A breast X-ray lesion segmentation method, system, device and storage medium

Abstract

The application discloses a breast X-ray lesion segmentation method, system, device and storage medium, which acquires images of the same side breast CC position and MLO position, extracts anatomical landmarks such as breast contour, nipple position and pectoralis major reference line to construct a normalized anatomical coordinate system across views, generates initial lesion masks of the CC position and MLO position by using an initial segmentation model, then performs coordinate parameter mapping and anatomical coding on the initial lesion masks, then identifies and divides the lesion area through bidirectional matching verification, and finally outputs the final lesion segmentation result after correcting the lesion masks of different areas according to the analysis result. The application is not simply double-view lesion detection or quadrant judgment, but implements cross-view consistency correction for pixel-level segmentation masks, which can effectively improve the boundary continuity, stability and accuracy of breast X-ray lesion segmentation, and reduce the risk of misclassification and missed classification under single-view conditions.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence-assisted system technology, specifically relating to a method, system, device and storage medium for segmenting breast X-ray lesions. Background Technology

[0002] Mammography is a crucial imaging tool for breast cancer screening and auxiliary diagnosis. However, due to the complex layered structure of breast tissue, overlapping and occlusion can occur between different tissues during projection. Furthermore, factors such as imaging noise, contrast variations, and blurred lesion boundaries make accurate segmentation of lesion areas on a single-position image challenging, easily leading to problems like discontinuous segmentation boundaries, localized expansion, false positives, and missed lesions. Standard clinical examination procedures typically involve acquiring both craniocaudal view (CC) and medial oblique view (MLO) images of the ipsilateral breast. This allows for observation of breast tissue structure and suspicious lesions from different projection angles, providing physicians with more comprehensive diagnostic information. Therefore, effectively integrating information from both positions to improve the accuracy of lesion analysis has become an important research direction in intelligent breast imaging analysis.

[0003] Current research on computer-aided analysis of mammograms using dual-view information fusion primarily focuses on improving the reliability of lesion identification by mining the correlation information between CC and MLO images. For example, a candidate solution region matching strategy is employed, extracting suspected lesion regions in both views and using geometric position constraints or feature similarity for correlation matching to reduce the false alarm rate in a single view. Alternatively, a multi-view feature fusion framework based on deep convolutional neural networks is used, extracting global features from CC and MLO positions in parallel through a dual-branch network and cascading them in fully connected layers or feature layers to achieve joint classification or quadrant discrimination based on dual views. Furthermore, some studies attempt to establish simple dual-view spatial correspondence models, utilizing the consistency of radial distance from the nipple to the lesion to help pinpoint the approximate search range of the lesion in different views. However, most of the above methods focus on lesion detection, quadrant judgment, and coarse-grained candidate region matching, lacking a mechanism to construct a unified reference coordinate system using key anatomical landmarks such as nipple position, chest wall reference line, and MLO pectoralis major muscle boundary. This leads to inconsistencies in the spatial location and boundary morphology of lesion segmentation results in different views. Moreover, due to the significant differences in projection methods and tissue deformation between two-position images, traditional rigid registration or simple feature stitching cannot accurately characterize the nonlinear correspondence between the two, thus limiting the effective use of two-position information in fine segmentation tasks. Summary of the Invention

[0004] To address the aforementioned shortcomings, this invention discloses a method, system, device, and storage medium for segmenting breast X-ray lesions. By utilizing cross-view anatomical consistency information, the initial lesion mask is corrected posteriorly, thereby improving the clinical reliability of the segmentation results and solving the problems of unstable boundaries, numerous false positives, and insufficient utilization of dual-position information in existing single-view segmentation methods.

[0005] This invention is achieved using the following technical solution:

[0006] A method for segmenting breast lesions via X-ray, comprising the following steps:

[0007] S1. Acquire CC and MLO images of the ipsilateral breast. After grayscale normalization of the original images, perform anatomical region localization and extraction to obtain the ROI image of the breast. Then, perform denoising and enhancement on the obtained ROI image of the breast to obtain the CC denoised and enhanced image and the MLO denoised and enhanced image.

[0008] S2. Based on the CC-position denoised and enhanced image and the MLO-position denoised and enhanced image obtained in step S1, extract anatomical landmark information respectively, and simultaneously obtain the corresponding lesion candidate regions to obtain the initial lesion mask of the CC position and the initial lesion mask of the MLO position. The anatomical landmark information includes the breast contour, nipple position and pectoralis major muscle reference line.

[0009] S3. Based on the anatomical landmark information obtained in step S2, construct an anatomical coordinate system framework, and then map the initial lesion mask at CC position and the initial lesion mask at MLO position to the anatomical coordinate system framework to finally generate the anatomical coding descriptor.

[0010] S4. Based on the anatomical coding descriptor and the CC-position denoised enhanced image and MLO-position denoised enhanced image obtained in step S1, extract the CC-position lesion descriptor and MLO-position lesion descriptor respectively, and calculate the anatomical equivalent position deviation between the lesions in the two views and the similarity of their local feature vectors. Based on the calculation results, refine the lesion region into consistent support region, boundary conflict region, single-view isolated region and low confidence boundary region, and perform region identification and division.

[0011] S5. Based on the classification results obtained in step S4, implement differentiated correction strategies for different types of regions. After completing the correction, construct and output multidimensional characterization information of the lesions to achieve a standardized description of the nature of the lesions. The multidimensional characterization information includes anatomical location encoding based on normalized geodesic distance and polar angle, multi-scale Gaussian feature vectors, normalized confidence scores generated based on anatomical equivalent location deviation and local feature vector similarity, and the final lesion segmentation mask.

[0012] Furthermore, in step S1, a grayscale normalization method based on Min-Max mapping is used to process the CC and MLO images of the ipsilateral breast in both positions, linearly mapping the grayscale values ​​of the original images to a uniform grayscale range to generate a normalized image I. norm (x,y); Using an automatic thresholding optimization method based on the image grayscale histogram, the global background suppression threshold T* is found by maximizing the variance between the background and tissue classes in the X-ray image:

[0013] The function `arg max(·)` represents the threshold parameter that maximizes the objective function within the grayscale region. `ω0(T)` represents the probability that pixels with grayscale values ​​less than the threshold `T` occupy the entire image. `ω1(T)` represents the probability that pixels with grayscale values ​​greater than or equal to the threshold `T` occupy the entire image. `μ0(T)` and `μ1(T)` represent the average grayscale values ​​of background pixels and tissue pixels, respectively. The grayscale-normalized image is binarized using an adaptive threshold `T*` to identify candidate tissue regions `M`. initial :

[0014] The morphological opening operation γ is used to break up small interfering connections, smoothing the tissue contour, and the final breast anatomical mask is determined by the maximum connected component operator LCC(·).

