A breast B-ultrasonic image recognition system based on deep learning and a construction method thereof

Abstract

This invention relates to the field of medical image processing and analysis technology, and discloses a breast ultrasound image recognition system and its construction method based on deep learning. The system includes: an image acquisition and preprocessing module, a candidate detection and multi-view registration module, a feature extraction and fusion module, a false positive suppression and consistency verification module, a topological reasoning and interpretable diagnosis module, and an uncertainty quantification and decision fusion module. By extracting ultrasound physical features such as posterior echo enhancement and acoustic attenuation and adaptively gating them, combined with multi-view geometric registration and physiological-artifact motion separation, pure physical-data joint features are obtained. Anatomical topological reasoning is performed using a graph attention network, and the decision fusion weights are dynamically adjusted using uncertainty metrics obtained through Bayesian sampling, triggering cross-sectional consistency verification. A feedback loop synchronously updates network parameters and the true positive feature library. This invention improves the accuracy and robustness of identifying benign and malignant breast lesions and provides interpretability.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and analysis technology, specifically relating to a breast ultrasound image recognition system and its construction method based on deep learning. Background Technology

[0002] Breast ultrasound (B-scan) has become an important tool for screening and diagnosing breast diseases due to its advantages such as being non-invasive, real-time, and low-cost. However, breast ultrasound images have inherent physical artifacts such as speckle noise, acoustic shadowing attenuation, and posterior echo enhancement. Furthermore, differences in acoustic impedance between different tissues lead to unstable image contrast, making the determination of benign or malignant lesions highly dependent on the doctor's clinical experience.

[0003] In recent years, deep learning-based medical image analysis technology has made significant progress. Existing technologies include methods using convolutional neural networks (CNNs) for end-to-end benign / malignant classification of breast ultrasound images, and schemes utilizing multi-view information or temporal dynamic features for assisted diagnosis. However, existing methods still have the following shortcomings: 1) High heterogeneity of lesions: Breast tumors vary significantly in morphology, size, margins, and acoustic characteristics, making it difficult to capture complete features with a single resolution or single viewpoint; 2) Artifacts and motion interference: During dynamic scanning with the probe, global rigid body motion caused by breathing, heartbeat, and probe jitter can mix in the true elastic deformation features of the lesion. Existing temporal modeling methods are unable to effectively separate these artifacts, reducing the purity of dynamic features; 3) Insufficient feature fusion: Traditional methods often rely on single-modality or single-network features, making it difficult to simultaneously utilize spatial, temporal, and physical features to improve classification accuracy; 4) High false positive rate: In the initial screening stage of candidate regions, single-resolution or single-path detection easily generates a large number of false positive regions. Existing post-processing methods (such as simple threshold segmentation or conditional random fields) lack the ability to jointly filter out multimodal features, easily leading to misdiagnosis; 5) Lack of interpretability and uncertainty quantification: Existing methods cannot clearly label the lesion contribution area, nor can they provide classification confidence and decision references. Clinicians find it difficult to trust their judgments, limiting practical application.

[0004] Therefore, there is an urgent need for a breast ultrasound image recognition system that can integrate multi-section, multi-resolution, and multi-modal features, combined with physical prior constraints, motion artifact separation, and uncertainty quantification, to achieve high-precision, low-false-positive, and interpretable intelligent breast ultrasound image recognition and analysis. Summary of the Invention

[0005] This invention provides a breast ultrasound image recognition system and construction method based on deep learning, aiming to solve the technical problems mentioned in the background art.

[0006] In a first aspect, the present invention provides a breast ultrasound image recognition system based on deep learning, comprising: The image acquisition and preprocessing module is used to acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and to perform noise reduction, standardization and enhancement processing on the acquired images; The candidate detection and multi-view registration module is used to perform multi-level pixel scale characterization analysis based on the preprocessed image to obtain candidate lesion regions. The feature extraction and fusion module is used to extract multi-faceted, multi-level pixel-scale spatial semantic features and elastic deformation dynamic features, regulate ultrasonic physical features, and generate physical-data joint features. The false positive suppression and consistency verification module is used to eliminate false positive candidate lesion regions and perform spatiotemporal consistency clustering. The topology reasoning and interpretable diagnosis module receives candidate lesion regions after false positive suppression and consistency verification, combines the physical-data joint features to construct a heterogeneous map of the lesion and surrounding tissues, performs topology relationship reasoning through a graph attention network, and outputs a malignancy probability and category activation heatmap. The uncertainty quantification and decision fusion module uses a Bayesian deep network to quantify and classify uncertainty, dynamically adjusts the weights of each feature and the strength of physical constraints, and generates the final benign / malignant identification results and confidence reports.

[0007] Furthermore, the image acquisition and preprocessing module includes: The noise reduction unit is used to filter out speckle noise and probe interference noise in the acquired images; Size normalization units are used to standardize the resolution of images across different sections. The data augmentation unit augments the training and inference datasets using image processing transformations and light processing transformations. The image processing transformations include scaling, flipping, rotating, and translating, while the light processing transformations include Gaussian blurring and sharpening.

[0008] Furthermore, the candidate detection and multi-view registration module includes: The multi-resolution candidate detection unit identifies abnormal regions by constructing an image pyramid and performing sliding window scanning and superpixel segmentation at each resolution level. The multi-view registration unit performs pixel-level geometric registration of candidate regions in longitudinal, transverse, and radial sections using probe spatial pose tracking data, generating multi-section aligned candidate lesion regions.

[0009] Furthermore, the feature extraction and fusion module includes: The spatial feature extraction branch is used to extract multi-faceted and multi-scale spatial semantic features based on a deep convolutional network with a squeeze-excitation attention module; The temporal feature extraction branch is used to model continuous frame motion images based on three-dimensional residual convolution and spatiotemporal separation attention mechanism, and extract dynamic features of lesion elastic deformation. The physical feature control unit is used to extract ultrasonic physical features and adaptively adjust their intensity based on spatial semantic features; The cross-scale feature fusion unit integrates spatial semantic features, temporal dynamic features, and gated physical features to generate physical-data joint features.

[0010] Furthermore, the false positive suppression and consistency verification module includes: The false positive elimination unit eliminates candidate regions that exceed the threshold by constructing a multimodal feature vector of the candidate region and comparing it with the true positive feature library using a weighted Mahalanobis distance. The spatiotemporal consistency verification unit uses mean-shift clustering to group candidate regions that are spatially and temporally close together, and deletes groups with fewer than a threshold number of elements within each group, outputting the final candidate lesion region that is spatiotemporally consistent.

[0011] Furthermore, the aforementioned topological reasoning and interpretable diagnostic module includes: Heterogeneous graph construction unit: Heterogeneous graph is constructed using candidate lesion regions and their surrounding tissues as nodes; The graph attention network unit is used to learn the topological relationships between nodes and higher-order adjacency weights, and output relationship reasoning features. The interpretable diagnostic unit obtains feature map category scores through global average pooling, generates the probability of malignancy, calculates category activation heatmaps, and marks the lesion contribution area.

[0012] Furthermore, the uncertainty quantification and decision fusion module includes: The Bayesian deep network unit uses a multi-head Bayesian neural network to perform Monte Carlo sampling on the physical-data joint features and relational reasoning features, and outputs the benign / malignant classification probability and uncertainty measure. The decision weight adaptive unit dynamically adjusts the weights of each feature and physical constraint in decision fusion based on the uncertainty metric, and strengthens the physical constraint strength and triggers cross-section consistency verification when the uncertainty exceeds the threshold. The final decision-making unit generates the benign or malignant lesion identification results and confidence reports based on the adjusted fusion weights.

[0013] Furthermore, it also includes a feedback closed-loop optimization module, which is used to update the parameters of the entire deep network, including the spatial feature network, temporal feature network, gated physical feature network, and graph attention network, based on the decision fusion result and uncertainty feedback, in order to improve recognition accuracy and robustness.

[0014] Furthermore, the system is further configured to update the typical true positive feature library online, record the physical and semantic feature mismatch patterns of the candidate regions to be eliminated, and optimize the false positive elimination performance based on the accumulated data.

[0015] Secondly, the present invention provides a method for constructing breast ultrasound images based on deep learning, comprising: Step S1: Acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and perform denoising, standardization and data augmentation on the images; Step S2: Construct a multi-resolution pyramid based on the preprocessed image, perform parallel analysis of sliding window and superpixel to obtain abnormal regions, and perform pixel-level registration of multi-faceted candidate regions using probe spatial pose data to generate multi-faceted aligned candidate lesion regions. Step S3: Extract multi-faceted, multi-scale spatial semantic features, elastic deformation dynamic features of continuous frame motion images, and ultrasonic physical features. Adaptively adjust the intensity based on spatial features, and fuse spatial, dynamic, and physical features across scales to generate physical-data joint features. Step S4: Construct multimodal feature vectors of candidate regions and compare them with the true positive feature library to eliminate false positive regions. Perform spatiotemporal clustering on the candidate regions, delete low-confidence clusters, and output the final candidate lesion regions that are consistent in time and space. Step S5: Construct a heterogeneous graph using candidate lesions and surrounding tissues as nodes, use a graph attention network to learn topological relationships, output relationship inference features, and generate a category activation heatmap and the probability of malignancy. Step S6: Perform Bayesian sampling on the physical-data joint features and relational reasoning features, quantify the classification uncertainty, dynamically adjust the feature fusion weights and physical constraint strength according to the uncertainty, and output the final benign / malignant identification results and confidence report; Step S7: Update the deep network parameters based on the decision results and uncertainty feedback to improve recognition accuracy and robustness, record the mismatch patterns of candidate regions online, update the true positive feature library, and optimize the false positive removal performance.

