The fused features can be input into multiple sequentially connected fully connected layers with different dimensions. The output of the fully connected layers is mapped to the [0, 1] interval by the Sigmoid activation function and then multiplied by 100 to obtain a quantized fit score of 0 to 100.

[0059] Furthermore, the fit score can be compared with the preset score. A fit score of 80 or higher is considered high fit, a fit score of 60 or higher but lower than 80 is considered medium fit, and a fit score of less than 60 is considered low fit.

[0060] Furthermore, the corresponding risk level is determined by the attention weight in the multi-directional, multi-head cross-attention layer, the bone metabolism vitality score, the elastic modulus of the bone repair material, and the preset risk scoring criteria.

[0061] It is understood that, in some embodiments of the present invention, the three types of data—collected 3D CT image data, blood test data, and parameter data of bone repair materials—can be combined to build a comprehensive predictive foundation for subsequent analysis. Dividing the target region in the collected 3D CT image data effectively captures the local microstructural features of the bone defect area and the host bone. Utilizing a multi-scale position-encoded 3D visual Transformer model to positionally encode multiple sub-blocks preserves the global anatomical context while capturing local microstructural details, further extracting anatomical and functional features. This allows for characterization of the structural morphology and bioactivity of the bone defect, providing a comprehensive and accurate local microenvironmental representation for the adaptability of bone defect repair materials. Processing the three modal features through a multi-directional, multi-head cross-attention layer allows for learning cross-modal associations between them, providing interpretable risk attribution while fusing multi-source information, significantly improving the accuracy and clinical reliability of bone repair material adaptability prediction. Finally, based on fusion features, attention weights in the multi-directional, multi-head cross-attention layer, and bone metabolic vitality scores, the suitability score and suitability risk of bone repair materials are determined. This not only allows for more accurate prediction of suitability results but also quantifies the correlation strength between different factors, intuitively revealing the sources of risk. In summary, this method can efficiently and accurately predict the suitability of bone defect repair materials.

[0062] In some embodiments of the present invention, S102 can be implemented through the following steps: inputting multiple sub-blocks into the block embedding layer of a 3D vision Transformer model, dividing them into multiple local blocks by a 3D convolutional layer containing a 3D convolutional kernel of a preset size, and mapping the multiple local blocks to construct an initial feature matrix; adding the 3D absolute position encoding and the local position encoding to the initial feature matrix element by element to obtain a first feature matrix; the 3D absolute position encoding is the absolute spatial position of each sub-block in the target region; the local position encoding is the absolute spatial position of each local block in the sub-block; adding learnable classification tokens to the first feature matrix, and inputting the added first feature matrix into the first encoder layer to obtain a second feature matrix fused with context information; inputting the second feature matrix into the second encoder layer again, until the last encoder layer outputs the target feature matrix; using a fully connected layer to perform dimensional mapping on the feature vectors of the classifiable tokens in the target feature matrix to obtain the first anatomical and functional features.

[0063] Specifically, the block embedding layer contains a 3D convolutional layer. Multiple 64×64×64 voxel sub-blocks are input into the block embedding layer and divided by a 3D convolutional layer containing 8×8×8 convolutional kernels, resulting in multiple local blocks. These are then linearly projected onto a 768-dimensional vector, yielding an initial feature matrix of shape [512, 768]. The 3D absolute position encoding represents the absolute spatial position of each sub-block within the target region (i.e., the target region is the reference point). The local position encoding represents the absolute spatial position of each local block within a sub-block (i.e., each 64×64×64 voxel sub-block is the reference point). Finally, the 3D absolute position encoding and the local position encoding are element-wise added to the initial feature matrix to obtain the final first feature matrix.

[0064] Furthermore, a learnable classification token, namely the [CLS] token, is added to the first feature matrix. This classification token can aggregate the global semantics of the local bone microenvironment. The first feature matrix with the added classification token is input into the first encoder layer. After processing through multi-head attention, residual connections, layer normalization, feedforward neural network, residual connections, and normalization, a second feature matrix with fused contextual information is obtained after processing by the first encoder layer. The second feature matrix is ​​then input into the second encoder layer and processed in the same way as the first encoder layer until the last encoder layer outputs the final target feature matrix. The internal processes of all encoder layers are consistent. After 12 iterations of encoder layers, the classification token continuously aggregates information from all local blocks, and finally, a global representation of the target region can be obtained.

