Training method, interpretation method and device of dynamic ultrasonic image interpretation model

Abstract

The invention relates to the technical field related to dynamic ultrasound, in particular to a training method, an interpretation method and a device of a dynamic ultrasound image interpretation model. The method comprises the following steps: acquiring dynamic ultrasonic sample data; annotating the dynamic ultrasonic sample data in combination with expert experience to obtain training data; wherein the annotation comprises a time sequence key point, a dynamic ROI track, a logic chain label and an interpretation data label used for representing output of an expectation model; based on the training data, training a pre-constructed deep learning model to obtain a dynamic ultrasonic image interpretation model; wherein the interpretation data is output of the dynamic ultrasonic image interpretation model based on an expert thinking chain and comprises dynamic change analysis data, and the interpretation data conforms to pathophysiologic phenotype rules defined by experts and diagnostic logic. Therefore, interpretation is carried out based on the model, and the problems of dependence on professional judgment of doctors, large subjectivity influence and unstable interpretation effect are avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of dynamic ultrasound, in particular to a training method and a reading method and device of a dynamic ultrasound image reading model. BACKGROUND

[0002] Critical ultrasound, as an important interpretation method of critical pathophysiology, has been widely used in clinical practice. It can reflect the functional status of important organs such as heart and lung in real time and dynamically, and provide key basis for diagnosis, treatment decision and prognosis evaluation of critical patients.

[0003] At present, the mainstream dynamic ultrasound image reading relies on the professional knowledge and clinical experience of critical doctors. Critical doctors need to undergo systematic ultrasound training to master image acquisition and interpretation skills. However, the existing mode has many limitations. On the one hand, the training of doctors is different, the operation and interpretation are separated, the training focuses on operation skills, and the depth of image interpretation training is insufficient, which leads to the limited recognition ability of doctors to complex lesions, rare signs or special sections. On the other hand, the image interpretation itself is complex, the same ultrasound performance may correspond to multiple causes, and needs to be comprehensively judged in combination with clinical background; some early or slight pathological changes of ultrasound performance are subtle and difficult to identify; the condition of critical patients changes rapidly, and the ultrasound performance changes dynamically, which requires deep pathophysiological knowledge and clinical experience for accurate interpretation.

[0004] The existing critical ultrasound image interpretation method simply relies on the professional judgment of doctors, is greatly affected by subjectivity, and the interpretation effect is unstable. SUMMARY

[0005] Therefore, the embodiments of the present application aim to provide a training method and a reading method and device of a dynamic ultrasound image reading model to interpret dynamic ultrasound, avoid relying on the professional judgment of doctors, and the problem of unstable interpretation effect caused by great subjectivity.

[0006] The present application provides a training method of a dynamic ultrasound image reading model, comprising: acquiring dynamic ultrasound sample data; annotating the dynamic ultrasound sample data in combination with expert experience to obtain training data; wherein the annotation includes: time sequence key points, dynamic ROI track, logic chain label and reading data label for representing expected model output; training a pre-constructed deep learning model based on the training data to obtain a dynamic ultrasound image reading model; wherein the reading data is the output of the dynamic ultrasound image reading model, including dynamic change analysis data, and the reading data conforms to the expert-defined pathophysiological phenotype rule and diagnosis logic.

[0007] In some embodiments, a pre-built deep learning model is trained based on the training data to obtain a dynamic ultrasound image interpretation model, including: Based on the dynamic ultrasound sample data and the timing key points and dynamic ROI trajectories in the corresponding annotations in the training data, the deep learning model is trained so that the deep learning model is used to identify the dynamic ROI trajectories and dynamic change analysis data to obtain a preliminary model; Adding a rule constraint layer to the preliminary model based on expert experience to obtain a target model; The rule constraint layer is used to interpret the dynamic ROI trajectory and dynamic change analysis data to obtain pathological information that meets the expert definition.

[0008] A target model is trained based on the training data to obtain a dynamic ultrasound image interpretation model.

[0009] In some embodiments, it further includes: Incremental learning is performed on the dynamic ultrasound image interpretation model.

[0010] In some embodiments, it further includes: Conduct clinical validation of the dynamic ultrasound image interpretation model; If the verification fails, the dynamic ultrasound image interpretation model is retrained.

[0011] In some embodiments, the dynamic ultrasound sample data further includes: dynamic ultrasound data of the heart, dynamic ultrasound data of the lungs, dynamic ultrasound data of blood vessels, dynamic ultrasound data of the gastrointestinal tract, dynamic ultrasound data of the kidneys, and dynamic ultrasound data of organs such as the brain.

[0012] The present application also provides a training device for a dynamic ultrasound image interpretation model, comprising: An acquisition module, used for acquiring dynamic ultrasound sample data; An annotation module, configured to annotate the dynamic ultrasound sample data based on expert experience to obtain training data; wherein the annotation includes: timing key points, dynamic ROI trajectories, logic chain labels, and interpretation data labels for representing expected model outputs; A training module, configured to train a pre-built deep learning model based on the training data to obtain a dynamic ultrasound image interpretation model; The interpretation data is the output of the dynamic ultrasound image interpretation model, including dynamic change analysis data, and the interpretation data conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic.

