Dynamic ultrasonic noise reduction model training method and dynamic ultrasonic noise reduction method and device

Abstract

The invention relates to the technical field of medical image processing, in particular to a dynamic ultrasonic noise reduction model training method and a dynamic ultrasonic noise reduction method and device. The method comprises the following steps: acquiring an ultrasonic video covering multiple clinical scenes such as antral pneumatosis, hydrops, solid contents, long antral antrum and cardiac anatomical occlusion; marking the precise boundary of the organ frame by frame by a clinical expert, segmenting and complementing a missing section, optimizing the contour of a fuzzy region, and inhibiting or erasing artifacts; a deep learning model is trained based on the labeled data, so that the model can call an exclusive noise reduction algorithm according to organ categories (the antral sinuses adopt morphological reconstruction filtering guided by ellipse / long axis prior, and the heart adopts time-space joint deblurring driven by a cardiac cycle and combining motion compensation and anisotropic filtering); and weighted noise reduction is carried out on the suspected lesion or the user indication area. In the reasoning stage, a high-fidelity and low-noise video stream can be output in real time by inputting any dynamic ultrasonic video, and organ boundary metadata is attached.

Description

Technical Field

[0001] The present application relates to the technical field related to medical image processing, and specifically to a training method for a dynamic ultrasound noise reduction model, a dynamic ultrasound noise reduction method and a device. Background Art

[0002] Dynamic ultrasound, with its radiation-free nature and real-time bedside imaging, has been widely used in scenarios such as gastric antral motility assessment and cardiac function monitoring. However, due to factors such as patient size, probe movement, gastrointestinal gas, obstruction by the bones surrounding the heart, and lung gas, ultrasound images often exhibit the following issues: blurred or partially missing boundaries; high-brightness / low-brightness artifacts; and low signal-to-noise ratios that make lesions difficult to identify. Summary of the Invention

[0003] In view of this, the embodiments of the present application are committed to providing a training method for a dynamic ultrasound noise reduction model, a dynamic ultrasound noise reduction method and a device to solve the following problems that often occur in ultrasound videos in the prior art: blurred or partially missing boundaries; high-brightness / low-brightness artifacts; and difficulty in identifying lesion areas due to low signal-to-noise ratio.

[0004] This application provides a training method for a dynamic ultrasonic noise reduction model, including: Obtaining a preset number of sample videos, the sample videos including ultrasound videos in different clinical scenarios; the different clinical scenarios including antral gas, antral fluid, solid antral contents, elongated antral gastric contents, and occlusion of anatomical structures around the heart; Annotating the sample video, wherein the annotation is used to indicate: boundaries of the subject organs in the sample video, filling of boundaries of incomplete images of the subject organs, optimization strategy of blurred images of the subject organs, and erasure of artifact images of the subject organs; Based on the sample videos and their corresponding annotations, a pre-built deep learning model is trained to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of incomplete images of the main organs, optimize the image based on the optimization strategy corresponding to the blurred image of the main organs, and erase the artifact images of the main organs to obtain the dynamic ultrasound video after noise reduction.

[0005] In some embodiments, the optimization strategy corresponding to the blurred image of the subject's organs includes: a noise reduction strategy for each subject's organ; The noise reduction strategy for each organ of the subject includes: When the subject organ is the gastric antrum, the noise reduction strategy set includes morphological reconstruction filtering based on the gastric antrum ellipse / major axis prior, When the subject organ is a heart, the noise reduction strategy set includes a spatiotemporal joint deblurring process based on a cardiac cycle phase.

[0006] In some embodiments, the optimization strategy corresponding to the blurred image of the subject's organs includes: When a suspected lesion region in a subject's organ is identified, a higher weight of noise reduction is performed on the suspected lesion region.

[0007] In some embodiments, further comprising: Adding indication information to the sample video, where the indication information is used to indicate a region of interest; Annotating the sample video to which the indication information is added, wherein the annotation is used to indicate that a higher weight of noise reduction is performed on the region of interest; Based on the sample video with added indication information and its corresponding annotations, the dynamic ultrasound noise reduction model is trained so that the dynamic ultrasound noise reduction model is used to input the dynamic ultrasound video with added indication information, perform higher-weighted noise reduction on the region of interest indicated by the indication information, and output the noise-reduced dynamic ultrasound video.

[0008] In some embodiments, further comprising: The dynamic ultrasound video after noise reduction includes indication information of the boundaries of the subject's organs.

[0009] The present application provides a dynamic ultrasonic noise reduction method, comprising: Acquire dynamic ultrasound video; Inputting the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; The dynamic ultrasonic noise reduction model is pre-trained using the above-mentioned training method for the dynamic ultrasonic noise reduction model.

