A deep learning-based cardiac ultrasound multi-task processing system
Patent Information
- Application Number
- CN202611152110.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-10-27
AI Technical Summary
1. 引入状态空间建模机制的深度神经网络结构以提升心包积液分割能力
Smart Images
Figure CN122991546A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ultrasound image processing technology, and more specifically to a deep learning-based cardiac ultrasound multi-task processing system. Background Technology
[0002] Echocardiography is a core tool for clinical assessment of cardiac structure and function, widely used in the diagnosis of coronary heart disease, cardiomyopathy, pericardial effusion, and other conditions. However, traditional ultrasound diagnosis has the following problems: 1. Cardiac section recognition relies on experience. Standard views of cardiac ultrasound (such as A4C, PLAX, PSAX, SC) require doctors to manually adjust the probe angle to obtain the images. Doctors in primary hospitals with insufficient experience are prone to misinterpreting the view type, leading to diagnostic errors.
[0003] 2. Difficulty in quantifying pericardial effusion The estimation of pericardial effusion volume relies on doctors manually delineating the effusion area, which is inefficient and highly subjective.
[0004] Existing technology: Patent CN118196180A proposes a CT image segmentation method based on U-Net, but ultrasound images differ from CT images in the following ways: Low contrast: The ultrasound image is noisy, and the boundary between the effusion and myocardium is blurred; Dynamics: The heartbeat causes drastic changes in the image sequence, and traditional segmentation models have poor adaptability.
[0005] 3. Cardiac tamponade has a high risk of being missed. Cardiac tamponade is a serious complication of pericardial effusion. It requires comprehensive judgment based on characteristics such as the amount of effusion and cardiac chamber collapse, and it is easily missed in clinical practice.
[0006] Existing methods rely on physician experience and lack automated risk assessment tools.
[0007] Echocardiography is a core technology for clinical assessment of cardiac structure and function; however, image quality and diagnostic results are highly dependent on operator experience, especially in acquiring specific sections and diagnosing effusions. Although some automated image processing systems are currently used for ultrasound imaging, they still suffer from low accuracy and fragmented model structures in multi-task fusion (such as section recognition, effusion detection, and cardiac tamponade discrimination), making it difficult to meet the needs of efficient and intelligent clinical diagnosis.
[0008] Therefore, there is an urgent need for a multi-tasking, high-precision auxiliary diagnostic system that adapts to the characteristics of cardiac ultrasound imaging. Summary of the Invention
[0009] The present invention is proposed to alleviate or solve at least one aspect or point of the above-mentioned problems.
[0010] To address the shortcomings of existing technologies, a deep learning-based multi-task processing system for cardiac ultrasound is provided. This system enables multi-task recognition of cardiac sections, quantification of pericardial effusion, and identification of tamponade, meeting the needs of clinical assessment. It solves problems such as low accuracy and fragmented model structures, fulfilling the clinical demand for efficient and intelligent diagnosis and filling a market gap.
[0011] The present invention provides a deep learning-based multi-task processing system for cardiac ultrasound, comprising an ultrasound image acquisition unit, an ultrasound image preprocessing unit, an ultrasound image recognition and processing unit, and a result export unit. The ultrasound image acquisition unit acquires ultrasound images using ultrasound equipment; The ultrasound image preprocessing unit preprocesses the acquired ultrasound images; The ultrasound image recognition and processing unit processes the preprocessed ultrasound images using a cardiac ultrasound processing model. The cardiac ultrasound processing model adopts a multi-task deep neural network design architecture, including encoder units, decoder units, skip connections, cardiac section classification branches, pericardial effusion segmentation branches, and pericardial tamponade classification branches; The result export unit is used to export the processing results obtained after processing the cardiac ultrasound processing model. The output of the ultrasound image acquisition unit is connected to the input of the ultrasound image preprocessing unit, the output of the ultrasound image preprocessing unit is connected to the input of the ultrasound image recognition and processing unit, and the output of the ultrasound image recognition and processing unit is connected to the input of the result export unit.
[0012] Preferably, the encoder unit includes a first encoder, a second encoder, a third encoder, and a fourth encoder, which have the same structure and adopt residual modules; The decoder unit includes a first decoder, a second decoder, a third decoder, and a fourth decoder, all of which have the same structure and use a state-space sequence modeling module.
