A computer-aided detection and diagnosis system for breast microcalcifications

Abstract

This application provides a computer-aided detection and diagnosis system for breast microcalcifications, comprising: an acquisition unit for acquiring at least two breast X-ray images of the same breast tissue from different projection angles as a target view group; a region detection unit for detecting suspected microcalcification candidate regions in each single view to obtain a candidate region set corresponding to each single view; based on the geometric projection relationship between different views in the target view group, performing cross-view pairing and association on candidate regions from different views to establish candidate region pairs characterizing the image appearance of the same potential microcalcification at different projection angles; an extraction module for extracting cross-view consistency features for each candidate region pair; a calculation unit for fusing the cross-view consistency features corresponding to each candidate region pair and its internal image features in each single view to form an enhanced feature vector of the potential microcalcification; and obtaining the benign or malignant classification result of the corresponding potential microcalcification.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical technology, specifically relating to a computer-aided detection and diagnosis system for breast microcalcifications. Background Technology

[0002] Mammography is the preferred method for early breast cancer screening. Microcalcifications are an important imaging marker for early breast cancer, especially ductal carcinoma in situ. However, microcalcifications are often small in size, have low contrast, and are easily confused with benign calcifications in breast tissue, vascular cross-sections, and image noise. This poses a significant challenge to radiologists' visual interpretation and can easily lead to missed or misdiagnosis.

[0003] To assist physicians in diagnosis, computer-aided detection and diagnosis (CAD) systems have been introduced into clinical workflows. Current mainstream CAD systems typically perform independent analysis of single mammogram images (such as cephalothorax (CC) or lateral oblique mammograms (MLO). Their workflow generally includes: detecting suspicious microcalcifications on a single view, extracting morphological and textural features of the region, and finally classifying the images based on these features to determine the probability of malignancy.

[0004] However, this single-view independent analysis approach has inherent limitations. Relying solely on information from a single perspective, the system struggles to effectively distinguish between genuine microcalcifications with three-dimensional spatial structures and "calcification-like" appearances caused by tissue overlap, image noise, or artifacts. This directly leads to two contradictory results: lowering the detection threshold to detect as many genuine lesions as possible introduces numerous false positives, increasing the unnecessary review burden on physicians; conversely, raising the detection threshold to improve specificity and reduce false positives easily misses early malignant calcifications that are subtle and atypical. Therefore, existing single-view CAD systems face a bottleneck in balancing sensitivity and specificity, limiting their clinical auxiliary value. Summary of the Invention

[0005] In view of the above-mentioned defects or deficiencies in the prior art, a computer-aided detection and diagnostic system for breast microcalcifications is provided, comprising: The acquisition unit is configured to acquire digital mammograms of the same breast tissue from at least two different projection angles, as a target view group. The area detection unit, electrically connected to the acquisition unit, is configured for: Image preprocessing and suspected microcalcification candidate region detection are performed on each single view in the target view group to obtain the candidate region set corresponding to each single view; Based on the geometric projection relationship between different views in the target view group, cross-view pairing and association are performed on candidate regions from different views to establish candidate region pairs that characterize the image performance of the same potential microcalcification foci under different projection angles. The extraction module, electrically connected to the region detection unit, is configured for: For each successfully paired candidate region pair, extract its cross-view... Figure 1 Consistency characteristics, the cross-view Figure 1 Consistency features include spatial projection error, morphological feature variability, or texture feature stability measures; The calculation unit, electrically connected to the extraction module, is configured for: Fuse each candidate region pair with the corresponding cross-view Figure 1 The homogeneous characteristics and their internal image features in each single view form the enhanced feature vector of the potential microcalcification foci; Based on the enhanced feature vector, the benign or malignant classification result of the corresponding potential microcalcification is obtained, along with the corresponding confidence score.

[0006] According to the technical solution provided in this application, the region detection unit is specifically configured for: For any candidate region in the first view, calculate its expected projection region in the second view based on the geometric projection relationship; Search for all candidate regions located within the expected projection area in the second view, and calculate the morphological similarity measure between each searched candidate region and the candidate region in the first view; Based on the weighted score of the morphological similarity measure and spatial projection error, the best matching object is determined from the candidate regions of the second view to form candidate region pairs; If the weighted score of the best matching object is lower than a preset threshold, it is determined that the candidate region in the first view failed to be successfully matched in the second view, and it is marked as a single-view suspicious point for independent processing.

[0007] According to the technical solution provided in this application, the extraction module is specifically configured for: Calculate at least one set of morphological parameters for the same candidate region in two views presented from different projection angles, the morphological parameters including area, perimeter, roundness, aspect ratio, or edge sharpness; Calculate the relative difference or standard deviation of each set of morphological parameters between the two views, as the variability of that set of parameters; The variability of each morphological parameter is normalized and weighted to obtain the morphological feature variability that characterizes the overall morphological stability of the candidate region.

[0008] According to the technical solution provided in this application, the geometric projection relationship is obtained based on an elastic deformation model; the region detection unit is specifically configured to: use the elastic deformation model to perform nonlinear deformation mapping on the position of the candidate region in the first view, so as to determine its corresponding expected search region in the second view.

[0009] According to the technical solution provided in this application, the elastic deformation model is constructed in the following manner: Multiple sets of paired, labeled, dual-view X-ray images of the breast were obtained as a training set; For each pair of images, a set of homologous anatomical landmarks is extracted, which includes the corresponding location points of the breast contour, the edge of the pectoralis major muscle, and the nipple. Based on all paired anatomical landmark sets in the training set, a deep learning network is trained through supervised learning, or a statistical deformation model is fitted through a group image registration algorithm, to obtain the elastic deformation model that can generalize to predict tissue deformation of different breasts from the first projection angle to the second projection angle.

[0010] According to the technical solution provided in this application, before determining the best matching object based on the weighted score, the region detection unit is further configured to: Determine whether a specific matching state exists. The specific matching state is: a pairing combination formed by a first candidate region in the first view and at least two different second candidate regions in the second view, the weighted scores of which are all higher than the preset threshold used to determine a successful pairing.

[0011] According to the technical solution provided in this application, the region detection unit is further configured to: If the specific matching state is determined to exist, the contextual attributes of the first candidate region in the first view are evaluated, and its cluster membership confidence as a member of the calcification cluster is calculated. If the cluster membership confidence is higher than the cluster determination threshold, then the cluster-aware matching mode is executed; The cluster-aware matching mode is executed as follows: In the first view, identify all other candidate regions that have been successfully paired within a preset neighborhood centered on the first candidate region, and form a confirmed cluster reference set; For each second candidate region in the second view that forms the specific matching state with the first candidate region, it is combined with all the paired objects in the second view of the confirmed cluster reference set to form multiple complete candidate clusters to be evaluated; Calculate the spatial distribution evaluation index for each of the complete candidate clusters to be evaluated; The second candidate region that optimizes the spatial distribution evaluation index is selected as the final pairing object with the first candidate region.

[0012] According to the technical solution provided in this application, the region detection unit is further configured to: If the cluster membership confidence is lower than the isolated point determination threshold, then a multi-dimensional re-evaluation mode is executed; The multi-dimensional re-evaluation mode is implemented as follows: Acquire additional image information associated with the target view group, including historical image sequences or dual-energy imaging data of the same breast tissue; Based on the additional image information, calculate the time stability measure or energy spectrum consistency measure of the pairing combination formed by the first candidate region and each conflicting second candidate region. The original weighted score is corrected using the time stability metric or energy spectrum consistency metric, and the final pairing is determined based on the corrected score.

