A method for scan depth assessment of a breast ABUS image

Abstract

This invention discloses a method for evaluating the scanning depth of breast ABUS images, belonging to the field of breast ABUS image technology. The method includes: S1: collecting three-dimensional DICOM data of breast ultrasound, annotating the ribs in the coronal images, and evaluating the scanning depth; S2: performing coronal slice processing and converting and generating model labels; S3: detecting the rib region based on a target detection architecture, and training a rib detection model capable of automatically locating the rib region in the image; S4: testing and evaluating the rib detection model; S5: establishing scanning depth judgment rules based on the number and position of rib frames to determine the scanning depth category. This invention solves the problems of significant differences in image coverage, lack of reliable reference, and difficulty in quickly determining the optimal scanning depth suitable for different breast sizes and shapes. This invention's automatic evaluation of ABUS image scanning depth enables standardization of ultrasound screening.

Description

Technical Field

[0001] This invention relates to the field of breast ABUS imaging technology, specifically to a method for evaluating the scanning depth of breast ABUS images. Background Technology

[0002] Breast cancer is the most common malignant tumor in women, and its early and accurate diagnosis is crucial. Automated breast volume ultrasound imaging (ABUS), as an advanced three-dimensional ultrasound technology, can generate transverse, sagittal, and coronal images of the entire breast through automatic scanning, greatly facilitating the screening and diagnosis of breast diseases. The quality of ABUS images is the foundation of diagnostic accuracy, and the appropriateness of the scan depth setting, as a core acquisition parameter, directly determines the completeness of the image and its diagnostic value. The scan depth parameter determines the extent to which the ultrasound beam penetrates the tissue and the displayed field of view. Insufficient depth setting will fail to fully cover the entire breast glandular tissue from the nipple to the chest wall, resulting in the absence of deep tissue in the image and causing information omission. Conversely, excessive depth setting, while ensuring complete tissue coverage, will reduce the overall resolution of the image, and the breast tissue will appear too small, increasing the difficulty of identifying subtle lesions. Inappropriate depth settings will result in images that do not meet quality control standards, requiring patients to be recalled for rescanning, which increases medical costs and reduces screening efficiency.

[0003] According to the 2022 expert consensus, a standard ABUS image should completely encompass the entire breast region from the nipple to the chest wall, clearly displaying both skin and glandular structures without artifact interference. In practice, physicians need to make real-time judgments on parameters such as scan depth to ensure complete imaging of glandular tissue at an appropriate depth, avoiding missing tissue information or decreased image resolution due to improper depth settings. Currently, ABUS scan depth settings heavily rely on the operator's immediate judgment and personal experience, lacking objective and unified evaluation standards. Although some studies have evaluated the quality of handheld ultrasound images, an objective quality control system specifically for ABUS scan depth is still lacking, exhibiting the following deficiencies:

[0004] 1) Different physicians have different standards for judging the appropriate depth, resulting in significant differences in the coverage of the image; 2) Manual assessment makes it difficult to accurately determine whether the image truly and completely includes all the necessary tissues from the superficial skin to the deep chest wall, lacking reliable reference; 3) Inexperienced operators have difficulty quickly determining the optimal scanning depth for different breast sizes and shapes, affecting the workflow. Summary of the Invention

[0005] The purpose of this invention is to provide a method for evaluating the scanning depth of breast ABUS images. By performing real-time and accurate detection of the rib region in the ABUS image, combined with a specially designed post-processing analysis module, it can automatically evaluate whether the scanning depth of the breast ultrasound image is qualified. This solves the problem of relying on experience-based adjustment of scanning depth parameters in clinical practice, and provides physicians with objective and quantitative operational guidance and suggestions, thus addressing the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for assessing the scanning depth of breast ABUS images, comprising:

[0008] S1: Collect multi-center, multi-device, and multi-parameter three-dimensional DICOM data of breast ultrasound, and annotate the ribs and assess the scanning depth of the coronal images;

[0009] S2: Perform coronal section processing and convert and generate model labels based on the makesense annotation platform;

[0010] S3: Based on the target detection architecture, rib region detection is performed. A complete process is constructed, including a backbone network, a neck network, detection output, and an error optimization module. A rib detection model that can automatically locate the rib region in the image is trained.

[0011] S4: Perform model testing and performance evaluation on the rib detection model, and output a visualized prediction result map;

[0012] S5: Establish scanning depth judgment rules based on rib frame number and position, extract coronal image sequence, detect rib region in each frame, obtain detection result sequence, identify the first consecutive detected rib frame segment and starting frame number, analyze the position of the 10 consecutive frames in the image where the starting frame number of the rib appears, and determine the scanning depth category.

[0013] Preferably, in step S1, the ribs in the coronal image are annotated and the scan depth is assessed by performing the following operations:

[0014] By systematically changing the scanning depth settings, a series of images with different scanning depths were acquired for each patient in addition to routine clinical examinations. The coronal plane was selected as the rib detection area, and an expert team annotated and evaluated the scanning depth of the ribs in the coronal plane images. Based on the physician evaluation criteria, bounding boxes were marked for the rib regions in each frame of the image on the MakeSense platform, forming training, validation, and test datasets.

[0015] Preferably, in step S2, coronal slicing is performed, and model labels are converted and generated based on the MakeSense annotation platform, by performing the following operations:

[0016] Breast ABUS data was acquired using clinical ultrasound equipment. The raw data was stored in DICOM format. The pydicom library of Python was used to read the metadata of the volume data. The three-dimensional images were serialized and sliced ​​along the coronal plane. The spatial mapping relationship was maintained by the orientation matrix. All slices were uniformly resampled to a standard size of 640×640 pixels. Finally, each DICOM generated 400 standard coronal slices. For the imbalance between positive and negative samples in the rib region detection task, the data enhancement was performed by horizontal or vertical flipping and brightness and contrast perturbation.