[0013] Preferably, one output of the first encoder is connected to the input of the second encoder, and the other output is connected to the input of the fourth decoder via a skip connection; One output of the second encoder is connected to the input of the third encoder, and the other output is connected to the input of the third decoder via a skip connection. One output of the third encoder is connected to the input of the fourth encoder, and the other output is connected to the input of the second decoder via a skip connection. The output of the fourth encoder is connected to the input of the first decoder.
[0014] Preferably, the outputs of the first decoder, second decoder, third decoder and fourth decoder are fused and connected to the input of the pericardial effusion segmentation head, and the output of the fourth encoder and the output of the fourth decoder are fused and connected to the input of the cardiac section classification head and the pericardial tamponade classification head, respectively.
[0015] Preferably, the first encoder includes a first convolution, batch normalization, activation function, and second convolution, with the output of the previous stage connected to the input of the next stage in sequence.
[0016] Preferably, the first decoder includes a first layer normalization, a first linear mapping, a depthwise separable convolution, element-wise addition of a two-dimensional state space model, a second layer normalization, and a second linear mapping. The output of the first layer normalization is connected to the input of the first linear mapping, and the output of the previous layer is connected to the input of the next layer. The input of the first layer normalization is fused with the output of the second linear mapping.
[0017] Preferably, the cardiac section classification branch includes a cardiac section classification head, which includes a first fully connected layer, batch normalization, activation function, random inactivation layer, and a second fully connected layer. The output of the previous level is connected to the input of the next level in sequence, and the output of the cardiac section classification head is the cardiac section classification result.
[0018] Preferably, the pericardial effusion segmentation branch includes a pericardial effusion segmentation head, which includes a first convolution, batch normalization, activation function, and a second convolution. The output of the previous level is connected to the input of the next level in sequence, and the output of the pericardial effusion segmentation head is a pericardial effusion segmentation mask.
[0019] Preferably, the cardiac tamponade classification branch includes a cardiac tamponade classification head, which includes a first fully connected layer, batch normalization, activation function, random inactivation layer, and a second fully connected layer. The output of the previous level is connected to the input of the next level in sequence, and the output of the cardiac tamponade classification head is the cardiac tamponade classification result.
[0020] Preferably, the ultrasound image recognition and processing unit needs to be trained and optimized before use. The model training and optimization system includes an ultrasound image acquisition and annotation unit, an ultrasound image preprocessing and enhancement unit, an ultrasound image recognition and processing unit, and a result export unit. The ultrasound image acquisition and annotation unit acquires ultrasound images using ultrasound equipment and annotates the ultrasound images. The ultrasound image preprocessing and enhancement unit performs preprocessing and enhancement on the labeled ultrasound images. The ultrasound image recognition and processing unit processes pre-processed and enhanced ultrasound images using a cardiac ultrasound processing model. The result export unit is used to export the processing results obtained after processing the cardiac ultrasound processing model. The output of the ultrasound image acquisition and annotation unit is connected to the input of the ultrasound image preprocessing unit; the output of the ultrasound image preprocessing and enhancement unit is connected to the input of the ultrasound image recognition and processing unit; and the output of the ultrasound image recognition and processing unit is connected to the input of the result export unit.
[0021] The present invention discloses a deep learning-based multi-task processing system for cardiac ultrasound, which has the following beneficial effects: 1. Introducing a deep neural network structure with a state-space modeling mechanism to improve pericardial effusion segmentation capabilities. This invention is the first to introduce the "state-space sequence modeling module (Mambablock)" into the task of pericardial effusion image segmentation, constructing a segmentation network structure with strong global modeling capabilities and fine-grained structure perception capabilities, which can effectively adapt to the characteristics of blurred boundaries, low contrast, and structural complexity of pericardial effusion regions. Specifically, it includes: (1) Enhance global perception capabilities to improve boundary discrimination accuracy By embedding a state-space sequence modeling module into the segmentation network, the network's ability to model long-distance dependencies in the spatial dimension is significantly improved by leveraging its state-space sequence modeling capabilities. This enables the network to accurately identify subtle boundary differences between pericardial effusion areas and cardiac, myocardial, or lung tissues in regions with complex structural backgrounds and weak signal contrast, thereby improving the accuracy of edge localization.