[0010] This application provides a training device for a dynamic ultrasonic noise reduction model, comprising: A sample acquisition module is configured to acquire a preset number of sample videos, the sample videos including ultrasound videos in different clinical scenarios; the different clinical scenarios including antral gas, antral fluid, solid antral contents, elongated antral gastric contents, and occlusion of anatomical structures surrounding the heart; a labeling module, configured to label the sample video, wherein the labeling is used to indicate: boundaries of subject organs in the sample video, filling of boundaries of incomplete images of subject organs, optimization strategy of blurred images of subject organs, and erasure of artifact images of subject organs; A training module, configured to train a pre-built deep learning model based on the sample video and its corresponding annotations to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of incomplete images of the main organs, optimize the image based on the optimization strategy corresponding to the blurred image of the main organs, and erase the artifact images of the main organs to obtain the dynamic ultrasound video after noise reduction.

[0011] The present application provides a dynamic ultrasonic noise reduction device, comprising: An acquisition module, used for acquiring dynamic ultrasound video; a noise reduction module, configured to input the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; The dynamic ultrasonic noise reduction model is pre-trained using the above-mentioned training method for the dynamic ultrasonic noise reduction model.

[0012] The present application provides an electronic device, including: A processor, and a memory for storing a program executable by the processor; The processor is used to implement the above-mentioned dynamic ultrasonic noise reduction model training method or the above-mentioned dynamic ultrasonic noise reduction method by running the program in the memory.

[0013] The present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor executes the training method of the dynamic ultrasonic noise reduction model or the dynamic ultrasonic noise reduction method.

[0014] The present application provides a training method for a dynamic ultrasound noise reduction model, comprising: obtaining a preset number of sample videos, wherein the sample videos include: ultrasound videos under different clinical scenarios; the different clinical scenarios include gastric antral gas, gastric antral fluid, solid gastric antral contents, a long linear gastric antrum, and occlusion of anatomical structures around the heart; annotating the sample videos, wherein the annotations are used to indicate: the boundaries of the main organs in the sample videos, the filling of the boundaries of incomplete images of the main organs, the optimization strategy of the blurred images of the main organs, and the erasure of artifact images of the main organs; based on the sample videos and their corresponding annotations, training a pre-constructed deep learning model to obtain a dynamic ultrasound noise reduction model; the dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of the incomplete images of the main organs, perform image optimization based on the optimization strategy corresponding to the blurred images of the main organs, and erase the artifact images of the main organs to obtain a dynamic ultrasound video after noise reduction. This setup allows for joint training of samples from diverse clinical scenarios (e.g., antral gas, effusion, solid contents, elongated antrum, cardiac anatomy occlusion, etc.) along with corresponding annotations. The model automatically identifies the current scene category based on the input video and applies the most appropriate noise reduction strategy, maintaining robustness in complex and changing clinical environments. The annotations also include four-dimensional information: organ boundaries, incomplete boundary completion, blur optimization, and artifact removal. The trained model prioritizes the preservation of true anatomical structures during the denoising process. Compared to existing technologies, edge localization error is reduced by ≥30%, effectively avoiding the "oversmoothing" and "false boundaries" caused by traditional filtering. All noise reduction steps are integrated into the same deep learning framework and can be fine-tuned online or offline with new samples and annotations. Hospitals can continuously improve model performance as data accumulates without hardware modifications, significantly reducing ongoing maintenance costs. Specifically, this dynamic ultrasound noise reduction model addresses existing ultrasound video challenges: blurred or partially missing boundaries; high-brightness / low-brightness artifacts; and difficulty identifying lesions due to low signal-to-noise ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] 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.

[0016] Figure 1 It is a flowchart of a training method for a dynamic ultrasonic noise reduction model provided in one embodiment of the present application.

[0017] Figure 2It is a flow chart of a dynamic ultrasonic noise reduction method provided in one embodiment of the present application.

[0018] Figure 3 It is a structural diagram of a training device for a dynamic ultrasonic noise reduction model provided in one embodiment of the present application.

[0019] Figure 4 It is a structural schematic diagram of a dynamic ultrasonic noise reduction device provided in one embodiment of the present application.

[0020] Figure 5 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] This application provides a comprehensive adaptive noise reduction solution for dynamic ultrasound imaging, aiming to overcome the shortcomings of existing technologies in complex clinical environments, such as unstable imaging quality, loss of anatomical details, and insufficient noise reduction in lesion areas. This solution is designed around the principle of "unified training, scenario adaptation, organ-specific, and lesion-prioritized," encompassing the entire chain from data acquisition, model training, inference deployment, and subsequent application.

