Super-resolution ultrasonic imaging microbubble positioning system and method based on deep learning

Abstract

The invention provides a super-resolution ultrasonic imaging microbubble positioning system based on deep learning, and the system comprises a model training module which is used for obtaining a real ultrasonic image, and taking the real ultrasonic image as training data to train a neural network model, thereby obtaining a deep learning model; the image positioning module is used for taking the segMama model as a baseline model, adding an attention mechanism to the deep learning model, carrying out image segmentation on the to-be-processed image based on an adding result in combination with the baseline model, and obtaining a microbubble position based on an image segmentation result; and the image tracking module is used for searching a global minimum value of the LSA cost function based on a Hungary algorithm, and tracking the microbubbles according to the positions of the microbubbles based on a searching result. According to the invention, more accurate microbubble positioning and tracking results can be provided, the application effect of ultrasonic imaging in the medical field can be improved, and a more reliable basis is provided for diagnosis and treatment of diseases.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and particularly relates to a super-resolution ultrasound imaging microbubble localization system and method based on deep learning. Background Art

[0002] Currently, in the field of medical ultrasound imaging, the precise localization of microbubbles is of great significance for disease diagnosis and treatment;

[0003] Traditional ultrasound imaging has certain limitations in terms of resolution and is achieved through traditional computer vision techniques such as local maximum detection, etc. It usually takes time and requires fine-tuning of multiple parameters to obtain the best results, and it is difficult to achieve high-precision localization of microbubbles. In addition, during the ultrasound imaging process, there are various interference factors and imaging quality problems such as noise and artifacts, and these influencing factors also affect the imaging effect and microbubble localization ability, greatly reducing the accuracy and effect of microbubble localization;

[0004] Therefore, in order to overcome the above defects, the present invention provides a super-resolution ultrasound imaging microbubble localization system based on deep learning. Summary of the Invention

[0005] The present invention provides a super-resolution ultrasound imaging microbubble localization system and method based on deep learning. By collecting a large number of real ultrasound images as training data and inputting them into a neural network model for training, the model learns the features and rules of ultrasound images, thereby obtaining a deep learning model capable of processing ultrasound images, ensuring the reliability of the obtained deep learning model. Training the model with a large number of real ultrasound images improves the model's processing ability and adaptability to ultrasound images, which helps to perform microbubble localization more accurately. At the same time, using the segMamba model as the baseline model and adding an attention mechanism on this basis can make the model focus more on the important feature regions related to microbubbles in the image, which can improve the accuracy and precision of image segmentation, better identify and locate microbubbles. Finally, combining the baseline model to precisely segment the input image to be processed, thereby determining the position of microbubbles and using the Hungarian algorithm to find the global minimum of the LSA cost function, and associating the positions of microbubbles at different times through this minimum value to achieve the tracking of microbubbles. Using the Hungarian algorithm for microbubble tracking can more efficiently and accurately track the movement trajectory of microbubbles, providing more valuable information for medical diagnosis and treatment. This system can provide more accurate microbubble localization and tracking results, which helps to improve the application effect of ultrasound imaging in the medical field, providing a more reliable basis for disease diagnosis and treatment. At the same time, the deep learning model can process ultrasound images more efficiently, reduce the need for parameter adjustment, and can provide higher imaging resolution. In addition, the method of the present invention can adapt to different data sets, including synthetic data and in-vivo data, and has a wide range of application prospects.

[0006] The present invention provides a super-resolution ultrasound imaging microbubble localization system based on deep learning, including:

[0007] A model training module, configured to obtain real ultrasound images and use the real ultrasound images as training data to train a neural network model to obtain a deep learning model;

[0008] An image localization module, configured to use the segMamba model as a baseline model. At the same time, an attention mechanism is added to the deep learning model, and based on the addition result, the baseline model is combined to perform image segmentation on the image to be processed, and the microbubble position is obtained based on the image segmentation result;

[0009] An image tracking module, configured to find the global minimum value of the LSA cost function based on the Hungarian algorithm, and track the microbubbles based on the search result according to the microbubble position.

[0010] Preferably, for the super-resolution ultrasound imaging microbubble localization system based on deep learning, the model training module includes:

[0011] An image acquisition unit, configured to obtain an image retrieval request based on a management terminal, access a historical image database based on the image retrieval request, and conditionally traverse the historical images in the historical image database based on the access result to obtain a set of real ultrasound images;

[0012] A training data determination unit, configured to perform image cropping on each real ultrasound image in the set of real ultrasound images based on a preset cropping mode, and obtain training data based on the image cropping result.

