Detection method of breast cancer lesion area in breast images based on convolutional network
Through a convolutional network-based method, Faster-RCNN and CDCN networks are used to extract and fuse higher-order features in breast images to achieve accurate classification and positioning of breast cancer lesion areas, solving the accuracy of breast cancer lesion areas in the prior art. and the problem of low accuracy significantly improves the accuracy of the detection results.
Patent Information
- Application Number
- CN202410082651.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-19
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-01-19
AI Technical Summary
The existing neural network model is inaccurate in the breast cancer lesion area, resulting in low detection accuracy and accuracy. In particular, the diversity distribution characteristics of malignant calcification make the existing model localize and identify the breast cancer lesion area. There are large deviations at the time.
Using a convolutional network-based method, the Faster-RCNN network is used to extract the four target class feature information in the higher-order feature map, and the feature fusion and reconstruction are performed through the CDCN network, combining the ROI classification linear layer and the ROI target box regression linear layer. Achieve accurate classification and precise positioning of breast cancer lesion areas.
This method can reduce the error detection rate of the real target area, improve the accuracy of the category of breast cancer lesion area, and significantly improve the accuracy of the detection results, especially in the identification and localization of malignant breast nodules and calcifications, which show higher results. accuracy.
Smart Images

Figure CN118037644B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to a method for detecting breast cancer lesion areas in breast images based on a convolutional network. Background Art
[0002] The incidence of breast cancer (BC) among women at home and abroad is increasing year by year. It is one of the three major malignant tumors in women. In China, due to the lack of relevant health awareness, the risk of breast cancer among women remains high.
[0003] Breast cancer lesions include breast nodules and calcification. Breast nodules are divided into benign and malignant. Usually, benign nodules do not have much impact on life, but there is a risk of developing into malignant nodules. Usually, calcification is benign, and malignant calcification is usually accompanied by malignant nodules.
[0004] At present, the analysis of breast cancer lesions relies more on human observation of medical images, and it is difficult to get results quickly. The reason is that the Hounsfield Unit (HU) value of breast nodules and surrounding tissues is less different, and the shape, distribution, texture, etc. of breast calcification are unpredictable. In addition, it requires doctors to have high medical knowledge and imaging knowledge. Figure 1 Schematic diagram of breast X-ray images under different conditions, Figure 1 A mammographic image of benign breast calcification. Figure 2 A schematic diagram of a mammary X-ray image of a malignant breast nodule. Figure 3 Schematic diagram of a mammographic image of a benign breast nodule.
[0005] To this end, some people have proposed to use the powerful generalization ability of neural networks to extract high-order breast X-ray features that cannot be recognized by the human eye to achieve the location and classification of target lesions. However, most of the current neural networks can only detect single target lesion areas such as nodules or calcifications. Although they can achieve good single lesion area location and recognition effects, in actual applications, since the basis for judging breast cancer lesion areas is no longer just malignant nodules, but also malignant breast calcifications, and since malignant calcifications are usually accompanied by malignant nodules, and due to the diverse distribution characteristics of calcifications, the existing neural network model will be inaccurate when performing breast cancer lesion area location and recognition detection, which will lead to a large deviation in the detection accuracy of the real target area. At the same time, since the real breast cancer lesion area in the breast image cannot be accurately detected, the accuracy of the detection results will be reduced. Summary of the invention
[0006] The purpose of the present invention is to provide a method for detecting breast cancer lesion areas in breast images based on a convolutional network, which can utilize a Faster-RCNN network and a CDCN network, firstly utilize the high-order feature map of the original Faster-RCNN network containing feature information of four target classes, and then utilize the CDCN convolutional network to achieve accurate classification and precise positioning of the target area, and finally, based on the classification result of the target class and the positioning result of the target class candidate box, select the breast cancer lesion area with accurate classification and accurate positioning, thereby reducing the error detection rate of the real target area and improving the accuracy of the category to which the real target area belongs.