[0017] The initial bounding box coordinates of the rib region are parsed from the annotation file exported from MakeSense. The integrity of the annotation region is verified using OpenCV's connected component analysis algorithm to ensure that each independent rib structure is correctly labeled. The verified absolute pixel coordinates are converted into the required normalized relative coordinates to generate a standardized label file that meets the requirements of the object detection model. Each label line contains the rib category identifier and the normalized bounding box center coordinates and width and height parameters. The label file and the corresponding ABUS image together construct the dataset.

[0018] Preferably, in S3, the rib detection model includes:

[0019] The backbone network uses the main network structure built into the model for feature extraction. This network constructs a feature extraction system that can efficiently capture multi-level information through multiple convolutional layers, cross-stage feature processing modules, and structural optimization techniques. The network adopts a hierarchical and progressive downsampling method, which gradually expands the receptive range while controlling the amount of computation. It is used to identify various features from fine textures to overall structure and capture rib region features in different images.

[0020] The neck network effectively integrates the rib region features extracted from different levels by the backbone network. It adopts a bidirectional multi-level feature fusion mechanism and uses bidirectional information flow paths to achieve deep fusion of features at different scales. This structure supports weighted fusion of features at each layer and establishes cross-layer connections to enhance information transmission efficiency and improve the model's ability to recognize rib regions of different sizes.

[0021] The detection output separates the classification task from the bounding box regression task. The detection head adopts a dynamic positive sample allocation strategy and improves training efficiency through an adaptive matching mechanism. Its output layer synchronously generates bounding box coordinates, target confidence scores, and probability distributions of specific categories, which together complete the rib localization and recognition task.

[0022] Error optimization, during the model training phase, aims to learn and accurately identify rib locations. This is achieved by comprehensively calculating multiple error components to optimize model performance. A task-aligned sample allocator matches the model's predicted bounding boxes with expert-annotated ground truth boxes, assigning appropriate learning targets to each predicted box. Based on the difference between the model's predicted rib identification results and the ground truth labels, a binary cross-entropy function is used to calculate the classification error, ensuring the model can accurately distinguish rib regions from background tissue. Finally, by combining overlap metrics and distribution distance methods, the regression error of the location boxes is calculated to precisely define rib locations.

[0023] Preferably, in step S3, a relationship-aware attention module is introduced into the backbone network to enhance the feature extraction capability of the rib region. In the rib detection task, this module integrates content attention and relative position attention. By explicitly encoding the relative position information between features, it models the spatial structural relationship between targets. When calculating attention, this module not only relies on feature content but also embeds relative position encoding to accurately capture the arrangement pattern and structural context of ribs and effectively distinguish adjacent ribs with similar appearances. It calculates the attention score of the spatial relationship between ribs. The calculation formula is as follows:

[0024] (1)

[0025] in, Indicates from the rib position Arrive at the location Relative position encoding, For position The query vector, For position The query vector, This is the feature dimension, used for stable training;

[0026] For attention score, For the first The attention weights are then calculated from the value vectors at each position. and output features The calculation formula is as follows:

[0027] (2)

[0028] (3)

[0029] This is used to enable networks to aggregate information based on both semantic content and structural priors;

[0030] In the backbone network, adaptive dynamic convolution is used as the convolution module. The dynamic convolution module adapts to the ribs of different shapes, generating large receptive field convolution kernels for thick rib regions, small receptive field convolution kernels for thin rib regions, and asymmetric convolution kernels for curved regions.

[0031] Input feature map Generate adaptive convolution kernels, For batch size, Input the number of channels. and The height and width of the feature map, and the kernel size. Number of output channels Then the dynamic convolution kernel The generation steps are as follows:

[0032] Global average pooling aggregates features for output. The calculation formula is as follows:

[0033] (4)

[0034] in, For the input feature map, ;

[0035] Attention weights are generated, and the formula for calculating weight α is as follows:

[0036] )) (5)

[0037] in, ∈ , ∈ For learnable parameters, For the hidden layer dimension, This represents the number of convolutional kernels;

[0038] Dynamic convolution kernel synthesis, state convolution kernel The calculation formula is as follows:

[0039] (6)

[0040] in, Indicates input The mixing coefficients of the corresponding network, For predefined basic convolutional kernels;

[0041] Dynamic convolution output:

[0042] (7)

[0043] Multi-level detection is achieved for ribs of different thicknesses.

[0044] Preferably, in step S3, a dilated depthwise separable convolutional module is introduced into the neck network to perform multi-level feature extraction and fusion, and input features... Macro-level input and output input at the meso-level and output Micro-level input and output The multi-level feature extraction calculation formula is as follows:

[0045] = ( (8)

[0046] (9)

[0047] = ( (10)

[0048] = ( (11)

[0049] = (12)

[0050] = ( (13)

[0051] Extracted different levels of output features , and Upsampling to a uniform size, the formula for calculating the uniform size of features is as follows:

[0052] = ( )+ ( )+ ( (14)

[0053] (15).

[0054] Preferably, in step S3, the total loss of rib identification decoupling is calculated. The formula is as follows:

[0055] = + + (16)

[0056] in, For classifying losses, For bounding box regression loss, For distributed focusing loss, , For balancing weighting coefficients;

[0057] For each real rib target, the most suitable prediction box is automatically selected from numerous prediction boxes as a positive sample for training. The selection criteria combine general prediction quality and the reasonableness of the rib region, and the allocation results are obtained using a dynamic positive sample allocation strategy. The calculation formula is as follows:

[0058] (17)

[0059] (18)

[0060] = × (19)

[0061] (20)

[0062] in, To predict the degree of deviation between the location and the actual rib, As the confidence weight for ribs, Indicates the first The prediction box and the first The matching quality of each ground truth bounding box under a general object detection standard. The final matching score for rib specialization, for each real target Select the score that matches its final result. The highest predicted bounding box will be assigned as a positive sample for calculating the loss.