[0013] Preferably, for the super-resolution ultrasound imaging microbubble localization system based on deep learning, the image acquisition unit includes:

[0014] An image retrieval subunit, configured to retrieve the obtained set of real ultrasound images and determine an image processing strategy for the set of real ultrasound images, where the image processing strategy includes contrast adjustment and image smoothing processing;

[0015] An image processing subunit, configured to process the set of real ultrasound images based on the image processing strategy.

[0016] Preferably, for the super-resolution ultrasound imaging microbubble localization system based on deep learning, the model training module includes:

[0017] A parameter setting unit, configured to obtain the obtained training data, and at the same time, obtain the configuration parameters of a Gaussian function, and set the configuration parameters of the Gaussian function to a 3*3 kernel and σ = 1 based on training requirements;

[0018] A model training unit, configured to:

[0019] Perform convolution operations on the Gaussian function after setting the configuration parameters with the training data respectively, obtain the heatmap corresponding to the training data based on the results of the convolution operations, and set the heatmap as the label of the training data;

[0020] Based on the set results, perform iterative training on the neural network model for a target number of times according to the labels and training data, and monitor the conditional convergence state of the neural network model in real time during the iterative training process. When the preset convergence condition is met, terminate the training to obtain the deep learning model.

[0021] Preferably, for the super-resolution ultrasonic imaging microbubble localization system based on deep learning, the model training unit includes:

[0022] The model testing subunit is used to obtain the obtained deep learning model. At the same time, retrieve the test data set from the historical database, input the test data set into the deep learning model for processing, and obtain the processing process parameters and target results of the deep learning model based on the processing results;

[0023] The model loss analysis subunit is used for:

[0024] Analyze the processing process parameters and target results of the deep learning model based on the preset loss function, determine the loss value of the deep learning model, and compare the loss value with the preset threshold;

[0025] If the loss value is less than or equal to the preset threshold, determine that the obtained deep learning model meets the preset convergence condition, and obtain the final deep learning model based on the determination result.

[0026] Preferably, for the super-resolution ultrasonic imaging microbubble localization system based on deep learning, the image localization module includes:

[0027] The parameter acquisition unit is used to obtain the segMamba model based on the management terminal;

[0028] The model optimization unit is used for:

[0029] Set the pre-trained linear time series of the segMamba model as the baseline model of the encoder backbone in the deep learning model, and receive and process the image to be processed based on the set results to extract the image features of the image to be processed;

[0030] At the same time, obtain the target requirements for microbubble localization based on the management terminal, and generate the attention mechanism corresponding to the image segmentation based on the target requirements;

[0031] Add the attention mechanism to the encoder branch in the deep learning model, and decode the extracted image features of the image to be processed based on the addition result;

[0032] A microbubble positioning unit, which is used to lock the key image regions of the decoding result based on the attention mechanism, and perform image segmentation on the image to be processed based on the key image region locking result to obtain the positions of microbubbles in the image to be processed.

[0033] Preferably, for the super-resolution ultrasonic imaging microbubble positioning system based on deep learning, the image tracking module includes:

[0034] A result analysis unit, which is used to:

[0035] Obtain the positioning results of different microbubble positions in the current frame of the image to be processed in the image sequence to be processed, and denote them as P{i1}, P{i2},..., P{in};

[0036] At the same time, obtain the positioning results of different microbubble positions in the next frame of the image to be processed based on the deep learning model, and denote them as P{j1}, P{j2},..., P{jm};

[0037] Pair the positioning results of different microbubble positions in the adjacent frames of the image to be processed based on the LSA cost function, and obtain the corresponding cost function values based on the pairing results;

[0038] An optimization unit, which is used to represent the cost function values as a target matrix, perform global minimum search on the target matrix based on the Hungarian algorithm, and obtain the pairing results between the positioning results of different microbubble positions in the adjacent frames of the image to be processed based on the search results;

[0039] A microbubble tracking unit, which is used to track the same microbubble in the images to be processed of different frames based on the pairing results to obtain the corresponding tracking paths.

[0040] Preferably, for the super-resolution ultrasonic imaging microbubble positioning system based on deep learning, the microbubble tracking unit includes:

[0041] A result analysis subunit, which is used to:

[0042] Obtain the obtained tracking paths of microbubbles, and sequentially determine the distribution and displacement of the same microbubble between the adjacent frames of the image to be processed based on the tracking paths;

[0043] Accumulate the distribution and displacement of the microbubble between the adjacent frames of the image to be processed to obtain the density distribution and velocity parameters of the microbubble;

[0044] An image generation subunit, which is used to generate the density and velocity images of microbubbles based on the density distribution and velocity parameters of microbubbles, and feed back the generated density and velocity images of microbubbles to the management terminal for recording.