[0023] First, at the data acquisition level, dynamic ultrasound videos covering various clinical scenarios are systematically collected, including at least antral gas, fluid accumulation, solid contents, elongated antrum, and occlusion of anatomical structures around the heart. With the help of experienced ultrasound physicians, each frame of the image is annotated in four dimensions, namely organ boundaries, missing boundary filling parameters, fuzzy optimization parameters, and artifact erasure parameters, to build a high-fidelity and scalable annotation system.

[0024] Secondly, at the model training level, based on the aforementioned labeled data, an end-to-end deep learning framework is used to preset differentiated noise reduction strategies for different organ types: For the gastric antrum, a morphological reconstruction filter with ellipse / major axis prior is used to constrain the reconstruction process with statistical anatomical proportions to ensure that the edge direction is consistent with the physiological structure. For the heart area, a spatiotemporal joint deblurring mechanism is constructed based on the phase of the cardiac cycle, and the fine structure blurred by motion is restored through motion compensation and anisotropic deconvolution.

[0025] At the same time, the system introduces a weighted loss function to give higher weights to suspected lesion areas or user-specified regions of interest, thereby prioritizing the enhancement of lesion details and avoiding the information loss caused by the "one-size-fits-all" smoothing of traditional methods.

[0026] Ultimately, the trained dynamic ultrasound noise reduction model was packaged as a lightweight inference engine deployable across multiple platforms. It supports seamless integration with ultrasound hosts, edge computing boxes, and cloud servers, with latency controlled to under 40 milliseconds. It also outputs a denoised video stream with organ boundary metadata in real time. Downstream AI algorithms or clinicians can directly access this metadata without the need for secondary annotation, significantly reducing review workload and improving diagnostic consistency.

[0027] In summary, this application achieves high-fidelity noise reduction of dynamic ultrasound images through scene adaptation, organ-specific strategies and lesion weighting mechanisms while ensuring real-time performance, laying a solid data foundation for subsequent intelligent analysis and precise diagnosis and treatment.

[0028] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0029] Figure 1 FIG. 1 is a flow chart of a training method for a dynamic ultrasonic noise reduction model provided by an embodiment of the present application. Figure 1 As shown, the method includes the following contents.

[0030] S101, obtaining a preset number of sample videos, the sample videos including ultrasound videos in different clinical scenarios; the different clinical scenarios including antral gas, antral effusion, solid antral contents, elongated antral gastric contents, and occlusion of anatomical structures around the heart; Specifically, the sample videos are in lossless DICOM format or demodulated B-mode images of raw RF data, with a frame rate of 15–60 fps, pixel resolution ranging from 640×480 to 1920×1080, and probe frequency band covering 2–12 MHz to ensure the generalization ability of subsequent models.

[0031] To fully simulate common clinical degradation patterns, the following five representative scenarios are preset: a) Antral gas: The gastric cavity contains a large amount of highly echogenic gas with significant acoustic shadowing behind it, resulting in complete or partial loss of the posterior wall of the gastric antrum; b) Gastric antral effusion: The gastric cavity presents an anechoic fluid dark area, and far-field enhancement leads to decreased wall echo contrast; c) The contents of the antrum are solid: the chyme is mixed echogenic, the interface echo is chaotic, and the wall boundary is discontinuous; d) Long linear antrum: more common in patients with slender body, the antrum is "tubular" in short axis section rather than the typical ellipse; e) Obstruction of the heart by surrounding anatomical structures: rib shadowing, lung air obstruction, or obese chest wall fat leading to partial non-visibility of the endocardium and epicardium.

[0032] Each type of scene is divided into training set, validation set and test set in a ratio of 7:2:1; the final sample videos included are no less than 12,000, with a total frame number of more than 1.5×10 7 frames, ensuring that the deep learning model can learn from enough cases.

[0033] Of course, in actual applications, different clinical scenarios can also be selected based on actual needs.

[0034] S102, annotating the sample video, wherein the annotation is used to indicate: boundaries of the subject organs in the sample video, filling of boundaries of incomplete images of the subject organs, optimization strategy of blurred images of the subject organs, and erasure of artifact images of the subject organs; Combining specific expert experience and the actual needs of doctors, sample videos are annotated so that the sample videos output by the trained model are standard videos that meet the needs of experts.

[0035] It should be noted that the specific marking forms are various. In some embodiments, the marking method can be as follows: We perform four layers of structured annotation on sample videos, corresponding to "boundary-filling-optimization-erasure." The annotation results are persisted in JSON format, with a confidence field attached for subsequent training and backtracking.

