Breast ultrasound automatic compression method and device based on image feedback

Abstract

The application provides a breast ultrasound automatic pressing method and device based on image feedback, which comprises the following steps: S1, a mechanical arm carries a control ultrasound probe to press down on a breast area at a preset initial pressure, and collects a transverse section ultrasound image of the breast; S2, a pre-trained target detection network is used to analyze the transverse section ultrasound image in real time, identify the position of the breast area, and calculate the offset of the area relative to the center of the image; S3, a feedback control signal is generated based on the offset, and the mechanical arm is dynamically adjusted to drive the ultrasound probe to move and / or change the applied pressure, so that the breast area is aligned to the center of the image; S4, steps S1 to S3 are repeatedly executed until the offset is less than a preset threshold, and an ultrasound image meeting the quality standard is output; through image feedback and closed-loop control, the application realizes automatic centering of the breast area in the ultrasound image and stable imaging, and can effectively improve the accuracy and reliability of breast ultrasound detection.

Description

Technical Field

[0001] This invention relates to the field of medical examination technology, and in particular to a method and apparatus for automatic pressure application of breast ultrasound based on image feedback. Background Technology

[0002] Breast ultrasound (BUS), as one of the core imaging technologies for breast disease screening and diagnosis, generates real-time images of breast tissue through high-frequency sound wave scanning. It effectively assists doctors in identifying lesions such as tumors and nodules, and has significant clinical application value. However, the quality and accuracy of ultrasound images are highly dependent on the pressure applied by the ultrasound probe and its positioning precision. Uneven pressure (such as local overpressure or underpressure) or probe position deviation (such as left-right / up-down offset) can easily lead to breast tissue deformation and abnormal acoustic impedance, resulting in image distortion (such as structural distortion and blurred boundaries). In severe cases, this may cause missed or misdiagnosed lesions.

[0003] In recent years, automated breast ultrasound (ABUS) systems have gradually emerged, using robotic arms or robots to control probe movement, aiming to replace manual operation and achieve precise control of pressure and position. However, current technology still faces the following key bottlenecks:

[0004] Insufficient pressure control precision: Most existing ABUS systems use preset fixed pressure values ​​to drive the probe, without taking into account individual patient differences (such as differences in breast density, gland thickness, and anatomical structure—dense breasts require lower pressure to avoid excessive compression and deformation, while loose breasts require appropriate pressure to ensure coupling effect), resulting in poor pressure and tissue compatibility and image quality that varies from person to person.

[0005] Image quality assessment is lagging: Current image quality assessment mainly relies on manual visual inspection or traditional image processing techniques, which makes it difficult to identify subtle problems (such as slight left / right / up / down offsets, local distortions caused by uneven local pressure) in real time and accurately. Especially during dynamic scanning, existing methods have low sensitivity to image offsets and are prone to missing the clear imaging requirements of critical areas.

[0006] Inadequate feedback mechanism: Although some ABUS systems attempt to combine image feedback to optimize pressure control, the image analysis technologies they rely on (such as edge detection and grayscale analysis) can only provide basic feature information. Furthermore, there is a time delay between pressure adjustment and image acquisition, and the adjustment strategy lacks dynamic adaptability, resulting in a discontinuous and inefficient optimization process that is difficult to meet the requirements of clinical diagnosis for high real-time performance and high accuracy.

[0007] Therefore, there is an urgent need for an automatic pressure application method and device for breast ultrasound that can achieve precise closed-loop control based on real-time image feedback, dynamic correlation between image quality and probe parameters, and improve the accuracy and reliability of breast ultrasound detection. Summary of the Invention

[0008] To address the aforementioned deficiencies in the prior art, the present invention aims to provide an automatic breast ultrasound pressure method and apparatus based on image feedback, thereby solving the technical problems of image offset, inaccurate pressure control, and feedback delay in traditional automatic breast ultrasound.

[0009] To achieve the above objectives, in a first aspect, the present invention provides a method for automatic breast ultrasound compression based on image feedback, the steps of which include:

[0010] S1. The robotic arm carries and controls the ultrasound probe to press down on the breast area with a preset initial pressure, and acquires a cross-sectional ultrasound image of the breast.

[0011] S2. The cross-sectional ultrasound image is analyzed in real time using a pre-trained target detection network to identify the location of the breast region and calculate the offset of the region relative to the image center.

[0012] S3. Generate a feedback control signal based on the offset, dynamically adjust the robotic arm to drive the ultrasound probe to move and / or change the applied pressure, so that the breast region is aligned with the image center;

[0013] S4. Repeat steps S1 to S3 until the offset is less than the preset threshold, and output an ultrasound image that meets the quality standard.

[0014] As a further improvement to the above scheme, the target detection network is a lightweight real-time target detection network for breast ultrasound regions based on YOLO, which includes an LRMBu backbone network, a neck network, and a segmentation head.

[0015] The LRMBu backbone network includes an initial convolutional block, a depth block, and an integrated multi-channel visual space (IMCVSS) block connected in sequence, used to extract multi-scale features from the input image;

[0016] The neck network is used to fuse the multi-scale features;

[0017] The segmentation head is used to output target detection box information for the breast region.

[0018] As a further improvement to the above scheme, the integrated multi-channel visual spatial block includes a dynamic scale channel convolution module (DSCC) and a diversified channel scanning mechanism (DCSM).

[0019] The dynamic scale channel convolution module is used to group input features by channel and fuse cross-group features; the diversified channel scanning mechanism is used to perform multi-path scanning of feature maps to fuse long-distance spatial dependencies.

[0020] As a further improvement to the above scheme, the operation of the dynamic scale channel convolution module includes: dividing the input feature map into multiple groups along the channel dimension; performing adaptive channel feature recalibration on each group of features; concatenating the calibrated feature groups and restoring the channel dimension through convolution operation to obtain the enhanced output features.