[0022] (2) Enhance the ability to model small and irregular fluid accumulation morphologies. The state-space sequence modeling module possesses continuous state representation capabilities, enabling effective modeling of thin, irregularly distributed early pericardial effusions in images while maintaining structural continuity and stability even with weak features. This feature significantly enhances the model's ability to perceive small-volume targets, making it suitable for sophisticated analysis tasks such as early pericardial effusion detection, quantification, and clinical trend tracking.
[0023] 2. Construct a multi-task learning architecture that collaboratively optimizes three tasks: segmentation, pericardial tamponade classification, and section classification. This invention proposes a multi-task learning framework for addressing diverse clinical needs, integrating three tasks: pericardial effusion region segmentation, pericardial tamponade classification, and sectional classification. This framework achieves complementary synergy at the feature level, effectively improving the overall model's discrimination robustness and clinical applicability. Specifically, it includes: (1) Introducing multi-dimensional auxiliary tasks to promote the learning of core features We introduce cardiac tamponade classification and cardiac section classification branches to assist in the discrimination of cardiac tamponade status and ultrasound section type, respectively. The cardiac tamponade classification branch uses the encoder's high-dimensional semantic features to assess whether there is a risk of cardiac tamponade in the image; the cardiac section classification branch determines the section of the ultrasound image based on structure and texture distribution, providing contextual structure guidance for the segmentation task and improving the model's generalization ability.
[0024] (2) Multi-task information flow interaction enhances spatial semantic modeling capabilities The three tasks achieve bidirectional information enhancement: the salient region features output by the pericardial effusion segmentation branch guide the classifier to focus on key regions and improve classification accuracy; the classification results in turn guide the pericardial effusion segmentation branch to optimize the spatial feature decoding path, making the network more structurally discernible and semantically robust when processing blurred boundaries or low-visibility images.
[0025] This multi-task structure achieves deep integration of semantic and spatial tasks, significantly enhancing the network's adaptability to different image structures, clinical conditions, and diagnostic scenarios, and assisting in practical applications such as automatic screening of pericardial effusion, risk assessment of cardiac tamponade, and automatic identification of ultrasound image quality and sections. Attached Figure Description
[0026] Figure 1 This is a framework diagram of a deep learning-based cardiac ultrasound multitasking system, which is an exemplary embodiment of the present invention.
[0027] Figure 2 This is a system architecture diagram of a cardiac ultrasound processing model, which is an exemplary embodiment of the present invention.
[0028] Figure 3 This is a system architecture diagram of a residual module as an exemplary embodiment of the present invention.
[0029] Figure 4 The system architecture diagram of the state space sequence modeling module is an exemplary embodiment of the present invention.
[0030] Figure 5 This is a system architecture diagram of a pericardial effusion segmentation head, which is an exemplary embodiment of the present invention.
[0031] Figure 6 This is a system architecture diagram of a cardiac section classification head, which is an exemplary embodiment of the present invention.
[0032] Figure 7 The diagram shows the analysis results of a deep learning multi-task network architecture as an exemplary embodiment of the present invention. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0035] According to an exemplary embodiment of the present invention: Figures 1-7 As shown, a deep learning-based cardiac ultrasound multi-task processing system includes an ultrasound image acquisition unit, an ultrasound image preprocessing unit, an ultrasound image recognition and processing unit, and a result export unit. The ultrasound image acquisition unit acquires ultrasound images using ultrasound equipment; For example, apical four-chamber view (A4C), left ventricular long axis (PLAX), left ventricular short axis (PSAX), subxiphoid plane (SC), and “Other” (referring to other cardiac ultrasound images besides the above sections) are acquired using ultrasound imaging equipment such as Philips Affiniti 70 and Mindray Resona 7. These ultrasound images are exported in DICOM format. Image quality is strictly controlled during ultrasound image acquisition, and relevant acquisition conditions are recorded.
[0036] The ultrasound image preprocessing unit preprocesses the acquired ultrasound images; The ultrasound image preprocessing unit performs standardization processing on the ultrasound images from the ultrasound image acquisition unit, including size uniformity and grayscale normalization. For example, OpenCV and NumPy are used to uniformly process the ultrasound images, and all ultrasound images are uniformly adjusted to 512×512 pixels, and normalization is performed by dividing each ultrasound image by 255.