[0063] Preferably, in step S3, the formula for calculating the number of samples is as follows:

[0064] | |= (twenty one)

[0065] | |= (twenty two)

[0066] in, The negative sample set represents the true rib targets, and the predicted bounding boxes that match the true bounding boxes are the predicted bounding boxes. The positive sample set represents the unmatched predicted bounding boxes. The number of negative samples The number of positive samples;

[0067] The formula for calculating the rib detection classification error of the model is as follows:

[0068] (twenty three)

[0069] The formula for calculating the confidence weight of ribs is as follows:

[0070] = × (twenty four)

[0071] The final classification error is calculated using the following formula:

[0072] ) -( ) (25).

[0073] Preferably, in step S4, the rib detection model is tested and its effectiveness evaluated by performing the following operations:

[0074] A test set was constructed using labeled image data. Each image underwent the same standardization process as during training and was then input into the trained model for analysis. Initial outputs, including location boxes, confidence scores, and class classifications, were obtained. A filtering algorithm was used to remove duplicate or highly overlapping prediction boxes based on preset overlap and confidence thresholds, ensuring that only the most accurate result was retained for each rib region. By comparing the overlap between the rib boxes output by the model and the ground truth boxes labeled by experts, the number of correctly identified ribs, the number of non-target regions misclassified as ribs, and the number of undetected real ribs were counted under different levels of stringency. Based on the statistical results, the mean accuracy (mAP), precision, recall, and F1 score were calculated as evaluation indicators to measure the detection accuracy of the model at different levels. A confusion matrix was generated to analyze classification errors in detail, and a visualized prediction result map was output to complete the evaluation of the model's generalization ability.

[0075] Preferably, in step S5, the scanning depth is evaluated based on the number and location of ribs. If the threshold range for rib appearance frames when the scanning depth of each DICOM is moderate is [a, b], then the scanning depth determination logic is as follows:

[0076] The scan depth prediction is too shallow, i.e., class 0: starting frame number If ribs are present in more than b frames of 10 consecutive images, the breast tissue is considered incomplete, and the probe position needs to be adjusted immediately.

[0077] The scan depth is predicted to be moderate, i.e., Class 1: starting frame number. If there are ribs in 10 consecutive frames of images within the frame range of [a, b], it is determined that the scanning range is appropriate and meets the basic screening requirements;

[0078] The scanning depth is predicted to be too deep, i.e., Category 2: starting frame number If there are ribs in 10 consecutive frames of images within the < a frame range, it is determined that the scanning range is too large, affecting the image resolution and diagnosis.

[0079] Compared with the prior art, the beneficial effects of the present invention are:

[0080] The present invention uses a target detection model to automatically locate the rib structure in ABUS images, thereby objectively judging whether the depth of a single scan fully covers all breast tissues. The position of the ribs is related to the scanning depth, so as to evaluate whether the scanning depth meets the quality control standard, providing a new technical path for standardizing the setting of the scanning depth. Automatically evaluating the scanning depth of ABUS images can achieve the standardization of ultrasound screening, avoid repeated scans caused by improper settings, and is the key to improving the efficiency of breast cancer screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a schematic diagram of the visualization of the coronal rib region ultrasound image of the present invention;

[0082] Figure 2 It is a schematic diagram of the visualization of the transverse rib region ultrasound image of the present invention;

[0083] Figure 3 It is a schematic diagram of the prediction result of the coronal rib in the frame region of the present invention;

[0084] Figure 4 It is a flowchart of the scanning depth evaluation of the breast ABUS image of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] To solve the problems that different physicians in the prior art have inconsistent judgment criteria for qualified depth, resulting in significant differences in the image coverage range, it is difficult for manual evaluation to accurately judge whether the image truly and completely contains all the necessary tissues from the superficial skin to the deep chest wall, lacking a reliable reference basis, and it is difficult for inexperienced operators to quickly determine the optimal scanning depth suitable for different breast sizes and shapes, affecting the workflow, please refer to Figures 1-4 , the present embodiment provides the following technical solutions:

[0087] A method for assessing the scanning depth of breast ABUS images, comprising:

[0088] S1: Data collection and physician evaluation;

[0089] Multi-center, multi-device, and multi-parameter 3D DICOM data of breast ultrasound were collected. The scan depth settings were systematically varied, and for each patient, in addition to routine clinical examinations, a series of images at different scan depths were acquired. Compared to the transverse rib region, the coronal rib region is more clearly visible, such as... Figure 1 and Figure 2 As shown, in this embodiment, the coronal plane is selected as the rib detection area. An expert team annotates the ribs in the coronal plane image and evaluates the scanning depth. Based on the physician evaluation standards, bounding boxes are annotated for the rib region in each frame of the image on the makesense platform to form training, validation and test datasets.

[0090] It should be noted that in breast imaging, the scanning depth parameter determines the extent to which the ultrasound beam penetrates the tissue and the displayed field of view. For most linear array probes, the ideal focal point is typically 1 to 2 cm away from the skin in the breast area. Placing the breast area at the center of the ultrasound image usually provides adequate depth, encompassing the entire breast and allowing for sufficient assessment of breast lesions. ABUS image scanning depth categories are defined as follows:

[0091] Category 0: Images with insufficient scanning depth. Due to the scanning depth value set too small on the ultrasound equipment, the rib area cannot be displayed in the image, and lesions in deeper areas such as the front of the chest wall may be missed. Incomplete breast coverage.

