A method for identifying benign and malignant breast tumors based on RepCSC-Net

Abstract

The present invention relates to the field of smart medical technology, and specifically to a method for identifying benign and malignant breast tumors based on RepCSC-Net, comprising: obtaining a breast mammography image data set, performing image enhancement processing on the breast mammography image data set, and dividing it into a training set, a validation set, and a test set; constructing a RepCSC module with a structural reparameterization strategy to obtain a RepCSC-Net model; obtaining a trained RepCSC-Net model, and then obtaining a verified RepCSC-Net model; inputting the test set into the verified RepCSC-Net model, and outputting the recognition result of the breast tumor. The present invention effectively improves the accuracy of identifying benign and malignant breast tumors by constructing a RepCSC-Net model, and can be applied to the field of identifying benign and malignant tumors.

Description

Technical Field

[0001] The present invention relates to the field of smart medical technology, and in particular to a method for identifying benign and malignant breast tumors based on RepCSC-Net. Background Art

[0002] Breast tumors are relatively common diseases in clinical practice, mainly including benign tumors and malignant tumors. Benign tumors such as breast fibroids and malignant tumors such as breast cancer. Regarding the identification of benign and malignant breast tumors, the traditional identification method is mainly imaging diagnosis, such as breast X-ray examination, breast MRI examination and breast color ultrasound examination. In imaging diagnosis, doctors need to spend a lot of energy on image comparison diagnosis, but in the case of dense and abundant breast glands, it is difficult to accurately distinguish between glands and lesions. In order to better assist doctors in the identification of benign and malignant breast tumors, computer technology and related technologies are currently introduced into the field of smart medical care. For example, deep neural networks (DNN) have been widely used in the field of benign and malignant tumor identification and are showing great potential.

[0003] A deep neural network is a multi-layer unsupervised neural network that uses the output features of the previous layer as the input of the next layer for feature learning. After layer-by-layer feature mapping, the features of the existing spatial samples are mapped to another feature space to learn the existing input and obtain better feature expression. However, deep neural networks have the defects of simple activation functions, inability to learn complex benign and malignant features of breast tumors, and easy loss of relevant information, which leads to poor recognition accuracy of benign and malignant breast tumors. Summary of the invention

[0004] In order to solve the above-mentioned technical problem of poor accuracy in identifying benign and malignant breast tumors, the object of the present invention is to provide a method for identifying benign and malignant breast tumors based on RepCSC-Net. The technical solution adopted is as follows:

[0005] An embodiment of the present invention provides a method for identifying benign and malignant breast tumors based on RepCSC-Net, the method comprising the following steps:

[0006] Step S1, obtaining a breast mammography image dataset, performing image enhancement processing on the breast mammography image dataset, and dividing the dataset into a training set, a validation set, and a test set;

[0007] Step S2, constructing a RepCSC module with a structure reparameterization strategy, combining the RepCSC module with a convolutional neural network RepVGG to obtain a RepCSC-Net model;

[0008] Step S3, using the training set to train the RepCSC-Net model to obtain a trained RepCSC-Net model;

[0009] Step S4, using the verification set, verifying the trained RepCSC-Net model to obtain a verified RepCSC-Net model;

[0010] Step S5, inputting the test set into the verified RepCSC-Net model, and outputting the recognition result of the breast tumor.

[0011] Furthermore, the content of step S2 includes:

[0012] A RepCSC module with a structural reparameterization strategy is constructed, and the optimization target of the RepCSC module is determined; the optimization target is solved by using an iterative soft threshold algorithm, so that the RepCSC module during training has a three-channel structure; during training, the three-channel structure of the RepCSC module is converted into a single-channel structure by merging dictionaries to obtain a new dictionary; and the first layer of the convolutional neural network RepVGG is replaced by the designed RepCSC module to obtain a RepCSC-Net model.

[0013] Furthermore, the input information of the RepCSC module is represented as follows:

[0014]

[0015] Where x∈R L×W×C is the input signal, L×W is the spatial dimension of the input signal, C is the number of channels of the input signal, d k is the kth atom in the dictionary, s k is the sparse coding coefficient corresponding to the kth atom in the dictionary, and * is the convolution operation.