[0021] As a further improvement to the above solution, the diversified channel scanning mechanism includes multi-channel scanning and multi-channel fusion, and the specific operations are as follows:

[0022] The input feature f[C,H,W] is first processed by a 1×1 convolution to reduce the channel dimension to C / 4, resulting in f1[C / 4,H,W].

[0023] During the multi-channel scanning stage, f1[C / 4,H,W] is copied into multiple copies, and various scanning modes are executed along multiple different scanning paths. The results are then concatenated into f2[C,H,W] to maintain the channel dimension while integrating multiple scanning information.

[0024] Then, the number of channels is expanded to 4C by 1×1 convolution, resulting in f3[4C,H,W];

[0025] Finally, the feature results are summed according to the scanning order to obtain the final output feature f'[C,H,W].

[0026] As a further improvement to the above solution, the specific method for calculating the offset in step S2 is as follows:

[0027] Based on the center coordinates (x) of the target detection box center ,y center Calculate the horizontal offset X based on the difference between the coordinates of the image center (W / 2, H / 2) and the coordinates of the image center. 偏移 and vertical offset Y 偏移 The specific calculation formula is as follows:

[0028] ;

[0029] ;

[0030] Where W is the width of the ultrasound image and H is the height of the ultrasound image;

[0031] The combined offset error Δd is calculated based on the horizontal and vertical offsets; specifically, the formula for calculating the combined offset error Δd is as follows:

[0032] .

[0033] As a further improvement to the above scheme, in step S3, the dynamic adjustment specifically includes: if Δd is greater than a preset threshold... Then according to X 偏移 The positive and negative controls allow the robotic arm to move the probe horizontally, according to Y. 偏移 The positive and negative controls allow the robotic arm to adjust the probe pressure in the vertical direction;

[0034] The moves Move(X) and Move(Y) are respectively given by the formula Move(X) = K × X 偏移 and Move(Y) = K × Y 偏移 Determined, where K is the adjustable proportional coefficient.

[0035] As a further improvement to the above scheme, a high-quality image selection step is included after step S4: from the N frames of images acquired by multiple scans, the image with the largest target detection box area w×h is selected as the final high-quality breast ultrasound image output.

[0036] In a second aspect, the present invention also provides an automatic breast ultrasound compression device based on the image feedback-based automatic breast ultrasound compression method provided in the first aspect, comprising:

[0037] The robotic arm is used to carry and precisely control the position and orientation of the ultrasound probe, and can move and adjust pressure in multiple degrees of freedom according to control commands.

[0038] An ultrasound probe is used to emit ultrasound waves and receive echo signals to generate a cross-sectional ultrasound image of the breast.

[0039] The image acquisition module is used to acquire ultrasound image data collected by the ultrasound probe in real time;

[0040] The processing control unit is configured to run a pre-trained target detection network, identify breast regions, calculate breast region offsets, and generate feedback signals for controlling the robotic arm; wherein the processing control unit is communicatively connected to the robotic arm and the image acquisition module to form a closed-loop control system.

[0041] As a further improvement to the above scheme, the target detection network is trained on a dataset containing cross-sectional images of breast tissue, which is augmented with data, including one or more operations such as rotation, translation, and scaling.

[0042] As a further improvement to the above solution, the device also includes a display module for displaying ultrasound images, breast region detection frames and offset information in real time, and providing image quality compliance prompts.

[0043] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows:

[0044] This invention provides an automated pressure application method for breast ultrasound based on image feedback. By synergistically combining robotic arm control, real-time image analysis, and closed-loop feedback adjustment, it effectively solves the problems of image shift, inaccurate pressure control, and feedback delay in traditional breast ultrasound imaging, thereby improving the quality and clinical applicability of automated breast ultrasound image acquisition. Specifically, on the one hand, by employing a lightweight real-time target detection network based on YOLO and a modified Mamba structure, the breast region in the breast ultrasound image can be quickly and accurately identified. Compared to traditional methods relying on edge detection or grayscale analysis, this improves the target detection accuracy in complex ultrasound image backgrounds, ensuring accurate framing of the breast region and providing a reliable basis for subsequent position adjustments. On the other hand, by analyzing the real-time acquired ultrasound image, the offset of the breast region relative to the image center is calculated, and a feedback control signal is generated based on this offset to dynamically adjust the position of the robotic arm and the probe pressure, gradually aligning the breast region with the image center. This closed-loop control mechanism effectively solves the image shift problem caused by manual operation or static settings in traditional methods, improving the spatial consistency of the image.

[0045] This invention achieves dynamic adjustment of probe pressure through real-time image feedback and robotic arm linkage, avoiding the problems of inaccurate pressure control and tissue deformation or image distortion caused by traditional automatic pressurization systems. It ensures appropriate and stable contact pressure is maintained for different patients or at different scanning stages, improving the controllability of image quality. By introducing an image feedback mechanism and automated adjustment process, it replaces the traditional manual adjustment method relying on physician experience, reducing image quality fluctuations caused by operator fatigue and subjective judgment differences. This makes the ultrasound image acquisition process more standardized and repeatable, contributing to improved objectivity and efficiency in clinical testing. Furthermore, this invention can dynamically optimize probe position and pressure parameters based on differences in breast morphology and location among patients through real-time image analysis and feedback adjustment, achieving personalized automatic adjustment. This enhances the system's adaptability to different individuals and helps obtain ultrasound images that better reflect individual differences. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0047] Figure 1 This is a schematic flowchart of an automatic ultrasound-assisted breast compression method based on image feedback disclosed in this invention.

[0048] Figure 2 This is a schematic diagram of the lightweight real-time target detection network for Mamba breast ultrasound region based on YOLO disclosed in this invention.

[0049] Figure 3 This is a schematic diagram of the network structure of the integrated multi-channel visual spatial block disclosed in this invention;

[0050] Figure 4 This is a schematic diagram of the network structure of dynamic scale channel convolution disclosed in this invention;

[0051] Figure 5 This is a schematic diagram of the scanning method of the diversified channel scanning mechanism disclosed in this invention;

[0052] Figure 6 This diagram illustrates a comparison between the method disclosed in this invention and the results of manual testing by a doctor.