[0016] This invention offers the following advantages: It achieves accurate detection and benign / malignant identification of breast tumors through multi-section, multi-resolution, and multi-modal feature fusion; it effectively reduces false-positive candidate regions and improves diagnostic reliability by combining physical constraints with a multi-modal false-positive elimination mechanism; it eliminates the influence of respiration and probe jitter by using continuous frame optical flow calculation and physiological-artifact separation, ensuring recognition stability in dynamic scanning scenarios; it enhances diagnostic credibility by annotating lesion areas with category activation maps and contribution region heatmaps, facilitating clinicians' understanding of the model's decision-making basis; it utilizes Bayesian deep networks to quantify classification uncertainty, enabling dynamic weight adjustment and cross-section consistency verification, thus improving decision reliability; and it adapts to different patient data and continuously improves recognition accuracy and robustness through feedback loop optimization of network parameters and online updates to the true positive feature library. Attached Figure Description

[0017] Figure 1 A block diagram of a breast ultrasound image recognition system based on deep learning provided by the present invention; Figure 2 The flowchart illustrates a method for constructing breast ultrasound image recognition based on deep learning, as provided in this invention. Detailed Implementation

[0018] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0019] Firstly, such as Figure 1 This embodiment provides a breast ultrasound image recognition system based on deep learning, including: The image acquisition and preprocessing module is used to acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and to perform noise reduction, standardization and enhancement processing on the acquired images.

[0020] The image acquisition and preprocessing module includes: a denoising unit, used to filter out speckle noise and probe interference noise in the acquired images; used to perform edge trimming, anonymization, and size standardization of each cross-sectional image to a standard resolution; and a data augmentation unit, which augments the training and inference datasets through image processing transformations and optical processing transformations, wherein the image processing transformations include at least one of scaling, flipping, rotating, and translating, and the optical processing transformations include at least one of Gaussian blur, sharpening, and deletion.

[0021] Specifically, the image acquisition and preprocessing module first acquires a series of consecutive multi-frame images of the same breast lesion in longitudinal, transverse, and radial sections using an ultrasound probe. During acquisition, due to the inherent acoustic characteristics of ultrasound imaging and the coupling state between the probe and the skin, speckle noise and high-frequency probe interference noise are inevitably mixed into the original images. The denoising unit uses a combination of adaptive median filtering and anisotropic diffusion filtering: the former is used to remove isolated high-intensity noise points, while the latter smooths speckle noise while preserving the edge structure of the lesion, thereby outputting multi-section images with improved signal-to-noise ratio.

[0022] The denoised images are then fed into a size normalization unit. This unit performs edge trimming on each cross-sectional image to remove invalid scan borders and annotation information around the image; subsequently, anonymization is performed to remove metadata containing patient identification information; finally, all cross-sectional images are uniformly scaled to a preset standard resolution, such as 512×512 pixels, using bilinear interpolation. The normalization process ensures that images from different cross-sections and different cases are spatially comparable, providing a unified input size for subsequent multi-view registration and feature extraction.

[0023] The data augmentation unit performs online augmentation based on standardized images. For image processing transformations, the unit performs horizontal and vertical flipping on the image with random probabilities to simulate the appearance of lesions under different probe scanning directions; rotation operations are performed randomly within the range of -15° to +15° to accommodate the rotational changes of lesions when the probe is tilted for acquisition; scaling operations are randomly selected between 0.8x and 1.2x to simulate scenarios where lesion size changes with depth or axial distance; translation operations shift the position of the lesion within the image window, enhancing the model's robustness to lesion localization deviations. All of the above image processing transformations are based on the image center and simultaneously adjust the coordinates of the corresponding lesion annotation boxes to maintain the correspondence between the labels and the transformed image.

[0024] For optical processing transformations, the data augmentation unit adds speckle noise to simulate the differences in graininess in ultrasound images under different equipment or scanning parameters. Sharpening uses the Laplacian operator to enhance the high-frequency components of lesion edges, making the model more focused on edge morphology features. Gaussian blurring uses convolution kernels with different standard deviations, such as σ=0.5~1.5, to smooth the image, simulating image blurring caused by inaccurate focus or uneven probe pressure. Deletion operations randomly erase rectangular regions in the image with a very low probability, such as 1%, covering 5%~20% of the original image area, and fill them with their neighborhood mean, simulating local information loss caused by tissue occlusion or acoustic attenuation. All images after optical processing transformations retain the original lesion category labels, and the transformation parameters are randomly generated during each training round, effectively expanding the dynamic range of image grayscale and texture.

[0025] Through the combined effects of the aforementioned geometric and optical transformations, this module generates an enhanced image set with diversity in scale, rotation, translation, grayscale, and texture. This enhanced set, together with the original normalized image, constitutes the input for subsequent candidate detection and multi-view registration modules. Taking a specific case as an example, when the lesion in the original acquired image is located at the image edge and is small in size, the data augmentation unit can move the lesion to the center of the image through translation operations, and then adjust it to an appropriate size through scaling operations, thereby avoiding the omission of edge areas by the fixed sliding window. At the same time, for the acoustic shadowing attenuation phenomenon that often occurs in deep lesions, the deletion operation simulates the loss of acoustic shadowing information, forcing the subsequent physical feature control unit to learn to rely on other physical features such as the posterior echo enhancement area and edge morphology for discrimination under incomplete information, thereby improving the robustness of the system when encountering different degrees of acoustic shadowing interference in real clinical settings.

[0026] After the aforementioned preprocessing and data augmentation, the standardized enhanced image set output by the image acquisition and preprocessing module is sent to the candidate detection and multi-view registration module. This module performs multi-level pixel-scale characterization analysis based on the preprocessed images, such as multi-resolution image pyramid detection and multi-view registration, to obtain candidate lesion regions. This module takes the enhanced longitudinal, transverse, and radial section images as input, first constructing a multi-resolution image pyramid, and then performing sliding window scanning and superpixel segmentation in parallel at each level. Because the enhancement processing has standardized the image size and simulated geometric changes and acoustic interference during probe scanning, the sliding window of the candidate detection module can stably cover lesion regions of different scales, and superpixel segmentation will not over-segment due to noise or uneven grayscale. The top layer of the pyramid is used for coarse localization, and the bottom layer for fine localization. The coordinates of abnormal regions detected at each level are merged after inverse mapping to form a multi-resolution fused candidate lesion region list. This list, along with the spatial correspondence between the sections, is passed to the multi-view registration unit to provide initial correspondence points for subsequent pixel-level geometric registration.

[0027] The candidate detection and multi-view registration module specifically includes: a multi-resolution candidate detection unit, which identifies abnormal regions by constructing an image pyramid and performing sliding window scanning and superpixel segmentation at each resolution level; and a multi-view registration unit, which performs pixel-level geometric registration of candidate regions in longitudinal, transverse and radial sections using probe spatial pose tracking data to generate multi-section aligned candidate lesion regions.

[0028] More specifically, the candidate detection and multi-view registration module receives a standardized enhanced image set output from the image acquisition and preprocessing module. This module first uses a multi-resolution candidate detection unit to process each frame of the longitudinal, transverse, and radial section images individually. The unit constructs an image pyramid: starting with the original resolution, e.g., 512×512 pixels, as the bottom layer, it successively downsamples to generate upper-layer images with half the resolution, setting three to five layers in total. Two paths are executed in parallel at each resolution level. The sliding window path uses a fixed-size search window, such as 64×64 pixels, to perform raster scanning at a set step size, extracting the grayscale statistical features and local binary pattern features of the image within the window, and inputting this into a pre-trained lightweight deep classifier to determine whether the window contains abnormal regions. The superpixel path first uses a simple linear iterative clustering algorithm to segment the current level image into multiple superpixel regions, each superpixel representing an image block with similar pixel attributes; then, it expands to a fixed-size rectangular region centered on the centroid of each superpixel, extracting its depth features and identifying anomalies. The detection results from both paths at different resolutions, such as anomalous superpixel bounding boxes, are mapped back to the original resolution coordinates using a coordinate scaling factor. All mapped bounding boxes are merged, and non-maximum suppression is applied to regions with overlap exceeding a set threshold, forming a multi-resolution fusion candidate lesion region list.

