Multi-objective architecture search method for osteoporosis image recognition based on structural entropy sampling

Abstract

The present invention discloses a multi-objective architecture search method for osteoporosis image recognition based on structural entropy sampling, comprising the following steps: 1. Collecting and preprocessing lumbar spine X-ray image data containing bone density information, and dividing it into a training set and a validation set; 2. Initializing a population of neural network structures; 3. Decoding all individuals in the population into corresponding neural network architectures, and training each neural network using the X-ray image training set, using the accuracy of different category recognition and the number of neural network parameters as fitness values; 4. Performing mating pool selection, crossover mutation based on structural entropy sampling, and environment selection based on the individual fitness values; 5. Repeating the above process to perform heuristic network structure search, ending evolution when pre-set conditions are met, and outputting the optimal network structure; 6. Retraining the optimal network structure obtained using the training set. The present invention can automatically search for a convolutional neural network architecture that is suitable for the lumbar spine X-ray image dataset, thereby achieving lumbar spine X-ray image classification.

Description

Technical Field

[0001] The present invention belongs to the field of automated machine learning and neural network architecture search in computer / machine learning, and specifically is a multi-objective architecture search osteoporosis image recognition method based on structural entropy sampling. Background Art

[0002] Medical image recognition utilizes imaging tests such as X-rays, ultrasound, CT, and MRI to assist doctors. With the advancement of deep learning, convolutional neural networks (CNNs) have become a mainstream technology in image recognition, providing significant diagnostic support for medical image recognition tasks. Therefore, designing a neural network architecture with superior performance can significantly improve the accuracy of medical image recognition.

[0003] Existing neural architecture search design methods include: network architecture design methods based on manual design and automated network architecture search methods (NAS); however, network architecture design methods based on manual design often require a lot of professional knowledge and experimental time. Classic convolutional neural network structures such as the linear structure model VGGNet, the residual structure model ResNet, and the multi-branch structure model GoogLeNet are all carefully designed by professionals.

[0004] The automated network architecture search method (NAS) mainly includes three search strategies: reinforcement learning-based, gradient optimization-based, and evolutionary algorithm-based.

[0005] Reinforcement learning-based approaches primarily utilize reinforcement learning as an optimization algorithm. They model the operators in a convolutional neural network structure as sequences and input them into an RNN. These serve as parameters that reinforcement learning continuously optimizes, ultimately searching for optimal operator structures and corresponding connections. While these approaches are superior to the aforementioned classical networks, they are very time-consuming due to the vast search space.

[0006] Gradient optimization-based methods mainly model the network structure as decision variables in a multidimensional optimization problem, and then use the gradient method to optimize the parameters in the search space, so that the structure can be explored continuously in the search space instead of discrete search. This method is more interpretable.

[0007] The method based on evolutionary algorithms mainly models the network structure into population individuals and performs heuristic exploration based on multi-objective evolutionary algorithms. Since evolutionary algorithms have significant advantages in multi-objective and global optimization, this method often has good performance.

[0008] The above network structure search methods have advantages over traditional manually designed networks, but they are very time-consuming and tend to ignore the distribution of solutions in the search space. Summary of the Invention

[0009] In order to address the shortcomings of the existing technology, the present invention provides a multi-objective architecture search osteoporosis image recognition method based on structural entropy sampling, hoping to search for a more effective neural network structure for the recognition and classification of osteoporosis images, thereby improving the classification accuracy.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0011] The multi-objective architecture search osteoporosis image recognition method based on structural entropy sampling of the present invention is characterized by the following steps:

[0012] Step 0: Collect n lumbar spine X-ray images containing bone density labels and preprocess them to obtain the dataset D all , and divided into training set D tra and validation set D vall , and there are I categories of lumbar spine X-ray images;

[0013] Step 1. Construct a convolutional neural network structure A consisting of NC convolution blocks and RC downsampling blocks; each convolution block and downsampling block has the same number of nodes and is composed of m nodes, each node represents an operation, and the types of operations include convolution operation, downsampling operation, full connection operation and channel attention operation. The connection between nodes is represented by "1", otherwise it is represented by "0"; the step size of all operations in the convolution block is 1, and the step size of all operations in the downsampling block is 2; the neural network structure A is composed of NC convolution blocks and RC downsampling blocks stacked in sequence; the density of the neural network structure A is represented by the number of "1"s in the connection relationship of all nodes;

