Breast cancer molecular subtype recognition and prediction method and system based on ultrasonic image

Abstract

The invention provides a breast cancer molecular subtype recognition and prediction method and system based on an ultrasonic image, and the method comprises the steps: carrying out the feature fusion preprocessing of a breast ultrasonic image based on acoustic impedance difference, and obtaining a standardized ultrasonic image; dividing the image into fan-shaped regions based on the radial fiber structure of the mammary gland, and dynamically adjusting the window size according to tissue characteristics to carry out adaptive window segmentation to obtain a window set carrying anatomical structure position information; inputting the window set into a pre-trained cross window Transform model for feature extraction, and analyzing a feature association mode between the windows through a cross window attention mechanism to obtain a multi-dimensional ultrasonic feature vector; and inputting the feature vector into a classifier for molecular subtype classification prediction to obtain a breast cancer molecular subtype identification prediction result. According to the method, through adaptive window segmentation, breast anatomical structure information is utilized, local and global feature association is captured through a cross window attention mechanism, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a method and system for identifying and predicting breast cancer molecular subtypes based on ultrasound images. Background Art

[0002] Accurate identification of breast cancer molecular subtypes is crucial for developing personalized treatment plans. Traditional molecular subtype identification relies primarily on histopathological examination and immunohistochemical staining, measuring the expression levels of molecular markers such as estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and the Ki-67 proliferation index to determine subtypes such as luminal A, luminal B, HER2-enriched, and triple-negative. This method requires tissue sampling, is invasive, and has a long detection cycle, making it inadequate for rapid clinical diagnosis.

[0003] In recent years, with the development of artificial intelligence (AI) technology, methods for predicting breast cancer molecular subtypes based on medical imaging have gradually emerged. Existing technologies primarily use traditional convolutional neural networks for feature extraction and classification of breast images, but these methods have significant shortcomings when processing ultrasound images. Ultrasound images are characterized by high noise, low contrast, and blurred tissue boundaries. Traditional networks struggle to effectively capture the subtle differences in features associated with molecular subtypes in ultrasound images, resulting in limited prediction accuracy and failing to meet the precision requirements for clinical applications. Summary of the Invention

[0004] The main purpose of the present invention is to solve the technical problem of limited prediction accuracy of existing methods for identifying breast cancer molecular subtypes based on ultrasound images; A first aspect of the present invention provides a method for identifying and predicting breast cancer molecular subtypes based on ultrasound images. The method comprises: Perform feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image; Dividing the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjusting the window size according to tissue characteristics, and performing adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; Inputting the window set into a pre-trained cross-window Transformer model for feature extraction, analyzing the feature correlation pattern between windows through a cross-window attention mechanism, and obtaining a multi-dimensional ultrasound feature vector; The multi-dimensional ultrasound feature vector is input into the classifier of the cross-window Transformer model for molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

[0005] Optionally, in a first implementation of the first aspect of the present invention, performing feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image includes: Based on the reflection characteristics of ultrasound in the skin layer, subcutaneous fat layer, glandular layer, and pectoral muscle layer, the acquired breast ultrasound image is processed for tissue interface boundary recognition to obtain tissue stratification information; Processing the tissue layer information using an acoustic shadow compensation algorithm to correct deep echo attenuation caused by changes in lesion tissue density to obtain an echo-corrected image; performing acoustic enhancement algorithm processing on the echo-corrected image to highlight tissue boundaries and internal structural features, thereby obtaining an acoustically enhanced image; performing texture directional feature extraction processing on the echo-corrected image to capture the anisotropic characteristics of breast cancer tissue and obtain a texture feature image; The echo-corrected image, the acoustically enhanced image, and the texture feature image are fused and standardized to a uniform size and contrast range to obtain a standardized ultrasound image.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the standardized ultrasound image is divided into sector-shaped regions based on the radial fiber structure of the breast, the window size is dynamically adjusted according to tissue characteristics, and the sector-shaped regions are adaptively window segmented based on the window size to obtain a window set carrying anatomical structure position information, including: Performing fiber bundle tracking processing on the standardized ultrasound image according to the direction of Cooper's ligaments of the radial fiber structure of the breast, dividing the image into multiple sector-shaped areas along the fiber direction to obtain sector-shaped partitioned images; Performing tissue characteristic analysis on each sector in the sector-shaped partitioned image, identifying the distribution characteristics of glandular tissue, adipose tissue, and fibrous tissue through echo intensity, texture complexity, and acoustic attenuation coefficient, and obtaining a tissue characteristic distribution map; Dynamically adjust the window size according to the distribution density of different tissue types in the tissue characteristic distribution map to obtain an adaptive window size parameter; Window segmentation processing is performed on each sector area according to the adaptive window size parameter, and position information and corresponding tissue characteristic type in the radial fiber structure are marked for each window to obtain a window set carrying anatomical structure position information.

[0007] Optionally, in a third implementation of the first aspect of the present invention, performing fiber bundle tracking processing on the standardized ultrasound image according to the direction of Cooper's ligaments of the radial fiber structure of the breast, dividing the image into multiple sector-shaped areas along the fiber direction, and obtaining the sector-shaped partitioned image includes: performing gradient direction calculation processing on the standardized ultrasound image, detecting the direction of the high-echo fiber bundles of the Cooper's ligament using a Sobel operator and a direction filter, and obtaining a fiber direction field map; Performing principal direction extraction processing on the fiber direction field map, identifying the main fiber bundle directions radially distributed from a preset position by using Hough transform and linear regression algorithm, and obtaining a radial principal direction set; Performing sector angle calculation processing based on the radial main direction set, calculating the angle between each main direction with the preset position as the center point and determining the angle range of the sector area to obtain sector angle parameters; The standardized ultrasound image is subjected to regional segmentation processing according to the sector angle parameter, the image is divided into a plurality of sector regions along the radial main direction and the angle range of each region is marked to obtain a sector partitioned image.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, inputting the window set into a pre-trained cross-window Transformer model for feature extraction processing, analyzing the feature correlation pattern between windows through a cross-window attention mechanism, and obtaining a multi-dimensional ultrasound feature vector includes: Performing position encoding and feature embedding processing on each window in the window set, converting the anatomical structure position information and image features of the window into a high-dimensional vector representation, and obtaining a window feature embedding sequence; Performing cross-window attention weight calculation processing on the window feature embedding sequence, calculating the attention distribution between different receptive field windows through linear transformation of the query matrix, the key matrix and the value matrix, and obtaining a cross-window attention weight matrix; Performing weighted fusion processing on the window features according to the cross-window attention weight matrix, performing cross-correlation calculation on the local window features and the global window features to obtain fused window correlation features; The window-related features are subjected to multi-layer Transformer encoding processing, and deep semantic features are extracted through a multi-head self-attention mechanism and a feedforward neural network layer to obtain a multi-dimensional ultrasound feature vector.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, inputting the multidimensional ultrasound feature vector into the classifier of the cross-window Transformer model for molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result includes: performing molecular marker feature mapping processing on the multi-dimensional ultrasound feature vector, and performing correlation analysis between the multi-dimensional ultrasound feature vector and the molecular marker expression pattern to obtain a molecular marker feature representation; Performing a fully connected layer linear transformation on the molecular marker feature representation to map the high-dimensional feature vector to the feature space of four molecular subtypes to obtain subtype feature scores; Performing softmax normalization on the subtype feature scores, calculating the probability distribution of each molecular subtype and generating a confidence evaluation index to obtain a probability distribution vector; The maximum probability classification decision processing is performed according to the probability distribution vector, the molecular subtype with the highest probability value is selected as the prediction result and the corresponding confidence evaluation is output to obtain the breast cancer molecular subtype identification prediction result.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, performing molecular marker feature mapping processing on the multidimensional ultrasound feature vector, performing association analysis on the multidimensional ultrasound feature vector and the molecular marker expression pattern to obtain the molecular marker feature representation includes: Performing feature dimension transformation processing on the multi-dimensional ultrasound feature vector, mapping the high-dimensional feature vector to a feature subspace corresponding to the molecular marker through a linear projection layer, and obtaining a molecular marker subspace feature; Performing molecular marker expression pattern matching processing on the molecular marker subspace features, calculating the similarity score between the feature vector and the positive expression pattern of each molecular marker through a pre-trained marker expression pattern library, and obtaining a marker expression similarity matrix; Performing weighted fusion processing according to the marker expression similarity matrix, linearly combining the expression characteristics of the molecular markers according to the clinical importance weights, and obtaining comprehensive molecular marker characteristics; The comprehensive molecular marker features are subjected to nonlinear activation processing, and the feature expression capability is enhanced through ReLU activation function and batch normalization operation to obtain molecular marker feature representation.

