Ultrasonic breast cancer focus self-attention identification method based on multi-dimensional deep convolution
By constructing a self-attention recognition model for ultrasound breast cancer lesions using multi-dimensional deep convolution and adaptive dimensional transformation attention mechanisms, the shortcomings of existing breast cancer lesion recognition methods in feature information capture and noise removal are addressed, achieving higher accuracy in lesion segmentation.
Patent Information
- Application Number
- CN202511717550.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-21
- Publication Date
- 2026-03-03
AI Technical Summary
Existing methods for identifying breast cancer lesions struggle to fully capture complex features in ultrasound images and lack multi-dimensional information fusion, resulting in inaccurate identification results and poor removal of noise and interference.
A self-attention recognition model for ultrasound breast cancer lesions based on multi-dimensional deep convolution was constructed. The original fusion features were extracted through multi-dimensional deep convolutional layers, and local texture features were processed by an adaptive dimension transformation attention mechanism module. Finally, a global feature map was generated to achieve lesion segmentation.
It significantly improved the segmentation accuracy and boundary preservation ability of breast cancer lesions, and enhanced the average intersection-union ratio (mIOU) of the model on different datasets, demonstrating its feasibility and effectiveness in practical applications.
Smart Images
Figure CN121599995A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing technology, specifically relating to a self-attention recognition method for ultrasound breast cancer lesions based on multi-dimensional depth convolution. Background Technology
[0002] Breast cancer, as one of the most common and serious malignant tumors threatening women's health, requires early diagnosis for improving patient survival and prognosis. Currently, ultrasound imaging technology, with its advantages of being non-invasive, convenient, and repeatable, has become an important tool for breast cancer screening and diagnosis. However, the characteristics of breast cancer lesions in ultrasound images are complex and diverse, and are often affected by noise, artifacts, and interference from surrounding tissues, making accurate identification and segmentation of lesions quite challenging.
[0003] Existing methods for identifying breast cancer lesions have some limitations. Traditional methods based on handcrafted features rely on manually designed features, making it difficult to comprehensively capture the complex features of the lesions, and their generalization ability is poor across different individuals and imaging conditions. While some deep learning-based methods have improved the accuracy of identification to some extent, they are still insufficient in mining the subtle textures and local features of lesions in ultrasound images, resulting in less precise identification results.
[0004] Meanwhile, existing network models often lack effective fusion of multi-dimensional information when processing ultrasound images, failing to fully utilize the features of the images in different dimensions, thus affecting the accurate identification of lesions. Furthermore, the removal of noise and interference information from the images during processing needs further improvement to better highlight lesion features. Therefore, developing a more effective ultrasound method for identifying breast cancer lesions has significant clinical importance and application value. Summary of the Invention
[0005] To address the problems existing in the prior art, this invention proposes a self-attention recognition method for ultrasound-guided breast cancer lesions based on multi-dimensional depth convolution. This method includes:
[0006] Constructing an ultrasound-guided self-attention model for breast cancer lesion recognition;
[0007] Acquire ultrasound images of breast cancer and preprocess them to remove noise and interference.
[0008] The preprocessed image is input into the encoder of the improved ultrasound breast cancer lesion self-attention recognition model. The original fusion features are extracted from different dimensions through multi-dimensional deep convolutional layers in the encoder, and the fusion features are obtained after basic processing. ;
[0009] Fusion features Local texture features are obtained through an adaptive dimensionality transformation attention mechanism module. ;
[0010] The upsampled fused features are concatenated with the local texture features to obtain the global feature map. The final lesion segmentation result is extracted after post-processing.