[0007] The present invention solves the technical problem and adopts the following technical solution:
[0008] A method for detecting breast cancer lesion areas in breast images based on a convolutional network comprises the following steps:
[0009] Download the breast image dataset;
[0010] The original Faster-RCNN network is trained by using a breast image dataset, so that after the breast image is input into the original Faster-RCNN network, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a uniform candidate box size, wherein the high-order feature map contains feature information of four target classes;
[0011] Construct and train a CDCN network, replace the hidden layer and input layer of the network structure of the last classification learning in the original Faster-RCNN network with the trained CDCN network, and retain the ROI classification linear layer and the ROI target frame regression linear layer. The trained CDCN network is two identical convolutional networks. The trained CDCN network is used for feature fusion of high-order feature maps, and after feature fusion, the classification of target classes is completed through the ROI classification linear layer and the positioning of candidate frames of each target class is completed through the ROI target frame regression linear layer;
[0012] Based on the classification result of the target class and the positioning result of the target class candidate box, the breast cancer lesion area in the breast image is selected to complete the detection of the breast cancer lesion area in the breast image.
[0013] As a further optimization, the downloaded breast image dataset includes a breast DDSM dataset and a breast VOC2007 dataset;
[0014] The breast VOC2007 dataset is a breast RGB dataset, and the breast DDSM dataset is a breast X-ray dataset;
[0015] The breast VOC2007 dataset is not processed in any way;
[0016] Each breast data in the breast DDSM dataset contains imaging in four directions, and the candidate box is calculated according to the mark in the overlay mark file, and the coordinates of the starting point and the ending point of the candidate box are recorded, and the data format is the same as that of the breast VOC2007 dataset;
[0017] In the overlay mark file, if an image has multiple marks, only one marked area is retained and the retained marked area is used as the candidate box.
[0018] As a further optimization, the high-order feature map contains feature information of four target classes, and the four target classes specifically refer to:
[0019] Benign breast nodules, malignant breast nodules, benign breast calcifications, and malignant breast calcifications.
[0020] As a further optimization, when training the CDCN network, the high-order feature map is input into the CDCN network, and the input of the CDCN network enters the structural feature branch and the detail feature branch respectively to obtain the structural feature vector and the detail feature vector of the high-order feature map, and then enters the MCB module and the RG module to connect the structural feature vector and the detail feature vector. After passing through the MSF module, the structural feature vector and the detail feature vector are interactively fused, and the structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are output, and the structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are input into the ROI classification linear layer and the ROI target frame regression linear layer respectively.
[0021] As a further optimization, the expression of the convolutional network is as follows:
[0022] Fs,Fd=Features(ROIinput)
[0023] The Features(.) function is a separation function used to separate ROI Input Separated into two feature domains, namely the structural feature domain F s , and the detail feature domain F d , ROI input Represents the input of the CDCN network, that is, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a unified candidate box size.
[0024] As a further optimization, the MCB module comes with a residual structure and attention mechanism. Each MCB module consists of two identical convolutional layers of 3×3 and an LReLU structure, where the LReLU structure is in the middle of the two convolutional layers.
[0025] The attention mechanism CA is composed of a pooling layer, two convolutional layers 1×1, an LReLU layer, and an activation layer, and is used to learn inter-channel information. Before the attention mechanism CA, the detail feature vector and the structure feature vector are connected to each other and input into the attention mechanisms CA of the two branches respectively. The data before and after the attention mechanism CA are residually connected using the residual structure, and the data feature vectors before and after the MCB module input are connected.
[0026] Multiple MCB modules form an RG module. The number of RG modules is 1. Multiple MCB modules are connected sequentially, and the output of the previous MCB module serves as the input of the next MCB.