[0013] The present application also provides a dynamic ultrasound image interpretation device, comprising: An acquisition module, used for acquiring dynamic ultrasound data; an interpretation module, configured to input the dynamic ultrasound data into a preset dynamic ultrasound image interpretation model to obtain interpretation data; The dynamic ultrasound image interpretation model is obtained by the training method of the dynamic ultrasound image interpretation model as described above, and is used to interpret the dynamic ultrasound data to obtain interpretation data that conforms to expert experience.

[0014] The present application also provides an electronic device, comprising: A processor, and a memory for storing a program executable by the processor; The processor is used to implement the above-mentioned training method of the dynamic ultrasound image interpretation model or the above-mentioned method of interpreting dynamic ultrasound images by running the program in the memory.

[0015] This application provides a method for training a dynamic ultrasound image interpretation model, comprising: obtaining dynamic ultrasound sample data; annotating the dynamic ultrasound sample data based on experience to obtain training data; wherein the annotations include: timing key points, dynamic ROI trajectories, logic chain labels, and interpretation data labels used to represent desired model outputs; and training a pre-built deep learning model based on the training data to obtain a dynamic ultrasound image interpretation model. The interpretation data, which is the output of the dynamic ultrasound image interpretation model, includes dynamic change analysis data, and the interpretation data conforms to expert-defined pathophysiological phenotype rules and diagnostic logic. This arrangement, combined with empirical annotation, provides high-quality training data, ensuring that the model outputs diagnostic results consistent with clinical practice. This captures dynamic physiological processes and improves the ability to recognize subtle changes in dynamic images. By obtaining dynamic change analysis data, and ensuring that the interpretation data conforms to expert-defined pathophysiological phenotype rules and diagnostic logic, model-based interpretation avoids reliance on physicians' professional judgment, which is subject to significant subjectivity and results in unstable interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 It is a flowchart of a training method for a dynamic ultrasound image interpretation model provided in one embodiment of the present application.

[0018] Figure 2is a flowchart of a training method of a dynamic ultrasound image interpretation model provided by an embodiment of the present application.

[0019] Figure 3 is a structural diagram of a dynamic ultrasound analysis device provided by an embodiment of the present application.

[0020] Figure 4 is a structural diagram of a dynamic ultrasound analysis device provided by an embodiment of the present application.

[0021] Figure 5 is a structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the scope of the present application.

[0023] Figure 1 is a training method of a dynamic ultrasound image interpretation model provided by an embodiment of the present application, comprising: Step S101, acquiring dynamic ultrasound sample data; Specifically, the dynamic ultrasound sample data is collected from the image database of a hospital, a clinical research dataset or a sample library provided by an ultrasound equipment manufacturer. These data cover various clinical scenarios, such as dynamic ultrasound data of the heart, the lungs, blood vessels, the gastrointestinal tract, the kidneys and the brain, etc. The dynamic ultrasound sample data is usually stored in the form of a video file, and common file formats include the DICOM (Digital Imaging and Communications) standard format, AVI, MP4, etc.

[0024] The collected dynamic ultrasound sample data is preprocessed, including operations such as removing noise, enhancing image contrast, normalizing image size, etc., to improve data quality and consistency, and to ensure the accuracy and efficiency of subsequent processing.

[0025] Step S102, combining experience, annotating the dynamic ultrasound sample data to obtain training data; wherein the annotation includes: time sequence key points, dynamic ROI trajectories, logical chain labels and interpretation data labels for representing expected model output. Specifically, ultrasound experts with rich clinical experience are invited to participate in the annotation process. These experts have deep professional knowledge and practical experience in critical ultrasound diagnosis, and can accurately identify and interpret various features and pathophysiological information in dynamic ultrasound images.

[0026] Timing key points: Label key time points in dynamic ultrasound videos, such as the start frame of systole and the peak frame of diastole. These key points help the model capture important changes in dynamic physiological processes, such as cardiac contraction and relaxation.

[0027] Dynamic ROI Trajectory: This tool annotates the motion trajectory of regions of interest (ROIs) in dynamic ultrasound videos. For example, it can annotate the motion trajectory of the ventricular endothelium or the motion path of the interventricular septum throughout the cardiac cycle. This helps the model focus on the dynamic changes of key anatomical structures and improves the recognition of subtle pathological features.

[0028] Logical chain labeling: Establishes a logical relationship between dynamic ultrasound image features and clinical diagnosis. For example, labeling the pathophysiological significance of a specific ROI trajectory feature and its relationship with other image features forms a complete diagnostic logic chain.

[0029] Interpretation data labeling: Define the interpretation results expected by the model, including dynamic change analysis data (such as systolic left ventricular ejection fraction, diastolic mitral blood flow velocity, etc.) and corresponding diagnostic conclusions (such as cardiac dysfunction, diastolic dysfunction, etc.), and convert them into a label format that the model can recognize.

[0030] Step S103: training a pre-built deep learning model based on the training data to obtain a dynamic ultrasound image interpretation model; The interpretation data is the output of the dynamic ultrasound image interpretation model, including dynamic change analysis data, and the interpretation data conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic.