[0011] A second aspect of the present invention provides a breast cancer molecular subtype identification and prediction system based on ultrasound images, the breast cancer molecular subtype identification and prediction system based on ultrasound images comprising: A preprocessing module is used to perform feature fusion preprocessing on the acquired breast ultrasound image based on acoustic impedance difference to obtain a standardized ultrasound image; a segmentation module, configured to divide the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjust the window size according to tissue characteristics, and perform adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; A feature extraction module is used to input the window set into a pre-trained cross-window Transformer model for feature extraction processing, analyze the feature correlation pattern between windows through the cross-window attention mechanism, and obtain a multi-dimensional ultrasound feature vector; The classification prediction module is used to input the multi-dimensional ultrasound feature vector into the classifier of the cross-window Transformer model to perform molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

[0012] The above-mentioned ultrasound image-based breast cancer molecular subtype identification and prediction method and system preprocesses the breast ultrasound image based on feature fusion based on acoustic impedance differences to obtain a standardized ultrasound image; divides the image into fan-shaped regions based on the radial fiber structure of the breast, and dynamically adjusts the window size according to tissue characteristics to perform adaptive window segmentation, obtaining a window set that carries anatomical structure location information; inputs the window set into a pre-trained cross-window Transformer model for feature extraction, and analyzes the feature correlation pattern between windows through a cross-window attention mechanism to obtain a multi-dimensional ultrasound feature vector; the feature vector is input into a classifier for molecular subtype classification prediction, obtaining a breast cancer molecular subtype identification prediction result. The present invention utilizes breast anatomical structure information through adaptive window segmentation and captures local and global feature correlations through a cross-window attention mechanism to improve prediction accuracy.

[0013] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The purposes and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of a first embodiment of a method for identifying and predicting breast cancer molecular subtypes based on ultrasound images in an embodiment of the present invention; Figure 2 Schematic diagram of an embodiment of a breast cancer molecular subtype identification and prediction system based on ultrasound images in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] The terms "including," "having," and any variations thereof, as used in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to the process, method, product, or device.

[0018] To facilitate understanding of this embodiment, a method for identifying and predicting breast cancer molecular subtypes based on ultrasound images disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, this method includes the following steps: 101. Perform feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image; In one embodiment of the present invention, the feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image includes: performing tissue interface boundary recognition processing on the acquired breast ultrasound image based on the reflection characteristics of ultrasound in the skin layer, subcutaneous fat layer, glandular layer, and pectoral muscle layer to obtain tissue stratification information; performing acoustic shadow compensation algorithm processing on the tissue stratification information to correct deep echo attenuation caused by changes in lesion tissue density to obtain an echo-corrected image; performing acoustic enhancement algorithm processing on the echo-corrected image to highlight tissue boundaries and internal structural features to obtain an acoustically enhanced image; performing texture directional feature extraction processing on the echo-corrected image to capture the anisotropic characteristics of breast cancer tissue to obtain a texture feature image; and fusing the echo-corrected image, acoustically enhanced image, and texture feature image and standardizing them to a uniform size and contrast range to obtain a standardized ultrasound image.

[0019] Specifically, during the feature fusion preprocessing of breast ultrasound images, tissue interface boundary identification is first required. Because ultrasound waves produce varying degrees of reflection when propagating through different tissues, this physical property can be exploited to identify tissue interfaces. The skin layer, due to its relatively high acoustic impedance, appears as bright, hyperechoic lines in ultrasound images, while the subcutaneous fat layer appears as a relatively uniform, hyperechoic area. The glandular layer, with its significantly lower acoustic impedance than adipose tissue, appears as hypoechoic or anechoic areas in the image, forming a stark contrast with the fat layer. The pectoral muscle layer, as a deep structure, has a relatively uniform, medium-echo intensity due to its acoustic impedance. By calculating the grayscale gradient change at each pixel in the image, locations with sudden changes in acoustic impedance can be detected. These locations correspond to the boundaries between different tissue layers. In implementation, an edge detection algorithm is used to calculate horizontal and vertical gradients. When the gradient value exceeds a preset threshold, the location is marked as a potential tissue interface. Subsequently, combined with the region growing algorithm, adjacent pixels with similar echo characteristics are classified into the same tissue layer, forming tissue stratification data containing the position information of the skin layer, fat layer, gland layer and pectoral muscle layer.

[0020] After obtaining tissue layer information, acoustic shadowing needs to be compensated. Acoustic shadowing is a common phenomenon in ultrasound imaging. When ultrasound waves encounter tissue structures or lesions with high acoustic impedance, the echo signal in the area behind them is significantly weakened. In breast cancer ultrasound images, the density variations of tumor tissue cause varying degrees of acoustic attenuation, resulting in an abnormal decrease in the echo intensity of normal tissue deep within the lesion. This phenomenon can interfere with feature extraction and analysis. The core of the acoustic shadowing compensation algorithm lies in establishing an attenuation model based on the physical laws of ultrasound propagation. The algorithm first analyzes the echo intensity variations along each beam path. By comparing the echo intensity differences between shallow and deep tissue types of the same type, the attenuation coefficient along the beam propagation direction is calculated. The theoretical echo intensity at different depths is then determined based on the tissue layer information and compared with the actual measured values. For areas where the echo intensity is significantly lower than the theoretical value, an interpolation algorithm is used to compensate for the signal, restoring the echo intensity of the deeper tissue to a near-normal level. This process generates echo-corrected images that not only preserve the original tissue structure information but also eliminate artifacts caused by acoustic shadowing.

[0021] In addition to echo correction, acoustic enhancement is required to highlight tissue boundaries and internal structural features. This acoustic enhancement algorithm primarily addresses the issue of insufficient contrast in ultrasound images by enhancing echogenicity differences between different tissues to improve image quality. This algorithm utilizes adaptive histogram equalization (AHE) technology to dynamically adjust contrast enhancement parameters based on the grayscale distribution characteristics of local image regions. Within tissue boundaries, the algorithm prioritizes echo gradient enhancement, making boundary contours clearer and sharper. Within tissue interiors, the algorithm focuses on improving texture contrast, highlighting subtle structural differences. In implementation, the algorithm divides the image into multiple overlapping subregions, calculates the grayscale histogram for each subregion, and then determines a contrast enhancement function based on the histogram distribution. The enhancement results from each subregion are combined using a weighted average, avoiding the over-processing problem caused by global enhancement. Furthermore, the algorithm specifically optimizes processing strategies for hypoechoic and hyperechoic regions, tailored to the characteristics of breast ultrasound images, ensuring contrast enhancement without losing important diagnostic information.

