A breast ultrasound auxiliary analysis method and system based on anatomical and quantitative consistency constraints

Abstract

This invention discloses a breast ultrasound-assisted analysis method and system based on anatomical and quantitative consistency constraints, belonging to the field of medical image processing technology. The method includes: acquiring breast ultrasound images; outputting a segmentation probability map through a pre-trained cascaded AI model, and calculating core quantitative indicators such as area, thickness, and glandular tissue proportion (GTC) based on physical resolution; finally, outputting a segmentation mask, quantitative indicators, and quality control prompts as auxiliary diagnostic references for doctors. The model innovatively introduces anatomical inclusion relationship constraints and quantitative consistency loss during the training phase, and optimizes and selects the model through a multi-stage weight scheduling strategy and a multi-indicator comprehensive evaluation mechanism. This solves the problems of strong subjectivity and poor consistency in traditional manual quantification by doctors, and the problems of existing AI baseline models emphasizing pixel-level segmentation while ignoring clinical quantitative errors and anatomical logic.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology. Specifically, it relates to a breast ultrasound AI-assisted quantitative analysis system and method, and more specifically, to a breast ultrasound-assisted analysis method and system based on anatomical and quantitative consistency constraints. Background Technology

[0002] Breast ultrasound is an important imaging tool for the clinical screening and diagnosis of breast diseases. In the clinical evaluation of breast ultrasound, the accurate identification and quantification of fibroglandular tissue (FGT) and glandular areas (such as thickness, area, and especially the core indicator of breast density—the proportion of glandular tissue (GTC)) is of great reference value for subsequent clinical evaluation and decision-making.

[0003] Currently, quantitative analysis of breast ultrasound images has the following main shortcomings:

[0004] In traditional clinical assessment procedures, doctors mainly rely on manual observation of ultrasound images combined with manual measurements. This method is highly subjective, and the measurement results at different sections and time points are prone to fluctuations. Furthermore, manual measurements are time-consuming and labor-intensive, making it difficult to scale up and accumulate structured data.

[0005] While existing AI-assisted tools (such as CascadedBaseline based on cascaded networks) achieve basic segmentation and quantization, their optimization objectives often focus on pixel-level segmentation evaluation metrics (such as the Dice coefficient), which is not equivalent to optimizing clinical quantification error. Furthermore, purely data-driven models lack anatomical and physical constraints, making them prone to localized prediction errors (such as glandular overflow into the FGT region), and their performance fluctuates significantly when handling difficult cases with ambiguous boundaries. Summary of the Invention

[0006] To address the technical problem of existing technologies that overemphasize pixel-level indicators while neglecting actual clinical quantitative errors and anatomical rationality, this invention provides a breast ultrasound-assisted analysis method and system based on anatomical and quantitative consistency constraints.

[0007] The technical solution of the present invention is as follows:

[0008] On one hand, this invention provides an AI-assisted quantitative analysis method for breast ultrasound, comprising: acquiring a breast ultrasound image to be analyzed and extracting image features based on the physical resolution of the ultrasound image; inputting the ultrasound image into a pre-trained AI analysis model deployed in a computing device, the AI ​​analysis model including a cascaded feature extraction network and a region segmentation network to output a prediction mask containing at least a first anatomical structure and a second anatomical structure; wherein the first anatomical structure is physically contained within the second anatomical structure; calculating clinical quantitative indicators for the first and second anatomical structures based on the prediction mask and the physical resolution; generating quality control prompts based on the local confidence distribution of the prediction mask, and outputting the prediction mask, clinical quantitative indicators, and quality control prompts; wherein the total loss function of the pre-trained AI analysis model during the training phase includes an anatomical constraint term and a quantization consistency constraint term; the anatomical constraint term is used to penalize abnormal regions where the predicted boundary of the first anatomical structure exceeds the predicted boundary of the second anatomical structure; the quantization consistency constraint term is used to minimize the error between the soft quantization features predicted by the model and the true labeled quantization features.

[0009] On the other hand, this invention also provides an AI-assisted quantitative analysis system for breast ultrasound, comprising: a data input and parsing module for acquiring the breast ultrasound image to be analyzed and its physical resolution; a cascaded inference module for running the pre-trained AI analysis model generated by the above training mechanism and outputting a predicted mask of the anatomical structure; a post-processing and quantification module for calculating clinical quantitative indicators based on the predicted mask and physical resolution; a quality control generation module for generating quality control prompts based on the confidence distribution of the inference process; and an interactive output module for presenting the predicted mask, clinical quantitative indicators, and quality control prompts to assist in manual confirmation. It should be noted that this invention is an AI-assisted quantitative tool, and its output segmentation results, quantitative indicators, and quality control information are only for auxiliary reference; the final clinical diagnosis is still completed by the doctor. The beneficial effects of this invention are:

[0010] Reduced quantization error: While maintaining stable basic segmentation performance, the introduction of quantization consistency constraints significantly reduced clinically important quantization errors such as area and thickness.

