Dynamic mammary gland density risk assessment method based on AI ultrasound

Abstract

The invention discloses a dynamic mammary gland density risk assessment method based on AI ultrasound, and belongs to the technical field of mammary gland risk assessment, and the method specifically comprises the steps: obtaining a multi-temporal mammary gland ultrasound image with a physiological cycle timestamp, and building a unified space-time coordinate system through the elastic registration of an anatomical structure reference point; gland morphological features are extracted through a deep convolutional neural network, and a spatial density feature map corresponding to each time node is generated; constructing a mammary gland tissue space-time evolution topological graph by taking the local features as vertexes and taking the deformation field and the feature similarity as edges; fusing space structure and time change information by using a space-time diagram convolutional neural network to obtain a dynamic evolution trajectory vector of hyperplasia and degeneration; and comparing the trajectory vector with a healthy population reference line, calculating a space-time deviation degree to obtain a dynamic risk coefficient, finally constructing a full life cycle prediction model, and outputting a dynamic mammary gland density risk assessment report.

Description

Technical Field

[0001] This invention relates to the field of breast risk assessment technology, specifically to a dynamic breast density risk assessment method based on AI ultrasound. Background Technology

[0002] Breast density is one of the core clinical indicators for assessing women's breast health and potential risk of breast cancer. With the deep integration of medical imaging technology and artificial intelligence, AI-based ultrasound imaging technology, due to its physical advantages such as no ionizing radiation, ease of operation, and real-time performance, has been widely applied in routine clinical breast screening and quantitative tissue density analysis. Existing AI-assisted assessment systems mainly utilize deep learning models (such as deep convolutional neural networks) to perform pixel-level processing on two-dimensional ultrasound cross-sectional images acquired in a single clinical session. By automatically extracting features such as echo grayscale, texture distribution, and anatomical edge contours within the image, it achieves precise segmentation of tissue structures such as the subcutaneous fat layer, glandular layer, and pectoralis major muscle of the breast. Based on this static image segmentation result, the system can objectively calculate the physical proportion of dense glandular tissue in the overall breast imaging area, and then automatically output the density grading index (such as BI-RADS grading) corresponding to the screening time, effectively improving the efficiency of radiologists' image interpretation and the objectivity of a single diagnosis.

[0003] However, most existing breast density assessment techniques rely on static cross-sectional image analysis at a single time point, lacking a longitudinal dynamic monitoring mechanism for the evolution of breast density over time. Breast tissue is influenced by endocrine and physiological cycles, exhibiting non-linear temporal changes in density. Traditional single-assessment methods only reflect the tissue state at a particular instant, failing to capture the rate and trend of density changes and ignoring the dynamic remodeling process of the breast stroma and glandular tissue. This static assessment model results in insufficient sensitivity to subtle proliferative or degenerative trends within the breast, masking the trajectory of potential physiological abnormalities. Consequently, the final risk assessment results are lagging and one-sided, failing to meet the clinical need for accurate prediction of breast health status throughout the entire lifespan. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic breast density risk assessment method based on AI ultrasound, thereby addressing the problems in the background art:

[0005] The objective of this invention can be achieved through the following technical solutions: A dynamic breast density risk assessment method based on AI ultrasound includes the following steps: S1: Acquire multi-temporal ultrasound image sequences with physiological cycle timestamps, extract anatomical structural reference points from the multi-temporal ultrasound image sequences, perform elastic registration, and establish a unified spatial coordinate system across time dimensions; S2: Input the registered multi-temporal ultrasound image sequence into a deep convolutional neural network to extract the morphological features of the mammary glands within each physiological cycle time stamp and generate a spatial density feature map corresponding to each time node. S3: Using local feature regions within the spatial density feature map as topological vertices, and utilizing the anatomical deformation field and feature similarity between adjacent physiological cycle timestamps as topological edges, a spatiotemporal evolution topological map mapping the nonlinear changes in breast tissue morphology is constructed. S4: Input the spatiotemporal evolution topology graph into the spatiotemporal graph convolutional neural network, simultaneously aggregate the structural features of the spatial dimension and the state transfer weights of the temporal dimension, and output the dynamic evolution trajectory vector representing the hyperplasia and degeneration process inside the breast. S5: Input the extracted evolution trajectory vector into the preset risk prediction classifier, calculate the spatiotemporal deviation between the vector and the standard physiological evolution baseline constructed based on the healthy population cohort, and then obtain the dynamic risk assessment coefficient of the subject. S6: Configure a full life-cycle breast health prediction model based on the calculated dynamic risk assessment coefficient, and output a dynamic breast density risk assessment report.

[0006] As a further aspect of the present invention: in step S1, the process of establishing a unified spatial coordinate system across the time dimension is as follows: Breast ultrasound images of subjects at different stages were acquired, and the physiological cycle timestamp data corresponding to the acquisition time were embedded into the breast ultrasound images to generate a multi-temporal ultrasound image sequence with physiological cycle timestamps. The feature recognition algorithm was used to analyze the multi-phase ultrasound image sequence with physiological cycle time stamps, locate the junction between the nipple and chest wall tissue, and extract the anatomical reference points of the multi-phase ultrasound image sequence. Based on the extracted anatomical reference points, a spatial deformation vector is calculated. This spatial deformation vector is then used to perform elastic registration on multi-temporal ultrasound image sequences, establishing a unified spatial coordinate system across time dimensions.

[0007] As a further aspect of the present invention: in step S2, the process of generating the spatial density feature map corresponding to each time node is as follows: The registered multi-temporal ultrasound image sequence is input into a deep convolutional neural network. The convolutional layers of the deep convolutional neural network are used to perform a sliding mapping operation on the image pixel matrix, and the initial feature tensor data is output.

[0008] Based on the initial feature tensor data of the output, a downsampling layer of a deep convolutional neural network is used to perform dimensionality reduction processing to extract the morphological features of mammary glands within each physiological cycle time stamp.

[0009] The morphological features of breast glands within each physiological cycle time stamp are input into the upsampling layer of a deep convolutional neural network to perform spatial distribution reconstruction, generating a spatial density feature map corresponding to each time node.

[0010] As a further aspect of the present invention: in step S3, the process of constructing a spatiotemporal evolution topology map that maps the nonlinear changes in breast tissue morphology is as follows: The spatial density feature map corresponding to each time node is divided into image blocks using a region segmentation algorithm. Local feature regions inside the spatial density feature map contained in each image block are extracted and defined as topological vertices. Extract the spatial coordinate matrix of the topological vertices corresponding to adjacent physiological cycle timestamps, and perform displacement mapping operation based on the coordinate offset vector of the spatial coordinate matrix to generate the anatomical structure deformation field between adjacent physiological cycle timestamps. Extract the pixel intensity distribution matrix of the corresponding topological vertices between adjacent physiological cycle timestamps, calculate the difference value of the pixel intensity distribution matrix through the distance metric function, and generate the feature similarity between adjacent physiological cycle timestamps; The generated anatomical deformation field and feature similarity are spliced ​​to generate a connection weight matrix. The connection weight matrix is ​​used as the topological edge connecting the topological vertices to construct a spatiotemporal evolution topological graph that maps the nonlinear changes in breast tissue morphology.