[0031] Deep learning models can choose architectures suitable for processing dynamic image data. Common choices include 3D convolutional neural networks (3DCNNs), recurrent neural networks (RNNs), and their variants (such as LSTMs and GRUs). 3DCNNs can effectively extract spatiotemporal features from dynamic ultrasound images, while RNNs excel at processing sequential data and can capture the temporal changes of dynamic ultrasound images.

[0032] Input training data: Annotated dynamic ultrasound sample data is fed into the deep learning model. The model automatically extracts features and establishes mapping relationships by learning from a large amount of training data to accurately interpret dynamic ultrasound images.

[0033] Loss Function Definition: Define an appropriate loss function to measure the difference between the model output and the expected result. For example, for classification tasks, you can use the categorical cross entropy loss function; for regression tasks, you can use the mean squared error loss function. You can also introduce regularization terms to prevent overfitting and improve the model's generalization ability.

[0034] Optimization Algorithm: Optimization algorithms (such as gradient descent and the Adam optimizer) are used to minimize the loss function. Through iterative optimization, the model continuously adjusts its internal parameters, gradually improving its ability to fit the training data and predict new data.

[0035] Model Validation and Adjustment: During training, the model is regularly evaluated using the validation set to monitor performance metrics (such as accuracy, recall, and F1 score). Based on the validation results, the model structure and hyperparameters (such as learning rate and batch size) are adjusted and optimized to improve performance and generalization.

[0036] Model Output and Validation: After training, the resulting dynamic ultrasound image interpretation model is able to analyze and interpret input dynamic ultrasound images, outputting interpretation data that conforms to expert-defined pathophysiological phenotype rules and diagnostic logic. Evaluation on a test set verifies the model's accuracy and reliability, ensuring its effectiveness in real-world clinical applications.

[0037] Finally, the beneficial effects of this method are emphasized again: improving the accuracy of ultrasound image interpretation, enhancing the perception of dynamic images, improving the interpretability of diagnosis, and continuously optimizing through clinical feedback to promote the development of intelligent medical care.

[0038] Specifically, based on the training data, a pre-built deep learning model is trained to obtain a dynamic ultrasound image interpretation model, including: Based on the dynamic ultrasound sample data and the timing key points and dynamic ROI trajectories in the corresponding annotations in the training data, the deep learning model is trained so that the deep learning model is used to identify the dynamic ROI trajectories and dynamic change analysis data to obtain a preliminary model; Adding a rule constraint layer to the preliminary model based on expert experience to obtain a target model; The rule constraint layer is used to interpret the dynamic ROI trajectory and dynamic change analysis data to obtain pathological information that meets the expert definition.

[0039] Specifically, in the solution provided in this application, the model is first trained into a preliminary model that can "observe trajectories and calculate indicators", and then it is transformed into a target model that can "draw conclusions according to clinical standards" through a rule constraint layer. In practical applications, the process of training the deep learning model based on the training data and obtaining the dynamic ultrasound image interpretation model is as follows: Preliminary model training based on temporal key points and dynamic ROI trajectories 1. Input data: The training data includes dynamic ultrasound videos (continuous frames), annotated timing key points (such as the systolic start frame and the diastolic peak frame), and annotated dynamic ROI trajectories (the pixel-level coordinate sequence of the endocardium, ventricular septum, valve edge, etc. in the entire video).

[0040] 2. Model Architecture Selection: Use a deep learning architecture suitable for processing dynamic image data, such as a 3D convolutional neural network (3DCNN). 3DCNN can simultaneously extract both spatial and temporal features of an image, making it suitable for analyzing dynamic ultrasound videos.

[0041] 3. Feature extraction: 3DCNN extracts features from the input dynamic ultrasound video and captures the spatiotemporal features in the video through multi-layer convolution operations, including grayscale changes in the ultrasound image, edge information, and change patterns during the dynamic process.

[0042] 4. Application of attention mechanism: Introducing the attention mechanism into the network enables the model to automatically focus on the key areas indicated by the dynamic ROI trajectory, enhance the ability to learn important features, and suppress interference from irrelevant areas.

[0043] 5. Training Objective: By defining appropriate loss functions, such as the classification cross entropy loss function (used to determine whether the time series key points are accurately identified) and the ROI trajectory regression loss function (used to accurately fit the dynamic ROI trajectory), the model is trained to enable the model to accurately identify dynamic ROI trajectories and obtain a preliminary model.

[0044] Through the above steps, a deep learning model can learn to understand what's happening in a dynamic cardiac ultrasound video. To this end, during the training phase, the network is fed the video itself, along with expert-annotated key time points during each cardiac cycle (e.g., systolic onset, diastolic peak), and the continuous motion trajectories (dynamic ROIs) of structures such as the endocardium, ventricular septum, and valves. Using three-dimensional convolution and an attention mechanism, the network learns both spatial and temporal features, ultimately accurately reproducing these trajectories. It can also calculate dynamic metrics of clinical interest, such as systolic left ventricular ejection fraction (LVEF) and diastolic mitral flow velocity (E / A). This step results in a "preliminary model" that possesses considerable image parsing capabilities but does not yet truly "think like an expert."