[0016] Furthermore, the optimization objective is expressed as:

[0017]

[0018] Where D = (d1, d2, ..., d C ) is the dictionary matrix, d C is the Cth atom in the dictionary, Γ=(s1,s2,…,s C ) T is the set of sparse coding coefficients, s C is the sparse coding coefficient corresponding to the Cth atom in the dictionary, X = (x1, x2, ..., x N ) T is the input information set, x Nis the Nth input information, min is the minimization function, λ is the sparsity parameter, is the reconstruction error, λ‖Γ‖1 is the sparsity constraint, and ‖‖ is the norm sign.

[0019] Furthermore, the optimization objective is solved by using an iterative soft threshold algorithm, and the iterative formula is:

[0020]

[0021] Among them, Γ k+1 is the k+1th sparse coding coefficient set, S β / L is the soft threshold operator, β is the sparsity parameter, is the iteration step length, D T is the transpose of the dictionary matrix, X is the input information set, Γ k is the kth sparse coding coefficient set, and D is the dictionary matrix;

[0022] By G 0 =X for the Γ k+1 Perform initialization processing to obtain a new iteration formula, which is:

[0023]

[0024] make C=-L, the new iterative formula is processed, and the new iterative formula after processing is: Γ 1 =S β / L (A T X+X+CA T AX).

[0025] Furthermore, the new dictionary is represented as follows:

[0026]

[0027] Among them, D ′ For the new dictionary, is the iteration step length, D T is the transpose of the dictionary matrix, I is the d×d identity matrix, and D is the dictionary matrix.

[0028] Furthermore, the content of step S1 includes:

[0029] Import the mammography image dataset and then read the mammography image dataset;

[0030] The image size of each image in the mammography dataset is adjusted, and the Lucy-Richardson algorithm is used to perform image enhancement on the mammography dataset.

[0031] According to the image quantity ratio specified by the user, the breast mammography image dataset after image enhancement is divided into a training set, a validation set, and a test set, and saved as corresponding XML tag files.

[0032] Furthermore, the content of step S3 includes:

[0033] Setting the number of iterations required for the training set and obtaining a preset number of mammographic images from the training set;

[0034] The Softmax classifier is used to receive the features output by the fully connected layer, output the classification prediction results after the probability value mapping of the label corresponding to the input target, and calculate the loss function value of the network model for each round of iteration until the minimum loss function value is obtained to obtain the trained RepCSC-Net model.

[0035] Furthermore, after obtaining the trained RepCSC-Net model, the last layer of the fully connected layer is used as the final extracted feature of the RepCSC-Net model.

[0036] Furthermore, the breast tumor recognition result includes tumor classification information and classification accuracy.

[0037] The present invention has the following beneficial effects:

[0038] The present invention provides a method for identifying benign and malignant breast tumors based on RepCSC-Net. The method applies the structural re-parameterization strategy to convolutional sparse coding to construct a new RepCSC structure. At the same time, the RepCSC structure is combined with the convolutional neural network RepVGG to obtain a RepCSC-Net network. The RepCSC-Net network has the characteristics of strong learning ability, fast reasoning speed and good interpretability. The RepCSC-Net network is applied to the identification of benign and malignant breast tumors, which helps to improve the accuracy of identifying benign and malignant breast tumors and obtain more accurate identification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0040] Figure 1 This is a flow chart of a method for identifying benign and malignant breast tumors based on RepCSC-Net in the present invention;

[0041] Figure 24 is a structural diagram of the RepCSC module in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0043] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0044] This embodiment provides a method for identifying benign and malignant breast tumors based on RepCSC-Net. Figure 1 As shown, the following steps are included:

[0045] Step S1, obtaining a mammography image dataset, performing image enhancement processing on the mammography image dataset, and dividing the dataset into a training set, a validation set, and a test set, the steps may include:

[0046] The first step is to import the mammography image dataset and then read the mammography image dataset.

[0047] In this embodiment, a breast mammography image dataset is collected using professional X-ray equipment in a hospital.

[0048] In the second step, the image size of each image in the mammography dataset is adjusted, and the Lucy-Richardson algorithm is used to perform image enhancement processing on the mammography dataset.

[0049] In this embodiment, since the field of view of the device is limited and the shaking or focusing of the instrument affects the clarity of the photo or video, it is necessary to perform image enhancement processing on each image in the mammography image dataset, that is, the image is iterated and deblurred 10 times using the Lucy-Richardson algorithm, and the image size of each image is adjusted to obtain the mammography image dataset after image enhancement processing.