[0011] Enhancing robustness and anatomical plausibility in challenging cases: By employing anatomical relationship constraints and a multi-index model selection strategy, local errors that violate anatomical norms are effectively reduced. Experiments show that in complex validation sets with ambiguous boundaries, this system effectively reduces the outside ratio of anatomical regions and improves the overall clinical score. Attached Figure Description

[0012] Figure 1 The above is an overall flowchart of a breast ultrasound AI-assisted quantitative analysis method provided in an embodiment of the present invention.

[0013] Figure 2 This is a module architecture diagram of an AI-assisted quantitative analysis system provided in an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to specific embodiments.

[0015] Example 1.

[0016] This embodiment provides an AI-assisted quantitative analysis method for breast ultrasound, including:

[0017] Acquire breast ultrasound images to be analyzed and extract image features based on the physical resolution of the ultrasound images;

[0018] Ultrasound images are input into a pre-trained AI analysis model deployed on a computing device. The AI ​​analysis model includes a cascaded feature extraction network and a region segmentation network to output a prediction mask containing at least a first anatomical structure and a second anatomical structure; wherein the first anatomical structure is physically contained within the second anatomical structure. Based on the prediction mask and physical resolution, clinical quantitative indicators for the first and second anatomical structures are calculated. Quality control prompts are generated based on the local confidence distribution of the prediction mask, and the prediction mask, clinical quantitative indicators, and quality control prompts are output. The total loss function of the pre-trained AI analysis model during the training phase includes an anatomical constraint term and a quantization consistency constraint term. The anatomical constraint term is used to penalize abnormal regions where the predicted boundary of the first anatomical structure exceeds the predicted boundary of the second anatomical structure. The quantization consistency constraint term is used to minimize the error between the soft quantization features predicted by the model and the true labeled quantization features.

[0019] The first anatomical structure is the glandular region, and the second anatomical structure is the fibroglandular tissue region. Clinical quantitative indicators include at least: the area of ​​the glandular region, the area of ​​the fibroglandular tissue region, the thickness of the glandular region, and the glandular tissue percentage (GTC), which characterizes the density of the breast. The glandular tissue percentage (GTC) is the ratio of the area or volume of the glandular region to the fibroglandular tissue region.

[0020] The calculation process of the quantization consistency constraint term includes: during the forward propagation process of model training, obtaining the pixel-level probability map of the network output for the anatomical structure; performing differentiable morphological feature approximation calculation on the pixel-level probability map to obtain the predicted soft area or soft thickness features; calculating the difference function between the predicted soft area or soft thickness features and the quantization index corresponding to the true annotation, and using it as the quantization consistency constraint term.

[0021] The pre-trained AI analysis model adopts a phased weight scheduling strategy during the training phase: In the first training phase, the total loss function is updated only based on the pixel-level basic segmentation error; in the second training phase, the weight values ​​of the dissection constraint and quantization consistency constraint in the total loss function are gradually increased according to the preset increment function until the preset upper limit is reached.

[0022] The quality control prompt information is generated based on the local confidence distribution of the prediction mask, including: calculating the information entropy of the model output probability map in the target boundary region; when the local information entropy of the boundary region exceeds the preset uncertainty threshold, generating a confidence anomaly identifier to prompt manual review, and outputting it as a quality control prompt information.

[0023] The optimal version of the pre-trained AI analysis model is selected through a multi-index comprehensive evaluation model. The selection steps include: on the validation set, calculating the independent scores of the candidate model on the pixel-level segmentation index, the absolute error of the quantitative measurement index, and the violation rate of dissecting inclusion relationships; weighting and summing the independent scores based on the preset weight coefficients to obtain the comprehensive score, and selecting the candidate model with the highest comprehensive score as the pre-trained AI analysis model for final deployment.

[0024] Application Example 1.1: Training and Model Building Process (Offline Training Phase)

[0025] This embodiment mainly illustrates the training mechanism of AI-assisted quantization models.

[0026] Input: A dataset of breast ultrasound images with labels and corresponding real clinical quantitative indicators.

[0027] Processing flow:

[0028] Cascaded segmentation module construction: A multi-stage cascaded segmentation network is constructed. The first-stage network is used for coarse localization of breast tissue; the second-stage network performs fine segmentation using FGT and Gland within the region of interest (ROI) output by the first stage. The network backbone has a replaceable architecture.