[0053] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0056] See Figures 1-5 The present invention provides an image feedback-based automatic breast ultrasound pressurization method that relies on an automatic breast ultrasound pressurization device integrating a robotic arm, an ultrasound probe, an image acquisition module, a target detection network, and a processing control unit. Specifically:

[0057] The robotic arm is used to carry and precisely control the position and orientation of the ultrasound probe, and can move and adjust pressure in multiple degrees of freedom according to control commands.

[0058] An ultrasound probe is used to emit ultrasound waves and receive echo signals to generate a cross-sectional ultrasound image of the breast.

[0059] The image acquisition module is used to acquire ultrasound image data collected by the ultrasound probe in real time;

[0060] The processing control unit is configured to run a pre-trained target detection network, identify breast regions, calculate breast region offsets, generate feedback control signals, and control the robotic arm to perform corresponding position and / or pressure adjustments; wherein, the processing control unit is communicatively connected to the robotic arm and the image acquisition module to form a closed-loop control system.

[0061] Specifically, the present invention provides an automatic breast ultrasound compression method based on image feedback, the specific implementation steps of which are as follows:

[0062] S1. Initial compression and image acquisition:

[0063] After the device is activated, the robotic arm is first moved to the approximate location of the breast area to be examined, and the ultrasound probe is gently pressed against the patient's breast surface with a preset initial pressure (e.g., a moderate contact pressure set based on clinical experience or system defaults). This initial pressure is designed to ensure good acoustic coupling between the probe and the skin, while avoiding excessive pressure that could cause discomfort or image artifacts.

[0064] While maintaining the initial pressure, the ultrasound probe performs a cross-sectional scan of the breast area, and the image acquisition module acquires the corresponding ultrasound images in real time, providing basic data for subsequent analysis.

[0065] S2. Breast region identification and offset calculation:

[0066] The acquired ultrasound images are input into a pre-trained target detection network. Preferably, the target detection network is based on the YOLO architecture and incorporates an improved Mamba structure, which can accurately identify and locate the breast region in the ultrasound image while ensuring real-time performance.

[0067] The target detection network outputs information such as the bounding box or center coordinates of the breast region in the image. The processing and control unit uses this information to calculate the offset of the center point of the breast region relative to the center point of the entire image. This offset reflects the degree of deviation of the current position of the breast region in the image and is an important basis for subsequent adjustments.

[0068] S3. Feedback Control and Dynamic Adjustment:

[0069] Based on the offset calculated in step S2, the processing control unit generates a corresponding feedback control signal. This signal is used to guide the robotic arm to perform fine adjustments, and may specifically include:

[0070] Adjusting the position of the robotic arm allows the ultrasound probe to move slightly in the horizontal or vertical direction, thereby causing the breast area to move closer to the center in the image;

[0071] Or / and adjust the pressure applied to the surface of the breast by the robotic arm to change the contact state between the probe and the breast, further assisting in the position adjustment of the breast area.

[0072] This adjustment process is dynamic and real-time, meaning that the control unit generates control commands in real time based on the current offset, driving the robotic arm to perform corresponding actions to gradually reduce the offset.

[0073] S4. Iterative optimization and image output:

[0074] The device will repeat steps S1 to S3, that is, after each adjustment, it will reacquire ultrasound images, identify the breast region again through the target detection network and calculate the new offset, and then continue to perform feedback adjustment of position and pressure.

[0075] This iterative process continues until the detected breast region offset is less than a preset threshold. The threshold can be set according to actual application requirements, such as within a certain pixel range near the image center, indicating that the breast region is now basically located in the image center and the image quality meets the expected standard.

[0076] Ultimately, the device outputs a breast ultrasound cross-sectional image that meets the current quality requirements, which can be used for subsequent clinical diagnosis or image analysis.

[0077] This invention effectively solves the problems of image offset, inaccurate pressure control, and feedback delay in traditional breast ultrasound imaging by synergistically combining robotic arm control, real-time image analysis, and closed-loop feedback adjustment, thereby improving the quality and clinical applicability of automated breast ultrasound image acquisition. Specifically, on the one hand, by employing a lightweight real-time target detection network based on YOLO and an improved Mamba structure, the breast region in breast ultrasound images can be quickly and accurately identified. Compared to traditional methods that rely on edge detection or grayscale analysis, this improves the target detection accuracy in complex ultrasound image backgrounds, ensuring that the breast region is accurately bounded and providing a reliable basis for subsequent position adjustment. On the other hand, by analyzing the real-time acquired ultrasound images, the offset of the breast region relative to the image center is calculated, and a feedback control signal is generated based on this offset to dynamically adjust the position of the robotic arm and the probe pressure, so that the breast region is gradually aligned with the image center. This closed-loop control mechanism effectively solves the image offset problem caused by manual operation or static settings in traditional methods, improving the spatial consistency of the image.

[0078] This invention achieves dynamic adjustment of probe pressure through real-time image feedback and robotic arm linkage, avoiding the problems of inaccurate pressure control and tissue deformation or image distortion caused by traditional automatic pressurization systems. It ensures appropriate and stable contact pressure is maintained for different patients or at different scanning stages, improving the controllability of image quality. By introducing an image feedback mechanism and automated adjustment process, it replaces the traditional manual adjustment method relying on physician experience, reducing image quality fluctuations caused by operator fatigue and subjective judgment differences. This makes the ultrasound image acquisition process more standardized and repeatable, contributing to improved objectivity and efficiency in clinical examinations. Furthermore, this invention can dynamically optimize probe position and pressure parameters based on the differences in breast morphology and location among different patients through real-time image analysis and feedback adjustment, achieving personalized automatic adjustment. This enhances the system's adaptability to different individuals and helps obtain ultrasound images that better reflect individual differences.