[0037] The ultrasound image recognition and processing unit processes the preprocessed ultrasound images using a cardiac ultrasound processing model. like Figures 1-6 As shown, the cardiac ultrasound processing model adopts a multi-task deep neural network design architecture, including encoder units, decoder units, skip connections, cardiac section classification branches, pericardial effusion segmentation branches, and pericardial tamponade classification branches. like Figure 2As shown, the cardiac ultrasound processing model adopts a backbone network that integrates residual modules, state-space sequence modeling modules, and U-Net structures. At the same time, it sets up cardiac section classification branches, pericardial effusion segmentation branches, and pericardial tamponade classification branches to achieve joint optimization of the three tasks.
[0038] In deep learning, a residual block is a special neural network structure that solves the gradient vanishing and degradation problems in deep neural networks by introducing skip connections, allowing the network to be trained to be deeper. The state-space sequence modeling module (MambaBlock) combines the advantages of state-space models (SSM) and neural networks, and shows performance that surpasses Transformer in long sequence modeling tasks.
[0039] U-Net is a convolutional neural network architecture specifically designed for image segmentation tasks. Its name comes from the U-shaped symmetrical design of its network structure. It mainly consists of an encoder (downsampling path) and a decoder (upsampling path), connected by skip connections. This structure is particularly suitable for medical image segmentation because it can accurately locate and segment target regions.
[0040] The overall architecture of U-Net consists of three key parts: encoder, decoder, and skip connections. The encoder (downsampling path) captures the semantic information of the image, the decoder (upsampling path) generates pixel-level segmentation masks using the semantic information learned by the encoder, and the skip connections directly pass the intermediate features (containing spatial detail information) of the encoder to the corresponding layer of the decoder to solve the problem of lost details during upsampling.
[0041] like Figure 2 , Figure 3 and Figure 4 As shown, the four encoders of the U-Net backbone network all use residual modules, which are sequentially set as the first convolution, batch normalization, activation function and second convolution along the data flow direction. The output of the previous level is connected to the input of the next level. The four decoders all use state space sequence modeling modules, which are sequentially set as the first layer normalization, first linear mapping, depthwise separable convolution, two-dimensional state space model element-level addition, second layer normalization and second linear mapping along the data flow direction. The output of the first layer normalization is connected to the input of the first linear mapping. The output of the previous level is connected to the input of the next level. The input of the first layer normalization is fused with the output of the second linear mapping, which enhances the model's ability to model long-distance dependencies and fuzzy boundaries, and improves the recognition and edge localization capabilities of small fluid accumulation structures.
[0042] The input ultrasound image is sequentially passed through the first convolution, batch normalization, activation function and second convolution of the first encoder. The output ultrasound image is connected to the input of the second encoder in one path and is connected to the input of the fourth decoder in another path through a skip connection. Similarly, one output of the second encoder is connected to the input of the third encoder as the input of the third encoder, and the other output is connected to the input of the third decoder via a skip connection as the input of the third decoder; One output of the third encoder is connected to the input of the fourth encoder as the input of the fourth encoder, and the other output is connected to the input of the second decoder via a skip connection as the input of the second decoder; The output of the fourth encoder is connected to the input of the first decoder.
[0043] The first decoder includes a first layer normalization, a first linear mapping, a depthwise separable convolution, element-wise addition of a two-dimensional state space model, a second layer normalization, and a second linear mapping. The output of the first layer normalization is connected to the input of the first linear mapping, and the output of the previous layer is connected to the input of the next layer. The input of the first layer normalization is fused with the output of the second linear mapping to serve as the output of the first decoder.
[0044] The outputs of the first, second, third, and fourth decoders are fused and connected to the input of the pericardial effusion segmentation head, which serves as the input of the pericardial effusion segmentation head. The output of the pericardial effusion segmentation head is the pericardial effusion segmentation mask. The outputs of the fourth encoder and the fourth decoder are fused and then connected to the inputs of the cardiac section classification head and the cardiac tamponade classification head, respectively. The output of the cardiac section classification head is the cardiac section classification result, and the output of the cardiac tamponade classification head is the cardiac tamponade classification result.