[0014] Step 2: Initialize the population P of the neural network structure t And perform iterative evolution to obtain the optimal network structure A best ;

[0015] Step 2.0, parameter definition, including: initial population size is N, candidate offspring population size is M, total evolutionary iterations is T, current iterations is t, maximum number of times each structure is trained is E max , the crossover probability is a, the mutation probability is b, and the interval for dividing the density of all individuals in the population is α;

[0016] Step 2.1, initialization t = 1; randomly initialize the tth generation population P of the convolutional neural network structure t , the t-th generation population P t It consists of the codes of N convolutional network structures, where the code corresponding to any convolutional network structure A is recorded as E, and the code E is A binary code of bits, where m represents the number of nodes in the convolution block and the downsampling block in the convolutional neural network structure A; the binary code includes the connection relationship between layers in the convolutional neural network structure and the operation type of the node;

[0017] Step 2.2: The t-th generation population P t The N codes in are mapped into the convolutional neural network structure information, thereby completing the decoding operation and building a complete convolutional neural network structure;

[0018] Step 2.3: Train the complete convolutional neural network structure and obtain the individual fitness value:

[0019] The training set D tra and validation set D val After horizontal flipping, angular rotation, Gaussian noise augmentation and pixel normalization, the processed training set D' tra The input is input into the complete convolutional neural network structure for forward propagation processing, and the stochastic gradient descent algorithm is used in the back propagation processing to the t-th generation population P corresponding to the neural network structure. t Each decoded neural structure in the iterative training is iteratively trained, and the cross entropy loss function is used to update the weight parameters of the convolutional neural network structure during the iterative training process until the maximum number of training times E is reached. max So far, N trained convolutional neural network structures are obtained;

[0020] The processed validation set D' val Input N trained convolutional neural network structures respectively to obtain the verification set D' val The N accuracy rates are used as the fitness values F1 of the N individuals, and the parameters of the trained convolutional neural network structure are used as the fitness value F2;

[0021] Step 2.4, mating pool selection:

[0022] In the tth generation, population P t Two parent individuals are randomly selected with replacement each time, and the individual with the larger fitness value F1 is retained using the binary tournament selection method, thereby obtaining the t-th generation mating pool P with a population size of N. t ';

[0023] Step 2.5: Individual crossover mutation based on structural entropy sampling:

[0024] Step 2.5.1, sequentially select the t-th generation mating pool P t 'Randomly select two parent individuals p1 and p2, perform coded crossover and mutation operations with crossover rate a and mutation rate b respectively, and thus obtain two offspring individuals o1 and o2;

[0025] Step 2.5.2: Repeat step 2.5.1 M / 2 times to obtain the t-th generation candidate offspring population Q with a population size of M. t ';

[0026] Step 2.5.3: Calculate the t-th generation parent population P according to the structural entropy formula shown in formula (1): t The initial structural entropy AE(P t ):

[0027]

[0028] In formula (1), F′ max and F′ min Indicates the maximum density value and the minimum density value of all individuals in the current population, α represents the interval of equal density values of all individuals in the current population, Q represents the number of intervals divided according to the equal density intervals of the individual, and N represents the size of the population. represents the jth interval of the tth generation population, It represents the proportion of the number of individuals in the jth interval in the tth generation population to the total number of the population;

[0029] Step 2.5.4: The t-th generation candidate offspring population Q t The kth individual q in ' k Join the tth generation parent population P t Constitute the mixed population P′ t , and then calculate the mixed population P' according to formula (1) t The structural entropy AE(P′ t );

[0030] Step 2.5.5, calculate the initial structural entropy AE (P t ) and the mixed population P′ t The structural entropy AE(P′ t ) to obtain the kth individual q k Individual entropy IE(q k ), used to indicate the quality of individual diversity;

[0031] Step 2.5.6: Calculate the candidate offspring population Q according to steps 2.5.3-2.5.5 t 'The individual entropy of each individual in is sorted in descending order, and the first N individuals are selected as the t-th generation offspring population Q t ;

[0032] Step 2.6, environment selection:

[0033] The t-th generation parent population P t and the t-th generation offspring population Qt Mix and get the t-th generation mixed population M with a population size of 2N t ; For the t-th generation mixed population M t All individuals in the t-th generation are sorted non-dominated based on the dual-objective fitness values F1 and F2, and the Pareto frontier surface index K corresponding to the j-th individual in the target space is obtained. j And the crowding distance D on the corresponding frontier surface j ; Select N individuals with smaller Pareto front surface index and larger crowding distance as the t+1 generation population P t+1 ;

[0034] Step 2.7: After assigning t+1 to t, determine whether t≤T holds. If so, repeat steps 2.2 to 2.7; otherwise, proceed to step 3.