[0029] Anatomical Constraint Optimization Module: Introduces anatomical logical constraints into the loss function. Since glands are necessarily contained within fibroglandular tissue (FGT), the system constructs a pixel-level asymmetric penalty function. Specifically, let the probability of the $i$-th pixel in the image being predicted as a gland be... The probability of being predicted as FGT is Then the anatomical constraint loss term The calculation formula is:

[0030]

[0031] in, This represents the total number of pixels in the image. When the model predicts that gland pixels exceed the FGT boundary (i.e., ...), ... When the term generates a loss gradient greater than 0, a penalty is applied; otherwise, the term is 0. The quantization-guided consistency optimization module (RatioConsistency) introduces a quantization-oriented loss term. During forward propagation, soft area features, such as the predicted glandular soft area, are approximated based on a probabilistic graph. FGT soft area Further calculations were performed to determine the predicted glandular tissue percentage (GTC), i.e., the area ratio. Let the GTC value actually labeled by professional doctors be... Then the quantification consistency loss term Defined as the absolute error or mean square error between the two, i.e. This forces the model to directly optimize the accuracy of GTC, a core clinical measurement indicator, during training.

[0032] Training scheduling module (Warmup): Employs phased weight scheduling. Since complex constraints can easily lead to gradient explosion in the initial stages, this invention exemplarily sets the total number of training rounds to [number missing]. (e.g., 200 rounds). In the first stage (e.g., the first 50 rounds), only basic segmentation losses such as DiceLoss are used, and the constraint weights are dissected at this time. Quantify Consistency Weights In the second stage (e.g., rounds 51 to 100), a linear incremental or cosine annealing strategy is employed to... and Gradually increase to a preset upper limit (e.g., preferred configuration) The upper limit is set to ratio_weight=0.10); in the third stage (round 101 and beyond), the weights are kept at the upper limit and fine-tuned.

[0033] Testing and Performance Validation: Based on the cascaded baseline model, validation was performed by setting a quantization consistency weight parameter (e.g., preferably ratio_weight=0.10). On a test set with relatively uniform data distribution, the model showed a stable improvement trend in area error, thickness error, and Dice index. On a complex test set with more blurred boundaries, while maintaining a relatively stable Dice, the model significantly reduced the outside ratio of anatomical regions, the quantization error converged significantly, and the ClinicalScore was significantly improved. This demonstrates the effectiveness of this invention in shifting the optimization objective from "Dice-oriented only" to "comprehensive orientation".

[0034] Application Example 1.2: Clinical Auxiliary Reasoning and Quantitative Output Process (Online Operation Phase)

[0035] This embodiment illustrates the system's workflow in a clinical setting.

[0036] Input: The breast ultrasound image to be analyzed and its physical resolution information (mm / pixel).

[0037] Processing flow:

[0038] [Step S101] Image acquisition and feature extraction: The system acquires the breast ultrasound image to be analyzed, performs preprocessing such as grayscale standardization, and extracts image features and regions of interest (ROI) based on the physical resolution of the ultrasound image.

[0039] [Step S102] Cascaded Inference: The ultrasound image is input into a pre-trained AI analysis model deployed on a computing device. The model includes a cascaded feature extraction network and a region segmentation network. The first-level network performs coarse localization, and the second-level network performs fine segmentation.

[0040] [Step S103] Output mask and verify anatomical constraints: The model outputs a prediction mask that includes at least the first anatomical structure (i.e., the glandular region) and the second anatomical structure (i.e., the fibroglandular tissue FGT region). Thanks to the anatomical constraints during the training phase, in the prediction mask output by the system, the first anatomical structure is strictly contained within the second anatomical structure in physical space, with no local overflow.

[0041] [Step S104] Quantitative Calculation: Based on the predicted mask and physical resolution, calculate the clinical quantitative indicators for the first and second anatomical structures. Specifically, this includes calculating the actual glandular area, FGT area, maximum thickness, and the percentage of glandular tissue (GTC), which is the core characterizer of breast density.

[0042] [Step S105] Generate quality control prompts: The system calculates the information entropy of the model output probability map in the target boundary region in real time. When the information entropy of the local boundary exceeds the preset uncertainty threshold, quality control prompt information (such as confidence anomaly identifier) ​​is generated based on the local confidence distribution of the prediction mask.

[0043] [Step S106] Interactive output: Overlay the output prediction mask, clinical quantitative indicators (especially GTC value) and quality control prompts on the UI interface to assist doctors in making the final diagnosis.