[0011] As a further aspect of the present invention: the process of generating the anatomical deformation field between adjacent physiological cycle timestamps is as follows: Read the pixel grid positions of topological vertices within adjacent physiological cycle timestamps, convert them into numerical sequences, and extract the spatial coordinate matrix of the corresponding topological vertices between adjacent physiological cycle timestamps; The extracted spatial coordinate matrix is ​​processed by matrix subtraction, and the coordinate difference between the corresponding vertex positions before and after the time nodes is calculated to generate the coordinate offset vector of the spatial coordinate matrix. Based on the coordinate offset vector of the spatial coordinate matrix, the spatial transformation function is called to perform displacement mapping operation, aggregate spatial deformation mesh data, and generate the anatomical structure deformation field between adjacent physiological cycle timestamps.

[0012] As a further aspect of the present invention: in step S4, the process of outputting the dynamic evolution trajectory vector characterizing the hyperplasia and degeneration process within the mammary gland is as follows: The spatiotemporal evolution topology graph is input into the spatiotemporal graph convolutional neural network. The spatial convolutional layer of the spatiotemporal graph convolutional neural network is used to traverse the topology vertices along the topology edges to extract the structural features of the spatial dimension. The temporal convolutional layer of the spatiotemporal graph convolutional neural network is used to extract the node feature differences of adjacent physiological cycle timestamps, generate state transit weights in the time dimension, and perform multiplication and addition operations on the state transit weights and the structural features in the spatial dimension to complete synchronous aggregation. The feature tensor data from the synchronous aggregation output is input into the fully connected layer of the spatiotemporal graph convolutional neural network to perform dimensionality reduction mapping, and outputs a dynamic evolution trajectory vector representing the hyperplasia and degeneration process inside the breast.

[0013] As a further aspect of the present invention: in step S5, the process of obtaining the dynamic risk assessment coefficient of the subject is as follows: The extracted evolution trajectory vector is input into a preset risk prediction classifier, the standard physiological evolution baseline constructed based on the healthy population cohort is read, and the feature distribution matrix of the standard physiological evolution baseline is extracted. Using the spatial distance calculation function of the risk prediction classifier, the spatial feature parameters of the evolution trajectory vector are compared with the spatial reference value of the feature distribution matrix to calculate the spatial deviation value. Using the time series alignment algorithm of the risk prediction classifier, the time feature parameters of the evolution trajectory vector are compared with the time reference values ​​of the feature distribution matrix to calculate the time deviation value; The spatial deviation value and the temporal deviation value are weighted and fused to generate the spatiotemporal deviation between the evolution trajectory vector and the standard physiological evolution baseline, thereby obtaining the dynamic risk assessment coefficient of the subject.

[0014] As a further aspect of the present invention: in step S6, the process of outputting the dynamic breast density risk assessment report is as follows: Configure a full life cycle breast health prediction model based on the calculated dynamic risk assessment coefficient, call the regression function of the full life cycle breast health prediction model to perform fitting operation, and generate the evolution trend of breast tissue morphology. The dynamic risk assessment coefficient is input into the classification node of the full life cycle breast health prediction model, the dynamic risk assessment coefficient is compared with the preset risk threshold value, and the disease warning status is output. The generated breast tissue morphological evolution trend and the disease warning status are imported into a preset layout framework and merged and spliced ​​to output a dynamic breast density risk assessment report that covers both the breast tissue morphological evolution trend and the disease warning status.

[0015] The beneficial effects of this invention are: This invention establishes a unified spatial coordinate system across time dimensions by acquiring multi-temporal ultrasound image sequences with physiological cycle timestamps and performing elastic registration, overcoming the spatial physical limitations of traditional single static assessments. Furthermore, it utilizes a deep convolutional neural network to generate a spatial density feature map, using local feature regions as topological vertices and combining the anatomical deformation field and feature similarity between adjacent physiological cycle timestamps to construct a spatiotemporal evolution topology map. Finally, it synchronously aggregates spatiotemporal features through a spatiotemporal graph convolutional neural network. This longitudinal dynamic monitoring mechanism can accurately capture the dynamic reconstruction process and nonlinear change patterns of breast tissue under the influence of endocrine and physiological cycles, outputting dynamic evolution trajectory vectors representing internal micro-proliferation and degeneration, significantly improving the sensitivity to the evolution trajectory of potential physiological abnormalities.

[0016] This invention inputs the extracted dynamic evolution trajectory vector into a preset risk prediction classifier. By calculating the spatiotemporal deviation between the subject's characteristics and the standard physiological evolution baseline constructed based on a healthy population cohort, the dynamic risk assessment coefficient of the subject is objectively obtained. This quantitative measurement method based on spatiotemporal deviation comparison effectively eliminates the lag and one-sidedness of absolute density assessment at a single time point, making the risk assessment accurately focus on the rationality of the tissue morphology evolution trend itself. On this basis, a full life cycle breast health prediction model is configured according to the dynamic risk assessment coefficient. The output assessment report not only covers highly sensitive disease warning states, but also comprehensively maps the long-term evolutionary characteristics of breast tissue. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating a dynamic breast density risk assessment method based on AI ultrasound according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is a dynamic breast density risk assessment method based on AI ultrasound, comprising the following steps: S1: Acquire multi-temporal ultrasound image sequences with physiological cycle timestamps, extract anatomical structural reference points from the multi-temporal ultrasound image sequences, perform elastic registration, and establish a unified spatial coordinate system across time dimensions; S2: Input the registered multi-temporal ultrasound image sequence into a deep convolutional neural network to extract the morphological features of the mammary glands within each physiological cycle time stamp and generate a spatial density feature map corresponding to each time node. S3: Using local feature regions within the spatial density feature map as topological vertices, and utilizing the anatomical deformation field and feature similarity between adjacent physiological cycle timestamps as topological edges, a spatiotemporal evolution topological map mapping the nonlinear changes in breast tissue morphology is constructed. S4: Input the spatiotemporal evolution topology graph into the spatiotemporal graph convolutional neural network, simultaneously aggregate the structural features of the spatial dimension and the state transfer weights of the temporal dimension, and output the dynamic evolution trajectory vector representing the hyperplasia and degeneration process inside the breast. S5: Input the extracted evolution trajectory vector into the preset risk prediction classifier, calculate the spatiotemporal deviation between the vector and the standard physiological evolution baseline constructed based on the healthy population cohort, and then obtain the dynamic risk assessment coefficient of the subject. S6: Configure a full life-cycle breast health prediction model based on the calculated dynamic risk assessment coefficient, and output a dynamic breast density risk assessment report.