[0035] Step 3: Get the optimal neural network structure individual:

[0036] Based on the fitness values of N individuals, the T-generation population P T Sort in descending order and take the individual with the largest fitness value F1 as the optimal convolutional neural network structure A best Output;

[0037] Step 4: The optimal convolutional neural network structure A best To train:

[0038] The processed training set D' tra and validation set D' val Input to the optimal convolutional neural network structure A best And use the stochastic gradient descent algorithm to optimize the optimal convolutional neural network structure A best Training is performed while using the weighted cross entropy loss function as shown in formula (2) The optimal convolutional neural network structure A in the training iteration best The weight parameters are updated until the total number of training times E is reached max So far, the optimal convolutional neural network structure A' after training is obtained best It is also used as a bone density classification prediction model to achieve classification prediction of X-ray images;

[0039]

[0040] In formula (2), y represents the true label of the input sample, represents the label predicted by the network model, Represents the label probability predicted by the network structure, N c Indicates the number of images in the cth category, N sumIndicates the total number of all images.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The present invention designs a structural entropy index to measure the distribution of each individual in the population (each individual represents a convolutional neural network structure) in the search space, and based on structural entropy sampling, it maintains the diversity and distribution of the population as much as possible without losing accuracy during the iteration process, so that the algorithm can more efficiently explore multiple potential optimal neural network structures in the search space during the search process. At the same time, in the search evolution process, both model accuracy and network structure complexity are considered, so that a neural network with a small number of parameters but high image recognition accuracy can be searched.

[0043] 2. The convolutional neural network structure searched by the present invention can better adaptively learn image features on a given osteoporosis data set, and the network structure searched by this method performs well in both parameter quantity and accuracy. The network structure that can be searched by the method designed by the present invention has the advantage of a small number of parameters and can obtain a higher osteoporosis classification accuracy on lumbar spine X-ray images.

[0044] 3. The network structure searched by the present invention has low redundancy and a small number of parameters, and can be easily integrated into the existing PACS system to achieve accurate classification of osteoporosis images. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 Diagram of the convolutional neural network structure designed for this invention. DETAILED DESCRIPTION

[0047] like Figure 1 As shown, in this embodiment, a multi-objective architecture search osteoporosis image recognition method based on structural entropy sampling is performed in the following steps:

[0048] Step 0: Collect a dataset D consisting of n lumbar spine X-ray images with bone density labels all , and after preprocessing operations, the training set D is obtained tra and validation set D val , and the lumbar spine X-ray images are divided into three categories of labels: normal bone mass, osteopenia and abnormal bone mass;

[0049] Step 0.1: Obtain n X-ray images, including any one X-ray image (h0, w0, c, l); where h0, w0, c, l represent the initial height, width, channel depth, and label category of the X-ray image, respectively;

[0050] Step 0.2: Input the X-ray image (h0, w0, c, l) into the preprocessing function R(h, w) for resizing, i.e., adjusting the image size to obtain the preprocessed X-ray image (h, w, c, l); where h and w represent the height and width of the image after the resizing operation, respectively;

[0051] Step 0.3: Input n X-ray images into the partitioning function S(r). The partitioning function S(r) will shuffle the input X-ray image data (i.e., disrupt the order of the images) and output the training set D. tra and validation set D val ; Among them, r represents the ratio of the training set and the validation set.

[0052] Step 1. Construct a neural network structure A consisting of NC convolution blocks and RC downsampling blocks; each convolution block and downsampling block has the same number of nodes and is composed of m nodes. Each node represents an operation. The types of operations include convolution operation, downsampling operation, full connection operation and channel attention operation. The connection between nodes is represented by 1, otherwise it is represented by 0; the step length of all operations in the convolution block is 1, and the step length of all operations in the downsampling block is 2; the neural network structure A is composed of NC convolution blocks and RC downsampling blocks stacked in sequence; the density of the neural network structure A is represented by the number of 1s between nodes. The more 1s there are, the denser the node connections are; set the parameters NC=6, RC=2, and m=6. Figure 2 As shown;