[0015] The final breast ROI image is obtained using pixel-by-pixel multiplication.

[0016] Because mammograms have extremely high bit depth and different exposure doses can cause huge differences in image brightness, this invention uses a Min-Max mapping grayscale normalization method. To eliminate non-diagnostic redundant information, this invention achieves anatomical region localization through global background suppression and accurately extracts anatomical regions through morphological opening operations and connected component analysis.

[0017] Then, the processed I is processed using binary wavelet transform. roi The image is decomposed into a multi-resolution representation of N levels, each level consisting of gradient components and approximate components; to suppress quantum noise while preserving edge information, a wavelet shrinkage operator C(x) is applied to the gradient components:

[0018] Where sgn(·) is the sign function, used to characterize the sign of the image gradient component coefficients, x represents the image gradient component coefficient, and H n The adaptive shrinkage threshold for the nth level is represented by the nonlinear mapping function E(x'), which is used to perform local gain control on the processed coefficients.

[0019] Where x' represents the denoised multi-scale feature coefficients, K represents the gain coefficient (K>1), and A represents the enhancement threshold. The processed scale coefficients are synthesized using inverse binary wavelet transform to obtain CC-bit denoised and enhanced images and MLO-bit denoised and enhanced images. This invention employs a multi-scale analysis method based on binary wavelet transform to enhance the local contrast of small lesions, targeting quantum noise commonly found in mammography.

[0020] Furthermore, in step S2, a feature extraction network is constructed based on the standard ResNet model. The first convolutional layer has 16 channels, and the architecture includes 5 ResNet blocks. Two fully connected layers are connected after the global average pooling layer. The model is trained using progressive weak supervision. First, the small-scale public dataset INbreast with pixel-level annotations is used for precise annotation and strong supervision, so that the model can initially grasp the basic geometric shape of the nipple, chest wall, and lesions. Then, on the large-scale BI-BADS labeled dataset, the image-level classification labels are converted into pixel-level probability heatmaps using Class Activation Mapping (CAM) technology, thereby solving the problem of the lack of precise breast cancer annotation datasets.

[0021] Furthermore, during training, a joint loss function is constructed. Physical constraints are applied to the prediction results to ensure that the model's predictions conform to human anatomical characteristics:

[0022] ,in For classification loss terms; For heatmap regression loss term, it represents the mean square error between the predicted heatmap and the actual anatomical landmarks or lesion center points; As an anatomical boundary constraint, a high penalty value is applied to predicted points that fall outside the tissue imaging area; As a preliminary consistency constraint across views, it represents the difference in normalized distance between candidate target points of the same side CC position and MLO position relative to the nipple, injecting geometric correlation awareness across body positions into the model;

[0023] During the anatomical landmark information extraction process, the denoised and enhanced image is input into the anatomical landmark regression branch. In the output K-channel heatmap, the coordinates of the pixel with the largest response value are found as the initial localization point.

[0024] Where H represents the probability distribution heatmap of the corresponding anatomical landmark, and to eliminate the positioning error caused by the discrete pixel grid, the response point and its neighboring pixel values ​​are used for center weighting to obtain sub-pixel level accurate coordinates:

[0025] Where Ω represents the local neighborhood centered at the peak point, and ω i,j p is the heatmap response value of the pixel at coordinate (i, j). i,j This represents the physical coordinates of the corresponding pixel. By geometrically mapping the specific points of a particular anatomical channel, the linear and contour references can be locked.

[0026] Simultaneously, using 256×256 pixels as the standard window, sliding window sampling with overlap is performed on the denoised and enhanced image. Each sampling window is input into the lesion extraction branch, and the probability value of the local region belonging to the candidate lesion category is calculated. Then, the local response results are stitched together to restore the lesion probability distribution map M at the full image size. prob After obtaining the probability distribution of the entire map, the probability map is binarized using an adaptive threshold operator to generate the initial lesion mask M. init :

[0027] , where τ is the threshold operator.

[0028] Furthermore, in step S3, an anatomical coordinate system framework is constructed based on the anatomical landmark information obtained in step S2. Specifically, the nipple point A is used as the local coordinate origin, and the pectoralis major reference line BC and the corresponding breast tissue boundary curve are obtained. Then, the original coordinates are transformed to the standard reference frame x' using the rotation matrix R and the translation variable t.

[0029] Where R is based on the tangent angle at the nipple boundary. With the direction angle of the parabola The rotation matrix is ​​expressed as:

[0030] ;

[0031] Next, using a family of second-order parabolas as coordinate lines, the tangent angle of the parabola at vertex A is defined as the polar angle parameter. The intersection point D of the intersection point D with the pectoralis major line BC is uniquely determined, and the position of the intersection point D on the line segment BC satisfies the following linear ratio constraint:

[0032] Furthermore, a radial distance parameter s is introduced, which is defined as the relative geodesic distance from nipple point A to the target point (x, y) along a specified parabolic trajectory. Its calculation formula is as follows:

[0033] , where sabs(ζ) represents the geodesic arc length function calculated along the parabolic trajectory, when s=0 corresponds to the nipple position, and when s=1 corresponds to the edge position of the pectoralis major muscle;

[0034] The final mapping function is: Then, the initial lesion masks at the CC and MLO positions are mapped onto the aforementioned anatomical coordinate system framework. Specifically, this involves mapping any lesion center point P detected in the original image coordinate system. i (x i , y i The normalized anatomical coordinates in the parameter space Ω are calculated using the established nonlinear mapping operator. Simultaneously, through the Jacobian matrix J of the mapping function T-1 Calculate the local unit vectors v1 and v2 aligned with the anatomical structures:

[0035] Where j1 and j2 are the Jacobian matrices J T-1 The column vectors are then used to extract the feature vector V from the original image based on this alignment direction. feat The final anatomical coding descriptor is generated as D=[V loc V feat ].

[0036] Furthermore, in step S4, lesion descriptors D are extracted based on the anatomical coding descriptor and the CC-position denoised enhanced image and MLO-position denoised enhanced image, respectively. CC =[V loc-CC V feat-CC ] with lesion descriptor D MLO =[V loc-MLO V feat-MLO ], and calculate the anatomical equivalent positional deviation of the lesion between the two views:

[0037] And calculate the similarity of their local feature vectors:

[0038] Δs is used to characterize the normalized geodesic distance deviation of candidate lesions along the direction from the nipple to the chest wall. The Corr(·) function is used to characterize the deviation of candidate lesions in anatomical orientation angles. It is used to detect the correlation between Gaussian derivative eigenvectors. Based on the calculated results, the lesion region is further subdivided into consistent support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions. Region identification is then performed, with the specific identification and subdivision decisions as follows:

[0039] (41) First, the anatomical location of the candidate lesion area is determined by boundary risk. If the normalized geodesic distance parameter s of the candidate area is close to 0 or 1, or the distance between the candidate area and the nipple, pectoralis major muscle reference line or breast tissue outline is less than the preset boundary tolerance threshold, then the candidate area is determined to be in the anatomical boundary sensitive area and classified as a low confidence boundary area.