[0021] In one embodiment of the present invention, the process of acquiring a multi-temporal ultrasound image sequence with physiological cycle timestamps in step S1, extracting anatomical reference points from the multi-temporal ultrasound image sequence, performing elastic registration, and establishing a unified spatial coordinate system across time dimensions is as follows: Acquiring breast ultrasound images of subjects at different stages involves acquiring cross-sectional data of breast tissue at different physiological stages using an ultrasound scanning device. During data acquisition, the specific time point of each ultrasound scan and the subject's corresponding physiological cycle status information are recorded simultaneously. The physiological cycle timestamp data corresponding to the acquisition time is embedded into the breast ultrasound image. This embedding process uses a data encapsulation protocol to convert the recorded physiological cycle timestamp data into a machine-readable binary identifier. This binary identifier is then bound and fused with the corresponding breast ultrasound image pixel matrix and written into the metadata identifier layer of the image file. This ensures that each frame of ultrasound image carries accurate time and physiological state labels. After data acquisition and identifier fusion processing at multiple time points, all labeled image data are integrated to finally generate a multi-temporal ultrasound image sequence with physiological cycle timestamps.

[0022] This study employs a feature recognition algorithm to analyze a multi-temporal ultrasound image sequence with physiological cycle timestamps. The algorithm uses a deep residual network model to perform layer-by-layer convolution and pooling operations on the input multi-temporal ultrasound image sequence, extracting grayscale echo gradient features and morphological edge contour features of different tissues in the image. Based on the extracted edge contour features, the algorithm traverses the entire image region to locate the boundary between the nipple and chest wall tissues. The localization process relies on comparing the high echo characteristics of the nipple region with the linear low echo characteristics of the chest wall tissue boundary, calculating the grayscale transition abrupt changes between the two tissue regions as anatomical boundaries. After accurately locating the aforementioned boundary locations and key extreme points on the boundaries, the row and column coordinates of these extreme points in the current image pixel grid are recorded. This extracts the anatomical structural reference points of the multi-temporal ultrasound image sequence, providing highly reliable anatomical anchor points for subsequent spatial alignment operations.

[0023] The spatial deformation vector is calculated based on the extracted anatomical reference points. Specifically, an ultrasound image at a certain time point in the sequence is selected as a static reference template. The coordinate offset of the anatomical reference points in the images at other time points relative to the corresponding reference points in the static reference template is calculated. Based on this coordinate offset, a dense deformation displacement field covering all image pixels is constructed and generated. This dense deformation displacement field is used to guide the redistribution of image pixel positions. The spatial deformation vector is used to perform elastic registration on the multi-temporal ultrasound image sequence. The elastic registration operation uses a non-rigid image transformation interpolation algorithm to ensure that the deformation of breast tissue undergoing nonlinear physiological changes can be accurately mapped and stretched to the template reference shape. This eliminates the macroscopic physical spatial deviation of tissue position and shape caused by the passage of time and the change of physiological cycle, and establishes a unified spatial coordinate system across the time dimension, so that ultrasound images acquired at different time points can be compared and analyzed pixel by pixel under the same spatial reference.

[0024] In one embodiment of the present invention, in step S2, the process of inputting the registered multi-temporal ultrasound image sequence into a deep convolutional neural network, extracting the morphological features of the mammary glands within each physiological cycle time stamp, and generating a spatial density feature map corresponding to each time node is as follows: The registered multi-temporal ultrasound image sequence is input into a deep convolutional neural network. This sequence has undergone spatial alignment processing, ensuring that anatomical structures at different time points are in a unified spatial reference coordinate system. The convolutional layers of the deep convolutional neural network perform a sliding mapping operation on the image pixel matrix. Specifically, each convolutional layer contains multiple convolutional kernels with learnable weights. These kernels act as local observation windows with specific receptive fields, moving row by row and column by column across the two-dimensional pixel matrix of the breast ultrasound image according to a preset stride. Within each sliding coverage area, the weight matrix of the convolutional kernel is multiplied and accumulated element-wise with the local pixel grayscale value matrix of the image. This local multiplication and accumulation operation allows for feature extraction because features within breast tissue... Tiny calcifications or glandular edges appear as dramatic changes in pixel grayscale values ​​in ultrasound images. The weight distribution of the convolutional kernel can resonate with these specific grayscale change patterns, thereby activating and amplifying these specific anatomical features. For example, when processing breast images recorded during a subject's ovulation period, a convolutional kernel specifically designed to identify the boundaries of hypoechoic regions will produce a high response value when sliding across the pixel matrix at the interface between glands and fat. This response mechanism can transform the original two-dimensional image space into a high-dimensional feature representation space. The two-dimensional feature maps calculated by sliding all convolutional kernels across the entire image matrix are stacked and stitched along the channel dimension to finally output the initial feature tensor data. This initial feature tensor data retains the basic texture and edge structure information in the multi-temporal ultrasound image sequence.

[0025] Based on the initial feature tensor data output, a downsampling layer of a deep convolutional neural network is used to perform dimensionality reduction. The downsampling layer typically employs max pooling or average pooling mechanisms. A fixed-size pooling window is defined on the feature map of each channel of the initial feature tensor data, and this window slides across the feature map. Within each local region, the downsampling layer selects the maximum feature activation value within that region or calculates the average of the feature values ​​within the region, compressing and merging multiple adjacent feature pixels into a single representative value. This regional numerical compression mechanism enables dimensionality reduction and extraction of effective features because ultrasound images often contain a large amount of speckle noise and minute probe pressure deformation artifacts. Dimensionality reduction effectively filters out these redundant high-frequency spatial detail interferences, while simultaneously expanding the image size exponentially. The receptive field space of the large network model on the original image enables the network to transcend local limitations and perceive a wider range of tissue structure morphology. For example, when processing feature data of the specific physiological cycle time stamp of the luteal phase, due to endocrine influences that cause mammary duct dilation and glandular congestion and thickening, the downsampling process retains and highlights the high activation signals of the boundaries of these macroscopically significantly thickened glandular structures, while discarding those subtle, non-essential local gray-level fluctuations. As the downsampling level continues to deepen, the spatial resolution of the feature tensor decreases step by step while the level of abstraction increases. The network gradually transitions from focusing on the bottom-level dots, lines, and patches to focusing on the overall tissue topology distribution. In this process, it completes the mapping and integration from low-level visual signals to high-level semantic information, and thus successfully extracts the morphological features of mammary glands within each physiological cycle time stamp.