[0065] Furthermore, feature vectors corresponding to classification tokens are extracted from the target feature matrix. These feature vectors are then input into a fully connected layer for dimensional mapping to obtain the first anatomical and functional features. These first anatomical and functional features encompass morphological features, dimensional features, and topological features. Morphological features include defect volume, surface area, and three-dimensional morphological complexity. Density features include the mean, standard deviation, and spatial distribution entropy of host bone mineral density. Topological features focus on the connectivity of microstructures, including trabecular connection density, number of branches, and network connectivity.

[0066] It is understood that, in some embodiments of the present invention, the addition of three-dimensional absolute position encoding and local position encoding to local blocks can both preserve the global anatomical context and capture local microstructural details. Through processing by multiple encoder layers, features with higher discriminative power for bone repair material adaptability can be extracted. By dimensionally mapping the features of the classification tokens, the first anatomical and functional features are obtained, and the global three-dimensional context information of the target region is automatically aggregated, which can provide highly discriminative input for subsequent adaptability scoring and risk prediction.

[0067] In some embodiments of the present invention, the bone metabolism vitality score and systemic inflammation score calculated based on blood test data in S103 can be achieved through the following steps: normalizing the inflammatory indicators C-reactive protein and tumor necrosis factor... Substituting interleukin-6 into the systemic inflammation scoring formula yields the systemic inflammation score; normalized bone metabolism indicators, including alkaline phosphatase, osteocalcin, and... Type I collagen is substituted into the bone metabolism vitality scoring formula to obtain the bone metabolism vitality score; the systemic inflammation scoring formula is as follows:

[0068] ;

[0069] In the above formula, S For systemic inflammation scores, It is C-reactive protein. Tumor necrosis factor , It is interleukin-6;

[0070] The formula for bone metabolism vitality scoring is as follows:

[0071] ;

[0072] In the above formula, B Bone metabolic activity score, It is an alkaline phosphatase. For osteocalcin, for Type II collagen.

[0073] Specifically, blood test data includes bone metabolism and inflammatory markers, as well as preset routine biochemical indicators, including creatinine, blood glucose, alanine aminotransferase (ALT), and hemoglobin. Bone metabolism markers include alkaline phosphatase, osteocalcin, and... Type II collagen, inflammatory markers including C-reactive protein and tumor necrosis factor. And interleukin-6. Missing values ​​for the three types of indicators can be filled using neighborhood mean, outliers are handled by Z-score truncation, and then normalized to obtain normalized values ​​for the three types of indicators. Systemic inflammation scoring formulas and bone metabolism vitality scoring formulas are set; the weights in these two formulas can be set by experts, including but not limited to experts. C-reactive protein and tumor necrosis factor are included in the normalized inflammatory indicators. Substituting interleukin-6 into the above systemic inflammation scoring formula yields the systemic inflammation score, along with normalized bone metabolism indicators such as alkaline phosphatase, osteocalcin, and... Substituting type 3 collagen into the above bone metabolism vitality scoring formula yields the bone metabolism vitality score.

[0074] The weights of 0.5, 0.3, 0.2, 0.4, 0.3, and 0.3 in the bone metabolism vitality scoring formula and the systemic inflammation scoring formula are determined by expert scoring for the scenario of predicting the suitability of bone defect repair materials. This method belongs to existing technology and will not be explained in detail here.

[0075] In some embodiments of the present invention, determining the first pathophysiological feature based on systemic inflammation score and bone metabolism activity score in S103 can be achieved by the following steps: combining systemic inflammation score, bone metabolism activity score, normalized bone metabolism index, normalized inflammation index and normalized preset conventional biochemical index to obtain the first pathophysiological feature.

[0076] Specifically, preset routine biochemical indicators are normalized, and systemic inflammation scores and bone metabolism activity scores are mapped to a range of 0-10. The mapped systemic inflammation scores, bone metabolism activity scores, normalized bone metabolism indicators, normalized inflammation indicators, and normalized preset routine biochemical indicators are combined. The first two dimensions can be the mapped scores, and the last 10 dimensions can be the indicators, finally yielding a 12-dimensional first pathophysiological feature.

