A breast lump detection method and system based on axial modulation and frequency decoupling

Abstract

This invention discloses a method and system for detecting breast masses based on axial modulation and frequency decoupling, belonging to the field of artificial intelligence medical diagnostic technology. The system includes a breast mass detection model, which mainly consists of a backbone network, a feature pyramid network, and a detection head network. The backbone network is based on the RepVGG backbone network and combined with AGLM, enhancing the perception of small lesions through adaptive gating fusion of axial perception and local high-frequency textures. HFEM is designed in the feature pyramid network, utilizing frequency decoupling and mask complementarity mechanisms to eliminate semantic aliasing and suppress background noise. ISADH is introduced in the detection head network, eliminating the misalignment problem between classification and localization tasks through a dynamic routing strategy. This invention effectively improves the detection accuracy of complex breast lesions, balancing computational and inference efficiency, and provides reliable technical support for computer-aided diagnosis of breast cancer.

Description

Technical Field

[0001] This invention relates to a method and system for detecting breast lumps based on axial modulation and frequency decoupling, belonging to the field of artificial intelligence technology. Background Technology

[0002] Breast cancer is one of the leading causes of death among women, and early detection and accurate diagnosis can effectively improve the survival rate of breast cancer patients. Mammography is the most commonly used screening method, but because mammograms mainly rely on imaging differences between tissue densities, the interpretation process faces significant challenges: the contrast between the breast mass and surrounding tissue is low, the boundaries are blurred, and the shape is irregular; the size of lesions varies greatly among different patients, increasing the rate of missed diagnoses through manual diagnosis.

[0003] In recent years, deep learning technologies such as the YOLO series and RetinaNet have been widely used in the automated localization of breast lesions. However, traditional CNN-based methods are still limited by local receptive fields, making it difficult to capture the semantic relationships between lesions and the surrounding large areas of breast tissue. In addition, the edge and texture features of low-contrast small masses are easily lost in deep networks. At the same time, traditional single-stage detectors have feature requirements that conflict between classification and regression localization tasks, leading to spatial feature misalignment and limiting the detection accuracy of complex lesions.

[0004] In view of this, it is indeed necessary to propose a breast mass detection method based on axial modulation and frequency decoupling to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings and problems of existing breast mass detection technologies, such as insufficient feature extraction, easy aliasing in multi-scale fusion, and misalignment of detection head tasks. This invention provides a single-stage breast mass detection method based on axial modulation and frequency decoupling. This method designs an axial global-local modulation module (AGLM) at the key level of feature extraction. Through parallel local gradient aggregation and orthogonal axial long-distance sensing, combined with a spatial dynamic gating mechanism, the network can adaptively rely on high-frequency geometric details in obvious edge regions and on broad macroscopic tissue context in blurred overlapping regions with extremely low computational and parameter overhead. This effectively solves the problem of easily missed detection of small, low-contrast masses.

[0006] The technical solution adopted by this invention to solve its technical problem is: a single-stage breast mass detection method based on axial modulation and frequency decoupling, the method comprising the following steps: Step 1: Obtain the mammogram image dataset and preprocess it. Use the preprocessed sample dataset to train the constructed single-stage breast mass detection network model AMFD-RetinaNet. Step 2: The input image is initially extracted through the backbone network of the detection model. An axial global-local modulation module (AGLM) is connected after the key layer output node of the RepVGG-based backbone network. Features are extracted by parallel local texture refinement and orthogonal axial long-distance modeling. Adaptive modulation is performed through a dynamic gating mechanism to obtain a multi-scale feature map. Step 3: The multi-scale feature map is received through the Feature Pyramid Network (FPN) of the detection model and fused across scales. In the feature fusion stage, a Hybrid Feature Extraction Module (HFEM) is added to each output layer. Based on the frequency domain decoupling idea, low-frequency global semantic features and high-frequency local edge features are extracted respectively, and mask complementation processing is performed to output a fused feature map that eliminates semantic aliasing. Step 4: The detection head network of the detection model performs target recognition on the fused feature map. The instance-specific asymmetric decoupling head ISADH is used to receive the feature map. The feature flow is allocated to the asymmetric dual-track topology through an adaptive feature dynamic routing mechanism. The confidence score of the breast mass category and the spatial bounding box coordinate offset corresponding to each anchor box are output in parallel.