[0045] The present invention provides a super-resolution ultrasonic imaging microbubble positioning method based on deep learning, which is characterized by including:

[0046] Step 1: Obtain a real ultrasonic image, and use the real ultrasonic image as training data to train a neural network model to obtain a deep learning model;

[0047] Step 2: Take the segMamba model as the baseline model. At the same time, add an attention mechanism to the deep learning model, and based on the addition result, combine the baseline model to perform image segmentation on the image to be processed, and obtain the microbubble position based on the image segmentation result;

[0048] Step 3: Based on the Hungarian algorithm, find the global minimum of the LSA cost function, and track the microbubbles according to the microbubble position based on the finding result.

[0049] Preferably, for the microbubble localization method of super-resolution ultrasonic imaging based on deep learning, in Step 1, obtaining a real ultrasonic image and using the real ultrasonic image as training data to train a neural network model includes:

[0050] Based on the management terminal, obtain an image retrieval request, access the historical image database based on the image retrieval request, and conditionally traverse the historical images in the historical image database based on the access result to obtain a set of real ultrasonic images;

[0051] Based on a preset cropping mode, crop each real ultrasonic image in the set of real ultrasonic images, and obtain training data based on the image cropping result.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] By collecting a large number of real ultrasound images as training data and inputting them into a neural network model for training, the model learns the features and patterns of ultrasound images, thereby obtaining a deep learning model capable of processing ultrasound images, ensuring the reliability of the obtained deep learning model. Training the model with a large number of real ultrasound images improves the model's processing ability and adaptability to ultrasound images, which helps to more accurately localize microbubbles. At the same time, using the segMamba model as a baseline model and adding an attention mechanism on this basis can make the model focus more on the important feature regions related to microbubbles in the image, which can improve the accuracy and precision of image segmentation and better identify and locate microbubbles. Finally, combining the baseline model to precisely segment the input image to be processed, thereby determining the position of the microbubbles and using the Hungarian algorithm to find the global minimum of the LSA cost function, and associating the positions of microbubbles at different times through this minimum value to achieve the tracking of microbubbles. Using the Hungarian algorithm for microbubble tracking can more efficiently and accurately track the movement trajectory of microbubbles, providing more valuable information for medical diagnosis and treatment. This system can provide more accurate microbubble localization and tracking results, which helps to improve the application effect of ultrasound imaging in the medical field and provide a more reliable basis for the diagnosis and treatment of diseases. At the same time, the deep learning model can process ultrasound images more efficiently, reduce the need for parameter adjustment, and can provide higher imaging resolution. In addition, the method of the present invention can adapt to different data sets, including synthetic data and in-vivo data, and has broad application prospects.

[0054] Other features and advantages of the present invention will be described in the following specification, and in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained by the structures specifically pointed out in this application document.

[0055] The technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0057] Figure 1 is a structural diagram of a microbubble localization system for super-resolution ultrasound imaging based on deep learning in an embodiment of the present invention;

[0058] Figure 2 is a structural diagram of a model training module in a microbubble localization system for super-resolution ultrasound imaging based on deep learning in an embodiment of the present invention;

[0059] Figure 3This is a structural diagram of an image positioning module in a super-resolution ultrasound imaging microbubble positioning system based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0061] This embodiment provides a super-resolution ultrasound imaging microbubble localization system based on deep learning, such as Figure 1 As shown, including:

[0062] A model training module is used to obtain real ultrasound images and use the real ultrasound images as training data to train the neural network model to obtain a deep learning model;

[0063] An image localization module is used to use the segMamba model as a baseline model. At the same time, an attention mechanism is added to the deep learning model, and image segmentation is performed on the image to be processed based on the added result combined with the baseline model, and the microbubble position is obtained based on the image segmentation result;

[0064] The image tracking module is used to find the global minimum of the LSA cost function based on the Hungarian algorithm, and track the microbubbles according to the microbubble positions based on the search results.

[0065] In this embodiment, the real ultrasound image refers to an ultrasound image collected from an actual image library, and the real ultrasound image is used as training data to train the neural network model, with the aim of ensuring the reliability of the final deep learning model.

[0066] In this embodiment, the deep learning model refers to a model obtained through training data that is ultimately capable of locating microbubbles and has image segmentation and image processing capabilities.

[0067] In this embodiment, the segMamba model serves as a baseline model to:

[0068] Improve segmentation accuracy: It can segment images or data more accurately, for example, to obtain more refined results in the fields of medical image segmentation and semantic segmentation, in order to improve the accuracy of microbubble positioning;

[0069] Adapt to different tasks and data characteristics: Targeted adjustments can be made according to specific application scenarios and data characteristics to better exert model performance;

[0070] Enhance the generalization ability of the model: Improve the applicability and robustness of the model in different data sets and situations by improving the structure.