[0079] As a preferred embodiment, see Figure 2 The target detection network for breast region identification employs a lightweight real-time target detection network based on the YOLO architecture and incorporating the Mamba structure, hereinafter referred to as "YOLO-based Mamba lightweight real-time target detection network for breast ultrasound regions". This network architecture balances detection accuracy and computational efficiency, making it suitable for medical image scenarios such as breast ultrasound images, which are characterized by complex textures and low contrast. It comprises three functional modules: the LRMBu backbone network, the neck network, and the segmentation head. The specific composition and function of each module are described in detail below.

[0080] The LRMBu (Lightweight Multi-scale Feature Extraction Backbone Unit) backbone network is the front-end feature extraction part of the target detection network, used to efficiently extract discriminative multi-scale feature information from the input breast ultrasound image.

[0081] The backbone network is composed of the following sub-modules connected in sequence:

[0082] The initial convolutional block, located at the network's input, is used for preliminary feature extraction and dimensionality reduction of the original ultrasound image. It consists of one or more standard convolutional layers, supplemented by batch normalization and activation functions. Its function is to perform preliminary feature mapping on the input image, laying the foundation for subsequent deep feature extraction.

[0083] Depth blocks, composed of multiple depthwise separable convolutional layers or similar lightweight convolutional structures, are used to further extract local and detailed features from images without significantly increasing computational cost. This design addresses the noise and weak boundary issues commonly found in ultrasound images, helping to enhance the network's ability to perceive the edges and shapes of breast regions.

[0084] The Integrated Multichannel Visual Space (IMCVSS) block is one of the key innovations of this backbone network, used to fuse and enhance information from different levels of features in multiple channels and spatial dimensions. This module, through the design of specific multichannel interaction mechanisms and spatial attention strategies, effectively improves the network's ability to represent targets such as the breast region in complex backgrounds, enhancing the expressiveness and robustness of features.

[0085] The LRMBu backbone network can extract multi-scale feature information from the input ultrasound image layer by layer, from low to high level and from local to global, providing rich and discriminative feature maps for subsequent feature fusion and target localization.

[0086] The neck network is a key component for achieving multi-scale feature fusion. It introduces structures such as SPPF (Spatial Pyramid Pooling Fast), upsampling, downsampling, channel concatenation, and CSPlayer (Cross Stage Partial Layer) to efficiently fuse the multi-scale features output by the LRMBu backbone network. This enhances the network's ability to express features of the breast region at different scales and locations, thereby improving the accuracy of target detection.

[0087] The neck network, located between the backbone network and the segmentation head, primarily functions to fuse multi-scale feature maps output by the LRMBu backbone network. Since the target region (i.e., breast tissue) in breast ultrasound images may be at different scales and locations, features at a single scale are often insufficient to fully represent its morphology and boundaries. Therefore, it is necessary to effectively fuse features from different levels of the backbone network to enhance the network's overall perception of the target.

[0088] The neck network employs a multi-layered, multi-path feature fusion strategy, which includes the following key components and processing steps:

[0089] The input to the neck network is multi-scale feature maps, which are derived from the multi-level feature maps output by the LRMBu backbone network. These feature maps typically correspond to different receptive fields of the image, including a variety of information from local details to overall semantics, and the resolution (i.e., feature map size) and number of channels of each layer of feature maps may vary.

[0090] The SPPF (Spatial Pyramid Pooling Fast) module is applied to a portion of the input feature maps (typically deep features with high semantics but low resolution) to perform multi-scale pooling on the feature maps without significantly increasing computational cost, capturing contextual information under different receptive fields. This module aggregates local regions of the feature map by concatenating multiple fixed-size pooling windows (such as 3×3, 5×5, etc.; in practice, continuous convolution is often used to simulate fast pooling), thereby enhancing the network's ability to perceive the overall structure of the target and its background environment. The introduction of SPPF helps improve the model's robustness to breast regions under different ultrasound imaging conditions.

[0091] To achieve the fusion of feature maps at different resolutions, the neck network employs upsampling and downsampling operations:

[0092] Upsampling involves using interpolation methods (such as nearest neighbor interpolation or bilinear interpolation) to enlarge low-resolution feature maps to higher resolution, so that they can be aligned and stitched with high-resolution features in shallow layers.

[0093] Downsampling typically employs stride convolution or pooling operations to reduce high-resolution feature maps to a lower resolution for matching and fusion with deeper low-resolution features.

[0094] These operations enable feature maps from different levels to be aligned in spatial dimensions, creating conditions for subsequent feature stitching and fusion.

[0095] After upsampling or downsampling and aligning the feature map sizes, the neck network concatenates feature maps from different layers along the channel dimension. This operation effectively integrates feature information from different receptive fields and resolutions, forming a composite feature representation containing richer semantic and location information. Through channel concatenation, the network can simultaneously utilize the detailed information of shallow features and the semantic information of deep features, thereby improving the accuracy of localizing the boundaries of the breast region.

[0096] The concatenated multi-scale fused features are further processed by CSPlayer (Cross Stage Partial Layer). CSPlayer is a lightweight feature processing module that enhances feature expressiveness and gradient fluidity while reducing computational redundancy, thus improving network training stability and detection performance. In the neck network, CSPlayer is used to perform further nonlinear transformations and channel adjustments on the concatenated multi-scale features, optimizing feature distribution and improving the regression and classification capabilities of subsequent detection heads for the target region (breast region).

[0097] The overall processing flow of the neck network can be summarized as follows:

[0098] The multi-scale features output from the LRMBu backbone network are input into the neck network. SPPF is applied to some features to enhance contextual modeling capabilities. Features at different resolutions are upsampled or downsampled as needed to achieve size alignment. The aligned features are then concatenated along the channel dimension. Finally, CSPlayer is used to further optimize and adjust the fused features, outputting the fused multi-scale comprehensive features. The fused features retain the location and shape information of the breast region at different scales and incorporate multi-level semantic information, providing a more comprehensive and robust input for subsequent segmentation or detection heads. This helps to accurately predict the location of the breast region and output corresponding target detection box information.

