Breast cancer pathological image pre-screening method based on deep learning
Through the deep learning model of multi-scale feature extraction and multiple attention mechanisms, the problem of insufficient feature fusion in pathological images is solved, and high-precision screening and risk assessment of breast cancer pathological images are realized.
Patent Information
- Application Number
- CN202510634359.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2025-09-02
AI Technical Summary
The existing subjective scoring methods for pathological images cannot fully extract effective features, the source data set of transfer learning is less similar to the target data set, and there is a lack of fusion processing of local and global features.
Using multi-scale feature extraction and multiple attention mechanisms, breast cancer pathological image features are extracted through DRBlock and CBAM modules, and deep learning models are trained using hybrid Dropout strategy to achieve fusion and accurate classification of multi-scale features.
It improves the screening accuracy of breast cancer pathological images, can accurately give pre-screening results of four levels, prompting breast cancer risk, and providing doctors with diagnostic reference.
Smart Images
Figure CN120580474A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for screening breast cancer pathology images, and in particular to a method for pre-screening breast cancer pathology images based on deep learning. Background Art
[0002] Breast cancer is one of the most common cancers in women worldwide, with the highest prevalence rate globally. Imaging tests based on X-rays, ultrasound, or thermal imaging can only provide preliminary screening, while pathology is the final basis for diagnosing benign or malignant breast lesions. Breast tissue obtained through a biopsy is processed and photographed to create pathological images, which are then diagnosed by a pathologist to confirm the diagnosis of breast cancer. In recent years, with the development of deep learning technology, computer-assisted diagnosis (CAD) has been used to classify breast cancer pathology images, improving diagnostic efficiency while ensuring a certain level of accuracy. This overcomes the traditional clinical diagnosis stage, which primarily relies on pathologists examining samples under a microscope, which is time-consuming and labor-intensive, and can lead to misdiagnosis.
[0003] Aiming at the problems that existing subjective scoring methods for pathological images cannot fully extract effective features, the source dataset and target dataset of transfer learning have little similarity, and there is a lack of fusion of local and global features, the present invention proposes a pre-screening method for breast cancer pathological images that integrates multi-scale features and multiple attention. Summary of the Invention
[0004] The purpose of this invention is to solve the problems existing in the existing image feature extraction process, such as the low similarity between the source dataset and the target dataset of transfer learning, and the lack of a processing method for fusing local features with global features, and to propose a breast cancer pathology image pre-screening method based on deep learning.
[0005] The above purpose is achieved through the following technical solutions: Step 1: Preprocess breast cancer pathology images. The original images are scaled to obtain images of 224×224×3, 160×160×3, and 112×112×3 dimensions, which are used as input for the deep learning network. Step 2: Build a pre-screening network model for breast cancer pathology images based on deep learning and extract image features; Step 3: Train the network model constructed in step 2; Step 4: Use the deep learning-based pre-screening network model for breast cancer pathology images trained in step 3 to pre-screen the image to be tested, and obtain pre-screening results corresponding to the four grades of breast cancer.
[0006] Furthermore, the construction of a pre-screening model for breast cancer pathology images based on deep learning described in step 2 is as follows: The pre-screening model includes three parts: multi-scale feature extraction, multi-scale feature fusion and image classification; The multi-scale feature extraction part uses DRBlock as the backbone network. DRBlock is divided into four stages when extracting features, corresponding to 2 DRBlocks, 4 DRBlocks, 4 DRBlocks, and 2 DRBlocks respectively. Multi-attention CBAM is introduced during feature extraction to extract important information contained in the features. Downsampling is completed through 1×1 convolution and the multi-scale features are fused with the features of the next level. The multi-scale feature maps are subjected to maximum pooling, feature map concatenation and addition operations to obtain the fused breast cancer pathology image features. Among them, DRBlock stands for Deformable Residual Block. DRBlock is a convolutional neural network module in deep learning. DRBlock combines convolution Conv, activation function LeakyReLU, residual connection ResidualConnection and deformable convolution module Deformable Convolution; CBAM integrates channel attention module Channel Attention module and spatial attention module Spatial Attention module; Channel Attention module models the importance of each channel and outputs a weight vector with the same number of channels to weight the original feature map; Spatial Attention module models the feature map spatially and outputs a weight matrix with the same spatial size as the feature map to weight the original feature map; Afterwards, 7×7 convolution and 3×3 maximum pooling were used to extract image features of breast cancer pathology images; Afterwards, the feature extraction network composed of DRBlock and CBAM is used to extract the image features of the preprocessed breast cancer pathology images.