[0071] In this embodiment, the purpose of adding the attention mechanism is to ensure that when the deep learning model performs image segmentation on an image, it can focus on sensitive regions or sensitive objects (microbubbles), so as to realize the localization of microbubbles through image segmentation.

[0072] In this embodiment, the Hungarian algorithm is an algorithm used to solve the assignment problem. Its main function is to achieve optimal task assignment or resource matching in some specific scenarios. For example, in problems such as personnel and task assignment, machine and work assignment, it can find an optimal assignment scheme to make the overall efficiency or cost reach the optimal. Through the Hungarian algorithm, the most suitable matching relationship can be quickly and effectively determined, so as to realize the efficient use of resources and the best task execution arrangement, and then realize the tracking of the same microbubble in the images to be processed in different frames.

[0073] In this embodiment, the LSA cost function (linear sum assignment) is a classic assignment algorithm. Using LSA to globally optimize the distance is to find the movement trajectory of microbubbles in adjacent data frames, that is, to realize the tracking of microbubbles.

[0074] The working principle and beneficial effects of the above technical solutions are as follows: By collecting a large number of real ultrasonic images as training data and inputting them into the neural network model for training, the model learns the characteristics and laws of ultrasonic images, so as to obtain a deep learning model capable of processing ultrasonic images, ensuring the reliability of the obtained deep learning model. By training the model with a large number of real ultrasonic images, the processing ability and adaptability of the model to ultrasonic images are improved, which helps to perform more accurate microbubble localization. At the same time, using the segMamba model as the baseline model and adding the attention mechanism on this basis can make the model more focused on the important feature regions related to microbubbles in the image, improve the accuracy and precision of image segmentation, better identify and locate microbubbles. Finally, combining the baseline model to accurately segment the input image to be processed, so as to determine the position of microbubbles and using the Hungarian algorithm to find the global minimum value of the LSA cost function, and associating the positions of microbubbles at different times through this minimum value to realize the tracking of microbubbles. Using the Hungarian algorithm for microbubble tracking can track the movement trajectory of microbubbles more efficiently and accurately, providing more valuable information for medical diagnosis and treatment. This system can provide more accurate microbubble localization and tracking results, helping to improve the application effect of ultrasonic imaging in the medical field and providing a more reliable basis for the diagnosis and treatment of diseases. At the same time, the deep learning model can process ultrasonic images more efficiently, reduce the need for parameter adjustment, and can provide higher imaging resolution. In addition, the method of the present invention can adapt to different data sets, including synthetic data and in-vivo data, and has a wide application prospect.

[0075] In one embodiment, a deep learning-based super-resolution ultrasound imaging microbubble localization system is provided, as Figure 2 shown, the model training module includes:

[0076] An image acquisition unit, configured to obtain an image retrieval request based on a management terminal, access a historical image database based on the image retrieval request, and perform conditional traversal on the historical images in the historical image database based on the access result to obtain a set of real ultrasound images;

[0077] A training data determination unit, configured to perform image cropping on each real ultrasound image in the set of real ultrasound images based on a preset cropping mode, and obtain training data based on the image cropping result.

[0078] In this embodiment, the historical image database is set in advance and is used to store different real ultrasound images.

[0079] In this embodiment, the conditional traversal refers to the reference basis for scheduling the historical images in the historical image database. For example, it can be the type of the image and the subject recorded in the image, etc.

[0080] In this embodiment, the image cropping refers to the segmentation of the obtained real ultrasound images, and the purpose is to enrich the training samples.

[0081] The working principle and beneficial effects of the above technical solution are: after receiving an image retrieval request from the management terminal, access the historical image database according to the request, and then perform conditional traversal on the historical images in the database to screen out the real ultrasound images that meet the requirements, form a set of real ultrasound images, and determine the training data by cropping the real ultrasound images, which can make the training data more standardized and concentrated, is beneficial to improving the effect and accuracy of model training, provides high-quality data support for model training, and further improves the performance of the entire system in ultrasonic image processing and analysis.

[0082] In one embodiment, a deep learning-based super-resolution ultrasound imaging microbubble localization system is provided, and the image acquisition unit includes:

[0083] An image retrieval subunit, configured to retrieve the obtained set of real ultrasound images and determine an image processing strategy for the set of real ultrasound images, where the image processing strategy includes contrast adjustment and image smoothing processing;

[0084] An image processing subunit, configured to process the set of real ultrasound images based on the image processing strategy.

[0085] In this embodiment, the image processing strategy refers to the scheme or rule for processing the set of real ultrasound images.