[0050] The specific calculation formula for deblurring an image using the Lucy-Richardson algorithm can be:

[0051]

[0052] Where f′(x,y) is the pixel value corresponding to the pixel at position (x,y) in the deblurred image, f(x,y) is the pixel value corresponding to the pixel at position (x,y) in the original image, g(x,y) is the pixel value corresponding to the pixel at position (x,y) in the degraded image, h(x,y) is the pixel value corresponding to the pixel at position (x,y) in the spatial function of the degradation function, * is the convolution operation, h(x,y) T is the transpose of the pixel value corresponding to the pixel at position (x, y).

[0053] It should be noted that the image enhancement processing method also includes: contrast transformation, spatial filtering and color transformation, etc. The implementation process of image enhancement processing is a prior art and is not within the scope of protection of the present invention, and will not be elaborated in detail here. The image size adjustment of the image can be set by the implementer according to the specific actual situation without specific limitation.

[0054] The third step is to divide the breast mammography image dataset after image enhancement into a training set, a validation set, and a test set according to the image quantity ratio specified by the user, and save them as corresponding extensible markup language (XML) markup files.

[0055] Step S2, construct a RepCSC module with a structural reparameterization strategy, and combine the RepCSC module with the convolutional neural network RepVGG to obtain the RepCSC-Net model.

[0056] First of all, it should be noted that the RepCSC module includes structural reparameterization (Rep) and convolutional sparse coding (CSC). The RepCSC module is a convolutional sparse coding module with a structural reparameterization strategy. The convolutional sparse coding module considers adding the idea of ​​sparse coding on the basis of the convolutional neural network (CNN), that is, decomposing the input signal into a linear combination of several convolution kernels. Compared with standard sparse coding, convolutional sparse coding takes into account the timing and locality of the signal, making it more effective in processing time series data, which helps to improve the accuracy of the output results of the RepCSC-Net model. The RepCSC module extracts the feature representation of the image by learning a set of sparse filters, specifically: the sparse filters can perform convolution operations on the input image to obtain a set of convolutional feature maps. Compared with traditional convolutional neural networks, the RepCSC-Net model can obtain a more compact and sparse feature representation, so as to better adapt to image processing tasks. Its steps may include:

[0057] The first step is to construct the RepCSC module of the structure reparameterization strategy and determine the optimization target of the RepCSC module.

[0058] In the first sub-step, the input information of the RepCSC module is represented as:

[0059]

[0060] Where x∈R L×W×C is the input signal, L×W is the spatial dimension of the input signal, C is the number of channels of the input signal, d k is the kth atom in the dictionary, one atom corresponds to one convolution kernel, s k is the sparse coding coefficient corresponding to the kth atom in the dictionary, and * is the convolution operation.

[0061] In the second sub-step, the optimization target can be obtained by minimizing the original loss function, where the optimization target is expressed as:

[0062]

[0063] The above optimization objective is simplified as:

[0064]

[0065] Where D = (d1, d2, ..., d C ) is the dictionary matrix, d C is the Cth atom in the dictionary, Γ=(s1,s2,…,s C ) T is the set of sparse coding coefficients, s C is the sparse coding coefficient corresponding to the Cth atom in the dictionary, X = (x1, x2, ..., x N ) T is the input information set, x N is the Nth input information, min is the minimization function, λ is the sparsity parameter, is the reconstruction error, λ‖Γ‖1 is the sparsity constraint, and ‖‖ is the norm sign.

[0066] The second step is to use an iterative soft threshold algorithm to solve the optimization target so that the RepCSC module during training is a three-channel structure. The steps may include:

[0067] In the first sub-step, the optimization target can use the iterative soft thresholding algorithm (ISTA) to obtain an iterative formula, which can be:

[0068]

[0069] Among them, Γ k+1 is the k+1th sparse coding coefficient set, S β / L is the soft threshold operator, β is the sparsity parameter, is the iteration step length, D T is the transpose of the dictionary matrix, X is the input information set, Γ k is the kth sparse coding coefficient set, and D is the dictionary matrix.

[0070] The second substep is to use Γ 0 =X for the Γ k+1 Perform initialization processing to obtain a new iteration formula, which is:

[0071] The third sub-step is to C=-L, the new iterative formula is processed, and the new iterative formula after processing is: Γ 1 =S β / L (A T X+X+CA T AX).