[0036] Organ Boundary Annotation: This uses a WebGL-based interactive annotation platform, supporting frame-by-frame polygon drawing and vertex dragging. Specific rules include: Antrum: Continuously trace along the mucosal-serosa echo interface to form a closed polygon; Heart: tracing the endocardium and epicardium separately, allowing for layered storage. Quality Control: Edge pixel error ≤ 1.5 mm (at a resolution of 0.3 mm / pixel).

[0037] Boundary Completion triggers: A contour discontinuity greater than 3 mm or an edge signal-to-noise ratio less than 10 dB is marked as "missing." The annotator places start and end points on either side of the gap, and the system automatically generates a cubic spline interpolation curve. The annotator can then fine-tune the curve by dragging the control points until it aligns with the ultrasound anatomy.

[0038] Metadata: records the starting coordinates, interpolation type, and curvature constraint level (Loose / Normal / Tight).

[0039] Blur Optimization strategy annotation area definition: Use alpha mask to mark the blur band, covering the wall layer or endocardial area; Parameter level: Provides sharpening strength suggestions between 1.0–2.5 in 0.5 steps, corresponding to mild, moderate, and severe blur; Auxiliary reference: The platform also displays the local gradient histogram for quantitative verification by the annotator.

[0040] Artifact Removal: Artifact categories include: highlight shadows, specular reflections, side lobes, and reverberation. Labeling method: binary mask (1 to erase, 0 to retain), allowing for multiple labels. Boundary feathering: 3-pixel Gaussian feathering is automatically applied to the mask edge to prevent hard cuts.

[0041] Specifically, a method of two people independently marking + third-party arbitration can be adopted to ensure consistency in marking.

[0042] Generate an independent JSON file for each frame, with fields including frame_id, polygon, completion_curve, blur_mask, artifact_mask, annotator_id, and timestamp; All labeled files are stored in the same path as the video frames and are encrypted and signed to ensure that the training phase is traceable and auditable.

[0043] S103, training a pre-built deep learning model based on the sample video and its corresponding annotations to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of incomplete images of the main organs, optimize the image based on the optimization strategy corresponding to the blurred image of the main organs, and erase the artifact images of the main organs to obtain the dynamic ultrasound video after noise reduction.

[0044] Specifically, the dynamic ultrasound noise reduction model adopts an encoder-decoder structure as a whole, including: encoder: ResNet-50 backbone, extracting multi-scale features; decoder: including five branches: organ classification branch: Softmax outputs organ category; boundary detection branch: Sigmoid outputs boundary probability map; missing boundary filling branch: U-Net++ structure, outputs continuous filling curve; deblurring branch: Residual Dense Net, outputs residual sharpening map; artifact erasure branch: U-Net, outputs binary erasure mask.

[0045] The total loss in the loss function can be the weighted sum of the data corresponding to the five branches: the weights of each part can be searched by Bayesian optimization on the validation set to obtain the optimal combination.

[0046] The model's inference process includes: Input dynamic ultrasound video, classify organs frame by frame, call the corresponding strategy set, perform boundary detection / padding / deblurring / erasing in parallel, and output the denoised video stream after fusion; The following functions can be achieved through the above training dynamic ultrasonic noise reduction model: The model performs the following functions during the inference phase: a) Organ determination: quickly distinguish between the gastric antrum and the heart based on the Softmax output; b) Boundary recognition: Output probability map, then binarize and vectorize; c) Gap filling: cubic spline interpolation is performed on the gaps in the probability map to maintain physiological curvature; d) Image optimization: loading sharpening intensity α or phase-driven deblurring kernel according to organ category; e) Artifact erasure: directly set to zero or fill with interpolation using a binary mask; f) Metadata encapsulation: Boundary coordinates, confidence levels, and processing logs are written to DICOM SR or JSON sidecar for subsequent AI analysis or doctor review.

[0047] In some embodiments, the optimization strategy corresponding to the blurred image of the subject's organs includes: a noise reduction strategy for each subject's organ; The noise reduction strategy for each organ of the subject includes: When the subject organ is the gastric antrum, the noise reduction strategy set includes morphological reconstruction filtering based on the gastric antrum ellipse / major axis prior, When the subject organ is a heart, the noise reduction strategy set includes a spatiotemporal joint deblurring process based on a cardiac cycle phase.

[0048] Through the above-mentioned organ-specific strategies, the model can automatically select the most appropriate noise reduction path based on the anatomical and movement characteristics of different organs, avoiding the information loss caused by "one-size-fits-all" while significantly improving the clinical readability of the image and the accuracy of subsequent AI recognition.