[0040] (42) For candidate lesions that are not classified as low-confidence boundary regions, the coordinates of the candidate lesion region in the normalized anatomical parameter space are used. Based on this, the anatomical equivalent positional deviation between it and the candidate region from another viewpoint is calculated respectively:

[0041] Δs=|s CC -s MLO |and If there is no image in another viewpoint that simultaneously satisfies Δs≤τ s and If a candidate region is not found, it is determined that the candidate region lacks cross-view support and is classified as a single-view isolated region.

[0042] (43) If there exists a view in another perspective that satisfies Δs≤τ s and If the candidate regions are selected, the candidate region with the smallest anatomically equivalent positional deviation is chosen as the priority matching object, and the local feature vector similarity between the two is further compared; if the local feature vector similarity satisfies:

[0043] If the candidate regions in the two perspectives are consistent in both spatial location and local morphological features, they are classified as consistent support regions.

[0044] (44) If the preferred matching object satisfies Δs≤τ s and However, the cosine similarity between local feature vectors satisfies:

[0045] If the two are in similar anatomically equivalent positions, but their boundary morphology, texture response, or local structural features are inconsistent, they are classified as boundary conflict areas.

[0046] Furthermore, in step S5, based on the classification results obtained in step S4, a differentiated correction strategy is implemented for different types of regions to improve the structural consistency and anatomical rationality of the segmentation results, specifically:

[0047] (51) For the consistent support region, the edge probability distribution of CC and MLO views is fused by the logistic regression operator to correct the boundary underfit caused by single-view projection occlusion and output an accurate segmentation mask with cross-view space consistency.

[0048] (52) For the boundary conflict area, use the Jacobian matrix J calculated in step S3. T-1 Nonlinear inverse transformation is performed on local pixels of the lesion to compensate for the shear deformation caused by breast compression, and the deformed contour is restored to the original aggregate shape under the standard anatomical framework.

[0049] (53) For isolated areas in a single view, use a mapping function:

[0050] The normalized anatomical coordinates of the lesion in another view are determined. By analyzing the local signal-to-noise ratio and the reciprocal Gaussian response value of the location, classification correction is performed. If the signal-to-noise ratio of the corresponding location is extremely low due to the high density of glands, the lesion is judged to be possibly obscured by tissue in another view and marked as a potential lesion. If the image of the corresponding location is clear and there is no anatomical feature disturbance, it is judged as a single-view artifact and removed.

[0051] (54) For low confidence boundary regions, a rigid constraint operator based on the anatomical boundary curve and the pectoralis major line BC is introduced to mechanically guide and correct the lesion contour close to the anatomical edge, so as to prevent the segmentation result from overflowing the anatomical boundary.

[0052] After correction, multidimensional characterization information of the lesion is constructed and output to achieve a standardized description of the lesion's properties. This multidimensional characterization information includes information based on normalized geodesic distance s and polar angle. Anatomical location encoding, multi-scale Gaussian feature vector V feat The system generates a normalized confidence score based on anatomical equivalent positional deviation and local feature vector similarity, as well as a final lesion segmentation mask.

[0053] This invention addresses the characteristics of breast imaging, including high resolution, small lesions, and large anatomical background areas. Based on the standard ResNet model, it designs a high-resolution preservation and dual-position feature interaction strategy, further constructing a spatially-aware customized residual network, named Spatial-Aware Customized Residual Network (SA-ResNet). To avoid the loss of small lesion features due to rapid downsampling in the early stages of the network, the number of channels in the first convolutional layer is reduced to 16, sacrificing initial channel capacity to preserve the high resolution of the original image. Simultaneously, to compensate for the limited feature expressiveness caused by the reduced number of channels in the first layer, the architecture includes 5 sets of ResNet blocks. Furthermore, after the global average pooling layer, two fully connected layers are added, allowing nonlinear interaction between CC and MLO features of the ipsilateral breast in depth space, thereby extracting semantic features with positional correlation. In addition, considering the ultra-high resolution characteristics of breast imaging, this invention employs a global semantic-guided local scanning strategy to avoid the loss of small lesion features during global downsampling.

[0054] A breast X-ray lesion segmentation system, applied to the aforementioned breast X-ray lesion segmentation method, specifically includes the following modules:

[0055] The dual-position image preprocessing module is used to acquire CC and MLO images of the ipsilateral breast and perform preprocessing, including grayscale normalization, anatomical region localization and extraction, and image denoising and enhancement.

[0056] The anatomical landmark information and lesion co-extraction module is used to extract anatomical landmark information from the CC-bit image and MLO-bit image respectively, and simultaneously obtain the corresponding lesion candidate region. The anatomical landmark information covers the breast contour, nipple position and pectoralis major muscle reference line.

[0057] The anatomical mapping and coding module is used to establish a normalized anatomical coordinate system based on the extracted anatomical landmark information, map the lesion candidate region to a unified anatomical parameter space, and finally generate an anatomical coding descriptor to achieve cross-view spatial alignment and anatomical consistency coding.

[0058] The consistency analysis module is used to perform bidirectional consistency analysis in the normalized anatomical coordinate system to identify and delineate consistency support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions.

[0059] The correction output module is used to perform differential correction operations on different categories of regions based on the identification and segmentation results, and output the final segmentation results and multidimensional lesion representation information. The differential correction operations include boundary smoothing, conflict region morphology correction, isolated region suppression, and boundary region constraint completion. The multidimensional lesion representation information includes anatomical location encoding, feature vector, normalized confidence score, and the final lesion segmentation mask.

[0060] This invention standardizes image quality through a dual-position image preprocessing module, effectively reducing imaging noise and eliminating redundant background features, thus laying a high-quality data foundation for subsequent processing. It simultaneously locks key anatomical benchmarks and candidate lesion regions using an anatomical landmark information and lesion co-extraction module, and establishes a unified coordinate system based on human anatomical landmarks by an anatomical mapping and coding module. By mapping lesions to a unified anatomical parameter space, it solves the problem of matching nonlinear deformation of the breast under different compression angles. A consistency analysis module performs bidirectional consistency analysis in the parameter space to identify and segment candidate regions, and a correction output module performs differential correction. This not only utilizes complementary information from both views to complete blurred boundaries but also effectively eliminates glandular superposition artifacts from a single-view perspective, thereby significantly improving the precision of lesion segmentation and diagnostic reliability.

