Radial grade subdivision method and device based on deep clustering
By combining deep learning and clustering algorithms, the CHN radius reference atlas is optimized, the radius bone grades are subdivided and scores are assigned, which solves the error problem caused by excessive grade differences in bone age assessment and improves the accuracy of the assessment.
Patent Information
- Application Number
- CN202411004597.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2025-01-21
AI Technical Summary
In the bone age assessment, the gap in the development indication of reference bone adjacent grades is too large, resulting in expert judgment errors and deep learning random misjudgment, affecting the accuracy of bone age assessment.
The combination of deep learning and clustering algorithms is used to optimize the CHN radial reference map, and the use of the main component analysis, self-encoder feature extraction, K-MEANS clustering and joint loss function, subdivide radial grade, and pass through Migration learning and interpolation method gives new level scores.
It improves the accuracy of bone age assessment, reduces errors, assists clinical applications, and enables the subdivision of bones with a large span of developmental indicators between adjacent levels.
Smart Images
Figure CN119339118A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and device for radial grade subdivision based on deep clustering. Background Art
[0002] Bone age is an important criterion for assessing the growth and development of adolescents and children. Domestic bone age experts often use the CHN method to assess bone age, which is a method of first distinguishing the grades of hand bones and then calculating bone age. The CHN method divides the development grades of the 14 reference bones in the wrist, and each grade is assigned a corresponding score. The final bone age value is obtained by summing up the scores of the corresponding grades of all reference bones and comparing the final score to the score table.
[0003] However, due to the large disparity in developmental indicators between adjacent levels of some reference bones, experts may make mistakes due to their personal experience when making judgments, and the scores corresponding to these adjacent levels are also very different, which ultimately leads to large errors in bone age results. The same problem exists when using deep learning to identify levels, and the randomness of deep learning can also lead to such misjudgments, thus affecting the assessment of bone age. It is obviously unrealistic to use artificial subdivision on the basis of the existing level division. Summary of the invention
[0004] In view of the above problems, the present invention proposes a hierarchical subdivision method and device based on deep clustering.
[0005] The present invention uses a method combining deep learning and clustering algorithms to optimize the CHN radius reference atlas, thereby improving the accuracy of bone age assessment.
[0006] The present invention solves the technical problem by adopting the following technical solution:
[0007] A radial bone grade subdivision method based on deep clustering comprises the following steps:
[0008] Step 1: Use principal component analysis (PCA) to reduce the dimension of the original data, and then use the autoencoder in the deep learning method to extract features of the radius;
[0009] Step 2: Classify the features extracted in step 1 through a clustering algorithm (K-means) to complete the segmentation of the radius grade, and add the reconstruction loss function of the autoencoder in step 1 and the clustering loss function as the joint loss;
[0010] Step 3: Using the idea of transfer learning, MobileNet-V2 is used to extract features from the subdivided levels, and the Euclidean distance between the subdivided levels and the adjacent levels is calculated. After standardizing the distance, the interpolation method is used to assign corresponding scores to the new levels generated after the subdivision.
[0011] Wherein, step 1 specifically includes:
[0012] 11) Use principal component analysis (PCA) to reduce the dimension of the original data while retaining the most important features. Then use the autoencoder in the deep learning method to extract features of the radius. The specific implementation method is as follows:
[0013] The input data is low-dimensional data processed by principal component analysis (PCA). The autoencoder network mainly consists of two parts, namely the encoder and the decoder. The encoder consists of three modules, each of which contains linear layers of different shapes, and a ReLU activation operation is added after each linear layer. The features obtained after the encoder are the input features of the clustering algorithm in step 2.
[0014] The structure of the decoder part is opposite to that of the encoder. It also consists of three modules. Each module contains linear layers of different shapes. A ReLU activation operation is added after each linear layer, and the final output is the reconstructed image.