[0045] like Figure 6 As shown, the cardiac section classification branch includes a cardiac section classification head for ultrasound image section classification. The cardiac section classification head adopts a fully connected + Softmax architecture. The cardiac section classification head includes a first fully connected layer, batch normalization, activation function, random inactivation layer, and a second fully connected layer. The output of the previous level is connected to the input of the next level in sequence, and its output is the cardiac section classification result.
[0046] The cardiac tamponade classification branch includes a cardiac tamponade classification head used for cardiac tamponade classification on ultrasound images. Its structure is the same as that of the cardiac cross-section classification head, and its output is the cardiac tamponade classification result.
[0047] The pericardial tamponade branch introduces the output from the fourth decoder, which is the attention region of the segmentation mask, to weight and enhance the classification features. The three tasks achieve feature sharing through a shared encoder and achieve synergistic improvement with semantic guidance.
[0048] like Figure 5 As shown, the pericardial effusion segmentation branch includes a pericardial effusion segmentation head for implementing pericardial effusion segmentation. The pericardial effusion segmentation head includes a first convolution, a batch normalization, an activation function, and a second convolution. The output of the previous level is connected to the input of the next level, and its output is the pericardial effusion segmentation mask.
[0049] like Figure 2 and Figure 7 As shown, the result export unit is used to export the processing results obtained after processing the cardiac ultrasound processing model.
[0050] Compile patient information, imaging data, and analysis results into a written report. Figure 7 (Left side) Doctors and patients can access written reports by visiting the website or viewing them on mobile devices. Figure 7 (Right side).
[0051] The output of the ultrasound image acquisition unit is connected to the input of the ultrasound image preprocessing unit, the output of the ultrasound image preprocessing unit is connected to the input of the ultrasound image recognition and processing unit, and the output of the ultrasound image recognition and processing unit is connected to the input of the result export unit.
[0052] Before using the cardiac ultrasound processing model, comprehensive model training and optimization are required. The model training and optimization system includes an ultrasound image acquisition and annotation unit, an ultrasound image preprocessing and enhancement unit, and an ultrasound image recognition and processing unit. The ultrasound image acquisition and annotation unit is used to acquire ultrasound images through ultrasound equipment and to annotate the ultrasound images. For the collected and organized ultrasound images, LabelMe is used for annotation. For example, three clinical professionals use polygonal regions to finely annotate the pericardial effusion cavity. Regions with consistent annotations are directly adopted, while regions with inconsistent annotations are evaluated by another clinical physician for final determination.
[0053] During the annotation process, the acquisition angles of individual ultrasound images were also annotated, namely the left ventricular long axis (PLAX), left ventricular short axis (PSAX), apical four-chamber view (A4C), subxiphoid plane (SC), and other non-standard planes.
[0054] The LabelMe tool was used for pixel-level segmentation of the effusion area, and three ultrasound physicians with associate chief physician qualifications determined cardiac tamponade. If the tamponade determination results were consistent, they were directly adopted; otherwise, another ultrasound physician evaluated and made the final determination. The labels were stored in JSON and binary mask formats to construct a standardized multi-label dataset, which served as the input images for ultrasound image preprocessing and enhancement units.
[0055] The ultrasound image preprocessing and enhancement unit is used to preprocess and enhance the labeled ultrasound images. The ultrasound image preprocessing and enhancement unit receives a standardized multi-label dataset from the ultrasound image acquisition and annotation unit, and performs standardization processing on the ultrasound images in the standardized multi-label dataset, including size unification and grayscale normalization. For example, OpenCV and NumPy are used to unify the processing of ultrasound images, and the size of all ultrasound images is uniformly adjusted to 512×512 pixels, and normalization processing is performed by dividing each ultrasound image by 255.