[0099] The segmentation head, the output module of the target detection network, predicts and outputs target detection bounding boxes for the breast region based on the multi-scale features fused from the neck network. In this invention, the segmentation head primarily outputs the location information of the breast region in the image, specifically represented by one or more bounding boxes, used to indicate the approximate extent of the breast region in the ultrasound image. The segmentation head typically includes several convolutional layers and upsampling layers, used to map the fused features back to the original image resolution and predict the center coordinates, width, height, or other relevant parameters of the target box through regression. In some embodiments, the segmentation head may also simultaneously output the confidence score of the target region to facilitate subsequent filtering out low-quality or falsely detected candidate regions. In actual deployment, the output format and accuracy of the segmentation head can be appropriately adjusted according to system resources and real-time requirements, for example, retaining only the bounding box regression branch to further simplify the network structure and improve inference speed.

[0100] In actual operation, the input breast ultrasound image first enters the LRMBu backbone network, and after processing through initial convolutional blocks, depth blocks, and IMCVS blocks, multi-scale feature information is extracted. Subsequently, these features are passed to the neck network for multi-scale fusion to form a comprehensive feature representation rich in semantic and location information. Finally, based on the fused features, the segmentation head predicts and outputs the target detection box of the breast region, providing accurate regional positioning information for subsequent image offset calculation and robotic arm feedback control.

[0101] As a preferred embodiment, the Integrated Multi-channel Visual Spatial Block (IMCVSS) aims to enhance the network's ability to model long-range spatial dependencies in breast ultrasound images while controlling computational complexity to meet lightweight deployment requirements. This module is built on the Mamba architecture, avoiding the excessive computational burden of traditional self-attention mechanisms, while effectively capturing global contextual information in the image.

[0102] The integrated multi-channel visual spatial block mainly consists of two core sub-modules: the Dynamic Scale Channel Convolution Module (DSCC) and the Diverse Channel Scanning Mechanism (DCSM). The specific structure and operation process of these two modules are described in detail below.

[0103] like Figure 3 As shown, the input to the integrated multi-channel visual spatial block is a feature map X (with dimensions C×H×W, representing the number of channels, height, and width, respectively) from the preceding layers of the backbone network. This feature map is first processed by the Dynamic Scale Channel Convolutional (DSCC) module to obtain an intermediate feature map Y. Subsequently, feature map Y is input to the Diverse Channel Scanning (DCSM) mechanism for multi-path, multi-sequence channel scanning and fusion to further model long-range spatial dependencies. Finally, after a series of processing steps (including residual connections, convolutional blocks, batch normalization, and SiLU activation function), the enhanced feature map is output.

[0104] X′ serves as the input to subsequent network layers.

[0105] This module, by integrating local convolution and a lightweight global scanning mechanism, effectively improves the network's ability to represent features of the breast region in complex ultrasound image backgrounds while maintaining low computational overhead.

[0106] The Dynamic Scale Channel Convolution (DSCC) module is a preprocessing unit that integrates multi-channel visual spatial blocks. Its main functions are to perform channel grouping, adaptive channel calibration, and cross-group feature fusion on the input feature map to enhance the information interaction between different channels, optimize feature representation capabilities, and reduce computational complexity.

[0107] Module input and channel grouping:

[0108] The input to DSCC is a feature map X. This feature map is first divided evenly into four groups along the channel dimension, each group containing different but complementary feature information. This grouping strategy helps to differentiate the features and enhances the model's attention to different feature dimensions.

[0109] Adaptive Channel Feature Recalibration (ACFR):

[0110] Each set of features then enters an Adaptive Channel Feature Recalibrator (ACFR). This component uses an adaptive weighting mechanism to recalibrate and optimize the channel features within each set, highlighting the information contribution of important channels and suppressing the influence of redundant or noisy channels. This process includes reshaping the channel dimensions and weighting calculations, enabling dynamic adjustment of each channel feature.

[0111] Cross-group feature fusion and dimensionality recovery:

[0112] After calibration, the features are recombined and stitched together in different arrangements to form a composite feature map containing multiple sets of interactive information. Then, a 1×1 convolutional layer and batch normalization are used to restore the stitched feature map to its original channel dimension, resulting in the enhanced output feature map Y.

[0113] In addition, DSCC also includes a short connection path: the input feature map is processed by a 1×1 convolution and batch normalization, then added to the output of the main path, and further fused through the ReLU activation function. This enhances feature representation while retaining key information of the original input, thereby improving the training stability and feature reuse efficiency of the network.

[0114] The Diverse Channel Scanning Mechanism (DCSM) is a core component integrating multi-channel visual spatial blocks. Its main goal is to perform multi-path, multi-sequence scanning of feature maps to fuse long-distance spatial dependencies while maintaining the consistency of feature map dimensions. This module enhances the model's ability to perceive long-distance positional information in images by simulating sequential scanning, and is suitable for situations where the target region in ultrasound images is not fixed in location and has a variable shape.

[0115] Input preprocessing: The input to DCSM is the intermediate feature map Y (dimension C×H×W) output by DSCC. First, the input feature f[C,H,W] is passed through a 1×1 convolutional layer to reduce the number of channels from C to C / 4, resulting in feature map f1 (dimension C / 4×H×W), thus reducing the computational burden of subsequent processing.

[0116] Multi-channel scanning and multi-path processing: In the multi-channel scanning stage, the feature map f1 is copied multiple times and processed along multiple different scanning paths using various scanning modes. Each path simulates a different method of spatial information extraction, thereby covering more positional relationships.

[0117] The multi-path results after scanning are stitched together into a new feature map f2 (the dimensions are restored to C×H×W), which integrates contextual information from different scanning paths while maintaining the original spatial resolution and number of channels.