[0022] Texture directional feature extraction is crucial for capturing the anisotropic characteristics of breast cancer tissue. Due to its pathological characteristics, breast cancer tissue exhibits a texture pattern distinct from that of normal tissue in ultrasound images. Malignant tumor tissue exhibits irregular internal structures and disordered cell arrangement, resulting in anisotropic texture features in ultrasound images. The texture directional feature extraction algorithm uses a Gabor filter bank to analyze the texture response of the image at different directions and frequencies. The algorithm constructs a Gabor filter bank consisting of eight different directions and four different frequencies, covering the directional range from 0 to 180 degrees. Each Gabor filter performs a convolution operation on the echo-corrected image to extract texture information at a specific direction and frequency. By calculating the energy distribution of the filter response in each direction, the directional characteristics of tissue texture can be quantified. For isotropic normal tissue, the response energy is relatively uniform across all directions, while for anisotropic tumor tissue, specific directions exhibit significantly higher response energy. The algorithm also introduces a texture directional consistency index, which quantifies the degree of texture directionality by calculating the energy ratio of the primary to secondary directions. The generated texture feature images show the texture directional characteristics of different regions in the form of pseudo-color, providing important texture information in molecular subtype identification.

[0023] Three-channel image fusion combines the echo-corrected image, acoustically enhanced image, and texture feature image to form a unified multidimensional feature representation. The fusion algorithm uses a weighted linear combination approach, assigning different weight coefficients to the three images based on their importance in molecular subtype identification. The echo-corrected image preserves the original and complete tissue structural information and is given a higher weight. The acoustically enhanced image highlights tissue boundaries and contrast information, which helps identify tumor morphological features. The texture feature image provides directional information on tissue microstructure, which is uniquely valuable for distinguishing different molecular subtypes. The fusion process first resizes the three images to the same size and then performs a pixel-wise weighted summation. The fused image is then contrast-normalized, mapping pixel values ​​to a standard grayscale range of 0-255 to ensure balanced contributions from the different source images. The resulting normalized ultrasound image retains the essential structural information of the original ultrasound image while incorporating the characteristic information of acoustic enhancement and texture directionality, playing a fundamental role in window segmentation and feature extraction.

[0024] 102. Divide the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjust the window size according to tissue characteristics, and perform adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; In one embodiment of the present invention, the standardized ultrasound image is divided into sector-shaped areas based on the radial fiber structure of the breast, and the window size is dynamically adjusted according to tissue characteristics. The sector-shaped areas are adaptively segmented based on the window size to obtain a window set carrying anatomical structure location information, including: performing fiber bundle tracking processing on the standardized ultrasound image according to the direction of Cooper's ligaments of the radial fiber structure of the breast, dividing the image into multiple sector-shaped areas along the fiber direction to obtain a sector-shaped partitioned image; performing tissue characteristic analysis processing on each sector-shaped area in the sector-shaped partitioned image, identifying the distribution characteristics of glandular tissue, adipose tissue and fibrous tissue through echo intensity, texture complexity and acoustic attenuation coefficient to obtain a tissue characteristic distribution map; dynamically adjusting the window size according to the distribution density of different tissue types in the tissue characteristic distribution map to obtain an adaptive window size parameter; performing window segmentation processing on each sector-shaped area according to the adaptive window size parameter, and marking the position information and corresponding tissue characteristic type in the radial fiber structure for each window to obtain a window set carrying anatomical structure location information.

[0025] Specifically, tracking Cooper's ligament fiber bundles is a core step in image segmentation based on the radial fiber structure of the breast. Cooper's ligaments are important support structures within the breast, composed of dense connective tissue. In ultrasound images, they appear as hyperechoic fiber bundles radiating from the nipple to the chest wall. These fiber bundles not only support breast tissue but, more importantly, segment the breast into relatively independent anatomical regions. Breast cancer in different regions exhibits distinct biological behaviors and molecular characteristics. The fiber bundle tracking algorithm first performs directional filtering on standardized ultrasound images, using a set of linear filters at different angles to detect linear structures within the image. Since Cooper's ligaments appear as continuous, hyperechoic linear structures in ultrasound images, the algorithm identifies the fiber orientation by calculating the echo intensity gradient of each pixel in different directions. In implementation, the algorithm sets a starting point at the center of the image and then continuously tracks along the detected hyperechoic direction. A dynamic threshold strategy is employed during tracking to automatically adjust tracking sensitivity based on changes in local echo intensity, ensuring accurate tracking of the complete path of the fiber bundles. When a significant drop in echo intensity or a sharp change in direction is detected, the algorithm concludes that the current fiber bundle has been traced and begins searching for the next one. In this way, the algorithm can identify all major radial fiber bundles in the image and divide the entire image into multiple sectors based on the angular relationships between adjacent fiber bundles.

[0026] After the sector regions are delineated, detailed tissue characteristic analysis is required for each region. Breast tissue primarily consists of glandular tissue, adipose tissue, and fibrous tissue, each of which exhibits distinct acoustic characteristics in ultrasound images. Due to its high cell density and abundant glandular ductal structures, glandular tissue appears as a relatively uniform hypoechoic region in ultrasound images, with echo intensity values ​​generally ranging from 40 to 80. Due to its low density and relatively simple internal structure, adipose tissue appears as a hyperechoic region in ultrasound images, with grayscale values ​​ranging from 120 to 200. Fibrous tissue, due to its dense arrangement of collagen fibers, exhibits hyperechoic characteristics, with grayscale values ​​ranging from 150 to 255. The tissue characteristic analysis algorithm statistically analyzes the echo intensity distribution histogram of pixels within each sector region to determine the primary tissue components. The algorithm calculates statistical parameters such as the histogram's peak position, distribution width, and skewness, which reflect the uniformity and complexity of the tissue within the region. Texture complexity analysis uses a gray-level co-occurrence matrix method to calculate image texture feature parameters at different directions and distances, including contrast, correlation, energy, and entropy. Glandular tissue exhibits high texture complexity due to its complex internal structure, while adipose tissue has a relatively simple texture. The acoustic attenuation coefficient is calculated by analyzing the echo intensity variations at different depths within the same sector. Glandular tissue exhibits a significantly higher acoustic attenuation coefficient than adipose tissue.

[0027] The dynamic adjustment of the window size is based on the complexity of the tissue distribution. In areas with dense glandular tissue, due to its complex internal structure and close association with breast cancer, a smaller window size is required to capture subtle structural changes and pathological features. The algorithm calculates the window size based on the proportion of glandular tissue within the sector. When the proportion of glandular tissue exceeds 60%, the window size is set to 8×8 pixels; when the proportion is between 30% and 60%, the window size is adjusted to 12×12 pixels. In areas dominated by adipose tissue, due to its relatively uniform structure and less pathological changes, a larger window size of 16×16 pixels can be used to extract macroscopic features. In areas with concentrated fibrous tissue, due to its linear structure, a rectangular window size of 6×18 pixels is more appropriate to accommodate the directional characteristics of the fibers. The algorithm also considers the mixing of different tissue types within the same sector, calculating the window size for mixed areas through a weighted average. Furthermore, the algorithm incorporates a boundary constraint mechanism to ensure that the window does not cross the boundaries between different tissue types, preventing the mixing of different tissue features and affecting feature extraction accuracy.