[0086] The beneficial effects of the above technical solution are as follows: By processing the obtained real ultrasonic images, the accuracy and reliability of the finally obtained real ultrasonic images are ensured, and reliable guarantee is also provided for model training.

[0087] In one embodiment, a super-resolution ultrasonic imaging microbubble localization system based on deep learning is provided, which is characterized in that the model training module includes:

[0088] A parameter setting unit, which is used to obtain the obtained training data. At the same time, it obtains the configuration parameters of the Gaussian function and sets the configuration parameters of the Gaussian function as a 3*3 kernel and σ = 1 based on the training requirements;

[0089] A model training unit, which is used for:

[0090] Performing convolution operations on the Gaussian function after setting the configuration parameters and the training data respectively, obtaining the heat map corresponding to the training data based on the convolution operation result, and setting the heat map as the label of the training data;

[0091] Based on the setting result, performing iterative training on the neural network model for a target number of times according to the label and the training data, and monitoring the conditional convergence state of the neural network model in real time during the iterative training process. When the preset convergence condition is satisfied, the training is terminated to obtain a deep learning model.

[0092] In this embodiment, the heat map is a two-dimensional matrix visualization method that represents the magnitude of data values with colors. In the context of training data, usually, a certain feature or the relationship between multiple features in the training data is presented in an intuitive color matrix.

[0093] In this embodiment, the purpose of setting the heat map as the label of the training data is to perform targeted training on the neural network according to the features presented by the heat map, aiming to improve the efficiency and effect of training.

[0094] In this embodiment, the target number of times is set in advance.

[0095] In this embodiment, the conditional convergence state is used to characterize the working effect of the neural network during the neural network training process. For example, it can be that the analysis accuracy rate reaches more than 90%. Among them, the preset convergence condition is set in advance and is used as a reference basis for measuring whether the training is completed and can be adjusted.

[0096] The beneficial effects of the above technical solution are as follows: By preprocessing the obtained training data, an effective determination of the heat map is realized. Secondly, based on the obtained heat map and training data, iterative training on the neural network model is performed for a target number of times, realizing effective training of the deep model, improving the convergence speed and accuracy during model training, and ensuring the microbubble localization effect of the deep learning model.

[0097] In one embodiment, a super-resolution ultrasound imaging microbubble localization system based on deep learning is provided. The model training unit includes:

[0098] A model testing subunit, which is used to obtain the obtained deep learning model. At the same time, it retrieves a test data set from the historical database, inputs the test data set into the deep learning model for processing, and obtains the processing process parameters and target results of the deep learning model based on the processing results;

[0099] A model loss analysis subunit, which is used for:

[0100] Analyze the processing process parameters and target results of the deep learning model based on a preset loss function, determine the loss value of the deep learning model, and compare the loss value with a preset threshold;

[0101] If the loss value is less than or equal to the preset threshold, it is determined that the obtained deep learning model meets the preset convergence condition, and the final deep learning model is obtained based on the determination result.

[0102] In this embodiment, the target result refers to the processing result of the deep learning model on the test data set.

[0103] In this embodiment, the preset loss function is set in advance and is used to measure whether the training effect of the deep learning model meets the requirements.

[0104] In this embodiment, the preset threshold is set in advance.

[0105] The beneficial effect of the above technical solution is that by retrieving the test data set from the historical database and analyzing the test data set through the obtained deep learning model, the performance of the deep learning model can be accurately and effectively verified, ensuring the effectiveness of the finally obtained deep learning model.

[0106] In one embodiment, a super-resolution ultrasound imaging microbubble localization system based on deep learning is provided, as Figure 3 shown. The image localization module includes:

[0107] A parameter acquisition unit, which is used to obtain the segMamba model based on the management terminal;

[0108] A model optimization unit, which is used for:

[0109] Set the pre-trained linear time series of the segMamba model as the baseline model of the encoder backbone in the deep learning model, and receive and process the image to be processed based on the setting result, and extract the image features of the image to be processed;

[0110] Meanwhile, based on the management terminal, obtain the target requirements for microbubble positioning, and generate an attention mechanism corresponding to image segmentation based on the target requirements;

[0111] Add the attention mechanism to the encoder branch in the deep learning model, and decode the image features of the to-be-processed image extracted based on the addition result;

[0112] A microbubble positioning unit, configured to lock the key image area based on the attention mechanism for the decoding result, and perform image segmentation on the to-be-processed image based on the key image area locking result to obtain the position of the microbubble in the to-be-processed image.

[0113] In this embodiment, the target requirements refer to the factors to be considered during microbubble positioning and requirements such as positioning accuracy during positioning.

[0114] In this embodiment, the key image area refers to the image area containing the microbubble.