[0026] The morphological features of breast glands within various physiological cycle time stamps are input into the upsampling layer of a deep convolutional neural network to perform spatial distribution reconstruction. The working mechanism of the upsampling layer is the opposite of the aforementioned downsampling process. It mainly restores the spatial resolution of the feature data step by step through transposed convolution or interpolation operations. Specifically, it uses a reverse mapping matrix with dilated weights to reproject and map the highly compressed low-dimensional abstract morphological feature vectors back into a high-dimensional two-dimensional image coordinate grid. The reason why spatial distribution reconstruction can be achieved through reverse spatial mapping is that although the core semantic information of glandular hyperplasia or degeneration is condensed in the morphological features extracted in the early stage, the precise physical location of these tissue changes is lost. The upsampling process, based on the tissue growth and diffusion laws learned by the network, concentrates the feature activation information... The signal is evenly and smoothly diffused to the corresponding anatomical area, thereby recovering the lost spatial positioning information. For example, when reconstructing the features of the menstrual period, since the mammary glands are usually in a relatively static or regressed state at this time, the upsampling layer will map the feature vectors representing low density back to the original ducts and lobules. After multiple resolution magnification operations, the density fluctuations of each tissue structure are accurately depicted on the two-dimensional plane. This refined restoration of spatial distribution not only marks the anatomical boundaries of the glandular tissue, but also quantifies the density of the tissue through different numerical intensities, thereby generating a spatial density feature map corresponding to each time point. This transforms the originally abstract deep learning features into a two-dimensional digital matrix that can intuitively reflect the distribution of substances inside the mammary gland at different physiological stages.

[0027] In one embodiment of the present invention, in step S3, the process of constructing a spatiotemporal evolution topology map that maps the nonlinear changes in breast tissue morphology, using local feature regions within the spatial density feature map as topological vertices and the anatomical deformation fields and feature similarities between adjacent physiological cycle time stamps as topological edges, is as follows: A region segmentation algorithm is used to divide the spatial density feature map corresponding to each time point into image blocks. This algorithm clusters similar feature pixels by calculating the gray-level gradient changes and spatial geometric distances between adjacent pixels within the feature map. Specifically, it expands outward from a certain pixel, packing sets of pixels with similar density distributions and physically connected locations into independent blocks. The reason why pixel clustering can achieve image block division is that the adipose tissue and glandular tissue inside the breast exhibit distinctly different activation intensities on the spatial density feature map. Isolating continuous regions with consistent intensities can accurately isolate tissue fragments with independent anatomical significance. For example, when processing regions with high-density nodule features, the segmentation algorithm will follow the nodule edges... High-gradient mutation locations are cut to separate the complete nodule into a single image block. Then, local feature regions within the spatial density feature map of each image block are extracted. The extraction process involves calibrating the center point coordinates of each segmented block and retaining the set of density feature values ​​of all pixels within that block. The aforementioned local feature regions containing specific spatial coordinates and corresponding density feature value sets are defined as topological vertices. The reason for defining topological vertices is to prepare for the construction of a topological graph is that the analytical framework of the topological graph needs to abstract continuous physical space into discrete mathematical representation entities. Each topological vertex at this time represents a small and specific physical tissue unit inside the breast, providing a fixed and independent data carrier for cross-time comparison.

[0028] Extract the spatial coordinate matrix of the topological vertices corresponding to adjacent physiological cycle timestamps, perform displacement mapping operation based on the coordinate offset vector of the spatial coordinate matrix, and generate the anatomical structure deformation field between adjacent physiological cycle timestamps. The specific content includes: The process involves reading the pixel grid positions of topological vertices within adjacent physiological cycle timestamps. Specifically, this involves traversing the two-dimensional physical structure space of ultrasound images spanning two adjacent physiological stages. On the row and column grid distribution plane of the image, the absolute physical coordinates of each topological vertex representing a local tissue feature are precisely located. For example, in an ovulation ultrasound image, the vertex of a specific breast lobule edge is identified, and its horizontal and vertical coordinates are recorded and converted into a numerical sequence. This conversion into a numerical sequence prepares the data for depth computation because discrete pixel physical position information can only directly participate in the algebraic matrix of the underlying computing hardware after being digitized and arranged into a mathematical vector column with a specific algebraic order. The operation extracts the spatial coordinate matrix of the topological vertices corresponding to adjacent physiological cycle timestamps. This extraction and stitching process involves orderly combining and stacking the coordinate value sequences of all vertices representing the same anatomical structure positions in the previous and subsequent physiological cycle images. Since a matrix is ​​essentially a structured collection carrier of multidimensional data, combining the coordinate values ​​of massive numbers of individual tissue physical points according to the corresponding topological connection order can form a complete mathematical matrix expression framework. Through this spatial coordinate matrix, the physical position distribution of breast tissue micro-units under different physiological endocrine states within the same reference space datum plane can be recorded completely and clearly.

[0029] The extracted spatial coordinate matrix is ​​processed using matrix subtraction. Specifically, the spatial coordinate matrix representing the corresponding topological vertex of the previous physiological cycle timestamp is subtracted from the spatial coordinate matrix representing the corresponding topological vertex of the next physiological cycle timestamp. This is achieved by strictly performing algebraic element-by-element subtraction at each row and column position within the corresponding matrix dimension, calculating the coordinate difference between the corresponding vertex positions at the previous and next time points. The reason matrix subtraction can accurately calculate this coordinate difference is that the uniformly established cross-time reference spatial coordinate system completely eliminates macroscopic geometric displacement caused by the physical movement of the external ultrasound probe. The relative coordinate changes of the same anatomical physical unit at different time points purely reflect the actual physiological deformation of the tissue under the influence of hormones, such as from the egg... During the transition from the vesicular phase to the luteal phase, mammary ducts may undergo physiological dilation and local edema. The coordinate values ​​of the vertices representing the duct wall's edge contour will inevitably expand outward physically. The difference obtained by algebraically subtracting the corresponding values ​​precisely quantifies the absolute physical distance and growth direction of this outward expansion, generating a coordinate offset vector of the spatial coordinate matrix. By summarizing and recombining the absolute differences in coordinates of all feature vertices, a mathematical set containing multiple vectors with specific physical directions and geometric sizes can be obtained. The reason why generating this coordinate offset vector can reflect the microscopic evolution of local tissues is that this vector set accurately depicts the true movement trajectory and deformation offset trend of each tiny tissue unit inside the mammary gland in the intuitive form of discrete arrows during the physiological cycle.