[0061] Furthermore, the anatomical mapping encoding module is configured with a nonlinear mapping operator. This nonlinear mapping operator maps candidate lesion regions in physical space to coordinates relative to the geodesic distance s and the polar angle by establishing a family of second-order parabolic coordinate lines. The normalized parameter space serves as the baseline to eliminate spatial ambiguity caused by differences in breast size and shape. The consistency analysis module employs bidirectional matching verification logic, which performs collaborative verification from two dimensions: spatial location consistency and local morphological feature consistency, based on the correlation between positional difference parameters in normalized anatomical coordinates and Gaussian derivative eigenvectors. The correction output module is equipped with a Jacobian matrix compensation unit, which uses the Jacobian matrix of the mapping function to calculate the local deformation tensor, performs nonlinear inverse transformation on boundary conflict areas, and compensates for shear deformation of the breast caused by body position compression, thus correcting the deformity. The altered lesion contour is restored to its original geometric shape within the standard anatomical framework. The correction output module performs implicit reprojection verification on isolated regions in a single view, uses a mapping function to determine the corresponding anatomical coordinates of the isolated region in another view, and combines local signal-to-noise ratio analysis to determine whether the region belongs to a potential lesion or imaging artifact caused by tissue occlusion, thereby achieving logical suppression of false positives. The correction output module introduces rigid constraint operators for low-confidence boundary regions, uses the breast contour curve and pectoralis major muscle reference line extracted in preprocessing as geometric boundary constraints, and performs mechanical guidance correction on lesion contours close to the anatomical edge to prevent the segmentation mask from overflowing the effective area of ​​biological tissue.

[0062] A device for segmenting breast lesions via X-ray includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method for segmenting breast lesions via X-ray. By integrating the memory and processor into the electronic device and pre-setting the computer program for implementing the aforementioned method for segmenting breast lesions via X-ray in the memory, the electronic device can automatically complete the acquisition, preprocessing, anatomical landmark extraction, initial lesion mask generation, cross-view space normalization alignment, pixel-level consistency analysis and correction, and final lesion segmentation result output when the processor executes the program. This device can encapsulate a dual-position geometric consistency constraint algorithm within itself, which facilitates direct embedding into a mammography system to achieve real-time, high-precision automatic segmentation of breast lesions, assisting radiologists in quickly locating lesions. Furthermore, it reduces subjective bias and time costs associated with manually outlining lesions, improving the standardization and efficiency of breast disease diagnosis.

[0063] A computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the aforementioned method for segmenting breast X-ray lesions. By storing the computer program implementing the aforementioned method for segmenting breast X-ray lesions on the computer-readable storage medium, it can be loaded and executed by different types of processors, thereby uniformly implementing the function of segmenting breast X-ray lesions on multiple hardware platforms. This facilitates modular encapsulation and cross-platform migration of the algorithm, ensures consistency of diagnostic logic in various deployment environments, shortens the system development and upgrade cycle, and enhances the versatility and engineering application value of the method for segmenting breast X-ray lesions.

[0064] Compared with existing technologies, this technical solution has the following advantages:

[0065] This invention does not perform simple lesion detection or quadrant localization for dual-view applications, but rather performs cross-view detection for pixel-level lesion masks. Figure 1 This invention employs a consistent correction method, focusing more on optimizing segmentation results. A normalized parameter space is constructed using key anatomical landmarks such as the nipple and pectoralis major muscle, and nonlinear deformation compensation is performed using a Jacobian matrix, effectively solving the problem of lesion alignment under different compression positions of the breast. Simultaneously, the anatomical correspondence between the ipsilateral CC and MLO positions is utilized to effectively reduce false-positive areas caused by missegmentation in single views, and it can improve problems such as blurred lesion boundaries, local defects, and boundary expansion, enhancing the continuity and stability of the segmentation boundaries. Furthermore, the method described in this invention is flexibly compatible with different types of initial segmentation models, facilitating engineering implementation and clinical deployment. Attached Figure Description

[0066] Figure 1 This is a flowchart of the breast X-ray lesion segmentation method described in Example 1.

[0067] Figure 2 This is a flowchart of the process for step S4, lesion area identification, and step S5, differential correction in the method described in Example 1.

[0068] Figure 3 This is a schematic diagram of the breast X-ray lesion segmentation system described in Example 2. Detailed Implementation

[0069] The present invention is further illustrated by the following examples, but these are not intended to limit the invention. Specific experimental conditions and methods not specified in the following examples are generally conventional methods well known to those skilled in the art.

[0070] Example 1: As Figure 1 As shown, a method for segmenting breast lesions using X-rays includes the following steps:

[0071] S1. Acquire CC and MLO images of the ipsilateral breast in both positions. Using a Min-Max mapping grayscale normalization method, process the CC and MLO images of the ipsilateral breast separately, linearly mapping the original image grayscale values ​​to a uniform grayscale range to generate a normalized image I. norm (x,y); Using an automatic thresholding optimization method based on the image grayscale histogram, the global background suppression threshold T* is found by maximizing the variance between the background and tissue classes in the X-ray image:

[0072] The function `arg max(·)` represents the threshold parameter that maximizes the objective function within the grayscale region. `ω0(T)` represents the probability that pixels with grayscale values ​​less than the threshold `T` occupy the entire image. `ω1(T)` represents the probability that pixels with grayscale values ​​greater than or equal to the threshold `T` occupy the entire image. `μ0(T)` and `μ1(T)` represent the average grayscale values ​​of background pixels and tissue pixels, respectively. The grayscale-normalized image is binarized using an adaptive threshold `T*` to identify candidate tissue regions `M`. initial :

[0073] The morphological opening operation γ is used to break up small interfering connections, smoothing the tissue contour, and the final breast anatomical mask is determined by the maximum connected component operator LCC(·).

[0074] The final breast ROI image is obtained using pixel-by-pixel multiplication.

[0075] ;

[0076] Then, the processed I is processed using binary wavelet transform. roi The image is decomposed into a multi-resolution representation of N levels, each level consisting of gradient components and approximate components; to suppress quantum noise while preserving edge information, a wavelet shrinkage operator C(x) is applied to the gradient components:

[0077] Where sgn(·) is the sign function, used to characterize the sign of the image gradient component coefficients, x represents the image gradient component coefficient, and H n The adaptive shrinkage threshold for the nth level is represented by the nonlinear mapping function E(x'), which is used to perform local gain control on the processed coefficients.

[0078] , where x' represents the multi-scale feature coefficient after denoising, K represents the gain coefficient and K>1, and A represents the enhancement threshold. The processed scale coefficients are synthesized by inverse binary wavelet transform to obtain CC-bit denoised and enhanced images and MLO-bit denoised and enhanced images.

