A breast image recognition and judgment method based on SFC-MSPCNN and improved ConvNeXt

Abstract

The present invention discloses a breast image recognition and determination method based on SFC-MSPCNN and improved ConvNeXt, and relates to the field of image analysis technology. Based on the comparative experimental analysis mode of breast image ultrasound database, MIAS database and DDSM database, according to the relationship between the characteristics of breast masses and calcification points and benign and malignant, the present invention proposes a pulse coupled neural network model based on a special ignition control mode and an improved next-generation convolutional neural network model to achieve benign and malignant classification of breast lesions, and uses comparative experimental methods to evaluate and analyze the classification effect.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of image analysis, and in particular to a breast image recognition and determination method based on SFC-MSPCNN and improved ConvNeXt. Background Art

[0002] The incidence and mortality of breast cancer are increasing year by year around the world. It is the leading disease threatening women's health. Therefore, early detection and treatment can effectively reduce the mortality of breast cancer patients. In breast images, due to the complexity of the tissue structure of breast images, the breast only occupies a part of the entire breast image. The irrelevant background area will interfere with the training results of the model, and the useless black background area in the image accounts for a large proportion. Many lesions grow in certain tissues or dense masses of the breast. The boundary between the lesion and the surrounding normal area is not clear, and the grayscale of the calcification lesion is very different from the surrounding tissue, which is easily submerged in high-frequency noise and low-frequency background. In addition, there is a high visual similarity between different categories of the same type of breast images, and conventional methods are difficult to accurately determine the benign and malignant nature of lumps and calcifications.

[0003] Based on the comparative experimental analysis mode of breast imaging ultrasound database, MIAS database and DDSM database, and according to the relationship between the characteristics of breast masses and calcification points and benign and malignant types, a pulse coupled neural network model based on a special ignition control mode (Special fire-controlled modified simplified pulse-coupled neural network, SFCMSPCNN) and an improved next-generation convolutional neural network model (Convolutional network net generation, ConvNeXt) are proposed to realize artificial intelligence (AI) classification of benign and malignant breast lesions, and a comparative experimental method is used to evaluate and analyze the classification effect. Summary of the invention

[0004] The purpose of the present invention is to solve the above-mentioned problems and to provide a breast image recognition and determination method based on SFC-MSPCNN and improved ConvNeXt.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] According to one aspect of the present invention, a method for identifying and determining breast images based on SFC-MSPCNN and improved ConvNeXt is provided.

[0007] Compared with the prior art, the present invention has the following beneficial effects:

[0008] (1) SFCMSPCNN model algorithm adaptively adjusts the internal activity item U ij [n] and dynamic threshold E ij The value of [n] adjusts the number of firing neurons and reduces the pixel value of the target pixel corresponding to the neuron, ultimately achieving accurate darkening of the pre-processed breast image. The darkened image highlights the characteristics of breast lesions more than the original image, providing a prerequisite for lesion information feature extraction and intelligent judgment of benign and malignant lesions.

[0009] (2) The primary feature extraction algorithm can extract and enhance image features, reduce network training parameters, reduce network computational complexity and error rate, and improve network accuracy and training speed; the advanced feature extraction algorithm has a unique inverse bottleneck residual structure that can expand the spatial dimension of the multi-layer perceptron, allowing the model to learn higher-level features as the network depth increases.

[0010] (3) The breast lesion classification algorithm helps to further reduce the number of model parameters, reduce the risk of overfitting, and quickly output the probability distribution of each category according to the category space, thus generating the final classification result more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of breast imaging lesion area classification of the present invention;

[0012] Figure 2 It is the CBAM attention mechanism module of the present invention;

[0013] Figure 3 It is the ConvNeXt-CBAM network model diagram of the present invention;

[0014] Figure 4 It is a mammary gland image learning rate attenuation effect diagram of the present invention;

[0015] Figure 5 It is a diagram showing changes in the loss value of breast image training of the present invention;

[0016] Figure 6 It is a schematic diagram of the prediction results of the breast ultrasound image three-classification experiment of the present invention;

[0017] Figure 7 It is a schematic diagram of the prediction results of the DDSM breast image three-classification experiment of the present invention;

[0018] Figure 8 It is a schematic diagram of the prediction results of the MIAS breast image five-classification experiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0020] The segmentation process of the technical solution of the present invention is as follows Figure 1 As shown:

[0021] exist Figure 1 In the figure, Stem represents the primary feature extraction layer, Stage represents the advanced feature extraction layer, and Dim represents the spatial dimension of the image block. The specific implementation steps of breast lesion classification are as follows:

[0022] Step 1: Preprocess the mammary gland image to darken it. The following algorithm is used:

[0023] ① According to the FC-MSPCNN algorithm model, the present invention proposes a SFCMSPCNN algorithm model, which is based on the internal activity item U ij [n] and dynamic threshold E ij The numerical change law of [n] dynamically adjusts the number of PCNN firing neurons and controls the pixel value size of the corresponding pixel of the neuron to achieve accurate darkening of the pre-processed breast image. The SFCMSPCNN algorithm model is as follows:

[0024] (8)

[0025] (9)

[0026] (10)

[0027] In formula (8)-formula (10), the neuron N at position (i, j) ij Contains two inputs: Feed input F ij and link input L ij , F ij Only external excitation S ij Effect, L ij is the link strength parameter β, the adjacent neuron output Y kl [n-1] and synaptic weight W ijkl The product and sum of internal activity terms U ij [n] consists of two terms, the first of which is the exponential decay factor e -α The internal activity item U at the previous moment of modulation ij [n-1], the second term is the link strength β modulated feed input F ij and link input L ij If the internal activity item U ij [n] is greater than the dynamic threshold E ij [n-1], then the neuron pulse output is Y ij[n] = 1, and the dynamic threshold increases with the increase of the amplitude parameter V, modulation parameter C, adjustment parameter Q and built-in parameter A. If the internal active item U ij [n] is less than the dynamic threshold E ij [n-1], the neuron has no impulse response output, and the dynamic threshold will decrease exponentially with the attenuation factor α.

[0028] ② Determine the key parameter expressions in the SFCMSPCNN algorithm, including the synaptic weight matrix W ijkl , link strength β, attenuation factor α, amplitude parameter V, modulation parameter C, adjustment parameter Q and built-in parameter A. Among them, the synaptic weight matrix W ijkl The calculation formula is as follows:

[0029] (1)

[0030] In formula (1), W ijkl A new weight matrix setting method is used to control the influence of neighboring neurons on central neurons. Because the new model has a lower algorithm calculation complexity, it can be achieved by reasonably setting the parameter W ijkl The value of , realizes the influence of the domain neuron on the central neuron. In addition, the link strength parameter β of SFCMSPCNN can be set to a mathematical form with stronger adjustment ability. The strength parameter β can adaptively adjust the interaction value between the central neuron and the surrounding neurons. Its mathematical expression is as follows:

[0031] (2)

[0032] In this model, the attenuation factor α adopts a new model parameter setting method, which ensures that the value range of the attenuation factor α is between 0 and 1. Its mathematical expression is as follows:

[0033] (3)

[0034] Among them, S max Represents the maximum pixel grayscale value of the entire image. Later, in order to ensure that the setting values ​​of the dynamic threshold and the internal activity item are more reasonable, the amplitude parameter V in the dynamic threshold is further simplified, and the expression of V is as follows:

[0035] (4)

[0036] The modulation parameter C can fine-tune the amplitude parameter V, further improving the numerical dynamic modulation capability of the amplitude parameter V. The expression of the modulation parameter C is as follows:

[0037] (5)

[0038] The main function of adjusting the parameter Q and the built-in parameter A is to correct the dynamic threshold. Through the joint action of the two parameters, the changing trend of the dynamic threshold can be accurately controlled. The expressions of adjusting the parameter Q and the built-in parameter A are as follows:

[0039] (6)

[0040] (7)

[0041] Adjusting the parameter Q and the built-in parameter A increases the complexity of the numerical changes of the SFCMSPCNN model.

[0042] Step 2: Primary feature extraction (performed at the Stem layer). The following algorithm is used:

[0043] Primary feature extraction is performed in the Stem layer, which is the input layer of the network and consists of a 3×3 convolutional layer with 64 channels, a normalization layer, an activation function, and a convolutional block attention module (CBAM). Its main function is to extract and enhance image features and suppress unnecessary features while maintaining the input image size. This design helps reduce the number of network parameters, reduce computational complexity and error rate, and improve network accuracy and training speed.

[0044] CBAM consists of input, channel attention module, spatial attention module and output. The implementation process diagram is shown in Figure 2 shown.