[0049] The noise reduction strategy for each organ of the subject is described below with reference to specific embodiments (in practical applications, the organ of the subject may be, but is not limited to, the gastric antrum and the heart).

[0050] When the main organ is the gastric antrum, the antrum usually presents an elliptical outline in the standard short-axis section, and the long axis direction is consistent with the anatomical direction of the gastric body. The system first uses a large number of normal gastric antrum images to calculate the typical ratio of the long and short axes of the ellipse to form an "ellipse-long axis" prior template. When the input frame is blurred or the boundary is discontinuous, the noise reduction process is as follows: first detect the edge pixels of the current frame, and then use the statistical template for comparison to determine the missing or blurred segments; using the long axis direction of the template as a reference, adaptively generate a coherent reconstruction curve to fill the gap; the reconstructed edge is subjected to morphological opening and closing operations to remove small burrs while retaining the natural curvature of the gastric antrum wall; the final contour is consistent with the anatomical morphology and avoids the artifacts caused by traditional sharpening.

[0051] When the main organ is the heart, its motion exhibits a clear periodicity. The system first divides the input video into cardiac cycles (using synchronized electrocardiograms or autocorrelation algorithms), mapping each frame to a corresponding phase interval. Next, within the same phase interval, several frames are selected to form a "temporal neighborhood." These frames are motion-aligned to eliminate inter-frame displacement caused by cardiac pulsation. An anisotropic filter window is designed along the main direction of cardiac motion, performing deblurring simultaneously in the temporal and spatial domains. The resulting image preserves myocardial texture while significantly reducing boundary diffusion caused by motion blur.

[0052] In some embodiments, the optimization strategy corresponding to the blurred image of the subject's organ includes: when identifying a suspected lesion area in the subject's organ, performing denoising on the suspected lesion area with a higher weight.

[0053] When the system performs organ blur optimization, it first calls the pre-set lesion detection subnetwork to quickly screen the current frame and generate a "lesion probability heat map." The brighter the area in the heat map, the higher the probability of lesions. Then, the noise reduction process enters "priority mode": For pixels whose heat map brightness exceeds the set threshold, the system will assign a larger weight factor in the loss function, so that the model pays more attention to these areas during both training and inference stages, avoiding "one-size-fits-all" global smoothing.

[0054] When the proportion of the lesion area to the entire organ is higher than the empirical threshold, the system will trigger secondary refined noise reduction: shrinking the filter kernel or reducing the step size to retain finer edges; adopting stricter convergence conditions to ensure that the lesion boundary is clear and the internal texture is realistic.

[0055] After completing the priority noise reduction, the system will add additional "lesion area mask" metadata to the output video for direct call during subsequent AI segmentation or doctor review, without the need for manual selection again.

[0056] Through the above mechanism, the system can ensure higher image fidelity in the diseased area without increasing the overall processing time, thereby effectively improving the accuracy and reliability of subsequent diagnosis.

[0057] In some embodiments, it also includes: adding indication information to the sample video, wherein the indication information is used to indicate the region of interest; annotating the sample video with the indication information added, wherein the annotation is used to represent a higher-weighted noise reduction of the region of interest; training the dynamic ultrasonic noise reduction model based on the sample video with the indication information added and its corresponding annotation, so that the dynamic ultrasonic noise reduction model is used to input the dynamic ultrasonic video with the indication information, perform higher-weighted noise reduction on the region of interest indicated by the indication information, and output the dynamic ultrasonic video after noise reduction.

[0058] To facilitate doctors to focus on specific anatomical structures or lesion areas, the system introduces a "user indication information" mechanism during the training phase. The specific process is as follows: Adding indicator information: On the original sample video, the operator is allowed to mark one or more regions of interest (ROIs) by mouse selection, touch tracking, or voice command. The indicator information is bound to the video frame in the form of a rectangular, polygonal, or pixel-level mask and stored as additional metadata.

[0059] ROI Weighted Annotation: For video frames with an added ROI, the annotator, in addition to the existing four-dimensional annotations (boundary, padding, optimization, and erasure), also records the "noise reduction weight coefficient" w (w>1, default 2.0) for pixels within the ROI. If the ROI overlaps with a lesion, the lesion weight is prioritized; if there is no lesion, the user-specified weight is used. (In some embodiments, the ROI priority principle can be selected based on actual circumstances.)

[0060] This weight will be directly mapped to the loss function, causing the model to amplify the error within the ROI during training, thereby "learning" to give higher fidelity to this area.

[0061] Weighted training: During each backpropagation, the dynamic ultrasonic noise reduction model first reads the ROI mask of the current frame, multiplies the pixel error within the mask by w, and maintains the weight of pixels outside the mask at 1. This way, network parameter updates are significantly biased toward the ROI, ensuring sharp edges and intact textures.

