Bone age CHN standard map optimization method and device oriented to intelligent evaluation

Through the improved deep learning network CS_Efficient and CHN method optimization standard maps, the time-consuming and laborious and subjective problems of traditional bone age assessment methods are solved. The attention mechanism and Focal Loss are used to optimize the EfficientNet-B0 network, and efficient and accurate bone age is achieved. Evaluation improves the accuracy and applicability of the evaluation.

CN119092089APending Publication Date: 2024-12-06ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411004598.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Traditional artificial bone age assessment methods are time-consuming and laborious and subjective. The existing intelligent bone age assessment methods have difficulties in selecting and building suitable neural network models, especially the lack of adaptability of the EfficientNet series network in bone age assessment.

Method used

By building an improved deep learning network CS_Efficient, combining CHN method to optimize the standard map, introducing attention mechanism and adaptive average pooling layer, using Focal Loss to solve the problem of data sample imbalance, and optimizing the EfficientNet-B0 network to improve the accuracy of bone age assessment and timeliness.

Benefits of technology

It has achieved efficient and accurate bone age assessment, significantly improved the accuracy and applicability of the assessment, and has high clinical application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119092089A_ABST
    Figure CN119092089A_ABST
Patent Text Reader

Abstract

The invention discloses a bone age assessment method and device based on a CHN method optimization standard map. The method comprises the steps that 1, a bone age assessment data set is prepared; step 2, the attention mechanism is introduced into the OfficientNet-B0 network; and step 3, adding a self-adaptive average pooling layer and a full connection layer for feature compression, and solving the problem of unbalanced number of data samples by using Focal Loss. According to the method, the improved deep learning network (CSEfficient) is constructed and combined with the CHN optimization standard, so that efficient and accurate bone age evaluation can be carried out.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method and device for optimizing a bone age CHN standard atlas for intelligent assessment. Background Art

[0002] The traditional method of bone age assessment by manual reading is time-consuming and laborious and consumes a lot of resources, and is highly subjective and has a high rate of misjudgment. With the development of artificial intelligence technology, the intelligent bone age assessment method based on computer technology is gradually replacing the traditional method, but it has the difficulty of selecting and building a suitable neural network model. In the prior art, the EfficientNet series network performs well in the field of image classification, so the present invention selects EfficientNet-B0 as the basic network in the bone age assessment task, and improves it to adapt to the CHN method standard atlas after the level optimization of the present invention, so as to improve the accuracy and timeliness of bone age assessment. Summary of the invention

[0003] The present invention proposes a bone age assessment method and device based on the CHN method optimized standard atlas, which can perform efficient and accurate bone age assessment by constructing an improved deep learning network (CS_Efficient) combined with the CHN optimization standard.

[0004] The first aspect of the present invention relates to a bone age assessment method based on the CHN method optimized standard atlas, and the specific steps are as follows:

[0005] Step 1: Prepare the bone age assessment dataset;

[0006] Step 2: Introduce the attention mechanism into the EfficientNet-B0 network;

[0007] Step 3: Add an adaptive average pooling layer and a fully connected layer for feature compression, and use Focal Loss to solve the problem of imbalance in the number of data samples.

[0008] Step 1 specifically includes:

[0009] 1) Collect data of 14 reference bones for grade identification, including proximal phalanx III, proximal phalanx V, metacarpal V, middle phalanx III, distal phalanx I, capitate, radius and hamate;

[0010] 2) In the grade recognition experiment, the data of each reference bone mentioned above were divided into a test set and a training set in a ratio of 2:8.