[0045] exist Figure 2 In the example, the input features are , where C represents the number of channels of the input image, H represents the length of the input image, and W represents the height of the input image. Then, a one-dimensional convolution is performed on the channel attention module. , multiply the convolution result F' by the original image and use the output of the channel attention as input to perform a two-dimensional convolution of the spatial attention module , get the output result F'', and then multiply it with the original image to get the accurate feature map. The implementation process of attention is shown in the following formula:

[0046] (11)

[0047] (12)

[0048] Step 3: Advanced feature extraction (performed at the Stage layer). The following algorithm is used:

[0049] The advanced feature extraction module consists of 4 Stage layers, each of which contains a ConvNeXt Block module. The schematic diagram of the ConvNeXt Block module is as follows: Figure 3 As shown. Figure 3 In the figure, Depthwise represents depthwise separable convolution, Gelu is the type of model activation function, Layer scale represents layer scale, and Drop path is the regularized pruning method used by the model.

[0050] The types and number of these modules can be adjusted according to the actual task needs, such as changing the block ratio and changing the size of the convolution kernel and the number of channels. Between different stages, a separate downsampling layer is used to reduce the size of the feature map and improve the feature abstraction level. In the training process, a deep separable convolution is introduced to improve the calculation speed of the model, and fewer normalization layers are used. The batch normalization (BN) layer is replaced by the Lay normalization (LN) layer to solve the problem of excessive deviation of the normalized statistics when the sample size is too small. The activation function uses GELU and reduces the number of activation functions. In addition, ConvNeXt has the inherent ability to inductively bias, and the unique inverse bottleneck residual structure can broaden the spatial dimension of the multi-layer perceptron, which helps to avoid information loss and complete different visual tasks. Through this design, the model can learn higher and higher-level features as the network depth increases.

[0051] Step 4: Classification of breast lesions. The following algorithm is used:

[0052] The end of the network is a classification module (the module structure diagram is as follows Figure 3 As shown in Figure 3, it includes three parts: global average pooling, fully connected layer, and activation function, which are used to generate the final classification results. Global average pooling averages the feature map in the spatial dimension to obtain a feature vector of fixed length, which helps to reduce the number of model parameters and reduce the risk of overfitting; the fully connected layer maps the feature vector obtained by global average pooling to the category space and outputs the probability distribution of each category; and the activation function converts the output of the fully connected layer into a probability distribution for easy interpretation and comparison, generating the final classification result.

[0053] (1) Experimental environment, evaluation indicators, and comparison algorithms

[0054] In order to comprehensively measure the effectiveness of the model of the present invention in the image classification task, the experiment of the present invention uses accuracy, specificity, precision and confusion matrix as evaluation indicators to quantitatively evaluate the classification performance of the model. Among them, TP (True Positive) represents the number of samples correctly predicted as positive examples by the model, TN (True Negative) represents the number of samples correctly predicted as negative examples by the model, FP (False Positive) represents the number of samples incorrectly predicted as positive examples by the model, and FN (False Negative) represents the number of samples incorrectly predicted as negative examples by the model.

[0055] The ConvNeXt-CBAM method (ConvNeXt model + CBAM convolutional attention module) proposed in this paper is compared with four classic and popular deep learning classification network models to verify the effectiveness of the method in determining the benign and malignant nature of breast imaging lesions. The four comparison models include the ConvNeXt model, the EfficientNet model, and the 、MobileViT model and SwinTransformer model .

[0056] Table 1 Experimental environment

[0057]

[0058] (2) Experimental results and analysis

[0059] The initial value of the learning rate of the experimental model of the present invention is 5e -4, the batch size is 8. In order to keep the model stable during training, the Warmup learning rate is selected to warm up the model so that the model starts training with a smaller learning rate. After the model becomes stable, in order to ensure that the entire network can converge quickly and stably during the training process, as the training iteration proceeds, the learning rate decays in the form of a cosine function value to accelerate the convergence speed of the model. The learning rate decreases linearly with the increase in the number of training batches. A total of 200 cycles, i.e. Epoch=200, are trained. After multiple iterations, it decays to 0 at about 188 rounds, and the training is terminated when the value no longer changes. During training and verification, the images must undergo preprocessing operations to extract the region of interest and convert it into image data of size 224×224 to improve network performance and generalization ability. Mixup and Cutmix data enhancement techniques are used in the training process to improve the robustness of the model, and a label smoothing regularization strategy is introduced to suppress overfitting during the training process. Since the grayscale image is a single-channel image, it is converted into a three-channel image suitable for network model input during training. The learning rate decay of the model of the present invention is as follows Figure 4 In the field of medical image processing, ImageNet pre-trained weights are usually used to train convolutional neural networks. Therefore, the experimental model of the present invention uses pre-trained weight parameters to reduce the training loss of the model and reach the minimum value quickly, thereby obtaining excellent performance. The training loss function curve of the model is shown in Figure 5 shown.