[0092] Class 1: Images with moderate scanning depth. The appropriate scanning depth is set according to the actual size and shape of the patient's breast. The rib line is clearly visible, and the breast tissue occupies 1 / 2 to 1 / 3 of the depth.

[0093] Category 2: Images with excessive scanning depth. This is due to an excessively large scanning depth value set on the ultrasound equipment. The images contain rib areas and a large amount of non-breast tissue such as intercostal muscles, pleura, and lung tissue, resulting in too much useless information.

[0094] It should be noted that in coronal views, the rib region is a characteristic structure at the bottom of the image, consisting of a hyperechoic arc and a clean, anechoic shadow behind it. The presence of ribs is a sign that the scan depth is sufficient and includes complete posterior breast tissue. In ABUS coronal images, ribs are defined as follows:

[0095] Category 0: Ribs. The ribs are located at the deepest layer of the image, corresponding to the chest wall region, and include all breast tissue up to the chest wall. The ABUS scan depth is assessed by the location and number of ribs in the coronal section.

[0096] S2: Data preprocessing and labeling;

[0097] 1) Coronal section preparation:

[0098] Breast ABUS data was acquired through clinical ultrasound scanning. The raw data was stored in DICOM format. The metadata of the volumetric data was read using the pydicom library of Python. The three-dimensional images were serialized and sliced ​​along the coronal plane, while the spatial mapping relationship was maintained through the orientation matrix. All slices were uniformly resampled to a standard size of 640×640 pixels. Finally, each DICOM could generate 400 standard coronal slices. To address the problem of positive and negative sample imbalance in the rib region detection task, such as the complete rib image in deep scans but the missing or partially visible ribs in shallow scans, data enhancement methods such as horizontal or vertical flipping and brightness and contrast perturbation were used.

[0099] 2) Tag generation:

[0100] After identifying the bounding boxes of the rib regions in the ABUS coronal images, model labels were converted and generated using the MakeSense annotation platform. First, the initial bounding box coordinates of the rib regions were parsed from the annotation files exported from MakeSense. The integrity of the annotated regions was verified using OpenCV's connected component analysis algorithm to ensure that each independent rib structure was correctly labeled. The verified absolute pixel coordinates were converted into the required normalized relative coordinates. Finally, a standardized label file that meets the requirements of the object detection model was generated. Each label line contains the rib category identifier and the normalized bounding box center coordinates and width and height parameters. The label file and the corresponding ABUS images together construct the dataset.

[0101] S3: Establish a rib detection model;

[0102] Rib region detection is performed based on an object detection architecture. A complete process, including constructing a backbone network, a neck network, detection output, and error optimization modules, is used to train a dedicated model capable of automatically locating rib regions in images. The rib detection model includes the following components:

[0103] 1) Backbone Network: Feature extraction is performed using the main network structure built into the model. This network constructs a feature extraction system that can efficiently capture multi-level information through multiple convolutional layers, cross-stage feature processing modules, and structural optimization techniques. The network adopts a hierarchical downsampling method, which gradually expands the receptive range while controlling the amount of computation, thereby effectively identifying various features from fine textures to overall structure and capturing rib region features in different images.

[0104] This embodiment introduces a relation-aware attention module into the backbone network to enhance the feature extraction capability of the rib region. Rib detection has its unique characteristics, namely, the existence of relatively fixed and regular spatial anatomical relationships between ribs. This module integrates content attention and relative position attention, effectively modeling the spatial structural relationships between targets by explicitly encoding the relative positional information between features. Furthermore, the ribs themselves have a long, arc-shaped structure, and perceiving their shape requires establishing long-distance dependencies, while the relative positions between ribs constitute important global contextual information.

[0105] Specifically, this module calculates attention not only based on feature content but also by embedding relative positional encoding. This allows for more accurate capture of the rib arrangement patterns and structural context, effectively distinguishing adjacent ribs with similar appearances and avoiding feature confusion. It calculates the attention score based on the spatial relationship between ribs. The calculation formula is as follows:

[0106] (1)

[0107] in Indicates from the rib position Arrive at the location Relative position encoding, For position The query vector, For position The query vector, This is the feature dimension, used to stabilize training and prevent the dot product result from becoming too large;

[0108] For attention score, For the first The attention weights are then calculated from the value vectors at each position. and output features The calculation formula is as follows:

[0109] (2)

[0110] (3)

[0111] This design ensures that the network can simultaneously rely on semantic content and structural priors when aggregating information, improving the robustness of detecting repetitive and regularly dissected targets. Ultimately, global contextual recall and relative location accuracy are improved, while the false detection rate is reduced.

[0112] In this embodiment, adaptive dynamic convolution is used as the convolution module in the backbone network to adapt to the diversity of rib morphology. Because ribs vary in shape, thickness, and curvature, and are partially obscured by soft tissue, the dynamic convolution module adapts to the diverse rib shapes, generating large receptive field convolution kernels for thicker rib regions, small receptive field convolution kernels for thinner rib regions, and asymmetric convolution kernels for curved regions, thus improving the model's adaptability and robustness to the characteristics of different rib regions. The input feature map... Generate adaptive convolution kernels, For batch size, Input the number of channels. and The height and width of the feature map, and the kernel size. Number of output channels Then the dynamic convolution kernel The generation steps are as follows:

[0113] Global average pooling aggregates features for output. The calculation formula is as follows:

[0114] (4)

[0115] in, For the input feature map, ;

[0116] Attention weights are generated, and the formula for calculating weight α is as follows:

[0117] )) (5)

[0118] in, ∈ , ∈ For learnable parameters, For the hidden layer dimension, This represents the number of convolutional kernels;

[0119] Dynamic convolution kernel synthesis, state convolution kernel The calculation formula is as follows:

[0120] (6)

[0121] in Indicates input The mixing coefficients of the corresponding network, For predefined basic convolutional kernels;

[0122] Dynamic convolution output:

[0123] (7)

[0124] Ultimately, multi-level detection of ribs of different thicknesses was achieved, improving the detection rate of rib-occluded areas and increasing the sensitivity of rib edge clarity.