[0045] In order to make the model’s conclusions consistent with clinical experience, we added a “rule constraint layer” after the preliminary model to write clinical experience into the network; Specifically, write expert experience into differentiable or executable rules, for example: If LVEF < 50 % → trigger “systolic dysfunction” If E / A > 2 and e' < 8 cm / s → trigger “diastolic dysfunction” If RV end-diastolic area / LV end-diastolic area > 1.0 and IVC > 2 cm → trigger “right heart volume overload” These rules can be implemented using logic gates, differentiable piecewise functions, or knowledge distillation losses.

[0046] The rule-constrained layer essentially converts empirical judgments used by experts in everyday diagnosis, such as "If LVEF < 50%, it indicates systolic dysfunction" and "If E / A > 2 and e' < 8 cm / s, it indicates diastolic dysfunction," into differentiable rule nodes. After the network outputs dynamic metrics, they are immediately checked by these nodes. Any results that do not conform to the rules are penalized and backpropagated, forcing the model to correct them in the next iteration. This fine-tuning step yields the final "target model." It not only produces precise trajectories and metrics but also outputs pathological conclusions that fully adhere to expert definitions, achieving seamless translation from images to clinical language.

[0047] Ultimately, the resulting dynamic ultrasound image interpretation model can accurately interpret the input dynamic ultrasound images and output interpretation data that conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic, providing reliable auxiliary support for clinical diagnosis.

[0048] Furthermore, the solution provided in this application also includes: performing incremental learning on the dynamic ultrasound image interpretation model.

[0049] Specifically, during clinical application, new dynamic ultrasound image data is continuously collected. This data, sourced from different patients, devices, and clinical scenarios, enriches the model's understanding of various situations. The collected new data is preprocessed and annotated to ensure its format and content are consistent with the training data, enabling smooth incremental learning of the model.

[0050] New data is fed into the existing dynamic ultrasound image interpretation model, and the model's parameters are updated using the new data. During this process, the model automatically adjusts its internal parameters, such as weights and biases, based on the features and annotations in the new data to adapt to the distribution and characteristics of the new data. Depending on the actual situation, one can choose online learning, where the model is updated using only one or a few new samples at a time, to quickly adapt to new data. Alternatively, one can choose small-batch learning, where a certain amount of new data is collected periodically and then used to update the model, to better balance learning efficiency and model stability.

[0051] Use the validation set to evaluate the performance of the updated model, focusing on metrics such as precision, recall, and F1 score on new data. Also observe whether the model's performance on the original data has degraded to ensure that incremental learning has not led to catastrophic forgetting. If the model's performance on new data is unsatisfactory or catastrophic forgetting occurs, further adjustments and optimizations are necessary. This may include adjusting the learning rate, adding regularization terms, modifying the model structure, and other measures to improve the model's incremental learning and generalization capabilities.

[0052] When the model achieves satisfactory performance after incremental learning, the updated model is deployed in a real-world clinical setting, replacing the original model and providing doctors with more accurate and reliable assistance in interpreting dynamic ultrasound images. After deployment, the model's performance in real-world applications is continuously monitored, and feedback from doctors and users is collected to promptly identify potential model issues and provide reference for the next incremental learning cycle.

[0053] The solution provided in this application also includes: clinically verifying the dynamic ultrasound image interpretation model; if the verification fails, retraining the dynamic ultrasound image interpretation model.

[0054] Specifically, dynamic ultrasound image data should be collected from real-world clinical scenarios. This data should cover a variety of pathological conditions and clinical scenarios to ensure comprehensive and representative validation results. Furthermore, this data should be accurately annotated and confirmed by experts for comparison with the model's output.

[0055] Determine metrics for evaluating model performance, such as accuracy, recall, F1 value, and area under the receiver operating characteristic (ROC) curve (AUC). These metrics can reflect the effectiveness and reliability of the model in clinical applications from different perspectives. Accuracy measures the model's ability to correctly diagnose, recall focuses on the model's ability to identify actual cases, F1 value comprehensively considers accuracy and recall, and AUC evaluates the model's ability to distinguish between different categories.

[0056] The prepared validation data is fed into the dynamic ultrasound image interpretation model, which then outputs a diagnostic interpretation of the data. The model's output is compared with the expert's diagnostic results, and the validation metrics identified above are calculated to quantify the model's performance.

[0057] If a model fails clinical validation, a detailed analysis of the validation results is required to identify the reasons for the model's performance shortfall. Possible reasons include insufficient training data, uneven data distribution, an irrational model structure, or overfitting or underfitting. For example, if a model has low diagnostic accuracy for certain pathological types, it may be because the training data contained relatively few data of that type, resulting in insufficient model learning of these features.

[0058] Based on the results of the problem analysis, supplement the training data to enrich the model's learning material. If performance issues are caused by insufficient training data, add more diverse dynamic ultrasound image data, especially data from case types where the model performed poorly during validation. Also, relabel and organize the training data to ensure data quality and consistency.