[0015] 12) Design reconstruction loss function:
[0016] The reconstruction loss function expression used in the present invention is as follows:
[0017]
[0018] where x i is the i-th data input, f(x i ) function is as follows:
[0019]
[0020] f(x i ) function itself represents a mapping relationship. It represents the network parameters in the encoder network. In the present invention, the encoding part of the autoencoder is used as a mapping function to map the original data to a low-dimensional space. i )) function is the function used to reconstruct the data, that is, the decoder part, and its complete function is parameter Represents various parameters in the reconstruction network. Function l is a specific loss function used in the present invention to measure the reconstruction error. The least squares loss is used as the loss function of this part. Its original formula is as follows:
[0021]
[0022] Wherein, step 2 specifically includes:
[0023] 21) Perform k-means clustering on the features after the dimension reduction process in step 1. In the k-means algorithm, the Euclidean distance is usually used to calculate the distance between data points. The formula is as follows:
[0024]
[0025] Where (x1, y1) and (x2, y2) represent the coordinate positions of two points respectively. It is an iterative process from the beginning to the final determination of clustering.
[0026] The clustering loss function is designed as follows:
[0027]
[0028] where s i is the assignment vector for data point i, which has only one non-zero element. This means that s i Is a sparse vector, which is used in the clustering algorithm to indicate which cluster the data point i is assigned to. The algorithm uses the assignment vector s i and the centroid matrix M to determine which cluster the data point i belongs to.
[0029] 22) Design a joint loss function.
[0030] The clustering loss of the clustering algorithm and the reconstruction loss of the autoencoder part are combined and jointly optimized to ensure that the model can comprehensively consider the requirements of different tasks during the training process, so as to more comprehensively learn the feature representation and clustering assignment of the data.
[0031] The joint loss function is the weighted addition of the above formulas (1) and (5), which is specifically expressed as follows:
[0032]
[0033] Here, λ is a hyperparameter used to balance the reconstruction error and the K-means clustering error. The value of λ is greater than or equal to 0, and in the present invention, it is set to 1. After the above-mentioned joint loss, the parameters in the model will be updated in a direction that is more conducive to clustering.
[0034] Wherein, step 3 specifically includes:
[0035] 31) Using the idea of transfer learning, the MobileNet-V2 network is used for feature extraction, and only the network structure and parameters of the feature extraction part are used. The three levels of images obtained after clustering are subjected to feature extraction respectively, and the extracted features are saved in different feature vector lists to obtain three vector lists. Then the vector lists of the newly added levels are matched with each data in the two pairs of feature vector lists of the other two levels, and then converted into array form, and then the Euclidean distance is calculated. The calculation formula of the Euclidean distance is shown in formula (4). The distances obtained from each list are then stored in their respective lists, and the distances between the three can be obtained. The obtained Euclidean distance is then standardized, and the standardized formula used is as follows:
[0036] f(x)=(x min ) / (x max -x min ) (7)
[0037] x min and x max are the minimum and maximum distances in the list, respectively, and x is the original distance value. After normalization, the original distance value can be mapped to the range of 0-1, where 0 corresponds to the minimum distance in the original value and 1 corresponds to the maximum distance in the original value. f(x) is the distance after normalization.
[0038] 2) Based on the above standardized distance, the interpolation formula can be used to obtain the score corresponding to the newly added level. Taking the first element in the three lists as an example, since only the newly added level has no corresponding score, an element in the list where it is located is set to x, and the elements in the corresponding positions in the other two lists are set to x1 and x0 respectively, and the values of x1 and x0 correspond to the level scores of the lower level and the higher level respectively, using the following formula:
[0039] f(x)=f(x0)*(x1-x) / (x1-x0)+f(x1)*(x-x0) / (x1-x0) (8)
[0040] According to this formula, the scores corresponding to all elements in the newly added level list can be calculated, and the final level score can be obtained by taking the average of these scores.
[0041] A second aspect of the present invention relates to a radius grade subdivision device based on deep clustering, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a radius grade subdivision method based on deep clustering described in the present invention.
[0042] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a radial grade subdivision method based on deep clustering of the present invention.