[0007] Preferably, in step 2 of this invention, RepVGG is performed by concatenating multiple RepVGG Stages to extract features. Before each RepVGG Stage begins, initial feature processing and spatial downsampling are first performed using a RepDownBlock, and then n consecutive RepBlocks are cascaded for nonlinear mapping of deep features. For a RepDownBlock, the input features are split into two parallel branches, which are then processed by... The combination of convolution and BN, and A combination of convolution and batch normalization (BN). The two feature streams are element-wise summed at the end, and then the output is activated by the ReLU function. For RepBlock, it further extends RepDownBlock into a three-branch structure: in addition to containing... Convolutional branching and In addition to the convolutional branch, an identity mapping branch containing only BN layers is introduced. These three feature paths are ultimately subjected to element-wise addition and ReLU activation.

[0008] Preferably, in step 2 of the present invention, the axial global-local modulation module AGLM assigns the input multi-scale feature map to two parallel feature extraction paths: a local texture branch and a global axial modeling branch; The local texture branch first aggregates spatial local gradient changes using depthwise separable convolution, then integrates cross-channel information through pointwise convolution, and finally completes nonlinear mapping by batch normalized normalization (BN) and GELU activation functions to output local features. The calculation formula for this process is as follows:

[0009] In the formula, The final output is for local features. The input is a multi-scale feature map. Indicates the core size is The depth of the separable convolutional layer, Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. This represents the GELU activation function.

[0010] The global axial modeling branch first halves the channel dimension through convolutional projection, then performs orthogonal axial scanning using vertical and horizontal convolutions respectively. After concatenating the features in the horizontal and vertical directions, it integrates the channel information through feature fusion and dimensionality-up projection convolution. Finally, it restores the original number of channels by sequentially applying batch normalization (BN) and GELU activation functions, outputting the global features. The calculation formula for this process is as follows:

[0011]

[0012] In the formula, These are the projection features after dimensionality reduction. For the final global feature output, The input is a multi-scale feature map. and These represent large-kernel depthwise convolutions in the vertical and horizontal directions, respectively. This indicates a splicing operation at the channel level. Features after splicing for Feature fusion and up-dimensional projection convolution.

[0013] Preferably, in step 2 of the present invention, the axial global-local modulation module (AGLM) introduces a dynamic gating fusion mechanism, and the specific calculation process is as follows: First, a spatially gated weight matrix is ​​generated from the original input feature map using a convolutional layer and a sigmoid activation function. :

[0014] In the formula, This is the Sigmoid activation function.

[0015] Then, the local features and the global features are weighted and modulated using the gated weight matrix to obtain the modulated fused features. :

[0016] In the formula, This is for element-wise multiplication.

[0017] Next, the modulated fused features are added to the input feature map using a residual structure:

[0018] Subsequently, the data is fed into a feedforward network (FFN) for deep nonlinear feature extraction. The mathematical expression for FFN is:

[0019] In the formula, Represents the GELU activation function; represent Upward convolution; represent Dimensionality-reduced projective convolution.

[0020] Finally, output the results. The complete mathematical expression is:

[0021] Preferably, in step 3, the hybrid feature extraction module HFEM assigns the input features to the low-frequency global semantic modeling branch based on Transformer and the high-frequency edge texture branch based on CNN; The low-frequency global semantic modeling branch based on Transformer described in this invention maps input features to a query matrix. Key matrix Sum matrix A linear attention mechanism is introduced, using the ReLU activation function as the kernel function to perform non-negative mapping on the query matrix and the key matrix. First, the dot product of the key matrix and the value matrix is ​​calculated to generate a global context matrix. Then, this context matrix is ​​multiplied by the query matrix and multiplied by a feature scaling factor to reconstruct low-frequency global semantic features. :

[0022] In the formula, Represents the ReLU activation function; Let be the channel dimension of the matrix. This is used to prevent gradient instability caused by excessively large inner product values. (After calculation) Will be reshaped back Spatial dimension.

[0023] The high-frequency edge texture branch based on CNN described in this invention uses depthwise separable convolution to aggregate micro-features within the local receptive field of the input features, performs cross-channel integration via batch normalization (BN) and pointwise convolution, and uses the Sigmoid activation function to generate high-frequency saliency edge masks with values ​​between 0 and 1. .

[0024]

[0025] In the formula, Represents the input feature map, Represents the Sigmoid activation function. Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. Indicates the core size is The depth of the separable convolutional layer.

[0026] After completing the dual-branch feature extraction, the hybrid feature extraction module HFEM of this invention uses the high-frequency salient edge mask extracted based on the CNN branch to perform pixel-by-pixel spatial filtering and complementary calculation on the low-frequency global semantic features generated based on the Transformer branch.

[0027] In the formula, For element-wise multiplication, This represents the features obtained after filtering.

[0028] The filtered features are passed through the output projection layer and fused with the original input features using a residual structure with a learnable scalar factor.

[0029] In the formula, express Output projection layer, This represents a learnable scalar factor.