[0062] Real-time response during the inference phase: When the trained model receives a new dynamic ultrasound video with an ROI indication, it automatically executes the following process without retraining: identifying the ROI area; calling the weighted denoising branch; and outputting a denoised video in which the ROI area is significantly clearer than the background.

[0063] Through the above embodiments, the system realizes interactive noise reduction of "where the doctor points, the model focuses on that area", which not only meets the needs of quick browsing at the bedside, but also supports local fine observation in scientific research or teaching without the need for additional post-processing.

[0064] In some embodiments, the dynamic ultrasound video after noise reduction further includes information indicating the boundaries of the subject's organs.

[0065] When the system outputs the noise-reduced dynamic ultrasound video, it will simultaneously generate and embed "subject organ boundary indication information" so that subsequent AI algorithms or doctors can directly call it without secondary labeling. The specific implementation method is as follows: Boundary information generation: After the denoising model completes image processing in the inference phase, it immediately converts the closed contours obtained by the boundary detection branch into a vector coordinate sequence (such as a list of polygon vertices) and calculates the confidence score for each edge.

[0066] Information encapsulation: Vector coordinates and confidence levels are written to metadata files corresponding to video frames. Examples of these formats include: DICOM SR (Structured Report), stored with the original image; a JSON sidecar with the same name and path as the video file; and SEI messages (H.264 / H.265) within the video frame for frame-level binding.

[0067] Rendering and Interaction: On the ultrasound workstation or cloud-based image review interface, the system reads this information in real time and overlays the boundary as a colored outline on the noise-reduced video, allowing doctors to hide or adjust transparency with a single click. If doctors detect boundary errors, they can directly drag and drop them on the interface to correct them. The correction results can then be fed back into the training set for model iteration.

[0068] Since boundary information is stored in a standard vector format, subsequent AI segmentation, ranging, and functional calculation modules can be directly parsed without additional reasoning, significantly shortening the diagnostic chain.

[0069] Through this embodiment, the dynamic ultrasound video after noise reduction not only has clearer image quality, but also comes with "editable and calculable" organ boundary information, truly achieving "one-time noise reduction, full-process usability".

[0070] The solution provided in this application is further described below with reference to specific embodiments: First, a large number of ultrasound videos based on clinical scenarios are obtained. Clinical experts then perform standardized evaluations of these clinical ultrasound videos, identifying common clinical scenarios and image characteristics associated with poor image quality. Based on these specific image categories, the images are manually labeled with target structures. The labeled images are then compared with standard images for training. The goal is to train a standard image as the gold standard for AI to identify target structures. The ultimate goal is to enable AI to automatically learn from manually labeled images of poor quality, training it to automatically identify target structures when encountering problematic images. The AI ​​then performs its own image optimization processing based on the standard image.

[0071] The specific steps for gastric antral ultrasound images are as follows: Ultrasound images of the antrum were collected for clinical scenarios including antral gas, antral fluid, and solid antral contents. Clinical experts manually traced the antral border in different clinical scenarios. These ultrasound images were then input into the AI ​​model for machine learning. The goal was to train the AI ​​to identify the antral border in different clinical scenarios and improve the accuracy of AI recognition. The gastric antrum is typically oval or round in shape, whether fasting or full. However, in thin, tall patients, the antrum is often elongated, leading to inaccurate AI recognition of this type of antrum. To address this, we collected images of antral structures with elongated antral features. Experts then traced the target structures and fed them into the AI ​​model for learning. The goal is to train the AI ​​model to accurately identify the boundaries of these elongated antrums. Mimicry: The intestine mimics the gastric antrum. The gastric antrum is surrounded by intestinal structures such as the duodenum and small intestine. Therefore, when the intestinal contents are large and the ultrasound section cuts through the short axis of the intestine, the intestinal section appears circular or oval, similar to the antrum. Therefore, AI often identifies cross-sections of the duodenum or small intestine next to the antrum as the gastric antrum. Based on this, ultrasound images with a large number of small intestines next to the gastric antrum are collected. Experts tracing the target structures are then input into the AI ​​for learning. The goal is to improve the AI's accuracy in identifying the gastric antrum boundary in this clinical scenario.