[0013] According to the technical solution provided in this application, the region detection unit is specifically configured for: For candidate regions marked as suspicious points in a single view, the local image features of the suspicious points in the single view are analyzed. If the feature pattern matches the feature library of known typical image noise, artifacts or blood vessel sections with a degree higher than the first matching threshold, then the region is filtered out.

[0014] According to the technical solution provided in this application, the region detection unit is further configured to: For the virtual cross-view attached Figure 1 For suspicious points in a single view with consistent characteristics, perform differential classification processing, which includes: A dedicated classifier with a higher sensitivity threshold, distinct from conventional paired calcification classification models, is used for processing. And / or, In the enhanced feature vector, the virtual cross-view is reduced. Figure 1 Weights of consistency features; Specifically, suspicious points in a single view that have undergone differential classification processing and whose classification confidence is within a specific uncertainty range are assigned a specific suspicious type label in their output results.

[0015] Compared with existing technologies, the beneficial effects of this application are as follows: This invention acquires at least two different projection angle views of the same breast and performs cross-view pairing association. This means that a region suspected of being a microcalcification needs to find a spatially corresponding and morphologically consistent matching region in another view at a different angle to be confirmed as a candidate region pair. This mechanism filters out a large amount of noise and artifacts that appear incidentally in one view but have no corresponding structure in another view, thereby reducing false positives at the source and enhancing the reliability of the detection results. Moreover, in addition to pairing, it further extracts cross-view associations from successfully paired region pairs. Figure 1 Consistency features, such as spatial projection error and morphological feature variability, are new dimensions of information that cannot be obtained from single-view analysis. They quantify the stability of the same potential calcification under different projections. True calcifications with solid structures show high consistency in these features, while artifacts often show large random variations. Fusing these consistency features with traditional single-view internal features forms a more informative enhanced feature vector, providing a stronger basis for subsequent classifiers, thus helping to more accurately distinguish between benign and malignant microcalcifications. Thus, this application ensures the specificity of detection (reducing false positives) and improves diagnostic accuracy by utilizing richer and more stable multi-view fusion features. It effectively solves the dilemma of balancing sensitivity and specificity in single-view CAD systems, effectively controlling the false positive rate while maintaining high sensitivity, thereby providing radiologists with more clinically valuable auxiliary diagnostic opinions and improving the detection efficiency and diagnostic confidence of early breast cancer. Attached Figure Description

[0016] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A schematic diagram of the computer-aided detection and diagnosis system for breast microcalcifications provided in this application; The text labels in the image represent: 1. Acquisition unit; 2. Region detection unit; 3. Extraction module; 4. Calculation unit. Detailed Implementation

[0017] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] As mentioned in the background section, this application proposes a computer-aided detection and diagnosis system for breast microcalcifications, such as... Figure 1 As shown, it includes: Acquisition unit 1 is configured to acquire digital mammograms of the same breast tissue from at least two different projection angles, as a target view group; Area detection unit 2, electrically connected to the acquisition unit 1, is configured for: Image preprocessing and suspected microcalcification candidate region detection are performed on each single view in the target view group to obtain the candidate region set corresponding to each single view; Based on the geometric projection relationship between different views in the target view group, cross-view pairing and association are performed on candidate regions from different views to establish candidate region pairs that characterize the image performance of the same potential microcalcification foci under different projection angles. Extraction module 3, electrically connected to the region detection unit 2, is configured for: For each successfully paired candidate region pair, extract its cross-view... Figure 1 Consistency characteristics, the cross-view Figure 1 Consistency features include spatial projection error, morphological feature variability, or texture feature stability measures; Calculation unit 4, electrically connected to extraction module 3, is configured for: Fuse each candidate region pair with the corresponding cross-view Figure 1 The homogeneous characteristics and their internal image features in each single view form the enhanced feature vector of the potential microcalcification foci; Based on the enhanced feature vector, the benign or malignant classification result of the corresponding potential microcalcification is obtained, along with the corresponding confidence score.

[0020] Specifically, the target view set refers to the collection of digital X-ray images of the same breast acquired by acquisition unit 1 at at least two different projection angles. The most typical type is the clinically standard cephalothorax and lateral oblique view image pair. The candidate region set refers to the collection of image regions identified by region detection unit 2 on each single view for all suspected microcalcifications. Each region is typically defined by a set of pixel coordinates and its bounding box or segmentation mask. A candidate region pair specifically refers to a pair of candidate regions that, after cross-view pairing and association, are confirmed to represent the image representation of the same three-dimensional spatial point (i.e., potential microcalcification) on two different two-dimensional projection views. Figure 1Consistency features are a set of metrics extracted from successfully paired candidate region pairs to quantify their consistency across multiple viewpoints. The main types include spatial projection error, morphological feature variability, and texture feature stability measures. The enhanced feature vector is a comprehensive feature representation generated by computation unit 4, which fuses cross-view... Figure 1 The consistency features are composed of the image features within each single view (such as grayscale statistics and local texture features), which serve as direct inputs for subsequent classification.

[0021] Detailed description of the implementation steps: The system implementation begins with acquisition unit 1, which is typically integrated into the interface of the hospital's image archiving and communication system. This unit automatically retrieves or receives dual-view images from the same examination sequence from a digital mammography device and performs necessary format standardization. Region detection unit 2 then works independently for each image. Image preprocessing steps include adjusting window width and level to optimize contrast, applying Gaussian filtering or nonlocal mean filtering to reduce noise, and possibly performing background segmentation to focus on glandular tissue regions. Detection of suspected microcalcification candidate regions can be achieved using various mature computer vision or deep learning methods, such as speckle detection algorithms based on multi-scale Gaussian Laplacian filtering, or training a convolutional neural network for pixel-level segmentation. This results in a set of candidate regions containing coordinates and preliminary features for each view.

[0022] Next, the system performs the core cross-view pairing and association. This requires pre-calibrating or calculating the geometric projection relationship between the two views, which can be calculated using the position parameters of the image equipment or estimated through feature point matching of the image itself. For a candidate region in the first view (e.g., head-to-tail view), the projection relationship is used to calculate the theoretical range in the second view (e.g., oblique view). Then, all candidate regions in the second view are searched within this range. For each second view candidate region within this range, a morphological similarity metric is calculated between it and the first view candidate region. This metric can be calculated based on the cosine similarity of the region shape descriptor (e.g., Hu moments). The system integrates the spatial projection error (e.g., the Euclidean distance between the actual and theoretical positions) and the morphological similarity metric, linearly weighted by a pre-set weight, to obtain a comprehensive score. Finally, the candidate region with the highest comprehensive score in the second view is selected and paired with it to form a candidate region pair. If the highest score is lower than an empirically preset threshold, the pairing is considered a failure, and the region in the first view is marked as a single-view suspicious point, pending further special processing.