[0079] S2. Based on the CC-position denoised and enhanced images and MLO-position denoised and enhanced images obtained in step S1, anatomical landmark information is extracted, and corresponding lesion candidate regions are simultaneously obtained to obtain the initial lesion masks for the CC-position and MLO-position. The anatomical landmark information includes the breast contour, nipple position, and pectoralis major muscle reference line. Specifically, a feature extraction network is constructed based on the standard ResNet model, where the first convolutional layer has 16 channels, and there are 5 sets of ResNet blocks in the architecture. Two fully connected layers are connected after the global average pooling layer. The model adopts progressive weak supervision training. First, the small-scale public dataset INbreast with pixel-level annotations is used for precise annotation and strong supervision, so that the model initially grasps the basic geometric shape of the nipple, chest wall, and lesions. Then, on the large-scale BI-BADS labeled dataset, the image-level classification labels are converted into pixel-level probability heatmaps using Class Activation Mapping (CAM) technology. During the training process, a joint loss function is constructed. Physical constraints are applied to the prediction results to ensure that the model's predictions conform to human anatomical characteristics:

[0080] ,in For classification loss terms; For heatmap regression loss term, it represents the mean square error between the predicted heatmap and the actual anatomical landmarks or lesion center points; As an anatomical boundary constraint, a high penalty value is applied to predicted points that fall outside the tissue imaging area; As a preliminary consistency constraint across views, it represents the difference in normalized distance between candidate target points of the same side CC position and MLO position relative to the nipple, injecting geometric correlation awareness across body positions into the model;

[0081] During the anatomical landmark information extraction process, the denoised and enhanced image is input into the anatomical landmark regression branch. In the output K-channel heatmap, the coordinates of the pixel with the largest response value are found as the initial localization point.

[0082] Where H represents the probability distribution heatmap of the corresponding anatomical landmark, and to eliminate the positioning error caused by the discrete pixel grid, the response point and its neighboring pixel values ​​are used for center weighting to obtain sub-pixel level accurate coordinates:

[0083] Where Ω represents the local neighborhood centered at the peak point, and ω i,j p is the heatmap response value of the pixel at coordinate (i, j). i,jThis represents the physical coordinates of the corresponding pixel. By geometrically mapping the specific points of a particular anatomical channel, the linear and contour references can be locked.

[0084] Simultaneously, using 256×256 pixels as the standard window, sliding window sampling with overlap is performed on the denoised and enhanced image. Each sampling window is input into the lesion extraction branch, and the probability value of the local region belonging to the candidate lesion category is calculated. Then, the local response results are stitched together to restore the lesion probability distribution map M at the full image size. prob After obtaining the probability distribution of the entire map, the probability map is binarized using an adaptive threshold operator to generate the initial lesion mask M. init :

[0085] , where τ is the threshold operator;

[0086] S3. Construct an anatomical coordinate system frame based on the anatomical landmark information obtained in step S2. Specifically, use nipple point A as the local coordinate origin, obtain the pectoralis major muscle reference line BC and the corresponding breast tissue boundary curve, and then transform the original coordinates to the standard reference frame x' using the rotation matrix R and the translation variable t:

[0087] Where R is based on the tangent angle at the nipple boundary. With the direction angle of the parabola The rotation matrix is ​​expressed as:

[0088] ;

[0089] Next, using a family of second-order parabolas as coordinate lines, the tangent angle of the parabola at vertex A is defined as the polar angle parameter. The intersection point D of the intersection point D with the pectoralis major line BC is uniquely determined, and the position of the intersection point D on the line segment BC satisfies the following linear ratio constraint:

[0090] Furthermore, a radial distance parameter s is introduced, which is defined as the relative geodesic distance from nipple point A to the target point (x, y) along a specified parabolic trajectory. Its calculation formula is as follows:

[0091] , where sabs(ζ) represents the geodesic arc length function calculated along the parabolic trajectory, when s=0 corresponds to the nipple position, and when s=1 corresponds to the edge position of the pectoralis major muscle;

[0092] The final mapping function is: Then, the initial lesion masks at the CC and MLO positions are mapped onto the aforementioned anatomical coordinate system framework. Specifically, this involves mapping any lesion center point P detected in the original image coordinate system. i (xi , y i The normalized anatomical coordinates in the parameter space Ω are calculated using the established nonlinear mapping operator. Simultaneously, through the Jacobian matrix J of the mapping function T-1 Calculate the local unit vectors v1 and v2 aligned with the anatomical structures:

[0093] Where j1 and j2 are the Jacobian matrices J T-1 The column vectors are then used to extract the feature vector V from the original image based on this alignment direction. feat The final anatomical coding descriptor is generated as D=[V loc V feat ];

[0094] S4, such as Figure 2 As shown, lesion descriptors D are extracted based on the anatomical coding descriptor and the CC-bit denoised enhanced image and MLO-bit denoised enhanced image obtained in step S1. CC =[V loc-CC V feat-CC ] with lesion descriptor D MLO =[V loc-MLO V feat-MLO ], and calculate the anatomical equivalent positional deviation of the lesion between the two views:

[0095] And calculate the similarity of their local feature vectors:

[0096] Δs is used to characterize the normalized geodesic distance deviation of candidate lesions along the direction from the nipple to the chest wall. The Corr(·) function is used to characterize the deviation of candidate lesions in anatomical orientation angles. It is used to detect the correlation between Gaussian derivative eigenvectors. Based on the calculated results, the lesion region is further subdivided into consistent support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions. Region identification is then performed, with the specific identification and subdivision decisions as follows:

[0097] (41) First, the anatomical location of the candidate lesion area is determined by boundary risk. If the normalized geodesic distance parameter s of the candidate area is close to 0 or 1, or the distance between the candidate area and the nipple, pectoralis major muscle reference line or breast tissue outline is less than the preset boundary tolerance threshold, then the candidate area is determined to be in the anatomical boundary sensitive area and classified as a low confidence boundary area.

[0098] (42) For candidate lesions that are not classified as low-confidence boundary regions, the coordinates of the candidate lesion region in the normalized anatomical parameter space are used. Based on this, the anatomical equivalent positional deviation between it and the candidate region from another viewpoint is calculated respectively:

[0099] Δs=|s CC -s MLO |and If there is no image in another viewpoint that simultaneously satisfies Δs≤τ s and If a candidate region is not found, it is determined that the candidate region lacks cross-view support and is classified as a single-view isolated region.

[0100] (43) If there exists a view in another perspective that satisfies Δs≤τ s and If the candidate regions are selected, the candidate region with the smallest anatomically equivalent positional deviation is chosen as the priority matching object, and the local feature vector similarity between the two is further compared; if the local feature vector similarity satisfies:

[0101] If the candidate regions in the two perspectives are consistent in both spatial location and local morphological features, they are classified as consistent support regions.

[0102] (44) If the preferred matching object satisfies Δs≤τ s and However, the cosine similarity between local feature vectors satisfies:

[0103] If the two are in similar anatomically equivalent positions, but their boundary morphology, texture response, or local structural features are inconsistent, they are therefore classified as boundary conflict areas.

