Breast molybdenum target calcification point recognition and segmentation method based on breast ultrasound image

Abstract

This invention relates to the field of image processing technology, specifically to a method for identifying and segmenting mammographic calcifications based on breast ultrasound images. The method includes: acquiring dynamic ultrasound image sequences of the patient to be analyzed and each historical patient in a database; extracting feature candidate points for each anatomical region in the ultrasound images, and obtaining calcification feature indices for each anatomical region in the ultrasound images; constructing a probability density image based on the feature density of the mapping results of calcifications in the mammographic images onto the ultrasound images; obtaining feature association confidence based on the correlation features between the calcification feature indices and the corresponding probability densities; training an image segmentation network by combining the feature association confidence and the calcification feature indices; and obtaining the segmentation results of mammographic calcifications using the image segmentation network based on the calcification feature indices and the feature association confidence of the patient to be analyzed. This invention achieves intelligent segmentation of mammographic calcifications in the patient's mammographic images solely based on the patient's ultrasound images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for identifying and segmenting calcifications on mammograms based on breast ultrasound images. Background Technology

[0002] Breast cancer is one of the most serious malignant tumors threatening women's physical and mental health. Accurate detection of early-stage, asymptomatic microcalcifications is a crucial step in achieving early screening, diagnosis, and treatment of breast cancer, thereby reducing mortality. Currently, clinical imaging screening for breast cancer mainly includes two mainstream methods: mammography and breast ultrasound. Each of these imaging modalities has its own technical advantages and limitations.

[0003] Mammography, with its high spatial resolution and high-density tissue identification, possesses extremely high detection sensitivity for breast microcalcifications ranging from 0.1 to 0.2 mm. It can clearly present the distribution morphology, aggregation range, and topological structure of dense microcalcifications, making it the clinically recognized gold standard for breast microcalcification detection. However, mammography involves ionizing radiation, making it unsuitable for pregnant women, breastfeeding mothers, and those requiring frequent follow-up examinations. Furthermore, its application is significantly limited by the patient's positioning and the density of the mammary glands.

[0004] Breast ultrasound examination boasts significant advantages, including no ionizing radiation, real-time dynamic imaging, convenient operation, repeated screening capability, and strong adaptability to dense breast tissue, making it the preferred method for large-scale routine breast screening in women. However, due to limitations in the ultrasound imaging mechanism, dense, tiny breast calcifications exhibit a gentle gray-scale gradient and extremely low pixel contrast with surrounding normal glandular tissue, making them easily masked by inherent ultrasound spot noise and tissue texture noise. Traditional image segmentation algorithms and conventional deep learning models rely solely on the gray-scale and edge features of a single ultrasound image for detection, making it difficult to capture subtle changes in local acoustic texture caused by microcalcification aggregation. This easily leads to problems such as missed calcification detection and localization errors, making it impossible to achieve refined identification and accurate segmentation of dense microcalcifications. Relying solely on ultrasound images is insufficient to achieve the microcalcification detection capability at the level of mammography.

[0005] Current multimodal breast imaging research largely remains at the level of simple registration and fusion, and direct feature overlay. It fails to effectively transfer the advantages of high-precision calcification annotation from mammography to ultrasound imaging, lacking transferable cross-modal feature mapping mechanisms and prior guidance strategies. This hinders the establishment of intelligent microcalcification recognition and segmentation models applicable to pure ultrasound imaging. Therefore, there is an urgent need to construct a cross-modal feature association learning method that leverages the gold standard calcification information from mammography to build prior features for ultrasound image regions. This would enable high-precision identification and segmentation of dense, micro-calcifications in the breast, previously only clearly visible on mammography, using only breast ultrasound images without the aid of mammography, thus overcoming the technical limitations of single-mode ultrasound imaging in microcalcification detection. Summary of the Invention

[0006] To address the low accuracy of existing methods that rely solely on ultrasound images for identifying and segmenting mammographic calcifications, this invention aims to provide a method for identifying and segmenting mammographic calcifications based on breast ultrasound images. The specific technical solution adopted is as follows: Acquire dynamic ultrasound image sequences of the patient to be analyzed and each historical patient in the database, the dynamic ultrasound image sequences comprising several ultrasound images and each of their anatomical regions; Feature candidate points of each anatomical region in the ultrasound image are extracted. Based on the pixel differences and distance distribution between the feature candidate points in the anatomical region, and combined with the pixel feature distribution of all feature candidate points, the calcification feature index of each anatomical region in the ultrasound image is obtained. Obtain mammograms for each historical patient and construct a probability density image based on the feature density of the mapping results of calcifications in the mammograms onto the ultrasound images; The feature association confidence is obtained by considering the correlation between the calcification feature index and the corresponding probability density of each anatomical region in the ultrasound images of each historical patient. The image segmentation network is then trained by combining the feature association confidence and the calcification feature index corresponding to each historical patient in all ultrasound images. Based on the calcification feature indices and feature association confidence scores corresponding to the patient in all ultrasound images, the image segmentation network is used to obtain the segmentation results of calcification points on the molybdenum target.

[0007] Preferably, the step of obtaining the calcification feature index of each anatomical region in the ultrasound image based on the pixel differences and distance distribution between feature candidate points within the anatomical region, combined with the pixel feature distribution of all feature candidate points, specifically includes: For any ultrasound image of any patient, any anatomical region is taken as the target anatomical region; the corresponding anatomical region in the ultrasound images of the time sequence adjacent to the target anatomical region is taken as the reference anatomical region. Based on the pixel differences and distance distribution between each feature candidate point in the target anatomical region and its nearest neighbor other feature candidate points, and combined with the pixel fluctuation characteristics of all feature candidate points in the target anatomical region, the state influence factor of the target anatomical region is obtained. The ultrasound response factor of the target anatomical region is obtained based on the comprehensive distribution of pixel values ​​of all feature candidate points in the target anatomical region and the area distribution of the target anatomical region. Based on the consistency of characteristic changes between the differences in state influencing factors and ultrasound response factors between the target anatomical region and the reference anatomical region, calcification characteristic indicators of the target anatomical region are obtained.

[0008] Preferably, the step of obtaining the state influence factor of the target anatomical region based on the pixel difference and distance distribution between each feature candidate point in the target anatomical region and its nearest neighbor other feature candidate points, combined with the pixel fluctuation characteristics of all feature candidate points in the target anatomical region, specifically includes: In the target anatomical region, a first feature coefficient is determined based on the pixel difference between each feature candidate point and its nearest neighbor other feature candidate points, and a second feature coefficient is determined based on the Euclidean distance between each feature candidate point and its nearest neighbor other feature candidate points; The fluctuation coefficient, first feature coefficient, and second feature coefficient of the pixel values ​​of all feature candidate points in the target anatomical region are fused to obtain the state influence factor of the target anatomical region.

[0009] Preferably, obtaining the ultrasound response factor of the target anatomical region based on the comprehensive distribution of pixel values ​​of all feature candidate points in the target anatomical region and the area distribution of the target anatomical region specifically includes: Based on the pixel values ​​of all feature candidate points in the target anatomical region, determine the cumulative pixel value of the target anatomical region; The ratio between the cumulative pixel value and the area of ​​the target anatomical region is used as the ultrasound response factor of the target anatomical region.