[0030] Based on the coordinate offset vector of the spatial coordinate matrix, spatial interpolation transformation is used to perform displacement mapping calculations. This process employs the fundamental mathematical and physical principles of spatial difference interpolation and geometric deformation mapping. The extracted discrete topological vertices with coordinate offset vector directions are used as local deformation control reference anchor points. Smooth force phase displacement calculations and algebraic interpolation are performed on the transition pixels in the blank image regions between adjacent vertices. For example, thin-plate spline interpolation is used to smoothly infer the passive force displacement direction and absolute distance of pixels in the ductal interior and surrounding interstitial adipose tissue. The reason why displacement mapping calculations can smoothly reconstruct global deformation features is that breast tissue is naturally a continuous viscoelastic medium. The traction and compression actions of local tissue control vertices will inevitably affect the surrounding tissue according to the laws of physical mechanical conduction. Some adjacent pixels undergo continuous and smooth cooperative deformation. Aggregating spatial deformation grid data involves mathematically splicing and summarizing the tiny displacement vector data of all discrete pixels obtained by interpolation according to their original two-dimensional planar grid coordinates. This process re-weaves the discrete deformation information of the local area into a continuous dynamic grid covering the entire two-dimensional cross-section of the breast image, generating an anatomical deformation field between adjacent physiological cycle timestamps. The reason why the final anatomical deformation field can be obtained by aggregating continuous grid data is that this continuous topological grid completely breaks the observation limitations of a single vertex. It provides a panoramic view of the entire nonlinear geometric evolution physical life process of breast tissue under the complex endocrine and biochemical environment changes, from local microscopic glandular hyperplasia and expansion to macroscopic overall tissue structure distortion and reshaping.

[0031] The pixel intensity distribution matrix is ​​extracted from the topological vertices corresponding to adjacent physiological cycle timestamps. This matrix records the density activation values ​​of all pixels within a specific tissue block, reflecting the microscopic physiological manifestations of glandular density or edema in that local area. After identifying the same tissue block in two adjacent time cycles, all intensity values ​​within it are extracted in their original spatial order to form a comparison matrix. The difference between the pixel intensity distribution matrices is calculated using a distance metric function. This calculation involves comparing the values ​​of elements at the same position in the two matrices one by one. For example, the absolute distance between two high-dimensional intensity vectors in numerical space is measured by calculating the difference of squares and then summing and taking the square root. If a glandular lobule undergoes [a certain process] during the luteal phase... Significant physiological hyperplasia enhances local echoes, causing an overall increase in the values ​​within the corresponding matrix. Consequently, the sum of squared differences calculated using distance metrics becomes significantly larger. The reason why calculating matrix differences can quantify changes in tissue state is that pixel intensity is directly related to the physical reflection characteristics within the breast, and fluctuations in intensity objectively correspond to cell-level proliferation or degeneration. Based on the calculated difference values, the reciprocal or exponential decay operation is used to generate feature similarity between adjacent physiological cycle timestamps. A larger difference value indicates more drastic changes in the internal state of the tissue over time, resulting in lower feature similarity. Conversely, a smaller difference value indicates a stable tissue state and higher similarity. This similarity accurately characterizes the potential non-structural physiological evolution features of breast tissue without significant geometric deformation.

[0032] The generated anatomical deformation field and feature similarity are used to perform data concatenation to generate a connection weight matrix. This operation involves concatenating the deformation field tensor, which represents the degree of distortion of the macroscopic geometric structure, and the feature similarity scalar, which represents changes in the physiological state of microscopic tissues, along the feature dimension according to a preset weight ratio, forming a new matrix containing multimodal change information. The reason why data concatenation can generate complete connection weights is that a single deformation field can only describe where the tissue has moved, while a single similarity can only describe what the tissue has become. Only by integrating the two can the true evolution of breast tissue under the dual influence of internal and external environments be fully reflected. For example, when a certain region not only undergoes significant outward expansion deformation, but also has extremely low pixel intensity similarity and shows a proliferative trend, the concatenated matrix will show a more complete picture. The matrix assigns a very high attention weight to the region in the corresponding dimension; using the connection weight matrix as the topological edge connecting the topological vertices, the extracted local tissue units are treated as isolated nodes in graph theory. They are interconnected by the connection weight matrix, which represents the evolutionary relationship in the time dimension and the adjacent influence in the spatial dimension. The topological edge with high weight indicates that there is a strong physiological causal relationship and state transition trend between two nodes or the same node over time; by establishing the network relationship between all vertices and edges, a spatiotemporal evolution topological graph that maps the nonlinear changes in breast tissue morphology is constructed. This topological graph completely breaks the static limitations of traditional single ultrasound images and uses a complex graph network structure to provide a complete and concrete mathematical expression of the dynamic reconstruction process of breast tissue in different physiological cycles and the potential pathological evolution trajectory.

[0033] In one embodiment of the present invention, step S4, which involves inputting the spatiotemporal evolution topology graph into a spatiotemporal graph convolutional neural network, simultaneously aggregating the structural features of the spatial dimension and the state transfer weights of the temporal dimension, and outputting a dynamic evolution trajectory vector representing the hyperplasia and degeneration process within the breast, is as follows: The constructed spatiotemporal evolution topology is input into a spatiotemporal graph convolutional neural network. This topology contains topological vertices representing local tissue units of the breast and topological edges representing the deformation field and feature similarity of inter-tissue anatomical structures. The spatial convolutional layer of the spatiotemporal graph convolutional neural network traverses the topological vertices along the topological edges. Specifically, a local receptive field is defined, and each topological vertex is used as a central node. The network searches outward along the topological edges connected to it and locks down neighboring nodes. For example, when processing the central node representing a breast lobule, the spatial convolutional layer finds the surrounding adipose tissue nodes or connected ductal tissue nodes based on the topological edges, collects the spatial density feature information carried by these neighboring nodes, and extracts the spatial dimension. The reason why spatial structural features can be extracted by traversing along topological edges is that the physical tissues inside the breast are not isolated. The densification or edema of a local area is often accompanied by stress changes and echo changes in the surrounding matrix. The spatial convolutional layer uses an aggregation matrix with learnable weights to perform a weighted summation operation on the features of the central node itself and the features of the collected neighboring nodes. This neighborhood feature weighted fusion mechanism on the graph structure can break the physical barriers of traditional discrete pixels and reweave the scattered local density features into a topological feature map with macroscopic anatomical significance, thereby accurately capturing the intricate spatial distribution patterns of tissues and glandular morphological features inside the breast under the same physiological time slice.