[0125] 2) Neck Network: This network effectively integrates the rib region features extracted from different levels of the backbone network. It employs a bidirectional, multi-level feature fusion mechanism, achieving deep fusion of features at different scales through bidirectional information flow paths. Simultaneously, this structure supports weighted fusion of features from each layer and establishes cross-layer connections to enhance information transmission efficiency, thereby improving the model's ability to recognize rib regions of different sizes.

[0126] This embodiment introduces a depthwise separable convolutional module into the neck network for multi-level feature extraction and fusion. Due to the multi-level characteristics of the ribs, the macroscopic level represents the overall shape of the thoracic cavity, the mesoscopic level represents the structure of a single rib, and the microscopic level represents the details of the rib edges; input features Macro-level input and output input at the meso-level and output Micro-level input and output The multi-level feature extraction calculation formula is as follows:

[0127] = ( (8)

[0128] (9)

[0129] = ( (10)

[0130] = ( (11)

[0131] = (12)

[0132] = ( (13)

[0133] Extracted different levels of output features , and Upsampling to a uniform size, the formula for calculating the uniform size of features is as follows:

[0134] = ( )+ ( )+ ( (14)

[0135] (15)

[0136] Ultimately, the recall rate for detecting small, distal ribs improved, and the accuracy of rib edge recognition was enhanced.

[0137] 3) Detection Output: The classification task and bounding box regression task are processed separately, effectively avoiding mutual interference between the two tasks during feature learning. This detection head adopts a dynamic positive sample allocation strategy and improves training efficiency through an adaptive matching mechanism. Its output layer synchronously generates bounding box coordinates (x, y, w, h), target confidence scores, and specific category probability distributions, jointly completing the rib localization and recognition task. The total loss is calculated decoupled for rib recognition. The formula is as follows:

[0138] = + + (16)

[0139] in, For classifying losses, For bounding box regression loss, For distributed focusing loss, , These are the balancing weighting coefficients.

[0140] For each real rib target, the most suitable prediction box is automatically selected from numerous prediction boxes as a positive sample for training. The selection criteria combine general prediction quality and the reasonableness of the rib region, and the allocation results are obtained using a dynamic positive sample allocation strategy. The calculation formula is as follows:

[0141] (17)

[0142] (18)

[0143] = × (19)

[0144] (20)

[0145] in, To predict the degree of deviation between the location and the actual rib, As the confidence weight for ribs, Indicates the first The prediction box and the first The matching quality of each ground truth bounding box under a general object detection standard. The final matching score for rib specialization, for each real target Select the score that matches its final result. The highest predicted bounding box will be assigned as a positive sample for calculating the loss.

[0146] 4) Error Optimization: During the model training phase, the goal is to learn and accurately identify rib locations. Model performance is optimized by comprehensively calculating multiple error components. First, a task-aligned sample allocator matches the model's predicted bounding boxes with expert-annotated ground truth boxes, assigning a suitable learning objective to each predicted box. Based on the difference between the model's predicted rib identification results and the ground truth labels, a binary cross-entropy function is used to calculate the classification error, ensuring the model can accurately distinguish the rib region from background tissue. The regression error of the location boxes is calculated using an overlap metric and a distribution distance metric to accurately define the rib locations. Position coordinate prediction is treated as a distribution learning problem. By calculating the difference between the predicted and ground truth distributions, the model focuses more on learning the positional uncertainty of the rib boundaries, thereby improving localization accuracy. The formula for calculating the number of samples is as follows:

[0147] | |= (twenty one)

[0148] | |= (twenty two)

[0149] in, The negative sample set represents the true rib targets, and the predicted bounding boxes that match the true bounding boxes are the predicted bounding boxes. The positive sample set represents the unmatched predicted bounding boxes. The number of negative samples The number of positive samples.

[0150] The formula for calculating the rib detection classification error of the model is as follows:

[0151] (twenty three)

[0152] The formula for calculating the confidence weight of ribs is as follows:

[0153] = × (twenty four)

[0154] The final classification error is calculated using the following formula:

[0155] ) -( ) (25).

[0156] It should be noted that the target recognition model extracts key information from medical images through hierarchical processing to help accurately locate the target region. It uses a series of adjustable filters to scan the original image and calculates basic visual features from edges and textures to brightness and darkness layer by layer. The system fuses these features at different levels into a multi-layer information map. This map not only accurately preserves the location data of the target region, but also integrates high-level semantic information for recognition, thus providing effective input for the final target localization and recognition steps.

[0157] It should be noted that feature extraction refers to selecting the most effective features or attributes from medical images to support the accurate localization and recognition of the rib region. This process is a hierarchical, progressive computational flow. The raw data is fed into a convolutional layer, and a set of learnable convolutional kernels are used to perform convolution operations on the input. By calculating the dot product between the convolutional kernel and the local data region, specific types of features such as edges, textures, brightness, and contrast are extracted. Finally, multi-level feature maps are output. These feature maps retain the precise spatial location information and high-level semantic content of the ribs, providing feature input for the subsequent rib detection head and ensuring the accurate regression and classification of the rib bounding boxes.