[0044] Application Example 1.3: Comparative Analysis of Real Clinical Cases

[0045] To further demonstrate the accuracy of the analytical conclusions of this invention and its superiority over traditional clinical methods, the following analysis and comparative verification are conducted using a real case of difficult breast ultrasound (with local acoustic shadowing interference and blurred glandular boundaries).

[0046] Case background: A patient's breast ultrasound image with a physical resolution of 0.08 mm / pixel. Due to the density of the patient's breast and the presence of local artifacts, the boundaries between fibroglandular tissue (FGT) and surrounding adipose tissue, and between glands and FGT, are difficult to distinguish accurately with the naked eye.

[0047] 1. Traditional Clinical Analysis Results (Subjective Visual Assessment): > In traditional routine clinical diagnosis, doctors, limited by time and tools, are usually unable to perform accurate mathematical calculations of the area of ​​anatomical structures. Instead, they rely solely on subjective feelings from visual observation of images to make empirical qualitative gradings of breast density (i.e., a rough reflection of GTC). In this difficult case, senior doctor A assessed it intuitively as "glandular tissue proportion of approximately 40% (high)," while doctor B considered it as "glandular tissue proportion of less than 30% (low)." This traditional method, relying on visual estimation, is highly subjective, suffers from poor consistency among doctors, and cannot output accurate quantitative calculations suitable for structured follow-up.

[0048] 2. Existing AI baseline model analysis results (comparative case): > This image was processed using a traditional CascadedBaseline model without the dual constraints of this invention. Driven only by the pixel-level Dice metric, the model misjudged local acoustic shadowing, predicting that the mask of some glands (first anatomical structures) exceeded the boundary of the FGT (second anatomical structure), violating basic clinical anatomical norms. Furthermore, the model's final output of glandular tissue percentage (GTC) was 41.5%, significantly deviating from the macroscopic perception of experienced physicians. The system did not provide any risk warnings, and directly using it as an auxiliary quantitative indicator could easily mislead subsequent diagnoses.

[0049] 3. System Analysis Results (Precise Quantification and Quality Control):

[0050] (1) Execution steps S101-S103: The system completes automatic parsing within 1 second. Due to the strict implementation of quantization consistency constraints and dissection constraints during the training phase, the prediction mask output by the system perfectly matches the dissection logic, and the Gland region is 100% physically contained within the FGT region, with a dissection violation rate of 0%.

[0051] (2) Step S104: With a resolution of 0.08 mm / pixel, the system instantly achieved a precise measurement that is impossible for humans to complete, objectively calculating the FGT area to be 4.52 square centimeters and the Gland area to be 1.42 square centimeters, and outputting the precise clinical core quantitative indicator: GTC=31.4%. This conclusion not only transforms the traditional doctor's "vague feeling" into "precise objective quantitative data", but the value also perfectly matches the "gold standard" range of multiple experts' meticulous review afterward.

[0052] (3) Execute steps S105-S106: In the area with the most severe sound and shadow interference in the lower right corner of the image, the system calculates that the local prediction information entropy reaches 0.72 (higher than the preset safety threshold). The UI interface immediately highlights the boundary of the segment and outputs a quality control prompt: "Local boundary confidence is low, manual review is recommended".

[0053] Comparative Conclusion: The analysis of the aforementioned real-case examples demonstrates that this embodiment successfully overcomes the pain point of traditional clinical practice, which relies solely on intuition for qualitative grading and cannot objectively calculate GTC. Through distributed step-by-step analysis, this system not only eliminates anatomical common-sense errors (overflow violations) inherent in existing purely data-driven AI but also achieves high-precision automated quantitative output (GTC, etc.), supplemented by quality control prompts for local uncertainties. This proves that this system, as an "AI-assisted quantitative tool," can provide doctors with objective, stable, and secure structured quantitative data that is unavailable through traditional methods, possessing extremely significant clinical application value and beneficial effects.

[0054] Example 2

[0055] This embodiment also provides a breast ultrasound AI-assisted quantitative analysis system, including:

[0056] The data input and parsing module is used to acquire the breast ultrasound image to be analyzed and its physical resolution;

[0057] A cascaded inference module is used to run a pre-trained AI analysis model generated using a training mechanism that employs any one of the features described in claims 1 to 4, and outputs a predicted mask of the anatomical structure.

[0058] The post-processing and quantization module is used to calculate clinical quantitative indicators based on the prediction mask and physical resolution;

[0059] The quality control generation module is used to generate quality control prompts based on the confidence distribution of the reasoning process.

[0060] The interactive output module is used to present prediction masks, clinical quantitative indicators, and quality control prompts to assist in manual confirmation.