[0104] S5. Based on the classification results obtained in step S4, implement differentiated correction strategies for different types of regions to improve the structural consistency and anatomical rationality of the segmentation results, specifically:

[0105] (51) For the consistent support region, the edge probability distribution of CC and MLO views is fused by the logistic regression operator to correct the boundary underfit caused by single-view projection occlusion and output an accurate segmentation mask with cross-view space consistency.

[0106] (52) For the boundary conflict area, use the Jacobian matrix J calculated in step S3. T-1 Nonlinear inverse transformation is performed on local pixels of the lesion to compensate for the shear deformation caused by breast compression, and the deformed contour is restored to the original aggregate shape under the standard anatomical framework.

[0107] (53) For isolated areas in a single view, use a mapping function:

[0108] The normalized anatomical coordinates of the lesion in another view are determined. By analyzing the local signal-to-noise ratio and the reciprocal Gaussian response value of the location, classification correction is performed. If the signal-to-noise ratio of the corresponding location is extremely low due to the high density of glands, the lesion is judged to be possibly obscured by tissue in another view and marked as a potential lesion. If the image of the corresponding location is clear and there is no anatomical feature disturbance, it is judged as a single-view artifact and removed.

[0109] (54) For low confidence boundary regions, a rigid constraint operator based on the anatomical boundary curve and the pectoralis major line BC is introduced to mechanically guide and correct the lesion contour close to the anatomical edge, so as to prevent the segmentation result from overflowing the anatomical boundary.

[0110] After correction, multidimensional characterization information of the lesion is constructed and output to achieve a standardized description of the lesion's properties. This multidimensional characterization information includes information based on normalized geodesic distance s and polar angle. Anatomical location encoding, multi-scale Gaussian feature vector V feat The system generates a normalized confidence score based on anatomical equivalent positional deviation and local feature vector similarity, as well as a final lesion segmentation mask.

[0111] Example 2: A breast X-ray lesion segmentation system, applied to the breast X-ray lesion segmentation method described in Example 1, specifically includes the following modules:

[0112] The dual-position image preprocessing module is used to acquire CC and MLO images of the ipsilateral breast and perform preprocessing, including grayscale normalization, anatomical region localization and extraction, and image denoising and enhancement.

[0113] The anatomical landmark information and lesion co-extraction module is used to extract anatomical landmark information from the CC-bit image and MLO-bit image respectively, and simultaneously obtain the corresponding lesion candidate region. The anatomical landmark information covers the breast contour, nipple position and pectoralis major muscle reference line.

[0114] The anatomical mapping and coding module is used to establish a normalized anatomical coordinate system based on extracted anatomical landmark information, and map candidate lesion regions to a unified anatomical parameter space, ultimately generating anatomical coding descriptors to achieve spatial alignment and anatomical consistency coding across views. This module is equipped with a nonlinear mapping operator, which establishes a family of second-order parabolic coordinate lines to map candidate lesion regions in physical space to a coordinate system based on relative geodesic distance *s* and polar angle. A normalized parameter space based on breast size and shape is used to eliminate spatial ambiguity caused by differences in breast size and shape;

[0115] The consistency analysis module is used to perform bidirectional consistency analysis in the normalized anatomical coordinate system to identify and classify consistency support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions. The consistency analysis module adopts bidirectional matching verification logic, which performs collaborative verification from two dimensions: spatial location consistency of lesions across views and local morphological feature consistency, based on the correlation between positional difference parameters in the normalized anatomical coordinates and Gaussian derivative feature vectors.

[0116] The correction output module is used to perform differential correction operations on different categories of regions based on the identification and segmentation results, and output the final segmentation results and multidimensional lesion representation information; the differential correction operations include boundary smoothing, conflict region morphology correction, isolated region suppression, and boundary region constraint completion; the multidimensional lesion representation information includes anatomical location encoding, feature vector, normalized confidence score, and final lesion segmentation mask;

[0117] The correction output module is equipped with a Jacobi matrix compensation unit. This unit uses the Jacobi matrix of the mapping function to calculate the local deformation tensor, performs a nonlinear inverse transformation on the boundary conflict area, compensates for the shear deformation of the breast caused by body position compression, and restores the distorted lesion contour to the original geometric shape under the standard anatomical framework. The correction output module performs implicit reprojection verification on isolated areas in a single view, uses the mapping function to determine the corresponding anatomical coordinates of the isolated area in another view, and combines local signal-to-noise ratio analysis to determine whether the area belongs to a potential lesion or imaging artifact caused by tissue occlusion, thereby achieving logical suppression of false positives. The correction output module introduces a rigid constraint operator for low confidence boundary areas, uses the breast contour curve and pectoralis major muscle reference line extracted by preprocessing as geometric boundary constraints, and performs mechanical guidance correction on the lesion contour close to the anatomical edge to prevent the segmentation mask from overflowing the effective area of ​​biological tissue.

[0118] Example 3: A device for segmenting breast lesions using X-ray, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the breast lesion segmentation method described in Example 1.

[0119] Example 4: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the breast X-ray lesion segmentation method described in Example 1.

[0120] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for segmenting breast lesions using X-rays, characterized in that: Includes the following steps: S1. Acquire CC and MLO images of the ipsilateral breast. After grayscale normalization of the original images, perform anatomical region localization and extraction to obtain the ROI image of the breast. Then, perform denoising and enhancement on the obtained ROI image of the breast to obtain the CC denoised and enhanced image and the MLO denoised and enhanced image. S2. Based on the CC-position denoised and enhanced image and the MLO-position denoised and enhanced image obtained in step S1, extract anatomical landmark information respectively, and simultaneously obtain the corresponding lesion candidate regions to obtain the initial lesion mask of the CC position and the initial lesion mask of the MLO position. The anatomical landmark information includes the breast contour, nipple position and pectoralis major muscle reference line. S3. Based on the anatomical landmark information obtained in step S2, construct an anatomical coordinate system framework, and then map the initial lesion mask at CC position and the initial lesion mask at MLO position to the anatomical coordinate system framework to finally generate the anatomical coding descriptor. S4. Based on the anatomical coding descriptor and the CC-position denoised enhanced image and MLO-position denoised enhanced image obtained in step S1, extract the CC-position lesion descriptor and MLO-position lesion descriptor respectively, and calculate the anatomical equivalent position deviation between the lesions in the two views and the similarity of their local feature vectors. Based on the calculation results, refine the lesion region into consistent support region, boundary conflict region, single-view isolated region and low confidence boundary region, and perform region identification and division. S5. Based on the classification results obtained in step S4, implement differentiated correction strategies for different types of regions. After completing the correction, construct and output multidimensional characterization information of the lesions. The multidimensional characterization information includes anatomical location encoding based on normalized geodesic distance and polar angle, multi-scale Gaussian feature vectors, normalized confidence scores generated based on anatomical equivalent location deviation and local feature vector similarity, and the final lesion segmentation mask.