[0010] Preferably, the step of obtaining calcification characteristic indicators of the target anatomical region based on the consistency of characteristic changes between the differences in state influencing factors and ultrasound response factors between the target anatomical region and the reference anatomical region specifically includes: The difference in state influencing factors between the target anatomical region and the reference anatomical region is taken as the first difference, and the difference in ultrasound response factors between the target anatomical region and the reference anatomical region is taken as the second difference. Based on the consistency of the changing trends between the first and second differences, the change direction coefficient is determined; based on the consistency of the changing magnitude between the first and second differences, the change magnitude coefficient is determined; the change direction coefficient and the change magnitude coefficient are fused to obtain the calcification characteristic index of the target anatomical region.

[0011] Preferably, the step of obtaining the feature association confidence based on the correlation between the calcification feature indicators and the corresponding probability density of each anatomical region in the ultrasound images of each historical patient specifically includes: Using any anatomical region as the calibration anatomical region, for any historical patient, the average level of calcification characteristic indicators of the calibration anatomical region in all ultrasound images is obtained as the calcification balance index of the historical patient in the calibration anatomical region. For any pixel, construct a calcification equilibrium index sequence by the calcification equilibrium index of the anatomical region where the pixel is located for each historical patient; construct a probability density sequence by the probability density of the probability density image of each historical patient at the location of the pixel. Calculate the correlation coefficient between the calcification equalization index sequence and the probability density sequence to obtain the feature correlation confidence of any pixel.

[0012] Preferably, training the image segmentation network by combining feature association confidence and the calcification feature index corresponding to each historical patient in all ultrasound images specifically includes: For any pixel, the product of the calcification balance index of the anatomical region where the pixel is located and the feature association confidence is used as the calcification significance of the pixel in each historical patient. A calcification feature image for each historical patient is constructed based on the pixel position of the ultrasound image, with each pixel representing the calcification significance for each historical patient. Using the calcification feature image and corresponding ultrasound image of each historical patient as input, a cross-entropy loss function is constructed based on the pixel difference between the calcification feature image and the corresponding probability density image of each historical patient. The training process of the image segmentation network is supervised by the cross-entropy loss function.

[0013] Preferably, the step of obtaining the segmentation result of mammography calcification points using the image segmentation network based on the calcification feature indicators and the feature association confidence scores corresponding to the patient in all ultrasound images specifically includes: For any pixel, the average level of calcification feature index of the anatomical region in which the pixel is located in all ultrasound images of the patient to be analyzed is obtained as the calcification balance index of the patient at that pixel. Based on the product of the calcification balance index and the feature association confidence of each pixel of the patient to be analyzed, a calcification feature image of the patient to be analyzed is constructed. The calcification feature image and the corresponding ultrasound image are input into the image segmentation network to obtain the segmentation result of calcification points on the mammogram.

[0014] Preferably, the step of extracting feature candidate points for each anatomical region in the ultrasound image specifically includes: For any ultrasound image, pixels with a value greater than or equal to a preset highlight threshold are designated as highlight pixels. If the pixel value of a highlight pixel is the maximum value of all pixel values ​​in its neighborhood, then the highlight pixel is designated as a feature candidate point.

[0015] Preferably, the construction of the probability density image based on the feature density of the mapping result of calcification points in the mammogram to the ultrasound image specifically includes: The molybdenum target image includes calcification points. Using a spatial registration algorithm, the molybdenum target image is mapped to the space of the ultrasound image. In the mapping result, each calcification point is used as the center of the Gaussian kernel. Based on the cumulative sum of the Gaussian kernel function values ​​from each pixel to each Gaussian kernel center, the probability density of each pixel is determined. Based on the probability density of all pixels in the mapping result, a probability density image is constructed.

[0016] The embodiments of the present invention have at least the following beneficial effects: This invention first constructs a complete multi-patient image dataset, uniformly collecting dynamic ultrasound image sequences of the patient to be analyzed and historical patients, and fully preserving the anatomical region segmentation information corresponding to each image, providing standardized original image samples for subsequent calcification feature extraction, cross-modal association modeling, and network training. Then, it mines the calcification texture and distribution features of anatomical regions from both pixel and spatial dimensions, transforming the image visual features into quantifiable calcification feature indicators to characterize the severity of calcification in different anatomical regions, providing quantitative feature support for cross-modal association analysis. Next, it achieves cross-modal spatial mapping of mammography calcification information to ultrasound images, quantifying the distribution probability of calcification points in different ultrasound regions, forming a density benchmark atlas that can be used for association matching, and opening up feature association channels between the two types of images. Furthermore, it quantifies the matching confidence of dual-modal image features, using association confidence as a constraint to complete the segmentation network training, allowing the network to learn the inherent mapping law between ultrasound calcification features and mammography calcification distribution, improving the network's ability to identify and adapt to calcification regions. Finally, it achieves intelligent segmentation of mammography calcification points in the patient to be analyzed solely based on the patient's ultrasound images. Attached Figure Description

[0017] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the steps of a method for identifying and segmenting calcifications in mammography based on breast ultrasound images provided by the present invention. Figure 2 This is a flowchart of the steps for obtaining calcification characteristic indicators provided by the present invention; Figure 3 This is a flowchart of the steps for obtaining feature association confidence provided by the present invention. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details a specific scheme for identifying and segmenting mammary calcifications based on breast ultrasound images, provided by this invention.

[0021] Please see Figure 1 The diagram illustrates a flowchart of a method for identifying and segmenting calcifications in mammography based on breast ultrasound images, according to an embodiment of the present invention. The method includes the following steps: Step S100: Obtain dynamic ultrasound image sequences of the patient to be analyzed and each historical patient in the database. The dynamic ultrasound image sequences contain several ultrasound images and each of their anatomical regions.

[0022] To achieve accurate identification and segmentation of breast microcalcifications, subsequent analysis based on the temporal texture features of dynamic ultrasound image sequences is necessary. In clinical breast ultrasound single-frame images, dense microcalcifications exhibit low contrast with glandular tissue and are easily masked by speckle noise. Relying solely on a single static image cannot capture the local texture and temporal fluctuations caused by probe micro-movements and beam angle changes. Furthermore, without anatomical region segmentation, refined analysis of local calcification features is difficult. Traditional identification methods often use scattered ultrasound samples, lacking historical patient medical record data, and cannot establish universal feature mapping rules, resulting in poor model generalization ability.

[0023] This invention considers that dynamic ultrasound image sequences contain multiple frames of image information from continuous scanning, which can completely preserve the temporal changes in grayscale and texture of each anatomical region. Simultaneously, the introduction of historical patient data from a database can accumulate sufficient samples, providing a data foundation for cross-modal feature association, regional parameter modeling, and network training. Therefore, it is necessary to acquire dynamic ultrasound image sequences of the patient to be analyzed and each historical patient in the database, and to divide the anatomical regions corresponding to each ultrasound image in the sequence. Under ideal data acquisition conditions, the acquired ultrasound sequence frames have high integrity, and the anatomical region division closely matches the physiological structure of the breast, accurately matching the requirements of subsequent mammography image registration and feature extraction, providing standardized and complete raw data support for subsequent training of the segmentation network.