[0029] The aforementioned multi-resolution detection mechanism addresses the issues of small lesions being easily overlooked and large lesions having coarse boundary localization under single-resolution conditions. Taking a common clinical example of a small lobular hyperplasia nodule with a diameter of approximately 5 mm, its pixel area is small in the low-resolution image at the bottom layer, making it easy for the sliding window to include it within the normal background. However, in the low-resolution image at the top of the pyramid, the lesion's size is relatively larger, making it easier for superpixel segmentation to cluster it into an independent region and classify it as abnormal. Conversely, for invasive ductal carcinoma with a diameter exceeding 20 mm, the low-resolution image at the bottom layer provides the fine structure of the lesion's edges, allowing it to be correctly captured by the sliding window. Multi-resolution fusion ensures a high detection rate for lesions of different sizes.

[0030] After obtaining the candidate region list, the multi-view registration unit uses the spatial pose tracking sensor built into the ultrasound probe, such as an electromagnetic or optical tracker, to record the probe's spatial coordinates (e.g., six-DOF pose) when acquiring each section image. Since the longitudinal, transverse, and radial sections are acquired sequentially during the same scan, the probe pose data corresponding to each section image is continuously recorded. The registration unit first uses the spatial coordinate system of the transverse section image as a reference and calculates the spatial transformation matrix from the pixel coordinates of each section image to the reference coordinate system based on the probe pose parameters of the longitudinal and radial sections. For each candidate lesion region, its midpoint pixel coordinates are extracted and mapped to the reference coordinate system using the transformation matrix to obtain a three-dimensional spatial point. Utilizing the geometric correspondence between lesion regions across sections, an iterative nearest-point algorithm is used to optimize the transformation parameters, achieving pixel-level registration. After registration, the corresponding positions of the same lesion in different sections are aligned, forming multi-section aligned candidate lesion regions. This alignment result is stored in the form of a spatial mapping table, containing the pixel coordinate set of each lesion region in each section image and its mutual transformation parameters.

[0031] In clinical practice, when the same lesion appears as a long strip on the longitudinal section and a near-circular shape on the transverse section, the registration unit can determine from spatial pose data that both actually describe different axial projections of the same lesion. After pixel-level geometric registration, the lesion regions on both sections occupy the same physical space in the reference coordinate system, thus providing spatially consistent multi-view input for subsequent feature extraction and fusion modules. This registration process and the multi-resolution detection function support each other: multi-resolution detection provides reliable initial positions for candidate regions, while the registration unit uses spatial pose data to accurately locate the correspondence between sections, jointly solving the technical problem of lack of geometric alignment and inability to fuse three-dimensional information in lesion regions in multi-view ultrasound images.

[0032] After multi-resolution candidate detection and multi-view registration, the candidate detection and multi-view registration module outputs multi-section aligned candidate lesion regions, while preserving the original multi-section image sequence, the pixel-level spatial geometric registration relationship between each section, and the motion image sequence of consecutive frames during probe scanning. This data is then passed to the feature extraction and fusion module, which extracts multi-section, multi-level pixel-scale spatial semantic features and elastic deformation dynamic features, modulates ultrasound physical features, and generates physical-data joint features. This module uses the multi-section aligned candidate lesion regions as the regions of interest for spatial feature extraction, uses geometric registration relationships as constraints for motion compensation and physiological-artifact separation in temporal feature extraction, and uses the motion image sequence as the input source for temporal modeling, thereby achieving the synergistic utilization of multi-view spatial information and dynamic temporal information.

[0033] Specifically, the feature extraction and fusion module includes: a spatial feature extraction branch, used to extract multi-faceted, multi-scale spatial semantic features based on a deep convolutional network with a squeeze-excitation attention module; and a temporal feature extraction branch, used to model continuous frame motion images based on three-dimensional residual convolution and a spatiotemporal separation attention mechanism, and extract dynamic features of lesion elastic deformation. The temporal feature extraction branch also includes a physiological-artifact motion separation unit, configured to: decompose the optical flow field between continuous frames using the spatial geometric registration relationship established by the multi-view registration unit; calculate the global rigid body motion transformation matrix to characterize artifact motion caused by probe jitter or breathing; subtract the global rigid body motion transformation matrix from the total motion field to obtain the local deformation motion field; and use the local deformation motion field as the temporal feature extraction branch. The modeling process involves several key inputs, including a physical feature control unit. The physical feature control unit extracts ultrasonic physical features and adaptively adjusts their intensity based on spatial semantic features. Specifically, it uses shallow features from the multi-scale spatial semantic features output by the spatial feature extraction branch as input, calculates a gate weight vector with the same dimension as the physical features in real-time using a lightweight gating network, multiplies the gate weight vector element-wise with the original physical constraint vector, outputs the gated physical features, and applies L1 sparse regularization to the gate weights during training. A cross-scale feature fusion unit fuses spatial semantic features, temporal dynamic features, and gated physical features to generate joint physical-data features. This cross-scale feature fusion unit uses a bidirectional feature pyramid structure to fuse the spatial semantic features and temporal dynamic features, and uses the gated physical features to recalibrate the channel attention of the fused feature map, generating joint physical-data features.

[0034] More specifically, the feature extraction and fusion module receives multi-faceted aligned candidate lesion regions, multi-faceted image sequences, spatial geometric registration relationships, and continuous frame motion image sequences from the candidate detection and multi-view registration module. This module collaboratively extracts heterogeneous features and completes the fusion through three branches.

[0035] The spatial feature extraction branch takes multi-section image sequences and aligned candidate lesion regions as input. The branch uses an improved EfficientNet-B4 as its backbone network, embedding a squeeze-excitation attention block after each MBConv module. This network is fine-tuned on breast ultrasound images based on ImageNet pre-training. When a longitudinal section ultrasound image, such as 512×512 pixels, is input, the network downsamples stepwise, outputting a 16×16 resolution feature map in the fifth stage and an 8×8 feature map in the sixth stage, representing superficial details of the lesion, such as edge spiculation and internal echo distribution, and deep semantics, such as overall morphological category, respectively. Multi-scale spatial semantic features consist of a set of feature maps at these two or more scales, each corresponding to a different receptive field and semantic level. For transverse and radial sections, the branches process in parallel, each outputting feature maps at its corresponding scale.

[0036] The temporal feature extraction branch receives a sequence of consecutive motion images, such as 30 consecutive frames, along with spatial geometric registration relationships. First, the physiological-artifact motion separation unit calculates a dense optical flow field for each adjacent frame pair to obtain the total motion vector field. Using probe spatial pose data provided by the multi-view registration unit, the unit estimates the global rigid body motion transformation matrix through least-squares fitting. This matrix describes the overall image displacement and rotation caused by probe jitter or patient breathing. The global rigid body motion component is subtracted pixel-by-pixel from the total motion field to obtain a local deformation motion field containing only the elastic deformation of the lesion itself. If the amplitude of the global rigid body motion exceeds a preset threshold, for example, a root mean square displacement greater than 5 pixels or a rotation angle greater than 3 degrees, the current frame is determined to be severely contaminated by motion artifacts, and the frame is automatically discarded or its weight is reduced to near zero in temporal modeling. The local deformation motion fields of the remaining frames are fed as effective dynamic features into the subsequent 3D residual convolution module. This module employs a three-dimensional convolutional kernel (3x3 in time and 3x3 in space) to extract the spatiotemporal patterns of inter-frame changes and overlays a spatiotemporal separation attention mechanism: first, self-attention is applied in the spatial dimension to focus on the deformed lesion region, and then self-attention is applied in the temporal dimension to correlate the deformation evolution process of different frames. The output is an elastic deformation dynamic feature tensor with dimensions of time length × spatial size × number of channels.

[0037] The physical feature control unit extracts ultrasound physical features from the preprocessed image: the posterior echo enhancement region is quantified by detecting areas with higher gray levels behind the lesion than in front; the acoustic shadow attenuation region is quantified by the area and gray level descent slope of the low-gray-level region below the lesion; lesion edge morphology parameters are quantified by calculating edge blurring and spur index; and acoustic impedance difference features are quantified by the gradient intensity at the interface between the lesion and surrounding tissue. These features are encoded into a fixed-length physical constraint vector, such as 16-dimensional. The unit receives shallow multi-scale features from the spatial feature extraction branch, such as a 56×56 feature map from the second stage of EfficientNet-B4, and inputs it into a lightweight gating network consisting of two fully connected layers after global average pooling. The gating network outputs a gating weight vector with the same dimension as the physical constraint vector, with each element ranging from 0 to 1. The gating weight vector is multiplied element-wise by the original physical constraint vector to obtain the gated physical features. During training, L1 regularization is applied to the gating weights, causing the network to output near-zero weights for irrelevant physical features, thereby achieving adaptive filtering. Taking cystic lesions in clinical practice as an example, the echo enhancement area behind them is the main basis for differentiation, while the acoustic shadow attenuation area is almost non-existent. The gating network will give high weight to the echo enhancement feature and suppress the weight of the acoustic shadow attenuation feature to near zero, thereby avoiding noise interference.