2. The method for segmenting breast lesions using X-rays according to claim 1, characterized in that: In step S1, a grayscale normalization method based on Min-Max mapping is used to process the CC and MLO images of the ipsilateral breast in both positions, linearly mapping the grayscale values ​​of the original images to a uniform grayscale range to generate a normalized image I. norm (x,y); The global background suppression threshold T* is found by maximizing the variance between the background and tissue classes in the X-ray image: The function `arg max(·)` represents the threshold parameter that maximizes the objective function within the grayscale region. `ω0(T)` represents the probability that pixels with grayscale values ​​less than the threshold `T` occupy the entire image. `ω1(T)` represents the probability that pixels with grayscale values ​​greater than or equal to the threshold `T` occupy the entire image. `μ0(T)` and `μ1(T)` represent the average grayscale values ​​of background pixels and tissue pixels, respectively. The grayscale-normalized image is binarized using an adaptive threshold `T*` to identify candidate tissue regions `M`. initial : The morphological opening operation γ is used to break up small interfering connections, smoothing the tissue contour, and the final breast anatomical mask is determined by the maximum connected component operator LCC(·). The final breast ROI image is obtained using pixel-by-pixel multiplication. ； Then, the processed I is processed using binary wavelet transform. roi The image is decomposed into a multi-resolution representation of N levels, each level consisting of gradient components and approximate components; the wavelet shrinkage operator C(x) is applied to the gradient components: Where sgn(·) is the sign function, used to characterize the sign of the image gradient component coefficients, x represents the image gradient component coefficient, and H n The adaptive shrinkage threshold for the nth level is represented by the nonlinear mapping function E(x'), which is used to perform local gain control on the processed coefficients. Where x' represents the multi-scale feature coefficients after denoising, K represents the gain coefficient and K>1, and A represents the enhancement threshold. The processed scale coefficients are synthesized by inverse binary wavelet transform to obtain CC-bit denoised and enhanced images and MLO-bit denoised and enhanced images.

3. The method for segmenting breast lesions using X-rays according to claim 1, characterized in that: In step S2, a feature extraction network is constructed based on the standard ResNet model. The first convolutional layer has 16 channels and five ResNet blocks are included in the architecture. Two fully connected layers are connected after the global average pooling layer. The model is trained with progressive weak supervision. First, the small-scale public dataset INbreast with pixel-level annotations is used for precise annotation and strong supervision, so that the model can initially grasp the basic geometric shape of the nipple, chest wall and lesions. Then, on the large-scale BI-BADS labeled dataset, the image-level classification labels are converted into pixel-level probability heatmaps using class activation mapping technology. Furthermore, during training, a joint loss function is constructed. Apply physical constraints to the prediction results: ,in For classification loss terms; For heatmap regression loss term, it represents the mean square error between the predicted heatmap and the actual anatomical landmarks or lesion center points; As an anatomical boundary constraint, a high penalty value is applied to predicted points that fall outside the tissue imaging area; For cross-view preliminary consistency constraints, it represents the difference in normalized distance between candidate target points of the same side CC position and MLO position relative to the nipple; During the anatomical landmark information extraction process, the denoised and enhanced image is input into the anatomical landmark regression branch. In the output K-channel heatmap, the coordinates of the pixel with the largest response value are found as the initial localization point. Where H represents the probability distribution heatmap of the corresponding anatomical marker, and sub-pixel-level precise coordinates are obtained by center-weighting the response points and their neighboring pixel values: Where Ω represents the local neighborhood centered at the peak point, and ω i,j p is the heatmap response value of the pixel at coordinate (i, j). i,j This represents the physical coordinates of the corresponding pixel. By geometrically mapping the specific points of a particular anatomical channel, the linear and contour references can be locked. Simultaneously, using 256×256 pixels as the standard window, sliding window sampling with overlap is performed on the denoised and enhanced image. Each sampling window is input into the lesion extraction branch, and the probability value of the local region belonging to the candidate lesion category is calculated. Then, the local response results are stitched together to restore the lesion probability distribution map M at the full image size. prob After obtaining the probability distribution of the entire map, the probability map is binarized using an adaptive threshold operator to generate the initial lesion mask M. init : , where τ is the threshold operator.

4. The method for segmenting breast lesions using X-rays according to claim 1, characterized in that: In step S3, an anatomical coordinate system framework is constructed based on the anatomical landmark information obtained in step S2. Specifically, nipple point A is used as the local coordinate origin, and the pectoralis major muscle reference line BC and the corresponding breast tissue boundary curve are obtained. Then, the original coordinates are transformed to the standard reference frame x' using the rotation matrix R and the translation variable t. Where R is based on the tangent angle at the nipple boundary. With the direction angle of the parabola The rotation matrix is ​​expressed as: ； Next, using a family of second-order parabolas as coordinate lines, the tangent angle of the parabola at vertex A is defined as the polar angle parameter. The intersection point D of the intersection point D with the pectoralis major line BC is uniquely determined, and the position of the intersection point D on the line segment BC satisfies the following linear ratio constraint: Furthermore, a radial distance parameter s is introduced, which is defined as the relative geodesic distance from nipple point A to the target point (x, y) along a specified parabolic trajectory. Its calculation formula is as follows: , where sabs(ζ) represents the geodesic arc length function calculated along the parabolic trajectory, when s=0 corresponds to the nipple position, and when s=1 corresponds to the edge position of the pectoralis major muscle; The final mapping function is: Then, the initial lesion masks at the CC and MLO positions are mapped onto the aforementioned anatomical coordinate system framework. Specifically, this involves mapping any lesion center point P detected in the original image coordinate system. i (x i , y i The normalized anatomical coordinates in the parameter space Ω are calculated using the established nonlinear mapping operator. Simultaneously, through the Jacobian matrix J of the mapping function T-1 Calculate the local unit vectors v1 and v2 aligned with the anatomical structures: Where j1 and j2 are the Jacobian matrices J T-1 The column vectors are then used to extract the feature vector V from the original image based on this alignment direction. feat The final anatomical coding descriptor is generated as D=[V loc V feat ].