[0023] Extraction module 3 works for each successfully paired candidate region pair. Spatial projection error is directly calculated by determining the coordinate difference between the center points of the two paired regions after transformation through geometric projection. The calculation of morphological feature variability requires first calculating the morphological parameters (such as area, perimeter, etc.) of the region in both views, and then calculating the relative rate of change of these parameters between the two views. Texture feature stability measurement can be obtained by comparing the differences in the local binary pattern histograms or gray-level co-occurrence matrix features of the two regions. Calculation unit 4 receives these cross-view... Figure 1 Consistent features are extracted and concatenated or fused early with the internal features extracted from each region on its respective single view (e.g., the region's gray-scale mean, standard deviation, entropy, etc.) to form a higher-dimensional enhanced feature vector. Finally, this enhanced feature vector is input into a pre-trained classification model, such as a support vector machine or a deep neural network classifier. The classifier outputs a binary classification result of the potential microcalcification as malignant or benign, and also outputs a confidence score between 0 and 1, quantifying the reliability of the judgment.

[0024] This implementation aims to address the core problem of traditional single-view computer-aided detection systems, which suffer from a difficulty in simultaneously achieving both sensitivity and specificity due to the limited information dimension. Its primary technical effect is the significant improvement in the specificity of detection results (i.e., reducing false positives) through cross-view pairing verification. The underlying principle is that genuine microcalcifications with three-dimensional solid structures will inevitably produce corresponding image representations in two different projection views. Many image noises, artifacts, or tissue overlap structures that resemble calcifications in a single view often lack stable corresponding representations in the corresponding locations in another view, thus being naturally filtered out during the pairing process. Secondly, this scheme extracts and utilizes cross-view... Figure 1 The improved consistency significantly enhances the system's diagnostic ability to differentiate between benign and malignant microcalcifications. The underlying principle is that genuine calcifications, especially malignant calcifications, should maintain relatively stable physical properties (such as morphology and density) across different projection angles; therefore, their cross-viewing characteristics... Figure 1 Consistent features (such as low morphological variability and projection errors within a reasonable range) will exhibit specific patterns. However, some benign changes or artifacts may show greater inconsistencies across multiple views. Fusing these consistent features with single-view features creates an enhanced feature vector that is more comprehensive and discriminative, enabling subsequent classifiers to make more accurate judgments.

[0025] In a preferred embodiment, the region detection unit 2 is specifically configured to: For any candidate region in the first view, calculate its expected projection region in the second view based on the geometric projection relationship; Search for all candidate regions located within the expected projection area in the second view, and calculate the morphological similarity measure between each searched candidate region and the candidate region in the first view; Based on the weighted score of the morphological similarity measure and spatial projection error, the best matching object is determined from the candidate regions of the second view to form candidate region pairs; If the weighted score of the best matching object is lower than a preset threshold, it is determined that the candidate region in the first view failed to be successfully matched in the second view, and it is marked as a single-view suspicious point for independent processing.

[0026] Specifically, the first view and the second view are images from two different projection angles arbitrarily specified in the target view group. For ease of description, the view processed first is usually referred to as the first view. The expected projection area refers to a rectangular or circular search range centered on the mapped point after mapping the coordinates of the center point of a candidate region in the first view to the coordinate system of the second view according to the geometric projection relationship. The size of this range needs to take into account the systematic error of the projection calculation itself and the slight deformation of breast tissue. The morphological similarity measure is a scalar value used to quantify the degree of similarity in shape between two different image regions. In specific implementations, Fourier descriptor similarity based on region contours or graph matching scores based on region skeleton maps can be used. The weighted score is a comprehensive evaluation index, which is formed by linearly combining the morphological similarity measure with the spatial projection error (usually in pixels) according to preset coefficients. Its purpose is to balance the two key factors of "shape similarity" and "positional pairing". The preset threshold is an empirical or statistically determined threshold value used to finally determine whether the pairing is successful. If the score is higher than the threshold, the pairing is accepted; otherwise, it is rejected.

[0027] Detailed description of the implementation steps: During processing, region detection unit 2 first traverses the candidate region set of the first view. For each candidate region, the system calculates the theoretical coordinates of its center point in the second view based on the established geometric projection relationship (e.g., an affine transformation matrix). Then, with these theoretical coordinates as the center, a square with a side length of W pixels is defined as the expected projection region, where the value of W is set according to the image resolution and the prediction error.

[0028] Next, the system examines the candidate region set of the second view, filtering out all candidate regions whose center points fall within the expected projection area. These regions form a candidate list to be matched. If the list is empty, the candidate regions of the current first view are directly marked as single-view suspicious points. If the list is not empty, for each second view candidate region in the list, a morphological similarity measure is calculated between it and the current first view candidate region. For example, the Zernike moment feature vectors of the two regions can be extracted, and the cosine of their included angle can be calculated as the similarity.

[0029] Simultaneously, the spatial projection error is calculated, which is the Euclidean distance between the actual center point of the candidate region in the second view and the theoretical point calculated based on the projection relationship. Then, a weighted score is calculated. The weight of the morphological similarity measure is set as α, and the weight of the spatial projection error is set as β (usually α+β=1). The normalized morphological similarity is multiplied by α, and the normalized projection error is multiplied by β (because the error is a negative indicator) is subtracted to obtain a comprehensive score. After traversing all regions to be matched, the one with the highest comprehensive score is selected as the best matching object.

[0030] Finally, the system compares the overall score of the best match with a preset threshold T. If the score is greater than or equal to T, a candidate region pair is successfully established and the pairing is recorded. If the score is less than T, the pairing is deemed a failure. For all first-view candidate regions that fail to pair, the system uniformly marks them as single-view suspicious points and stores their metadata (such as coordinates and features) in a separate list so that subsequent processes can perform special processing to distinguish them from paired points, such as using a more stringent single-view feature classifier or directly prompting doctors to focus on reviewing them.

[0031] This implementation upgrades the pairing process from simple nearest-neighbor matching to a multi-factor comprehensive decision-making process, effectively reducing erroneous pairings caused by points that are accidentally close in location but have vastly different shapes, or points that are similar in shape but have significantly different projected positions. Its technical principle lies in simulating the cognitive process of a radiologist comparing two views: the doctor not only checks whether the positions of two points correspond but also judges whether their shapes are similar. This algorithm quantifies and fuses this dual judgment, controlling the strictness of the matching through a threshold. Ultimately, this refined pairing strategy provides a high-quality, highly reliable input data foundation for core multi-view analysis, fundamentally ensuring accurate subsequent feature extraction and classification.

[0032] In a preferred embodiment, the extraction module 3 is specifically configured to: Calculate at least one set of morphological parameters for the same candidate region in two views presented from different projection angles, the morphological parameters including area, perimeter, roundness, aspect ratio, or edge sharpness; Calculate the relative difference or standard deviation of each set of morphological parameters between the two views, as the variability of that set of parameters; The variability of each morphological parameter is normalized and weighted to obtain the morphological feature variability that characterizes the overall morphological stability of the candidate region.