[0053] Step 2: Initialize the population P of the neural network structure t And perform iterative evolution to obtain the optimal neural network structure A best ;

[0054] Step 2.0, parameter definition and initialization, including: initial population size N = 20, sampling population size M = 100, total evolution iterations T = 25, current iteration number t = 1, maximum number of training times for each structure E max =100, crossover probability a=0.3, mutation probability b=0.2, density partition interval is α=5, number of categories I=3;

[0055] Step 2.1: Randomly initialize the t-th generation population P of the neural network structure t , the t-th generation population P t It consists of the codes of N neural network structures, where the code corresponding to any neural network structure A is recorded as E. The code E is A binary code of bits, where m represents the number of nodes in the convolution block and the downsampling block in the neural network structure A; the binary code includes the connection relationship between layers in the neural network structure and the operation type of the node;

[0056] Step 2.2: For the t-th generation population P t The N individual codes in are mapped into neural network structure information, thereby completing the decoding of each individual code and building a complete convolutional neural network structure;

[0057] Step 2.3: Train the complete convolutional neural network structure and obtain the individual fitness value:

[0058] The training set D tra and validation set D val After horizontal flipping, angular rotation, Gaussian noise augmentation and pixel normalization, the processed training set D' tra The input is input into the complete convolutional neural network structure for forward propagation processing, and the stochastic gradient descent (SGD) algorithm is used in the back propagation processing to perform the t-th generation population P corresponding to the neural network structure. t Each decoded neural structure in the iterative training is iteratively trained, and the cross entropy loss function is used to update the weight parameters of the convolutional neural network structure during the iterative training process until the maximum number of training times E is reached. max So far, N trained convolutional neural network structures are obtained;

[0059] The processed validation set D' val Input N trained convolutional neural network structures respectively to obtain the verification set D' val The N accuracy rates are used as the fitness values F1 of the N individuals, and the parameters of the trained convolutional neural network structure are used as the fitness value F2;

[0060] Step 2.4, mating pool selection:

[0061] In the tth generation, population P t Two parent individuals are randomly selected with replacement each time, and the individual with the larger fitness value F1 is retained using the binary tournament selection method, thereby obtaining the t-th generation mating pool P with a population size of N. t ';

[0062] Step 2.5: Individual crossover mutation based on structural entropy sampling:

[0063] Step 2.5.1, sequentially select the t-th generation mating pool P t'Randomly select two parent individuals p1 and p2, perform coded crossover and mutation operations with crossover rate a and mutation rate b respectively, and thus obtain two offspring individuals o1 and o2;

[0064] Step 2.5.2: Repeat step 2.5.1 M times to obtain the t-th generation candidate offspring population Q with a population size of 2M. t ';

[0065] Step 2.5.3: Calculate the t-th generation parent population P according to the structural entropy formula shown in formula (1): t The initial structural entropy AE(P t ):

[0066]

[0067] In formula (1), P t represents the t-generation population, where F' represents the node connection density of the neural network individual, F' max and F min Indicates the maximum density value and the minimum density value of all individuals in the current population, α represents the interval of equal density values of all individuals in the current population, Q represents the number of intervals divided according to the equal density intervals of the individual, and N represents the size of the population. represents the jth interval of the tth generation population, It represents the proportion of the number of individuals in the jth interval in the tth generation population to the total number of the population;

[0068] Step 2.5.4: The t-th generation candidate offspring population Q t The kth individual q in ' k Join the tth generation parent population P t Constitute the mixed population P′ t , and then calculate the mixed population P' according to formula (1) t The structural entropy AE(P′ t );

[0069] Step 2.5.5, calculate the initial structural entropy AE (P t ) and the mixed population P′ t The structural entropy AE(P′ t ), thereby obtaining the kth individual q k Individual entropy IE(q k ), used to indicate the quality of individual diversity;

[0070] Step 2.5.6: Calculate the candidate offspring population Q according to steps 2.5.3-2.5.5 t 'The individual entropy of each individual in is sorted in descending order, and the first N individuals are selected as the t-th generation offspring population Qt ;

[0071] Step 2.6, environment selection:

[0072] The t-th generation parent population P t and the t-th generation offspring population Q t Mix and get the t-th generation mixed population M with a population size of 2N t ; For the t-th generation mixed population M t All individuals in the t-th generation are sorted non-dominated based on the dual-objective fitness values F1 and F2, and the Pareto frontier surface index K corresponding to the j-th individual in the target space is obtained. j And the crowding distance D on the corresponding frontier surface j ; Select N individuals with smaller Pareto front surface index and larger crowding distance as the t+1 generation population P t+1 ;

[0073] Step 2.7: After assigning t+1 to t, determine whether t≤T holds. If so, repeat steps 2.2 to 2.7; otherwise, proceed to step 3.

