Ultrasound artery segmentation method, system, and apparatus based on knowledge-driven positioning

Abstract

The application provides an ultrasound artery segmentation method, system and device based on positioning knowledge driving, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring an original ultrasound image, wherein the original ultrasound image is an image containing an arterial lumen; based on a positioning prior learning model, rough region mask mapping of the original ultrasound image is realized to obtain a positioning map containing an arterial highlight region; the original ultrasound image and the positioning map are input into a fusion model to obtain a fusion feature map; based on a segmentation network model, the fusion feature map is segmented to obtain a target segmentation image; wherein the fusion model is a model of guided adaptive fusion. The method effectively solves the positioning ambiguity problem of the arterial lumen in the ultrasound image, ensures the correct positioning of the target, thereby making the subsequent segmentation more stable and accurate, and effectively avoiding missed segmentation and erroneous segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method, system, and device for ultrasound artery segmentation based on localization knowledge. Background Technology

[0002] Ultrasound imaging, due to its advantages of being non-invasive, safe, economical, and real-time, has become a first-line tool for arterial assessment, clinical screening, and health monitoring. Precise segmentation of the arterial lumen in ultrasound images is fundamental to assessing vascular stenosis, measuring cross-sectional area, and estimating blood flow, directly impacting the accuracy of clinical decisions.

[0003] However, automatic segmentation of arterial lumens faces significant challenges, primarily due to the inherent characteristics of ultrasound images. First, the arterial target is highly similar in features to surrounding tissues (such as veins) and imaging artifacts, and the low image contrast introduces considerable interference and ambiguity to accurate lumen localization—a phenomenon known as "localization ambiguity." Second, because the ultrasound beam is nearly perpendicular to the vessel wall, the lumen outline, especially the wall, often becomes blurred or lacks signal, further complicating localization and precise contouring.

[0004] To address these challenges, existing technologies primarily employ implicit strategies to handle the localization problem, but with limited effectiveness. For example, data augmentation-based methods rely on techniques such as random cropping and translation to implicitly enhance the model's learning of target location information during training. However, this approach cannot fundamentally resolve the inherent localization ambiguity of the image content itself. Attention-based methods introduce location or spatial attention modules into the segmentation network, attempting to enable the model to autonomously learn and focus on the target region. However, these methods often lack generalization ability, and their performance significantly degrades when faced with poor image quality or complex backgrounds. Shape-prior-based methods introduce fixed shape priors, such as assuming an elliptical cross-section of a blood vessel, to constrain the segmentation results. This approach lacks flexibility and struggles to handle various irregular blood vessel morphologies caused by pathological changes or imaging angles. Furthermore, regarding cascaded or coarse-to-fine processing, some existing methods employ a two-stage process: first, a detection network locates a region of interest, and then a segmentation network performs fine segmentation within that region. The final performance of such methods depends heavily on the accuracy of the first-stage localization, and is prone to error propagation problems.

[0005] Therefore, most existing technologies handle the localization problem implicitly, failing to effectively and directly solve the ambiguity of arterial lumen localization in ultrasound images, thus limiting the accuracy and robustness of segmentation. Summary of the Invention

[0006] A first aspect of the present invention provides a localization knowledge-driven ultrasound artery segmentation method, comprising: acquiring an original ultrasound image, wherein the original ultrasound image contains an arterial lumen; performing coarse region mask mapping on the original ultrasound image based on a localization prior learning model to obtain a localization map containing a highlighted region of the artery; inputting the original ultrasound image and the localization map into a fusion model to obtain a fusion feature map; and segmenting the fusion feature map based on a segmentation network model to obtain a target segmented image; wherein the fusion model is a guided adaptive fusion model.

[0007] According to an embodiment of the present invention, the localization prior learning model is obtained through pre-training. The training process of the localization prior learning model includes: acquiring a first training set, which is a coarsely labeled training set, and the first training set includes multiple training images containing arterial lumens; inputting the first training set into the localization prior learning network, and mapping the coarse region mask of each training image through a composite loss function.

[0008] According to an embodiment of the present invention, the training process of the localization prior learning model further includes: acquiring a second training set, wherein the second training set is a finely annotated training set; selecting a set of sample images from the second training set according to a preset ratio; generating a corresponding pseudo-label dataset from the sample image set based on a preset geometric algorithm; fine-tuning the localization prior learning model based on the pseudo-label dataset; and outputting a localization map containing the artery highlight region.

[0009] According to an embodiment of the present invention, the fusion model is a multilayer perceptron structure. The original ultrasound image and the localization map are input into the fusion model to obtain a fusion feature map, including: calculating the average value of the original ultrasound image and the average value of the localization map; inputting the average value of the original ultrasound image and the average value of the localization map into the multilayer perceptron structure to generate a dynamic weight map containing two channels; and performing channel stitching based on the dynamic weight map to obtain the fusion feature map.

[0010] According to an embodiment of the present invention, the segmentation network model is pre-trained, and the training process of the segmentation network model includes: regularizing the segmentation network model by smoothing the label boundaries.

[0011] According to an embodiment of the present invention, regularizing a segmentation network model by smoothing label boundaries includes: obtaining a third training set, the third training set including multiple finely labeled binary ground truth masks; determining the boundary of each binary ground truth mask using morphological gradients; performing convolution processing on the boundary of each binary ground truth mask using a preset filter; integrating the convolution-processed boundary and the binary ground truth mask to generate soft label data; and regularizing the segmentation network model based on the soft label data.