[0027] As a further optimization, the input of the MSF module is the structural feature vector, detail feature vector, and structural feature vector + detail feature vector before interactive feature fusion;
[0028] The MSF module contains a multi-scale feature fusion module. When performing interactive feature fusion, the interactive feature fusion of detail feature vectors and structural feature vectors will reference each other. At the same time, the above three vectors are input into three feature extractors respectively. The extractors are composed of convolutional layers 3×1×1, 5×1×2, and 7×1×2+LReLU layers respectively.
[0029] When performing interactive feature fusion, the Concat operation is used to link the three output vectors and input them into three convolutional layers 3×1×1, 5×1×2, and 7×1×2+LReLU layers respectively. Six convolutional layers 3×1×1+LReLU layers are defined. The input of the first five layers is the output of each previous layer for Concat operation, and the last layer is the output of the fifth layer. At the same time, the input of the six convolutional layers is the Concat operation of three components. Finally, the output of the MSF module is Concat, that is, detail feature vector + structural feature vector + output feature vector of the sixth layer.
[0030] As a further optimization, the MSF module includes an upsampling structure consisting of a convolutional layer 3×2×1 and a PixelShuffle layer.
[0031] As a further optimization, when the structural feature vector and the detail feature vector are interactively fused, a loss adjustment parameter r is introduced, where r=0.9. The total loss function of the CDCN network is as follows:
[0032]
[0033] Among them, L cls and L locrepresents the target detection optimization function term, which represents the target classification and target positioning loss respectively, Θ c , Θ g and They represent the CDCN sub-network optimization function items, namely the loss after the fusion of edge features, structural features and interaction features. The range of the γ parameter is [0,1], which represents the weight ratio parameter and is mainly used to adjust the optimization direction of the entire network.
[0034] The beneficial effects of the present invention are as follows: through the above-mentioned convolutional network-based breast cancer lesion area detection method in breast images, on the one hand, the Faster-RCNN network can be used to extract a high-order feature map containing feature information of four target classes, and on the other hand, the CDCN convolutional network can be used to reconstruct and fuse the high-order feature map, thereby achieving accurate classification and precise positioning of the target area. Finally, based on the classification result of the target class and the positioning result of the target class candidate box, the breast cancer lesion area with accurate classification and accurate positioning can be selected, thereby reducing the error detection rate of the real target area and improving the accuracy of the category to which the real target area belongs. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A schematic diagram of a mammary gland X-ray image of benign breast calcification in the prior art;
[0036] Figure 2 A schematic diagram of a mammary gland X-ray image of a malignant breast nodule in the prior art;
[0037] Figure 3 A schematic diagram of a mammary gland X-ray image of a benign breast nodule in the prior art;
[0038] Figure 4 This is a schematic diagram of the overall network structure of the original Faster-RCNN network + CDCN network in an embodiment of the present invention. DETAILED DESCRIPTION
[0039] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, 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 part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0040] Example
[0041] This embodiment provides a method for detecting breast cancer lesion areas in breast images based on a convolutional network, comprising the following steps:
[0042] S1. Download breast image dataset;
[0043] S2, training the original Faster-RCNN network with a breast image dataset, so that after the breast image is input into the original Faster-RCNN network, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a uniform candidate box size, wherein the high-order feature map contains feature information of four target classes;
[0044] S3. Construct and train a CDCN network, replace the hidden layer and input layer of the network structure of the last classification learning in the original Faster-RCNN network with the trained CDCN network, and retain the ROI classification linear layer and the ROI target frame regression linear layer. The trained CDCN network is two identical convolutional networks. The trained CDCN network is used for feature fusion of high-order feature maps, and after feature fusion, the classification of target classes is completed through the ROI classification linear layer and the positioning of candidate frames of each target class is completed through the ROI target frame regression linear layer;
[0045] S4. Based on the classification result of the target class and the positioning result of the target class candidate box, the breast cancer lesion area in the breast image is selected to complete the detection of the breast cancer lesion area in the breast image.