[0028] Window segmentation and position labeling are the final stages of the entire processing workflow. Based on the calculated adaptive window size parameters, the algorithm segments each sector into non-overlapping windows. The segmentation process employs a sliding window strategy, starting from the vertex of each sector and sequentially placing windows in the radial and tangential directions. Due to the geometric characteristics of the sectors, the radial window spacing remains fixed, while the tangential spacing is dynamically adjusted based on the distance from the center point to ensure uniform window distribution. Each window is assigned a series of identifying information upon creation, including the sector number to which it belongs, its relative position within the sector, the window's geometric dimensions, and the predominant tissue type at that location. Position labeling uses a polar coordinate system, with the image center as the origin, recording the angle and radial distance of each window's center point. This labeling method accurately reflects the anatomical position of the window within the radial fibrous structure of the breast. Tissue type labeling, based on the previously analyzed tissue characteristics, categorizes each window as glandular, fat, fibrous, or mixed. The algorithm also records the proportion of the predominant tissue type within each window, facilitating subsequent feature extraction to adjust the processing strategy based on tissue characteristics. The generated window set not only preserves the spatial structural information of the image, but also incorporates rich anatomical and histological prior knowledge, which plays an important role in the cross-window attention mechanism.

[0029] Furthermore, the standardized ultrasound image is subjected to fiber bundle tracking processing according to the direction of Cooper's ligament of the radial fiber structure of the breast, and the image is divided into multiple sector-shaped areas along the fiber direction to obtain the sector-shaped partitioned image, including: performing gradient direction calculation processing on the standardized ultrasound image, detecting the direction of the high-echo fiber bundle of the Cooper's ligament by the Sobel operator and the direction filter, and obtaining a fiber direction field map; performing main direction extraction processing on the fiber direction field map, identifying the main fiber bundle direction radially distributed from the preset position by the Hough transform and the linear regression algorithm, and obtaining a radial main direction set; performing sector angle calculation processing according to the radial main direction set, calculating the angle between each main direction with the preset position as the center point and determining the angle range of the sector area, and obtaining the sector angle parameter; performing region segmentation processing on the standardized ultrasound image according to the sector angle parameter, dividing the image into several sector-shaped areas along the radial main direction and marking the angle range of each area, and obtaining the sector-shaped partitioned image.

[0030] Specifically, the Sobel operator, a classic edge detection tool, detects the intensity and direction of grayscale changes by calculating the first-order derivatives of the image in the horizontal and vertical directions. Because Cooper's ligament appears as a hyperechoic linear structure in ultrasound images, creating a significant grayscale contrast with surrounding tissue, the Sobel operator is effective in identifying these linear edge features. In its implementation, the operator convolves the standardized ultrasound image with 3×3 horizontal and vertical convolution kernels, respectively. The horizontal Sobel kernel primarily detects vertical edges, while the vertical Sobel kernel detects horizontal edges. By calculating the magnitude and angle of the two directional gradient components, the gradient strength and direction at each pixel can be determined. However, the Sobel operator alone can only detect local edge features and cannot accurately identify fiber structures with specific directionality. Therefore, a directional filter is needed to enhance linear features in specific directions. The directional filter consists of a set of linear filters with different angles, each corresponding to a specific angle range. Twelve directional filters are constructed at intervals of 15 degrees from 0 to 180 degrees. Each directional filter uses a rectangular kernel structure with a length of 21 pixels and a width of 3 pixels, with weights uniformly distributed along the kernel's length. The maximum response is achieved when the directional filter coincides with a linear structure in the image with the same orientation. By comparing the response strengths of each directional filter, the most likely fiber orientation at each pixel can be determined. Combining the gradient information of the Sobel operator with the directional responses of the directional filters, the resulting fiber orientation field map contains fiber orientation information and confidence values ​​for each location in the image.

[0031] After the fiber orientation field map is established, the principal direction extraction process is required to identify the true orientation of the Cooper ligament. The Hough transform is an effective method for detecting straight line features in images. Its basic principle is to transform the line detection problem in image space into a peak detection problem in parameter space. In polar coordinate representation, each line in the image can be uniquely identified by its perpendicular distance ρ from the origin and the angle θ between the perpendicular and the x-axis. The Hough transform maps each edge point in the image to a sine curve in parameter space. When multiple sine curves intersect at a point in parameter space, it indicates that a line exists in the image space passing through these edge points. The algorithm first binarizes the fiber orientation field map and marks pixels with a fiber orientation confidence score above a set threshold as candidate edge points. The Hough transform is then performed on each candidate edge point, accumulating votes in the parameter space. The parameter space is divided into a grid of ρ and θ, where ρ ranges from the negative to the positive image diagonal length and θ ranges from 0 to 180 degrees. After voting, the algorithm searches for local maxima in the parameter space. These peaks correspond to straight line features in the image. Because the Cooper's ligaments of the breast radiate from a central location, the algorithm focuses specifically on lines passing through a pre-set central location. By constraining the Hough transform's search range, only candidate lines passing through the central location are retained. A linear regression algorithm further optimizes the detected line parameters, fitting the pixels on each line using the least squares method to obtain a more accurate line equation. This process generates a set of radial principal directions that encompasses all the major fiber bundle directions radiating outward from the central location, with each direction represented by an angle value and a confidence level.

[0032] The sector angle calculation process determines the angular parameters for image segmentation based on a set of radial principal directions. Since the distribution of Cooper's ligaments exhibits certain anatomical regularities, the angular spacing between adjacent fiber bundles is relatively stable, providing a theoretical basis for sector segmentation. The algorithm first sorts the angle values ​​in the set of radial principal directions and then calculates the angular differences between adjacent directions. Normally, the main Cooper's ligaments are distributed within the four quadrants of the breast, with each quadrant containing two to three major fiber bundles, and the angular spacing between adjacent fiber bundles ranges from 20 to 45 degrees. The algorithm identifies the boundaries of the sector regions by analyzing the distribution pattern of these angular differences. When the angular difference between two adjacent principal directions is significantly greater than the average spacing, the algorithm demarcates the sector region. The algorithm also considers the confidence level of fiber bundle detection, assigning less weight to directions with lower confidence levels in sector segmentation. For adjacent directions with too small an angular spacing, the algorithm merges them into a single direction using a weighted average to avoid creating overly narrow sectors. The fan angle parameters not only include the starting and ending angles of each fan-shaped region, but also record the direction and distribution density of the main fiber bundles in the region. These parameters play a guiding role in regional segmentation, ensuring that each fan-shaped region corresponds to a relatively independent anatomical unit.

[0033] The region segmentation process uses a preset center position as the common vertex of all sector regions. Segmentation lines are drawn on the standardized ultrasound image based on the sector angle parameters. Each segmentation line originates from the center point and extends along the corresponding angle to the image boundary. Segmentation line drawing utilizes a linear rasterization algorithm to ensure the continuity and accuracy of the segmentation boundary. Because the rectangular boundary of the ultrasound image does not perfectly match the sector geometry, the algorithm needs to handle intersections between the sector and the image boundary. When a segmentation line reaches the image boundary, the area between two adjacent segmentation lines is defined as a sector. The algorithm assigns a unique identification number to each sector region and records its angular range, the set of pixel coordinates contained within it, and the dominant fiber direction within the region. The region labeling process utilizes a region filling algorithm, starting from the center of each sector region and extending outward until it encounters the segmentation boundary. The resulting sector-partitioned image uses different grayscale values ​​or pseudocolors to distinguish each sector region, with each region's pixel value corresponding to its region number. This labeling method facilitates subsequent processing algorithms to quickly locate and access specific sectors based on region numbers. The sector-partitioned image retains all structural information of the original ultrasound image while incorporating anatomically informed regional delineation information.