[0011] The beneficial effects of this invention are:
[0012] The multi-dimensional deep convolutional network framework proposed in this invention, which integrates multi-dimensional deep convolution and adaptive dimensionality transformation attention mechanism, demonstrates superior performance in breast ultrasound image segmentation tasks. It achieves the highest mean intersection-over-union (mIOU) on both datasets A and B, and also shows significant advantages in segmentation accuracy and boundary preservation, proving its feasibility and effectiveness in practical applications. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the multi-dimensional depth convolution processing implementation process of the present invention;
[0014] Figure 2 This is a schematic diagram of the implementation process of the adaptive dimension transformation attention mechanism of the present invention;
[0015] Figure 3 This is a schematic diagram of the ultrasound breast cancer lesion self-attention recognition network based on multi-dimensional deep convolution of the present invention.
[0016] Figure 4 This is a table comparing the effects of the present invention;
[0017] Figure 5 This is a flowchart of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] A self-attention recognition method for ultrasound breast cancer lesions based on multi-dimensional deep convolution is proposed. The method includes: constructing an ultrasound breast cancer lesion self-attention recognition model; acquiring ultrasound breast cancer images and preprocessing them to remove noise and interference; inputting the preprocessed images into the encoder of the improved ultrasound breast cancer lesion self-attention recognition model, extracting original fusion features from different dimensions through multi-dimensional deep convolutional layers in the encoder, and obtaining the fusion features after basic processing. ; to integrate features Local texture features are obtained through an adaptive dimensionality transformation attention mechanism module. The upsampled fused features are concatenated with local texture features to obtain the global feature map. The final lesion segmentation result is extracted after post-processing.
[0020] In this embodiment, a self-attention recognition method for ultrasound breast cancer lesions based on multi-dimensional deep convolution includes: acquiring ultrasound breast images and preprocessing the ultrasound breast images; inputting the preprocessed images into a trained ultrasound breast cancer lesion self-attention recognition network based on multi-dimensional deep convolution to obtain a recognition result; the ultrasound breast cancer lesion self-attention recognition network based on multi-dimensional deep convolution adds a multi-dimensional deep convolution layer to the encoder, and the output of the encoder is connected to the corresponding decoder via an adaptive dimensionality transformation attention mechanism module. Training the ultrasound breast cancer lesion self-attention recognition network based on multi-dimensional deep convolution includes: acquiring an ultrasound breast image dataset and preprocessing the images in the dataset; inputting the preprocessed images into the encoder to obtain a fused feature map. ; Fuse feature maps The input is fed into the adaptive dimensionality transformation attention mechanism module to obtain local texture features. The upsampled fused features are concatenated with local texture features to obtain the global feature map. The final lesion segmentation result is extracted after post-processing, and the loss function of the model is calculated based on the segmentation result. The model parameters are adjusted, and the training of the model is completed when the loss function converges.
[0021] In this embodiment, the encoder includes multi-dimensional deep convolutional layers, 2dConv layers, normalization layers, and ReLU activation layers; as shown Figure 1 As shown, the multi-dimensional depth convolutional layer contains a multi-dimensional depth convolution module. The multi-dimensional depth convolutional layer processes the preprocessed image. The processing includes dividing the preprocessed image into four feature sub-images according to the channel dimension. , The original features are obtained through identity mapping. , Three feature sub-maps were obtained by performing depthwise convolutions in three different dimensions. and with The images are stitched together to obtain a fused feature map. The fused feature map is then normalized, and the normalized feature map is processed by a multilayer perceptron. The perceived feature map is then fused with the input image to obtain a fused feature map.
[0022] Specifically, the different-dimensional depth convolution processing proposed in this invention employs three convolution kernels with different receptive fields, including a square convolution kernel for extracting spatial dimension features, a strip convolution kernel for extracting horizontal dimension features, and a strip convolution kernel for extracting vertical dimension features.