[0038] The physical-data joint features, multi-scale spatial semantic feature subsets, and physical constraint vectors output by the feature extraction and fusion module, along with the candidate lesion region list generated by the multi-resolution candidate detection unit, are passed to the false positive suppression and consistency verification module. The false positive suppression and consistency verification module is used to eliminate false positive candidate lesion regions and perform spatiotemporal consistency clustering. It uses the candidate region list as its processing object, extracting the corresponding spatial semantic feature slices and physical constraint vectors for each candidate region from the feature extraction and fusion module to construct a multimodal feature vector, which serves as the input to the false positive elimination unit. Meanwhile, the physical-data joint features are retained for subsequent topological inference and do not participate in false positive suppression.

[0039] Specifically, the false positive suppression and consistency verification module includes: a false positive elimination unit, which constructs a multimodal feature vector of candidate regions and compares it with a weighted Mahalanobis distance from the true positive feature library to eliminate candidate regions exceeding the threshold; the weighting coefficient of the weighted Mahalanobis distance in the false positive elimination unit is dynamically adjusted according to the uncertainty metric value currently output by the uncertainty quantification and decision fusion module; and a spatiotemporal consistency verification unit, which groups candidate regions that are spatially and temporally close together through mean-shift clustering and deletes groups with fewer than a threshold number of elements within the group, outputting the final candidate lesion region that is spatiotemporally consistent.

[0040] More specifically, the false positive suppression and consistency verification module receives a multi-scale spatial semantic feature subset, a physical constraint vector subset, and a list of candidate lesion regions output by the multi-resolution candidate detection unit from the feature extraction and fusion module. This module consists of two cascaded parts: a false positive elimination unit and a spatiotemporal consistency verification unit.

[0041] The false positive elimination unit constructs a multimodal feature vector for each candidate lesion region. Specifically, the unit obtains multi-scale spatial semantic feature slices corresponding to the region from the feature extraction and fusion module, for example, extracting scales from the third and fifth stages of the improved EfficientNet-B4. Figure 1 The feature map regions of / 8 and 1 / 16 are processed by global average pooling to form a 64-dimensional vector, along with sub-features corresponding to these regions in the physical constraint vector, such as the quantization value of posterior echo enhancement, the area ratio of the acoustic shadow attenuation zone, the edge spur index, and the acoustic impedance gradient magnitude, totaling 16 dimensions. These two feature parts are then concatenated to form an 80-dimensional multimodal feature vector. The unit pre-stores a feature library of typical true positive lesions, which is composed of clusters of multimodal feature vectors from pathologically confirmed benign and malignant lesion samples. Each category stores the mean vector and covariance matrix. For the input multimodal feature vector, the unit calculates its Mahalanobis distance with each category in the feature library. The Mahalanobis distance calculation formula can be described as follows: first, calculate the difference between the feature vector and the category mean, then perform a quadratic operation on this difference and the inverse of the category covariance matrix. Since the Mahalanobis distance considers the correlation between features, it can effectively distinguish lesions with similar morphology but different physical properties.

[0042] The weighting coefficient of the Mahalanobis distance is dynamically adjusted based on the uncertainty metric currently output by the uncertainty quantification and decision fusion module. When the uncertainty metric is high, for example, greater than 0.7, it indicates that the classification confidence of the current sample to be identified is low. In this case, the false positive rejection should be more stringent, and the weighting coefficient should be larger, such as 2.0, so that the distance threshold is effectively reduced and more candidate regions are rejected. When the uncertainty metric is low, such as less than 0.3, the weighting coefficient should be smaller, such as 0.5, to reduce the risk of false rejection. The weighted Mahalanobis distance is compared with the preset threshold. If it exceeds the threshold, the candidate region is judged as a false positive and rejected. All rejected regions record the mismatch patterns between their multimodal feature vectors and the most matching category in the feature library, that is, the direction and magnitude of the deviation of each dimension of the feature, for subsequent online updates of the feature library.

[0043] The spatiotemporal consistency verification unit receives the remaining candidate regions after false positives are removed. Since multiple consecutive frames of images recorded the location of the same lesion at different time points during ultrasound scanning, the unit obtains the center pixel coordinates and corresponding frame numbers of all candidate regions in the current frame and previous consecutive frames (e.g., the first 5 frames). For each candidate region, its spatial coordinates and frame number are combined to form a three-dimensional data point (x, y, t). The mean-shift clustering algorithm iterates in three-dimensional space with a set bandwidth (20 pixels spatially and 2 frames temporally), grouping points that are spatially close and temporally consecutive. After clustering, the number of points in each group is counted. If the number of elements in a group is less than a preset threshold, such as 3 points, the group is determined to be isolated noise or transient artifacts and is deleted. Representative candidate regions (such as the region with the highest frequency within the group) are selected from the remaining groups as the final candidate lesion regions for spatiotemporal consistency verification.

[0044] Taking a common vascular cross-section in clinical scans as an example, its morphology resembles small solid nodules, which may lead to misjudgment as an abnormal area by the multi-resolution detection unit. The vascular cross-section rapidly moves out of the field of view due to probe movement across multiple frames, appearing only in 1-2 frames in the time dimension; while the actual lesion area usually exists stably across multiple frames, such as 5 or more. Therefore, the spatiotemporal consistency verification unit can classify and eliminate vascular cross-sections into groups with fewer than a threshold elements. Simultaneously, for complex lesions with significant acoustic attenuation and blurred edges, the false positive elimination unit uses a stricter criterion for the weighted Mahalanobis distance when the uncertainty metric is high, further screening for atypical lesions.

[0045] The two units mentioned above support each other: false positive elimination utilizes physical-optical multimodal features and dynamically weighted Mahalanobis distance to achieve single-frame image-based screening of individual regions; spatiotemporal consistency verification utilizes the spatiotemporal continuity between consecutive frames to achieve multi-frame-based group verification. The cascading of these two units solves the problem of excessively high false positive rates in candidate detection modules caused by inherent noise in ultrasound images, interference from vascular cross-sections, and fat particles. Taking 100 clinical data tests as an example, after using this module, the false positive rate decreased from an average of 5.2 per image before suppression to 0.8, while the true positive rate remained above 92%. This technical effect stems from the deep fusion of algorithmic features (Mahalanobis distance, mean-shift clustering) with the spatiotemporal characteristics and physical properties of breast ultrasound imaging: the uncertainty metric comes from a Bayesian network and has feedback linkage with gated physical features; spatiotemporal clustering utilizes the clinical prior of stable lesions in ultrasound scanning.

[0046] The final candidate lesion region with spatiotemporal consistency output by the false positive suppression and consistency verification module, along with the physical-data joint features generated by the feature extraction and fusion module, are passed to the topology reasoning and interpretable diagnosis module. The topology reasoning and interpretable diagnosis module receives the candidate lesion region after false positive suppression and consistency verification, and, in conjunction with the physical-data joint features, constructs a heterogeneous graph of the lesion and surrounding tissues. It then performs topological relationship reasoning through a graph attention network, outputting a malignancy probability and category activation heatmap. This module uses the final candidate lesion region as the graph center node, extracts the corresponding feature vector from the physical-data joint features as the initial node embedding, and utilizes multi-view registration relationships to obtain the anatomical location information of surrounding ducts, fat layers, and glandular tissues to construct neighboring nodes, thereby forming a heterogeneous graph topology.

[0047] The topology reasoning and interpretable diagnosis module includes: a heterogeneous graph construction unit, which constructs a heterogeneous graph using candidate lesion regions and surrounding tissues as nodes; a graph attention network unit, which learns the topological relationships and higher-order adjacency weights between nodes and outputs relationship reasoning features; and an interpretability diagnosis unit, which obtains feature map category scores through global average pooling, generates the probability of malignancy, calculates category activation heatmaps, and identifies lesion contribution areas. The interpretability diagnosis unit is further configured to: scale the generated category activation map using min-max normalization and binarize it using thresholding to determine the contribution areas.

[0048] More specifically, the topology reasoning and interpretable diagnosis module receives the spatiotemporally consistent final candidate lesion region output from the false positive suppression and consistency verification module, as well as the physical-data joint features generated by the feature extraction and fusion module.

[0049] The heterogeneous graph construction unit uses each final candidate lesion region as a central node and the anatomical structures within a predetermined range surrounding the lesion as neighboring nodes. Specifically, based on the spatial geometric registration relationship provided by the multi-view registration unit, ductal structures, fat layers, and glandular tissues are detected within a 5 mm radius centered on the lesion in the reference coordinate system. Ductal structures are identified through tubular hypoechoic regions in ultrasound images, fat layers are located through hyperechoic bands and their posterior echo attenuation features, and glandular tissues are defined by non-uniform medium-to-hypoechoic regions. Each identified anatomical structure is treated as an independent node. The initial feature vectors of all nodes, including the central node and neighboring nodes, are obtained from the feature map regions corresponding to the spatial locations in the physical-data joint features through bilinear interpolation sampling, with a dimension of 256. The heterogeneous graph includes two node types (lesion nodes and tissue nodes) and three edge types (lesion-duct, lesion-fat, lesion-gland), with initial edge weights set according to the reciprocal of the Euclidean distance between nodes.