[0024] As a concrete example, in the database, mammograms containing calcification density annotations are acquired for each historical patient. Simultaneously, continuous exposure ultrasound dynamic images covering the corresponding areas of the calcification density annotations are extracted from the ultrasound examination data of the historical patients. Medical image registration and uniform region segmentation are performed on all continuous ultrasound dynamic images of the historical patients to obtain a dynamic ultrasound image sequence for each historical patient. Preferably, to avoid noise interference with subsequent feature calculation results, this embodiment performs filtering processing on the acquired dynamic ultrasound image sequence. The filtering method is a known technique, such as the Gaussian filtering algorithm, which will not be elaborated further here.

[0025] It should be noted that the multimodal breast imaging data used in this embodiment of the invention comes from publicly available breast imaging databases. The cases in these databases all include mammograms and corresponding dynamic ultrasound sequences (or publicly available resources for ultrasound images) of the same patient. All data has been anonymized, does not contain patient privacy information, and has been authorized for use. Calcification density marking refers to the target area containing microcalcification clusters manually circled by the physician. Ultrasound data comes from historical examination records stored in the same publicly available database or related clinical cohorts. These ultrasound images were acquired during the same period or visit as the patient's mammogram examination. Since this ultrasound data is historical data, this embodiment of the invention only extracts local segments from the sequence corresponding spatially to the mammogram-marked lesions during the data processing stage.

[0026] Mammography images provide a global static spatial coordinate reference including microcalcifications, while dynamic ultrasound sequences provide dynamic temporal texture features of local areas in the corresponding anatomical coordinates. Both point to the same target lesion area in the patient's physical anatomical space.

[0027] Preferably, in one embodiment of the present invention, mapping the original ultrasound images of different patients to a common standard anatomical template enables cross-sample feature statistics of the same anatomical region from different patients based on a fixed anatomical region. Furthermore, the specific method for obtaining a dynamic ultrasound image sequence for each historical patient by performing medical image registration and uniform region segmentation on continuous dynamic ultrasound images of all historical patients includes: For all historical patients' continuous dynamic ultrasound images, a medical image registration method was used to uniformly map them onto a standard anatomical template and divide them into a fixed number of anatomical regions. It should be noted that the medical image registration method is a well-known technique and will only be briefly introduced here.

[0028] Specifically, the nipple center, chest wall line, and epidermal contour line from the original ultrasound images of each historical patient were extracted as anatomical anchor points. A standard anatomical template for the breast was established, with the reference nipple as the origin of a two-dimensional coordinate system and the reference chest wall line as the boundary. Using affine transformation, the translation, rotation, and scaling parameter matrices from the anatomical anchor points of each historical patient to the standard template anchor points were calculated. This matrix was used to map and align the original ultrasound images of each historical patient to the standard template space. A polar coordinate system was established on the aligned standard template for partitioning. Centered on the nipple origin, multiple concentric rings were divided according to a preset physical distance step. Twelve angular sectors were then divided equidistantly along the circumference according to the clinical clock orientation. Each sector grid formed by the intersection of the concentric rings and the angular sectors constitutes an independent anatomical region.

[0029] It should be noted that the mapping method for subsequent probability density images is the same, and will not be repeated here. The method for dividing the anatomical region can be set according to the specific implementation scenario. For example, the ultrasound image after aligning with the standard anatomical template can be uniformly divided to obtain several anatomical regions. It should be understood that in the embodiments of the present invention, several ultrasound images in the dynamic ultrasound image sequence of each historical patient are ultrasound images after aligning with the standard anatomical template.

[0030] Step S200: Extract feature candidate points for each anatomical region in the ultrasound image. Based on the pixel differences and distance distribution between feature candidate points within the anatomical region, and combined with the pixel feature distribution of all feature candidate points, obtain the calcification feature index for each anatomical region in the ultrasound image.

[0031] To accurately quantify the distribution characteristics and acoustic response properties of microcalcifications in different anatomical regions of the breast, it is necessary to comprehensively characterize them based on multidimensional information from candidate feature points within the region. Traditional breast ultrasound calcification identification relies solely on single-point grayscale thresholds, ignoring grayscale pixel differences, spatial arrangement patterns, and overall pixel distribution characteristics between candidate points. This makes it impossible to distinguish between dense clusters of microcalcifications and ordinary glandular texture or ultrasound artifacts, easily leading to missed detections of microcalcifications and misjudgments of false lesions, thus hindering the formation of quantifiable regional calcification assessment criteria.

[0032] In this embodiment of the invention, candidate feature points within the anatomical region are considered to be strong echogenic key points suspected of microcalcification. Pixel differences between candidate points can reflect local grayscale fluctuations, distance distribution can reflect the density of calcification aggregation, and the overall pixel feature distribution characterizes the overall intensity of the acoustic response of the region. Multi-dimensional information fusion is necessary to objectively characterize the calcification properties of the region. Therefore, candidate feature points are extracted from each anatomical region of the ultrasound image. Combining the pixel differences between candidate points, spatial distance distribution, and overall pixel feature distribution, calcification feature indices corresponding to each anatomical region are calculated. These indices characterize the temporal consistency between the dynamic ultrasound texture changes and the acoustic response caused by microcalcification aggregation within the target anatomical region; higher values ​​indicate a greater likelihood of the presence of genuine microcalcification features in the region.

[0033] Preferably, in one embodiment of the present invention, the specific method for extracting feature candidate points for each anatomical region in the ultrasound image includes: For any anatomical region in any ultrasound image, pixels with a value greater than or equal to a preset highlight threshold are designated as highlight pixels. If the pixel value of a highlight pixel is the maximum value of all pixel values ​​in its neighborhood, then the highlight pixel is designated as a feature candidate point.

[0034] In breast ultrasound images, microcalcifications appear as strong echo spots with local gray-scale abrupt changes. However, due to the inherent spot noise in ultrasound imaging and the low contrast between microcalcifications and surrounding glandular tissue, this embodiment combines the dual criteria of spatial local brightness and absolute brightness compliance to screen for false maxima caused by random noise, while retaining key strong echo points with pathological significance.

[0035] As a concrete example, for an ultrasound image that has undergone spatial normalization and registration, feature candidate point extraction is performed region by region for each anatomical region. Specifically, for any anatomical region in any ultrasound image, the average value of all pixels within that anatomical region is calculated. A preset brightness threshold is set to a preset multiple of the average pixel value of that anatomical region, such as 1.5 to 2.0 times, so that the threshold is slightly higher than the average pixel level of the anatomical region, thereby effectively filtering out relatively bright, strong echo regions within the anatomical region.

[0036] Furthermore, in this embodiment, if the pixel value of the highlighted pixel is the maximum value of all pixel values ​​within the 8-neighborhood of the highlighted pixel, then the highlighted pixel is determined as a feature candidate point, indicating that the pixel value of the highlighted pixel belongs to a local maximum value within the neighborhood.

[0037] The first grayscale threshold screening ensures that the feature candidate points have sufficient echo intensity, eliminating low-brightness background noise; the second local maximum screening ensures that the feature candidate points are the brightest structures in their neighborhood space, avoiding the mistaken inclusion of non-peak edge bright areas into the candidate set.

[0038] Preferably, such as Figure 2 As shown, in one embodiment of the present invention, the method for obtaining calcification feature indicators of each anatomical region in the ultrasound image can be specifically implemented by steps S201 to S204.