[0034] This study utilizes the temporal convolutional layer of a spatiotemporal graph convolutional neural network to extract the node feature differences between adjacent physiological cycle timestamps. Specifically, after locking onto the same anatomical topological vertex in the spatial dimension, the feature vectors of that vertex at different time points, such as the follicular and luteal phases, are aligned along the temporal evolution axis. The algebraic distance and directional shift values ​​between the feature vectors of the preceding and following time points are calculated in the multidimensional feature space. For example, a node representing the edge of a gland may experience an increase in density during the luteal phase due to congestion and hyperplasia. The temporal convolutional layer can sensitively capture this abrupt change in feature value from low to high through vector subtraction and nonlinear activation operations, generating temporal state transfer weights. The reason why calculating feature differences can generate state transfer weights is that the fluctuation amplitude of features objectively reflects the intensity of the microscopic physiological evolution of that local tissue. The greater the difference between nodes, the more crucial their dynamic indicative role becomes throughout the entire life cycle. Neural networks assign them higher numerical weight matrices to highlight their importance. The state transfer weights and spatial structural features are multiplied and added together to achieve synchronous aggregation. This process involves algebraically multiplying the scalar weights representing the intensity of temporal changes with the tensor features representing the spatial distribution patterns element by element. The result of the multiplication is then accumulated within a sliding window of the temporal receptive field. The reason why synchronous aggregation can be achieved through multiplication and addition is that isolated spatial features cannot independently reflect dynamic processes, while single temporal differences lack anatomical localization. The multiplication and addition fusion mechanism endows the originally static spatial feature map with the direction and rate attributes of temporal flow and evolution, constructing a complete high-dimensional spatiotemporal feature tensor that includes both where the organization is and how it changes.

[0035] The feature tensor data from the synchronous aggregation output is input into the fully connected layer of the spatiotemporal graph convolutional neural network for dimensionality reduction mapping. Specifically, the high-dimensional multi-channel feature tensor, which integrates temporal weights and spatial structure, is flattened into a one-dimensional elongated data vector. The fully connected layer contains a dense connection matrix composed of numerous neurons. These neurons utilize weight distribution parameters learned in advance through training on a large number of healthy and abnormal samples to perform global linear combination and nonlinear compression calculations on the flattened one-dimensional data vector. For example, a huge vector containing tens of thousands of spatiotemporal feature values ​​is compressed and condensed into a compact mathematical vector containing only a few hundred core principal component factors. The reason why the core principal components can be extracted through the dimensionality reduction mapping of the fully connected layer is that the original high-dimensional tensor contains a large amount of redundant common background physiological information. The dimensionality reduction process, which excludes minor non-critical noise fluctuations, is essentially an information purification and feature distillation mechanism. It forcibly filters out marginal parameters that have little impact on the macroscopic evolution trend, retaining only the core evolutionary markers that best represent the substantial abnormal proliferation or atrophy and degeneration of breast tissue. It outputs a dynamic evolutionary trajectory vector that characterizes the internal proliferation and degeneration process of the breast. The reason why this evolutionary trajectory vector can be obtained through the above dimensionality reduction operation is that the mathematical vector after extreme compression and purification has formed a coordinate evolutionary trajectory with a clear direction in algebraic space. The starting point of this trajectory represents the initial physiological state of the subject's breast, and the extension direction and length of the trajectory precisely quantify the complete dynamic process of nonlinear proliferation mutation or degeneration and reversion of the breast gland due to endocrine influences or potential pathological factors.

[0036] In one embodiment of the present invention, step S5 involves inputting the extracted evolution trajectory vector into a preset risk prediction classifier, calculating the spatiotemporal deviation between the vector and the standard physiological evolution baseline constructed based on a healthy population cohort, and then obtaining the dynamic risk assessment coefficient of the subject. The extracted evolutionary trajectory vector is input into a pre-defined risk prediction classifier. This classifier is pre-loaded with a baseline model trained using long-term follow-up data from a large-scale healthy subject population. It reads the standard physiological evolution baseline constructed based on the healthy population cohort. Specifically, the feature comparison module loads the mathematical expectation trajectory of the regular changes in breast tissue morphology in healthy individuals at different physiological cycle stages. For example, healthy women typically experience mild physiological congestion and increased density in the breast glands during ovulation to the luteal phase, while involution and decreased density occur during menstruation. These universally observed regular physiological fluctuations are quantified into a base in multidimensional space. The quasi-reference frame is used to extract the feature distribution matrix of the standard physiological evolution baseline. This extraction process involves discretizing the continuous mathematical expectation trajectory according to the corresponding time nodes, and converting the standard tissue density range and the allowable limit of minor deformation of anatomical structures under different physiological cycle time stamps into a digital matrix with fixed dimensions. The reason why the feature distribution matrix can be extracted through discretization sampling is that the original standard physiological evolution baseline is a smooth theoretical curve, which must be converted into an algebraic matrix form that is completely aligned with the dimension of the subject's evolution trajectory vector in order to provide an objective and quantifiable health reference coordinate template for risk prediction at the digital processing level.

[0037] The input data is processed using the spatial geometric distance algebraic operation logic built into the risk prediction classifier. The spatial feature parameters of the evolution trajectory vector are compared with the spatial baseline values ​​of the feature distribution matrix. Specifically, the feature vector elements representing the density of local breast tissue and the extent of anatomical contour expansion of the subject at specific time points are extracted one by one. These vector elements are then compared with the standard density mean and normal contour deformation variance of healthy individuals at the same physiological cycle stage, and a multidimensional spatial difference calculation is performed. For example, if a local tissue block in a subject shows an abnormally high-density cluster on an ultrasound image, while the same location in the baseline matrix of healthy individuals only shows a uniformly scattered low density... When comparing the density of glandular tissue, a significant distance difference will occur between the two in the algebraic coordinates of the spatial feature dimension. The reason why the spatial deviation value can be obtained through the above-mentioned multidimensional spatial difference calculation is that the potential micro-tumor lesions or abnormal proliferative tissues inside the breast will inevitably break the normal physiological spatial distribution law in terms of physical morphology. Projecting these abnormal mutations in anatomical morphology onto the high-dimensional feature space will manifest as the subject's data points being far away from the cluster center of the healthy benchmark dataset. The absolute magnitude of this algebraic distance objectively and accurately quantifies the severity of the subject's breast tissue deviating from the normal physiological range in terms of physical structure and density distribution.

[0038] Using a pre-set time-series dynamic alignment matching algorithm within the risk prediction classifier, the time feature parameters of the evolution trajectory vector are compared with the time baseline values ​​of the feature distribution matrix. This process synchronously maps and matches the time node distribution sequence of changes in the subject's breast tissue density with the physiological rhythm evolution time axis of a standard healthy population. For example, in healthy women, breast glands typically regress rapidly within two to three days after menstruation. However, if the subject's evolution trajectory shows that the high density of their glands remains firm and shows no downward trend a week after menstruation ends, the time alignment algorithm... This will capture obvious lag and misalignment phenomena in the comparison matrix of the time dimension, and calculate the time deviation value. The reason why the time deviation value can be calculated through dynamic alignment matching is that the nonlinear morphological reconstruction of breast tissue is not only spatially specific, but its evolution rate and synchronous rhythm with the fluctuation of endocrine hormones are also core indicators for measuring health status. By calculating the phase difference and duration difference between the peak of the change in the subject's tissue state and the peak of the standard reference time, we can accurately capture the potential pathological evolution precursor signals that have not yet shown obvious abnormalities in spatial morphology but whose physiological rhythms have been severely disrupted.