[0072] The third step is to transform the three-channel structure of the RepCSC module into a single-channel structure by merging dictionaries during training to obtain a new dictionary.

[0073] In this embodiment, after obtaining the three-path network structure during training, it is necessary to consider converting the three-path structure into a single-path structure during inference, that is, by merging dictionaries. The representation of the new dictionary is:

[0074]

[0075] Among them, D ′ For the new dictionary, is the iteration step length, D T is the transpose of the dictionary matrix, I is the d×d identity matrix, and D is the dictionary matrix.

[0076] It is worth noting that the iterative formula transforms Γ 1 S β / L (D′X), the network at this time is a single-path structure, which completes the goal of training the three-path reasoning single-path RepCSC structure. The structure diagram of the RepCSC module is as follows Figure 2 shown.

[0077] The fourth step is to use the designed RepCSC module to replace the first layer of the convolutional neural network RepVGG to obtain the RepCSC-Net model.

[0078] In this embodiment, in order to incorporate sparse modeling into the RepVGG network architecture, the first layer of the convolutional neural network RepVGG is replaced with the designed RepCSC module. At the same time, all other layers in the convolutional neural network RepVGG, such as normalization layers, nonlinear layers, and fully connected layers, are retained.

[0079] It should be noted that the RepCSC-Net model allows the CSC layer to have wide applicability. The RepVGG network with the RepCSC layer is called RepCSC-Net. The trade-off factors of all RepCSC modules are set to 0.03, and the forward propagation of each RepCSC layer is iteratively performed through the soft threshold algorithm.

[0080] Step S3, using the training set to train the RepCSC-Net model to obtain a trained RepCSC-Net model, the steps may include:

[0081] The first step is to set the number of iterations required for the training set and obtain a preset number of breast mammography images from the training set.

[0082] In this embodiment, the number of iterations required for the training set is set to p, and the preset number of mammary gland mammography images acquired in the training set is set to q, where the mammary gland mammography images are tumor images, wherein p≥1, q≥1.

[0083] The second step is to obtain the trained RepCSC-Net model.

[0084] In this embodiment, a Softmax classifier is used to receive the features output by the fully connected layer, output the classification prediction results after the probability value mapping of the label corresponding to the input target, and calculate the loss function value of the network model for each iteration until the minimum loss function value is obtained to obtain the trained RepCSC-Net model. After obtaining the trained RepCSC-Net model, the last layer of the fully connected layer is used as the final extracted feature of the RepCSC-Net model.

[0085] Among them, the RepCSC-Net model with the RepCSC layer is trained, and the training data is recorded as Among them, x i is the i-th training data, y i is the i-th label data, N is the number of training data, and the RepCSC-Net model is trained by solving the following optimization problem. The calculation formula of the optimization problem can be:

[0086]

[0087] Among them, θ is the set of parameters that can be learned, l CEdenotes the cross entropy loss and f(;θ) denotes the mapping performed by the neural network.

[0088] It should be noted that the training process of the RepCSC-Net model is a prior art and is not within the scope of protection of the present invention, and will not be elaborated here.

[0089] Step S4, using the verification set, verifying the trained RepCSC-Net model to obtain a verified RepCSC-Net model.

[0090] In this embodiment, the trained RepCSC-Net model is verified using the verification set to adjust the model parameters. Specifically, the user can participate in parameter adjustment and test the verification set. If the user agrees with the tumor detection and recognition results, the parameter adjustment is stopped and the data is saved. If there is any objection, the user can continuously adjust the model parameters as needed until the recognition result of the model that satisfies the user is achieved, and the verified RepCSC-Net model is obtained.

[0091] Step S5, input the test set into the verified RepCSC-Net model, and output the recognition result of breast tumor.

[0092] In this embodiment, the test set is input into the verified RepCSC-Net model, and the RepCSC-Net model outputs the recognition result of the breast tumor, and the recognition result is displayed and saved. The recognition result of the breast tumor includes tumor classification information and classification accuracy.

[0093] The present invention provides a method for identifying benign and malignant breast tumors based on RepCSC-Net. The method first applies the structural re-parameterization strategy to convolutional sparse coding to construct a new RepCSC structure. Then, the RepCSC structure is combined with the convolutional neural network RepVGG to construct a RepCSC-Net network with strong feature representation learning ability, fast reasoning speed and good interpretability. Finally, the RepCSC-Net network is used for intelligent identification of benign and malignant breast tumors, which effectively enhances the accuracy of identifying benign and malignant breast tumors.