[0118] Channel expansion and summation fusion: To further enhance feature representation, feature map f2 is then passed through a 1×1 convolutional layer to expand the number of channels to 4C, resulting in feature map f3 (dimension 4C×H×W). Next, the features are summed according to the original scanning order to reintegrate the multi-channel information, and finally the fused feature map f′ (dimension restored to C×H×W) is output.

[0119] Residual Connections and Subsequent Processing: Finally, f′ and the original input Y are fused through residual connections to preserve the basic information in the original features. The fused features are then processed through a ConvBlock (containing convolution, batch normalization, and SiLU activation function) for nonlinear transformation and feature enhancement, resulting in the final output X′ of the integrated multi-channel visual spatial block.

[0120] The integrated multi-channel visual spatial block in this invention achieves inter-channel grouping and adaptive calibration through a Dynamic Scale Channel Convolution (DSCC) module, enhancing feature expressiveness and interactivity. A Diverse Channel Scanning (DCSM) mechanism introduces multi-path, multi-sequence long-distance spatial dependency scanning and fusion, effectively compensating for the shortcomings of convolution operations in global modeling. This module, built on the Mamba architecture, avoids the high computational complexity of traditional self-attention mechanisms, making it suitable for lightweight deployment in high-resolution, low-contrast medical image scenarios such as breast ultrasound images. This provides a crucial feature foundation for subsequent precise breast region localization and image quality improvement.

[0121] In a preferred embodiment, the offset calculation is primarily based on the positional information of the breast region detection box output by the target detection network, particularly the relative positional relationship between the center coordinates of the breast region detection box and the geometric center coordinates of the breast ultrasound image itself. By comparing the difference between the two, the degree of offset of the breast region in the image can be quantified, providing a basis for subsequent feedback control.

[0122] Specifically, assuming the width of the acquired breast ultrasound image is W and the height is H, then the coordinates of the geometric center of the image are:

[0123] The horizontal coordinate of the image center is W / 2; the vertical coordinate of the image center is H / 2.

[0124] After processing the ultrasound image, the target detection network outputs a detection box representing the breast region. This detection box is typically a rectangle, used to define the approximate extent of the breast region within the image. This detection box has its own geometric center, denoted as:

[0125] The horizontal coordinate of the center of the detection frame is x center The vertical coordinate of the center of the detection frame is y. center ;

[0126] The coordinates above all have the top left corner of the image as the origin (0,0), with the positive x-axis pointing horizontally to the right and the positive y-axis pointing vertically downwards.

[0127] Based on the difference between the center coordinates of the detection box and the center coordinates of the image, the offset of the breast region in the horizontal and vertical directions is calculated, i.e., the horizontal offset X. 偏移 and vertical offset Y 偏移 The specific calculation formula is as follows:

[0128] ;

[0129] ;

[0130] Calculated X 偏移 This represents the number of pixels the breast region is offset from the image center in the horizontal direction: a positive value indicates the detection box is located to the right of the image center; a negative value indicates it is located to the left. Similarly, Y... 偏移 The value indicates the offset in the vertical direction: a positive value indicates that the center of the detection box is below the center of the image, and a negative value indicates that it is above the center.

[0131] To comprehensively evaluate the overall offset of the breast region in the image, this invention further calculates a comprehensive offset error Δd based on the aforementioned horizontal and vertical offsets, which characterizes the overall degree to which the breast region deviates from the image center. Specifically, the formula for calculating the comprehensive offset error Δd is as follows:

[0132] ;

[0133] The above formula essentially calculates the Euclidean distance between the center point of the detection box and the center point of the image. Its value directly reflects the degree of displacement of the breast region in the image: a larger Δd value indicates a more significant displacement; a smaller Δd value indicates that the breast region is closer to the image center. In practical applications, a preset threshold Δd is set. thre (For example, a pixel distance threshold determined based on image resolution and clinical needs), when the calculated Δd is greater than Δd thre If the position of the ultrasound probe is not adjusted, it indicates that the breast area is still off-center from the image, and the probe needs to be repositioned until the offset Δd is less than Δd. thre If the image is in the ideal position and meets the image quality requirements, the adjustment process can be ended and the current ultrasound image can be output.

[0134] This invention calculates the differences in horizontal and vertical coordinates between the center of the target detection box and the center of the image to obtain X. 偏移 With Y 偏移 This is further aggregated into a Δd offset error, used to quantitatively describe the positional deviation of the breast region in the ultrasound image. This offset calculation method is simple and intuitive, facilitating subsequent control modules to generate feedback signals and drive the robotic arm to perform precise position and / or pressure adjustments, thereby achieving image centering and quality optimization of the breast region.

[0135] In a preferred embodiment, the trigger for dynamic adjustment is the comprehensive offset error Δd calculated in step S2. When the offset error Δd is greater than a preset threshold Δd... thre If the image shows a significant deviation of the breast region from the image center, it does not meet the image quality requirements. In this case, a feedback control mechanism needs to be activated to make appropriate adjustments to the robotic arm.

[0136] Conversely, if Δd is less than or equal to the preset threshold Δd thre If the image is considered to be centered and meets the quality standards, the system can stop adjusting and output the currently acquired ultrasound image.

[0137] If the adjustment is triggered, the horizontal offset X calculated in step S2 will be used. 偏移 With vertical offset Y 偏移 The positive and negative values ​​are used to determine the offset direction of the breast region in the image, and the robotic arm is then used to adjust the position or pressure of the probe along the corresponding direction.

[0138] X 偏移 It reflects the offset of the breast region in the horizontal direction (i.e., the left-right direction) of the image;

[0139] If X 偏移 A value >0 indicates that the center of the detection frame is biased to the right side of the image. In this case, the robotic arm needs to be controlled to move the probe to the left so that the breast area moves closer to the center of the image.

[0140] If X 偏移 A value less than 0 indicates that the center of the detection frame is biased to the left side of the image. In this case, the robotic arm needs to be controlled to move the probe to the right.