[0043] The present invention has the following beneficial effects:
[0044] (1) A method combining deep learning and clustering algorithms was used to optimize the CHN radius reference atlas, and the bones with a large span of developmental indicators between adjacent levels were subdivided into different levels. A joint loss function was used to jointly optimize deep learning and clustering.
[0045] (2) Use the interpolation method to assign corresponding scores to the new levels generated after subdivision.
[0046] (3) Using the segmented atlas to assess bone age can improve the accuracy of the assessment and assist in clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is the overall flow chart of the present invention.
[0048] Figure 2 It is a diagram of the autoencoder network structure used in the present invention.
[0049] Figure 3 It is a flow chart of the interpolation method used in the present invention. DETAILED DESCRIPTION
[0050] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0051] Example 1
[0052] Reference Figure 1 to Figure 3 , a radial grade subdivision method based on deep clustering, comprising the following steps:
[0053] Step 1: Use principal component analysis (PCA) to reduce the dimension of the original data, and then use the autoencoder in the deep learning method to extract features of the radius;
[0054] Step 2: Classify the features extracted in step 1 through a clustering algorithm (k-means) to complete the segmentation of the radius grade, and add the reconstruction loss function of the autoencoder in step 1 and the clustering loss function as the joint loss;
[0055] Step 3: Using the idea of transfer learning, MobileNet-V2 is used to extract features from the subdivided levels, and the Euclidean distance between the subdivided levels and the adjacent levels is calculated. After standardizing the distance, the interpolation method is used to assign corresponding scores to the new levels generated after the subdivision.
[0056] Step 1 specifically includes:
[0057] 1) Use principal component analysis (PCA) to reduce the dimension of the original data while retaining the most important features. Then use the autoencoder in the deep learning method to extract features of the radius. The specific implementation method is as follows:
[0058] The input data is low-dimensional data processed by principal component analysis (PCA). The autoencoder network mainly consists of two parts, namely the encoder and the decoder. The encoder consists of three modules, each of which contains linear layers of different shapes, and a ReLU activation operation is added after each linear layer. The features obtained after the encoder are the input features of the clustering algorithm in step 2.
[0059] The structure of the decoder part is opposite to that of the encoder. It also consists of three modules. Each module contains linear layers of different shapes. A ReLU activation operation is added after each linear layer, and the final output is the reconstructed image.
[0060] 2) Design reconstruction loss function:
[0061] The reconstruction loss function expression used in the present invention is as follows:
[0062]
[0063] where x i is the i-th data input, f(x i ) function is as follows:
[0064]
[0065] f(x i ) function itself represents a mapping relationship. It represents the network parameters in the encoder network. In the present invention, the encoding part of the autoencoder is used as a mapping function to map the original data to a low-dimensional space. i )) function is the function used to reconstruct the data, that is, the decoder part, and its complete function is parameter Represents various parameters in the reconstruction network. Function l is a specific loss function used in the present invention to measure the reconstruction error. The least squares loss is used as the loss function of this part. Its original formula is as follows:
[0066]
[0067] Step 2 specifically includes:
[0068] 1) Perform k-means clustering on the features after the dimension reduction process in step 1. In the k-means algorithm, Euclidean distance is usually used to calculate the distance between data points. The formula is as follows:
[0069]
[0070] Where (x1, y1) and (x2, y2) represent the coordinate positions of two points respectively. It is an iterative process from the beginning to the final determination of clustering.
[0071] The clustering loss function is designed as follows:
[0072]
[0073] where s i is the assignment vector for data point i, which has only one non-zero element. This means that s i Is a sparse vector, which is used in the clustering algorithm to indicate which cluster the data point i is assigned to. The algorithm uses the assignment vector s i and the centroid matrix M to determine which cluster the data point i belongs to.
[0074] 2) Design a joint loss function.
[0075] The clustering loss of the clustering algorithm and the reconstruction loss of the autoencoder part are combined and jointly optimized to ensure that the model can comprehensively consider the requirements of different tasks during the training process, so as to more comprehensively learn the feature representation and clustering assignment of the data.