[0059] Retrain the dynamic ultrasound image interpretation model using the supplemented and adjusted training data. During retraining, consider adjusting the model structure or training parameters, such as increasing the model's depth or width, changing the learning rate, or adjusting the regularization term, to improve model performance. For example, if the model is overfitting, increase the weight of the regularization term to make the model smoother and reduce overfitting to the noise in the training data.

[0060] The retrained model is then clinically validated again, and the validation process is repeated until the model's performance meets the requirements for clinical application. This process may require multiple iterations, with continuous optimization and validation of the model to ensure its accuracy and reliability in actual clinical applications.

[0061] Analyze the calculated validation metrics to determine whether the model meets the requirements for clinical application. If the validation metrics meet the predetermined thresholds, such as an accuracy rate exceeding 90% or an AUC exceeding 0.9, the model is considered clinically validated and can be used for actual clinical auxiliary diagnosis. If the validation metrics are unsatisfactory, the model is considered unvalidated.

[0062] Specifically, the dynamic ultrasound sample data may be, but is not limited to, cardiac dynamic ultrasound data.

[0063] The solution provided by this application is described below with reference to specific embodiments: The present invention intends to establish a special model training method, adopt expert thinking chain embedding, model expert interpretation cognitive path, train AI to identify subtle signs and perform dynamic time series analysis, and finally output a standardized report. It also combines clinical information and patient conditions for identification support and feedback learning. At the same time, expert quality control supports selection to establish a competitive learning path for large models and vertical models.

[0064] In the solution of this application, Expert Chain of Thought (Expert CoT) refers to the explicit simulation and recording of an expert's thought process when solving a problem, and its integration into the training process of a dynamic ultrasound image interpretation model. This chain of thought can help the model more accurately understand and interpret the complex features and pathophysiological information in dynamic ultrasound images. The specific content and function of the Expert Chain of Thought are as follows: Expert Chain of Thought is a technology that simulates the thinking process of experts when solving complex problems. It gradually builds a logical chain from problem to answer, enabling AI to more deeply understand the problem and provide solutions. Unlike traditional pattern recognition or statistical learning methods, Expert Chain of Thought emphasizes the transparency and explainability of the reasoning process, making the AI ​​decision-making process closer to the way human experts think. In dynamic ultrasound image interpretation, Expert Chain of Thought can help the model identify and understand subtle features and dynamic changes in the image, thereby more accurately diagnosing and analyzing.

[0065] 1. Dynamic annotation process: Using a standardized interpretation template for dynamic videos, experts will interpret and interpret dynamic images based on the standardized interpretation template to ensure annotation based on the dynamic information of the video and standardization of the interpretation content.

[0066] For example: Step 1: Experts define critical care ultrasound pathophysiology standard fields Experts develop standard fields for different critical ultrasound pathophysiological abnormalities and conduct training to identify them, such as: 1. Systolic left ventricular ejection fraction (LVEF) <50% → triggers the conclusion of "heart function impairment"; 2. Diastolic mitral blood flow velocity E / A ratio > 2 → triggers the conclusion of "diastolic dysfunction and increased left atrial pressure"; 3. Abnormal ROI dynamic trajectory (e.g., ventricular septal motion presenting a central concave pattern) → triggers the conclusion of "high-pressure right ventricular dysfunction."

[0067] Step 2: Dynamic ROI and label annotation Experts marked in the ultrasound video: 1. Timing key points (such as the start frame of cardiac systole and the peak frame of diastole); 2. Dynamic ROI trajectory (e.g. full-motion cycle trajectory of the endocardium, continuous coordinate points of the septal motion path, dynamic lung ultrasound sign changes such as tidal recruitment, dynamic bronchial inflation sign): The advantage is that it can avoid the interference of single-frame image artifacts, and the position coordinates of each frame image can be derived from the clear edge trajectory image of the front and back time sequence and then connected to form the motion characteristics. The second point is that the trajectory based on the full cardiac cycle has different characteristics under different pathophysiological conditions, which can be used for learning based on these characteristics. 3. Logical chain label (binding relationship between conclusion and ROI / time sequence, e.g. pathophysiological characteristics represented by the motion trajectory characteristics of the continuous coordinate points of the septal motion path, and the relationship between the motion of the septal and the systolic / diastolic period to determine the severity of pathophysiology).

[0068] The annotation result contains spatio-temporal joint information, not independent frame static annotation, which can avoid local image blur and artifact interference, and the spatio-temporal joint information can prompt special pathophysiological characteristics.