[0141] Y 偏移 It reflects the offset of the breast region in the vertical direction (i.e., the up-down direction) of the image;

[0142] If Y 偏移 A value >0 indicates that the center of the detection frame is biased towards the bottom of the image. In this case, the robotic arm can be controlled to reduce the downward pressure or the probe position can be finely adjusted upwards to adjust the position of the breast area.

[0143] If Y 偏移 A value <0 indicates that the center of the detection frame is biased towards the top of the image. In this case, the robotic arm can be controlled to increase the downward pressure or the probe position can be finely adjusted downwards to move the breast area down to the vicinity of the image center.

[0144] It should be noted that in the actual system, horizontal adjustment is mainly achieved by controlling the probe to move horizontally using a robotic arm; vertical adjustment can be achieved by controlling the probe to move vertically using a robotic arm, by adjusting the contact pressure of the probe on the breast, or by a combination of both. The specific implementation method can be flexibly configured according to the mechanical structure design and clinical needs.

[0145] To control the adjustment range and response speed, the probe position or pressure adjustment amount in this invention is not directly equal to the offset amount, but is calculated through a linear mapping relationship to ensure a smooth and controllable adjustment process. Specifically, the adjustment amount (i.e., the movement amount) in the horizontal and vertical directions are determined by the following formulas:

[0146] Horizontal movement Move(X): Move(X) = K × X 偏移 ;

[0147] Horizontal movement Move(Y): Move(Y) = K × Y 偏移 ;

[0148] K is an adjustable proportional coefficient, which is a preset empirical parameter or an adjustable parameter used to control the magnitude of the adjustment, so that the adjustment process is neither too sensitive nor too slow.

[0149] Move(X) and Move(Y) represent the amount of adjustment that the robotic arm should perform in the horizontal and vertical directions, respectively. The units correspond to the offset, such as the physical displacement or pressure change corresponding to a pixel.

[0150] The value of the scaling factor K can be adjusted according to factors such as the actual application scenario, the control precision of the robotic arm, the movement range of the ultrasound probe, and the clinical image quality requirements. By adjusting the K value, the system's response speed and adjustment stability can be balanced, avoiding overshoot or oscillation caused by excessive adjustment in a single step, and ensuring that the breast area moves smoothly and accurately toward the image center.

[0151] After each adjustment, the process returns to step S1 to reacquire ultrasound images and continues to step S2 to calculate the new offset and the overall offset error Δd. If Δd is still greater than the preset threshold, the process continues based on the new X... 偏移 With Y 偏移 Proceed to the next round of adjustments; if Δd is less than or equal to the preset threshold, it is determined that the current breast region location meets the image quality requirements, the adjustment process ends, and the current standard-compliant ultrasound image is output.

[0152] The dynamic adjustment process achieves precise adjustment of probe position and pressure through multiple small-amplitude, high-frequency feedback controls, effectively ensuring the centered display of the breast region in the ultrasound image and the consistency of image quality.

[0153] As a preferred embodiment, the present invention further includes a high-quality image selection step after step S4, which is used to select the best frame from several frames of images acquired by multiple scans as the final breast ultrasound image output.

[0154] In actual ultrasound scanning, to ensure accurate centering of the breast region and satisfactory imaging results, the system typically performs multiple scans and feedback adjustments, repeating steps S1 to S4 several times to gradually optimize the probe position and pressure, while simultaneously acquiring multiple frames of breast ultrasound images. These images are collectively referred to as the N frames acquired through multiple scans, where N is an integer greater than or equal to 1, representing the total number of breast ultrasound images acquired during the entire automatic pressurization and feedback control process. Each frame contains detection box information for the breast region, which is identified and output in real time by the target detection network in step S2, used to define the position and extent of the breast region in the image.

[0155] In the high-quality image selection step, the system selects the frame with the largest target detection box area from the aforementioned N frames, based on the area of ​​the target detection box for the breast region in each frame. This frame is then used as the final high-quality breast ultrasound image output. Specifically, for each frame, the system acquires the breast region detection box output by the target detection network. The breast region detection box is typically a rectangle, and its size is determined by the width w and height h of the detection box. The area of ​​the detection box is: Area = w × h. The system iterates through all N frames, calculates the area of ​​the breast region detection box in each frame, and records the frame with the largest area. This frame typically indicates that at that moment, the breast region occupies a wider imaging area in the image, the target region is displayed more completely, and the image content is richer, which is beneficial for subsequent clinical observation and diagnosis.

[0156] The specific steps for selecting high-quality images are as follows:

[0157] Image and detection box information storage: After each step S4 is completed and an image that meets the offset requirements is output, the system stores the image frame and its corresponding target detection box parameters (including the top left corner coordinates, width w, and height h of the detection box, etc.) to form a set of N frames of images and their detection box data.

[0158] Area Calculation: For each frame of an image, the system calculates the area of ​​the detection box for the breast region in that frame based on the width w and height h of the detection box, i.e., S. i =w i ×h i (i=1,2,...,N);

[0159] Among them, S i w represents the area of ​​the detection box in the i-th frame image. i and h i These are the width and height of the detection box in that frame, respectively.

[0160] Maximum area filtering: The system iterates through the detection box area S of all N frames of images. i Find the frame with the largest area, i.e., the one that satisfies: S max =max(S1,S2,...,S N );

[0161] The corresponding image is the one with the largest target detection bounding box area, and it is marked as the final output image.

[0162] Output high-quality images: The selected frame image will be used as the final output of this breast ultrasound automatic pressure process for subsequent display, storage or diagnostic analysis.

[0163] By selecting the image with the largest target detection box area as the final output, this invention can further improve the practicality and image quality of the output image based on multiple feedback adjustments and image acquisitions. Generally speaking, a larger detection box area means that the breast region is displayed more completely in the image, with better centering and richer image details, which helps doctors to observe the morphology and structure of breast tissue more accurately and improves the reliability of detection. This selection method is simple to implement, computationally inexpensive, and does not require the introduction of complex image quality evaluation algorithms. It can effectively filter out images with better visual effects and coverage from multiple scan results, making it suitable for widespread application in clinical ultrasound examinations.