[0074] Step 3: Get the optimal neural network structure individual:

[0075] Based on the fitness values of N individuals, the T-generation population P T Sort in descending order and take the individual with the largest fitness value F1 as the optimal convolutional neural network structure A best Output;

[0076] Step 4: The optimal convolutional neural network structure A best Training: The processed training set D' tra and validation set D' val Input to the optimal convolutional neural network structure A best And use the Stochastic Gradient Descent (SGD) algorithm to optimize the optimal convolutional neural network structure A best Training is performed while using the weighted cross entropy loss function as shown in formula (2) The optimal convolutional neural network structure A in the training iteration best The weight parameters are updated until the total number of training times E is reached max So far, the optimal convolutional neural network structure A' after training is obtained best It is also used as a bone density classification prediction model to achieve classification prediction of X-ray images;

[0077]

[0078] In formula (2), y represents the true label of the input sample, represents the label predicted by the network model, Represents the label probability predicted by the network structure, where N c Indicates the number of images in the cth category, N sum Indicates the total number of all images.

[0079] Step 5. Save the trained neural network model A best And the corresponding weight parameters are used to obtain the bone density classification prediction model, thereby realizing the bone type classification prediction of X-ray images.

[0080] Table 1

[0081]

[0082] As shown in Table 1, the convolutional neural network structure obtained by the architecture search method based on structural entropy sampling of the present invention and the osteoporosis image classification accuracy of other different models are shown. AEMONAS represents the convolutional neural network obtained by the architecture search method based on structural entropy sampling of the present invention; ResNet and DenseNet represent artificially designed convolutional neural networks; ENAS represents the convolutional neural network obtained by the network structure search method based on reinforcement learning; DARTS represents the convolutional network structure obtained by the network structure search method based on gradient optimization; NSGANet and AmoebaNet represent the convolutional neural networks obtained by the network structure search method based on evolutionary algorithm. It can be seen that the method of the present invention has significantly improved in accuracy compared with other methods, and the number of model parameters is smaller, which is more advantageous and convenient to integrate in an environment with limited computing resources.