[0069] Step 3: Expert defines pathophysiological phenotype rules and diagnostic logic The expert formulates a dynamic diagnostic logic chain for the target sign (e.g. shock), for example: 1. The diameter of the vena cava is <1.5 cm, and the vena cava is short-axis water droplet / linear, and the right heart to left heart diastolic area ratio is <0.6, and the left heart systolic function is >50%, and CO is <4→ triggers the conclusion of "left-right heart matching, low blood volume, left heart high power, low output"; 2. The right heart to left heart diastolic area ratio is >1, and the vena cava diameter is >2 cm, and the vena cava is short-axis circular, and the left heart systolic function is >50%, and CO is <4→ triggers the conclusion of "left-right heart mismatch, right heart volume overload, left heart low volume and low output"; 3. The diameter of the vena cava is >2 cm, and the right heart to left heart diastolic area ratio is <0.6, and the left ventricular ejection fraction (LVEF) is <50%, and CO is <4→ triggers the conclusion of "left-right heart matching, volume overload, left heart systolic function decline, low output";

[0070] 4. The diameter of the vena cava is >1.5 cm, and the right heart to left heart diastolic area ratio is <0.6, and the left ventricular ejection fraction (LVEF) is >50%, and CO is >6, and the snuffbox presents a low-tension spectrum / resistance index reduction→ triggers the conclusion of "left-right heart matching, volume not low, left heart systolic function not poor, non-low output, low tension";

[0071] Step 4: Clinical post-feedback learning The initially trained model is trained on clinical cases, and the conclusions after judgment are fed back to the clinic. Clinicians and experts comprehensively evaluate whether it is consistent or not, and combine the targeted treatment given and indicators such as effects and outcomes to feedback the degree of compliance of the initial judgment, so as to conduct post-clinical feedback learning in real scenarios.

[0072] For example, for target symptoms (such as shock), the model has triggered the conclusion of "left-right heart matching, low blood volume, left heart hyperdynamics, and low output" based on the following conditions: "vena cava diameter <1.5cm, vena cava short-axis water drop / linearity, right heart to left heart end-diastolic area ratio <0.6, left heart systolic function >50%, and CO <4". Clinicians will judge and correct the details, such as whether the phenotype is clinically consistent or not. If it is not clinically consistent, the phenotypic characteristics will be divided into different short fields, and the non-compliant fields will be selected for reverse correction, and the correction content will be recorded and learned. If it is clinically consistent, the treatment effect will continue to be observed to see if it meets expectations. If it meets expectations, intensive learning will be carried out, and if it does not meet expectations, the specific content that needs to be corrected will be reviewed.

[0073] 2. The model training process includes: Phase 1: ROI conclusion association pre-training phase: The input of the model in this stage is: dynamic ultrasound video + expert-annotated ROI trajectory + conclusion label; the model network is 3DCNN to extract spatiotemporal features → the attention mechanism focuses on the ROI area → the classification layer outputs the conclusion probability; the model loss function includes: classification cross entropy loss + ROI trajectory regression loss (L1Loss) Stage 2: Expert Rule Knowledge Distillation The inputs to this stage of the model are: the inputs from stage 1 and the expert rules (e.g., "if ROI_A expansion rate > X%, then weight + α"). The network of this stage of the model: Based on the stage 1 model, a rule constraint layer (e.g., a logic gate module) is added. The loss function of this stage of the model includes: adding a penalty term for rule violation (e.g., doubling the loss value when the rules conflict). Phase 3: Incremental Learning and Clinical Validation Dynamic update: New case data, after being reviewed by experts, triggers incremental model training; Verification mechanism: The Kappa consistency coefficient between the model output conclusion and the expert diagnosis must be ≥0.9.

[0074] Further, refer to Figure 2 , the present application provides a method for interpreting dynamic ultrasound images, comprising: Step S201, acquiring dynamic ultrasound data; Dynamic ultrasound data were collected from hospital imaging databases, clinical research datasets, or sample libraries provided by ultrasound equipment manufacturers.

[0075] Dynamic ultrasound data is usually stored in the form of video files. Common file formats include DICOM, AVI, MP4, etc.

[0076] The collected dynamic ultrasound data are preprocessed, including noise removal, image contrast enhancement, image size normalization and other operations, to improve data quality and consistency and ensure the accuracy and efficiency of subsequent processing.

[0077] Step S202: inputting the dynamic ultrasound data into a preset dynamic ultrasound image interpretation model to obtain interpretation data; The dynamic ultrasound image interpretation model is obtained through the above-mentioned training method of the dynamic ultrasound image interpretation model, and is used to interpret the dynamic ultrasound data to obtain interpretation data that conforms to the expert's thinking chain.

[0078] Specifically, first ensure that a dynamic ultrasound image interpretation model, obtained through the above training method, is available. This model, trained on professionally labeled data, is capable of processing dynamic ultrasound data and outputting interpretation results that are consistent with the expert's thought chain.

[0079] The preprocessed dynamic ultrasound data were input into the model.

[0080] The model analyzes and interprets the input dynamic ultrasound data, and outputs corresponding interpretation data based on the knowledge and logic learned during its training process.

[0081] Model output interpretation data: This interpretation data includes dynamic change analysis data that conforms to expert-defined pathophysiological phenotype rules and diagnostic logic. Specifically, this data may involve analysis of dynamic changes in different parts of the image, identification and judgment of specific pathological features, etc., providing doctors with valuable diagnostic reference information.

[0082] Through the above steps, the trained dynamic ultrasound image interpretation model can be used to accurately interpret the actual dynamic ultrasound data, assisting doctors in making diagnosis and treatment decisions.