[0039] Step S201: For any ultrasound image of any patient, any anatomical region is taken as the target anatomical region; the corresponding anatomical region in the ultrasound image of the time sequence adjacent to the target anatomical region is taken as the reference anatomical region.

[0040] As a specific example, the methods for obtaining calcification feature indicators for each anatomical region are the same for both the patient to be analyzed and historical patients. This embodiment uses any anatomical region in any ultrasound image of any patient as an example for illustration. Furthermore, it should be understood that the ultrasound images of each patient are arranged in chronological order to form a dynamic ultrasound image sequence. In this embodiment, the same anatomical region in the previous ultrasound image adjacent to the target anatomical region is used as the reference anatomical region for the target anatomical region. It should be understood that since the ultrasound images have undergone spatial normalization and registration, and the anatomical regions are fixedly divided, the correspondence between anatomical regions in different ultrasound images specifically represents the anatomical region at the same location in different ultrasound images.

[0041] Step S202: Based on the pixel differences and distance distribution between each feature candidate point in the target anatomical region and its nearest neighbor other feature candidate points, and combined with the pixel fluctuation characteristics of all feature candidate points in the target anatomical region, the state influence factor of the target anatomical region is obtained.

[0042] In densely microcalcified regions, calcification points are scattered in a gravel-like pattern. The varying particle density leads to inconsistent echo intensities, while dramatic brightness jumps occur between the calcification points and the surrounding tiny gaps. Therefore, it is necessary to simultaneously quantify the gray-level differences and distance distribution among feature candidate points (reflecting spatial contrast intensity) and the overall gray-level fluctuation of the feature candidate points (reflecting particle inhomogeneity). Both methods work together to characterize the microcalcification aggregation topology of this anatomical region.

[0043] Preferably, in one embodiment of the present invention, the method for obtaining the state influencing factors of the target anatomical region specifically includes: In the target anatomical region, a first feature coefficient is determined based on the pixel difference between each feature candidate point and its nearest neighbor other feature candidate points, and a second feature coefficient is determined based on the Euclidean distance between each feature candidate point and its nearest neighbor other feature candidate points. The fluctuation coefficient of the pixel values ​​of all feature candidate points in the target anatomical region, the first feature coefficient, and the second feature coefficient are fused to obtain the state influence factor of the target anatomical region.

[0044] Specifically, within the target anatomical region, each feature candidate point is paired with its nearest other feature candidate point to form a feature point pair. A feature point pair represents pairing a feature candidate point with its nearest unvisited feature candidate point, forming a series of ordered point pairs. Optionally, starting from any feature candidate point within the target anatomical region, such as the feature candidate point closest to the image origin, the nearest unvisited feature candidate point to the current feature candidate point is found by calculating the Euclidean distance between the two feature candidate points, thus forming a feature point pair.

[0045] As a concrete example, for any pair of feature points, the absolute value of the difference in pixel values ​​between the two candidate feature points is calculated as the first feature coefficient, and the Euclidean distance between the two candidate feature points is calculated as the second feature coefficient. The first feature coefficient characterizes the gradient of echo intensity variation among candidate feature points within the same anatomical region. True microcalcifications have similar gray levels and small gradients, while spurious peaks caused by noise exhibit drastic gray level jumps. The second feature coefficient characterizes the spatial clustering density of candidate points within the anatomical region; denser calcifications result in shorter distances, while scattered noise leads to dispersed distances. The standard deviation of pixel values ​​for all candidate feature points within the target anatomical region is calculated as the fluctuation coefficient, reflecting the consistency or dispersion of candidate point pixels throughout the entire anatomical region.

[0046] Furthermore, the ratio of the first and second characteristic coefficients of the same feature point pair is calculated, and the product of the sum of these ratios for all feature point pairs and the fluctuation coefficient is determined as the state influence factor. This embodiment integrates the feature analysis results from three aspects through a product, which can significantly enhance the response when spatial compactness, grayscale smoothness, and overall fluctuation are simultaneously abnormal, thereby improving the sensitivity for discriminating true calcified regions.

[0047] Preferably, if the nth anatomical region in the t-th ultrasound image of any patient is taken as the target anatomical region, then the method for obtaining the state influencing factor of the target anatomical region can be expressed by the formula:

[0048] in, The state influence factor of the target anatomical region is represented, which is the state influence factor of the nth anatomical region in the tth ultrasound image, where t represents the tth ultrasound image, n represents the nth anatomical region, and x represents the xth feature candidate point. The fluctuation coefficient represents the target anatomical region, which is also the standard deviation of the pixel values ​​of all feature candidate points within the target anatomical region. This represents the pixel value of the x-th feature candidate point within the target anatomical region. This represents the pixel value of the feature candidate point (i.e., the (x+1)th feature candidate point) within the target anatomical region that is closest to the xth feature candidate point. This represents the Euclidean distance between the x-th and (x+1)-th feature candidate points within the target anatomical region, which is also known as the second feature coefficient. Denotes the first characteristic coefficient. This represents the total number of candidate feature points within the target anatomical region.

[0049] It should be noted that, specifically, when At this time, it is impossible to perform distance and pixel difference analysis between feature candidate points, and the number of feature candidate points in the region is small, making it less likely to exhibit high calcification features. In this embodiment, the state influence factor of the target anatomical region is set to 0. Since the feature candidate points in the same anatomical region are all different pixel positions, the Euclidean distance in the denominator cannot be 0 in this implementation scenario.

[0050] A larger fluctuation coefficient indicates that the echo intensity among multiple candidate feature points within the target anatomical area is extremely uneven. Microcalcifications appear as scattered, gravel-like points, with varying particle densities leading to uneven echoes; conversely, large, uniform, dense tissues or large calcification patches tend to have smaller fluctuation coefficients.

[0051] The smaller the distance but the greater the grayscale difference, the larger the ratio of the first feature coefficient to the second feature coefficient, which essentially indicates a maximum spatial grayscale gradient. This suggests a dramatic jump in brightness within a very small spatial range, consistent with the aggregated topology where calcification points are extremely bright, while the tiny gaps between points are relatively dark. The multiplication of these two factors reflects that the anatomical region possesses both global granular inhomogeneity and localized high spatial contrast.

[0052] A higher value for the state influence factor indicates a coarser local texture and more dramatic contrast, which closely matches the ultrasound physical characteristics of densely clustered microcalcifications and can effectively distinguish between ordinary tissue and large benign calcifications. It should be noted that the state influence factor characterizes the sum of squares of the degree of grayscale variation per unit distance. Although this parameter has the dimension of grayscale squared per pixel distance, its value reflects the contrast intensity of pixel distance space and has practical physical significance. Furthermore, this parameter is used for subsequent inter-frame image comparison and has undergone normalization processing; therefore, the dimension of this parameter does not affect the feature analysis results of this embodiment.

[0053] Step S203: Based on the comprehensive distribution of pixel values ​​of all feature candidate points in the target anatomical region and the area distribution of the target anatomical region, the ultrasound response factor of the target anatomical region is obtained.

[0054] In breast ultrasound images, microcalcifications, besides their spatial clustering and contrast, also possess an important acoustic property: due to the strong reflection and attenuation of ultrasound waves by calcifications, the echo energy contributed by calcification points per unit area is significantly higher than that of normal glandular tissue. Based on this, this embodiment quantifies the intensity density by analyzing the cumulative pixel values ​​and area of ​​all candidate feature points within the anatomical region.