[0060] ①Experimental results and analysis of breast ultrasound image database

[0061] The breast ultrasound image database contains 710 breast image data. The experiment of the present invention expands 133 normal images into 200 image data through data enhancement, randomly selects 200 image data from 437 benign lesion images, and randomly selects 200 image data from 210 malignant lesion images to form a breast ultrasound data set of 600 image data, and randomly divides the data set according to a ratio of 4:1, that is, 480 breast images in the training set and 120 breast images in the verification set for experimental verification of the present invention. In order to effectively verify the classification and judgment ability of the improved model for breast lesions, the experiment of the present invention adopts a comparative experimental method for experimental verification. The objective experimental evaluation results are shown in Table 2.

[0062] Table 2 Breast image classification comparison experimental results (ultrasound image database)

[0063]

[0064] As shown in Table 2, the classification accuracy, specificity and precision of the ConvNeXt-CBAM method of the present invention are 95%, 97.5% and 95.2%, respectively. The classification performance is obviously superior, meeting the requirements for ultrasound imaging breast lesion classification, with a low misdiagnosis rate, and able to find the location of specific lesions, which plays an auxiliary guiding role in the diagnosis of clinical breast diseases. However, breast ultrasound images have a high similarity and there is also a certain classification error.

[0065] ② Experimental results and analysis of DDSM image database

[0066] The DDSM image database contains 2620 breast image data. The experiment of the present invention randomly selects 200 image data from 695 normal images, 200 image data from 1011 benign lesion images, and 200 image data from 914 malignant lesion images to form a DDSM breast image data set of 600 image data, and randomly divides the data set according to a ratio of 4:1, that is, 480 breast images in the training set and 120 breast images in the validation set for the experimental verification of the present invention. The objective experimental evaluation results are shown in Table 3.

[0067] Table 3 Breast image classification comparison experimental results (DDSM image database)

[0068]

[0069] As shown in Table 3, the classification accuracy of the ConvNeXt-CBAM of the present invention for the classification of DDSM breast lesions can reach 87.5%, the specificity of 93.8%, and the precision of 87.6%. Overall, the objective evaluation index has an advantage over other network models, and the classification performance and performance are relatively ideal.

[0070] ③Experimental results and analysis of MIAS image database

[0071] The MIAS breast image database contains 322 breast image data. The experiment of the present invention randomly selects 120 image data from 209 normal images, enhances and expands 10 benign calcification image data to 120, expands 52 benign mass image data to 120, expands 12 malignant calcification image data to 120, and expands 39 malignant mass image data to 120, which together form a 600 experimental data set, and randomly divides the data set according to a ratio of 4:1, that is, 480 breast images are used as a training set and 120 breast images are used as a validation set to verify the experiment of the present invention. The objective experimental evaluation results are shown in Table 4.

[0072] Table 4 Breast image classification comparison experimental results (MIAS image database)

[0073]

[0074] As can be seen from Table 4, the ConvNeXt-CBAM of the present invention can obtain 86.7% classification accuracy, 96.7% specificity and 88.4% precision, which are significantly higher than the other three classification network models with different structures. In the five-classification experiment of lesions in the MIAS image dataset, it has a great advantage in correctly identifying and classifying five types of breast images, and the network performance is relatively better.

[0075] The technology of the present invention can provide doctors with valuable information and evaluation criteria in the early detection and screening stage of breast diseases, greatly reducing the workload of doctors in diagnosis and treatment. The method of the present invention is trained and tested in the breast ultrasound database, MIAS database and DDSM database. The three databases include a total of 3,600 experimental images (2,880 training images and 720 test images). The effectiveness and rationality of the model are verified and evaluated in combination with evaluation indicators and prediction data. The experimental data show that the method of the present invention has a high accuracy rate and low computational complexity in determining the nature of breast lesions. It has certain application value for the early diagnosis and auxiliary diagnosis and treatment of different types of breast cancer lesions, and can assist doctors in the analysis and treatment of breast cancer lesions.