[0046] In the above method, the downloaded breast image dataset may include a breast DDSM dataset and a breast VOC2007 dataset; the breast VOC2007 dataset is a breast RGB dataset, and the breast DDSM dataset is a breast X-ray dataset; the breast VOC2007 dataset is not processed in any way; each breast data in the breast DDSM dataset contains imaging in four directions, and it is necessary to calculate the candidate box according to the mark in the overlay mark file, and record the starting point coordinates and the end point coordinates of the candidate box, and form the same data format as the breast VOC2007 dataset; in the overlay mark file, if an image has multiple marks, only one mark area is retained, and the retained mark area is used as the candidate box.
[0047] After downloading the breast image dataset, the effectiveness of the overall method of this embodiment can be verified by the breast VOC2007 dataset, and the recognition and positioning of the four major categories of breast nodules and benign and malignant breast calcifications can be verified by the breast DDSM dataset.
[0048] In this embodiment, the original Faster-RCNN network training can be trained using the breast VOC2007 dataset. The mAP index is calculated every 5 iterations, and the weight is retained for each iteration. The training optimizer uses adam, the momentum parameter is 0.9, the initial learning rate is 0.00001, and the minimum learning rate is 0.0000001.
[0049] Moreover, after the CDCN network structure is introduced, it replaces the RCNN in the Faster-RCNN network, and retains the last two layers (ROI classification linear layer, ROI target box regression linear layer). At the same time, the output of the ROIpool layer is used as the input of the CDCN network, where the input channel, output channel, number of features, number of RGs, number of MCB modules, and downsampling scale parameters of the CDCN network are 1024, 256, 2, 1, and 1 respectively. The goal of the ROIPooling structure is to convert features of different scales into features of the same scale. Its MCB and RG group parameters are set to 1, the input channel parameter is 1024, and the output channel parameter is 256. This mAP indicator is calculated every 5 iterations, and this weight is retained for each iteration. The training optimizer uses adam, the momentum parameter is 0.9, the initial learning rate is 0.00001, the minimum learning rate is 0.0000001, the r loss ratio parameter is set to 0.9, and the number of layers of the edge gradient convolution block and the blur convolution block is set to 1. All other parameters are consistent with the Faster-RCNN model training.
[0050] In addition, the training results of the Faster-RCNN model and the Faster-RCNN+CDCN model on the breast VOC2007 dataset can be compared. The Faster-RCNN+CDCN network can have lower loss and map value, and the map value is increased by about 5.623%. The experimental results show that Faster-RCNN+CDCN is feasible for target detection tasks.
[0051] Furthermore, we can compare the training results of Faster-RCNN and Faster-RCNN+CDCN on the breast DDSM dataset. The experimental results show that when nms is equal to 0., 0.3, 0.5, and 0.7, Faster-RCNN+CDCN has higher accuracy than Faster-RCNN in all categories, and its detection rate is more balanced in the four categories, where 0. indicates floating point type.
[0052] See also Figure 4 After ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a uniform candidate box size, in order to achieve accurate recognition and positioning of different target classes, it is necessary to confirm the target categories contained in the high-order feature map. Specifically, the high-order feature map contains feature information of four target classes, and the four target classes specifically refer to:
[0053] Benign breast nodules, malignant breast nodules, benign breast calcifications, and malignant breast calcifications.
[0054] See also Figure 4, the network structure before ROIPooing is the structure of the original Faster-RCNN network. After the image passes through 4 Block modules, the high-order feature map Feature map is obtained. Among them, the First branch of RPN indicates whether the candidate box contains the target information, and the Second branch of RPN indicates that the original candidate box is used as input and input into ROIPooing. Here, ROIPooing can unify the sizes of different candidate boxes into one standard;
[0055] Then, they are respectively sent to the structure feature branch Structure Feature and the detail feature branch Detail Feature of the CDCN network. In this embodiment, MCB, RG and MSF can represent the alternating fusion module of detail and structure features, where MSF includes upsampling, SR represents the features after high-order feature reconstruction; ROI_CLS represents the target classification head, and ROI_LOC represents the target coordinate regression classification head.