[0115] In this embodiment, the decoder branch is a part of the deep learning model. In some deep learning models with specific architectures, such as models for tasks like image generation and semantic segmentation, they often include encoder and decoder structures. The decoder branch usually undertakes the task of decoding, reconstructing, or generating the final output for the feature representation obtained after being processed by the encoder. It cooperates with the encoder and, through a series of operations and conversions, gradually restores or generates the features to the desired output form, such as generating an image, a segmentation result, etc.

[0116] In this embodiment, the encoder plays an important role in the deep learning model. The encoder is mainly responsible for feature extraction and compressed representation of the input data. Through a series of operations such as convolution and pooling, the original complex input data is converted into a more abstract and representative feature vector. Deep learning models usually aim to solve various complex tasks, and the features extracted by the encoder provide the basis for the subsequent processing and decision-making of the model. These features can be further used in different task operations such as classification, regression, and generation.

[0117] The beneficial effects of the above technical solution are: By adding the segMamba model and the attention mechanism to the deep learning model, it is convenient for the deep learning model to more accurately extract image features and focus on microbubbles when performing image segmentation on the to-be-processed image, thereby achieving accurate and effective segmentation of the to-be-processed image, and further improving the accuracy of microbubble positioning.

[0118] In one embodiment, a microbubble positioning system for super-resolution ultrasound imaging based on deep learning is provided. The image tracking module includes:

[0119] A result analysis unit, configured to:

[0120] Obtain the positioning results of different microbubble positions in the to-be-processed image of the current frame in the image sequence to be processed, and denote them as P{i1}, P{i2},..., P{in};

[0121] Meanwhile, obtain the positioning results of different microbubble positions in the to-be-processed image of the next frame based on the deep learning model, and denote them as P{j1}, P{j2},..., P{jm};

[0122] Pair the positioning results of different microbubble positions in the to-be-processed images of adjacent frames based on the LSA cost function, and obtain the corresponding cost function values based on the pairing results;

[0123] An optimization unit, configured to represent the cost function values as a target matrix, perform a global minimum search on the target matrix based on the Hungarian algorithm, and obtain the pairing results between the positioning results of different microbubble positions in the to-be-processed images of adjacent frames based on the search results;

[0124] A microbubble tracking unit, configured to track the same microbubble in the to-be-processed images of different frames based on the pairing results, and obtain the corresponding tracking paths.

[0125] In this embodiment, the target matrix refers to converting the different results corresponding to the cost function values into the corresponding matrix form, so as to facilitate training the global minimum value, and further realizing pairing of microbubbles and achieving tracking.

[0126] In this embodiment, LSA is to pair the positions in two frames of images, and the paired microbubbles are regarded as the same microbubble. For example, it is calculated through LSA that P{i1} and P{j2} are paired, that is, the same microbubble. In this way, the displacement of this microbubble in adjacent two frames can be calculated (because the position of each frame is known). Knowing the displacement can calculate the speed, and the pairing is realized through the global optimization distance function.

[0127] The beneficial effects of the above technical solution are: By using the Hungarian algorithm to optimize the pairing results of the same microbubble in the to-be-processed images of different frames by the LSA cost function, the effective pairing of the same microbubble in the to-be-processed images of different frames is realized. Through pairing, the specific position distribution and displacement parameters of the same microbubble in the to-be-processed images of different frames are determined, thereby improving the positioning effect and reliability of microbubbles.

[0128] In one embodiment, a microbubble positioning system for super-resolution ultrasound imaging based on deep learning is provided, and the microbubble tracking unit includes:

[0129] A result analysis subunit, configured to:

[0130] Obtain the tracking path of the microbubbles, and sequentially determine the distribution and displacement of the same microbubble between the to-be-processed images of adjacent frames based on the tracking path;

[0131] Accumulate the distribution and displacement of the microbubbles between the to-be-processed images of adjacent frames to obtain the density distribution and velocity parameters of the microbubbles;

[0132] An image generation subunit, configured to generate a density and velocity image of the microbubbles based on the density distribution and velocity parameters of the microbubbles, and feedback the generated density and velocity images of the microbubbles to the management terminal for recording.

[0133] The beneficial effects of the above technical solution are: By analyzing the tracking path of the microbubbles, the distribution and displacement of the same microbubble between the to-be-processed images of adjacent frames are effectively determined, and further, the density and velocity images of the microbubbles are effectively generated, which is convenient to more directly and effectively understand the position and distribution of the microbubbles at different times, and improves the effect of microbubble positioning.