[0030] Preferably, in step 4 of the present invention, the instance-specific asymmetric decoupling header ISADH includes a shared backbone layer and an instance bootstrap module IGM; The shared backbone layer extracts basic shared semantic representations through continuous convolutions, batch normalized normalization (BN), and ReLU activation functions. :

[0031] In the formula, Represents shared convolution. This represents the ReLU activation function.

[0032] The instance guidance module (IGM) dynamically generates spatially assigned weights for specific lesions within the receptive field, mapping the shared semantic representation into a pixel-by-pixel dual-channel spatial gating matrix. :

[0033] In the formula, Represents the part used to generate the route mask. convolution, It operates along the channel dimension to ensure that the sum of the allocation probabilities of all spatial pixels is 1.

[0034] The instance guidance module (IGM) of this invention applies adaptive instance-level modulation to the shared semantic representation, and applies a classification mask. and regression mask By performing element-wise multiplication with the shared semantic representation and adding a zero-centered translation constant, classification and regression information are effectively separated in the feature space.

[0035]

[0036] The ISADH specific asymmetric decoupling head described in this invention enters an asymmetric dual-track topology based on the separated feature information; The asymmetric dual-track topology described in this invention includes a classification branch and a regression branch; The classification branch described in this invention employs a topological link containing consecutive dilated convolutional layers and introduces a global semantic bypass. It utilizes GAP to extract global feature vectors and injects them back into the backbone via a fully connected MLP network.

[0037] In the formula, Represents a continuous, hollow convolutional layer. Represents a fully connected network. This represents the global average pooling operation.

[0038] The regression branch described in this invention employs a shallow structure with two standard convolutional layers, and a coordinate attention module is cascaded at the end of the link to independently capture the spatial distribution characteristics of lesion boundaries along the horizontal and vertical directions, respectively.

[0039] In the formula, Representing two Convolutional layer This represents the coordinate attention module.

[0040] The ends of the classification branch and the regression branch are respectively connected to the output of the prediction convolutional layer.

[0041] Preferably, in step 4 of this invention, the classification loss function used for model training is Focal Loss, which is used to alleviate the imbalance between positive and negative samples and to encourage the network to focus on optimizing weak lesions that are difficult to detect. Its calculation formula is as follows:

[0042] In the formula, This represents the model's predicted probability of the true class. The lower the value, the more difficult it is for the sample to be correctly classified; The representative parameter is used to dynamically adjust the loss weights of samples, forcing the network to focus on optimizing weak lesions that are difficult to detect. This represents a balancing parameter used to mitigate the imbalance in the ratio of positive to negative samples.

[0043] The regression loss function used in the model training of this invention is the CIoU loss function, and its calculation formula is as follows:

[0044] In the formula, The ratio of the intersection area to the union area of ​​the predicted bounding box and the actual lesion bounding box; Representing the Euclidean distance from the center point, in breast imaging, this term can guide the model to quickly locate the core area of ​​the mass; It represents the diagonal length of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; This represents the aspect ratio penalty term, used to constrain the geometric consistency between the predicted bounding box and the ground truth bounding box.

[0045] This invention also proposes a breast mass detection system based on axial modulation and frequency decoupling. The system includes: an input unit for inputting the breast X-ray image data to be detected; a model prediction unit, which incorporates the aforementioned single-stage breast mass detection network model AMFD-RetinaNet; wherein the backbone network enhances its sensitivity to irregular small lesions in a hidden dense background by integrating the axial global-local modulation module AGLM; the feature pyramid network eliminates semantic aliasing caused by upsampling and suppresses background noise from dense glands by integrating the hybrid feature extraction module HFEM; the detection head network achieves spatial alignment of classification and regression features by integrating the instance-specific asymmetric decoupling head ISADH; and an output unit for parsing the output of the model prediction unit, generating and displaying the category confidence score and high-precision bounding box coordinates of the target breast mass on the original image.

[0046] Beneficial effects: This invention designs an Axial Global-Local Modulation (AGLM) module at the key level of feature extraction. Through parallel local gradient aggregation and orthogonal axial long-distance perception, combined with a spatial dynamic gating mechanism, the network can adaptively rely on high-frequency geometric details in obvious edge regions and on a broad macroscopic organizational context in blurred overlapping regions with extremely low computational and parameter overhead. This effectively solves the problem of easy missed detection of small, low-contrast tumors.

[0047] This invention designs a hybrid feature extraction module HFEM in the multi-scale fusion stage. Taking the perspective of frequency domain decoupling, it uses the high-frequency saliency edge mask extracted by the CNN path to perform pixel-by-pixel spatial filtering and mask complementation on the low-frequency global macroscopic semantic features generated by the Transformer path. This not only successfully eliminates the multi-scale semantic feature aliasing caused by cross-level upsampling operations, but also greatly enhances the network's ability to remove harsh noise and dense glandular tissue background interference, and significantly suppresses false positives caused by background noise.