[0056] Data augmentation is based on the Albumentations library. Specifically, `RandomCrop(height=448, width=448, p=0.5)` randomly selects a region in the image and crops it to a size of 448 pixels high and 448 pixels wide, with a probability (p) of 0.5 for applying this transformation; `HorizontalFlip` is used to horizontally flip the image to augment the training data, with a probability (p) of 0.5 for applying this transformation; `CLAHE(clip_limit=2.0, p=0.3)` optimizes local contrast, enhances image details, and avoids the noise amplification problem that may be caused by traditional histogram equalization, where the histogram clipping threshold (`clip_limit`) is 2, and the probability (p) for applying this transformation is 0.3; `GaussianNoise(var_limit=(10.0, 50.0), p=0.3)` adds Gaussian noise to the image to simulate random interference in real-world scenes, where the noise variance range (`var_limit`) is (10.0, ... The probability (p) of applying this transformation is 0.3; ElasticTransform(alpha=1, sigma=50, alpha_affine=50, p=0.2) is used to generate more natural and diverse training samples by simulating elastic deformation, with a deformation intensity factor (alpha) of 1, a Gaussian kernel standard deviation (sigma) of 50, an affine transformation component (alpha_affine) of 50, and a probability (p) of applying this transformation of 0.2.
[0057] Enhanced samples are constructed using various methods such as affine transformation, random cropping, contrast perturbation, and artifact enhancement involved in the above process. This improves the robustness of the model to images acquired by different devices and patient differences. The preprocessed and enhanced images are then used as input images for the ultrasound image recognition and processing unit.
[0058] The ultrasound image recognition and processing unit processes pre-processed and enhanced ultrasound images using a cardiac ultrasound processing model. The preprocessed and enhanced ultrasound images were trained using a cardiac ultrasound processing model that integrates a fusion residual module, a state-space sequence modeling module, and a U-Net backbone network.
[0059] The cardiac ultrasound processing model must simultaneously perform the following three functions: Cardiac section classification: Automatically identifies the five section categories to which ultrasound images belong; Pericardial effusion segmentation, precisely locating the pericardial effusion area; Cardiac tamponade classification to determine the presence of cardiac tamponade risk in ultrasound imaging.
[0060] During training, the model was optimized using the training set, and its performance was measured using an independent test set. A detailed analysis of the metrics for the three functions was conducted, including the accuracy of slice classification, the Dice coefficient for pericardial effusion segmentation, and the accuracy and sensitivity of pericardial tamponade classification.
[0061] During training, for example, the cardiac segmentation classification branch and the pericardial tamponade classification branch are trained using the Focal Loss loss function, while the pericardial effusion segmentation branch is trained using the cross-entropy, Dice, and edge joint loss function. The final loss function is a weighted sum of the loss functions for each classification and segmentation. Learning rate scheduling and early stopping strategies are used for optimization training. The optimizer used is the AdamW adaptive learning rate optimizer with an initial learning rate of 1e-4, CosineAnnealingLR learning rate scheduling, batch size = 8, epoch = 120, and the training stops when there is no improvement in the Dice coefficient on the validation set for 10 epochs. The training platform of this invention uses an NVIDIA A100 40G graphics card.
[0062] The results of training the cardiac ultrasound processing model are as follows: The accuracy rate of cardiac section classification reached 98.7%, significantly improving the efficiency and consistency of section recognition.
[0063] The pericardial effusion segmentation achieved a Dice coefficient of 0.91, demonstrating clinically acceptable accuracy, and was superior to traditional models, especially in identifying small effusions.
[0064] The AUC for cardiac tamponade classification reached 0.95, with an accuracy of 94.2%, providing reliable support for high-risk assessment.
[0065] The system as a whole possesses end-to-end intelligent analysis capabilities and exhibits excellent deployability, scalability, and diagnostic consistency.
[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that variations and combinations of elements may be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A deep learning-based multi-task processing system for cardiac ultrasound, characterized in that, It includes an ultrasound image acquisition unit, an ultrasound image preprocessing unit, an ultrasound image recognition and processing unit, and a result export unit; The ultrasound image acquisition unit acquires ultrasound images using ultrasound equipment; The ultrasound image preprocessing unit preprocesses the acquired ultrasound images; The ultrasound image recognition and processing unit processes the preprocessed ultrasound images using a cardiac ultrasound processing model. The cardiac ultrasound processing model adopts a multi-task deep neural network design architecture, including encoder units, decoder units, skip connections, cardiac section classification branches, pericardial effusion segmentation branches, and pericardial tamponade classification branches; The result export unit is used to export the processing results obtained after processing the cardiac ultrasound processing model. The output of the ultrasound image acquisition unit is connected to the input of the ultrasound image preprocessing unit, the output of the ultrasound image preprocessing unit is connected to the input of the ultrasound image recognition and processing unit, and the output of the ultrasound image recognition and processing unit is connected to the input of the result export unit.