5. The method for segmenting breast X-ray lesions according to claim 4, characterized in that: In step S4, based on the anatomical coding descriptor, lesion descriptors D are extracted from the CC-position denoised enhanced image and the MLO-position denoised enhanced image, respectively. CC =[V loc-CC V feat-CC ] with lesion descriptor D MLO =[V loc-MLO V feat-MLO ], and calculate the anatomical equivalent positional deviation of the lesion between the two views: And calculate the similarity of their local feature vectors: Δs is used to characterize the normalized geodesic distance deviation of candidate lesions along the direction from the nipple to the chest wall. The Corr(·) function is used to characterize the deviation of candidate lesions in anatomical orientation angles. It is used to detect the correlation between Gaussian derivative eigenvectors. Based on the calculated results, the lesion region is further subdivided into consistent support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions. Region identification is then performed, with the specific identification and subdivision decisions as follows: (41) First, the anatomical location of the candidate lesion area is determined by boundary risk. If the normalized geodesic distance parameter s of the candidate area is close to 0 or 1, or the distance between the candidate area and the nipple, pectoralis major muscle reference line or breast tissue outline is less than the preset boundary tolerance threshold, then the candidate area is determined to be in the anatomical boundary sensitive area and classified as a low confidence boundary area. (42) For candidate lesions that are not classified as low-confidence boundary regions, the coordinates of the candidate lesion region in the normalized anatomical parameter space are used. Based on this, the anatomical equivalent positional deviation between it and the candidate region from another viewpoint is calculated respectively: Δs=|s CC -s MLO | and ; If there is no image from another perspective that simultaneously satisfies Δs≤τ s and If a candidate region is not found, it is determined that the candidate region lacks cross-view support and is classified as a single-view isolated region. (43) If there exists a view in another perspective that satisfies Δs≤τ s and If the candidate regions are selected, the candidate region with the smallest anatomically equivalent positional deviation is chosen as the priority matching object, and the local feature vector similarity between the two is further compared; if the local feature vector similarity satisfies: If the candidate regions in the two perspectives are consistent in both spatial location and local morphological features, they are classified as consistent support regions. (44) If the preferred matching object satisfies Δs≤τ s and However, the cosine similarity between local feature vectors satisfies: If the two are in similar anatomically equivalent positions, but their boundary morphology, texture response, or local structural features are inconsistent, they are classified as boundary conflict areas.

6. The method for segmenting breast lesions by X-ray according to claim 5, characterized in that: In step S5, based on the classification results obtained in step S4, a differentiated correction strategy is implemented for different types of regions, specifically as follows: (51) For the consistent support region, the edge probability distribution of CC and MLO views is fused by the logistic regression operator to correct the boundary underfit caused by single-view projection occlusion and output an accurate segmentation mask with cross-view space consistency. (52) For the boundary conflict area, use the Jacobian matrix J calculated in step S3. T-1 Nonlinear inverse transformation is performed on local pixels of the lesion to compensate for the shear deformation caused by breast compression, and the deformed contour is restored to the original aggregate shape under the standard anatomical framework. (53) For isolated areas in a single view, use a mapping function: The normalized anatomical coordinates of the lesion in another view are determined. By analyzing the local signal-to-noise ratio and the reciprocal Gaussian response value of the location, classification correction is performed. If the signal-to-noise ratio of the corresponding location is extremely low due to the high density of glands, the lesion is judged to be possibly obscured by tissue in another view and marked as a potential lesion. If the image of the corresponding location is clear and there is no anatomical feature disturbance, it is judged as a single-view artifact and removed. (54) For low confidence boundary regions, a rigid constraint operator based on the anatomical boundary curve and the pectoralis major line BC is introduced to mechanically guide and correct the lesion contour close to the anatomical edge, so as to prevent the segmentation result from overflowing the anatomical boundary. After correction, multidimensional characterization information of the lesion is constructed and output. This multidimensional characterization information includes information based on normalized geodesic distance s and polar angle. Anatomical location encoding, multi-scale Gaussian feature vector V feat The system generates a normalized confidence score based on anatomical equivalent positional deviation and local feature vector similarity, as well as a final lesion segmentation mask.

7. A breast X-ray lesion segmentation system, applied to the breast X-ray lesion segmentation method of claim 6, characterized in that: Specifically, it includes the following modules: The dual-position image preprocessing module is used to acquire CC and MLO images of the ipsilateral breast and perform preprocessing, including grayscale normalization, anatomical region localization and extraction, and image denoising and enhancement. The anatomical landmark information and lesion co-extraction module is used to extract anatomical landmark information from the CC-bit image and MLO-bit image respectively, and simultaneously obtain the corresponding lesion candidate region. The anatomical landmark information covers the breast contour, nipple position and pectoralis major muscle reference line. The anatomical mapping and coding module is used to establish a normalized anatomical coordinate system based on the extracted anatomical landmark information, map the lesion candidate region to a unified anatomical parameter space, and finally generate an anatomical coding descriptor to achieve cross-view spatial alignment and anatomical consistency coding. The consistency analysis module is used to perform bidirectional consistency analysis in the normalized anatomical coordinate system to identify and delineate consistency support regions, boundary conflict regions, single-view isolated regions, and low-confidence boundary regions. The correction output module is used to perform differential correction operations on different categories of regions based on the identification and segmentation results, and output the final segmentation results and multidimensional lesion representation information. The differential correction operations include boundary smoothing, conflict region morphology correction, isolated region suppression, and boundary region constraint completion. The multidimensional lesion representation information includes anatomical location encoding, feature vector, normalized confidence score, and the final lesion segmentation mask.

8. The breast X-ray lesion segmentation system according to claim 7, characterized in that: The anatomical mapping encoding module is equipped with a nonlinear mapping operator. This nonlinear mapping operator maps candidate lesion regions in physical space to coordinates relative to the geodesic distance s and the polar angle by establishing a family of second-order parabolic coordinate lines. The normalized parameter space is used as the reference. The consistency analysis module adopts a two-way matching verification logic, which performs collaborative verification from two dimensions: spatial location consistency and local morphological feature consistency of lesions across views, based on the correlation between the positional difference parameters in normalized anatomical coordinates and the Gaussian derivative feature vectors. The correction output module is equipped with a Jacobi matrix compensation unit. The Jacobi matrix compensation unit uses the Jacobi matrix of the mapping function to calculate the local deformation tensor, performs nonlinear inverse transformation on the boundary conflict area, compensates for the shear deformation of the breast caused by body position compression, and restores the distorted lesion outline to the original geometric shape under the standard anatomical framework. The correction output module performs implicit reprojection verification on isolated regions in a single view, uses a mapping function to determine the corresponding anatomical coordinates of the isolated region in another view, and combines local signal-to-noise ratio analysis to determine that the region belongs to a potential lesion or imaging artifact caused by tissue occlusion. The correction output module introduces a rigid constraint operator for low-confidence boundary regions and uses the breast contour curve and pectoralis major muscle reference line extracted by preprocessing as geometric boundary constraints to mechanically guide and correct the lesion contour close to the anatomical edge.

9. A device for segmenting breast X-ray lesions, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the breast X-ray lesion segmentation method as described in claim 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the breast X-ray lesion segmentation method as described in claim 6.

Images