[0033] Specifically, morphological parameters are quantitative indicators used to describe the shape, size, and boundary characteristics of an image region. These include area (the total number of pixels within the region), perimeter (the length of the line connecting the region's boundary pixels), circularity (measures how close the region is to a circle, calculated as 4π * area / perimeter squared), aspect ratio (the ratio of the longer side to the shorter side of the region's circumscribed rectangle), and edge sharpness (measured by calculating the average gradient magnitude of the region's boundary pixels). Relative variability refers to the percentage change obtained by dividing the difference in values ​​of the same morphological parameter across two views by the average value of the parameter in both views or the parameter value in the first view. Standard deviation, when calculating variability, refers to treating the values ​​of the same parameter across two views as a sample set and calculating its standard deviation. Normalization transforms the variability values ​​calculated for different morphological parameters into a uniform numerical range (e.g., [0, 1]) using methods such as min-max scaling or Z-score standardization. Weighted fusion refers to assigning different importance weights to the variability of different morphological parameters, and then obtaining a single, comprehensive morphological feature variability value through linear weighted summation.

[0034] Detailed description of the implementation steps: When calculating the morphological feature variability, extraction module 3 performs the following operations for each successfully paired candidate region pair. First, it calculates the same set of morphological parameters for both image regions of the calcification presented in the first and second views. For example, it fixes the calculation of three parameters: area, roundness, and edge sharpness.

[0035] Next, for each calculated morphological parameter, its difference between the two views is calculated. The system can employ one of two strategies: one is to calculate the relative difference, for example, for the area parameter, its relative difference = |area - view Figure 1 -Area_View2| / ((Area_View) Figure 1 + Area_View2) / 2). Secondly, the parameter values ​​of the two views are treated as a set containing two samples, and their standard deviation is calculated. Regardless of the strategy used, each parameter will receive an initial variability value characterizing its degree of variation between the two views.

[0036] Since different morphological parameters have different dimensions and numerical ranges (e.g., area may be in the thousands, while circularity is between 0 and 1), direct combination is meaningless. Therefore, it is necessary to normalize the initial variability values ​​of all parameters. For example, min-max normalization can be used, based on the minimum and maximum values ​​of the variability of each parameter obtained from the training set, to map the variability of each parameter in the current instance to a range of 0 to 1.

[0037] Finally, a weighted fusion is performed. The system predefines a weight vector, for example, setting the weight of area variation to 0.3, roundness variation to 0.4, and edge sharpness variation to 0.3. The normalized variation values ​​of each parameter are multiplied by their corresponding weights and then summed, i.e.: Comprehensive morphological feature variation = (Normalized area variation * 0.3) + (Normalized roundness variation * 0.4) + (Normalized edge sharpness variation * 0.3). This final scalar value represents the overall stability of the candidate region across views in terms of morphological features; the smaller the value, the more stable the shape.

[0038] The underlying principle is based on a key clinical observation: real, existing microcalcifications with fixed physical morphology, although they may show slight morphological variations on a two-dimensional projection due to different angles, exhibit limited and regular changes in their core morphological attributes (such as size and shape regularity). Conversely, transient artifacts formed by noise or tissue overlap may show randomness and significant inconsistency in their "morphology" under different viewpoints. By systematically calculating the cross-view variation of multiple morphological parameters and fusing them into a comprehensive score, this approach transforms this intuitive physician experience into a calculable and repeatable objective feature. This feature, when fed into a classifier, effectively helps the model learn the rule that "morphologically stable lesions are more likely to be true, while morphologically variable lesions are more likely to be false," thus directly contributing to reducing the system's false positive rate and improving the diagnostic specificity for malignant calcifications (whose morphology may be more irregular but maintains this irregularity under multiple viewpoints).

[0039] In a preferred embodiment, the geometric projection relationship is obtained based on an elastic deformation model; the region detection unit 2 is specifically configured to: use the elastic deformation model to perform nonlinear deformation mapping on the position of the candidate region in the first view, so as to determine its corresponding expected search region in the second view.

[0040] Specifically, the elastic deformation model is a mathematical model that describes the complex, nonlinear spatial transformations experienced by breast soft tissue during dual-view imaging. Unlike simple rigid or affine transformations, it simulates the localized stretching, compression, and shearing that occur when tissue is compressed. This model can be represented as a dense displacement field, where each point contains a displacement vector from the first view coordinates to the second view coordinates; or it can be a parameterized function, such as a transformation function based on B-splines or thin-plate splines. Nonlinear deformation mapping refers to the coordinate transformation process using the elastic deformation model. For an input point (such as the center of a candidate region in the first view), the output is not the result of a simple linear formula calculation, but rather the coordinates of the corresponding point, considering local tissue deformation, obtained by querying the displacement field or evaluating a complex transformation function. The expected search area is the range delineated in the second view based on the nonlinear deformation mapping result, used to find matching candidate regions. Because the mapping is nonlinear, this area may not be a regular rectangle; its shape and orientation will adaptively adjust with local deformation.

[0041] Detailed description of the implementation steps First, a suitable elastic deformation model needs to be built or obtained before system deployment. During actual operation, when the region detection unit 2 performs pairing, for each candidate region in the first view, it no longer uses a simple linear transformation to calculate the expected position, but instead calls this preloaded elastic deformation model.

[0042] The specific operation steps are as follows: Let the coordinates of the center point of a candidate region in the first view be P1(x1, y1). Input this coordinate P1 into the elastic deformation model. If the model is in the form of a displacement field, the system will find the displacement vector (dx, dy) corresponding to the position P1 in the displacement field, and then calculate the expected point P2' = (x1+dx, y1+dy). If the model is a parametric function F, then directly calculate P2' = F(x1, y1). This point P2' is the most likely corresponding point predicted after fully considering the local tissue deformation caused by the different compression methods and angles of the breast from the first view to the second view.

[0043] Due to the inherent minor errors in model prediction and imaging, the system defines a circular region with a radius of R pixels, centered at point P2', as the final expected search area. Compared to the region obtained by the linear model, the center position P2' of this region is closer to the actual corresponding point, thus improving the accuracy of the search starting point. Subsequent steps, such as finding candidate regions for the second view within the search area and calculating morphological similarity metrics, are consistent with the process described in claim 2.

[0044] This implementation improves the accuracy and recall of cross-view candidate region matching, especially for microcalcifications located in areas prone to significant deformation, such as the breast margin and dense glandular areas. The underlying principle is that it employs an elastic deformation model that more closely resembles biomechanical properties to approximate actual tissue movement. In this way, the system can more accurately predict the location where a calcification should appear in another view, allowing the "expected search area" to more tightly encompass the actual matching point. This reduces false exclusion of matches (improved sensitivity), narrows the search range, and lowers the probability of introducing incorrect candidates (improved specificity), thus ensuring the effectiveness of the entire multi-view analysis process from a fundamental geometric perspective.

[0045] In a preferred embodiment, the elastic deformation model is constructed in the following manner: Multiple sets of paired, labeled, dual-view X-ray images of the breast were obtained as a training set; For each pair of images, a set of homologous anatomical landmarks is extracted, which includes the corresponding location points of the breast contour, the edge of the pectoralis major muscle, and the nipple. Based on all paired anatomical landmark sets in the training set, a deep learning network is trained through supervised learning, or a statistical deformation model is fitted through a group image registration algorithm, to obtain the elastic deformation model that can generalize to predict tissue deformation of different breasts from the first projection angle to the second projection angle.