[0134] This embodiment provides a microbubble localization method for super-resolution ultrasound imaging based on deep learning, including:

[0135] Step 1: Obtain a real ultrasound image, and use the real ultrasound image as training data to train a neural network model to obtain a deep learning model;

[0136] Step 2: Use the segMamba model as a baseline model. At the same time, add an attention mechanism to the deep learning model, and perform image segmentation on the to-be-processed image based on the addition result in combination with the baseline model, and obtain the microbubble positions based on the image segmentation result;

[0137] Step 3: Find the global minimum value of the LSA cost function based on the Hungarian algorithm, and track the microbubbles according to the microbubble positions based on the finding result.

[0138] The working principle and beneficial effects of the above technical solution are as follows: By collecting a large number of real ultrasound images as training data and inputting them into a neural network model for training, the model learns the features and rules of ultrasound images, thereby obtaining a deep learning model capable of processing ultrasound images, ensuring the reliability of the obtained deep learning model. Training the model with a large number of real ultrasound images improves the model's processing ability and adaptability to ultrasound images, which helps to more accurately locate microbubbles. At the same time, using the segMamba model as the baseline model and adding an attention mechanism on this basis can make the model focus more on the important feature regions related to microbubbles in the image, improve the accuracy and precision of image segmentation, and better identify and locate microbubbles. Finally, combining the baseline model to accurately segment the input image to be processed, thereby determining the position of the microbubbles and using the Hungarian algorithm to find the global minimum of the LSA cost function, and associating the positions of microbubbles at different times through this minimum value to achieve the tracking of microbubbles. Using the Hungarian algorithm for microbubble tracking can more efficiently and accurately track the movement trajectory of microbubbles, providing more valuable information for medical diagnosis and treatment. This system can provide more accurate microbubble positioning and tracking results, helping to improve the application effect of ultrasound imaging in the medical field and providing a more reliable basis for the diagnosis and treatment of diseases. At the same time, the deep learning model can more efficiently process ultrasound images, reduce the need for parameter adjustment, and provide higher imaging resolution. In addition, the method of the present invention can adapt to different data sets, including synthetic data and in-vivo data, and has broad application prospects.

[0139] In one embodiment, a deep learning-based super-resolution ultrasound imaging microbubble localization method is provided. In step 1, real ultrasound images are obtained, and the real ultrasound images are used as training data to train a neural network model, including:

[0140] Based on the management terminal, an image retrieval request is obtained, the historical image database is accessed based on the image retrieval request, and the historical images in the historical image database are conditionally traversed based on the access result to obtain a set of real ultrasound images;

[0141] Based on a preset cropping mode, each real ultrasound image in the set of real ultrasound images is cropped, and training data is obtained based on the image cropping result.

[0142] The working principle and beneficial effects of the above technical solution are as follows: After receiving an image retrieval request from the management terminal, the historical image database is accessed according to the request, and then by traversing the historical images in the database under certain conditions, the real ultrasonic images that meet the requirements are screened out to form a set of real ultrasonic images. By cropping the real ultrasonic images to determine the training data, the training data can be made more standardized and concentrated, which is beneficial to improving the effect and accuracy of model training, providing high-quality data support for model training, and further enhancing the performance of the entire system in ultrasonic image processing and analysis.

[0143] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and modifications.

Claims

1. Super-resolution ultrasound imaging microbubble localization system based on deep learning, characterized in that: include: A model training module is used to obtain real ultrasound images and use the real ultrasound images as training data to train the neural network model to obtain a deep learning model; An image localization module is used to use the segMamba model as a baseline model. At the same time, an attention mechanism is added to the deep learning model, and image segmentation is performed on the image to be processed based on the added result combined with the baseline model, and the microbubble position is obtained based on the image segmentation result; The image tracking module is used to find the global minimum of the LSA cost function based on the Hungarian algorithm, and track the microbubbles according to the microbubble positions based on the search results.

2. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 1, characterized in that: Model training module, including: An image acquisition unit, configured to acquire an image acquisition request based on a management terminal, access a historical image database based on the image acquisition request, and conditionally traverse historical images in the historical image database based on the access result to obtain a set of real ultrasound images; The training data determination unit is used to perform image cropping on each real ultrasound image in the real ultrasound image set based on a preset cropping mode, and obtain training data based on the image cropping result.

3. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 2, characterized in that: An image acquisition unit, comprising: An image retrieval subunit, used to retrieve the acquired real ultrasound image set and determine an image processing strategy for the real ultrasound image set, wherein the image processing strategy includes contrast adjustment and image smoothing processing; The image processing subunit is used to process the real ultrasound image set based on the image processing strategy.

4. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 1, characterized in that: Model training module, including: A parameter setting unit, used to obtain the obtained training data, and at the same time, obtain the configuration parameters of the Gaussian function, and set the configuration parameters of the Gaussian function to a 3*3 kernel and σ=1 based on the training requirements; Model training unit, used to: Convolve the Gaussian function after the configuration parameters are set with the training data respectively, and obtain the heat map corresponding to the training data based on the convolution operation result, and set the heat map as the label of the training data; Based on the set results, the neural network model is iteratively trained for the target number of times according to the labels and training data, and the conditional convergence state of the neural network model is monitored in real time during the iterative training process. When the preset convergence conditions are met, the training is terminated to obtain a deep learning model.

5. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 4, characterized in that: Model training unit, including: The model testing subunit is used to obtain the obtained deep learning model, and at the same time, retrieve the test data set from the historical database, and input the test data set into the deep learning model for processing, and obtain the processing process parameters and target results of the deep learning model based on the processing results; Model loss analysis subunit, used to: Analyze the processing parameters and target results of the deep learning model based on a preset loss function, determine the loss value of the deep learning model, and compare the loss value with a preset threshold; If the loss value is less than or equal to the preset threshold, the deep learning model is judged to meet the preset convergence condition, and the final deep learning model is obtained based on the judgment result.

6. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 1, characterized in that: Image positioning module, including: A parameter acquisition unit, used for acquiring a segMamba model based on a management terminal; Model optimization unit for: The linear time series pre-trained by the segMamba model is set as the baseline model of the encoder backbone in the deep learning model, and the image to be processed is received and processed based on the set result to extract the image features of the image to be processed; At the same time, the target requirements for microbubble positioning are obtained based on the management terminal, and the attention mechanism corresponding to the image segmentation is generated based on the target requirements; The attention mechanism is added to the encoder branch of the deep learning model, and the image features of the extracted image to be processed are decoded based on the added result; The microbubble positioning unit is used to lock the key image area of ​​the decoding result based on the attention mechanism, and to perform image segmentation on the image to be processed based on the key image area locking result to obtain the microbubble position in the image to be processed.

7. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 1, characterized in that: Image tracking module, including: Result parsing unit, used to: Obtaining positioning results of different microbubble positions in the image to be processed of the current frame in the image sequence to be processed, and recording them as P{i1}, P{i2}, ..., P{in}; At the same time, the positioning results of different microbubble positions in the next frame of the image to be processed are obtained based on the deep learning model and recorded as P{j1}, P{j2}, ..., P{jm}; Pairing the localization results of different microbubble positions in the to-be-processed images of adjacent frames based on the LSA cost function, and obtaining corresponding cost function values ​​based on the pairing results; An optimization unit, used for expressing the cost function value as a target matrix, searching for a global minimum of the target matrix based on the Hungarian algorithm, and obtaining a pairing result between the positioning results of different microbubble positions in the to-be-processed images of adjacent frames based on the searching result; The microbubble tracking unit is used to track the same microbubble in the to-be-processed images of different frames based on the pairing result to obtain a corresponding tracking path.

8. The super-resolution ultrasound imaging microbubble localization system based on deep learning according to claim 7, characterized in that: Microbubble tracking unit, including: The result analysis subunit is used to: Acquire the tracking path of the microbubble, and determine the distribution and displacement of the same microbubble between the images to be processed in adjacent frames based on the tracking path; Accumulating the distribution and displacement of microbubbles between the images to be processed in adjacent frames to obtain the density distribution and velocity parameters of the microbubbles; The image generation subunit is used to generate density and velocity images of microbubbles based on the density distribution and velocity parameters of the microbubbles, and feed the generated density and velocity images of the microbubbles back to the management terminal for recording.

9. A super-resolution ultrasound imaging microbubble localization method based on deep learning, characterized in that: include: Step 1: Obtain a real ultrasound image and use the real ultrasound image as training data to train the neural network model to obtain a deep learning model; Step 2: Use the segMamba model as the baseline model. At the same time, add an attention mechanism to the deep learning model, and perform image segmentation on the processed image based on the added results combined with the baseline model, and obtain the microbubble position based on the image segmentation results; Step 3: Find the global minimum of the LSA cost function based on the Hungarian algorithm, and track the microbubbles according to their positions based on the search results.

10. The method for microbubble localization based on super-resolution ultrasound imaging based on deep learning according to claim 9, characterized in that: In step 1, a real ultrasound image is obtained and used as training data to train the neural network model, including: Obtaining an image retrieval request based on a management terminal, accessing a historical image database based on the image retrieval request, and conditionally traversing historical images in the historical image database based on the access result to obtain a set of real ultrasound images; Image cropping is performed on each real ultrasound image in the real ultrasound image set based on a preset cropping mode, and training data is obtained based on the image cropping result.

Images