[0050] The graph attention network unit receives the heterogeneous graph and uses a multi-head graph attention mechanism (setting up 4 attention heads) to learn the topological relationships between nodes. For each node, the unit calculates the weighted sum of the features of its neighboring nodes, with the weights determined by the attention coefficients between the node features. The specific process can be described as follows: for the center node and a certain neighboring node, the initial feature vectors of both are mapped to a new space through a shared linear transformation. Then, the inner product of the mapped vectors is calculated and nonlinear activation is performed to obtain the original attention coefficients. These coefficients are normalized on all neighboring nodes by dividing them by the sum of the exponents of all neighboring node coefficients, ensuring that the sum of the neighborhood weights for the same center node is 1. To capture higher-order adjacency relationships, the unit stacks two graph attention layers: the first layer outputs updated node features (128 dimensions), and the second layer takes the output of the first layer as input and further expands the receptive field of each node to second-order neighboring nodes. Finally, the center node features output by the second layer are used as the relation inference features, with a dimension of 128. This feature not only includes the physical-data attributes of the lesion itself but also incorporates the topological influence of surrounding ducts, fat, and glands on the lesion. Taking invasive ductal carcinoma in clinical practice as an example, this type of tumor is often accompanied by ductal structural distortion and traction of surrounding glandular tissue. The attention weight between lesion nodes and ductal nodes and glandular nodes in the heteromorphism diagram will be significantly increased, thus reflecting this pathological association in relational reasoning features.

[0051] The interpretability diagnostic unit performs operations on the output feature map of the last convolutional layer of the feature extraction and fusion module, namely the cross-scale feature fusion unit. The unit performs global average pooling on the feature map output by this convolutional layer, assuming a size of 8×8×256, to obtain a 256-dimensional feature vector. This vector is then mapped to a 2-dimensional category score (benign / malignant) through a fully connected layer. The softmax function is applied to the category score to obtain the probability of malignancy. The category activation map is calculated by multiplying the weight vector (256-dimensional) corresponding to the malignant category in the fully connected layer weight matrix with each channel of the feature map channel-by-channel, and summing the results across the channel dimensions to obtain an 8×8 heatmap. This heatmap is then upsampled, such as by bilinear interpolation, to restore it to the original input image size (512×512). The interpretability diagnostic unit further performs min-max normalization on the category activation map, linearly scaling all pixel values ​​to the [0,1] interval. Subsequently, a preset threshold (e.g., 0.5) is used for binarization. Pixel regions larger than the threshold are marked as contribution regions, which are the image regions that the model focuses on when determining benignity or malignancy. This contribution region is output as a binary mask and overlaid with the original ultrasound image to generate a semi-transparent thermal image for physicians to visualize and review.

[0052] Taking the relational reasoning features output by the graph attention network as an example, when the lesion is a cystic benign lesion, the topological association with the surrounding ducts and fat layer is usually weak, the attention weights are evenly distributed, and the corresponding relational reasoning features tend to be benign. However, invasive carcinoma has a strong abnormal association with surrounding tissues (ductal invasion, glandular structure destruction), the attention weights are concentrated on a few abnormal nodes, and the contribution area heatmap shows spiculation at the lesion edge and the area connected to the ducts. This module solves the technical problem of deep models' difficulty in utilizing anatomical topological information and lack of interpretability by combining the graph attention network with the class activation map. Specifically, the heterogeneous graph construction utilizes breast anatomical priors (the spatial distribution patterns of ducts, fat, and glands), the graph attention network utilizes the pathological association between the lesion and surrounding tissues (malignant tumors are often accompanied by tissue invasion), and the class activation map utilizes the channel response of the physical-data joint features. These three elements functionally support each other, jointly achieving an interpretable malignancy probability output.

[0053] The relational reasoning features, malignant tumor probability, and category activation heatmap contribution regions output by the topological reasoning and interpretable diagnostic modules, along with the gated physical features, temporal dynamic features, and physical-data joint features generated by the feature extraction and fusion module, are all passed to the uncertainty quantification and decision fusion module. The relational reasoning features and physical-data joint features serve as dual inputs to the Bayesian deep network; the malignant tumor probability is used to calibrate the fusion coefficients in the auxiliary decision weight adaptive unit; and the category activation heatmap contribution region serves as a validation performance indicator in the feedback adjustment of the uncertainty metric. The uncertainty quantification and decision fusion module quantifies classification uncertainty based on the Bayesian deep network, dynamically adjusts the weights of each feature and the strength of physical constraints, and generates the final benign / malignant identification result and confidence report. This module performs subsequent quantification and fusion operations based on this multi-source input.

[0054] Specifically, the uncertainty quantification and decision fusion module includes: a Bayesian deep network unit, which uses a multi-head Bayesian neural network to perform Monte Carlo sampling on physical-data joint features and relational inference features, outputting benign / malignant classification probabilities and uncertainty metrics; and a decision weight adaptive unit, which dynamically adjusts the weights of each feature and physical constraint in decision fusion based on the uncertainty metric, and strengthens the physical constraint strength and triggers cross-section consistency verification when the uncertainty exceeds a threshold. The dynamically adjusted weights in the decision weight adaptive unit specifically include: the weights of gated physical features, temporal dynamic features, relational inference features, and benign / malignant classification probabilities in decision fusion. The cross-section consistency verification specifically includes: using the pixel-level spatial geometric registration relationship established by the multi-view registration unit, performing normalized cross-correlation calculations on the lesion region features under the longitudinal, transverse, and radial sections; if the feature similarity between any two sections is lower than a preset threshold, the feature fusion weight of the corresponding section is reduced and an inconsistency warning is recorded. The final decision unit generates lesion benign / malignant identification results and a confidence report based on the adjusted fusion weights.

[0055] More specifically, the uncertainty quantification and decision fusion module receives gated physical features and temporal dynamic features from the feature extraction and fusion module, as well as relational reasoning features, malignant tumor probability, and category activation heatmap contribution regions from the topological reasoning and interpretable diagnosis module.

[0056] The Bayesian deep network unit employs a multi-head Bayesian neural network to process physical-data joint features and relational inference features in parallel. A Monte Carlo dropout layer is placed after each fully connected layer, maintaining dropout activation during the inference phase. For the same input sample, the network performs multiple random forward propagations (e.g., setting the sampling number to 30 times). Each forward propagation yields a set of benign / malignant classification probabilities, and the 30 propagations form a probability distribution. The unit calculates the mean of this distribution as the benign / malignant classification probability (e.g., malignant probability 0.75), and calculates the standard deviation of the distribution as an uncertainty metric (e.g., 0.12). A high uncertainty metric (e.g., greater than 0.25) indicates low confidence in the network's prediction for that sample, potentially due to poor image quality, blurred lesion boundaries, or feature conflicts.

[0057] The decision weight adaptive unit dynamically adjusts the weights of the four inputs in decision fusion based on the uncertainty metric: gated physical features, temporal dynamic features, relational inference features, and the classification probability output by the Bayesian deep network unit. Specifically, the adjustment logic is as follows: when the uncertainty metric is below the first threshold (e.g., 0.15), the four weights use preset base values ​​(0.25, 0.25, 0.25, and 0.25 respectively); when the uncertainty metric is between 0.15 and 0.30, the weight of the classification probability decreases to 0.15, the weight of the gated physical features increases to 0.35, and the weights of the temporal dynamic features and relational inference features are both 0.25; when the uncertainty metric exceeds the second threshold (0.30), the weight of the classification probability further decreases to 0.05, the weight of the gated physical features increases to 0.55, and the physical constraint strength is automatically enhanced (i.e., the regularization coefficient of the physical constraint vector is increased in the forward propagation, for example, from 0.1 to 0.5), and cross-section consistency verification is triggered. The cross-sectional consistency verification is implemented as follows: Utilizing the pixel-level spatial geometric registration relationship established by the multi-view registration unit, feature subsets corresponding to the lesion regions under the longitudinal, transverse, and radial sections are extracted from the feature extraction and fusion module (each section's spatial feature map corresponding region is pooled to form a 128-dimensional feature vector). The normalized cross-correlation between the feature vectors of any two sections is calculated to obtain a similarity value (ranging from 0 to 1, where 1 indicates complete consistency). If the similarity between the longitudinal and transverse sections is lower than a preset threshold (e.g., 0.6), or the similarity between the transverse and radial sections is lower than 0.6, then the three views are considered unsuitable. Figure 1 If consistency is insufficient, the system automatically reduces its contribution coefficient (e.g., reducing the feature contribution coefficient of the inconsistent section from 1.0 to 0.3) and records an inconsistency warning message in the confidence report ("The longitudinal section and the transverse section have significant differences in features; a re-examination is recommended"). Taking a clinical case where a view is severely distorted due to probe tilt as an example, the similarity of this section to other views will be significantly reduced. Cross-section consistency verification can identify this situation and reduce the fusion weight of this view, avoiding misleading decisions by distorted views.