[0011] Step 2 specifically includes:

[0012] 1) Add channel attention mechanism. The goal of the channel attention mechanism is to emphasize important feature channels and suppress less important feature channels. Its structure is as followsFigure 2 As shown. On a given input feature map X, the channel attention mechanism can be described by the following formula:

[0013] A avg =GlobalAvgPool(X) (2-1)

[0014] A max =GlobalMaxPool(X) (2-2)

[0015] Among them A avg and A max Represent the feature vectors obtained by global average pooling and global maximum pooling respectively. F is connected through two fully connected layers with shared weights (1×1 convolution is usually used in actual implementation) avg and F max To process:

[0016] F avg =ReLU(W1(A avg )) (2-3)

[0017] F max =ReLU(W1(A max )) (2-4)

[0018] Among them, W1 is the weight of the fully connected layer. The final channel attention feature F ca By adding F avg and F max Add together and pass through a 1×1 convolution (W2) and the Sigmoid activation function, i.e., the σ function, to get:

[0019] F ca =σ(W2(F avg +F max )) (2-5)

[0020] 2) Adding a spatial attention mechanism, the spatial attention mechanism aims to highlight the key spatial regions in the image, so that the model can focus on the most important parts. Its structure is as follows Figure 3 As shown. The spatial attention mechanism generates two two-dimensional feature maps by performing average pooling and maximum pooling of the channel dimension on the input feature map X, which can be described by the equation of the process:

[0021] S avg =AvgPool channel (X) (2-6)

[0022] S max =MaxPool channel (X) (2-7)

[0023] Savg and S max Concatenate in the channel dimension and then process through a 1×1 convolution layer (or a convolution layer with other kernel sizes depending on the actual situation):

[0024] S merged =Concat(S avg ,S max ) (2-8)

[0025] S conv =Conv(S merged ) (2-9)

[0026] The final spatial attention map F sa Through S conv Applying the Sigmoid activation function yields:

[0027] F sa =σ(S conv ) (2-10)

[0028] Step 3 specifically includes:

[0029] 1) For feature value compression, an adaptive average pooling layer and a fully connected layer are added after the attention mechanism. The main purpose of this fully connected layer is to map the deep features to a new space, the dimension of which is specified by a variable bits. By changing the size of the variable bits, the network can be adapted to different reference bones. In addition, the activation function is used for further activation before outputting the final result.

[0030] 2) Focal Loss is used to effectively solve the problem of imbalanced number of data samples encountered in reference bone grade recognition. The specific formula of Focal Loss is as follows:

[0031] FL(p t )=-α t (1-p t ) γ log(p t ) (3-1)

[0032] Among them, p t is the probability of the model predicting the correct category, for the positive class p t =p, for negative class p t =1-p. α t is a weight factor to balance positive and negative samples. γ is an adjustment factor used to reduce the weight of easy-to-classify samples and determines the rate at which the weight of easy-to-classify samples decreases. As γ increases, the penalty for easy-to-classify samples increases, making the model pay more attention to difficult-to-classify samples.

[0033] The second aspect of the present invention relates to a bone age assessment device based on a CHN method optimized standard atlas, 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 the bone age assessment method based on the CHN method optimized standard atlas of the present invention.

[0034] A second aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the bone age assessment method of the present invention based on the CHN method optimized standard atlas.

[0035] The present invention has the following beneficial effects:

[0036] The present invention proposes a bone age assessment method based on the CHN method to optimize the standard atlas. By improving the neural network structure, enhancing the feature extraction capability, and optimizing the loss function to solve the data imbalance problem, efficient and accurate bone age assessment is achieved. In actual tests, this method significantly improves the accuracy and applicability of bone age assessment and has high clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is the EfficientNet-B0 network structure diagram of the present invention.

[0038] Figure 2 It is a diagram of the channel attention mechanism of the present invention.

[0039] Figure 3 It is a diagram of the spatial attention mechanism of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0041] Example 1

[0042] Reference Figure 1 to Figure 3 , a bone age assessment method based on the CHN method to optimize the standard atlas, the specific steps are as follows:

[0043] Step 1: Collect bone age assessment dataset;

[0044] Step 2: Introduce the attention mechanism into the EfficientNet-B0 network;

[0045] Step 3: Add an adaptive average pooling layer and a fully connected layer for feature compression, and use Focal Loss to solve the problem of imbalance in the number of data samples.