[0055] Preferably, in one embodiment of the present invention, the method for obtaining the ultrasound response factor of the target anatomical region specifically includes: Based on the pixel values ​​of all feature candidate points in the target anatomical region, the cumulative pixel value of the target anatomical region is determined; the ratio between the cumulative pixel value and the area of ​​the target anatomical region is used as the ultrasound response factor of the target anatomical region.

[0056] As a concrete example, the sum of the pixel values ​​of all candidate feature points in the target anatomical region is used as the cumulative pixel value of the target anatomical region. Preferably, the total number of all pixels in the target anatomical region is used as the area of ​​the target anatomical region. The ultrasound response factor of the target anatomical region characterizes the pixel grayscale accumulation per unit area, reflecting the overall echo intensity within the anatomical region.

[0057] Step S204: Based on the consistency of characteristic changes between the differences in state influencing factors and ultrasound response factors between the target anatomical region and the reference anatomical region, the calcification characteristic index of the target anatomical region is obtained.

[0058] Calcification characteristic index is a measure of the degree of synergistic resonance between the fluctuation of the overall echo intensity and the perturbation of local texture roughness in the same anatomical region in a dynamic ultrasound image sequence.

[0059] Specifically, the difference in state influencing factors between the target anatomical region and the reference anatomical region is taken as the first difference, and the difference in ultrasound response factors between the target anatomical region and the reference anatomical region is taken as the second difference; based on the consistency of the change trend between the first difference and the second difference, the change direction coefficient is determined; based on the consistency of the change amplitude between the first difference and the second difference, the change amplitude coefficient is determined; the change direction coefficient and the change amplitude coefficient are fused to obtain the calcification characteristic index of the target anatomical region.

[0060] As a specific example, this embodiment uses the nth anatomical region of the preceding ultrasound image adjacent to the t-th ultrasound image of the same patient as the reference anatomical region. The method for obtaining the calcification feature index of the target anatomical region can be expressed by the formula:

[0061] in, The calcification characteristic index represents the calcification characteristic index of the target anatomical region, that is, the calcification characteristic index of the nth anatomical region in the tth ultrasound image. This indicates the first difference corresponding to the target anatomical region. This indicates the second difference corresponding to the target anatomical region. For symbolic functions, To find the minimum value function, To find the maximum value function, The preset minimum positive number is 0.1 in this embodiment. The purpose of this parameter is to prevent the formula from becoming invalid due to a denominator of 0.

[0062] ,in Let represent the ultrasound response factor of the nth anatomical region in the t-th ultrasound image. Let represent the ultrasound response factor of the nth anatomical region in the (t-1)th ultrasound image. ,in This represents the state influence factor of the nth anatomical region in the t-th ultrasound image. This represents the state influence factor of the nth anatomical region in the (t-1)th ultrasound image. This represents the normalization function.

[0063] This represents the direction coefficient of change. When the probe is slightly moved across the real microcalcification group, the acoustic response intensity will increase, which will inevitably be accompanied by a sharp increase in the contrast between particles. This item is used to determine whether the trends of the first difference and the second difference are in the same direction. If they increase or decrease together, it is +1, which is consistent with the physical characteristics of calcification; if one increases and the other decreases, it is -1, which is usually an artifact of overall whitening caused by the change of angle in uniform tissue.

[0064] This represents the amplitude coefficient, which measures the synchronicity of the disturbance amplitudes of the two variables by calculating the ratio of their absolute values. If the overall echo brightens but the texture contrast remains largely unchanged (a single variable abrupt change), this ratio tends to 0. When both variables experience simultaneous, drastic fluctuations of the same magnitude, this ratio approaches 1.

[0065] Calcification characteristic indicators The value range is limited to [-1, 1]. The closer the value is to 1, the more closely the temporal changes in brightness and roughness are linked and coordinated, indicating that the area is highly likely to be a real, dense microcalcification. A negative value indicates that the changes in the ultrasonic response factor (echo intensity) and the state influence factor (texture contrast) are opposite, and their characteristic changes are asynchronous. Real calcification should increase or decrease together when the probe is slightly moved. This reverse fluctuation does not conform to the physical response law of real microcalcification aggregation. Therefore, a negative value indicates that the area is more likely to be an artifact, noise interference, or non-calcified homogeneous tissue.

[0066] It should be noted that when or According to the above calculation formula, the results of the calcification characteristic indicators are all 0. Specifically, when... hour, and If both are 0, the calcification feature index is directly set to 0, without using the above formula, to avoid a denominator of 0. When the result of the calcification feature index is 0, it means that the anatomical region does not show coordinated temporal fluctuations between adjacent frames, that is, the echo intensity and texture contrast either do not change, or only one changes while the other remains unchanged. In this case, it does not have the typical dynamic characteristics of microcalcification aggregation.

[0067] In order to eliminate and The dimensional differences between the data are preserved while maintaining the direction of change of the data differences. For the same anatomical region of a single patient, the first difference of the nth anatomical region in all ultrasound images under dynamic ultrasound image sequence is extracted. Form the first set, and all second differences A second set is formed, and the elements in both sets are divided by the maximum absolute value of that set to achieve normalization. In other words, the first difference and the second difference in this embodiment are the results of normalizing the first difference and the second difference. Specifically, when the maximum absolute value of the corresponding set is 0 because all elements in the first or second set are 0, the above normalization division operation is not performed; instead, the derivation result of the corresponding difference in that set is directly assigned the value 0.

[0068] Step S300: Obtain mammograms for each historical patient and construct a probability density image based on the feature density of the mapping results of calcifications in the mammograms onto the ultrasound images.

[0069] To establish a benchmark labeling system for cross-modal feature mapping between mammography and ultrasound, probability density modeling needs to be completed by leveraging the high-precision microcalcification annotation advantage of mammography images. Mammography images can clearly capture microcalcification sites and are the gold standard for clinical calcification detection. This invention considers accurately locating calcification sites through mammography images, then mapping and matching them to the corresponding ultrasound space. Based on the spatial distribution density of the mapped sites, the probability of calcification in different regions can be quantified, generating a continuous probability representation image.

[0070] Therefore, by acquiring mammograms for each historical patient and using the spatial mapping of mammogram calcifications to ultrasound images, the characteristic density distribution patterns of these mapped points are statistically analyzed to construct a standardized microcalcification probability density image. Under ideal conditions, this probability density image can accurately reproduce the degree and probability of calcification aggregation in each anatomical region, compensating for the locational uncertainties introduced by registration, and replacing traditional discrete single-point annotation methods. This provides a reliable gold standard reference for subsequent anatomical region feature association analysis and segmentation network training.

[0071] Preferably, in one embodiment of the present invention, the specific method for constructing a probability density image includes: The molybdenum target image includes calcification points. Using a spatial registration algorithm, the molybdenum target image is mapped to the space of the ultrasound image. In the mapping result, each calcification point is used as the center of the Gaussian kernel. Based on the cumulative sum of the Gaussian kernel function values ​​from each pixel point to each Gaussian kernel center, the probability density of each pixel point is determined. Based on the probability density of all pixels points in the mapping result, a probability density image is constructed.