[0083] Specifically, the dynamic ultrasound data includes: dynamic cardiac ultrasound data; The interpretation data includes: diagnostic index data and diagnostic conclusions corresponding to the diagnostic index data; The diagnostic indicator data at least includes dynamic change analysis data; The dynamic change analysis data includes at least one of: systolic left ventricular ejection fraction, diastolic mitral valve blood flow velocity, and ROI dynamic trajectory.

[0084] Specifically, dynamic ultrasound data can be, but is not limited to, dynamic cardiac ultrasound data; dynamic cardiac ultrasound data consists of a series of continuous ultrasound image frames, usually recorded in video form, reflecting the morphological and motion changes of organs such as the heart during dynamic processes. Each frame of the image contains a pixel value matrix, which represents the reflection intensity of the tissue interface and can be used to distinguish different tissue structures, such as the myocardium, cardiac chambers, valves, etc. Each frame of the image has a corresponding timestamp, which accurately records the acquisition time of the image during the dynamic process, which helps to analyze the time characteristics of dynamic physiological processes such as the heart's contraction and relaxation cycle. It includes dynamic cardiac ultrasound images of different sections, such as the long axis section, short axis section, four-chamber heart section, etc. Different sections can show the morphology and movement of various parts of the heart, providing a multi-angle perspective for a comprehensive assessment of cardiac function.

[0085] The diagnostic indicator data in the interpretation data include: Dynamic change analysis data: including systolic left ventricular ejection fraction, diastolic mitral valve blood flow velocity, ROI dynamic trajectory, etc.

[0086] Other diagnosis-related data: In addition to the above-mentioned dynamic change analysis data, it may also include other diagnosis-related features and indicators extracted from dynamic ultrasound data, such as heart rate, heart rhythm, size and volume of each chamber of the heart, etc.

[0087] Diagnostic conclusions are based on the analysis of diagnostic indicator data, such as cardiac function assessment (whether left ventricular systolic function is normal, whether diastolic function is impaired, etc.), valvular function assessment (whether the heart valve is stenotic or insufficiency, etc.), myocardial disease diagnosis (whether myocardial ischemia, infarction, or cardiomyopathy, etc.), heart disease diagnosis (such as coronary heart disease, dilated cardiomyopathy, rheumatic heart valve disease, etc.), disease severity assessment (such as heart disease stage, disease severity, etc.), and treatment recommendations (drug treatment plan, surgical treatment indications, etc.).

[0088] The device embodiments of this application can be used to execute the method embodiments of this application. For details not disclosed in the device embodiments of this application, please refer to the method embodiments of this application.

[0089] Figure 3 FIG. 1 is a block diagram of a training device for a dynamic ultrasound image interpretation model provided by an embodiment of the present application. Figure 3 As shown, the device includes: An acquisition module 31 is used to acquire dynamic ultrasound sample data; Annotation module 32 is used to annotate the dynamic ultrasound sample data in combination with expert experience to obtain training data; wherein the annotation includes: timing key points, dynamic ROI trajectories, logic chain labels and interpretation data labels for representing the expected model output; A training module 33 is configured to train a pre-built deep learning model based on the training data to obtain a dynamic ultrasound image interpretation model; The interpretation data is the output of the dynamic ultrasound image interpretation model, including dynamic change analysis data, and the interpretation data conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic.

[0090] Figure 4 FIG. 1 is a block diagram of a dynamic ultrasound image interpretation device provided by an embodiment of the present application. Figure 4 As shown, the device includes: An acquisition module 41 is used to acquire dynamic ultrasound data; An interpretation module 42 is used to input the dynamic ultrasound data into a preset dynamic ultrasound image interpretation model to obtain interpretation data; The dynamic ultrasound image interpretation model is obtained through the training method of the dynamic ultrasound image interpretation model described above, and is used to interpret the dynamic ultrasound data to obtain interpretation data that conforms to the expert's thinking chain.

[0091] Below, reference Figure 5 To describe the electronic device according to the embodiment of the present application. Figure 5 The figure shows a block diagram of an electronic device according to an embodiment of the present application.

[0092] like Figure 5 As shown, electronic device 500 includes one or more processors 510 and memory 520 .

[0093] The processor 510 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 500 to perform desired functions.

[0094] The memory 520 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 510 may execute the program instructions to implement the dynamic ultrasound image interpretation model training method, dynamic ultrasound image interpretation method, and / or other desired functions described in the various embodiments of the present application described above. The computer-readable storage medium may also store various contents, such as category correspondences.

[0095] In one example, the electronic device 500 may further include an input device 530 and an output device 540 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0096] In addition, the input device 530 may also include, for example, a keyboard, a mouse, an interface, etc. The output device 540 may output various information to the outside, including analysis results, etc. The output device 540 may include, for example, a display, a speaker, a printer, a communication network and its connected remote output device, etc.

[0097] Of course, to simplify, Figure 5 Only some of the components in the electronic device related to the present application are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, the electronic device may further include any other appropriate components according to specific application scenarios.

[0098] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps in the training method of the dynamic ultrasound image interpretation model or the dynamic ultrasound image interpretation method according to various embodiments of the present application described in the above-mentioned "Exemplary Method" section of this specification.