[0048] This invention innovatively designs an instance-specific asymmetric decoupling head (ISADH) in the detection head section. Through the instance guidance module (IGM) dynamic routing mechanism, the shared feature stream is accurately distributed to the asymmetric dual-track topology link. This design simultaneously satisfies the translation invariance requirement of the classification task and the spatial variance sensitivity requirement of the regression task, and achieves a high degree of fit between the class confidence and the lesion bounding box, generating a high-precision predicted bounding box that fits the real irregular contour of the lesion. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the breast mass detection model based on axial modulation and frequency decoupling described in this invention.

[0050] Figure 2 This is a schematic diagram of the RepVGG Stage, RepDownBlock, and RepBlock structures in the backbone network described in this invention.

[0051] Figure 3 This is a schematic diagram of the axial global-local modulation module (AGLM) structure described in this invention.

[0052] Figure 4 This is a schematic diagram of the HFEM hybrid feature extraction module described in this invention.

[0053] Figure 5 This is a schematic diagram of a specific asymmetric decoupling head ISADH structure for an example of the present invention.

[0054] Figure 6 This is a schematic diagram of the ablation experiment of the breast mass detection model of the present invention. Detailed Implementation

[0055] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0056] like Figure 1 As shown, this invention provides a single-stage breast mass detection method based on axial modulation and frequency decoupling, comprising the following steps: Step 1: Obtain the mammogram image dataset and preprocess it. Use the preprocessed sample dataset to train the constructed single-stage breast mass detection network model AMFD-RetinaNet. Step 2: The input image is initially extracted through the backbone network of the detection model. An axial global-local modulation module (AGLM) is connected after the key layer output node of the RepVGG-based backbone network. Features are extracted by parallel local texture refinement and orthogonal axial long-distance modeling. Adaptive modulation is performed through a dynamic gating mechanism to obtain a multi-scale feature map. Step 3: The multi-scale feature map is received through the Feature Pyramid Network (FPN) of the detection model and fused across scales. In the feature fusion stage, a Hybrid Feature Extraction Module (HFEM) is added to each output layer. Based on the frequency domain decoupling idea, low-frequency global semantic features and high-frequency local edge features are extracted respectively, and mask complementation processing is performed to output a fused feature map that eliminates semantic aliasing. Step 4: The detection head network of the detection model performs target recognition on the fused feature map. The instance-specific asymmetric decoupling head ISADH is used to receive the feature map. The feature flow is allocated to the asymmetric dual-track topology through an adaptive feature dynamic routing mechanism. The confidence score of the breast mass category and the spatial bounding box coordinate offset corresponding to each anchor box are output in parallel.

[0057] like Figure 2 As shown, RepVGG extracts features by concatenating multiple RepVGG Stages. Before each RepVGG Stage begins, initial feature processing and spatial downsampling are performed using a RepDownBlock. Then, n consecutive RepBlocks are cascaded for nonlinear mapping of deep features. For a RepDownBlock, the input features are split into two parallel branches, each processed by... The combination of convolution and BN, and A combination of convolution and batch normalization (BN). The two feature streams are element-wise summed at the end, and then the output is activated by the ReLU function. For RepBlock, it further extends RepDownBlock into a three-branch structure: in addition to containing... Convolutional branching and In addition to the convolutional branch, an identity mapping branch containing only BN layers is introduced. These three feature paths are ultimately subjected to element-wise addition and ReLU activation.

[0058] like Figure 3 As shown, the Axial Global-Local Modulation Module (AGLM) assigns the input multi-scale feature map to two parallel feature extraction paths: a local texture branch and a global axial modeling branch. The local texture branch first aggregates spatial local gradient changes using depthwise separable convolution, then integrates cross-channel information through pointwise convolution, and finally completes nonlinear mapping by batch normalized normalization (BN) and GELU activation functions to output local features. The calculation formula for this process is as follows:

[0059] In the formula, The final output is for local features. The input is a multi-scale feature map. Indicates the core size is The depth of the separable convolutional layer, Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. This represents the GELU activation function.

[0060] The global axial modeling branch first halves the channel dimension through convolutional projection, then performs orthogonal axial scanning using vertical and horizontal convolutions respectively. After concatenating the features in the horizontal and vertical directions, it integrates the channel information through feature fusion and dimensionality-up projection convolution. Finally, it restores the original number of channels by sequentially applying batch normalization (BN) and GELU activation functions, outputting the global features. The calculation formula for this process is as follows:

[0061]

[0062] In the formula, These are the projection features after dimensionality reduction. For the final global feature output, The input is a multi-scale feature map. and These represent large-kernel depthwise convolutions in the vertical and horizontal directions, respectively. This indicates a splicing operation at the channel level. Features after splicing for Feature fusion and up-dimensional projection convolution.