[0158] It should be noted that multi-level feature fusion enhances the system's ability to recognize objects of different sizes by fusing information from different levels in an image. Through top-down paths and cross-layer connections, it combines deep features containing rich semantics with shallow features that preserve fine location information to construct a comprehensive multi-level feature map. In recognition tasks, it can capture both large-scale overall structures and small-scale local details, thus accurately finding the target area under different scanning conditions.

[0159] It should be noted that the error function represents the difference or error between the model prediction and the true value, and is a function used to optimize the model. In object detection tasks, the error function is divided into category error and location error. This study uses binary cross-entropy to identify the rib region, focusing on distinguishing the rib region from the background tissue. The bounding box regression loss is used to consider the overlap area between the predicted box and the true box, and to evaluate the consistency of the center point distance and aspect ratio to ensure the spatial positioning accuracy of the rib bounding box.

[0160] It's important to note that in neural networks, activation functions are responsible for non-linearly processing the input signal and passing the results to the next layer, enabling the entire network to learn and express complex features and relationships. Different activation functions, specifically designed for different tasks, are used in different parts of the network. In the main feature extraction and feature fusion stages, a ReLU function with a small negative slope is used; while in the final output stage, the part used to predict position coordinates uses a sigmoid function, limiting its output value to between 0 and 1. This directly maps to the relative position within each unit, ensuring the reasonableness and stability of the positional information.

[0161] S4: Model testing and performance evaluation;

[0162] To test the performance of the trained model in locating ribs in specified images, a test set was constructed using labeled image data. Each image underwent the same standardization process as during training and was then input into the trained model for analysis, yielding initial outputs including bounding boxes, confidence scores, and class classifications. A filtering algorithm was employed to remove duplicate or highly overlapping predicted boxes based on preset overlap and confidence thresholds, ensuring that only the most accurate result was retained for each rib region. By comparing the overlap between the model's output rib boxes and expert-annotated ground truth boxes, the number of correctly identified ribs, the number of non-target regions misclassified as ribs, and the number of undetected real ribs were counted under different levels of stringency. Based on the statistical results, mean accuracy (mAP), precision, recall, and F1 score were calculated as evaluation indicators to measure the model's detection accuracy at different levels. A confusion matrix was generated to analyze classification errors in detail, and a visualized prediction result map was output to evaluate the model's generalization ability. Figure 3 As shown.

[0163] S5: Establish rules for determining scan depth;

[0164] A scanning depth determination rule is established based on the number and location of rib frames, and coronal image sequences are extracted. , ,..., }, detect each frame of image The rib region was used to obtain the detection result sequence. ), )=1 indicates that ribs were detected. )=0 indicates no ribs detected, and the first consecutive frame segment with detected ribs is identified. and starting frame number Analysis of the starting frame number of rib appearance The location of 10 consecutive frames in the image is used to determine the scan depth category.

[0165] Evaluate the scanning depth based on the number and position of ribs. Assume that when the scanning depth of each DICOM is moderate, the rib appearance frame threshold range is [a, b]. Then the scanning depth judgment logic is as follows:

[0166] 1) The predicted scanning depth is too shallow (class 0): If there are ribs in the continuous 10 frames starting from the starting frame number within the range of >b frames, it is determined that the breast tissue is incomplete, and the probe position needs to be adjusted immediately;

[0167] 2) The predicted scanning depth is moderate (class 1): If there are ribs in the continuous 10 frames starting from the starting frame number within the range of [a, b] frames, it is determined that the scanning range is appropriate and meets the basic screening requirements;

[0168] 3) The predicted scanning depth is too deep (class 2): If there are ribs in the continuous 10 frames starting from the starting frame number within the range of <a frames, it is determined that the scanning range is too large, affecting image resolution and diagnosis.

[0169] It should be noted that recognition involves using artificial intelligence algorithms and technologies to automatically recognize input data samples; when recognized into two categories, it is called binary classification, and it can also be divided into multiple categories (such as three categories). The goal of this type of task is to train an artificial intelligence model to accurately recognize and classify input data into predefined categories.

[0170] It should be noted that the scanning depth refers to the detection range required to completely cover the breast tissue from the nipple to the chest wall. The scanning depth is automatically evaluated by the post-processing analysis module, which determines whether the breast tissue is sufficiently imaged by identifying the position and number of ribs in the image and combining anatomical knowledge, and completes the determination of the scanning depth qualification.

[0171] Specifically, the scanning depth evaluation task usually involves the following aspects:

[0172] 1) Definition of evaluation criteria: In the position evaluation task, the criteria for qualified scanning positions must be clearly defined. These criteria are based on the rib number and position characteristics, and the scanning depth quality evaluation results are divided into too shallow, moderate, and too deep scanning depths.

[0173] 2) Data input: The input of the evaluation task is ABUS coronal plane images, which need to be properly preprocessed and standardized so that they can be accepted and processed by the object detection model.

[0174] 3) Model Training: ABUS image scanning depth evaluation is achieved by training an automatic rib target detection model and a post-processing analysis module. This is accomplished through supervised learning, using a dataset with rib bounding box annotations for training. During training, the model continuously adjusts its internal parameters through backpropagation and optimization algorithms to minimize the difference between the predicted bounding boxes and the true annotations.

[0175] 4) Model Evaluation: After training, the model's performance needs to be evaluated using an independent test set. Evaluation metrics typically include detection accuracy (mAP), recall, and precision, which help assess the model's performance on the rib detection task.

[0176] 5) Scan depth assessment: The rib region predicted by the model is used for scan depth discrimination, which transforms the appearance of ribs into an automatic judgment of the scan depth qualification. By using the rib position, it is ensured that each scan includes all tissues with diagnostic value in the depth direction, thereby ensuring the quality and integrity of the image.