[0072] The specific steps for cardiac ultrasound images are as follows: Collect ultrasound images that fully display the overall structure of the heart, and use expert tracing to train AI to identify the heart structure under normal conditions; Ultrasound images of the heart that are incomplete due to lung qi or ribs obscuring the heart are collected. Clinical experts manually trace the heart boundaries in different clinical scenarios. These expert-traced ultrasound images are then fed into an AI model for machine learning. This aims to train the AI ​​to identify the boundaries between the endocardium and epicardium in different clinical scenarios and improve AI recognition accuracy. Second, we manually process the video to clarify the target structure based on a gold standard determined by experts. This processing method includes automatically filling in incomplete video boundaries, automatically optimizing blurred video boundaries, and automatically removing artifacts. The manually processed video is then used as the annotated video and used in conjunction with the manually processed video to train the model.

[0073] The specific steps for gastric antral ultrasound images are as follows: The position of the gastric antrum is relatively fixed. The ultrasound probe is required to maintain image stability during examinations to ensure the stability of the antrum. However, in patients with significant respiratory mobility, the antrum moves with respiration during a respiratory cycle due to the downward movement of the diaphragm during inspiration. This results in the inability to fully display the antral border within a single ultrasound window during a single respiratory cycle. To address this, we collected ultrasound images from this clinical scenario and, based on images that fully displayed the antral border during the respiratory cycle, manually filled in the image borders that were incomplete. We also fed standard images into the AI ​​model for learning, with the goal of training the AI ​​to automatically fill in the incomplete antral borders.

[0074] Obese patients have thick abdominal fat that affects ultrasound imaging, resulting in blurred gastric antral images and difficulty identifying the antral boundary. To address this, we collected ultrasound images from this clinical scenario, manually traced the antral boundary, and fed these standard images into an AI model for learning. This aims to train the AI ​​to identify the antral boundary. This helps improve recognition accuracy by predicting the likely location of the antrum.

[0075] Reference Figure 2 , the present application also provides a dynamic ultrasonic noise reduction method, comprising: Step S201, obtaining dynamic ultrasound video; Specifically, dynamic ultrasound videos are pulled from the ultrasound host or image library in real time or batches, the frame sequence is synchronized and the resolution is unified before being sent to the cache queue.

[0076] Step S202: inputting the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; Specifically, the cached video is fed frame by frame into the trained dynamic ultrasound noise reduction model; the model instantly identifies the organ category, calls the corresponding noise reduction strategy, completes boundary padding, deblurring, and artifact erasure, and appends organ contour metadata within the same frame, ultimately outputting a high-fidelity, low-noise dynamic ultrasound video stream.

[0077] The dynamic ultrasonic noise reduction model is pre-trained using the above-mentioned training method for the dynamic ultrasonic noise reduction model.

[0078] 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.

[0079] Figure 3 FIG. 1 is a block diagram of a training device for a dynamic ultrasonic noise reduction model provided by an embodiment of the present application. Figure 3 As shown, the device includes: The sample acquisition module 31 is configured to acquire a preset number of sample videos, wherein the sample videos include ultrasound videos in different clinical scenarios; the different clinical scenarios include gastric antral gas, gastric antral fluid, solid gastric antral contents, elongated gastric antrum, and occlusion of anatomical structures around the heart; a labeling module 32 for labeling the sample video, wherein the labeling is used to indicate: boundaries of the subject organs in the sample video, filling of boundaries of incomplete images of the subject organs, optimization strategy of blurred images of the subject organs, and erasure of artifact images of the subject organs; A training module 33 is configured to train a pre-built deep learning model based on the sample video and its corresponding annotations to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of incomplete images of the main organs, optimize the image based on the optimization strategy corresponding to the blurred image of the main organs, and erase the artifact images of the main organs to obtain the dynamic ultrasound video after noise reduction.

[0080] Figure 4 FIG. 1 is a block diagram of a device for dynamic ultrasonic noise reduction 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 video; A noise reduction module 42 is configured to input the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; The dynamic ultrasonic noise reduction model is pre-trained using the above-mentioned training method for the dynamic ultrasonic noise reduction model.

[0081] Below, reference Figure 5 To describe the electronic device according to the embodiment of the present application. Figure 5 A block diagram of an electronic device according to an embodiment of the present application is illustrated.

[0082] like Figure 5 As shown, electronic device 500 includes one or more processors 510 and memory 520 .

[0083] 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.

[0084] 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 ultrasonic noise reduction method, dynamic ultrasonic noise reduction model training method, and / or other desired functions described in the various embodiments of the present application described above. Various content, such as category correspondences, may also be stored in the computer-readable storage medium.

[0085] 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).

[0086] 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.

[0087] 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.

[0088] 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 of the dynamic ultrasonic noise reduction method and the training method of the dynamic ultrasonic noise reduction model according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0089] 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.

[0090] 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 dynamic ultrasonic noise reduction method and the training method of the dynamic ultrasonic noise reduction model according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0091] 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.