[0076] The joint loss function is the weighted addition of the above formulas (1) and (5), which is specifically expressed as follows:
[0077]
[0078] Here, λ is a hyperparameter used to balance the reconstruction error and the K-means clustering error. The value of λ is greater than or equal to 0, and in the present invention, it is set to 1. After the above-mentioned joint loss, the parameters in the model will be updated in a direction that is more conducive to clustering.
[0079] Step 3 specifically includes:
[0080] 1) Using the idea of transfer learning, the MobileNet-V2 network is used for feature extraction, and only the network structure and parameters of the feature extraction part are used. Taking the radius as an example, the images of level 5, level 5_5 and level 6 obtained after clustering are respectively subjected to feature extraction, and the extracted features are saved in different feature vector lists to obtain three vector lists. Since level 5_5 is a newly added level, the vector list of level 5_5 is matched with each data in the two pairs of feature vector lists of levels 5 and 6, and then converted into array form, and then the Euclidean distance is calculated. The calculation formula of the Euclidean distance is shown in formula (4). The distance obtained from each list is then stored in their respective lists, and the distance between the three can be obtained. The obtained Euclidean distance is then standardized, and the standardized formula used is as follows:
[0081] f(x)=(x min ) / (x max -x min ) (7)
[0082] x min and x max are the minimum and maximum distances in the list, respectively, and x is the original distance value. After normalization, the original distance value can be mapped to the range of 0-1, where 0 corresponds to the minimum distance in the original value and 1 corresponds to the maximum distance in the original value. f(x) is the distance after normalization.
[0083] 2) Based on the above standardized distance, the interpolation formula can be used to obtain the score corresponding to the newly added level. Take the first element of the three lists as an example. Since only the newly added level 5_5 has no corresponding score, it is set to x. The first elements of the other two lists are set to x1 and x0 respectively, and the values of x1 and x0 correspond to the level scores of level 5 and level 6 respectively. Use the following formula:
[0084] f(x)=f(x0)*(x1-x) / (x1-x0)+f(x1)*(x-x0) / (x1-x0) (8)
[0085] According to this formula, the scores corresponding to all elements in the level 5_5 list can be calculated, and the final level score can be obtained by averaging these scores.
[0086] Example 2
[0087] The present embodiment relates to a radius grade subdivision device based on deep clustering, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a radius grade subdivision method based on deep clustering described in Example 1.
[0088] Example 3
[0089] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a radius grade subdivision method based on deep clustering in embodiment 1 is implemented.
[0090] Finally, it should be noted that the above implementation modes are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned implementation modes, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned implementation modes, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the implementation modes of the present invention.
Claims
1. A radial grade subdivision method based on deep clustering, characterized in that: The following steps are involved: Step 1: Use principal component analysis (PCA) to reduce the dimension of the original data, and then use the autoencoder in the deep learning method to extract features of the radius; Step 2: Classify the features extracted in step 1 through a clustering algorithm (K-means) to complete the segmentation of the radius grade, and take the weighted sum of the reconstruction loss function and clustering loss function of the autoencoder in step 1 as the joint loss; Step 3: Using the idea of transfer learning, the MobileNet-V2 network is used to extract features from the subdivided levels, and the Euclidean distance between the subdivided levels and the adjacent levels is calculated. After standardizing the distance, the interpolation method is used to assign corresponding scores to the new levels generated after the subdivision.