[0072] As a concrete example, it should be understood that mapping a mammogram to the ultrasound image space includes the precise two-dimensional coordinates of all calcifications in the static, high-resolution mammogram. Then, a non-rigid deformable registration algorithm (such as the DIR algorithm based on a breast biomechanical model) can be used, with anatomical structures such as the breast contour, chest wall line, and nipple position as anchor points, to calculate the deformation field from the mammogram to the ultrasound image (or ultrasound 3D space). Through this deformation field, the precise two-dimensional coordinates of each calcification on the mammogram are mapped to the corresponding coordinate position in the ultrasound image, enabling subsequent kernel density estimation to generate a probability density image consistent with the pixel space of the ultrasound image. Essentially, this process transforms the anatomical space of the mammogram to the imaging space of the ultrasound, achieving cross-modal pixel-level correspondence.

[0073] In the mapping results, all pixels on the molybdenum target image, including calcification points, are mapped to the ultrasound image space. A two-dimensional Gaussian kernel function is placed with each mapped calcification point as the center to perform kernel density estimation. This method is a well-known technique and will not be described in detail here.

[0074] For any pixel in the mapping result, calculate the sum of the Gaussian kernel function values ​​from the pixel to the centers of all Gaussian kernels, and then normalize the sum to obtain the probability density of that pixel. The probability density represents the density and distance of the mapped calcification points within the local region centered on that pixel; the closer the distance and the more center points there are, the greater the probability density.

[0075] The normalization process involves using the maximum sum of the Gaussian kernel function values ​​corresponding to all pixels as the denominator and the sum of the Gaussian kernel function values ​​corresponding to each pixel as the numerator. The ratio of the numerator to the denominator is then used as the probability density for that pixel. Based on the mapping result, an initial probability density image is generated for all pixels in the ultrasound image space. The pixel value represents the probability of calcification at that location and is directly used as the gold standard label for the subsequent ultrasound segmentation network. It should be understood that the probability density image and the ultrasound image have the same region location correspondence.

[0076] The initial probability density images generated above, located in the respective ultrasound spaces of different patients, are spatially normalized and transformed to the same anatomical template to obtain the final probability density image. This ensures that the same anatomical region in different patients has a spatial correspondence. The method for spatial registration of mammogram images in this embodiment is consistent with the aforementioned spatial registration process for ultrasound images, and will not be repeated here. It should be noted that the spatial registration operation in this embodiment includes two aspects: firstly, intra-individual cross-modal registration, mapping the coordinates of precisely labeled calcification points in the mammogram image to the patient's own ultrasound image space; secondly, individual-to-standard template registration, registering the initial probability density image to the same anatomical template.

[0077] Step S400: Based on the correlation between the calcification feature index and the corresponding probability density of each anatomical region in the ultrasound images of each historical patient, obtain the feature association confidence. Combine the feature association confidence and the calcification feature index corresponding to each historical patient in all ultrasound images to train the image segmentation network.

[0078] To establish the intrinsic correlation between breast ultrasound calcification features and the actual calcification distribution on mammography, and to optimize the training of a high-precision image segmentation network, it is necessary to quantify the degree of matching between the two and integrate multi-sample temporal features for model learning. Traditional breast calcification segmentation networks rely solely on single ultrasound image features for training, failing to utilize mammography images as a gold standard to construct feature associations. This results in an inability to uncover the correspondence between changes in ultrasound texture in anatomical regions and the actual distribution of microcalcifications. Furthermore, they are susceptible to ultrasound speckle noise and glandular artifacts, lack cross-patient universal feature support, and have limited generalization ability.

[0079] In this embodiment of the invention, calcification feature indicators of each anatomical region characterize the texture, grayscale, and calcification aggregation characteristics of ultrasound images, while probability density images reflect the confidence level of the existence of real microcalcifications mapped by the mammogram. The correlation features between the two can quantify the matching confidence level between regional features and the real calcification distribution, forming feature correlation confidence. Furthermore, by fusing calcification feature indicators from multiple frames of ultrasound images of the same patient, temporal fluctuation features can be incorporated to further enhance the model's ability to learn microcalcification features.

[0080] Based on this, the feature association confidence is calculated by relying on the correlation features between historical patients' anatomical calcification feature indicators and corresponding probability densities. Combined with the calcification feature indicators of full-sequence ultrasound images, iterative training of the image segmentation network is completed. Under ideal training conditions, the network can learn the intrinsic mapping relationship between ultrasound texture changes and the presence of microcalcifications in different anatomical regions, effectively shielding ultrasound artifacts and noise interference. After training, the segmentation network can accurately identify and segment microcalcifications using only a single frame of ultrasound image, without the assistance of mammogram images, significantly improving the clinical applicability and segmentation accuracy of the model.

[0081] Preferably, such as Figure 3 As shown, in one embodiment of the present invention, the method for obtaining feature association confidence specifically includes: Step S401: Take any anatomical region as the calibration anatomical region. For any historical patient, obtain the average level of calcification characteristic indicators of the calibration anatomical region in all ultrasound images, and use it as the calcification balance indicator of the historical patient in the calibration anatomical region.

[0082] Specifically, any anatomical region in the standard anatomical template is used as the calibration anatomical region. By analyzing the average level of calcification characteristic indicators of the calibration anatomical region in the dynamic ultrasound image sequence of a single historical patient, a stable scalar representing the overall condition of that patient can be obtained. More specifically, for any historical patient, the average value of the calcification characteristic indicators of the calibration anatomical region in all ultrasound images is used as the calcification balance index of that historical patient in the calibration anatomical region.

[0083] The calcification equilibrium index reflects the average intensity level of dynamic fluctuations in ultrasound texture caused by microcalcification accumulation within a specific anatomical region during continuous ultrasound scans of a single historical patient. A higher calcification equilibrium index value indicates more significant and stable spatiotemporal fluctuations caused by microcalcifications in the ultrasound images of that anatomical region; a lower value indicates that the ultrasound appearance of that anatomical region is close to normal glandular tissue or exhibits weak fluctuations. This index eliminates transient noise and inter-frame random differences, providing a unified individual-level characteristic benchmark for statistical correlation analysis across patients and regions.

[0084] Step S402: For any pixel, construct a calcification balance index sequence from the calcification balance index of each historical patient in the anatomical region where the pixel is located; construct a probability density sequence from the probability density image of each historical patient at the pixel location.

[0085] It should be understood that the ultrasound images and probability density images of different patients have been registered, and the correspondence of the pixel locations is clear. Therefore, for any pixel in the ultrasound image, the calcification balance index and probability density of each historical patient at that pixel location can be obtained. It should be understood that the calcification balance index of all pixels belonging to the same anatomical region is equal, but the probability density may not be equal. Therefore, feature analysis is performed through single pixels.

[0086] Step S403: Calculate the correlation coefficient between the calcification equalization index sequence and the probability density sequence to obtain the feature correlation confidence of any pixel.