[0063] To adaptively combine the features extracted from the two branches at different spatial locations, the Axial Global-Local Modulation (AGLM) module introduces a dynamic gating fusion mechanism, the specific calculation process of which is as follows: First, a spatially gated weight matrix is ​​generated from the original input feature map using a convolutional layer and a sigmoid activation function. :

[0064] In the formula, This is the Sigmoid activation function.

[0065] Then, the local features and the global features are weighted and modulated using the gated weight matrix to obtain the modulated fused features. :

[0066] In the formula, This is for element-wise multiplication.

[0067] Next, the modulated fused features are added to the input feature map using a residual structure:

[0068] Subsequently, the data is fed into a feedforward network (FFN) for deep nonlinear feature extraction. The mathematical expression for FFN is:

[0069] In the formula, Represents the GELU activation function; represent Upward convolution; represent Dimensionality-reduced projective convolution.

[0070] Finally, output the results. The complete mathematical expression is:

[0071] like Figure 4 As shown, the hybrid feature extraction module HFEM assigns input features to a low-frequency global semantic modeling branch based on Transformer and a high-frequency edge texture branch based on CNN. The Transformer-based low-frequency global semantic modeling branch maps the input features to a query matrix. Key matrix Sum matrix A linear attention mechanism is introduced, using the ReLU activation function as the kernel function to perform non-negative mapping on the query matrix and the key matrix. First, the dot product of the key matrix and the value matrix is ​​calculated to generate a global context matrix. Then, this context matrix is ​​multiplied by the query matrix and multiplied by a feature scaling factor to reconstruct low-frequency global semantic features. :

[0072] In the formula, Represents the ReLU activation function; Let be the channel dimension of the matrix. This is used to prevent gradient instability caused by excessively large inner product values. (After calculation) Will be reshaped back Spatial dimension.

[0073] The CNN-based high-frequency edge texture branch uses depthwise separable convolution to aggregate micro-features within the local receptive field of the input features. It then integrates these features across channels via batch normalization (BN) and pointwise convolution, and uses a sigmoid activation function to generate high-frequency saliency edge masks with values ​​between 0 and 1. .

[0074]

[0075] In the formula, Represents the input feature map, Represents the Sigmoid activation function. Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. Indicates the core size is The depth of the separable convolutional layer.

[0076] After the dual-branch feature extraction is completed, the hybrid feature extraction module HFEM uses the high-frequency salient edge mask extracted based on the CNN branch to perform pixel-by-pixel spatial filtering and complementary calculation on the low-frequency global semantic features generated based on the Transformer branch:

[0077] In the formula, For element-wise multiplication, This represents the features obtained after filtering.

[0078] The filtered features are passed through the output projection layer and fused with the original input features using a residual structure with a learnable scalar factor.

[0079] In the formula, express Output projection layer, This represents a learnable scalar factor.

[0080] like Figure 5 As shown, the instance-specific asymmetric decoupling header ISADH contains a shared backbone layer and an instance bootstrap module (IGM). The shared backbone layer extracts basic shared semantic representations through continuous convolutions, batch normalized normalization (BN), and ReLU activation functions. :

[0081] In the formula, Represents shared convolution. This represents the ReLU activation function.

[0082] The instance guidance module (IGM) dynamically generates spatially assigned weights for specific lesions within the receptive field, mapping the shared semantic representation into a pixel-by-pixel dual-channel spatial gating matrix. :

[0083] In the formula, Represents the part used to generate the route mask. convolution, It operates along the channel dimension to ensure that the sum of the allocation probabilities of all spatial pixels is 1.

[0084] The instance guidance module (IGM) applies adaptive instance-level modulation to the shared semantic representation and applies a classification mask. and regression mask By performing element-wise multiplication with the shared semantic representation and adding a zero-centered translation constant, classification and regression information are effectively separated in the feature space.

[0085]

[0086] The instance-specific asymmetric decoupling head ISADH enters an asymmetric dual-track topology based on the separated feature information. The asymmetric dual-track topology includes classification branches and regression branches; The classification branch employs a topology link containing consecutive dilated convolutional layers and introduces a global semantic bypass. It utilizes GAP to extract global feature vectors and injects them back into the backbone via a fully connected MLP network.

[0087] In the formula, Represents a continuous, hollow convolutional layer. Represents a fully connected network. This represents the global average pooling operation.