[0077] It is understood that, in some embodiments of the present invention, systemic inflammation scores and bone metabolic activity scores are calculated and integrated with three types of indicators to construct a first pathophysiological feature, which enables the construction of a comprehensive, quantitative and individualized biological state profile that not only reflects the current balance between immunity and osteogenic formation, but also integrates systemic influencing factors, enabling efficient and accurate prediction of the suitability of bone defect repair materials.

[0078] In some embodiments of the present invention, S104 can be implemented through the following steps: normalizing, one-hot encoding, and extracting key features from the parameter data to construct a first material feature; unifying the dimensions of the first anatomical and functional feature, the first pathophysiological feature, and the first material feature to obtain a second anatomical and functional feature, a second pathophysiological feature, and a second material feature; inputting the second anatomical and functional feature, the second pathophysiological feature, and the second material feature into the self-attention layer of the multimodal Transformer model to obtain a third anatomical and functional feature, a third pathophysiological feature, and a third material feature that integrates the information above and below; inputting the multimodal features composed of the third anatomical and functional feature, the third pathophysiological feature, and the third material feature into a multi-directional multi-head cross-attention layer for multi-round interaction, in each round, selecting one modal feature as the query vector, and selecting any modal feature from the remaining modal features as the key vector and value vector, until the interaction features after each round are obtained.

[0079] Specifically, the porosity and specific surface area in the parameter data are normalized, and the surface functional groups are individually thermally encoded. The degradation kinetic curve is resampled to 12 time points using cubic spline interpolation to construct a smooth degradation kinetic curve. Then, the loss rate in week 1, the maximum loss rate, and the loss rate in week 8 are calculated based on the smooth degradation kinetic curve as key features. At the same time, based on predefined fields, three key features are extracted from the parameter data: initial elastic modulus, modulus retention rate after 4 weeks, and fracture toughness. The six key features are normalized and integrated with the normalized porosity, specific surface area, and surface functional groups to construct the first material feature.

[0080] Furthermore, three independent linear layers are used to unify the dimensions of the first anatomical and functional features, the first pathophysiological features, and the first material features, respectively, to obtain the second anatomical and functional features, the second pathophysiological features, and the second material features. Self-attention layers are then used to fuse the context of the second anatomical and functional features, the second pathophysiological features, and the second material features, respectively, to obtain the fused third anatomical and functional features, the third pathophysiological features, and the third material features. There can be two self-attention layers. When there are two, the output of the first self-attention layer can be input again into the second attention layer for processing to obtain the final third anatomical and functional features, the third pathophysiological features, and the third material features.

[0081] Furthermore, the third anatomical and functional features, the third pathophysiological features, and the third material features are input into a multi-directional, multi-head cross-attention layer for multi-round interactions. In the first round, the third anatomical and functional features can be used as query vectors, and the third pathophysiological features as key and value vectors. Based on the query vectors, key vectors, and value vectors, the features of the fusion context are calculated. The features output by multiple heads are fused to obtain the interaction features after the first round. The second and third rounds are similar. In the second round, the third pathophysiological features can be used as query vectors, and the third material features as key and value vectors. In the third round, the third material features can be used as query vectors, and the third anatomical and functional features as key and value vectors.

[0082] It is understood that, in some embodiments of the present invention, the three modal features are first fully modeled within their respective modal contexts through self-attention to form high-quality three-modal representations; then, cross-modal alignment and interaction are performed through multi-directional multi-head cross-attention to achieve more accurate, stable and efficient multimodal fusion, thereby ultimately improving the prediction accuracy, robustness and interpretability of bone repair material fit scores and risk levels.

[0083] In some embodiments of the present invention, the determination of the compatibility score and risk level of the bone repair material based on fusion features, attention weights in the multi-directional multi-head cross-attention layer, and bone metabolic activity score in S105 is achieved through the following steps: determining the predicted compatibility score of the bone repair material based on fusion features; and in the second round of interaction in the multi-directional multi-head cross-attention layer, extracting the weights between inflammatory indicators and surface chemistry in the material features from the attention weight matrix; the pathophysiological features in the second round are query vectors, and the material features are bond vectors and value vectors; determining the immune rejection risk level based on the sum of the weights and the risk threshold; determining the mechanical mismatch risk level through the cosine similarity between the elastic modulus of the parameter data and the patient's host bone density distribution; and determining the degradation synchronicity risk level through the decomposition rate and bone metabolic activity score in the parameter data.