[0056] Therefore, when training the CDCN network, the high-order feature map is input into the CDCN network, and the input of the CDCN network enters the structural feature branch and the detail feature branch respectively to obtain the structural feature vector and the detail feature vector of the high-order feature map, and then enters the MCB module and the RG module to connect the structural feature vector and the detail feature vector. After passing through the MSF module, the structural feature vector and the detail feature vector are interactively fused, and the structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are output, and the structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are input into the ROI classification linear layer and the ROI target frame regression linear layer respectively.
[0057] See also Figure 4 , Figure 4The input of the CDCN network in will enter two branches respectively, a structural feature branch and a detail feature branch. The network is two identical tiny convolutional networks, which consists of three convolutional layers (3×3) and two LReLU (Leaky ReLU) layers. In order to better extract edge information and structural information, fixed convolution kernel canny edge operator and fuzzy convolution kernel are introduced. In the four-category screening of benign and malignant nodules and calcifications, malignant nodules are generally accompanied by calcification. Therefore, the extraction of edge and structural features of CDCN can better characterize the relevant features of calcification and nodules. The two modules MCB (Multiple Mutual Collaboration) and RG (Residual Groups) are a system. Multiple MCBs form an RG. CDCN contains a specified number of RGs for network training. The feature extraction of high-order features does not require multiple RGs for cascade operations. Therefore, when Faster-RCNN introduces the CDCN network, the number of RGs is set to 1. The convolutional network formula is as follows:
[0058] Fs,Fd=Features(ROIinput)
[0059] The Features(.) function is a separation function used to separate ROI Input Separated into two feature domains, namely the structural feature domain F s , and the detail feature domain F d , ROI input Represents the input of the CDCN network, that is, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a unified candidate box size.
[0060] Here, MCB comes with a residual structure and attention mechanism. MCB consists of two identical two-layer convolutional networks (3×3) and an LReLU structure, with the LReLU structure in the middle of the two convolutional layers. The attention mechanism CA (ChannelAttention) consists of a pooling layer, two convolutional layers (1×1), an LReLU layer, and an activation layer (Sigmoid). Its purpose is to learn inter-channel information. Before CA, the detail vector and the structure vector will be connected to each other (Concat) and input into the two branches of CA respectively. The data before and after CA are residually connected, and the feature vectors of the data before and after the MCB input are connected. It is worth noting that multiple MCB modules and RG modules are connected sequentially. The output of the previous MCB is used as the input of the next MCB, and the output of the previous RG is used as the input of the next RG. This design pattern ensures that the data can interact fully.
[0061] Here, the MSF (Multi-scale fusion) module is to interact the details and structural features and form outputs. MSF has three outputs, and a new output of fusion features is added. Therefore, the input of MSF is detail features, structural features, detail features + structural features (summation operation), and MSF has a multi-scale feature fusion module. Therefore, in this embodiment, CDCN is essentially a reconstruction network. During the reconstruction process, the reconstruction of detail features and structural features will refer to each other (as parameters). At the same time, the above three vectors are respectively input into three feature extractors, and the extractors are respectively composed of convolutional layers (3×1×1, 5×1×2, 7×1×2) + LReLU layers. In order to better perform feature fusion, the Concat operation is used to link the three output vectors and input them into three convolutional layers (3×1×1, 5×1×2, 7×1×2) + LReLU layers. In order to enhance the features, 6 convolutional layers (3×1×1) + LReLU layers are defined later. The input of the first five layers is the output of each previous layer for Concat operation, and the last layer is the output of the fifth layer. At the same time, the input of the six convolutional layers is the Concat operation of three components. Finally, the output of MSF is Concat (detail features + structural features + output features of the sixth layer). In order to output high resolution, the network designs an upsampling structure, which consists of a convolutional layer (3×2×1) and a PixelShuffle layer. At this point, the CDCN network design is complete. At the same time, in order to optimize the network in the direction of target detection and reconstruction, the loss adjustment parameter r is introduced when the structural feature vector and the detail feature vector are interactively fused, where r=0.9. The total loss function of the CDCN network is as follows:
[0062]
[0063] Among them, L cls and L loc represents the target detection optimization function term, which represents the target classification and target positioning loss respectively, Θ c , Θ g and They represent the CDCN sub-network optimization function items, namely the loss after the fusion of edge features, structural features and interaction features. The range of the γ parameter is [0,1], which represents the weight ratio parameter and is mainly used to adjust the optimization direction of the entire network.