[0046] Step 1 specifically includes:

[0047] 1) The 14 reference bones collected for grade identification, including proximal phalanx III, proximal phalanx V, metacarpal V, middle phalanx III, distal phalanx I, capitate, radius and hamate, were graded and the standard atlas was optimized.

[0048] 2) In the grade recognition experiment, the data of each reference bone mentioned above were divided into a test set and a training set in a ratio of 2:8.

[0049] Step 2 specifically includes:

[0050] 1) Add channel attention mechanism. The goal of the channel attention mechanism is to emphasize important feature channels and suppress less important feature channels. Its structure is as follows Figure 2 As shown. On a given input feature map X, the channel attention mechanism can be described by the following formula:

[0051] A avg =GlobalAvgPool(X) (2-1)

[0052] A max =GlobalMaxPool(X) (2-2)

[0053] Among them A avg and A max Represent the feature vectors obtained by global average pooling and global maximum pooling respectively. F is connected through two fully connected layers with shared weights (1×1 convolution is usually used in actual implementation) avg and F max To process:

[0054] F avg =ReLU(W1(A avg )) (2-3)

[0055] F max =ReLU(W1(A max )) (2-4)

[0056] Among them, W1 is the weight of the fully connected layer. The final channel attention feature F ca By adding F avg and F max Add together and pass through a 1×1 convolution (W2) and the Sigmoid activation function, i.e., the σ function, to get:

[0057] F ca =σ(W2(F avg +F max )) (2-5)

[0058] 2) Adding a spatial attention mechanism, the spatial attention mechanism aims to highlight the key spatial regions in the image, so that the model can focus on the most important parts. Its structure is as follows Figure 3 As shown. The spatial attention mechanism generates two two-dimensional feature maps by performing average pooling and maximum pooling of the channel dimension on the input feature map X, which can be described by the equation of the process:

[0059] S avg =AvgPool channel (X) (2-6)

[0060] S max =MaxPool channel (X) (2-7)

[0061] S avg and S max Concatenate in the channel dimension and then process through a 1×1 convolution layer (or a convolution layer with other kernel sizes depending on the actual situation):

[0062] S merged =Concat(S avg ,S max ) (2-8)

[0063] S conv =Conv(S merged ) (2-9)

[0064] The final spatial attention map F sa Through S conv Applying the Sigmoid activation function yields:

[0065] F sa =σ(S conv ) (2-10)

[0066] Step 3 specifically includes:

[0067] 1) For feature value compression, an adaptive average pooling layer and a fully connected layer are added after the attention mechanism. The main purpose of this fully connected layer is to map the deep features to a new space, the dimension of which is specified by a variable bits. By changing the size of the variable bits, the network can be adapted to different reference bones. In addition, the activation function is used for further activation before outputting the final result.

[0068] 2) Focal Loss is used to effectively solve the problem of imbalanced number of data samples encountered in reference bone grade recognition. The specific formula of Focal Loss is as follows:

[0069] FL(p t )=-αt (1-p t ) γ log(p t ) (3-1)

[0070] Among them, p t is the probability of the model predicting the correct category, for the positive class p t =p, for negative class p t =1-p. α t is the weight factor for balancing positive and negative samples. γ is an adjustment factor used to reduce the weight of easy-to-classify samples and determines the rate at which the weight of easy-to-classify samples decreases. As γ increases, the penalty for easy-to-classify samples increases, making the model pay more attention to difficult-to-classify samples.

[0071] The present invention has the following beneficial effects:

[0072] The present invention proposes a bone age assessment method based on the CHN method to optimize the standard atlas. By improving the neural network structure, enhancing the feature extraction capability, and optimizing the loss function to solve the data imbalance problem, efficient and accurate bone age assessment is achieved. In actual tests, this method significantly improves the accuracy and applicability of bone age assessment and has high clinical application value.

[0073] Example 2

[0074] The present embodiment relates to a bone age assessment device based on a CHN method optimized standard atlas, 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 the bone age assessment method based on the CHN method optimized standard atlas of Example 1.

[0075] Example 3

[0076] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the bone age assessment method based on the CHN method optimized standard atlas of Example 1 is implemented.