[0039] The spatial and temporal deviation values ​​are weighted and fused. Specifically, based on the clinical pathological patterns of breast disease evolution, different importance algebraic scaling factors are assigned to the degree of spatial distortion and temporal rhythm disorder. For example, considering that morphological solid mass abnormalities have a higher probability of malignancy than simple involution lag, the algorithm assigns a higher computational weight to the spatial deviation value than the temporal deviation value. The values ​​of both are multiplied by their corresponding weight factors and then summed to generate the spatiotemporal deviation between the evolutionary trajectory vector and the standard physiological evolution baseline. The reason why weighted fusion can generate a comprehensive spatiotemporal deviation is that... A single spatial or temporal dimension can only reflect a partial slice of breast health. Only by combining abnormal deformation of physical structure with dynamic misalignment of physiological cycle can the true evolution of the subject's breast tissue be fully restored in a higher-dimensional spatiotemporal continuum, and the subject's dynamic risk assessment coefficient be obtained. At this time, the deviation value obtained by comprehensive calculation has been mapped to a continuous probability interval. This coefficient directly breaks the blind spot of missed diagnosis that is easily caused by a single static screening, and accurately identifies the potential comprehensive risk probability of the subject's breast tissue evolving into malignant lesions or irreversible abnormal hyperplasia in the future with a quantitative continuous algebraic scalar.

[0040] In one embodiment of the present invention, step S6, which involves configuring a full-life-cycle breast health prediction model based on the calculated dynamic risk assessment coefficient and outputting a dynamic breast density risk assessment report, is as follows: A full-life-cycle breast health prediction model is configured based on the calculated dynamic risk assessment coefficient. This configuration process involves substituting this coefficient as a key initial variable parameter into the model's underlying architecture. This ensures that the model's internal calculation weights closely align with the subject's current individual health baseline. For example, if a subject's risk assessment coefficient is high, the model will automatically adjust its sensitivity response to aging and hormone level fluctuations to accommodate potential abnormal tissue hyperplasia. The model then employs regression mathematical logic within the full-life-cycle breast health prediction model to perform a fitting operation. Specifically, this fitting operation uses the subject's current age and the percentage of dense tissue extracted from previous imaging records as the time axis. The independent variable is represented by the probability of variation of the calculated dynamic risk assessment coefficient as the core constraint. A smooth and continuous evolution prediction curve is drawn in a multidimensional coordinate system. The reason why regression fitting can generate the evolution trend of breast tissue morphology is that regression calculation can deeply explore and connect the deep nonlinear correlation between historical tissue density changes and the passage of time. It can also reasonably extend the small risk deviations found at present to the future life evolution axis. For example, it can predict that the rate of glandular degeneration in subjects before and after menopause will be significantly slower than that in normal people. Thus, it can clearly show the density trend and morphological reconstruction development direction of breast tissue at different age stages in the future with a continuous trajectory.

[0041] The dynamic risk assessment coefficient is input into the classification node of the full-life-cycle breast health prediction model. This classification node is essentially a set of decision boundaries trained on massive amounts of real clinical pathological data. Internally, it constructs a mathematical hyperplane that divides different health risk levels. The input process involves mapping a single scalar coefficient, condensed through the aforementioned complex spatiotemporal topological evolution calculation, to specific coordinate points in this multidimensional decision space. A comparison operation logic is used to compare the dynamic risk assessment coefficient with a preset risk threshold value. This preset risk threshold value represents a digital watershed between healthy physiological fluctuations and pathological abnormal hyperplasia, such as a low-risk threshold for benign breast cystic hyperplasia and a high-risk boundary for severe atypical hyperplasia. The comparison process directly calculates the input dynamic risk coefficient. The reason why we can draw conclusions about which two critical watersheds the state coefficient falls between or whether it exceeds the highest warning line is because the occurrence and evolution of any disease in the mathematical feature space involves a threshold crossing process from quantitative accumulation to qualitative change. When the coefficient breaks through a certain statistical red line, it means that the variation of its tissue microstructure has exceeded the limit of the body's self-repair and physiological compensation. The model classification node makes logical judgments based on the relative range of the coefficient and outputs the disease warning status. This warning status is presented in the form of highly instructive category labels, such as classifying it as a low-risk stage that requires close long-term observation or a high-risk stage that requires immediate intervention with biopsy puncture, thereby realizing the clinical interpretation and risk qualitative classification of abstract feature data.

[0042] The generated breast tissue morphological evolution trend and the disease warning status are imported into a preset layout framework for merging and splicing. This preset layout framework is a visual organizational template containing multiple blank information-carrying blocks. The import operation assigns the abstract mathematical evolution trajectory curve to the visual analysis section for displaying long-term trends, while qualitative classification warning labels and corresponding clinical intervention guidance text are filled into the conclusion summary block at the front. The merging and splicing operation uses image-text fusion to eliminate display gaps between different data sources, logically corresponding and physically aligning the dynamic evolution prediction details on the timeline with the current risk qualitative conclusion on the same reading plane. For example, high-risk warning nodes are directly marked with a striking color on the morphological evolution... The reason why reports can be generated by merging and splicing the age scale corresponding to the steep slope change on the trend curve is that isolated mathematical curves and individual warning labels are difficult for non-professional subjects or clinicians to understand intuitively and comprehensively. Through framework-based visual reconstruction and logical splicing, cold data can be transformed into a health record with a clear narrative structure. Finally, a dynamic breast density risk assessment report covering the evolution trend of breast tissue morphology and disease warning status is generated and output. This report is not just a static routine screening notice, but a comprehensive health guidance blueprint that deeply integrates spatiotemporal topological analysis features and individual life evolution patterns. It directly provides a solid and intuitive reference for subjects to formulate personalized lifelong cancer screening plans and clinical intervention programs.