[0087] The calcification equilibrium index is derived from the dynamic temporal characteristics of ultrasound images, reflecting the patient's historical ultrasound findings. The probability density, derived from the gold standard of mammogram mapping, reflects the objective probability of actual calcification at that pixel location in that historical patient. For the same pixel location within the same anatomical region, individual differences exist between patients; some patients have strong ultrasound features and a high probability of calcification, while others have weak ultrasound features and a low probability of calcification. If these two sets of sequences (across patients) show a highly consistent trend—that is, patients with strong ultrasound features also have a high probability of mammogram calcification, and patients with weak ultrasound features also have a low probability of calcification—then it indicates that the dynamic ultrasound features at that pixel location can indeed effectively predict the presence of actual calcification.

[0088] In this embodiment of the invention, for any pixel in an ultrasound image, the Pearson correlation coefficient between the calcification equalization index sequence and the probability density sequence is calculated as the feature association confidence score. The feature association confidence score characterizes the cross-patient statistical correlation strength between the calcification equalization index extracted from the ultrasound image and the calcification probability density in the gold standard mammography at a specific pixel location. A stronger correlation indicates that the ultrasound features at that pixel location are more suitable as a priori confidence region for the subsequent ultrasound segmentation network; a weaker correlation indicates that the ultrasound features at that pixel location have a weaker ability to predict calcification.

[0089] It should be understood that the feature association confidence score is a statistical indicator pre-calculated and fixed based on historical patient population data. This indicator is calculated only from historical training set data and serves as a static, resident population experience dictionary system after model deployment. In subsequent use, correlation analysis can be periodically (e.g., every 1000 new calcified patients) to update the indicator, allowing the dictionary to adapt to equipment differences or population feature drift, but it remains unchanged during a single clinical inference process. Its core meaning is the reliability of the ultrasound dynamic fluctuation features at each pixel location in the ultrasound image in predicting the presence of actual calcification. The higher the confidence score, the more consistent the spatiotemporal variation of the ultrasound texture at that location is with the probability of calcification, and the more suitable it is as a reliable prior to guide the segmentation network.

[0090] It should be noted that since the calcification probability density and ultrasound calcification characteristic indicators of the historical patient population are derived from real clinical samples, there are natural individual differences between different patients. Therefore, the standard deviation of the calcification equilibrium index sequence and the probability density sequence cannot be zero in actual calculations, and the denominator of the Pearson correlation coefficient is always valid.

[0091] Furthermore, in one embodiment of the present invention, the specific method for training an image segmentation network includes: The first step is to use the product of the calcification balance index of the anatomical region where the pixel is located and the feature association confidence score for each historical patient as the calcification significance of the pixel in each historical patient.

[0092] The physical meaning of calcification significance lies in the fact that the confidence level verified by the population statistics is used to weight the dynamic features of individual ultrasound. Only pixels that simultaneously satisfy the conditions of significant individual fluctuation and high correlation with calcification in historical data will obtain high significance. The significance of pixels with strong individual fluctuation but low confidence (such as areas where artifacts are common) or areas with high confidence but weak individual fluctuation is suppressed.

[0093] The second step is to construct a calcification feature image for each historical patient based on the pixel position of the ultrasound image, using the calcification significance of each pixel in each historical patient.

[0094] For each historical patient, the calcification significance of all pixels is arranged according to the spatial location of the ultrasound image to generate a calcification feature image with the exact same size as the original ultrasound image. Each pixel value in the calcification feature image represents the calcification significance at its corresponding location. The calcification feature image numerically integrates the patient's own dynamic fluctuation intensity with the population's prior confidence level, exhibiting highly significant clustered regions in spatial distribution. These regions are statistically most likely to be the locations where microcalcifications actually exist.

[0095] The third step involves using the calcification feature image and corresponding ultrasound image of each historical patient as input, and constructing a cross-entropy loss function based on the pixel difference between the calcification feature image and the corresponding probability density image of each historical patient. This cross-entropy loss function is then used to supervise the training process of the image segmentation network.

[0096] The calcification feature image of each historical patient and the corresponding original ultrasound image are used as input to the image segmentation network. Specifically, the calcification feature image of each historical patient and the corresponding original ultrasound image are concatenated along the channel dimension to form a dual-channel input tensor, which is then input to the image segmentation network (such as U-Net). The network outputs a prediction probability map with the same resolution as the input image. Each pixel value in the prediction probability map represents the probability that the network predicts the presence of a microcalcification at that pixel location. During training, the prediction error is calculated using the mammogram mapping probability density image corresponding to each historical patient. Specifically, a pixel-wise binary cross-entropy loss function is used.

[0097] in, Represents the loss function. This represents the total number of pixels in the ultrasound image, and M represents the total number of historical patients involved in the training. This represents the probability density of the j-th pixel in the probability density image of the m-th historical patient. This represents the predicted probability value of the j-th pixel in the predicted probability map output by the m-th historical patient. This is a preset, extremely small positive number, used only to ensure the validity of logarithmic operations and without affecting the actual meaning of the loss value. Its value can be set to... .

[0098] The loss function is minimized using gradient descent, and the network parameters are updated via backpropagation, enabling the network to accurately segment the spatial distribution of microcalcifications from the calcification feature image and the original ultrasound image. The summation operation covers all patients and all pixels, ensuring that the model fully learns the common features in the group data while avoiding overfitting due to single-frame noise.

[0099] The essence of this training process is to enable the network to learn the mapping relationship between high-significance regions in calcification feature images and high-probability regions in the gold standard. This allows the network to accurately segment the spatial distribution of microcalcifications during actual inference, based solely on the calcification feature images of the patient to be analyzed and the original ultrasound images. After training is completed, the network parameters are fixed, and it can then be used in the clinical inference stage of step S500.

[0100] Step S500: Based on the calcification feature index and the feature association confidence level corresponding to the patient in all ultrasound images, the image segmentation network is used to obtain the segmentation result of the calcification points on the mammogram.

[0101] Specifically, for any pixel, the average level of calcification feature indicators of the anatomical region in which the pixel is located in all ultrasound images of the patient to be analyzed is obtained as the calcification balance index of the patient to be analyzed at that pixel; based on the product between the calcification balance index of the patient to be analyzed at each pixel and the feature association confidence, a calcification feature image of the patient to be analyzed is constructed; the calcification feature image and the corresponding ultrasound image are input into an image segmentation network to obtain the segmentation result of the calcification point of the mammogram.

[0102] In practical clinical applications, there is no corresponding mammogram as the gold standard for the patients to be analyzed, so the mammogram probability density map cannot be directly reused as in the training phase. Instead, the dynamic characteristics of the individual patient and the prior knowledge of the group are weighted pixel-by-pixel to generate a calcification feature image spatially aligned with the original ultrasound image. This image, along with the original ultrasound image, serves as the input to the segmentation network (stitched along the channel dimension) to obtain the probability representation of calcification points for the patient to be analyzed based solely on the ultrasound image.

[0103] It should be noted that the methods for obtaining the ultrasound images, calcification equilibrium indicators, and calcification feature images of the patient to be analyzed are exactly the same as those for obtaining the corresponding parameters of the historical patients mentioned above, and will not be repeated here. The calcification feature images of the patient to be analyzed are input into a trained image segmentation network, and the output is a predicted probability map of the patient to be analyzed. This output can then be directly provided to clinicians as an auxiliary diagnostic basis. By combining the patient's clinical history, physical examination, and the highlighted information in the predicted probability map, doctors can quickly locate suspected microcalcification lesions and assess the degree of aggregation, distribution range, and density trend of calcification points.