[0164] To further illustrate the effectiveness of the automated ultrasound compression method for breast cancer disclosed in this invention, see [link to relevant documentation]. Figure 6 , Figure 6 This diagram illustrates a comparison between cross-sectional breast ultrasound images obtained using the method provided in this invention and cross-sectional breast ultrasound images obtained manually by a physician. The breast region was detected and calibrated using the YOLOv12 model. The results show that the images obtained using the method provided in this invention are not only clearer, but also have the breast region precisely centered in the image, resulting in superior display quality compared to traditional manual detection methods.

[0165] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct or indirect applications in other related technical fields, are within the patent protection scope of the present invention.

Claims

1. A method for automatic ultrasound-guided breast compression based on image feedback, characterized in that, The steps and methods include: S1. The robotic arm carries and controls the ultrasound probe to press down on the breast area with a preset initial pressure, and acquires a cross-sectional ultrasound image of the breast. S2. The cross-sectional ultrasound image is analyzed in real time using a pre-trained target detection network to identify the location of the breast region and calculate the offset of the region relative to the image center. The target detection network is a lightweight real-time target detection network for breast ultrasound regions based on YOLO, which includes an LRMBu backbone network, a neck network, and a segmentation head. The LRMBu backbone network includes an initial convolutional block, a depth block, and an integrated multi-channel visual spatial block connected in sequence, used to extract multi-scale features from the input image; The neck network is used to fuse the multi-scale features; The segmentation head is used to output target detection box information for the breast region; The integrated multi-channel visual spatial block includes a dynamic-scale channel convolution module and a diversified channel scanning mechanism; The dynamic scale channel convolution module is used to group input features by channel and fuse cross-group features; the diversified channel scanning mechanism is used to perform multi-path scanning of feature maps to fuse long-distance spatial dependencies; The diversified channel scanning mechanism includes multi-channel scanning and multi-channel fusion, and the specific operations are as follows: The input feature f[C,H,W] is first processed by a 1×1 convolution to reduce the channel dimension to C / 4, resulting in f1[C / 4,H,W]. During the multi-channel scanning stage, f1[C / 4,H,W] is copied into multiple copies, and various scanning modes are executed along multiple different scanning paths. The results are then concatenated into f2[C,H,W] to maintain the channel dimension while integrating multiple scanning information. Then, the number of channels is expanded to 4C by 1×1 convolution, resulting in f3[4C,H,W]; Finally, the feature results are summed according to the scanning order to obtain the final output feature f'[C,H,W]; S3. Generate a feedback control signal based on the offset, dynamically adjust the robotic arm to drive the ultrasound probe to move and / or change the applied pressure, so that the breast region is aligned with the image center; S4. Repeat steps S1 to S3 until the offset is less than the preset threshold, and output an ultrasound image that meets the quality standard.

2. The method for automatic breast ultrasound compression based on image feedback according to claim 1, characterized in that, The operations of the dynamic-scale channel convolution module include: The input feature map is divided into multiple groups along the channel dimension; adaptive channel feature recalibration is performed on each group of features. The calibrated feature sets are concatenated, and the channel dimensions are restored through convolution operations to obtain the enhanced output features.

3. A method for automatic breast ultrasound compression based on image feedback according to claim 1 or 2, characterized in that, In step S2, the specific method for calculating the offset is as follows: Based on the center coordinates (x) of the target detection box center ,y center Calculate the horizontal offset X based on the difference between the coordinates of the image center (W / 2, H / 2) and the coordinates of the image center. 偏移 and vertical offset Y 偏移 The specific calculation formula is as follows: ； ； Where W is the width of the ultrasound image and H is the height of the ultrasound image; The combined offset error Δd is calculated based on the horizontal and vertical offsets; specifically, the formula for calculating the combined offset error Δd is as follows: 。 4. The method for automatic breast ultrasound compression based on image feedback according to claim 3, characterized in that, In step S3, the dynamic adjustment specifically includes: If Δd is greater than the preset threshold Then according to X 偏移 The positive and negative controls move the probe of the robotic arm horizontally, according to... Y 偏移 The positive and negative controls allow the robotic arm to adjust the probe pressure in the vertical direction; The moves Move(X) and Move(Y) are respectively given by the formula Move(X) = K × X 偏移 and Move(Y) = K × Y 偏移 Determined, where K is the adjustable proportional coefficient.

5. A method for automatic breast ultrasound compression based on image feedback according to claim 1 or 2, characterized in that, After step S4, a high-quality image selection step is also included: from the N frames of images acquired through multiple scans, the image with the largest target detection box area is selected as the final high-quality breast ultrasound image output.

6. A breast ultrasound automatic pressurization device according to any one of claims 1-5, characterized in that, include: The robotic arm is used to carry and precisely control the position and orientation of the ultrasound probe, and can move and adjust pressure in multiple degrees of freedom according to control commands. An ultrasound probe is used to emit ultrasound waves and receive echo signals to generate a cross-sectional ultrasound image of the breast. The image acquisition module is used to acquire cross-sectional ultrasound images of the breast region; The processing control unit is configured to run a pre-trained target detection network, identify breast regions, calculate breast region offsets, and generate feedback signals for controlling the robotic arm; wherein the processing control unit is communicatively connected to the robotic arm and the image acquisition module to form a closed-loop control system.

7. The automatic breast ultrasound pressurization device according to claim 6, characterized in that, The target detection network is trained on a dataset including cross-sectional images of breast tissue, which is augmented with data, including one or more operations such as rotation, translation and scaling. The device also includes a display module for displaying ultrasound images, breast region detection frames and offset information in real time, and providing image quality compliance prompts.

Images