[0034] 103. Input the window set into a pre-trained cross-window Transformer model for feature extraction, analyze the feature correlation pattern between windows through a cross-window attention mechanism, and obtain a multi-dimensional ultrasound feature vector; In one embodiment of the present invention, the window set is input into a pre-trained cross-window Transformer model for feature extraction processing, and the feature association pattern between windows is analyzed through a cross-window attention mechanism to obtain a multi-dimensional ultrasound feature vector, including: performing position encoding and feature embedding processing on each window in the window set, converting the anatomical structure position information and image features of the window into a high-dimensional vector representation, and obtaining a window feature embedding sequence; performing cross-window attention weight calculation processing on the window feature embedding sequence, calculating the attention distribution between windows with different receptive fields through linear transformation of the query matrix, key matrix and value matrix, and obtaining a cross-window attention weight matrix; performing weighted fusion processing on the window features according to the cross-window attention weight matrix, performing cross-correlation calculation on the local window features and the global window features, and obtaining the fused window association features; performing multi-layer Transformer encoding processing on the window association features, extracting deep semantic features through a multi-head self-attention mechanism and a feedforward neural network layer, and obtaining a multi-dimensional ultrasound feature vector.

[0035] Specifically, during feature extraction in the cross-window Transformer model, position encoding and feature embedding processes transform the spatial position information and image content of each window into numerical representations that the model can understand. Position encoding utilizes a learnable embedding approach, mapping the polar coordinate position of each window within the breast anatomy into a 128-dimensional vector space. Angular information is encoded using a combination of sine and cosine functions. Specifically, the angle values ​​are divided by different powers of 10,000 to generate periodic functions of varying frequencies, ensuring smooth similarity between the encoded vectors of adjacent angular positions. Radial distance information is encoded using a logarithmic scale, as the biological characteristics of breast tissue vary nonlinearly at different depth levels. The algorithm takes the logarithm of the radial distance values ​​and linearly maps them into a 64-dimensional subspace. In addition to spatial position information, position encoding also incorporates the window's tissue type identity, mapping the four types of tissue types—glandular, adipose, fibrous, and mixed—to different 64-dimensional encoding subspaces. Feature embedding extracts image features within each window and converts them into fixed-dimensional vector representations. The algorithm first extracts features from the pixels within a window using multiple convolution kernels of varying sizes. These include 16 3×3 kernels to capture detailed textures, 8 5×5 kernels to extract mid-scale structural features, and 4 7×7 kernels to capture macroscopic morphological information. Each kernel outputs a feature map, which is then compressed into a 512-dimensional feature vector using global average pooling. This feature vector is concatenated with a 128-dimensional positional encoding vector to form a 640-dimensional feature embedding for the entire window. The feature embeddings for all windows are then arranged according to their angular order within the sector, forming a feature embedding sequence with a length equal to the total number of windows.

[0036] Cross-window attention weight calculation captures cross-regional feature dependencies by establishing correlations between windows with different receptive fields. The algorithm divides the window feature embedding sequence into two parallel processing paths. The backbone path contains windows with local receptive fields, which are small and focus on detailed features within a single sector. The branch path contains windows with global receptive fields, which span multiple sectors to capture a wide range of anatomical information. The query matrix Q, key matrix K, and value matrix V are constructed using linear transformation layers. For the local receptive field window, the algorithm uses three different linear transformation layers to map the 640-dimensional window feature embedding vector to a 64-dimensional query space, a 64-dimensional key space, and a 512-dimensional value space, respectively. For the global receptive field window, due to the need to handle more complex cross-regional information, the dimensions of its query, key, and value matrices are set to 128, 128, and 768, respectively. Attention weights are calculated using a scaled dot product method: the algorithm calculates the dot product between the local window query vector and the global window key vector, and then scales it by dividing it by the square root of the key vector's dimensions to avoid vanishing gradients caused by excessively large dot products. The obtained similarity scores are normalized by the softmax function to ensure that the sum of the attention weights of each local window to all global windows is equal to 1. The generated cross-window attention weight matrix is ​​an M×N matrix, where M represents the number of local windows, N represents the number of global windows, and each element in the matrix represents the strength of the association between the corresponding local window and the global window.

[0037] The weighted fusion of window features converts attention weights into specific feature representation updates. This fusion process is implemented through matrix multiplication: the M×N attention weight matrix is ​​multiplied by the N×768 global window value matrix to produce an M×768 fused feature matrix. This calculation effectively performs a weighted average of the global window value vector for each local window, based on its attention weight to the global window. The fused feature vector contains information from the global context, but its dimensionality differs from that of the original local window features. Therefore, a linear projection layer is used to compress the 768-dimensional fused features back to 512 dimensions to allow subsequent operations with the original local window features. The algorithm also incorporates a residual connection mechanism to add the compressed fused features to the original local window feature vector. This design ensures that the original local information is preserved during the fusion process. Weighted fusion also takes into account the unique processing requirements of different tissue type windows. Glandular tissue windows give higher weight to information from other glandular region windows during fusion, while adipose tissue windows prioritize structural information from tissue boundary regions. The algorithm achieves this differentiated processing by introducing a tissue type similarity weighting term in the attention weight calculation, and the attention weights between windows of the same tissue type are additionally enhanced.

[0038] The multi-layer Transformer encoding process uses a six-layer encoder structure to extract deep semantic information from window-related features. Each encoder layer consists of a multi-head self-attention sublayer and a position feedforward network sublayer, both equipped with residual connections and layer normalization. The multi-head self-attention mechanism projects the 640-dimensional window-related features into eight different attention heads, each with an 80-dimensional query, key, and value vector. The eight attention heads compute their own attention outputs in parallel, and all outputs are concatenated to produce a 640-dimensional multi-head attention result. Different attention heads focus on different aspects of the features. The first two heads primarily learn the changing patterns of texture details, the middle three heads focus on morphological features and boundary information, and the last three heads specialize in spatial relationships and hemodynamic-related signal features. The position feedforward network uses a two-layer fully connected structure to achieve nonlinear feature transformation. The first layer expands the 640-dimensional input to 2048 dimensions, introducing nonlinearity using the GELU activation function. The second layer compresses the feature dimensions back to 640. This expansion-then-compression design increases the model's parameter capacity, enabling it to learn more complex feature transformation relationships. Layer normalization normalizes the feature vector at each location, calculating its mean and variance, and then normalizing it to zero mean and unit variance. Through layer-by-layer processing by the six-layer encoder, the initial feature representation of each window gradually evolves into an abstract feature vector containing rich semantic information. These multidimensional ultrasound feature vectors not only retain the texture and morphological information of the original image but also incorporate cross-regional spatial correlation information and high-level semantic features learned through deep network learning.

[0039] 104. Input the multi-dimensional ultrasound feature vector into the classifier of the cross-window Transformer model to perform molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

[0040] In one embodiment of the present invention, the step of inputting the multidimensional ultrasound feature vector into the classifier of the cross-window Transformer model for molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result includes: performing molecular marker feature mapping processing on the multidimensional ultrasound feature vector, performing correlation analysis on the multidimensional ultrasound feature vector and the molecular marker expression pattern to obtain a molecular marker feature representation; performing fully connected layer linear transformation processing on the molecular marker feature representation, mapping the high-dimensional feature vector to the feature space of four molecular subtypes to obtain a subtype feature score; performing softmax normalization processing on the subtype feature score, calculating the probability distribution of each molecular subtype and generating a confidence assessment index to obtain a probability distribution vector; performing maximum probability classification decision processing based on the probability distribution vector, selecting the molecular subtype with the highest probability value as the prediction result and outputting the corresponding confidence assessment to obtain a breast cancer molecular subtype identification prediction result.