[0061] Application Example 2.1:

[0062] Embodiments of the present invention also provide an electronic device (such as a clinical ultrasound workstation or a cloud edge computing server). This electronic device includes a processor, a memory, a communication interface, and a communication bus. The processor, memory, and communication interface communicate with each other via the communication bus. The memory stores a computer program, and when the processor executes the computer program, it implements the breast ultrasound AI-assisted quantitative analysis method described in Embodiments 1 and 2 above. The processor can be a central processing unit (CPU), graphics processing unit (GPU), application-specific integrated circuit (ASIC), or digital signal processor (DSP), etc. The memory may include high-speed random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

Claims

1. A breast ultrasound-assisted analysis method based on anatomical and quantitative consistency constraints, characterized in that, include: Acquire breast ultrasound images to be analyzed, and extract image features based on the physical resolution of the ultrasound images; The ultrasound image is input into a pre-trained AI analysis model deployed in a computing device. The AI ​​analysis model includes a cascaded feature extraction network and a region segmentation network to output a prediction mask that includes at least a first anatomical structure and a second anatomical structure. The first anatomical structure is physically contained within the second anatomical structure. Based on the prediction mask and the physical resolution, calculate clinical quantitative indicators for the first anatomical structure and the second anatomical structure; Based on the local confidence distribution of the prediction mask, quality control prompts are generated, and the prediction mask, the clinical quantitative indicators, and the quality control prompts are output. The total loss function of the pre-trained AI analysis model during the training phase includes an anatomical constraint term and a quantization consistency constraint term. The anatomical constraint term is used to penalize abnormal regions where the predicted boundary of the first anatomical structure exceeds the predicted boundary of the second anatomical structure. The quantization consistency constraint term is used to minimize the error between the soft quantization features predicted by the model and the true labeled quantization features.

2. The method according to claim 1, characterized in that, The first anatomical structure is a glandular region, and the second anatomical structure is a fibroglandular tissue region; The clinical quantitative indicators include at least: the area of ​​the glandular region, the area of ​​the fibroglandular tissue region, the thickness of the glandular region, and the glandular tissue percentage (GTC) which characterizes the density of the breast. The glandular tissue percentage (GTC) is the ratio of the area or volume of the glandular region to the fibroglandular tissue region.

3. The method according to claim 1, characterized in that, The calculation process for the quantization consistency constraint term includes: During the forward propagation of model training, a pixel-level probability map of the anatomical structure is obtained from the network output. Differentiable morphological feature approximation calculations are performed on the pixel-level probability map to obtain the predicted soft area or soft thickness features. Calculate the difference function between the predicted soft area or soft thickness feature and the quantization index corresponding to the actual annotation, and use it as the quantization consistency constraint term.

4. The method according to claim 1, characterized in that, The pre-trained AI analysis model employs a phased weight scheduling strategy during the training phase. In the first training phase, the total loss function is updated with parameters based only on the pixel-level basic segmentation error; In the second training phase, the weights of the anatomical constraint and the quantization consistency constraint in the total loss function are gradually increased according to a preset increment function until a preset upper limit is reached.

5. The method according to claim 1, characterized in that, The generation of quality control prompt information based on the local confidence distribution of the predicted mask includes: Calculate the information entropy of the probability map output by the model in the target boundary region; When the local information entropy of the boundary region exceeds a preset uncertainty threshold, an abnormal confidence indicator is generated to prompt manual review, and is output as the quality control prompt information.

6. The method according to any one of claims 1 to 5, characterized in that, The optimal version of the pre-trained AI analysis model was selected through a comprehensive evaluation of the model using multiple metrics. The selection steps included: On the validation set, independent scores of candidate models were calculated on pixel-level segmentation metrics, absolute error of quantization measurement metrics, and violation rate of dissecting inclusion relationships. The independent scores are weighted and summed based on preset weight coefficients to obtain a comprehensive score, and the candidate model with the highest comprehensive score is selected as the pre-trained AI analysis model for final deployment.

7. A breast ultrasound-assisted analysis system based on anatomical and quantitative consistency constraints, characterized in that, include: The data input and parsing module is used to acquire the breast ultrasound image to be analyzed and its physical resolution; A cascaded inference module is used to run a pre-trained AI analysis model generated using the training mechanism described in any one of claims 1 to 4, and output a predicted mask of the anatomical structure. The post-processing and quantization module is used to calculate clinical quantitative indicators based on the prediction mask and the physical resolution. The quality control generation module is used to generate quality control prompts based on the confidence distribution of the reasoning process. The interactive output module is used to present the prediction mask, clinical quantitative indicators, and quality control prompts to assist in manual confirmation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images