[0177] 6) Model Application: The model achieves rib region detection through training and validation. A post-processing analysis module is added to form a complete scan depth evaluation architecture. This model is integrated into the ABUS scanning workstation to analyze rib imaging in real time during the scanning process. Post-processing analysis automatically evaluates whether the scan depth is acceptable and provides adjustment suggestions to guide the operator in obtaining qualified images that meet diagnostic requirements.

[0178] In summary, this invention utilizes a target detection model to automatically locate rib structures in ABUS images, thereby objectively determining whether the depth of a single scan adequately covers all breast tissue. The position of the ribs is related to the scan depth, thus assessing whether the scan depth meets quality control standards. This provides a new technical approach for standardizing scan depth settings. Automatically evaluating the scan depth of ABUS images can standardize ultrasound screening and avoid repeated scans caused by improper settings, which is key to improving the efficiency of breast cancer screening.

[0179] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0180] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the scanning depth of breast ABUS images, characterized in that, include: S1: Collect multi-center, multi-device, and multi-parameter three-dimensional DICOM data of breast ultrasound, and annotate the ribs and assess the scanning depth of the coronal images; S2: Perform coronal section processing and convert and generate model labels based on the makesense annotation platform; S3: Based on the target detection architecture, rib region detection is performed. A complete process is constructed, including a backbone network, a neck network, detection output, and an error optimization module. A rib detection model that can automatically locate the rib region in the image is trained. S4: Perform model testing and performance evaluation on the rib detection model, and output a visualized prediction result map; S5: Establish scanning depth judgment rules based on rib frame number and position, extract coronal image sequence, detect rib region in each frame, obtain detection result sequence, identify the first consecutive detected rib frame segment and starting frame number, analyze the position of the 10 consecutive frames in the image where the starting frame number of the rib appears, and determine the scanning depth category.

2. The method for evaluating the scanning depth of breast ABUS images according to claim 1, characterized in that, In step S1, the ribs in the coronal image are labeled and the scanning depth is assessed by performing the following operations: By systematically changing the scanning depth settings, a series of images with different scanning depths were acquired for each patient in addition to routine clinical examinations. The coronal plane was selected as the rib detection area, and an expert team annotated and evaluated the scanning depth of the ribs in the coronal plane images. Based on the physician evaluation criteria, bounding boxes were marked for the rib regions in each frame of the image on the MakeSense platform, forming training, validation, and test datasets.

3. The method for evaluating the scanning depth of breast ABUS images according to claim 2, characterized in that, In step S2, coronal slicing is performed, and model labels are converted and generated based on the MakeSense annotation platform. The following operations are performed: Breast ABUS data was acquired using clinical ultrasound equipment. The raw data was stored in DICOM format. The pydicom library of Python was used to read the metadata of the volume data. The three-dimensional images were serialized and sliced ​​along the coronal plane. The spatial mapping relationship was maintained by the orientation matrix. All slices were uniformly resampled to a standard size of 640×640 pixels. Finally, each DICOM generated 400 standard coronal slices. For the imbalance between positive and negative samples in the rib region detection task, the data enhancement was performed by horizontal or vertical flipping and brightness and contrast perturbation. The initial bounding box coordinates of the rib region are parsed from the annotation file exported from MakeSense. The integrity of the annotation region is verified using OpenCV's connected component analysis algorithm to ensure that each independent rib structure is correctly labeled. The verified absolute pixel coordinates are converted into the required normalized relative coordinates to generate a standardized label file that meets the requirements of the object detection model. Each label line contains the rib category identifier and the normalized bounding box center coordinates and width and height parameters. The label file and the corresponding ABUS image together construct the dataset.

4. The method for evaluating the scanning depth of breast ABUS images according to claim 3, characterized in that, In S3, the rib detection model includes: The backbone network uses the main network structure built into the model for feature extraction. This network constructs a feature extraction system that can efficiently capture multi-level information through multiple convolutional layers, cross-stage feature processing modules, and structural optimization techniques. The network adopts a hierarchical and progressive downsampling method, which gradually expands the receptive range while controlling the amount of computation. It is used to identify various features from fine textures to overall structure and capture rib region features in different images. The neck network effectively integrates the rib region features extracted from different levels by the backbone network. It adopts a bidirectional multi-level feature fusion mechanism and uses bidirectional information flow paths to achieve deep fusion of features at different scales. This structure supports weighted fusion of features at each layer and establishes cross-layer connections to enhance information transmission efficiency and improve the model's ability to recognize rib regions of different sizes. The detection output separates the classification task from the bounding box regression task. The detection head adopts a dynamic positive sample allocation strategy and improves training efficiency through an adaptive matching mechanism. Its output layer synchronously generates bounding box coordinates, target confidence scores, and probability distributions of specific categories, which together complete the rib localization and recognition task. Error optimization, during the model training phase, aims to learn and accurately identify rib locations. This is achieved by comprehensively calculating multiple error components to optimize model performance. A task-aligned sample allocator matches the model's predicted bounding boxes with expert-annotated ground truth boxes, assigning appropriate learning targets to each predicted box. Based on the difference between the model's predicted rib identification results and the ground truth labels, a binary cross-entropy function is used to calculate the classification error, ensuring the model can accurately distinguish rib regions from background tissue. Finally, by combining overlap metrics and distribution distance methods, the regression error of the location boxes is calculated to precisely define rib locations.