[0041] Specifically, at the initial stage of molecular subtype classification prediction, molecular marker feature mapping involves converting abstract multidimensional ultrasound feature vectors into representations associated with molecular biological features of breast cancer. This mapping process is based on a library of molecular marker expression patterns established from extensive clinical pathological data. This library contains statistical patterns of key molecular markers, such as estrogen receptor, progesterone receptor, human epidermal growth factor receptor 2, and Ki-67 proliferation index, in different ultrasound imaging manifestations. Molecular marker feature mapping is implemented through four parallel feature mapping branches, one for each key molecular marker. The estrogen receptor mapping branch utilizes a network structure consisting of two fully connected layers. The first layer compresses the 640-dimensional multidimensional ultrasound feature vector to 256 dimensions, and the second layer further compresses it to 128 dimensions. This 128-dimensional vector represents the ultrasound feature pattern associated with estrogen receptor expression. The progesterone receptor mapping branch adopts a similar structure, but uses different weighting parameters in the first layer to accommodate the specific expression patterns of progesterone receptor. Because human epidermal growth factor receptor 2 (HER2) often appears in ultrasound images as correlated with angiogenesis and tissue invasiveness, its mapping branch incorporates an attention mechanism layer into the network structure. This layer automatically identifies and enhances feature components related to the degree of vascularization. The Ki-67 proliferation index mapping branch specifically focuses on texture features associated with the activeness of cell proliferation, using a specially designed convolutional layer to extract high-frequency texture information. The outputs of the four mapping branches are concatenated into a 512-dimensional comprehensive molecular marker feature vector, which contains expression predictions for all major molecular markers. A cross-validation mechanism is also introduced into the mapping process to verify the consistency of the mapping results by comparing the correlations between different molecular markers. For example, estrogen receptors and progesterone receptors show a positive correlation in most cases, while triple-negative breast cancer shows low expression of all receptors.

[0042] After generating the molecular marker feature representation, these biologically relevant features need to be linearly transformed into a specific molecular subtype feature space through fully connected layers. This transformation process utilizes a multi-layer perceptron architecture consisting of three fully connected layers and two intermediate dropout layers. The first fully connected layer expands the 512-dimensional molecular marker feature vector to 1024 dimensions. This dimensionality expansion increases the capacity of the feature representation, enabling the model to learn more complex molecular subtype discriminative patterns. Reinforced linear unit (ReLU) activation functions introduce nonlinear transformations while maintaining computational simplicity and numerical stability. The first dropout layer randomly sets some neuron outputs to zero with a probability of 0.3. This regularization technique prevents the model from overfitting the training data. The second fully connected layer compresses the feature dimensionality from 1024 to 256 dimensions. This process acts as feature selection and dimensionality reduction, retaining the most discriminative feature components. The second dropout layer also performs random dropout with a probability of 0.3. The third fully connected layer is the classification output layer, which directly maps the 256-dimensional features into a 4-dimensional output space, with each dimension corresponding to a molecular subtype. These four output dimensions represent the characteristic scores for Luminal A, Luminal B, HER2-enriched, and triple-negative subtypes, respectively. The weight parameters of the linear transformation are optimized during training using a backpropagation algorithm, ensuring that the output subtype characteristic scores accurately reflect the likelihood that the input sample belongs to each molecular subtype. The resulting subtype characteristic scores are unnormalized real-valued values ​​that reflect the degree of match between the sample and the corresponding molecular subtype characteristic pattern. Positive values ​​indicate a strong match, while negative values ​​indicate a weak match.

[0043] Softmax normalization of subtype feature scores converts the original real-number scores into a standard probability distribution. The softmax function is a commonly used activation function in multi-class classification problems. Its mathematical principle is to amplify the differences between class scores using an exponential function, and then normalize the scores so that the sum of the probabilities of all classes equals 1. In implementation, the algorithm first calculates the exponential value of each subtype feature score. This operation amplifies the differences between classes with higher scores and classes with lower scores. Each class exponential value is then divided by the sum of all class exponential values ​​to obtain the predicted probability for each molecular subtype. A key feature of softmax normalization is that it preserves the relative magnitude of the original scores; the class with the highest score still has the highest probability after normalization. Normalization also generates confidence metrics that measure the reliability of the prediction results. The first confidence metric is the maximum probability value itself. A maximum probability close to 1 indicates that the model is very confident in the prediction, while a maximum probability close to 0.25 indicates that the model cannot make a clear decision between the four classes. The second confidence metric is the entropy of the probability distribution. A lower entropy indicates a more concentrated probability distribution and a more certain prediction. The third confidence metric is the difference between the highest and second highest probabilities. A larger difference indicates a greater degree of discrimination between the predicted class and the other classes. The resulting probability distribution vector contains four elements, one for each of the four molecular subtypes. All elements of this vector have values ​​between 0 and 1, and their sum equals 1.

[0044] The classification decision process is the decision-making stage of the entire prediction process. It involves determining the sample's molecular subtype based on the probability distribution vector and assessing the reliability of the prediction. Maximum probability classification is the most straightforward decision-making strategy. The algorithm searches for the element with the largest value in the probability distribution vector, and the molecular subtype corresponding to that element is selected as the prediction. However, a simple maximum probability decision can produce unstable results when the probabilities are close. Therefore, the algorithm incorporates a confidence threshold mechanism. When the maximum probability falls below the preset confidence threshold, the algorithm outputs an "uncertain" prediction, indicating that further investigation or analysis is required. The confidence threshold is set differently based on the clinical importance of different molecular subtypes. For subtypes with significantly different treatment options, a higher confidence threshold is set to ensure prediction accuracy. The algorithm also considers the biological relevance of molecular subtypes. If the prediction result is Luminal B but the confidence is low, the algorithm checks the probability of Luminal A. If the two probabilities are similar, the algorithm outputs a "Luminal type, further subtype determination required" result. For the prediction of HER2-enriched subtypes, the algorithm pays special attention to distinguishing them from triple-negative subtypes because these two subtypes have significant differences in treatment strategies. The output of the prediction results contains three parts of information: molecular subtype category, corresponding probability value, and comprehensive confidence score. The confidence score is based on the weighted average of multiple confidence indicators, ranging from 0 to 1, and the higher the value, the more reliable the prediction. The generated breast cancer molecular subtype identification prediction results not only provide clear classification information, but also contain detailed probability distribution and reliability assessment. This information provides important reference for clinicians' diagnosis and treatment decisions.

[0045] Furthermore, the multi-dimensional ultrasound feature vector is subjected to molecular marker feature mapping processing, and the multi-dimensional ultrasound feature vector is subjected to correlation analysis with the molecular marker expression pattern to obtain the molecular marker feature representation, including: performing feature dimension transformation processing on the multi-dimensional ultrasound feature vector, mapping the high-dimensional feature vector to the feature subspace corresponding to the molecular marker through a linear projection layer, and obtaining the molecular marker subspace feature; performing molecular marker expression pattern matching processing on the molecular marker subspace feature, and calculating the similarity score between the feature vector and the positive expression pattern of each molecular marker through a pre-trained marker expression pattern library to obtain a marker expression similarity matrix; performing weighted fusion processing according to the marker expression similarity matrix, linearly combining the expression features of the molecular markers according to the clinical importance weight, and obtaining the comprehensive molecular marker feature; performing nonlinear activation processing on the comprehensive molecular marker feature, enhancing the feature expression ability through the ReLU activation function and batch normalization operation, and obtaining the molecular marker feature representation.