[0046] Specifically, the training set is a collection of numerous data pairs, each containing rigorously registered and labeled X-ray images of the same breast taken in standard cephalothorax and lateral oblique views. Labeled key anatomical landmarks refer to locations on these images that are manually marked by experts or automatically marked using high-precision algorithms, possessing clear anatomical significance and corresponding one-to-one between the two views. Homologous anatomical landmark sets refer to two sets of landmark coordinate data extracted from a pair of images, representing the same anatomical structure. In this context, "deep learning network" specifically refers to a neural network structure used for regression tasks, such as variants of U-Net or dedicated spatial transformation networks, whose input is an image and output is the displacement vector field of each pixel to the other image. Group image registration algorithms are a class of algorithms that discover common deformation patterns by aligning multiple images. The statistical deformation model is its output. It uses methods such as principal component analysis to represent the deformation observed in the training set with a set of average deformation fields and main deformation patterns (i.e., feature vectors), thereby enabling the generation of specific deformation fields for new individuals by adjusting a small number of parameters.

[0047] Detailed description of the implementation steps: The construction of the elastic deformation model is an offline, data-driven training process. This process begins with data preparation. It requires obtaining a large number (e.g., thousands of pairs) of high-quality dual-view mammograms from collaborating medical institutions or public databases. Each pair of images must be precisely aligned with the examination ID and ensure they belong to the same breast from the same examination. Subsequently, key anatomical landmarks need to be labeled on these image pairs. This can be done manually by organizing radiologists using annotation tools, or it can be initially screened using automated detection algorithms followed by expert correction. The goal of annotation is to mark corresponding anatomical features on each image. These points typically include: specific inflection points on the breast contour (such as the apex of the inframammary fold, the apex of the glandular margin), a clear inflection point at the junction of the anterior border of the pectoralis major muscle and the posterior border of the breast, and the nipple center point. These points, due to their anatomical stability, serve as reliable anchor points for constructing the transformation relationship between the dual views.

[0048] After data labeling is completed, the model training phase begins. The system provides two parallel technical approaches.

[0049] Path 1: Deep Learning Network Based on Supervised Learning. First, all images in the training set undergo normalization preprocessing (e.g., resolution unified to the same DPI, grayscale normalization). Each pair of images and its corresponding two sets of landmark coordinates are used as a training sample. The network input can be an image from the first view. The training objective is for the network to predict a dense displacement field such that when the landmarks in the first view are deformed using this displacement field, the mean square error between their positions and the corresponding landmark positions in the second view is minimized. Common network structures include encoder-decoder convolutional neural networks, which use a specific layer (such as a spatial transformation layer) at the decoder end to output the displacement field. Through iterative training with a large number of samples, the network learns the complex mapping relationship from single-view images to cross-view deformation fields.

[0050] Path Two: Statistical Deformation Model Based on Group Image Registration. This method does not directly train an end-to-end network, but instead performs mathematical modeling based on landmark data. First, using all training image pairs, a precise elastic deformation field based on its own landmarks is calculated for each image pair through feature point matching and optimization algorithms (such as thin-plate spline interpolation). This yields a "deformation field library." Next, statistical shape analysis is performed on this deformation field library, specifically using principal component analysis. The average deformation field of all deformation fields is calculated, and then the covariance matrix between the deformation fields is analyzed to obtain a series of eigenvectors (i.e., principal components) representing the most dominant deformation patterns. Finally, the deformation of any new breast in both views can be approximated by this average deformation field plus a linear combination of several dominant deformation patterns. The model parameters are the weight coefficients of these combinations.

[0051] Regardless of the path taken, a deployable elastic deformation model file will ultimately be obtained. This model file can receive newly captured, unlabeled dual-view images of the breast (or one of them) and, based on learned generalization rules, estimate the nonlinear deformation relationship from one view to the other, which can then be used by region detection unit 2.

[0052] In a preferred embodiment, before determining the best matching object based on a weighted score, the region detection unit 2 is further configured to: Determine whether a specific matching state exists. The specific matching state is: a pairing combination formed by a first candidate region in the first view and at least two different second candidate regions in the second view, the weighted scores of which are all higher than the preset threshold used to determine a successful pairing.

[0053] Specifically, a specific matching state is a complex scenario automatically identified by the system during cross-view matching. Its core characteristic is that a candidate region from the first view, when attempting to match with candidate regions from the second view, does not have only one clearly superior match, but simultaneously forms a preliminary matching relationship with two or more different candidate regions from the second view, exhibiting a high degree of credibility. Weighted scores exceeding the preset threshold for determining successful matching mean that the comprehensive scores calculated by each of these multiple matching combinations (such as the aforementioned weighted scores based on shape and position) all exceed the basic threshold value set by the system for determining whether a preliminary matching is acceptable. This indicates that relying solely on basic scores and threshold comparisons, the system cannot automatically determine which is the only correct match, thus falling into a matching conflict or ambiguity state.

[0054] Detailed description of the implementation steps: In the process of determining the best matching object during the execution of the regional detection unit 2, the system embeds a pre-processing state judgment module. Its specific implementation steps are as follows: When processing a candidate region A in the first view, the system first determines its expected search area in the second view based on geometric projection relationships, and finds all candidate regions in the second view within that region, denoted as set B = {B1, B2, …, Bn}. Next, the system calculates the morphological similarity metric and spatial projection error between candidate region A and each element Bi in set B according to a predetermined algorithm, thereby obtaining the weighted score Score_i for each pair (A, Bi).

[0055] The system then compares these weighted scores to a preset "successful match threshold" T_success. This threshold T_success is typically determined during model development by balancing the match success rate and the false match rate on the validation set. Normally, if only one Score_i > T_success, Bi is directly determined as the best match. If all Score_i <= T_success, A is marked as a single-view suspicious point.

[0056] A specific matching state occurs when the system detects at least two distinct candidate regions, such as Bp and Bq, such that Score_p > T_success and Score_q > T_success, and p is not equal to q. In this case, the system triggers a flag, recording the ID of candidate region A in the first view, and a list of IDs of all candidate regions in the second view that satisfy Score_i > T_success. This means that for point A, there are multiple seemingly reasonable matching options, creating a one-to-many competitive situation.

[0057] Instead of immediately forcing an selection, the system identifies this state and transfers control or triggers a more complex conflict resolution mechanism. The identification itself is automated; the system completes this by iterating through and comparing the scores with thresholds. This step essentially adds an intelligent decision switch to the entire pairing process. When encountering complex situations that simple rules cannot handle, the system can recognize the situation and initiate a backup advanced analysis process.

[0058] In a preferred embodiment, the region detection unit 2 is further configured to: If the specific matching state is determined to exist, the contextual attributes of the first candidate region in the first view are evaluated, and its cluster membership confidence as a member of the calcification cluster is calculated. If the cluster membership confidence is higher than the cluster determination threshold, then the cluster-aware matching mode is executed; The cluster-aware matching mode is executed as follows: In the first view, identify all other candidate regions that have been successfully paired within a preset neighborhood centered on the first candidate region, and form a confirmed cluster reference set; For each second candidate region in the second view that forms the specific matching state with the first candidate region, it is combined with all the paired objects in the second view of the confirmed cluster reference set to form multiple complete candidate clusters to be evaluated; Calculate the spatial distribution evaluation index for each of the complete candidate clusters to be evaluated; The second candidate region that optimizes the spatial distribution evaluation index is selected as the final pairing object with the first candidate region.