[0007] Furthermore, the steps of training the deep learning-based pre-screening model for breast cancer pathology images constructed in step 2 described in step 3 are as follows: Download and organize public datasets, build a deep learning-based pre-screening model image dataset for breast cancer pathology images, and train the model in step 2. During training, a hybrid Dropout strategy is adopted, that is, the Dropout function and the Dropout2d function are randomly used. The processing flow is as follows: generate a random number in [0,1] and determine whether the random number is greater than 0.5. If it is greater than 0.5, execute the Dropout function; otherwise, execute the Dropout2d function.
[0008] Furthermore, the process of obtaining the pre-screening results corresponding to the four grades of breast cancer described in step 4 is to convert the breast cancer pathology image features into a probability distribution ranging from [0, 1] and 1 through the Softmax function, that is, the pre-screening results.
[0009] The beneficial effects of the present invention are: This deep learning-based breast cancer image pre-screening method utilizes multi-scale features and a multi-attention mechanism to focus on both local and global features at multiple scales. This multi-feature fusion enables more comprehensive and accurate feature information acquisition, resulting in a more precise model. The pre-screening and classification of breast cancer images using the fused features yields four levels of pre-screening results, improving the accuracy of the classification. The levels are used to indicate the risk of breast cancer, providing guidance and reference for doctors in breast cancer pathology diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 It is a flow chart of the method of the present invention; Figure 2 The present invention relates to a pre-screening model for breast cancer pathology images based on deep learning; Figure 3 This is a schematic diagram of the DRBlock module structure involved in the present invention; Figure 4 This is a schematic diagram of the structure of the CBAM module (convolutional block attention module) involved in the present invention; Figure 5 This is a pre-screening model training curve for breast cancer pathology images based on deep learning involved in the present invention; Figure 6 The present invention relates to a breast cancer pathology diagnosis image. Figure 7-10 Schematic diagram of samples at different magnifications of the BreakHis dataset involved in the present invention; Figure 7 It is 40 times, Figure 8 It is 100 times, Figure 9 It is 200 times, Figure 10 It is a multiple of 400; Figure 11 Pre-screening results of the screening model for breast cancer pathology images at different scales. DETAILED DESCRIPTION
[0011] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention. Specific implementation method one:
[0013] A deep learning-based breast cancer pathology image pre-screening method in this embodiment, such as Figure 1 As shown, the method is implemented by the following steps: Step 1: Preprocess breast cancer pathology images. The original images are scaled to obtain images of 224×224×3, 160×160×3, and 112×112×3 dimensions, which are used as input for the deep learning network. Step 2: Build a pre-screening network model for breast cancer pathology images based on deep learning and extract image features; Step 3: Train the network model constructed in step 2; Step 4: Use the deep learning-based pre-screening network model for breast cancer pathology images trained in step 3 to pre-screen the images to be tested, and obtain pre-screening results corresponding to the four levels of breast cancer. The higher the level, the greater the risk of breast cancer, which provides doctors with prompts and references for breast cancer pathology diagnosis. Specific implementation method two:
[0015] Different from the first embodiment, the present embodiment is a method for pre-screening breast cancer pathology images based on deep learning. Step 2 of the method is to construct a pre-screening model for breast cancer pathology images based on deep learning, specifically as follows: The pre-screening model includes three parts: multi-scale feature extraction, multi-scale feature fusion and image classification; The multi-scale feature extraction part mainly uses DRBlock as the backbone network. DRBlock is divided into four stages when extracting features. The four stages correspond to 2 DRBlocks, 4 DRBlocks, 4 DRBlocks, and 2 DRBlocks respectively. In order to focus on important multi-scale features, multi-attention CBAM is introduced when extracting features to fully extract the important information contained in the features. After downsampling through 1×1 convolution, the multi-scale features are fused with the features of the next level. The multi-scale feature maps are subjected to maximum pooling, feature map connection and addition operations to obtain the fused breast cancer pathology image features. DRBlock, short for Deformable Residual Block, is a convolutional neural network module used in deep learning. Combining Convolution (Conv), the LeakyReLU activation function, Residual Connection (ResidualConnection), and Deformable Convolution (Deformable Convolution), DRBlock aims to improve the pre-screening model's ability to model geometric transformations, enabling the network to better handle object deformation and scale changes in images. CBAM effectively enhances feature representation capabilities by integrating the Channel Attention (Channel Attention) and the Spatial Attention (Spatial Attention) modules. The Channel Attention module focuses on the relationship between different channels in the input feature map, specifically which channels are important for the task. It models the importance of each channel and outputs a weight vector equal to the number of channels to weight the original feature map, enhancing feature channels that are beneficial to the task and suppressing channels that are not. The Spatial Attention module focuses on the importance of each location in the feature map. It spatially models the feature map and outputs a weight matrix with the same spatial dimensions to weight the original feature map. Let the model focus more on the key areas in the image and ignore unimportant background information; Step 21: Use 7×7 convolution and 3×3 maximum pooling to extract image features of breast cancer pathology images; Step 22: Use the feature extraction network composed of DRBlock and CBAM to extract the image features of the preprocessed breast cancer pathology image. Specific implementation method three:
[0017] Different from the first embodiment, the present embodiment is a method for pre-screening breast cancer pathology images based on deep learning. The steps of step 3 for training the pre-screening model for breast cancer pathology images based on deep learning constructed in step 2 are as follows: Download and organize public datasets, build a deep learning-based pre-screening model image dataset for breast cancer pathology images, and train the model in step 2. During training, a hybrid Dropout strategy is adopted, that is, the Dropout function and the Dropout2d function are randomly used. The processing flow is as follows: generate a random number in [0,1] and determine whether the random number is greater than 0.5. If it is greater than 0.5, execute the Dropout function; otherwise, execute the Dropout2d function. Specific implementation method four:
[0019] Unlike the third embodiment, in this embodiment, a deep learning-based pre-screening method for breast cancer pathology images is used. In step 4, the process of obtaining the pre-screening results corresponding to the four grades of breast cancer is to convert the breast cancer pathology image features into a probability distribution in the range [0, 1] and 1 using a Softmax function, i.e., the pre-screening result.