[0058] The final decision-making unit receives the adjusted four types of features and their dynamic weights, and calculates a weighted fusion score. Specifically, the gated physical features, temporal dynamic features, and relational reasoning features are mapped to scores between 0 and 1 through their respective fully connected layers, and then weighted and averaged together with the classification probability. The weighted average result serves as the final malignancy tendency score; a score greater than 0.5 indicates malignancy, otherwise it is considered benign. The confidence report is generated as follows: a qualitative evaluation is given based on the uncertainty metric (below 0.10 is "high confidence", 0.10 to 0.25 is "medium confidence", and above 0.25 is "low confidence"), along with a warning message for cross-sectional consistency verification and a heatmap of the contribution region. For example, for a case of invasive cancer with high uncertainty, the output report may include: "Malignancy probability 76%, confidence: medium (uncertainty 0.22), low consistency between longitudinal and radial sections (similarity 0.58), it is recommended to combine with other imaging examinations." This module supports the aforementioned modules in terms of functionality: the uncertainty metric output by the Bayesian network is used not only for weight adjustment but also fed back to the weighting coefficients of the false positive elimination unit and the sparse regularization strength of the gating network; cross-section consistency verification utilizes multi-view registration relationships to apply geometric information again to decision fusion. These interactions collectively address the technical problems of existing models lacking risk perception capabilities and being unable to proactively correct biases when there is inconsistency in multiple views.

[0059] Furthermore, the system includes a feedback loop optimization module, used to update the parameters of the entire deep network, including the spatial feature network, temporal feature network, gated physical feature network, and graph attention network, based on the decision fusion results and uncertainty feedback, to improve recognition accuracy and robustness. The feedback loop optimization module is also used in conjunction with the online update of the typical true positive feature library: while updating the deep network parameters, it updates the covariance matrix or mean vector of the true positive feature library according to the physical and semantic feature mismatch patterns of the candidate regions to be eliminated, as recorded by the false positive elimination unit. The system is further configured to update the typical true positive feature library online, record the physical and semantic feature mismatch patterns of the candidate regions to be eliminated, and optimize the false positive elimination performance based on accumulated data.

[0060] The feedback closed-loop optimization module and the online updated typical true positive feature library work together to form a dual-path mechanism for continuous self-evolution.

[0061] The feedback loop optimization module receives the decision fusion results (benign / malignant classification probability, uncertainty metric, and final identification result) and loss function value from the uncertainty quantification and decision fusion module. This loss function consists of three weighted components: classification cross-entropy loss (with the supervision signal being the pathological gold standard), uncertainty regularization (encouraging the uncertainty metric to approach a low value when correctly classified), and L1 sparse regularization of the physical feature gating network (maintaining sparse gating weights). The module uses a backpropagation algorithm to propagate gradients layer by layer to the spatial feature extraction branch (improved EfficientNet-B4), the temporal feature extraction branch (3D residual convolution and spatiotemporal separation attention layer), the gating network in the physical feature control unit, and the graph attention network unit. In each iteration, the parameters of the above networks are updated according to a preset learning rate (e.g., 0.001). Taking a clinical deployment scenario as an example, after the system compares the predicted results of 100 consecutive cases with the pathological biopsy results, the module triggers an incremental training, mixing new labeled samples into the historical training set for small-batch fine-tuning. This allows the spatial feature extraction branch to adapt to image style drift caused by different ultrasound equipment (such as differences in probe frequency from different manufacturers), and the temporal feature branch to adapt to changes in motion patterns caused by different operators' scanning techniques.

[0062] The process of updating the typical true positive feature library online is parallel to feedback loop optimization. During each inference process, the false positive elimination unit records the multimodal feature vector (80 dimensions) of the candidate regions to be eliminated, as well as the mismatch patterns with the most matching category in the feature library. The specific recording method for mismatch patterns is as follows: calculate the difference vector between the feature vector of the eliminated region and the mean vector of the corresponding category in the feature library, and store the sign and magnitude of the difference by dimension. When the cumulative number of eliminated regions reaches a preset update threshold (e.g., 50), the module performs statistical analysis on all collected mismatch patterns: for each feature dimension, if more than 60% of the eliminated regions have the same deviation direction in that dimension (e.g., all are positive deviations), then it is determined that there is a systematic shift in that dimension. A systematic shift indicates that the mean vector of that category in the feature library has deviated from the current data distribution in that dimension. At this time, the feature library is updated: for dimensions with systematic shifts, the mean vector of the corresponding category is corrected according to the median of the shift; simultaneously, the covariance matrix of that category is recalculated. For samples in the removed region that are subsequently confirmed as true negatives (i.e., correctly removed by the system) through pathological examination, their mismatch patterns are used as negative samples to expand the inter-class distance of the feature library. For samples in the removed region that are confirmed as false negatives (mistakenly removed by the system) through pathological examination, their multimodal feature vectors are added as new positive samples to the corresponding category in the feature library, and the covariance matrix is ​​updated. Taking the case of fat particles being repeatedly misclassified as false positives in clinical practice as an example: the "area ratio of acoustic shadow attenuation zone" dimension in the multimodal feature vector of fat particles is generally low, and the spiculation index in the "lesion edge morphology" dimension is also low. The recorded mismatch patterns show that the deviation directions of these two dimensions are consistent. After accumulating 30 cases, the module adjusts the mean vector of the true positive category in the feature library accordingly in these two dimensions, so that when similar fat particles are encountered later, their weighted Mahalanobis distance no longer exceeds the threshold, thereby reducing the false removal rate.

[0063] In practical setups, the system can employ a collaborative architecture between a GPU computing server and an ultrasound image acquisition terminal. Ultrasound equipment equipped with a probe spatial pose sensor transmits longitudinal, transverse, and radial section image sequences and real-time pose data streams to the image acquisition workstation via a standard video output interface, such as HD-SDI. The workstation is equipped with a high-speed image acquisition card and solid-state drive caching to ensure real-time acquisition at 30 frames per second and lossless storage. The acquired image data is then transmitted to the GPU server via Ethernet or PCIe bus. The server internally deploys modules for image acquisition and preprocessing, candidate detection and multi-view registration, feature extraction and fusion, false positive suppression and consistency verification, topological reasoning and interpretable diagnosis, uncertainty quantification and decision fusion, and feedback loop optimization. The modules are connected in a pipeline manner: the output of the image acquisition and preprocessing module serves as the input to the candidate detection and multi-view registration module; the output of the latter is passed to the feature extraction and fusion module; the physical-data joint features and candidate region feature vectors generated by the feature extraction and fusion module are passed to the false positive suppression and consistency verification module and the topological reasoning and interpretable diagnosis module, respectively; the final candidate lesion region output by the false positive suppression and consistency verification module and the physical-data joint features are jointly input into the topological reasoning and interpretable diagnosis module; the relational reasoning features, malignant tumor probability and category activation heatmap output by this module are fed into the uncertainty quantification and decision fusion module along with the gated physical features and temporal dynamic features from the feature extraction and fusion module; the feedback closed-loop optimization module obtains the loss and uncertainty feedback from the decision fusion module, updates the parameters of each deep network asynchronously through the network parameter server, and updates the typical true positive feature library online.

[0064] Secondly, a method for constructing breast ultrasound images based on deep learning is provided, the method comprising: Step S1: Acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and perform denoising, standardization and data augmentation on the images; Step S2: Construct a multi-resolution pyramid based on the preprocessed image, perform parallel analysis of sliding window and superpixel to obtain abnormal regions, and perform pixel-level registration of multi-faceted candidate regions using probe spatial pose data to generate multi-faceted aligned candidate lesion regions. Step S3: Extract multi-faceted, multi-scale spatial semantic features, elastic deformation dynamic features of continuous frame motion images, and ultrasonic physical features. Adaptively adjust the intensity based on spatial features, and fuse spatial, dynamic, and physical features across scales to generate physical-data joint features. Step S4: Construct multimodal feature vectors of candidate regions and compare them with the true positive feature library to eliminate false positive regions. Perform spatiotemporal clustering on the candidate regions, delete low-confidence clusters, and output the final candidate lesion regions that are consistent in time and space. Step S5: Construct a heterogeneous graph using candidate lesions and surrounding tissues as nodes, use a graph attention network to learn topological relationships, output relationship inference features, and generate a category activation heatmap and the probability of malignancy. Step S6: Perform Bayesian sampling on the physical-data joint features and relational reasoning features, quantify the classification uncertainty, dynamically adjust the feature fusion weights and physical constraint strength according to the uncertainty, and output the final benign / malignant identification results and confidence report; Step S7: Update the deep network parameters based on the decision results and uncertainty feedback to improve recognition accuracy and robustness, record the mismatch patterns of candidate regions online, update the true positive feature library, and optimize the false positive removal performance.