[0023] Preprocessed image The input tensor is proportionally split along the channel dimension, keeping the batch, height, and width dimensions unchanged, and only the number of channels is adjusted to divide it into four feature sub-maps. The implementation process is as follows:
[0024]
[0025] Among them, feature subgraph As an identity mapping branch, it is passed directly to the output without any convolution operation. Its purpose is to effectively preserve the original information and promote the proper functioning of gradients in deep networks. That is:
[0026]
[0027] The remaining three smaller feature submaps with the same number of channels are then input into three depthwise convolution branches with different kernel sizes for feature transformation. By using a standard square convolution kernel, local, isotropic features in space are extracted. That is:
[0028]
[0029] Feature subgraph Vertical texture and edge features are extracted by using long, strip-shaped convolutional kernels in the vertical direction. That is:
[0030]
[0031] Feature subgraph Horizontal texture and edge features are extracted through horizontally oriented elongated convolutional kernels. That is:
[0032]
[0033] Ultimately, In terms of channel dimension The resulting fused feature sub-map is obtained by stitching together the features. The entire multi-dimensional depthwise convolution operation is completed through normalization and multilayer perceptron processing. That is:
[0034]
[0035]
[0036] Where Split represents the partitioning operation. Indicates the kernel size as Depth convolution operation, This indicates a splicing operation. Represents the original fused feature map. This indicates normalization processing. This indicates multilayer perceptron processing. This represents a multi-dimensional depthwise convolution operation.
[0037] Obtain the original fused feature map Subsequently, basic processing is performed through a 2dConv layer, a normalization layer, and a ReLU activation layer to obtain the fused features. .
[0038] In this embodiment, the DTrans dimensional transformation is implemented based on the spatial structure of the input features. First, the two-dimensional feature vector is transformed... It is reshaped into a four-dimensional spatial form, and the channel dimension is adjusted to the front end, that is... Where B represents the batch size, i.e., the number of data samples processed at one time, and C represents the number of feature channels, which is the feature dimension contained in each spatial location of the original feature vector. Subsequently, adaptive average pooling is used to... Replace to obtain , The output size, determined by adaptive average pooling, exhibits dynamic variation. Finally, it is flattened to obtain... This completes the dimensionality transformation of the feature vectors. That is:
[0039]
[0040] in, Indicates the flattening operation. This indicates an adaptive average pooling operation.
[0041] like Figure 2 As shown, the adaptive dimensionality transformation attention mechanism processes the input fused feature map by: based on the fused feature map... The K, Q, and V matrices are obtained. These matrices are then partitioned dimensionally according to the number of attention heads using a linear layer, thus separating the K, Q, and V vectors. The Q vector is assigned to a pooling layer and processed by DTrans to obtain the A vector.
[0042]
[0043] in, For query vector, This indicates a dimensional change operation. This indicates a pooling operation.
[0044] Simultaneously, the K and V vectors are assigned to the first softmax layer and then... Processing yields V A vector:
[0045]
[0046] Subsequently, the original Q vector, A vector, and V... A The vector is input to the second softmax layer, and the original V vector is simultaneously subjected to depthwise convolution. The outputs of the two layers are concatenated to generate a local feature map.
[0047]
[0048] in, For adaptive dimensionality transformation attention mechanism processing, This is handled using the softmax attention mechanism. This indicates depthwise convolution processing.
[0049] In this embodiment, as Figure 3 As shown, the baseline U-net network is improved by adding a multi-dimensional deep convolutional layer to the encoder, and the output of the encoder is connected to the corresponding decoder through an adaptive dimension transformation attention mechanism module.
[0050] The improved U-net network processes the pre-processed image in the following ways:
[0051]
[0052]
[0053]
[0054] Where X represents the preprocessed image. This represents the i-th fused feature map, output by the i-th encoder; This represents multi-dimensional depthwise convolution processing; This indicates ordinary convolution processing; This indicates normalization processing; Indicates activation function processing; This represents the i-th local feature map, output by the i-th adaptive dimension transformation attention mechanism module; This is handled by an adaptive dimensionality transformation attention mechanism. Represents the i-th global feature map, by After sampling and obtained by piecing together; This indicates upsampling processing.