[0094] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for identifying benign and malignant breast tumors based on RepCSC-Net, characterized in that: The following steps are involved: Step S1, obtaining a breast mammography image dataset, performing image enhancement processing on the breast mammography image dataset, and dividing the dataset into a training set, a validation set, and a test set; Step S2, constructing a RepCSC module with a structure reparameterization strategy, combining the RepCSC module with a convolutional neural network RepVGG to obtain a RepCSC-Net model; Step S3, using the training set to train the RepCSC-Net model to obtain a trained RepCSC-Net model; Step S4, using the verification set, verifying the trained RepCSC-Net model to obtain a verified RepCSC-Net model; Step S5, inputting the test set into the verified RepCSC-Net model, and outputting the recognition result of the breast tumor; The content of step S2 includes: A RepCSC module of a structural reparameterization strategy is constructed, and an optimization target of the RepCSC module is determined; an iterative soft threshold algorithm is used to solve the optimization target, so that the RepCSC module during training is a three-channel structure; during training, the three-channel structure of the RepCSC module is converted into a single-channel structure by merging dictionaries to obtain a new dictionary; and the first layer of the convolutional neural network RepVGG is replaced by the designed RepCSC module to obtain a RepCSC-Net model; Among them, the RepCSC module includes structural reparameterization and convolutional sparse coding, and the RepCSC module is a convolutional sparse coding module with a structural reparameterization strategy; The input information of the RepCSC module is expressed as: Where x∈R L×W×C is the input signal, L×W is the spatial dimension of the input signal, C is the number of channels of the input signal, d k is the kth atom in the dictionary, s k is the sparse coding coefficient corresponding to the kth atom in the dictionary, and * is the convolution operation; The optimization objective is expressed as: Where D = (d1, d2, ..., d C ) is the dictionary matrix, d C is the Cth atom in the dictionary, Γ=(s1,s2,…,s C ) T is the set of sparse coding coefficients, s C is the sparse coding coefficient corresponding to the Cth atom in the dictionary, X = (x1, x2, ..., x N ) T is the input information set, x N is the Nth input information, min is the minimization function, λ is the sparsity parameter, is the reconstruction error, λ||Γ||1 is the sparsity constraint, and || || is the norm sign.

2. A method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 1, characterized in that: The optimization objective is solved by using an iterative soft threshold algorithm. The iterative formula is: Among them, Γ k+1 is the k+1th sparse coding coefficient set, S β / L is the soft threshold operator, β is the sparsity parameter, is the iteration step length, D T is the transpose of the dictionary matrix, X is the input information set, Γ k is the kth sparse coding coefficient set, and D is the dictionary matrix; By G 0 =X for the Γ k+1 Perform initialization processing to obtain a new iteration formula, which is: make C=-L, the new iterative formula is processed, and the new iterative formula after processing is: Γ 1 =S β / L (A T X+X+CA T AX).

3. The method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 1, characterized in that: The new dictionary is represented as follows: Among them, D ′ For the new dictionary, is the iteration step length, D T is the transpose of the dictionary matrix, I is the d×d identity matrix, and D is the dictionary matrix.

4. The method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 1, characterized in that: The content of step S1 includes: Import the mammography image dataset and then read the mammography image dataset; The image size of each image in the mammography dataset is adjusted, and the Lucy-Richardson algorithm is used to perform image enhancement on the mammography dataset. According to the image quantity ratio specified by the user, the breast mammography image dataset after image enhancement is divided into a training set, a validation set, and a test set, and saved as corresponding XML tag files.

5. The method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 1, characterized in that: The content of step S3 includes: Setting the number of iterations required for the training set and obtaining a preset number of mammographic images from the training set; The Softmax classifier is used to receive the features output by the fully connected layer, output the classification prediction results after the probability value mapping of the label corresponding to the input target, and calculate the loss function value of the network model for each round of iteration until the minimum loss function value is obtained to obtain the trained RepCSC-Net model.

6. The method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 5, characterized in that: After obtaining the trained RepCSC-Net model, the last layer of the fully connected layer is used as the final extracted feature of the RepCSC-Net model.

7. The method for identifying benign and malignant breast tumors based on RepCSC-Net according to claim 1, characterized in that: The breast tumor recognition result includes tumor classification information and classification accuracy.

Images