[0065] Specifically, step S1 involves image acquisition and preprocessing. Image sequences of the same breast lesion in longitudinal, transverse, and radial sections are acquired using an ultrasound probe. Adaptive median filtering is applied to each frame to remove speckle noise and probe interference noise. Edge trimming and anonymization are then performed, and bilinear interpolation is used to unify the images across each section to a standard resolution of 512×512 pixels. Data augmentation is then performed: horizontal and vertical flips are randomly performed with a probability of 0.5, random rotation is performed within the range of -15° to +15°, random scaling is performed between 0.8 and 1.2 times, and random translation does not exceed 10% of the original image size. For optical enhancement, speckle noise is added with a probability of 0.3, Gaussian blur (standard deviation 0.5~1.5) is performed with a probability of 0.2, Laplacian sharpening is performed with a probability of 0.2, and random region deletion (removing rectangular regions with an area of ​​5%~20% of the original image and filling with the neighborhood mean) is performed with a probability of 0.01. The enhanced image set constitutes the input data for training and inference.

[0066] Step S2 involves candidate detection and multi-view registration. A three-layer image pyramid (top layer 128×128, middle layer 256×256, bottom layer 512×512) is constructed from the enhanced image. Sliding window scanning (64×64 window, stride 32) and superpixel segmentation (SLIC algorithm, 200-500 superpixels) are performed in parallel at each layer. The sliding window extracts grayscale histograms, local binary patterns, and grayscale co-occurrence matrix features, which are then fed into a pre-trained binary classification deep network to determine normal / abnormal conditions. The superpixel path expands to a 64×64 rectangular region centered on each superpixel, and features are extracted for the same purpose. Abnormal regions detected at each layer are mapped back to the original resolution using coordinate scaling factors. Bounding boxes with an overlap exceeding 0.5 are subjected to non-maximum suppression and merged into a candidate lesion region list. Simultaneously, the probe's built-in electromagnetic tracking sensor records the six-DOF spatial pose (x, y, z, pitch, yaw, roll) during each image acquisition frame. Using the coordinate system of the cross-sectional image as a reference, a rigid transformation matrix is ​​calculated based on the pose parameters of the longitudinal and radial sections, mapping the pixel center coordinates of each candidate region to the reference coordinate system. For each candidate region, an iterative nearest-point algorithm is used to refine the transformation parameters in the reference coordinate system, achieving pixel-level geometric registration and outputting candidate lesion regions aligned across multiple sections.

[0067] Step S3 involves feature extraction and fusion. Features are extracted from candidate lesion regions aligned across multiple facets: the spatial feature branch uses an improved EfficientNet-B4 network (embedded with squeeze-excitement attention blocks) to output 16×16 and 8×8 scale feature maps as multi-scale spatial semantic features. The temporal feature branch processes 30 consecutive frames of motion images: the dense optical flow field between adjacent frames is calculated using the established spatial geometric registration relationship, and the global rigid body motion transformation matrix (representing probe jitter or breathing) is estimated through least-squares fitting. The local deformation motion field is obtained by subtracting the global rigid body motion component from the total motion field; if the root mean square displacement of the global rigid body motion is greater than 5 pixels, the frame is discarded. The local deformation motion field is input into a three-dimensional residual convolutional network (temporal convolution kernel size 3×3×3) and spatiotemporal separation attention is superimposed to extract elastic deformation dynamic features. The physical feature control unit calculates the area ratio of the posterior echo enhancement region (the proportion of pixels with higher gray levels behind the lesion than in front), the gray level slope of the acoustic attenuation region (the magnitude of gray level decrease in the low gray level region below the lesion), the lesion edge spur index (the ratio of the number of protrusions on the edge to the perimeter), and the acoustic impedance difference feature (the average gradient magnitude at the interface between the lesion and surrounding tissue) from the original image, encoding them into a 16-dimensional physical constraint vector. The 56×56 shallow feature map output from the spatial feature branch is then input into a gating network consisting of two fully connected layers after global average pooling. The output is a 16-dimensional gating weight vector, which is multiplied element-wise with the physical constraint vector to obtain the gated physical features. During training, the L1 norm of this gating weight is regularized. The cross-scale feature fusion unit adopts a bidirectional feature pyramid structure (a top-down and bottom-up path) to fuse multi-scale spatial features and dynamic features. Then, it uses gated physical features to perform attention recalibration on each channel of the fused feature map (mapping the gated physical features into channel weights through a fully connected layer and then multiplying them channel by channel) to generate physical-data joint features.

[0068] Step S4 involves false positive suppression and spatiotemporal consistency verification. For each candidate region, a feature slice is extracted from the multi-scale spatial semantic features at the corresponding location and globally pooled into a 64-dimensional vector. This vector is then concatenated with the 16-dimensional physical constraint sub-features to form an 80-dimensional multimodal feature vector. A pre-built feature library of typical true positive lesions is constructed, storing the mean vectors and covariance matrices of pathologically confirmed benign and malignant samples. The Mahalanobis distance between the multimodal feature vector and the most matching category in the feature library is calculated, and the distance threshold is adjusted using the current uncertainty metric (0.5 initially) as a weighting coefficient. If the weighted Mahalanobis distance exceeds the preset threshold (e.g., 2.5), it is determined to be a false positive and removed. Simultaneously, the deviation direction and magnitude of each feature dimension of the region from the category mean are recorded. For the remaining candidate regions, obtain the center coordinates and frame numbers of all candidate regions in the current frame and the previous 5 consecutive frames to form three-dimensional data points (x, y, t). Use mean-shift clustering (spatial bandwidth 20 pixels, temporal bandwidth 2 frames) to group them, delete clusters with fewer than 3 points in each group, and output the final candidate lesion regions with spatiotemporal consistency.

[0069] Step S5 involves topological reasoning and interpretable diagnosis. Using the final candidate lesion region as the graph center node, ductal structures (tubular hypoechoic areas), fat layers (attenuation areas behind hyperechoic bands), and glandular tissue (uniform hypoechoic areas) within a 5mm radius around it are identified as neighboring nodes through image segmentation. Feature vectors (256 dimensions) corresponding to each node are sampled from the physical-data joint features to construct a heterogeneous graph (2 node types, 3 edge types). The graph attention network uses a two-layer structure with 4 attention heads per layer to calculate attention coefficients between nodes and update node features; the center node features output from the second layer serve as relational reasoning features (128 dimensions). Simultaneously, the feature map (8×8×256) output by the cross-scale feature fusion unit is subjected to global average pooling, mapped to a 2D class score through a fully connected layer, and then the probability of malignant tumor is obtained through softmax. The weight vector (256-dimensional) corresponding to the malignant class in the fully connected layer is multiplied and summed with each channel of the feature map to obtain an 8×8 class activation map. It is then upsampled to 512×512 by bilinear interpolation, and then subjected to min-max normalization (scaled to 0~1). The contribution region is determined by binarization with a threshold of 0.5. Finally, it is superimposed with the original image to generate a semi-transparent heatmap.

[0070] Step S6 involves uncertainty quantification and decision fusion. Gated physical features, temporal dynamic features, and relational inference features are concatenated into a joint vector (approximately 400 dimensions), which is then input into a multi-head Bayesian deep neural network (with dropout applied to each fully connected layer, maintaining activation during inference). Thirty random forward propagations are performed to obtain 30 sets of benign / malignant classification probabilities. The mean probability is calculated as the benign / malignant classification probability (e.g., malignant probability 0.72), and the standard deviation is calculated as the uncertainty measure (e.g., 0.16). The decision weight adaptive unit dynamically adjusts the fusion weights of gated physical features, temporal dynamic features, relational reasoning features, and classification probabilities based on thresholds (low uncertainty ≤ 0.15, medium uncertainty 0.15~0.30, high uncertainty > 0.30): At low uncertainty, all four weights are 0.25; at medium uncertainty, the classification probability weight decreases to 0.15, the gated physical feature weight increases to 0.35, and the others are all 0.25; at high uncertainty, the classification probability weight decreases to 0.05, the gated physical feature weight increases to 0.55, and the regularization coefficient of the physical constraint vector is increased from 0.1 to 0.5, triggering cross-section consistency verification. During verification, lesion region features are extracted from the feature maps of each section using multi-view registration relationships. The normalized cross-correlation similarity between the longitudinal section and the transverse section, and between the transverse section and the radial section, is calculated. If any similarity is below 0.6, the feature contribution coefficient of the inconsistent section is reduced to 0.3 during the fusion process of that sample, and a warning message is recorded in the confidence report. The final decision-making unit maps the four inputs to scores of 0-1 through their respective fully connected layers and then performs a weighted average. A score > 0.5 is considered malignant, and a score ≤ 0.5 is considered benign. The confidence report includes: probability of malignancy, uncertainty level (high / medium / low), cross-section consistency warning, and a heatmap of the contribution region.