[0055] In this embodiment, a self-attention network for ultrasound breast cancer lesion recognition based on multi-dimensional deep convolution is used, such as... Figure 3 As shown, in the encoding stage, the network gradually extracts high-level features of the image through a series of downsampling operations and embeds multi-dimensional deep convolutional layers into the encoder part. These multi-dimensional deep convolutional layers, with convolutional kernels of three different receptive fields, effectively capture local features and global contextual information in the image, significantly enhancing the model's ability to recognize complex image structures. In each layer of the encoder, the input feature map is first processed by the multi-dimensional deep convolutional layer, then by a 2dConv layer, a normalization layer, and a ReLU activation layer to obtain a fused feature map. In the decoding stage, the network gradually restores the image resolution through upsampling operations, while combining low-level features from the encoder to finely depict the boundaries and textures of the target. Simultaneously, the fused feature map is input to the adaptive dimensionality transformation attention mechanism module. Through dimensionality transformation of the feature vectors and processing by the softmax attention mechanism, it adaptively focuses on important feature regions, achieving effective fusion and propagation of global information, further enhancing the model's ability to segment lesion regions. In the post-processing stage, the feature map output by the decoder is further processed. Morphological transformation and feature fusion techniques are used to optimize the boundary and shape of the lesion region, generating a binarized lesion segmentation mask, which is then mapped to the original ultrasound image space to achieve precise localization and contour delineation of the lesion region. Experimental results show that this method performs excellently in terms of segmentation accuracy and adaptability to lesion regions with different shapes and textures, providing strong support for the early diagnosis and treatment of ultrasound-guided breast cancer.
[0056] In this embodiment, as Figure 5 As shown, the data processing of the ultrasound breast cancer lesion self-attention recognition network includes: preprocessing the ultrasound breast cancer image to remove noise and interference information; inputting the preprocessed image into the encoder, and extracting fusion features from different dimensions through multi-dimensional deep convolutional layers; and processing the extracted fusion features. Local texture features are obtained through an adaptive dimensionality transformation attention mechanism. ; local texture features and the fusion features after upsampling The global feature map is obtained by splicing. The final lesion segmentation result is extracted after post-processing.
[0057] This invention uses a public breast ultrasound image dataset and a private breast ultrasound image dataset to evaluate the performance of the proposed image segmentation network. Specifically, dataset A (BUSI) consists of 780 breast ultrasound images acquired using two different ultrasound devices. Dataset B contains 156 breast ultrasound images and is a private dataset.
[0058] This invention provides a comprehensive experimental evaluation of the proposed self-attention network for ultrasound breast cancer lesion recognition based on multi-dimensional deep convolution and adaptive dimensionality transformation attention mechanism. First, to verify the model's effectiveness, various comparative experiments were conducted. The proposed model was compared with traditional image segmentation models based on convolutional neural networks, network models without deep convolution decomposition strategies, and models without adaptive dimensionality transformation attention mechanisms, such as U-Net, Attention U-Net, ResUNet, and TransUNet. Training and testing were performed on datasets A and B, respectively, and key metrics such as segmentation accuracy, recall, and F1 score were recorded for each model. Experimental results are as follows: Figure 4 As shown, the results demonstrate that, compared to these classic models, the U-net network framework proposed in this study, which integrates multi-dimensional deep convolution and surrogate attention mechanisms, exhibits superior performance in breast ultrasound image segmentation tasks.
[0059] For model optimization, this invention employs the AdamW optimization algorithm with an initial learning rate of 0.001, a weight decay parameter of 1e−4, and applies a cosine annealing strategy to dynamically adjust the learning rate. The batch size is fixed at 8. A weighted cross-entropy loss function is used during training. All models undergo 200 training epochs without early termination. To prevent underfitting or overfitting, model parameters are saved after each training epoch, provided the validation set performance (mIOU) is improved. All models used in the experiments are based on the officially released code.
[0060] In summary, the self-attention recognition method for ultrasound breast cancer lesions based on multi-dimensional deep convolution proposed in this invention effectively improves the model's performance and segmentation accuracy through the innovative application of multi-dimensional deep convolution strategy and adaptive dimensional transformation attention mechanism. It provides a new and effective method for the accurate identification and segmentation of ultrasound breast cancer lesions and has broad clinical application prospects.