[0088] The regression branch employs a shallow structure with two standard convolutional layers, and a coordinate attention module is cascaded at the end of the link to independently capture the spatial distribution characteristics of lesion boundaries along the horizontal and vertical directions, respectively.

[0089] In the formula, Representing two Convolutional layer This represents the coordinate attention module.

[0090] The ends of the classification branch and the regression branch are respectively connected to the output of the prediction convolutional layer.

[0091] The classification loss function used in model training is Focal Loss, which is used to alleviate the imbalance between positive and negative samples and to encourage the network to focus on optimizing weak lesions that are difficult to detect. Its calculation formula is as follows:

[0092] In the formula, This represents the model's predicted probability of the true class. The lower the value, the more difficult it is for the sample to be correctly classified; The representative parameter is used to dynamically adjust the loss weights of samples, forcing the network to focus on optimizing weak lesions that are difficult to detect. This represents a balancing parameter used to mitigate the imbalance in the ratio of positive to negative samples.

[0093] The regression loss function used for model training is the CIoU loss function, and its calculation formula is as follows:

[0094] In the formula, The ratio of the intersection area to the union area of ​​the predicted bounding box and the actual lesion bounding box; Representing the Euclidean distance from the center point, in breast imaging, this term can guide the model to quickly locate the core area of ​​the mass; It represents the diagonal length of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; This represents the aspect ratio penalty term, used to constrain the geometric consistency between the predicted bounding box and the ground truth bounding box.

[0095] Finally, the INbreast dataset was used to train the breast lump detection network model. 80% of the dataset was used as the training set and 20% as the test set. The trained detection model was then used to make predictions on the test set.

[0096] The simulation experimental data of this invention include the following: like Figure 5 As shown, in order to verify the effectiveness of the improved method proposed in this invention for the breast mass detection model and the specific performance of the improved module, a model ablation simulation experiment was conducted. The training process lasted for 300 epochs, with the learning rate set to 0.0001 and decayed in segments at epochs 20, 40, 60, and 80.

[0097] The ablation simulation experiment described in this invention uses precision, recall, and mean average precision (mAP) as evaluation metrics for model performance. Precision is the accuracy when the intersection-over-union (IOU) ratio between the predicted and ground truth boxes is above 0.5. mAP is the average value calculated with a step size of 0.05 for IOU thresholds ranging from 0.5 to 0.95. By calculating the average precision under different thresholds, mAP can provide a more comprehensive performance evaluation.

[0098] like Figure 5As shown, the ablation simulation experiment of this invention consists of 4 groups of experiments. Among them, as shown in Experiment A, the baseline model achieved a mAP of 77.42%. Although its precision was acceptable, its recall was only 73.18%. This indicates that relying solely on the traditional baseline network still presents a significant bottleneck in feature extraction and missed detection when dealing with small breast lesions with extremely low contrast or blurred morphology. As shown in Experiment B, when the Axial Global-Local Modulation (AGLM) module was introduced into the network feature extraction layer, the model's mAP improved by 3.95%, and the recall improved by 6.64%. This demonstrates that AGLM enhances the network's sensitivity to irregular lesions hidden in dense backgrounds by combining orthogonal axial long-range modeling with dynamic gating fusion of local high-frequency textures. As shown in Experiment C, the hybrid feature extraction module HFEM was further added to the multi-scale fusion stage of FPN in Experiment B, which further improved mAP by 3.80% and significantly increased precision by 5.84%. This indicates that the HFEM module can effectively eliminate semantic aliasing caused by upsampling and suppress noise interference caused by dense glandular tissue in the background, thereby reducing the occurrence of false positives. As shown in Experiment D, replacing the traditional detection head with the instance-specific asymmetric decoupling head ISADH resulted in a model mAP of 88.52% and a peak F1-Score of 0.8538. This shows that adding ISADH achieved an optimal balance between precision and recall, enabling the network to not only detect complex lesions but also achieve high-precision bounding box fitting.

[0099] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting breast masses based on axial modulation and frequency decoupling, characterized in that, The method includes the following steps: Step 1: Obtain the mammogram image dataset and preprocess it. Train the constructed single-stage breast mass detection network model AMFD-RetinaNet using the preprocessed sample dataset, and then proceed to Step 2. Step 2: The input image is initially extracted by the backbone network of the detection model. After the key layer output node of the RepVGG-based backbone network, the Axial Global-LocalModulator (AGLM) module is connected. Features are extracted by parallel local texture refinement and orthogonal axial long-distance modeling. Adaptive modulation is performed through dynamic gating mechanism to obtain multi-scale feature maps, and then proceed to step 3. Step 3: The multi-scale feature maps are received through the Feature Pyramid Network (FPN) of the detection model and fused across scales. In the feature fusion stage, a Hybrid Feature Extraction Module (HFEM) is added to each output layer. Based on the frequency domain decoupling idea, low-frequency global semantic features and high-frequency local edge features are extracted respectively, and mask complementation processing is performed to output a fused feature map that eliminates semantic aliasing, and then proceed to step 4. Step 4: The detection head network of the detection model performs target recognition on the fused feature map. The feature map is received using an instance-specific asymmetric decoupled head (ISADH). The feature flow is allocated to the asymmetric dual-track topology through an adaptive feature dynamic routing mechanism. The confidence score of the breast mass category and the spatial bounding box coordinate offset corresponding to each anchor box are output in parallel.

2. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 2, RepVGG performs feature extraction by concatenating multiple RepVGG Stages. Before each RepVGG Stage begins, initial feature processing and spatial downsampling are performed using a RepDownBlock. Then, n consecutive RepBlocks are cascaded for nonlinear mapping of deep features. For a RepDownBlock, the input features are split into two parallel branches, which are then processed by... The combination of convolution and batch normalization (BN), and The combination of convolution and batch normalization (BN) involves element-wise addition of the two feature streams at the end, followed by output through the ReLU activation function. For RepBlock, it further extends RepDownBlock into a three-branch structure: in addition to containing... Convolutional branching and In addition to the convolutional branch, an identity mapping branch containing only BN layers is introduced. These three feature paths are ultimately subjected to element-wise addition and ReLU activation.

3. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 2, the axial global-local modulation module AGLM assigns the input multi-scale feature map to two parallel feature extraction paths: a local texture branch and a global axial modeling branch. The local texture branch first aggregates spatial local gradient changes using depthwise separable convolution, then integrates cross-channel information through pointwise convolution, and finally completes nonlinear mapping by batch normalized normalization (BN) and GELU activation functions to output local features. The calculation formula for this process is as follows: In the formula, The final output is for local features. The input is a multi-scale feature map. Indicates the core size is The depth of the separable convolutional layer, Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. Represents the GELU activation function; The global axial modeling branch first halves the channel dimension through convolutional projection, then performs orthogonal axial scanning using vertical and horizontal convolutions respectively. After concatenating the features in the horizontal and vertical directions, it integrates the channel information through feature fusion and dimensionality-up projection convolution, and then restores the original number of channels by batch normalization (BN) and GELU activation functions to output global features. The calculation formula for this process is as follows: In the formula, These are the projection features after dimensionality reduction. For the final global feature output, The input is a multi-scale feature map. and These represent large-kernel depthwise convolutions in the vertical and horizontal directions, respectively. This indicates a splicing operation at the channel level. Features after splicing for Feature fusion and up-dimensional projection convolution.

4. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 2, the axial global-local modulation module (AGLM) introduces a dynamic gating fusion mechanism, and the specific calculation process is as follows: First, a spatially gated weight matrix is ​​generated from the original input feature map using a convolutional layer and a sigmoid activation function. : In the formula, Use the Sigmoid activation function; Then, the local features and the global features are weighted and modulated using the gated weight matrix to obtain the modulated fused features. : In the formula, For element-wise multiplication; Next, the modulated fused features are added to the input feature map using a residual structure: Subsequently, the data is fed into a feed-forward network (FFN) for deep nonlinear feature extraction. The mathematical expression for the FFN is as follows: In the formula, Represents the GELU activation function; represent Upward convolution; represent Dimensionality reduction projective convolution, Finally, output the results. The complete mathematical expression is: 。 5. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 3, the hybrid feature extraction module HFEM assigns the input features to the low-frequency global semantic modeling branch based on Transformer and the high-frequency edge texture branch based on Convolutional Neural Networks (CNN). The Transformer-based low-frequency global semantic modeling branch maps the input features to a query matrix. Key matrix Sum matrix A linear attention mechanism is introduced, using the ReLU activation function as the kernel function to perform non-negative mapping on the query matrix and the key matrix. First, the dot product of the key matrix and the value matrix is ​​calculated to generate a global context matrix. Then, this context matrix is ​​multiplied by the query matrix and multiplied by a feature scaling factor to reconstruct low-frequency global semantic features. : In the formula, Represents the ReLU activation function; Let be the channel dimension of the matrix. To prevent gradient instability caused by excessively large inner product values, after calculation... Will be reshaped back Spatial dimension, The CNN-based high-frequency edge texture branch uses depthwise separable convolution to aggregate micro-features within the local receptive field of the input features. It then integrates these features across channels via batch normalization (BN) and pointwise convolution, and uses a sigmoid activation function to generate high-frequency saliency edge masks with values ​​between 0 and 1. , In the formula, Represents the input feature map, Represents the Sigmoid activation function. Indicates the core size is Pointwise convolutional layers, This indicates a batch normalization operation. Indicates the core size is The depth of the separable convolutional layer, After completing the dual-branch feature extraction, the hybrid feature extraction module HFEM uses the high-frequency salient edge mask extracted based on the CNN branch to perform pixel-by-pixel spatial filtering and complementation calculation on the low-frequency global semantic features generated based on the Transformer branch: In the formula, For element-wise multiplication, This represents the features obtained after filtering; The filtered features are passed through the output projection layer and fused with the original input features using a residual structure with a learnable scalar factor. In the formula, express Output projection layer, This represents a learnable scalar factor.

6. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 4, the instance-specific asymmetric decoupling header ISADH includes a shared backbone layer and an instance guidance module (IGM). The shared backbone layer extracts basic shared semantic representations through continuous convolutions, batch normalized normalization (BN), and ReLU activation functions. : In the formula, Represents shared convolution. Represents the ReLU activation function; The instance guidance module (IGM) dynamically generates spatially assigned weights for specific lesions within the receptive field, mapping the shared semantic representation into a pixel-by-pixel dual-channel spatial gating matrix. : In the formula, Represents the part used to generate the route mask. convolution, It operates along the channel dimension to ensure that the sum of the allocation probabilities of all spatial pixels is 1; The instance guidance module (IGM) applies adaptive instance-level modulation to the shared semantic representation and applies a classification mask. and regression mask By performing element-wise multiplication with the shared semantic representation and adding a zero-centered translation constant, classification and regression information are effectively separated in the feature space. In the formula, This represents element-wise multiplication. It is a zero-centered translation constant used to ensure the stability of the feature flow and gradient smoothing in the early stages of network training.

7. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 4, the instance-specific asymmetric decoupling head ISADH enters the asymmetric dual-track topology based on the separated feature information. The asymmetric dual-track topology includes classification branches and regression branches; The classification branch employs a topological link containing consecutive dilated convolutional layers and introduces a global semantic bypass. Global average pooling (GAP) is used to extract global feature vectors, which are then injected back into the backbone via a fully connected network (Multilayer Perceptron, MLP). In the formula, Represents a continuous, hollow convolutional layer. Represents a fully connected network. This represents a global average pooling operation; The regression branch employs a shallow structure with two standard convolutional layers, and a coordinate attention module is cascaded at the end of the link to independently capture the spatial distribution characteristics of lesion boundaries along the horizontal and vertical directions, respectively. In the formula, Representing two Convolutional layer Represents the coordinate attention module; The classification branch and the regression branch are respectively connected to the output of the prediction convolutional layer.

8. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 1, characterized in that: In step 4, the classification loss function used for model training is Focal Loss, which is used to alleviate the imbalance between positive and negative samples and to encourage the network to focus on optimizing weak lesions that are difficult to detect. Its calculation formula is as follows: In the formula, This represents the model's predicted probability of the true class. The lower the value, the more difficult it is for the sample to be correctly classified; The representative parameter is used to dynamically adjust the loss weights of samples, forcing the network to focus on optimizing weak lesions that are difficult to detect. This represents a balancing parameter used to mitigate the imbalance in the ratio of positive to negative samples.

9. The method for detecting breast masses based on axial modulation and frequency decoupling according to claim 8, characterized in that: The regression loss function used in the model training is the CIoU loss function, and its calculation formula is as follows: In the formula, The ratio of the intersection area to the union area of ​​the predicted bounding box and the actual lesion bounding box; Representing the Euclidean distance from the center point, in breast imaging, this term can guide the model to quickly locate the core area of ​​the mass; It represents the diagonal length of the smallest bounding rectangle that can simultaneously contain both the predicted bounding box and the ground truth bounding box; This represents the aspect ratio penalty term, used to constrain the geometric consistency between the predicted bounding box and the ground truth bounding box.

10. A breast mass detection system based on axial modulation and frequency decoupling, characterized in that, include: The input unit is used to input the mammogram data to be detected. The model prediction unit incorporates the AMFD-RetinaNet single-stage breast mass detection network model as described in any one of claims 1 to 9; wherein the backbone network enhances the sensitivity of perceiving irregular small lesions in a hidden dense background by integrating the axial global-local modulation module AGLM; the feature pyramid network eliminates semantic aliasing caused by upsampling and suppresses background noise of dense glands by integrating the hybrid feature extraction module HFEM; and the detection head network achieves spatial alignment of classification and regression features by integrating the instance-specific asymmetric decoupling head ISADH. The output unit is used to parse the output of the model prediction unit, generate and display the category confidence score and high-precision bounding box coordinates of the target breast mass on the original image.

Images