[0077] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A bone age assessment method based on the CHN method to optimize the standard atlas, the specific steps are as follows: Step 1: Prepare the bone age assessment dataset; Step 2: Introduce the attention mechanism into the EfficientNet-B0 network; Step 3: Add an adaptive average pooling layer and a fully connected layer for feature compression, and use FocalLoss to solve the problem of imbalance in the number of data samples.

2. A bone age assessment method based on the CHN method optimized standard atlas as claimed in claim 1, characterized in that: Step 1 specifically includes: 11) Collect data of 14 reference bones for grade identification, including proximal phalanx III, proximal phalanx V, metacarpal V, middle phalanx III, distal phalanx I, capitate, radius and hamate; 12) In the grade recognition experiment, the data of each reference bone mentioned above were divided into a test set and a training set in a ratio of 2:

8.

3. A bone age assessment method based on the CHN method optimized standard atlas as claimed in claim 1, characterized in that: Step 2 specifically includes: 21) Add a channel attention mechanism. The goal of the channel attention mechanism is to emphasize important feature channels and suppress less important feature channels. Its structure is shown in Figure 2. On a given input feature map X, the channel attention mechanism can be described by the following formula: A avg =GlobalAvgPool(X) (2-1) A max =GlobalMaxPool(X) (2-2) Among them A avg and A max Represent the feature vectors obtained by global average pooling and global maximum pooling respectively; F is connected by two fully connected layers with shared weights (1×1 convolution is usually used in actual implementation) avg and F max To process: F avg =ReLU(W1(A avg )) (2-3) F max =ReLU(W1(A max )) (2-4) Among them, W1 is the weight of the fully connected layer; the final channel attention feature F ca By adding F avg and F max Add together and pass through a 1×1 convolution (W2) and the Sigmoid activation function, i.e., the σ function, to get: F ca =σ(W2(F avg +F max )) (2-5) 22) Add a spatial attention mechanism, which aims to highlight the key spatial areas in the image and make the model focus on the most important parts. Its structure is shown in Figure 3. The spatial attention mechanism generates two two-dimensional feature maps by performing average pooling and maximum pooling of the channel dimension on the input feature map X. The process can be described by the equation: S avg =AvgPool channel (X) (2-6) S max =MaxPool channel (X) (2-7) S avg and S max Concatenate in the channel dimension and then process through a 1×1 convolution layer (or a convolution layer with other kernel sizes depending on the actual situation): S merged =Concat(S avg ,S max ) (2-8) S conv =Conv(S merged ) (2-9) The final spatial attention map F sa Through S conv Applying the Sigmoid activation function yields: F sa =σ(S conv ) (2-10) 4. A bone age assessment method based on the CHN method optimized standard atlas as claimed in claim 1, characterized in that: Step 3 specifically includes: 31) In order to compress the feature values, an adaptive average pooling layer and a fully connected layer are added after the attention mechanism. The main purpose of this fully connected layer is to map the deep features to a new space, the dimension of which is specified by a variable bits; by changing the size of the variable bits, the network can be adapted to different reference bones; in addition, the activation function is used for further activation before outputting the final result; 32) Focal Loss is used to effectively solve the problem of imbalanced number of data samples encountered in reference bone grade recognition. The specific formula of Focal Loss is as follows: FL(p t )=-a t (1-p t ) γ log(p t ) (3-1) Among them, p t is the probability of the model predicting the correct category, for the positive class p t =p, for negative class p t =1-p; α t is a weight factor that balances positive and negative samples; γ is an adjustment factor used to reduce the weight of easy-to-classify samples and determines the rate at which the weight of easy-to-classify samples decreases; as γ increases, the penalty for easy-to-classify samples increases, making the model pay more attention to difficult-to-classify samples.

5. A bone age assessment device based on the CHN method to optimize the standard atlas, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the bone age assessment method based on the CHN method optimized standard atlas 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, the bone age assessment method based on the CHN method optimized standard atlas described in any one of claims 1-4 is implemented.