[0061] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A self-attention recognition method for ultrasound breast cancer lesions based on multi-dimensional depth convolution, characterized in that, The method includes: Constructing an ultrasound-guided self-attention model for breast cancer lesion recognition; Acquire ultrasound images of breast cancer and preprocess them to remove noise and interference. The preprocessed image is input into the encoder of the improved ultrasound breast cancer lesion self-attention recognition model. The original fusion features are extracted from different dimensions through multi-dimensional deep convolutional layers in the encoder, and the fusion features are obtained after basic processing. ; Fusion features Local texture features are obtained through an adaptive dimensionality transformation attention mechanism module. ; The upsampled fused features are concatenated with the local texture features to obtain the global feature map. The final lesion segmentation result is extracted after post-processing.
2. The method for self-attention recognition of ultrasound breast cancer lesions based on multi-dimensional depth convolution according to claim 1, characterized in that, The ultrasound breast cancer lesion self-attention recognition model is an improved U-net network, which adds a multi-dimensional deep convolutional layer to the encoder, and the output of the encoder is connected to the corresponding decoder through an adaptive dimension transformation attention mechanism module.
3. The method for self-attention recognition of ultrasound breast cancer lesions based on multi-dimensional depth convolution according to claim 2, characterized in that, The improved U-net network processes the pre-processed image in the following ways: ； ； ； Where X represents the preprocessed image. This represents the i-th fused feature map, output by the i-th encoder; This represents multi-dimensional depthwise convolution processing; This indicates ordinary convolution processing; This indicates normalization processing; Indicates activation function processing; This represents the i-th local feature map, output by the i-th adaptive dimension transformation attention mechanism; This is handled by an adaptive dimensionality transformation attention mechanism. Represents the i-th global feature map, by After sampling and obtained by piecing together; This indicates upsampling processing.
4. The method for self-attention recognition of ultrasound breast cancer lesions based on multi-dimensional depth convolution according to claim 2, characterized in that, The encoder includes multi-dimensional deep convolutional layers, 2dConv layers, normalization layers, and ReLU activation layers; the multi-dimensional deep convolutional layers include multi-dimensional deep convolutional modules.
5. The ultrasound-based self-attention recognition method for breast cancer lesions based on multi-dimensional depth convolution according to claim 4, characterized in that, The multi-dimensional depthwise convolution module processes the preprocessed image by dividing it into four sub-images according to the channel dimension. , The original features are obtained through identity mapping. , The subgraphs are processed by three different depthwise convolutions to obtain three feature subgraphs. , and The fused feature sub-image is obtained by splicing the images together.
6. The method for self-attention recognition of ultrasound breast cancer lesions based on multi-dimensional depth convolution according to claim 5, characterized in that, The different dimensions of depth convolution processing employ three convolution kernels with different receptive fields, including a square convolution kernel for extracting spatial dimension features, a strip convolution kernel for extracting horizontal dimension features, and a strip convolution kernel for extracting vertical dimension features.
7. The method for self-attention recognition of ultrasound breast cancer lesions based on multi-dimensional depth convolution according to claim 1, characterized in that, The adaptive dimensionality transformation attention mechanism processes the input fused feature map in the following ways: Based on the fusion feature map Obtain the K, Q, and V matrices. These matrices are then partitioned dimensionally according to the number of attention heads using a linear layer, thus separating the K, Q, and V vectors. The Q vector is then assigned to a pooling layer and processed by DTrans to obtain the A vector. ； in, For query vector, This indicates a dimensional change operation. Indicates pooling operation; After assigning the K and V vectors to the first softmax layer, ... Processing yields V A vector: ； The original Q vector, A vector, and V A The vector is input to the second softmax layer, and the original V vector is simultaneously subjected to depthwise convolution. The outputs of the two layers are concatenated to generate a local feature map. ； in, For adaptive dimensionality transformation attention mechanism processing, This is handled using the softmax attention mechanism. This indicates depthwise convolution processing.