[0104] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for identifying and segmenting mammary calcifications based on breast ultrasound images, characterized in that, The method includes: Acquire dynamic ultrasound image sequences of the patient to be analyzed and each historical patient in the database, the dynamic ultrasound image sequences comprising several ultrasound images and each of their anatomical regions; Feature candidate points of each anatomical region in the ultrasound image are extracted. Based on the pixel differences and distance distribution between the feature candidate points in the anatomical region, and combined with the pixel feature distribution of all feature candidate points, the calcification feature index of each anatomical region in the ultrasound image is obtained. Obtain mammograms for each historical patient and construct a probability density image based on the feature density of the mapping results of calcifications in the mammograms onto the ultrasound images; The feature association confidence is obtained by considering the correlation between the calcification feature index and the corresponding probability density of each anatomical region in the ultrasound images of each historical patient. The image segmentation network is then trained by combining the feature association confidence and the calcification feature index corresponding to each historical patient in all ultrasound images. Based on the calcification feature indices and feature association confidence scores corresponding to the patient in all ultrasound images, the image segmentation network is used to obtain the segmentation results of calcification points on the molybdenum target.

2. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 1, characterized in that, The step of obtaining calcification feature indices for each anatomical region in the ultrasound image based on the pixel differences and distance distribution between feature candidate points within the anatomical region, combined with the pixel feature distribution of all feature candidate points, specifically includes: For any ultrasound image of any patient, any anatomical region is taken as the target anatomical region; the corresponding anatomical region in the ultrasound images of the time sequence adjacent to the target anatomical region is taken as the reference anatomical region. Based on the pixel differences and distance distribution between each feature candidate point in the target anatomical region and its nearest neighbor other feature candidate points, and combined with the pixel fluctuation characteristics of all feature candidate points in the target anatomical region, the state influence factor of the target anatomical region is obtained. The ultrasound response factor of the target anatomical region is obtained based on the comprehensive distribution of pixel values ​​of all feature candidate points in the target anatomical region and the area distribution of the target anatomical region. Based on the consistency of characteristic changes between the differences in state influencing factors and ultrasound response factors between the target anatomical region and the reference anatomical region, calcification characteristic indicators of the target anatomical region are obtained.

3. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 2, characterized in that, The process involves obtaining the state influence factor of the target anatomical region based on the pixel differences and distance distribution between each feature candidate point in the target anatomical region and its nearest neighbor, combined with the pixel fluctuation characteristics of all feature candidate points in the target anatomical region. Specifically, this includes: In the target anatomical region, a first feature coefficient is determined based on the pixel difference between each feature candidate point and its nearest neighbor other feature candidate points, and a second feature coefficient is determined based on the Euclidean distance between each feature candidate point and its nearest neighbor other feature candidate points; The fluctuation coefficient, first feature coefficient, and second feature coefficient of the pixel values ​​of all feature candidate points in the target anatomical region are fused to obtain the state influence factor of the target anatomical region.

4. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 2, characterized in that, The process of obtaining the ultrasound response factor of the target anatomical region based on the comprehensive distribution of pixel values ​​of all feature candidate points in the target anatomical region and the area distribution of the target anatomical region specifically includes: Based on the pixel values ​​of all feature candidate points in the target anatomical region, determine the cumulative pixel value of the target anatomical region; The ratio between the cumulative pixel value and the area of ​​the target anatomical region is used as the ultrasound response factor of the target anatomical region.

5. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 2, characterized in that, The calcification characteristic indicators of the target anatomical region are obtained based on the consistency of characteristic changes between the differences in state influencing factors and ultrasound response factors between the target anatomical region and the reference anatomical region. Specifically, these indicators include: The difference in state influencing factors between the target anatomical region and the reference anatomical region is taken as the first difference, and the difference in ultrasound response factors between the target anatomical region and the reference anatomical region is taken as the second difference. Based on the consistency of the changing trends between the first and second differences, the change direction coefficient is determined; based on the consistency of the changing magnitude between the first and second differences, the change magnitude coefficient is determined; the change direction coefficient and the change magnitude coefficient are fused to obtain the calcification characteristic index of the target anatomical region.

6. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 1, characterized in that, The feature association confidence level is obtained based on the correlation between the calcification feature indicators and the corresponding probability density of each anatomical region in the ultrasound images of each historical patient. Specifically, this includes: Using any anatomical region as the calibration anatomical region, for any historical patient, the average level of calcification characteristic indicators of the calibration anatomical region in all ultrasound images is obtained as the calcification balance index of the historical patient in the calibration anatomical region. For any pixel, construct a calcification equilibrium index sequence by the calcification equilibrium index of the anatomical region where the pixel is located for each historical patient; construct a probability density sequence by the probability density of the probability density image of each historical patient at the location of the pixel. Calculate the correlation coefficient between the calcification equalization index sequence and the probability density sequence to obtain the feature correlation confidence of any pixel.

7. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 6, characterized in that, The process of training an image segmentation network by combining feature association confidence scores and the calcification feature indices corresponding to each historical patient in all ultrasound images specifically includes: For any pixel, the product of the calcification balance index of the anatomical region where the pixel is located and the feature association confidence is used as the calcification significance of the pixel in each historical patient. A calcification feature image for each historical patient is constructed based on the pixel position of the ultrasound image, with each pixel representing the calcification significance for each historical patient. Using the calcification feature image and corresponding ultrasound image of each historical patient as input, a cross-entropy loss function is constructed based on the pixel difference between the calcification feature image and the corresponding probability density image of each historical patient. The training process of the image segmentation network is supervised by the cross-entropy loss function.

8. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 6, characterized in that, The step of obtaining the segmentation result of mammography calcification points using the image segmentation network based on the calcification feature indicators and feature association confidence scores corresponding to all ultrasound images of the patient to be analyzed includes: For any pixel, the average level of calcification feature index of the anatomical region in which the pixel is located in all ultrasound images of the patient to be analyzed is obtained as the calcification balance index of the patient at that pixel. Based on the product of the calcification balance index and the feature association confidence of each pixel of the patient to be analyzed, a calcification feature image of the patient to be analyzed is constructed. The calcification feature image and the corresponding ultrasound image are input into the image segmentation network to obtain the segmentation result of calcification points on the mammogram.

9. The method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 1, characterized in that, The extraction of feature candidate points for each anatomical region in the ultrasound image specifically includes: For any ultrasound image, pixels with a value greater than or equal to a preset highlight threshold are designated as highlight pixels. If the pixel value of a highlight pixel is the maximum value of all pixel values ​​in its neighborhood, then the highlight pixel is designated as a feature candidate point.

10. A method for identifying and segmenting mammary calcifications based on breast ultrasound images according to claim 1, characterized in that, The construction of a probability density image based on the feature density of calcification points mapped from the molybdenum target image to the ultrasound image specifically includes: The molybdenum target image includes calcification points. Using a spatial registration algorithm, the molybdenum target image is mapped to the space of the ultrasound image. In the mapping result, each calcification point is used as the center of the Gaussian kernel. Based on the cumulative sum of the Gaussian kernel function values ​​from each pixel to each Gaussian kernel center, the probability density of each pixel is determined. Based on the probability density of all pixels in the mapping result, a probability density image is constructed.

Images