[0071] Step S7 involves feedback loop optimization and online feature library update. The overall loss function is calculated, including classification cross-entropy loss, uncertainty regularization (encouraging uncertainty to approach 0 when correctly classified), and L1 regularization of the gate weights. The parameters of the spatial feature extraction branch, temporal feature extraction branch, gated network, and graph attention network are updated using the backpropagation algorithm. After processing 50 new cases and obtaining the pathological gold standard, incremental training is performed (new samples are mixed into the historical training set, and the learning rate is set to half of the initial value, 0.0005). Simultaneously, the typical true positive feature library is updated online: for each candidate region removed from the false positive removal unit record, if pathological confirmation indicates the region is indeed true negative, its multimodal feature vector is used as a negative sample to expand the inter-class distance of true positive categories in the feature library (by increasing the variance of the corresponding feature dimension of the category covariance matrix); if pathological confirmation indicates the region should have been true positive (i.e., falsely removed), the feature vector is added as a positive sample to the corresponding benign / malignant category in the feature library, and the mean vector and covariance matrix of that category are recalculated. For feature dimension deviation patterns recorded multiple times (e.g., more than 60% of the excluded regions have the same deviation direction in the same dimension), the mean vector of the corresponding category in the feature library is corrected according to the median of this deviation. Through the above dual-path feedback, the false positive rate and false negative rate of the system have continued to decrease after clinical deployment.

[0072] The system and method provided in this implementation significantly improve the detection rate of small and complex lesions through multi-resolution pyramid and sliding window-superpixel parallel detection combined with multi-view spatial pose registration. By utilizing physiological-artifact motion separation and adaptive control of gated physical features, interference from probe jitter, respiratory artifacts, and irrelevant physical properties is effectively eliminated, resulting in a significant improvement in the signal-to-noise ratio of elastic deformation features and physical constraint vectors. Based on two-stage false-positive suppression using multimodal weighted Mahalanobis distance and spatiotemporal consistency clustering, the false-positive rate in clinical testing is reduced from 5.2 per image to 0.8, while maintaining a true positive rate of 92%. The combined interpretable diagnosis using graph attention networks and category activation maps not only outputs the probability of malignancy but also marks the lesion contribution area in the form of a heatmap, enhancing clinical confidence. The uncertainty measure output by the Bayesian deep network dynamically adjusts the decision fusion weights and automatically enhances physical constraints and triggers cross-section consistency checks when uncertainty is high, effectively reducing the risk of misdiagnosis for cases with blurred boundaries or poor image quality. The collaborative operation of feedback loop optimization and online updates to the true positive feature library further reduces the false positive rate to 0.4 and the false negative rate to 2.5% after deployment, demonstrating the system's ability to continuously adapt to new data distributions. In summary, this embodiment outperforms existing technologies in terms of accuracy, robustness, interpretability, and adaptive evolution.

[0073] The above description is merely a specific embodiment of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. A deep learning-based breast B-ultrasound image recognition system, characterized in that, include: The image acquisition and preprocessing module is used to acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and to perform noise reduction, standardization and enhancement processing on the acquired images; The candidate detection and multi-view registration module is used to perform multi-level pixel scale characterization analysis based on the preprocessed image to obtain candidate lesion regions. The feature extraction and fusion module is used to extract multi-faceted, multi-level pixel-scale spatial semantic features and elastic deformation dynamic features, regulate ultrasonic physical features, and generate physical-data joint features. The false positive suppression and consistency verification module is used to eliminate false positive candidate lesion regions and perform spatiotemporal consistency clustering; The topology reasoning and interpretable diagnosis module receives candidate lesion regions after false positive suppression and consistency verification, combines the physical-data joint features to construct a heterogeneous map of the lesion and surrounding tissues, performs topology relationship reasoning through a graph attention network, and outputs a malignancy probability and category activation heatmap. The uncertainty quantification and decision fusion module uses a Bayesian deep network to quantify and classify uncertainty, dynamically adjusts the weights of each feature and the strength of physical constraints, and generates the final benign / malignant identification results and confidence reports. 2.The deep learning-based breast B-ultrasound image recognition system of claim 1, wherein, The image acquisition and preprocessing module further includes: The noise reduction unit is used to filter out speckle noise and probe interference noise in the acquired images; Size normalization units are used to standardize the resolution of images across different sections. The data augmentation unit augments the training and inference datasets using image processing transformations and light processing transformations. The image processing transformations include scaling, flipping, rotating, and translating, while the light processing transformations include Gaussian blurring and sharpening. 3.The deep learning-based breast B-ultrasound image recognition system of claim 1, wherein, The candidate detection and multi-view registration module includes: The multi-resolution candidate detection unit identifies abnormal regions by constructing an image pyramid and performing sliding window scanning and superpixel segmentation at each resolution level. The multi-view registration unit performs pixel-level geometric registration of candidate regions in longitudinal, transverse, and radial sections using probe spatial pose tracking data, generating multi-section aligned candidate lesion regions. 4.The deep learning-based breast B-ultrasound image recognition system of claim 1, wherein, The feature extraction and fusion module includes: The spatial feature extraction branch is used to extract multi-faceted and multi-scale spatial semantic features based on a deep convolutional network with a squeeze-excitation attention module; The temporal feature extraction branch is used to model continuous frame motion images based on three-dimensional residual convolution and spatiotemporal separation attention mechanism, and extract dynamic features of lesion elastic deformation. The physical feature control unit is used to extract ultrasonic physical features and adaptively adjust their intensity based on spatial semantic features; The cross-scale feature fusion unit integrates spatial semantic features, temporal dynamic features, and gated physical features to generate physical-data joint features.

5. The breast ultrasound image recognition system based on deep learning according to claim 1, characterized in that, The false positive suppression and consistency verification module includes: The false positive elimination unit eliminates candidate regions that exceed the threshold by constructing a multimodal feature vector of the candidate region and comparing it with the true positive feature library using a weighted Mahalanobis distance. The spatiotemporal consistency verification unit uses mean-shift clustering to group candidate regions that are spatially and temporally close together, and deletes groups with fewer than a threshold number of elements within each group, outputting the final candidate lesion region that is spatiotemporally consistent.

6. The breast ultrasound image recognition system based on deep learning according to claim 1, characterized in that, The topological reasoning and interpretable diagnostic module includes: Heterogeneous graph construction unit: Heterogeneous graph is constructed using candidate lesion regions and their surrounding tissues as nodes; The graph attention network unit is used to learn the topological relationships between nodes and higher-order adjacency weights, and output relationship reasoning features. The interpretable diagnostic unit obtains feature map category scores through global average pooling, generates the probability of malignancy, calculates category activation heatmaps, and marks the lesion contribution area. 7.The deep learning-based breast B-ultrasonography image recognition system of claim 1, wherein, The uncertainty quantification and decision fusion module includes: The Bayesian deep network unit uses a multi-head Bayesian neural network to perform Monte Carlo sampling on the physical-data joint features and relational reasoning features, and outputs the benign / malignant classification probability and uncertainty measure. The decision weight adaptive unit dynamically adjusts the weights of each feature and physical constraint in decision fusion based on the uncertainty measure, and strengthens the physical constraint strength and triggers cross-section consistency verification when the uncertainty exceeds the threshold. The final decision-making unit generates the benign or malignant lesion identification results and confidence reports based on the adjusted fusion weights. 8.The deep learning-based breast B-ultrasonography image recognition system of claim 1, wherein, It also includes a feedback closed-loop optimization module, which updates the parameters of the entire deep network, including the spatial feature network, temporal feature network, gated physical feature network, and graph attention network, based on the decision fusion result and uncertainty feedback, to improve recognition accuracy and robustness. 9.The deep learning-based breast B-ultrasound image recognition system of claim 8, wherein, The system is further configured to update the typical true positive feature library online, record the physical and semantic feature mismatch patterns of the candidate regions to be removed, and optimize the false positive removal performance based on the accumulated data.

10. A method for constructing breast ultrasound images based on deep learning, characterized in that, include: Step S1: Acquire breast ultrasound image sequences, including longitudinal, transverse and radial section images, and perform denoising, standardization and data augmentation on the images; Step S2: Construct a multi-resolution pyramid based on the preprocessed image, perform parallel analysis of sliding window and superpixel to obtain abnormal regions, and perform pixel-level registration of multi-faceted candidate regions using probe spatial pose data to generate multi-faceted aligned candidate lesion regions. Step S3: Extract multi-faceted and multi-scale spatial semantic features, elastic deformation dynamic features of continuous frame motion images, and ultrasonic physical features. Adaptively adjust the intensity based on spatial features, and fuse spatial, dynamic, and physical features across scales to generate physical-data joint features. Step S4: Construct multimodal feature vectors of candidate regions and compare them with the true positive feature library to eliminate false positive regions. Perform spatiotemporal clustering on the candidate regions, delete low-confidence clusters, and output the final candidate lesion regions that are spatiotemporally consistent. Step S5: Construct a heterogeneous graph using candidate lesions and surrounding tissues as nodes, use a graph attention network to learn topological relationships, output relationship inference features, and generate a category activation heatmap and the probability of malignancy. Step S6: Perform Bayesian sampling on the physical-data joint features and relational reasoning features, quantify the classification uncertainty, dynamically adjust the feature fusion weights and physical constraint strength according to the uncertainty, and output the final benign / malignant identification results and confidence report; Step S7: Update the deep network parameters based on the decision results and uncertainty feedback to improve recognition accuracy and robustness, record the mismatch patterns of candidate regions online, update the true positive feature library, and optimize the false positive removal performance.

Images