Claims

1. A multi-objective architecture search osteoporosis image recognition method based on structural entropy sampling, characterized by Follow these steps: Step 0: Collect n lumbar spine X-ray images containing bone density labels and preprocess them to obtain the dataset D all , and divided into training set D tra and validation set D vall , and there are I categories of lumbar spine X-ray images; Step 1: Construct a convolutional neural network structure A consisting of NC convolution blocks and RC downsampling blocks; each convolution block and downsampling block has the same number of nodes and is composed of m nodes, each node represents an operation, and the types of operations include convolution operation, downsampling operation, full connection operation and channel attention operation. The connection between nodes is represented by "1", otherwise it is represented by "0"; the step size of all operations in the convolution block is 1, and the step size of all operations in the downsampling block is 2; the neural network structure A is composed of NC convolution blocks and RC downsampling blocks stacked in sequence; the density of the neural network structure A is represented by the number of "1"s in the connection relationship of all nodes; Step 2: Initialize the population P of the neural network structure t And perform iterative evolution to obtain the optimal network structure A best ; Step 2.0, parameter definition, including: initial population size is N, candidate offspring population size is M, total evolutionary iterations is T, current iterations is t, maximum number of times each structure is trained is E max , the crossover probability is a, the mutation probability is b, and the interval for dividing the density of all individuals in the population is α; Step 2.1, initialization t = 1; randomly initialize the tth generation population P of the convolutional neural network structure t , the t-th generation population P t It consists of the codes of N convolutional network structures, where the code corresponding to any convolutional network structure A is recorded as E, and the code E is A binary code of bits, where m represents the number of nodes in the convolution block and the downsampling block in the convolutional neural network structure A; the binary code includes the connection relationship between layers in the convolutional neural network structure and the operation type of the node; Step 2.2: The t-th generation population P t The N codes in are mapped into the convolutional neural network structure information, thereby completing the decoding operation and building a complete convolutional neural network structure; Step 2.3: Train the complete convolutional neural network structure and obtain the individual fitness value: The training set D tra and validation set D val After horizontal flipping, angular rotation, Gaussian noise augmentation and pixel normalization, the processed training set D' tra The input is input into the complete convolutional neural network structure for forward propagation processing, and the stochastic gradient descent algorithm is used in the back propagation processing to the t-th generation population P corresponding to the neural network structure. t Each decoded neural structure in the iterative training is iteratively trained, and the cross entropy loss function is used to update the weight parameters of the convolutional neural network structure during the iterative training process until the maximum number of training times E is reached. max So far, N trained convolutional neural network structures are obtained; The processed validation set D' val Input N trained convolutional neural network structures respectively to obtain the verification set D' val The N accuracy rates are used as the fitness values F1 of the N individuals, and the parameters of the trained convolutional neural network structure are used as the fitness value F2; Step 2.4, mating pool selection: In the tth generation, population P t Two parent individuals are randomly selected with replacement each time, and the individual with the larger fitness value F1 is retained using the binary tournament selection method, thereby obtaining the t-th generation mating pool P with a population size of N. t '; Step 2.5: Individual crossover mutation based on structural entropy sampling: Step 2.5.1, sequentially select the t-th generation mating pool P t 'Randomly select two parent individuals p1 and p2, perform coded crossover and mutation operations with crossover rate a and mutation rate b respectively, and thus obtain two offspring individuals o1 and o2; Step 2.5.2: Repeat step 2.5.1 M / 2 times to obtain the t-th generation candidate offspring population Q with a population size of M. t '; Step 2.5.3: Calculate the t-th generation parent population P according to the structural entropy formula shown in formula (1): t The initial structural entropy AE(P t ): In formula (1), F′ max and F′ min Indicates the maximum density value and the minimum density value of all individuals in the current population, α represents the interval of equal density values of all individuals in the current population, Q represents the number of intervals divided according to the equal density intervals of the individual, and N represents the size of the population. represents the jth interval of the tth generation population, It represents the proportion of the number of individuals in the jth interval in the tth generation population to the total number of the population; Step 2.5.4: The t-th generation candidate offspring population Q t The kth individual q in ' k Join the tth generation parent population P t Constitute the mixed population P′ t , and then calculate the mixed population P' according to formula (1) t The structural entropy AE(P′ t ); Step 2.5.5, calculate the initial structural entropy AE (P t ) and the mixed population P′ t The structural entropy AE(P′ t ) to obtain the kth individual q k Individual entropy IE(q k ), used to indicate the quality of individual diversity; Step 2.5.6: Calculate the candidate offspring population Q according to steps 2.5.3-2.5.5 t 'The individual entropy of each individual in is sorted in descending order, and the first N individuals are selected as the t-th generation offspring population Q t ; Step 2.6, environment selection: The t-th generation parent population P t and the t-th generation offspring population Q t Mix and get the t-th generation mixed population M with a population size of 2N t ; For the t-th generation mixed population M t All individuals in the t-th generation are sorted non-dominated based on the dual-objective fitness values F1 and F2, and the Pareto frontier surface index K corresponding to the j-th individual in the target space is obtained. j And the crowding distance D on the corresponding frontier surface j ; Select N individuals with smaller Pareto front surface index and larger crowding distance as the t+1 generation population P t+1 ; Step 2.7: After assigning t+1 to t, determine whether t≤T holds. If so, repeat steps 2.2 to 2.7; otherwise, proceed to step 3. Step 3: Get the optimal neural network structure individual: Based on the fitness values of N individuals, the T-generation population P T Sort in descending order and take the individual with the largest fitness value F1 as the optimal convolutional neural network structure A best Output; Step 4: The optimal convolutional neural network structure A best To train: The processed training set D' tra and validation set D' val Input to the optimal convolutional neural network structure A best And use the stochastic gradient descent algorithm to optimize the optimal convolutional neural network structure A best Training is performed while using the weighted cross entropy loss function as shown in formula (2) The optimal convolutional neural network structure A in the training iteration best The weight parameters are updated until the total number of training times E is reached max So far, the optimal convolutional neural network structure A' after training is obtained best It is also used as a bone density classification prediction model to achieve classification prediction of X-ray images; In formula (2), y represents the true label of the input sample, represents the label predicted by the network model, Represents the label probability predicted by the network structure, N c Indicates the number of images in the cth category, N sum Indicates the total number of all images.

Images