5. The method for evaluating the scanning depth of breast ABUS images according to claim 4, characterized in that, In step S3, a relationship-aware attention module is introduced into the backbone network to enhance the feature extraction capability of the rib region. In the rib detection task, this module integrates content attention and relative position attention. By explicitly encoding the relative position information between features, it models the spatial structural relationship between targets. When calculating attention, this module not only relies on feature content but also embeds relative position encoding to accurately capture the arrangement pattern and structural context of ribs and effectively distinguish adjacent ribs with similar appearances. It calculates the attention score for the spatial relationship between ribs. The calculation formula is as follows: （1）； in, Indicates from the rib position Arrive at the location Relative position encoding, For position The query vector, For position The query vector, This is the feature dimension, used for stable training; For attention score, For the first The attention weights are then calculated from the value vectors at each position. and output features The calculation formula is as follows: ) （2）； （3）； This is used to enable networks to aggregate information based on both semantic content and structural priors; In the backbone network, adaptive dynamic convolution is used as the convolution module. The dynamic convolution module adapts to the ribs of different shapes, generating large receptive field convolution kernels for thick rib regions, small receptive field convolution kernels for thin rib regions, and asymmetric convolution kernels for curved regions. Input feature map Generate adaptive convolution kernels, For batch size, Input the number of channels. and The height and width of the feature map, and the kernel size. Number of output channels Then the dynamic convolution kernel The generation steps are as follows: Global average pooling aggregates features for output. The calculation formula is as follows: （4）； in, For the input feature map, ; Attention weights are generated, and the formula for calculating weight α is as follows: )) （5）； in, ∈ , ∈ For learnable parameters, For the hidden layer dimension, This represents the number of convolutional kernels; Dynamic convolution kernel synthesis, state convolution kernel The calculation formula is as follows: （6）； in, Indicates input The mixing coefficients of the corresponding network, For predefined basic convolutional kernels; Dynamic convolution output: （7）； Multi-level detection is achieved for ribs of different thicknesses.

6. The method for evaluating the scanning depth of a breast ABUS image according to claim 5, characterized in that, In step S3, a dilated depthwise separable convolutional module is introduced into the neck network to perform multi-level feature extraction and fusion, and input features. Macro-level input and output input at the meso-level and output Micro-level input and output The multi-level feature extraction calculation formula is as follows: = ( ) （8）； （9）； = ( ) （10）； = ( ) （11）； = （12）； = ( ) （13）； Extracted different levels of output features , and Upsampling to a uniform size, the formula for calculating the uniform size of features is as follows: = ( )+ ( )+ ( ) （14）； （15）。 7. The method for evaluating the scanning depth of a breast ABUS image according to claim 6, characterized in that, In step S3, the total loss of rib identification decoupling is calculated. The formula is as follows: = + + （16）； in, For classifying losses, For bounding box regression loss, For distributed focusing loss, , For balancing weighting coefficients; For each real rib target, the most suitable prediction box is automatically selected from numerous prediction boxes as a positive sample for training. The selection criteria combine general prediction quality and the reasonableness of the rib region, and the allocation results are obtained using a dynamic positive sample allocation strategy. The calculation formula is as follows: （17）； （18）； = × （19）； （20）； in, To predict the degree of deviation between the location and the actual rib, For rib confidence weights, Indicates the first The prediction box and the first The matching quality of each ground truth bounding box under a general object detection standard. The final matching score for rib specialization, for each real target Select the score that matches its final result. The highest predicted bounding box will be assigned as a positive sample for calculating the loss.

8. The method for evaluating the scanning depth of a breast ABUS image according to claim 7, characterized in that, In S3, the formula for calculating the number of samples is as follows: | |= （21）； | |= （22）； in, The negative sample set represents the true rib targets, and the predicted bounding boxes that match the true bounding boxes are the predicted bounding boxes. The positive sample set represents the unmatched predicted bounding boxes. The number of negative samples The number of positive samples; The formula for calculating the rib detection classification error of the model is as follows: （23）； The formula for calculating the confidence weight of ribs is as follows: = × （24）； The final classification error is calculated using the following formula: ) -( ) （25）。 9. The method for evaluating the scanning depth of a breast ABUS image according to claim 8, characterized in that, In step S4, the rib detection model is tested and its effectiveness is evaluated by performing the following operations: A test set was constructed using labeled image data. Each image underwent the same standardization process as during training and was then input into the trained model for analysis. Initial outputs, including location boxes, confidence scores, and class classifications, were obtained. A filtering algorithm was used to remove duplicate or highly overlapping prediction boxes based on preset overlap and confidence thresholds, ensuring that only the most accurate result was retained for each rib region. By comparing the overlap between the rib boxes output by the model and the ground truth boxes labeled by experts, the number of correctly identified ribs, the number of non-target regions misclassified as ribs, and the number of undetected real ribs were counted under different levels of stringency. Based on the statistical results, the mean accuracy (mAP), precision, recall, and F1 score were calculated as evaluation indicators to measure the detection accuracy of the model at different levels. A confusion matrix was generated to analyze classification errors in detail, and a visualized prediction result map was output to complete the evaluation of the model's generalization ability.

10. The method for evaluating the scanning depth of a breast ABUS image according to claim 9, characterized in that, In step S5, the scanning depth is evaluated based on the number and location of ribs. If the threshold range for rib appearance frames when the scanning depth of each DICOM is moderate is [a, b], then the scanning depth determination logic is as follows: The scan depth prediction is too shallow, i.e., class 0: starting frame number If ribs are present in more than b frames of 10 consecutive images, the breast tissue is considered incomplete, and the probe position needs to be adjusted immediately. The scan depth is predicted to be moderate, i.e., Class 1: starting frame number. If ribs are found in 10 consecutive frames within the range of [a, b], then the scanning range is deemed appropriate and meets the basic screening requirements. The predicted scan depth is too deep, i.e., Class 2: starting frame number If there are ribs in the continuous 10-frame images within the frame range, it is determined that the scan range is too large, affecting the image resolution and diagnosis.

Images