[0064] In the specific application process, in order to connect to the Faster-RCNN network, this embodiment directly uses ROI as the input of the CDCN network to purify the features of ROI. Unlike the original CDCN network, it is necessary to reset the parameters of the depth of the network, mainly the number of MCB modules and RG modules. In this embodiment, MCB group is set to 1, RG goup is set to 1, the input channel is 1024, and the output channel is 256. In the original CDCN network, feature fusion uses the Concat operation. In order to reduce the dimension, this embodiment uses the add operation to directly add the output of the CDCN network. At the same time, the three output dimensions in the original network are inconsistent. In this embodiment, the dimensions are consistent after parameter adjustment. The final output is applied to the two linear layers of the original Faster-RCNN network, one for ROI classification and the other for ROI coordinate frame regression.
[0065] Therefore, CDCN can further distill and purify the high-order features of the Faster-RCNN candidate boxes, remove irrelevant features and reconstruct them. Since the RCNN network in Faster-RCNN is connected to a small convolutional network, the high-order features cannot express the representation of the target area of the original image as much as possible due to the image convolution again. The CDCN network introduces the idea of image reconstruction to purify and distill the features while performing convolution, extracting and reconstructing the feature information of the four target classes in the same transformation field to avoid the imaging of irrelevant features.
[0066] At present, in existing studies, screening for nodules or calcifications alone can achieve good results. However, in real applications, the determination of the area of malignant lesions depends on whether there are calcification points and nodules around them. In addition, the diverse distribution of calcification is one of the main reasons for the extremely low detection rate of existing neural network models.
[0067] Compared with the Faster-RCNN model, the Faster-RCNN+CDCN model in this embodiment can achieve a better balance in the detection rate of multiple targets in the four-category target detection task, and more accurately determine the benign and malignant nature of the lesion area.
[0068] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for detecting breast cancer lesion areas in breast images based on a convolutional network, characterized in that: The steps include: Download the breast image dataset; The original Faster-RCNN network is trained by using a breast image dataset, so that after the breast image is input into the original Faster-RCNN network, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a uniform candidate box size, wherein the high-order feature map contains feature information of four target classes; Construct and train a CDCN network, replace the hidden layer and input layer of the network structure of the last classification learning in the original Faster-RCNN network with the trained CDCN network, and retain the ROI classification linear layer and the ROI target frame regression linear layer. The trained CDCN network is two identical convolutional networks. The trained CDCN network is used for feature fusion of high-order feature maps, and after feature fusion, the classification of target classes is completed through the ROI classification linear layer and the positioning of candidate frames of each target class is completed through the ROI target frame regression linear layer; Based on the classification result of the target class and the positioning result of the target class candidate box, the breast cancer lesion area in the breast image is selected to complete the detection of the breast cancer lesion area in the breast image.
2. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 1, characterized in that: The downloaded breast image datasets include a breast DDSM dataset and a breast VOC2007 dataset; The breast VOC2007 dataset is a breast RGB dataset, and the breast DDSM dataset is a breast X-ray dataset; The breast VOC2007 dataset is not processed in any way; Each breast data in the breast DDSM dataset contains imaging in four directions, and the candidate box is calculated according to the mark in the overlay mark file, and the coordinates of the starting point and the ending point of the candidate box are recorded, and the data format is the same as that of the breast VOC2007 dataset; In the overlay mark file, if an image has multiple marks, only one marked area is retained and the retained marked area is used as the candidate box.
3. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 1, characterized in that: The high-order feature map contains feature information of four target classes, and the four target classes specifically refer to: Benign breast nodules, malignant breast nodules, benign breast calcifications, and malignant breast calcifications.
4. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 1, characterized in that: When training the CDCN network, the high-order feature map is input into the CDCN network. The input of the CDCN network enters the structural feature branch and the detail feature branch respectively to obtain the structural feature vector and the detail feature vector of the high-order feature map. Then, the MCB module and the RG module are entered to connect the structural feature vector and the detail feature vector. After passing through the MSF module, the structural feature vector and the detail feature vector are interactively fused, and the structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are output. The structural feature vector, detail feature vector and structural feature vector + detail feature vector after feature fusion are input into the ROI classification linear layer and the ROI target frame regression linear layer respectively.
5. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 4, characterized in that: The expression of the convolutional network is as follows: F s ,F d =Features(ROI input ) The Features(.) function is a separation function used to separate ROI Input Separated into two feature domains, namely the structural feature domain F s , and the detail feature domain F d , ROI input Represents the input of the CDCN network, that is, the ROIPooling in the original Faster-RCNN network outputs a high-order feature map of a unified candidate box size.
6. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 4, characterized in that: The MCB module comes with a residual structure and attention mechanism. Each MCB module consists of two identical convolutional layers of 3×3 and an LReLU structure, with the LReLU structure in the middle of the two convolutional layers. The attention mechanism CA is composed of a pooling layer, two convolutional layers 1×1, an LReLU layer, and an activation layer, and is used to learn inter-channel information. Before the attention mechanism CA, the detail feature vector and the structure feature vector are connected to each other and input into the attention mechanisms CA of the two branches respectively. The data before and after the attention mechanism CA are residually connected using the residual structure, and the data feature vectors before and after the MCB module input are connected. Multiple MCB modules form an RG module. The number of RG modules is 1. Multiple MCB modules are connected sequentially, and the output of the previous MCB module serves as the input of the next MCB module.
7. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 4, characterized in that: The input of the MSF module is the structural feature vector, detail feature vector, and structural feature vector + detail feature vector before interactive feature fusion; The MSF module contains a multi-scale feature fusion module. When performing interactive feature fusion, the interactive feature fusion of detail feature vectors and structural feature vectors will reference each other. At the same time, the above three vectors are input into three feature extractors respectively. The extractors are composed of convolutional layers 3×1×1, 5×1×2, and 7×1×2+LReLU layers respectively. When performing interactive feature fusion, the Concat operation is used to link the three output vectors and input them into three convolutional layers 3×1×1, 5×1×2, and 7×1×2+LReLU layers respectively. Six convolutional layers 3×1×1+LReLU layers are defined. The input of the first five layers is the output of each previous layer for Concat operation, and the last layer is the output of the fifth layer. At the same time, the input of the six convolutional layers is the Concat operation of three components. Finally, the output of the MSF module is Concat, that is, detail feature vector + structural feature vector + output feature vector of the sixth layer.
8. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 7, characterized in that: The MSF module includes an upsampling structure consisting of a convolutional layer 3×2×1 and a PixelShuffle layer.
9. The method for detecting breast cancer lesion areas in breast images based on convolutional networks according to claim 4, characterized in that: When the structural feature vector and the detail feature vector are interactively fused, a loss adjustment parameter r is introduced, where r=0.
9. The total loss function of the CDCN network is as follows: Among them, L cls and L loc represents the target detection optimization function term, which represents the target classification and target positioning loss respectively, Θ c , Θ g and They represent the CDCN sub-network optimization function terms, namely the loss after the fusion of edge features, structural features and interaction features. The range of the γ parameter is [0,1], which represents the weight ratio parameter.