[0046] Specifically, in the initial stage of molecular marker feature mapping, feature dimensionality transformation involves decomposing the 640-dimensional multidimensional ultrasound feature vector into feature subspaces corresponding to the four primary molecular markers. This transformation is implemented using four parallel linear projection layers, each dedicated to extracting features for a specific molecular marker. The estrogen receptor projection layer uses a 640×160 weight matrix to map the original feature vector into a 160-dimensional estrogen receptor subspace. The progesterone receptor projection layer similarly maps the 640-dimensional features into a 160-dimensional subspace, but its weight matrix is ​​specifically trained to identify characteristic patterns associated with progesterone receptor expression. The output dimension of the human epidermal growth factor receptor 2 projection layer is set to 128 dimensions because the ultrasound appearance of HER2 is relatively specific, primarily related to tumor vascularization and invasiveness. The Ki-67 proliferation index projection layer outputs 192-dimensional features. This larger dimension is due to the correlation between the proliferation index and multiple ultrasound features, including texture complexity, boundary definition, and internal echo homogeneity. The computational process of each of the four projection layers is a simple matrix multiplication operation, multiplying the input feature vector by the corresponding weight matrix to produce the projection result. The projection process also includes the addition of a bias term, with each subspace having an independent bias vector to adjust the baseline level of the feature.

[0047] Molecular marker expression pattern matching uses a pre-trained marker expression pattern library to assess the similarity of each subspace feature to known expression patterns. This expression pattern library, established by analyzing a large number of immunohistochemically validated breast cancer samples, contains typical ultrasound feature patterns corresponding to each molecular marker's positive expression state. For the estrogen receptor, the pattern library includes three typical patterns: high expression, moderate expression, and low expression, each represented by a 160-dimensional feature vector. High expression patterns reflect the typical ultrasound characteristics of hormone-sensitive tumors, such as relatively regular morphology, clear boundaries, and uniform internal echogenicity. The pattern library for human epidermal growth factor receptor 2 specifically emphasizes features related to angiogenesis, including enhanced blood flow signals around the tumor, irregular boundaries, and heterogeneous internal echogenicity. The pattern library for the Ki-67 proliferation index includes ultrasound feature patterns corresponding to low, moderate, and high proliferation states. Similarity is calculated using the cosine similarity method, measuring the degree of similarity by calculating the cosine of the angle between the subspace feature vector and each pattern vector in the pattern library. The computational process first L2-normalizes the feature vector and pattern vector, then calculates the dot product of the two normalized vectors. Each molecular marker subspace feature is then compared to all patterns in the corresponding pattern library to generate a similarity score vector.

[0048] The weighted fusion process assigns weights and combines features based on the clinical importance of each molecular marker in the classification of breast cancer molecular subtypes. In the molecular classification system for breast cancer, estrogen receptor and progesterone receptor are primary determinants of the luminal subtype and are therefore given higher weights. The weight coefficient for estrogen receptor is set at 0.35, and the weight coefficient for progesterone receptor is set at 0.25, with the combined weight accounting for 60% of the total weight. Human epidermal growth factor receptor 2 is a decisive marker for the HER2-enriched subtype, with a weight coefficient of 0.3. The Ki-67 proliferation index is primarily used to distinguish between luminal A and luminal B subtypes, with a weight coefficient of 0.1. The fusion process uses a weighted average method, multiplying the highest similarity score for each molecular marker by the corresponding weight coefficient, and then summing the results to obtain a composite score. The algorithm first selects the maximum value from the similarity score vector for each molecular marker, which represents the similarity between the marker and its most closely matched expression pattern. The maximum similarity of each of the four molecular markers is then multiplied by its weight coefficient to obtain a weighted similarity. The algorithm also takes into account the interaction between molecular markers. When both estrogen receptors and progesterone receptors show high similarity, additional synergistic weighting will be given to reflect the synergistic effect of the two receptors in biological functions.

[0049] Nonlinear activation processing further processes and optimizes the integrated molecular marker features through the ReLU activation function and batch normalization. The ReLU activation function sets all negative values ​​to zero and retains only positive values. This processing method is consistent with the biological characteristics of molecular marker expression, as the expression intensity of the marker does not have a negative value. Batch normalization is performed after ReLU activation and normalizes the features by calculating the mean and variance of the samples in the current batch. The batch normalization calculation process includes four steps: first, the mean of the features of all samples in the current batch is calculated, then the variance is calculated, and then the mean and variance are used to normalize the features to zero mean and unit variance. Learnable scaling and offset parameters are applied to the normalized results. This processing method ensures that the feature distribution of different samples is consistent, avoiding training instability caused by differences in the input data distribution. After ReLU activation and batch normalization, the generated molecular marker feature representation has good numerical properties and biological interpretability.

[0050] In this embodiment, a standardized ultrasound image is obtained by preprocessing the breast ultrasound image with features based on acoustic impedance differences. The image is divided into fan-shaped regions based on the radial fiber structure of the breast, and the window size is dynamically adjusted according to tissue characteristics to perform adaptive window segmentation, thereby obtaining a window set that carries anatomical structure location information. The window set is input into a pretrained cross-window Transformer model for feature extraction, and the feature correlation pattern between windows is analyzed through a cross-window attention mechanism to obtain a multi-dimensional ultrasound feature vector. The feature vector is input into a classifier for molecular subtype classification prediction, thereby obtaining a breast cancer molecular subtype identification prediction result. The present invention utilizes breast anatomical structure information through adaptive window segmentation and captures local and global feature correlations through a cross-window attention mechanism to improve prediction accuracy.

[0051] The above describes the method for identifying and predicting breast cancer molecular subtypes based on ultrasound images in an embodiment of the present invention. The following describes the system for identifying and predicting breast cancer molecular subtypes based on ultrasound images in an embodiment of the present invention. Figure 2 In one embodiment of the present invention, a breast cancer molecular subtype identification and prediction system based on ultrasound images includes: A preprocessing module 201 is configured to perform feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image; a segmentation module 202 for dividing the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjusting the window size according to tissue characteristics, and performing adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; A feature extraction module 203 is configured to input the window set into a pre-trained cross-window Transformer model for feature extraction, analyze the feature correlation pattern between windows through a cross-window attention mechanism, and obtain a multi-dimensional ultrasound feature vector; The classification prediction module 204 is configured to input the multi-dimensional ultrasound feature vector into the classifier of the cross-window Transformer model to perform molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

[0052] In an embodiment of the present invention, the ultrasound image-based breast cancer molecular subtype identification and prediction system runs the above-mentioned ultrasound image-based breast cancer molecular subtype identification and prediction method. The ultrasound image-based breast cancer molecular subtype identification and prediction system performs feature fusion preprocessing based on acoustic impedance differences on the breast ultrasound image to obtain a standardized ultrasound image; divides the image into fan-shaped areas based on the radial fiber structure of the breast, and dynamically adjusts the window size according to tissue characteristics to perform adaptive window segmentation to obtain a window set carrying anatomical structure location information; inputs the window set into a pre-trained cross-window Transformer model for feature extraction, analyzes the feature correlation pattern between windows through the cross-window attention mechanism, and obtains a multi-dimensional ultrasound feature vector; inputs the feature vector into a classifier for molecular subtype classification prediction, and obtains a breast cancer molecular subtype identification prediction result. The present invention utilizes breast anatomical structure information through adaptive window segmentation, captures local and global feature associations through the cross-window attention mechanism, and improves prediction accuracy.