[0043] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A dynamic breast density risk assessment method based on AI ultrasound, characterized in that, Includes the following steps: S1: Acquire multi-temporal ultrasound image sequences with physiological cycle timestamps, extract anatomical structural reference points from the multi-temporal ultrasound image sequences, perform elastic registration, and establish a unified spatial coordinate system across time dimensions; S2: Input the registered multi-temporal ultrasound image sequence into a deep convolutional neural network to extract the morphological features of the mammary glands within each physiological cycle time stamp and generate a spatial density feature map corresponding to each time node. S3: Using local feature regions within the spatial density feature map as topological vertices, and utilizing the anatomical deformation field and feature similarity between adjacent physiological cycle timestamps as topological edges, a spatiotemporal evolution topological map mapping the nonlinear changes in breast tissue morphology is constructed. S4: Input the spatiotemporal evolution topology graph into the spatiotemporal graph convolutional neural network, simultaneously aggregate the structural features of the spatial dimension and the state transfer weights of the temporal dimension, and output the dynamic evolution trajectory vector representing the hyperplasia and degeneration process inside the breast. S5: Input the extracted evolution trajectory vector into the preset risk prediction classifier, calculate the spatiotemporal deviation between the vector and the standard physiological evolution baseline constructed based on the healthy population cohort, and then obtain the dynamic risk assessment coefficient of the subject. S6: Configure a full life-cycle breast health prediction model based on the calculated dynamic risk assessment coefficient, and output a dynamic breast density risk assessment report.

2. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S1, the process of establishing a unified spatial coordinate system across the time dimension is as follows: Breast ultrasound images of subjects at different stages were acquired, and the physiological cycle timestamp data corresponding to the acquisition time were embedded into the breast ultrasound images to generate a multi-temporal ultrasound image sequence with physiological cycle timestamps. The feature recognition algorithm was used to analyze the multi-phase ultrasound image sequence with physiological cycle time stamps, locate the junction between the nipple and chest wall tissue, and extract the anatomical reference points of the multi-phase ultrasound image sequence. Based on the extracted anatomical reference points, a spatial deformation vector is calculated. This spatial deformation vector is then used to perform elastic registration on multi-temporal ultrasound image sequences, establishing a unified spatial coordinate system across time dimensions.

3. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S2, the process of generating the spatial density feature map corresponding to each time node is as follows: The registered multi-temporal ultrasound image sequence is input into a deep convolutional neural network. The convolutional layers of the deep convolutional neural network are used to perform a sliding mapping operation on the image pixel matrix and output the initial feature tensor data. Based on the output initial feature tensor data, the downsampling layer of the deep convolutional neural network is used to perform dimensionality reduction processing to extract the morphological features of mammary glands within each physiological cycle time stamp. The morphological features of breast glands within each physiological cycle time stamp are input into the upsampling layer of a deep convolutional neural network to perform spatial distribution reconstruction, generating a spatial density feature map corresponding to each time node.

4. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S3, the process of constructing a spatiotemporal evolution topology map that maps the nonlinear changes in breast tissue morphology is as follows: The spatial density feature map corresponding to each time node is divided into image blocks using a region segmentation algorithm. Local feature regions inside the spatial density feature map contained in each image block are extracted and defined as topological vertices. Extract the spatial coordinate matrix of the topological vertices corresponding to adjacent physiological cycle timestamps, and perform displacement mapping operation based on the coordinate offset vector of the spatial coordinate matrix to generate the anatomical structure deformation field between adjacent physiological cycle timestamps. Extract the pixel intensity distribution matrix of the corresponding topological vertices between adjacent physiological cycle timestamps, calculate the difference value of the pixel intensity distribution matrix through the distance metric function, and generate the feature similarity between adjacent physiological cycle timestamps; The generated anatomical deformation field and feature similarity are spliced ​​to generate a connection weight matrix. The connection weight matrix is ​​used as the topological edge connecting the topological vertices to construct a spatiotemporal evolution topological graph that maps the nonlinear changes in breast tissue morphology.

5. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 4, characterized in that, The process of generating the anatomical deformation field between adjacent physiological cycle timestamps is as follows: Read the pixel grid positions of topological vertices within adjacent physiological cycle timestamps, convert them into numerical sequences, and extract the spatial coordinate matrix of the corresponding topological vertices between adjacent physiological cycle timestamps; The extracted spatial coordinate matrix is ​​processed by matrix subtraction, and the coordinate difference between the corresponding vertex positions before and after the time nodes is calculated to generate the coordinate offset vector of the spatial coordinate matrix. Based on the coordinate offset vector of the spatial coordinate matrix, the spatial transformation function is called to perform displacement mapping operation, aggregate spatial deformation mesh data, and generate the anatomical structure deformation field between adjacent physiological cycle timestamps.

6. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S4, the process of outputting the dynamic evolution trajectory vector representing the hyperplasia and degeneration process within the mammary gland is as follows: The spatiotemporal evolution topology graph is input into the spatiotemporal graph convolutional neural network. The spatial convolutional layer of the spatiotemporal graph convolutional neural network is used to traverse the topology vertices along the topology edges to extract the structural features of the spatial dimension. The temporal convolutional layer of the spatiotemporal graph convolutional neural network is used to extract the node feature differences of adjacent physiological cycle timestamps, generate state transit weights in the time dimension, and perform multiplication and addition operations on the state transit weights and the structural features in the spatial dimension to complete synchronous aggregation. The feature tensor data from the synchronous aggregation output is input into the fully connected layer of the spatiotemporal graph convolutional neural network to perform dimensionality reduction mapping, and outputs a dynamic evolution trajectory vector representing the hyperplasia and degeneration process inside the breast.

7. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S5, the process of obtaining the subject's dynamic risk assessment coefficient is as follows: The extracted evolution trajectory vector is input into a preset risk prediction classifier, the standard physiological evolution baseline constructed based on the healthy population cohort is read, and the feature distribution matrix of the standard physiological evolution baseline is extracted. Using the spatial distance calculation function of the risk prediction classifier, the spatial feature parameters of the evolution trajectory vector are compared with the spatial reference value of the feature distribution matrix to calculate the spatial deviation value. Using the time series alignment algorithm of the risk prediction classifier, the time feature parameters of the evolution trajectory vector are compared with the time reference values ​​of the feature distribution matrix to calculate the time deviation value; The spatial deviation value and the temporal deviation value are weighted and fused to generate the spatiotemporal deviation between the evolution trajectory vector and the standard physiological evolution baseline, thereby obtaining the dynamic risk assessment coefficient of the subject.

8. The method for dynamic breast density risk assessment based on AI ultrasound according to claim 1, characterized in that, In step S6, the process of outputting the dynamic breast density risk assessment report is as follows: Configure a full life cycle breast health prediction model based on the calculated dynamic risk assessment coefficient, call the regression function of the full life cycle breast health prediction model to perform fitting operation, and generate the evolution trend of breast tissue morphology. The dynamic risk assessment coefficient is input into the classification node of the full life cycle breast health prediction model, the dynamic risk assessment coefficient is compared with the preset risk threshold value, and the disease warning status is output. The generated breast tissue morphological evolution trend and the disease warning status are imported into a preset layout framework and merged and spliced ​​to output a dynamic breast density risk assessment report that covers both the breast tissue morphological evolution trend and the disease warning status.

Images