[0099] The computer program product may be written in any combination of one or more programming languages ​​to implement the program code for performing the operations of the embodiments of the present application, including object-oriented programming languages ​​such as Java, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0100] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, enables the processor to execute the steps of the training method of the dynamic ultrasound image interpretation model or the dynamic ultrasound image interpretation method according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0101] The computer-readable storage medium may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0102] The above description has been provided for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A training method for a dynamic ultrasound image interpretation model, characterized in that: include: Acquire dynamic ultrasound sample data; In combination with expert experience, the dynamic ultrasound sample data is annotated to obtain training data; wherein the annotations include: timing key points, dynamic ROI trajectories, logic chain labels, and interpretation data labels for representing expected model outputs; Based on the training data, a pre-built deep learning model is trained to obtain a dynamic ultrasound image interpretation model; The interpretation data is the output of the dynamic ultrasound image interpretation model, including dynamic change analysis data, and the interpretation data conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic.

2. The training method of the dynamic ultrasound image interpretation model according to claim 1, characterized in that: Based on the training data, a pre-built deep learning model is trained to obtain a dynamic ultrasound image interpretation model, including: Based on the dynamic ultrasound sample data and the timing key points and dynamic ROI trajectories in the corresponding annotations in the training data, the deep learning model is trained so that the deep learning model is used to identify the dynamic ROI trajectories and dynamic change analysis data to obtain a preliminary model; Adding a rule constraint layer to the preliminary model based on expert experience to obtain a target model; The rule constraint layer is used to interpret the dynamic ROI trajectory and dynamic change analysis data to obtain pathological information that meets the expert definition; A target model is trained based on the training data to obtain a dynamic ultrasound image interpretation model.

3. The training method of the dynamic ultrasound image interpretation model according to claim 1, characterized in that: Also includes: Incremental learning is performed on the dynamic ultrasound image interpretation model.

4. The training method of the dynamic ultrasound image interpretation model according to claim 1, characterized in that: Also includes: Conduct clinical validation of the dynamic ultrasound image interpretation model; If the verification fails, the dynamic ultrasound image interpretation model is retrained.

5. The training method of the dynamic ultrasound image interpretation model according to claim 1, characterized in that: Also includes: The dynamic ultrasound sample data includes cardiac dynamic ultrasound data, lung dynamic ultrasound data, blood vessel dynamic ultrasound data, gastrointestinal tract dynamic ultrasound data, kidney dynamic ultrasound data and cranial dynamic ultrasound data.

6. A method for interpreting dynamic ultrasound images, characterized in that: include: Acquire dynamic ultrasound data; Inputting the dynamic ultrasound data into a preset dynamic ultrasound image interpretation model to obtain interpretation data; The dynamic ultrasound image interpretation model is obtained by the training method of the dynamic ultrasound image interpretation model according to any one of claims 1 to 5, and is used to interpret the dynamic ultrasound data to obtain interpretation data that conforms to the expert thinking chain.

7. The method for interpreting dynamic ultrasound images according to claim 6, wherein: The dynamic ultrasound data includes: dynamic ultrasound data of the heart, dynamic ultrasound data of the lungs, dynamic ultrasound data of blood vessels, dynamic ultrasound data of the gastrointestinal tract, dynamic ultrasound data of the kidneys and dynamic ultrasound data of the brain; The interpretation data includes: diagnostic index data and diagnostic conclusions corresponding to the diagnostic index data; The diagnostic indicator data at least includes dynamic change analysis data; The dynamic change analysis data includes at least one of: systolic left ventricular ejection fraction, diastolic mitral valve blood flow velocity, and ROI dynamic trajectory.

8. A training device for a dynamic ultrasound image interpretation model, characterized in that: include: An acquisition module, used for acquiring dynamic ultrasound sample data; An annotation module, configured to annotate the dynamic ultrasound sample data based on expert experience to obtain training data; wherein the annotation includes: timing key points, dynamic ROI trajectories, logic chain labels, and interpretation data labels for representing expected model outputs; A training module, configured to train a pre-built deep learning model based on the training data to obtain a dynamic ultrasound image interpretation model; The interpretation data is the output of the dynamic ultrasound image interpretation model, including dynamic change analysis data, and the interpretation data conforms to the expert-defined pathophysiological phenotype rules and diagnostic logic.

9. A dynamic ultrasound image interpretation device, characterized in that: include: An acquisition module, used for acquiring dynamic ultrasound data; an interpretation module, configured to input the dynamic ultrasound data into a preset dynamic ultrasound image interpretation model to obtain interpretation data; The dynamic ultrasound image interpretation model is obtained by the training method of the dynamic ultrasound image interpretation model according to any one of claims 1 to 5, and is used to interpret the dynamic ultrasound data to obtain interpretation data that conforms to the expert thinking chain.

10. An electronic device, characterized in that: include: A processor, and a memory for storing a program executable by the processor; The processor is configured to implement the training method for the dynamic ultrasound image interpretation model according to any one of claims 1 to 5, or the dynamic ultrasound image interpretation method according to claim 6 or 7, by running the program in the memory.

Images