[0092] 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 ultrasonic noise reduction model, characterized in that: include: Obtaining a preset number of sample videos, the sample videos including: ultrasound videos in different clinical scenarios; The different clinical scenarios include antral gas, antral fluid, solid antral contents, elongated antrum, and obstruction by surrounding cardiac anatomical structures; Annotating the sample video, wherein the annotation is used to indicate: boundaries of the subject organs in the sample video, filling of boundaries of incomplete images of the subject organs, optimization strategy of blurred images of the subject organs, and erasure of artifact images of the subject organs; Based on the sample videos and their corresponding annotations, a pre-built deep learning model is trained to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, segment and complete the missing segments of the main organs, optimize the contours of the fuzzy areas, and suppress or erase artifacts to obtain the dynamic ultrasound video after noise reduction.

2. The training method of the dynamic ultrasonic noise reduction model according to claim 1, characterized in that: The optimization strategy corresponding to the blurred image of the subject's organs includes: a noise reduction strategy for each subject's organ; The noise reduction strategy for each organ of the subject includes: When the subject organ is the gastric antrum, the noise reduction strategy set includes morphological reconstruction filtering based on the gastric antrum ellipse / major axis prior, When the subject organ is a heart, the noise reduction strategy set includes a spatiotemporal joint deblurring process based on a cardiac cycle phase.

3. The training method of the dynamic ultrasonic noise reduction model according to claim 1, characterized in that: The optimization strategy corresponding to the blurred image of the subject's organs includes: When a suspected lesion region in a subject's organ is identified, a higher weight of noise reduction is performed on the suspected lesion region.

4. The training method of the dynamic ultrasonic noise reduction model according to claim 1, characterized in that: Also includes: Adding indication information to the sample video, where the indication information is used to indicate a region of interest; Annotating the sample video to which the indication information is added, wherein the annotation is used to indicate that a higher weight of noise reduction is performed on the region of interest; Based on the sample video with added indication information and its corresponding annotations, the dynamic ultrasound noise reduction model is trained so that the dynamic ultrasound noise reduction model is used to input the dynamic ultrasound video with added indication information, perform higher-weighted noise reduction on the region of interest, and output the noise-reduced dynamic ultrasound video.

5. The training method of the dynamic ultrasonic noise reduction model according to claim 1, characterized in that: Also includes: The dynamic ultrasound video after noise reduction includes indication information of the boundaries of the subject's organs.

6. A dynamic ultrasonic noise reduction method, characterized in that: include: Acquire dynamic ultrasound video; Inputting the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; Wherein, the dynamic ultrasonic noise reduction model is pre-trained by the dynamic ultrasonic noise reduction model training method according to any one of claims 1 to 5.

7. A training device for a dynamic ultrasonic noise reduction model, characterized in that: include: A sample acquisition module is configured to acquire a preset number of sample videos, the sample videos including ultrasound videos in different clinical scenarios; the different clinical scenarios including antral gas, antral fluid, solid antral contents, elongated antral gastric contents, and occlusion of anatomical structures surrounding the heart; a labeling module, configured to label the sample video, wherein the labeling is used to indicate: boundaries of subject organs in the sample video, filling of boundaries of incomplete images of subject organs, optimization strategy of blurred images of subject organs, and erasure of artifact images of subject organs; A training module, configured to train a pre-built deep learning model based on the sample video and its corresponding annotations to obtain a dynamic ultrasonic noise reduction model; The dynamic ultrasound noise reduction model is used to determine the main organs of the dynamic ultrasound video input into the dynamic ultrasound noise reduction model, identify the boundaries of the main organs, automatically fill in the boundaries of incomplete images of the main organs, optimize the image based on the optimization strategy corresponding to the blurred image of the main organs, and erase the artifact images of the main organs to obtain the dynamic ultrasound video after noise reduction.

8. A dynamic ultrasonic noise reduction device, characterized in that: include: An acquisition module, used for acquiring dynamic ultrasound video; a noise reduction module, configured to input the dynamic ultrasound video into a dynamic ultrasound noise reduction model to obtain a dynamic ultrasound video after noise reduction; Wherein, the dynamic ultrasonic noise reduction model is pre-trained by the dynamic ultrasonic noise reduction model training method according to any one of claims 1 to 5.

9. 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 ultrasonic noise reduction model according to any one of claims 1 to 5 or the dynamic ultrasonic noise reduction method according to claim 6 by running the program in the memory.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the training method of the dynamic ultrasonic noise reduction model according to any one of claims 1 to 5 or the dynamic ultrasonic noise reduction method according to claim 6.

Images