[0084] Specifically, the fused features are processed based on multiple sequentially connected fully connected layers and a Sigmoid activation function to determine the suitability score. In the second round of interaction in the multi-directional multi-head cross-attention layer, all weights between inflammatory indicators and surface chemistry in material features are extracted from the attention weight matrix. The extracted weights are summed, and the total weight is compared with a risk threshold to determine the immune rejection risk level. If the total weight is less than 0.3, it is considered low risk; if the total weight is not less than 0.3 and less than 0.6, it is considered medium risk; and if the total weight is not less than 0.6, it is considered high risk.

[0085] Furthermore, fields can be read from the material's parameter data to extract the elastic modulus. Additionally, the HU values ​​of 3D CT images can be converted to bone mineral density (BMD) values ​​based on a calibration phantom, and voxel-by-voxel mapping can be performed on these BMD values ​​to obtain the host's BMD distribution. The cosine similarity between the elastic modulus and the host BMD distribution is used to determine the risk level of mechanical mismatch. A cosine similarity of no more than 0.5 indicates high risk; no less than 0.5 and less than 0.8 indicates medium risk; and greater than 0.8 indicates low risk.

[0086] Furthermore, the Pearson correlation coefficient between the degradation rate and bone metabolic activity score in the material parameter data is calculated to determine the concurrent degradation risk level. If the Pearson correlation coefficient is not greater than 0.3, it is considered high risk; if it is not less than 0.3 and less than 0.7, it is considered medium risk; and if it is not less than 0.7, it is considered low risk.

[0087] In some embodiments of the present invention, when the fit score is high but there is a relatively high risk of immune rejection, mechanical mismatch, or degradation synchronization, the data can be matched using a preset clinical rule base to make improvements. The optimization strategy is determined by matching the fit score and the specific risk level with the preset clinical rule base. For example, when the fit score indicates high fit and the immune rejection risk is also high, it may be recommended to add chitosan or dexamethasone surface modification measures for optimization.

[0088] It is understood that, in some embodiments of the present invention, predicting the fit score by incorporating semantic fusion features from multiple modal data can improve the accuracy and robustness of the prediction, avoiding misjudgments caused by single-modal bias. The immune rejection risk level is determined by the weighted sum of inflammatory indicators and surface chemistry in material characteristics, enabling interpretable and mechanism-guided precise assessment of immune rejection risk. The mechanical mismatch risk level is determined by calculating the cosine similarity between the material's elastic modulus and the patient's host bone density distribution, effectively capturing the overall matching degree between material stiffness and the host's bone mechanical environment, thereby objectively identifying stress shielding or micromotion risks. The degradation synchronicity risk level is determined by calculating the Pearson correlation coefficient between the material's degradation rate and the patient's bone metabolic activity score, quantifying the synergy between material degradation rate and new bone formation rhythm in a statistically rigorous manner, thereby objectively identifying the risk of repair failure due to excessively rapid or slow degradation, achieving precise assessment.

[0089] In some embodiments of this invention, a unified experimental dataset is used, such as collecting complete clinical data from a total of 1000 patients with bone defects, including 3D CT images, blood test reports, and parameter data of various common bone repair materials, covering three systems: bioceramics, metals, and polymers. All cases have a gold standard label for bone integration rate at 6 months post-surgery, independently evaluated by three senior orthopedic surgeons, with consensus values ​​adopted. The experimental dataset is divided in an 8:1:1 ratio to obtain training data, test data, and validation data. Table 1 below shows the comparison results of various models in predicting the compatibility of bone repair materials. Model 1 is a model combining a 3D convolutional neural network and two modalities of data, i.e., inputting data from two different sources into the 3D convolutional neural network for prediction. Model 2 is a baseline model combining a 3D convolutional neural network and a shallow fusion of three modalities, i.e., using a 3D convolutional neural network to extract features, and directly concatenating the three modal features before prediction without cross-modal interaction. The compatibility prediction correlation in Table 1 measures the linear correlation between the compatibility score output by this invention and the gold standard score assessed by experts based on the actual post-operative bone integration effect. The closer r is to 1, the better the sorting ability and clinical discriminative power of the present invention. The postoperative material repair rate is the proportion of bone defect repair materials recommended by the present invention that are determined by clinicians to require replacement or secondary intervention due to poor healing, loosening, infection or rejection during postoperative follow-up. The average accuracy rate of multiple risks is the average of the accuracy rates of immune rejection risk, mechanical mismatch risk, and degradation synchronicity risk.