2. The deep learning-based multi-task processing system for cardiac ultrasound according to claim 1, characterized in that: The encoder unit includes a first encoder, a second encoder, a third encoder, and a fourth encoder, which have the same structure and use residual modules; The decoder unit includes a first decoder, a second decoder, a third decoder, and a fourth decoder, all of which have the same structure and use a state-space sequence modeling module.
3. The deep learning-based multi-task processing system for cardiac ultrasound according to claim 2, characterized in that: One output of the first encoder is connected to the input of the second encoder, and the other output is connected to the input of the fourth decoder via a skip connection; One output of the second encoder is connected to the input of the third encoder, and the other output is connected to the input of the third decoder via a skip connection. One output of the third encoder is connected to the input of the fourth encoder, and the other output is connected to the input of the second decoder via a skip connection. The output of the fourth encoder is connected to the input of the first decoder.
4. The deep learning-based multi-task processing system for cardiac ultrasound according to claim 3, characterized in that: The outputs of the first, second, third, and fourth decoders are fused and connected to the input of the pericardial effusion segmentation head. The outputs of the fourth encoder and the fourth decoder are fused and connected to the inputs of the cardiac section classification head and the pericardial tamponade classification head, respectively.
5. The deep learning-based cardiac ultrasound multi-task processing system according to claim 4, characterized in that: The first encoder includes a first convolution, batch normalization, activation function, and a second convolution, with the output of the previous stage connected to the input of the next stage in sequence.
6. The deep learning-based cardiac ultrasound multi-task processing system according to claim 5, characterized in that: The first decoder includes a first layer normalization, a first linear mapping, a depthwise separable convolution, element-wise addition of a two-dimensional state space model, a second layer normalization, and a second linear mapping. The output of the first layer normalization is connected to the input of the first linear mapping, and the output of the previous layer is connected to the input of the next layer. The input of the first layer normalization is fused with the output of the second linear mapping.
7. The deep learning-based cardiac ultrasound multi-task processing system according to claim 6, characterized in that: The cardiac section classification branch includes a cardiac section classification head; the cardiac section classification head includes a first fully connected layer, batch normalization, activation function, random inactivation layer, and a second fully connected layer, with the output of the previous level connected to the input of the next level in sequence; the output of the cardiac section classification head is the cardiac section classification result.
8. The deep learning-based cardiac ultrasound multi-task processing system according to claim 7, characterized in that: The pericardial effusion segmentation branch includes a pericardial effusion segmentation head, which includes a first convolution, batch normalization, activation function, and a second convolution. The output of the previous level is connected to the input of the next level in sequence. The output of the pericardial effusion segmentation head is the pericardial effusion segmentation mask.
9. The deep learning-based multi-task processing system for cardiac ultrasound according to claim 8, characterized in that: The cardiac tamponade classification branch includes a cardiac tamponade classification head, which consists of a first fully connected layer, batch normalization, activation function, random inactivation layer, and a second fully connected layer. The output of the previous level is connected to the input of the next level in sequence. The output of the cardiac tamponade classification head is the cardiac tamponade classification result.
10. The deep learning-based multi-task processing system for cardiac ultrasound according to claim 9, characterized in that: Before the ultrasound image recognition and processing unit is used, model training and optimization are required. The model training and optimization system includes an ultrasound image acquisition and annotation unit, an ultrasound image preprocessing and enhancement unit, and an ultrasound image recognition and processing unit. The ultrasound image acquisition and annotation unit acquires ultrasound images using ultrasound equipment and annotates the ultrasound images. The ultrasound image preprocessing and enhancement unit performs preprocessing and enhancement on the labeled ultrasound images. The ultrasound image recognition and processing unit processes pre-processed and enhanced ultrasound images using a cardiac ultrasound processing model. The output of the ultrasound image acquisition and annotation unit is connected to the input of the ultrasound image preprocessing unit, and the output of the ultrasound image preprocessing and enhancement unit is connected to the input of the ultrasound image recognition and processing unit.