2. A radius grade subdivision method based on deep clustering according to claim 1, characterized in that: The step 1 specifically includes: Step 1.1: Use principal component analysis (PCA) to reduce the dimension of the original data while retaining the most important features; then use the autoencoder in the deep learning method to extract features of the radius. The specific implementation method is as follows: The input data is low-dimensional data processed by principal component analysis (PCA). The autoencoder network mainly consists of two parts, namely the encoder part (Encoder) and the decoder part (Decoder); the encoder part consists of three modules, each module contains linear layers of different shapes, and a ReLU activation operation is added after each linear layer; the features obtained after the encoder are the input features of the clustering algorithm in step 2; The structure of the decoder is opposite to that of the encoder. It also consists of three modules. Each module contains linear layers of different shapes. A ReLU activation operation is added after each linear layer, and the final output is the reconstructed image. Step 1.2, design reconstruction loss function: The reconstructed loss function expression is as follows: where x i is the i-th data input, f(x i ) function is as follows: f(x i ) function itself represents a mapping relationship. It represents the network parameters in the encoder network. The encoding part of the autoencoder is used as a mapping function to map the original data to a low-dimensional space. i )) function is the function used to reconstruct the data, that is, the decoder part, and its complete function is parameter Represents the various parameters in the reconstruction network; the l function is a specific loss function to measure the reconstruction error. The least squares loss is used as the loss function of this part. Its original formula is as follows:
3. The method for radial grade subdivision based on deep clustering according to claim 1, characterized in that: The step 2 specifically includes: Step 2.1: Perform k-means clustering on the features after the dimension reduction process in step 1. In the k-means algorithm, Euclidean distance is usually used to calculate the distance between data points. The formula is as follows: Where (x1, y1) and (x2, y2) represent the coordinates of two points respectively. It is an iterative process from the beginning to the final determination of clustering. The clustering loss function is designed as follows: where s i is the assignment vector for data point i, which has only one non-zero element; this means that s i Is a sparse vector, which is used in the clustering algorithm to indicate to which cluster the data point i is assigned; the algorithm uses the assignment vector s i and centroid matrix M to determine which cluster the data point i belongs to; Step 2.2, design a joint loss function; The clustering loss of the clustering algorithm and the reconstruction loss of the autoencoder are combined and jointly optimized to ensure that the model can comprehensively consider the requirements of different tasks during training, so as to learn the feature representation and clustering assignment of the data more comprehensively. The joint loss function is the weighted addition of the above formulas (1) and (5), which is specifically expressed as follows: Here, λ is a hyperparameter used to balance the reconstruction error and the K-means clustering error. The value of λ is greater than or equal to 0, and its value is 1. After the above joint loss, the parameters in the model will be updated in a direction that is more conducive to clustering.
4. The method for radial grade subdivision based on deep clustering according to claim 1, characterized in that: The step 3 specifically includes: Step 3.1, using the idea of transfer learning, using the MobileNet-V2 network for feature extraction, only using the network structure and parameters of the feature extraction part; extracting features from the three levels of images obtained after clustering, and saving the extracted features in different feature vector lists to obtain three vector lists; then the vector lists of the newly added level are matched with each data in the two pairs of feature vector lists of the other two levels, and then converted into array form, and then the Euclidean distance is calculated. The calculation formula of the Euclidean distance is shown in formula (4); the distance obtained from each list is stored in their respective lists, and the distance between the three can be obtained; then the obtained Euclidean distance is standardized, and the standardized formula used is as follows: f(x)=(x-x min ) / (x max -x min ) (7) x min and x max are the minimum and maximum distances in the list, respectively, and x is the original distance value; after normalization, the original distance value can be mapped to the range of 0-1, where 0 corresponds to the minimum distance in the original value and 1 corresponds to the maximum distance in the original value, and f(x) is the distance after normalization; Step 3.2, according to the above standardized distance and then use the interpolation formula, we can get the score corresponding to the new level; since only the new level has no corresponding score, set an element in its list to x, and set the elements in the corresponding positions in the other two lists to x1 and x0 respectively, and the values of x1 and x0 correspond to the level scores of the lower level and the higher level respectively, using the following formula: f(x)=f(x0)*(x1-x) / (x1-x0)+f(x1)*(x-x0) / (x1-x0) (8) According to this formula, the scores corresponding to all elements in the newly added level list can be calculated, and the final level score can be obtained by taking the average of these scores.
5. A radial grade subdivision device based on deep clustering, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement a radial grade subdivision method based on deep clustering as described in any one of claims 1-4.
6. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a radius grade subdivision method based on deep clustering as described in any one of claims 1 to 4 is implemented.