[0090] Table 1. Comparison of the predictive performance of various models on the compatibility of bone repair materials

[0091] ;

[0092] Table 1 shows that the adaptability prediction relevance of this invention is higher than that of Model 1 and Model 2. This indicates that relying solely on data from two modalities, or even using data from three modalities without deep dynamic correlation modeling, cannot accurately predict the adaptability of bone repair materials. In contrast, this invention, based on a cross-modal deep interaction mechanism, captures the conditional dependencies between different factors, rather than simply superimposing them, thus improving the adaptability prediction relevance. The adaptability prediction error of this invention is 3.2%, a reduction of 62.4% and 53.6% compared to Model 1 and Model 2, respectively. This demonstrates that this invention can more stably generate quantitative scores consistent with clinical gold standards, and that the dynamic interaction mechanism can adaptively adjust the contributions of each modality, improving prediction accuracy. The average multi-risk accuracy of this invention is 89.8%, a relative improvement of 18.0% and 12.8% compared to Model 1 and Model 2, respectively. This indicates that Model 2, by simply increasing the number of modalities without improving the fusion method, cannot improve fine-grained discrimination capabilities; while the multi-directional, multi-head cross-attention of this invention can focus on key risk areas, possessing stronger fine-grained risk analysis capabilities and achieving accurate attribution. The postoperative revision rate of this invention was 7.8%, which is a reduction of 57.4% and 46.6% compared to Model 1 and Model 2, respectively. This demonstrates that the deep interactive fusion of this invention can more effectively translate predictive capabilities into clinical benefits, while the two modalities in Model 1 and the shallow fusion in Model 2 have limitations in complex decision-making scenarios. In summary, Table 1 demonstrates the superiority of this invention in predicting the compatibility of bone repair materials.

[0093] Reference Figure 2 The diagram shows a structural schematic of an electronic device according to an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the electronic device.

[0094] like Figure 2 As shown, the electronic device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0095] in:

[0096] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508.

[0097] Communication interface 504 is used to communicate with other electronic devices or servers.

[0098] The processor 502 is used to execute program 510, specifically the relevant steps in the above method embodiments.

[0099] Specifically, program 510 may include program code that includes computer operation instructions.

[0100] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The smart device may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0101] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0102] Specifically, program 510 can be used to cause processor 502 to perform the operations corresponding to the methods described in the above method embodiments.

[0103] The specific implementation of each step in program 510 can be found in the corresponding descriptions of the steps and units in the above method embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0104] It should be noted that, depending on the implementation needs, the various components / steps described in the embodiments of the present invention can be broken down into more components / steps, or two or more components / steps or parts of the operation of components / steps can be combined into new components / steps to achieve the purpose of the embodiments of the present invention.

[0105] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0106] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the embodiments of the present invention.

[0107] The above embodiments are only used to illustrate the embodiments of the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of the present invention, and the patent protection scope of the embodiments of the present invention should be defined by the claims.