[0020] The embodiments disclosed in the present invention are preferred embodiments, but are not limited to them. Ordinary technicians in this field can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. As long as they do not deviate from the spirit of the present invention, they are all within the scope of protection of the present invention.
Claims
1. A deep learning-based pre-screening method for breast cancer pathology images, characterized by: The method is implemented by the following steps: Step 1: Preprocess breast cancer pathology images. The original images are scaled to obtain images of 224×224×3, 160×160×3, and 112×112×3 dimensions, which are used as input for the deep learning network. Step 2: Build a pre-screening network model for breast cancer pathology images based on deep learning and extract image features; Step 3: Train the network model constructed in step 2; Step 4: Use the deep learning-based pre-screening network model for breast cancer pathology images trained in step 3 to pre-screen the image to be tested, and obtain pre-screening results corresponding to the four grades of breast cancer.
2. The method for pre-screening breast cancer pathology images based on deep learning according to claim 1, characterized in that: The construction of a deep learning-based pre-screening model for breast cancer pathology images described in step 2 is as follows: The pre-screening model includes three parts: multi-scale feature extraction, multi-scale feature fusion and image classification; The multi-scale feature extraction part uses DRBlock as the backbone network. DRBlock is divided into four stages when extracting features, corresponding to 2 DRBlocks, 4 DRBlocks, 4 DRBlocks, and 2 DRBlocks respectively. Multi-attention CBAM is introduced during feature extraction to extract important information contained in the features. Downsampling is completed through 1×1 convolution and the multi-scale features are fused with the features of the next level. The multi-scale feature maps are subjected to maximum pooling, feature map concatenation and addition operations to obtain the fused breast cancer pathology image features. Among them, DRBlock stands for Deformable Residual Block. DRBlock is a convolutional neural network module in deep learning. DRBlock combines convolution Conv, activation function LeakyReLU, residual connection ResidualConnection and deformable convolution module Deformable Convolution; CBAM integrates channel attention module Channel Attention module and spatial attention module Spatial Attention module; Channel Attention module models the importance of each channel and outputs a weight vector with the same number of channels to weight the original feature map; Spatial Attention module models the feature map spatially and outputs a weight matrix with the same spatial size as the feature map to weight the original feature map; Afterwards, 7×7 convolution and 3×3 maximum pooling were used to extract image features of breast cancer pathology images; Afterwards, the feature extraction network composed of DRBlock and CBAM is used to extract the image features of the preprocessed breast cancer pathology images.
3. The deep learning-based breast cancer pathology image pre-screening method according to claim 2, characterized in that: The steps of training the deep learning-based pre-screening model for breast cancer pathology images constructed in step 2 described in step 3 are as follows: Download and organize public datasets, build a deep learning-based pre-screening model image dataset for breast cancer pathology images, and train the model in step 2. During training, a hybrid Dropout strategy is adopted, that is, the Dropout function and the Dropout2d function are randomly used. The processing flow is as follows: generate a random number in [0,1] and determine whether the random number is greater than 0.
5. If it is greater than 0.5, execute the Dropout function; otherwise, execute the Dropout2d function.
4. The deep learning-based breast cancer pathology image pre-screening method according to claim 3, characterized in that: The process of obtaining the pre-screening results corresponding to the four grades of breast cancer described in step 4 is to convert the breast cancer pathology image features into a probability distribution ranging from [0, 1] and 1 through the Softmax function, that is, the pre-screening results.