[0059] Specifically, cluster membership confidence is a quantitative indicator used to assess the probability that a candidate region belongs to a certain calcification cluster. Its value range is usually between 0 and 1, and it can be output by a machine learning classifier (such as a gradient boosting tree based on the local features of the point and its relationship with surrounding points) or calculated by rule-based methods (such as the density of existing points in the neighboring region). The cluster determination threshold is a preset threshold value used to determine whether to enable the cluster-aware mode. If the cluster membership confidence is higher than this threshold, the point is considered to be likely a member of a cluster. The confirmed cluster reference set refers to the set of other candidate regions that have successfully completed cross-view pairings within a certain neighborhood of the currently processed conflict point (first candidate region) in the first view. These points and their pairing information constitute a reliable local spatial context. The complete candidate cluster to be evaluated is a virtual construction. For each conflict option in the second view, it is formed by combining the current option with all paired objects in the confirmed cluster reference set in the second view, resulting in the complete calcification cluster morphology that should be presented if the option is a correct match. Spatial distribution evaluation index is a scalar value used to assess the rationality of the spatial distribution pattern of a point set. For example, it can calculate the ratio of the area of ​​the convex hull formed by all points in the point set to the number of points (evaluating dispersion), or calculate the average distance from the centroid of the point set to each point (evaluating compactness). Its purpose is to quantify which conflicting option, when added, makes the morphology of the entire cluster look most consistent with the biological characteristics of real calcified clusters (usually exhibiting focal compact aggregation).

[0060] Detailed description of the implementation steps: When the system identifies a specific matching state and calculates the cluster membership confidence C_A of the first candidate region A, if C_A is higher than the preset cluster determination threshold (e.g., 0.7), the system enters the cluster-aware matching mode.

[0061] The first step in this pattern is to construct the context. The system defines a circular neighborhood with radius R (e.g., 20 mm) in the first view, centered on point A. Within this neighborhood, it searches for all other candidate regions that have been successfully paired (i.e., have found a unique match in the second view). Assume M such points are found, denoted as set S = {A1, A2, …, Am}. This set S is the "confirmed cluster reference set". Simultaneously, the system learns the paired objects of these M points in the second view, denoted as set T = {B1, B2, …, Bm}, which are distributed at their respective positions in the second view.

[0062] The second step is to generate a hypothetical cluster morphology. Assume that in the second view, there are K candidate regions that conflict with point A, i.e., the conflict set is C = {C1, C2, …, Ck}. For each candidate Cj in the conflict set, the system performs the following operation: merges Cj with the set of paired objects T of the confirmed cluster reference set to form a hypothetical complete cluster Hj = {Cj} ∪ T. This Hj represents the point set distribution of the entire calcified cluster that should be observed in the second view if Cj is a correct match for A.

[0063] The third step is evaluation and selection. The system calculates the spatial distribution evaluation index I_j for each hypothetical cluster Hj. For example, it calculates the covariance matrix of the two-dimensional coordinates of all points in Hj, finds its eigenvalues, and divides the smaller eigenvalue by the larger eigenvalue to obtain the eigenvalue ratio. The closer this ratio is to 0, the more the point set tends to cluster linearly or in clusters; the closer it is to 1, the more dispersed the point set is. Alternatively, it calculates the average Euclidean distance between all pairs of points in Hj as the average intra-cluster distance. Assuming we use the eigenvalue ratio as the index I_j, the smaller the value, the stronger the clustering and the more reasonable the distribution. The system then iterates through all I_j (j=1 to k) and selects the conflict candidate Cj that minimizes the I_j value. That is, Cj = argmin(I_j). This Cj* is determined to be the object that best matches the candidate region A in the first view, because when combined with the surrounding confirmed paired points, the spatial morphology formed best matches the expectation of a real, compact calcification cluster.

[0064] In a preferred embodiment, the region detection unit 2 is further configured to: If the cluster membership confidence is lower than the isolated point determination threshold, then a multi-dimensional re-evaluation mode is executed; The multi-dimensional re-evaluation mode is implemented as follows: Acquire additional image information associated with the target view group, including historical image sequences or dual-energy imaging data of the same breast tissue; Based on the additional image information, calculate the time stability measure or energy spectrum consistency measure of the pairing combination formed by the first candidate region and each conflicting second candidate region. The original weighted score is corrected using the time stability metric or energy spectrum consistency metric, and the final pairing is determined based on the corrected score.

[0065] This implementation provides an alternative solution for matching conflicts of isolated suspicious points. The isolated point determination threshold is a threshold lower than the cluster determination threshold. When the cluster membership confidence is below this value, the system considers the candidate region an isolated point, not belonging to any obvious cluster structure, and therefore employs a different strategy. Additional imaging information refers to other imaging data related to the same breast that the system can access besides the target view group (the dual views of this examination), which is the current subject of analysis. This mainly includes two categories: historical image sequences (i.e., mammograms from multiple past examinations of the patient) and dual-energy imaging data (taken with X-rays of different energy levels, providing additional information about tissue composition). Temporal stability measures quantify whether a hypothetical pairing persists over time. For example, it can check whether morphologically similar candidate regions can always be detected near the corresponding anatomical location in images from several past examinations. Spectral consistency measures quantify whether a hypothetical pairing is consistent in the material composition properties revealed by dual-energy imaging. For example, it calculates the pixel value similarity of two candidate regions on the "calcium content map." The revised score is a more comprehensive final score obtained by recalculating the original weighted score after incorporating time stability or energy spectrum consistency measures.

[0066] Detailed description of the implementation steps: When the first candidate region A is determined to be a specific matching state, and its cluster membership confidence C_A is lower than the isolated point determination threshold (e.g., 0.3), the system enters the multi-dimensional re-evaluation mode.

[0067] The activation of this mode depends on acquiring additional imaging information. The system indexes the medical image archive and communication system through the patient ID and attempts to retrieve two types of data: 1) the patient's previous, available mammogram images, sorted by time; 2) if the current examination includes dual-energy imaging, it simultaneously acquires the basic material breakdown images generated, such as "calcium material map" and "soft tissue map".

[0068] Next, the system calculates an additional confidence metric for each conflict candidate region Cj (j=1 to k) in the second view.

[0069] If historical image sequences are invoked, a temporal stability metric is calculated. The system searches for morphologically similar candidate regions near point A (or its deformed location) in corresponding views of historical images (e.g., the previous two inspections) using established temporal registration relationships. If stable corresponding points are found in a majority (e.g., over 70%) of the historical images, and these points exhibit some morphological continuity with the current conflict candidate Cj, the pairing (A, Cj) is considered temporally stable and assigned a high stability score. Conversely, if the corresponding location in the historical images is empty or inconsistent, the score is lower.

[0070] If dual-energy imaging data is used, an energy spectrum consistency metric is calculated. The system extracts the statistical features of calcium content (such as mean and maximum values) of the region where point A is located from the calcium content map in the first view, and extracts the corresponding features of the region where point Cj is located from the calcium content map in the second view. Then, the similarity (such as cosine similarity) between these two sets of features is calculated. If both A and Cj show typical high calcium content regions, the similarity is high; if one of them shows a weak similarity in the calcium content map, the similarity is low.