Claims

1. A deep learning-based method for predicting the compatibility of bone defect repair materials, characterized in that, include: Acquire the patient's 3D CT image data, blood test data, and parameter data of bone repair materials; The target region in the preprocessed 3D CT image data is then uniformly divided into overlapping blocks to obtain multiple sub-blocks; the target region includes bone defect areas and host bone areas. The positional encoding and feature extraction of multiple sub-blocks are performed using a 3D visual Transformer model with multi-scale positional encoding to obtain the first anatomical and functional features; Bone metabolism activity score and systemic inflammation score were calculated based on blood test data; and the first pathophysiological feature was determined based on the systemic inflammation score and bone metabolism activity score. Based on the first anatomical and functional features, the first pathophysiological features, parameter data, and the multi-directional multi-head cross-attention layer of the multimodal Transformer model, multiple rounds of interaction are performed. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector to obtain the interaction features after each round. The interaction features are fused to obtain the fused features; The compatibility score and risk level of bone repair materials were determined based on fusion features, attention weights in the multi-directional multi-head cross-attention layer, and bone metabolic activity scores, respectively. The blood test data includes bone metabolism indicators, inflammation indicators, and preset routine biochemical indicators; The bone metabolism vitality score and systemic inflammation score are calculated based on blood test data, including: C-reactive protein and tumor necrosis factor were normalized inflammatory markers. Substituting interleukin-6 into the systemic inflammation scoring formula yields the systemic inflammation score; Normalized bone metabolism indicators include alkaline phosphatase, osteocalcin, and... Type I collagen is substituted into the bone metabolism vitality scoring formula to obtain the bone metabolism vitality score; The formula for the systemic inflammation scoring system is as follows: ； In the above formula, S For systemic inflammation scores, It is C-reactive protein. Tumor necrosis factor , It is interleukin-6; The formula for bone metabolism vitality scoring is as follows: ； In the above formula, B Bone metabolic activity score, It is an alkaline phosphatase. For osteocalcin, for Type II collagen; The multi-scale position-encoded 3D vision Transformer model includes a block embedding layer, multiple encoder layers, and a fully connected layer. The three-dimensional visual Transformer model, which utilizes multi-scale positional encoding, performs positional encoding and feature extraction on multiple sub-blocks to obtain the first anatomical and functional features, including: Multiple sub-blocks are input into the block embedding layer of the 3D vision Transformer model, and then divided by a 3D convolutional layer containing a 3D convolutional kernel of a preset size to obtain multiple local blocks. The multiple local blocks are then mapped to construct an initial feature matrix. The three-dimensional absolute position code and the local position code are added element-wise to the initial feature matrix to obtain the first feature matrix; the three-dimensional absolute position code is the absolute spatial position of each sub-block in the target region; the local position code is the absolute spatial position of each local block in the sub-block; Learnable classification tokens are added to the first feature matrix, and the first feature matrix with the added tokens is input into the first encoder layer to obtain the second feature matrix with fused context information; the second feature matrix is ​​then input into the second encoder layer, and so on, until the last encoder layer outputs the target feature matrix. The first anatomical and functional features are obtained by using a fully connected layer to perform dimension mapping on the feature vectors of classifiable tokens in the target feature matrix. The compatibility score and risk level of bone repair materials are determined based on fusion features, attention weights in a multi-directional, multi-head cross-attention layer, and bone metabolic activity scores, including: The compatibility score of the predicted bone repair material is determined based on the fusion features; and in the second round of interaction in the multi-directional multi-head cross attention layer, the weights between inflammatory indicators and surface chemistry in the material features are extracted from the attention weight matrix; the pathophysiological features in the second round are query vectors, and the material features are key vectors and value vectors. The risk level of immune rejection is determined by summing the weights and the risk threshold; the risk level of mechanical mismatch is determined by the cosine similarity between the elastic modulus of the parameter data and the bone density distribution of the patient host; and the risk level of degradation synchronicity is determined by the decomposition rate and bone metabolic activity score in the parameter data.

2. The method according to claim 1, characterized in that, The determination of the first pathophysiological feature based on systemic inflammation score and bone metabolic activity score includes: The first pathophysiological characteristic was obtained by combining systemic inflammation score, bone metabolism activity score, normalized bone metabolism index, normalized inflammation index, and normalized preset routine biochemical index.

3. The method according to claim 1, characterized in that, The multi-directional, multi-head cross-attention layer based on the first anatomical and functional features, the first pathophysiological features, parametric data, and the multimodal Transformer model performs multi-round interactions. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector, resulting in the interaction features after each round, including: The parameter data were normalized, uniquely thermally encoded, and key features were extracted to construct the first material features; and the first anatomical and functional features, the first pathophysiological features, and the first material features were unified in dimensionality to obtain the second anatomical and functional features, the second pathophysiological features, and the second material features. The second anatomical and functional features, the second pathophysiological features, and the second material features are input into the self-attention layer of the multimodal Transformer model to obtain the third anatomical and functional features, the third pathophysiological features, and the third material features, which respectively integrate the information from the upper and lower parts. The multimodal features, consisting of the third anatomical and functional features, the third pathophysiological features, and the third material features, are input into a multi-directional, multi-head cross-attention layer for multi-round interaction. In each round, one modal feature is selected as the query vector, and any modal feature from the remaining modal features is selected as the key vector and value vector, until the interaction features after each round are obtained.

Images