[0076] (1) Randomly select different categories of image data from the breast ultrasound image database and use the trained model to predict the type. The predicted classification results are shown in the figure below. Figure 6 As shown. Compared with other medical images, breast ultrasound images have lower contrast, blurred edges and contain noise. Therefore, there will be certain errors in the prediction of breast image categories. It is necessary to further study the benign and malignant judgment experiments of breast lesions in combination with the characteristics of breast ultrasound images. Among them, class 0 in sub-image (a) represents the image that is correctly predicted as a normal image, prob represents the probability of correct prediction is 0.997, and the probability of being incorrectly predicted as other categories is 0.003. Class 1 in sub-image (b) represents the image that is correctly predicted as a benign lesion, the correct probability is 0.995, and the probability of being incorrectly predicted as other categories is 0.005. Class 2 in sub-image (c) represents the image that is correctly predicted as a malignant lesion, and the correct probability is 1.0.

[0077] (2) Randomly select different categories of image data from the DDSM image database and use the trained model to predict the category. The predicted classification results are shown in the figure below. Figure 7As shown in the figure, most images can be correctly classified, and the probability of being correctly classified is relatively high, but some images are also incorrectly predicted as other categories. This is due to the image properties of breast images and breast lesions, which is within an acceptable range. Among them, class 0 in sub-image (a) represents the image that is correctly predicted as a normal image, prob represents the probability of correct prediction is 1.0, class 1 in sub-image (b) represents the image that is correctly predicted as a benign lesion, and the correct probability is 1.0, and class 2 in sub-image (c) represents the image that is correctly predicted as a malignant lesion, and the correct probability is 0.782, and the probability of being incorrectly predicted as other categories is 0.218.

[0078] (3) We randomly selected different categories of image data from the MIAS breast image database and used the trained model to make predictions. Most images were correctly predicted to belong to the same category. The prediction classification results are shown in the figure below. Figure 8 As shown. In sub-image (a), class0 represents the image correctly predicted as a normal image, prob represents the probability of correct prediction 1.0, sub-image (b) class 1 represents the image correctly predicted as a benign calcification lesion, sub-image (c) class 2 represents the image correctly predicted as a benign mass lesion, sub-image (d) class 3 represents the image correctly predicted as a malignant calcification lesion, sub-image (e) class 4 represents the image correctly predicted as a malignant mass lesion, and the correct prediction probabilities are all 1.0.

[0079] This experiment uses three image datasets, including:

[0080] (1) The breast ultrasound image database contains 710 breast image data. The experiment of the present invention expands 133 normal images into 200 image data through data enhancement, randomly selects 200 image data from 437 benign lesion images, and randomly selects 200 image data from 210 malignant lesion images to form a breast ultrasound data set of 600 image data. The data set is randomly divided into 480 breast images as a training set and 120 breast images as a validation set according to a ratio of 4:1 for experimental verification of the present invention.

[0081] (2) The DDSM image database contains 2620 breast image data. The experiment of the present invention randomly selected 200 image data from 695 normal images, randomly selected 200 image data from 1011 benign lesion images, and randomly selected 200 image data from 914 malignant lesion images to form a DDSM breast image data set of 600 image data. The data set is randomly divided into 480 breast images for training and 120 breast images for validation for the experiment of the present invention.

[0082] (3) The MIAS breast image database contains 322 breast image data. The experiment of the present invention randomly selects 120 image data from 209 normal images, enhances and expands 10 benign calcification image data to 120 image data, expands 52 benign mass image data to 120 image data, expands 12 malignant calcification image data to 120 image data, and expands 39 malignant mass image data to 120 image data, which together constitute 600 experimental data sets, and randomly divide the data sets according to the ratio of 4:1, that is, 480 breast images are used as the training set and 120 breast images are used as the validation set for experimental verification of the present invention.