[0071] The system then uses this calculated additional metric (temporal stability metric S_j or spectral consistency metric E_j) to correct the original weighted score Score_j. For example, a linear weighted correction can be used: FinalScore_j = α * Score_j + β * S_j (or E_j), where α and β are coefficients that adjust the weights of the two parts, and α + β = 1. In this way, the final score not only considers the morphological and positional matching degree in this examination, but also incorporates its temporal continuity or physical composition consistency.

[0072] Finally, the system compares the FinalScore_j of all conflicting candidates and determines the candidate Cj* with the highest FinalScore_j as the final pairing with A. If all FinalScore_j are still below a certain acceptance threshold, the system may ultimately determine that a pairing is not possible and classify A as an uncertainty point requiring special attention.

[0073] In a preferred embodiment, the region detection unit 2 is specifically configured to: For candidate regions marked as suspicious points in a single view, the local image features of the suspicious points in the single view are analyzed. If the feature pattern matches the feature library of known typical image noise, artifacts or blood vessel sections with a degree higher than the first matching threshold, then the region is filtered out.

[0074] Specifically, local image features refer to a set of quantitative indicators extracted from the image of the region containing the suspicious point to describe its visual attributes. Examples include the region's grayscale histogram statistical features (mean, variance, skewness, kurtosis), gradient magnitude features, and local binary pattern features. The known typical image noise pattern library is a pre-built knowledge base containing image feature templates or feature distribution ranges of various artifacts and noises commonly found in mammograms and confirmed by experts. Examples include the striped texture features of motion artifacts, the fixed-position high-brightness features caused by detector defects, and the specific morphological and texture features of blood vessel cross-sections. Feature library matching degree is a similarity score used to quantify the similarity between the feature vector of the single-view suspicious point to be evaluated and a certain type of typical artifact feature template in the feature library. It can be calculated using Euclidean distance, cosine similarity, or probability output based on a machine learning classifier. The first matching threshold is a pre-set numerical limit used to determine whether the matching degree is high enough to identify the suspicious point as an artifact and filter it out.

[0075] Detailed description of the implementation steps: This filtering mechanism is an important quality control step in the processing flow of the area detection unit 2. When the system determines that a candidate area P in the first view fails to be successfully matched with any candidate area in the second view (i.e., the weighted score of its best matching object is lower than a preset threshold), P is marked as a "single-view suspicious point" and enters this processing branch.

[0076] The system first performs feature extraction on the suspicious point P. A fixed-size image patch (e.g., 32×32 pixels) is cropped from the original resolution single-view image, centered on the coordinates of P. After normalization preprocessing of this image patch, a set of carefully designed local image features is calculated. These features need to be able to distinguish between genuine microcalcifications and common artifacts. Typical feature combinations include: Texture features: Calculate the gray-level co-occurrence matrix of image patches and extract statistics such as contrast, correlation, energy, and homogeneity from it.

[0077] Shape and edge features: Threshold segmentation of image blocks is performed to obtain binary regions, and the eccentricity, roundness, and average gradient intensity of the edges of these regions are calculated.

[0078] Frequency domain characteristics: Perform a fast Fourier transform on the image patch and analyze the energy distribution of its spectrum. Some periodic noises (such as fence artifacts) will show spikes at specific frequencies.

[0079] Gray-scale distribution characteristics: Calculate the mean and standard deviation of gray-scale values ​​of pixels within an image block, as well as higher-order statistical moments.

[0080] Next, the system calculates the matching degree. These extracted features are combined into a feature vector. The system inputs this feature vector into a pre-trained "artifact discriminator." This discriminator can be a multi-class support vector machine classifier, whose training data consists of a large number of image patches and their feature vectors that have been labeled by experts as categories such as "real microcalcification," "image noise," "vascular cross-section," "motion artifact," and "metal artifact." The sum of the probabilities corresponding to each artifact category output by the classifier can be used as the "matching degree with a feature library of known typical image noise, artifacts, or vascular cross-sections." Another more direct method is to calculate the similarity between the feature vector of the suspicious point and the average vector (or representative vector) of each artifact feature stored in the feature library, and take the maximum value as the matching degree.

[0081] Finally, the system makes a decision based on a comparison between the matching degree and a first matching threshold. The first matching threshold is determined during the model validation phase to filter out obvious artifacts as much as possible while ensuring high specificity. For example, a threshold of 0.85 might be set. If the calculated matching degree is higher than 0.85, the system is confident that the suspicious point P is actually a known type of artifact or noise, rather than a potential microcalcification. In this case, the system removes P from the further analysis queue; that is, it is no longer passed to subsequent classification units for benign or malignant determination, nor is it considered a key area requiring doctor review. This directly reduces the noise in the system output and lowers the unnecessary workload for doctors.

[0082] Specifically, if the matching degree is lower than the threshold, it means that the feature pattern of P is not typical enough and cannot be simply identified as an artifact. The system will retain P and pass it to the following analysis process: In a preferred embodiment, the region detection unit 2 is further configured to: For the virtual cross-view attached Figure 1 For suspicious points in a single view with consistent characteristics, perform differential classification processing, which includes: A dedicated classifier with a higher sensitivity threshold, distinct from conventional paired calcification classification models, is used for processing. And / or, In the enhanced feature vector, the virtual cross-view is reduced. Figure 1 Weights of consistency features; Specifically, suspicious points in a single view that have undergone differential classification processing and whose classification confidence is within a specific uncertainty range are assigned a specific suspicious type label in their output results.

[0083] This implementation occurs after the system has completed the virtual feature generation for suspicious points in a single view (for example, when the distance between a suspicious point and a neighboring paired cluster is less than the association radius, a virtual matching point is calculated based on the position of the neighboring paired points in another view, and then the virtual spatial projection error and morphological variability are calculated).

[0084] First, computational unit 4 performs differential classification processing. The system employs one or a combination of the following two technical approaches: Path 1: Using a dedicated classifier. The system maintains two pre-trained classification models. The main classification model handles calcifications with reliable features and true pairing information. The dedicated classifier is trained using a specially constructed training set, which not only contains standard paired calcifications but also artificially constructs a large number of samples with "virtual features" or missing features. During training, the loss function is adjusted to make the model more sensitive to the "malignant" category (i.e., lowering its threshold for detecting malignant lesions). When it is necessary to classify a suspicious point in a single view with virtual features, the system calls this dedicated classifier, inputs its enhanced feature vector (containing true single-view features and virtual consistency features), and obtains a preliminary classification probability.

[0085] Path Two: Adjust Feature Weights. The system still uses the main classification model, but when constructing the input feature vector for this suspicious point, the virtual cross-view weights are adjusted. Figure 1 Consistency features are weighted less. For example, if the weight factor for spatial projection error is 1.0 in the regular process, the weight for virtual projection error might be set to 0.3 or 0.5. Other virtual features, such as morphological variability, are treated similarly. This is equivalent to explicitly telling the classifier during feature fusion: "This part of the consistency information has low reliability; please refer to it with caution." Then, this weighted feature vector is input into the main classification model for classification.

[0086] Regardless of the path used, the classifier will output a result, which typically includes a benign / malignant label (or a malignant probability value) and a confidence score.