[0053] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0054] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0055] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and predicting breast cancer molecular subtypes based on ultrasound images, characterized in that: The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images includes: Perform feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image; Dividing the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjusting the window size according to tissue characteristics, and performing adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; Inputting the window set into a pre-trained cross-window Transformer model for feature extraction, analyzing the feature correlation pattern between windows through a cross-window attention mechanism, and obtaining a multi-dimensional ultrasound feature vector; The multi-dimensional ultrasound feature vector is input into the classifier of the cross-window Transformer model for molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

2. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 1, characterized in that: The step of performing feature fusion preprocessing based on acoustic impedance difference on the acquired breast ultrasound image to obtain a standardized ultrasound image includes: Based on the reflection characteristics of ultrasound in the skin layer, subcutaneous fat layer, glandular layer, and pectoral muscle layer, the acquired breast ultrasound image is processed for tissue interface boundary recognition to obtain tissue stratification information; Processing the tissue layer information using an acoustic shadow compensation algorithm to correct deep echo attenuation caused by changes in lesion tissue density to obtain an echo-corrected image; performing acoustic enhancement algorithm processing on the echo-corrected image to highlight tissue boundaries and internal structural features, thereby obtaining an acoustically enhanced image; performing texture directional feature extraction processing on the echo-corrected image to capture the anisotropic characteristics of breast cancer tissue and obtain a texture feature image; The echo-corrected image, the acoustically enhanced image, and the texture feature image are fused and standardized to a uniform size and contrast range to obtain a standardized ultrasound image.

3. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 1, characterized in that: The standardized ultrasound image is divided into sector-shaped regions based on the radial fiber structure of the breast, and the window size is dynamically adjusted according to tissue characteristics. The sector-shaped regions are adaptively segmented based on the window size to obtain a window set carrying anatomical structure position information, including: Performing fiber bundle tracking processing on the standardized ultrasound image according to the direction of Cooper's ligaments of the radial fiber structure of the breast, dividing the image into multiple sector-shaped areas along the fiber direction to obtain sector-shaped partitioned images; Performing tissue characteristic analysis on each sector in the sector-shaped partitioned image, identifying the distribution characteristics of glandular tissue, adipose tissue, and fibrous tissue through echo intensity, texture complexity, and acoustic attenuation coefficient, and obtaining a tissue characteristic distribution map; Dynamically adjust the window size according to the distribution density of different tissue types in the tissue characteristic distribution map to obtain an adaptive window size parameter; Window segmentation processing is performed on each sector area according to the adaptive window size parameter, and position information and corresponding tissue characteristic type in the radial fiber structure are marked for each window to obtain a window set carrying anatomical structure position information.

4. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 3, characterized in that: The fiber bundle tracking process is performed on the standardized ultrasound image according to the direction of Cooper's ligaments of the radial fiber structure of the breast, and the image is divided into a plurality of sector-shaped areas along the fiber direction to obtain the sector-shaped partitioned images. performing gradient direction calculation processing on the standardized ultrasound image, detecting the direction of the high-echo fiber bundles of the Cooper's ligament using a Sobel operator and a direction filter, and obtaining a fiber direction field map; Performing principal direction extraction processing on the fiber direction field map, identifying the main fiber bundle directions radially distributed from a preset position by using Hough transform and linear regression algorithm, and obtaining a radial principal direction set; Performing sector angle calculation processing based on the radial main direction set, calculating the angle between each main direction with the preset position as the center point and determining the angle range of the sector area to obtain sector angle parameters; The standardized ultrasound image is subjected to regional segmentation processing according to the sector angle parameter, the image is divided into a plurality of sector regions along the radial main direction and the angle range of each region is marked to obtain a sector partitioned image.

5. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 1, characterized in that: The window set is input into a pre-trained cross-window Transformer model for feature extraction, and the feature correlation pattern between windows is analyzed through a cross-window attention mechanism to obtain a multi-dimensional ultrasound feature vector including: Performing position encoding and feature embedding processing on each window in the window set, converting the anatomical structure position information and image features of the window into a high-dimensional vector representation, and obtaining a window feature embedding sequence; Performing cross-window attention weight calculation processing on the window feature embedding sequence, calculating the attention distribution between different receptive field windows through linear transformation of the query matrix, the key matrix and the value matrix, and obtaining a cross-window attention weight matrix; Performing weighted fusion processing on the window features according to the cross-window attention weight matrix, performing cross-correlation calculation on the local window features and the global window features to obtain fused window correlation features; The window-related features are subjected to multi-layer Transformer encoding processing, and deep semantic features are extracted through a multi-head self-attention mechanism and a feedforward neural network layer to obtain a multi-dimensional ultrasound feature vector.

6. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 1, characterized in that: Inputting the multi-dimensional ultrasound feature vector into the classifier of the cross-window Transformer model to perform molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result includes: performing molecular marker feature mapping processing on the multi-dimensional ultrasound feature vector, and performing correlation analysis between the multi-dimensional ultrasound feature vector and the molecular marker expression pattern to obtain a molecular marker feature representation; Performing a fully connected layer linear transformation on the molecular marker feature representation to map the high-dimensional feature vector to the feature space of four molecular subtypes to obtain subtype feature scores; Performing softmax normalization on the subtype feature scores, calculating the probability distribution of each molecular subtype and generating a confidence evaluation index to obtain a probability distribution vector; The maximum probability classification decision processing is performed according to the probability distribution vector, the molecular subtype with the highest probability value is selected as the prediction result and the corresponding confidence evaluation is output to obtain the breast cancer molecular subtype identification prediction result.

7. The method for identifying and predicting breast cancer molecular subtypes based on ultrasound images according to claim 6, characterized in that: The performing molecular marker feature mapping processing on the multi-dimensional ultrasound feature vector, and performing correlation analysis between the multi-dimensional ultrasound feature vector and the molecular marker expression pattern to obtain the molecular marker feature representation includes: Performing feature dimension transformation processing on the multi-dimensional ultrasound feature vector, mapping the high-dimensional feature vector to a feature subspace corresponding to the molecular marker through a linear projection layer, and obtaining a molecular marker subspace feature; Performing molecular marker expression pattern matching processing on the molecular marker subspace features, calculating the similarity score between the feature vector and the positive expression pattern of each molecular marker through a pre-trained marker expression pattern library, and obtaining a marker expression similarity matrix; Performing weighted fusion processing according to the marker expression similarity matrix, linearly combining the expression characteristics of the molecular markers according to the clinical importance weights, and obtaining comprehensive molecular marker characteristics; The comprehensive molecular marker features are subjected to nonlinear activation processing, and the feature expression capability is enhanced through ReLU activation function and batch normalization operation to obtain molecular marker feature representation.

8. A breast cancer molecular subtype identification and prediction system based on ultrasound images, characterized in that: The breast cancer molecular subtype identification and prediction system based on ultrasound images includes: A preprocessing module is used to perform feature fusion preprocessing on the acquired breast ultrasound image based on acoustic impedance difference to obtain a standardized ultrasound image; a segmentation module, configured to divide the standardized ultrasound image into sector-shaped regions based on the radial fiber structure of the breast, dynamically adjust the window size according to tissue characteristics, and perform adaptive window segmentation on the sector-shaped regions based on the window size to obtain a window set carrying anatomical structure position information; A feature extraction module is used to input the window set into a pre-trained cross-window Transformer model for feature extraction processing, analyze the feature correlation pattern between windows through the cross-window attention mechanism, and obtain a multi-dimensional ultrasound feature vector; The classification prediction module is used to input the multi-dimensional ultrasound feature vector into the classifier of the cross-window Transformer model to perform molecular subtype classification prediction processing to obtain a breast cancer molecular subtype identification prediction result.

Images