[0083] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0084] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A breast image recognition and determination method based on SFC-MSPCNN and improved ConvNeXt, including a primary feature extraction layer and an advanced feature extraction layer, characterized in that: The specific steps are as follows: Step 1: Preprocessing mammary image darkening ① According to the FC-MSPCNN algorithm model, a SFCMSPCNN algorithm model is proposed. The model is based on the internal activity item U ij [n] and dynamic threshold E ij The numerical change law of [n] dynamically adjusts the number of PCNN firing neurons and controls the pixel value size of the corresponding pixel of the neuron to achieve accurate darkening of the pre-processed breast image. The SFCMSPCNN algorithm model is as follows: (8) (9) (10) In formula (8)-formula (10), the neuron N at position (i, j) ij Contains two inputs: Feed input F ij and link input L ij , F ij Only external excitation S ij Effect, L ij is the link strength parameter β, the adjacent neuron output Y kl [n-1] and synaptic weight W ijkl The product and sum of Internal Activity ItemU ij [n] consists of two terms, the first of which is the exponential decay factor e -α The internal activity item U at the previous moment of modulation ij [n-1], the second term is the link strength β modulated feed input F ij and link input L ij ; If the internal activity item U ij [n] is greater than the dynamic threshold E ij [n-1], then the neuron pulse output is Y ij [n]=1, and the dynamic threshold increases with the increase of the amplitude parameter V, modulation parameter C, adjustment parameter Q and built-in parameter A; If the internal activity item U ij [n] is less than the dynamic threshold E ij [n-1], the neuron has no impulse response output, and the dynamic threshold will decrease exponentially with the attenuation factor α; ② Determine the key parameter expressions in the SFCMSPCNN algorithm, including the synaptic weight matrix W ijkl , link strength β, attenuation factor α, amplitude parameter V, modulation parameter C, adjustment parameter Q and built-in parameter A, where the synaptic weight matrix W ijkl The calculation formula is as follows: (1) In formula (1), W ijkl A new weight matrix setting method is adopted to control the influence of neighborhood neurons on central neurons. The link strength parameter β of SFCMSPCNN can be set to a mathematical form with stronger adjustment ability. The strength parameter β can adaptively adjust the interaction value between the central neuron and the surrounding neurons. Its mathematical expression is as follows: (2) In this model, the attenuation factor α adopts a new model parameter setting method, which ensures that the value range of the attenuation factor α is between 0 and 1. Its mathematical expression is as follows: (3) Among them, S max Represents the maximum pixel grayscale value of the entire image; the amplitude parameter V in the dynamic threshold is further simplified, and the expression of V is as follows: (4) The modulation parameter C can fine-tune the amplitude parameter V, further improving the numerical dynamic modulation capability of the amplitude parameter V. The expression of the modulation parameter C is as follows: (5) The main function of adjusting the parameter Q and the built-in parameter A is to correct the dynamic threshold. Through the joint action of the two parameters, the changing trend of the dynamic threshold can be accurately controlled. The expressions of adjusting the parameter Q and the built-in parameter A are as follows: (6) (7) Adjusting the parameter Q and the built-in parameter A increases the complexity of the numerical changes of the SFCMSPCNN model; Step 2: Primary feature extraction Primary feature extraction is performed in the Stem layer, which is the input layer of the network and consists of a 3×3 convolutional layer with 64 channels, a normalization layer, an activation function, and a convolutional attention module CBAM; CBAM consists of input, channel attention module, spatial attention module and output. The input feature is , where C represents the number of channels of the input image, H represents the length of the input image, and W represents the height of the input image. Then, a one-dimensional convolution of the channel attention module is performed. , multiply the convolution result F' by the original image and use the output of the channel attention as input to perform a two-dimensional convolution of the spatial attention module , get the output result F'', and then multiply it with the original image to get the accurate feature map; the implementation process of attention is shown in the following formula: (11) (12) Step 3: Advanced Feature Extraction The high-level feature extraction module consists of 4 stages. Between different stages, a separate downsampling layer is used to reduce the feature map size and improve the feature abstraction level. Each Stage contains a ConvNeXt Block module. The ConvNeXt module has an inherent ability to inductively bias. The unique inverse bottleneck residual structure can broaden the spatial dimension of the multi-layer perceptron, which helps to avoid information loss and complete different visual tasks. Through this design, the model can learn higher and higher level features as the network depth increases. Each ConvNeXt Block module includes a depthwise separable convolution module, a normalization layer, a Gelu activation function, a layer scale module, and a Drop path regularization pruning module. In the training process, the depthwise separable convolution module is first introduced to improve the calculation speed of the model. Then, fewer normalization layers are used, and the batch normalization layer is replaced by the Lay normalization layer to solve the problem of excessive deviation of normalized statistics when the sample size is too small. Next, the activation function GELU is selected. This type of activation function is smoother and helps to improve the convergence speed of the function during the training process. Finally, the layer scale module and the Drop path regularization pruning module are designed to reduce the size and calculation amount of the model, thereby achieving model compression and acceleration. Step 4: Classification of breast lesions At the end of the network is a classification module, which consists of three parts: global average pooling, fully connected layer, and activation function, which are used to generate the final classification results; Global average pooling averages the feature map in the spatial dimension to obtain a feature vector of fixed length, which helps to reduce the number of model parameters and reduce the risk of overfitting; The fully connected layer maps the feature vector obtained by global average pooling to the category space and outputs the probability distribution of each category; The activation function converts the output of the fully connected layer into a probability distribution for easy interpretation and comparison, generating the final classification result.

Images