[0087] Next, the system assigns and labels the results. The output processing module checks the classification confidence of the point. The system defines a "specific uncertainty interval," for example, when the malignancy probability value P is between 0.25 and 0.65, the result is considered to have high uncertainty. If the classification result of the suspicious point falls into this interval, the system will not simply output it as benign or malignant, but will assign it a specific suspicious type label. This label may be an enumerated constant named "UNCERTAIN_WITH_CONTEXT" internally. This label carries multiple meanings: it indicates that the point originated from a single-view detection and the pairing failed; it underwent context-based virtual feature compensation and conservative classification processing; and its final classification result has insufficient confidence.

[0088] Finally, when generating structured reports or visualizations, this "specific suspicious type label" is mapped to a physician-friendly prompt. For example, in the text description section of the report, the system might generate a standardized description such as, "A cluster of isolated microcalcifications was found at position X. Due to the lack of a clear correspondence between the two views, the risk of malignancy is uncertain. It is recommended to combine this with 3D breast tomography or clinical review." In image annotations, this point may be displayed with a special color (such as orange) and an icon (such as a question mark) to distinguish it from high-confidence malignant (red) or benign (green) markings.

[0089] This implementation presents a more collaborative human-computer interaction model: the system's role extends beyond detection and diagnosis, intelligently screening and identifying complex cases that require in-depth interpretation by doctors using their comprehensive experience, and providing them with structured analytical context. This leverages AI's advantages in image scanning and quantitative analysis while respecting the irreplaceable role of doctors in complex and comprehensive judgments, thereby improving the efficiency and safety of the entire diagnosis and treatment process.

[0090] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A computer-aided detection and diagnostic system for breast microcalcifications, characterized in that, include: The acquisition unit (1) is configured to acquire at least two different projection angles of the same breast tissue as a target view group; The area detection unit (2), electrically connected to the acquisition unit (1), is configured for: Image preprocessing and suspected microcalcification candidate region detection are performed on each single view in the target view group to obtain the candidate region set corresponding to each single view; Based on the geometric projection relationship between different views in the target view group, cross-view pairing and association are performed on candidate regions from different views to establish candidate region pairs that characterize the image performance of the same potential microcalcification foci under different projection angles. Extraction module (3), electrically connected to the region detection unit (2), is configured for: For each successfully paired candidate region pair, its cross-view consistency features are extracted, including spatial projection error, morphological feature variability, or texture feature stability measure. The calculation unit (4), electrically connected to the extraction module (3), is configured to: The cross-view consistency features of each candidate region and its internal image features in each single view are fused to form the enhanced feature vector of the potential microcalcification. Based on the enhanced feature vector, the benign or malignant classification result of the corresponding potential microcalcification is obtained, along with the corresponding confidence score.

2. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 1, characterized in that, The specific configuration of the region detection unit (2) is as follows: For any candidate region in the first view, calculate its expected projection region in the second view based on the geometric projection relationship; Search for all candidate regions located within the expected projection area in the second view, and calculate the morphological similarity measure between each searched candidate region and the candidate region in the first view; Based on the weighted score of the morphological similarity measure and spatial projection error, the best matching object is determined from the candidate regions of the second view to form candidate region pairs; If the weighted score of the best matching object is lower than a preset threshold, it is determined that the candidate region in the first view failed to be successfully matched in the second view, and it is marked as a single-view suspicious point for independent processing.

3. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 1, characterized in that, The extraction module (3) is specifically configured for: Calculate at least one set of morphological parameters for the same candidate region in two views presented from different projection angles, the morphological parameters including area, perimeter, roundness, aspect ratio, or edge sharpness; Calculate the relative difference or standard deviation of each set of morphological parameters between the two views, as the variability of that set of parameters; The variability of each morphological parameter is normalized and weighted to obtain the morphological feature variability that characterizes the overall morphological stability of the candidate region.

4. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 2, characterized in that, The geometric projection relationship is obtained based on the elastic deformation model; the region detection unit (2) is specifically configured to: use the elastic deformation model to perform nonlinear deformation mapping on the position of the candidate region in the first view, so as to determine its corresponding expected search region in the second view.

5. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 4, characterized in that, The elastic deformation model is constructed in the following manner: Multiple sets of paired, labeled, dual-view X-ray images of the breast were obtained as a training set; For each pair of images, a set of homologous anatomical landmarks is extracted, which includes the corresponding location points of the breast contour, the edge of the pectoralis major muscle, and the nipple. Based on all paired anatomical landmark sets in the training set, a deep learning network is trained through supervised learning, or a statistical deformation model is fitted through a group image registration algorithm, to obtain the elastic deformation model that can generalize to predict tissue deformation of different breasts from the first projection angle to the second projection angle.

6. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 2, characterized in that, Before determining the best matching object based on the weighted score, the region detection unit (2) is also configured to: Determine whether a specific matching state exists. The specific matching state is: a pairing combination formed by a first candidate region in the first view and at least two different second candidate regions in the second view, the weighted scores of which are all higher than the preset threshold used to determine a successful pairing.

7. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 6, characterized in that, The region detection unit (2) is also configured to: If the specific matching state is determined to exist, the contextual attributes of the first candidate region in the first view are evaluated, and its cluster membership confidence as a member of the calcification cluster is calculated. If the cluster membership confidence is higher than the cluster determination threshold, then the cluster-aware matching mode is executed; The cluster-aware matching mode is executed as follows: In the first view, identify all other candidate regions that have been successfully paired within a preset neighborhood centered on the first candidate region, and form a confirmed cluster reference set; For each second candidate region in the second view that forms the specific matching state with the first candidate region, it is combined with all the paired objects in the second view of the confirmed cluster reference set to form multiple complete candidate clusters to be evaluated; Calculate the spatial distribution evaluation index for each of the complete candidate clusters to be evaluated; The second candidate region that optimizes the spatial distribution evaluation index is selected as the final pairing object with the first candidate region.

8. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 7, characterized in that, The region detection unit (2) is also configured to: If the cluster membership confidence is lower than the isolated point determination threshold, then a multi-dimensional re-evaluation mode is executed; The multi-dimensional re-evaluation mode is implemented as follows: Acquire additional image information associated with the target view group, including historical image sequences or dual-energy imaging data of the same breast tissue; Based on the additional image information, calculate the time stability measure or energy spectrum consistency measure of the pairing combination formed by the first candidate region and each conflicting second candidate region. The original weighted score is corrected using the time stability metric or energy spectrum consistency metric, and the final pairing is determined based on the corrected score.

9. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 2, characterized in that, The specific configuration of the region detection unit (2) is as follows: For candidate regions marked as suspicious points in a single view, the local image features of the suspicious points in the single view are analyzed. If the feature pattern matches the feature library of known typical image noise, artifacts or blood vessel sections with a degree higher than the first matching threshold, then the region is filtered out.

10. The computer-aided detection and diagnosis system for breast microcalcifications according to claim 9, characterized in that, The region detection unit (2) is also configured to: For single-view suspicious points with the aforementioned virtual cross-view consistency feature, differential classification processing is performed, which includes: A dedicated classifier with a higher sensitivity threshold, distinct from conventional paired calcification classification models, is used for processing. And / or, In the enhanced feature vector, the weight of the virtual cross-view consistency feature is reduced; Specifically, suspicious points in a single view that have undergone differential classification processing and whose classification confidence is within a specific uncertainty range are assigned a specific suspicious type label in their output results.

Images