[0012] According to an embodiment of the present invention, a target segmented image is obtained by segmenting a fused feature map based on a segmentation network model, including: inputting the fused feature map into a pre-trained segmentation network model, and using a composite loss function based on soft-label data as a standard to complete the segmentation of the fused feature map and obtain the predicted target segmented image.

[0013] According to an embodiment of the present invention, the target segmentation image is a probability map, and the method further includes: judging the target segmentation image based on a preset threshold to obtain a target binary segmentation mask.

[0014] A second aspect of the present invention provides a localization knowledge-driven ultrasound artery segmentation system, comprising: a raw image acquisition module for acquiring a raw ultrasound image, wherein the raw ultrasound image is an image containing the arterial lumen; a localization prior learning module for performing coarse region mask mapping on the raw ultrasound image based on a localization prior learning model to obtain a localization map containing the arterial highlight region; a fusion module for inputting the raw ultrasound image and the localization map into a fusion model to obtain a fused feature map; and a segmentation module for segmenting the fused feature map based on a segmentation network model to obtain a target segmented image; wherein the fusion model is a guided adaptive fusion model.

[0015] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0016] The technical solutions of the embodiments of the present invention have at least the following beneficial effects:

[0017] (1) This ultrasound artery segmentation method based on localization knowledge significantly improves segmentation accuracy and robustness. Compared with existing techniques that rely on implicit strategies, this method fundamentally solves the localization ambiguity problem through explicit localization guidance provided by a localization prior learning model. Even in challenging images with low contrast, high noise, and blurred boundaries, it can first ensure the correct localization of the target, thereby making subsequent segmentation more stable and accurate, and effectively avoiding missed segmentation and incorrect segmentation.

[0018] (2) The ultrasound artery segmentation method based on localization knowledge achieves excellent data efficiency and generalization ability. Compared with traditional methods that require a large amount of target domain data for training, the proposed few-sample pseudo-label fine-tuning strategy has great advantages. It can quickly adapt general localization knowledge to new domains with only a very small amount of target data, greatly reducing the cost of data annotation and demonstrating strong domain adaptability.

[0019] (3) This ultrasound artery segmentation method based on localization knowledge improves training efficiency and model stability. This method decouples the complex segmentation task into two relatively simple sub-tasks: localization and localization-guided segmentation. This decoupling makes the learning process more stable and efficient. The localization prior learning model provides a powerful prior, which accelerates the convergence of the segmentation network model and reduces training time.

[0020] (4) The ultrasound artery segmentation method based on localization knowledge has stronger interpretability and flexibility. The method can intuitively understand the model through the visualization of localization map, which provides interpretability for the model's decision-making. At the same time, the guided attention fusion mechanism is more flexible than simple feature stitching or fixed shape prior, and can adaptively adjust the intensity of localization guidance and the weight of image details according to the specific content of each image. Attached Figure Description

[0021] Figure 1 A flowchart illustrating a location knowledge-driven ultrasound artery segmentation method according to an embodiment of the present invention is shown schematically.

[0022] Figure 2 A schematic diagram of a location-knowledge-driven ultrasonic artery segmentation system according to an embodiment of the present invention is shown.

[0023] Figure 3 A block diagram of an electronic device suitable for implementing a location knowledge-driven ultrasonic artery segmentation method according to an embodiment of the present invention is shown schematically. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0025] Figure 1 A flowchart illustrating a location-knowledge-driven ultrasound artery segmentation method according to an embodiment of the present invention is shown.

[0026] Please refer to the following for details. Figure 1 In an embodiment of the present invention, the specific process of the ultrasound artery segmentation method driven by localization knowledge includes operations S110 to S140.

[0027] In operation S110, a raw ultrasound image is acquired, which is an image containing the arterial lumen.

[0028] In embodiments of the present invention, such as in clinical screening or arterial assessment scenarios, after obtaining the patient's original ultrasound images (which contain arterial lumens), it is necessary to accurately segment the arterial lumens in the ultrasound images in order to assess clinical indicators such as vascular stenosis.

[0029] In operation S120, based on the localization prior learning model, a coarse region mask mapping is performed on the original ultrasound image to obtain a localization map containing the artery highlight region.

[0030] In embodiments of the present invention, the localization prior learning model is pre-trained, and the training process of the localization prior learning model includes a training phase and a fine-tuning phase. The training process of the localization prior learning model is described in detail below.

[0031] According to an embodiment of the present invention, the training phase of the localization prior learning model includes: acquiring a first training set, wherein the first training set is a coarsely labeled training set, and the first training set includes multiple training images containing arterial lumens; inputting the first training set into the localization prior learning network, and mapping coarse region masks to each training image through a composite loss function.

[0032] For example, the first training set here can be a publicly available arterial ultrasound image dataset (i.e., multiple images containing arterial lumens to be trained). This arterial ultrasound dataset is a source dataset with coarse annotations, such as a simple circular region YC (Yaw Circle, YC) defined by a center point and radius.

[0033] Furthermore, the coarsely labeled arterial ultrasound dataset is input into the constructed localization prior learning network for model training. For example, the localization prior learning network can adopt a lightweight U-shaped structure or a segmentation network with an encoder-decoder structure.

[0034] Furthermore, during model training, the network is trained to learn the mapping from the input ultrasound image to its coarse region mask by minimizing the segmentation loss function (i.e., the composite loss function).

[0035] Understandably, the training phase of the above model enables the localization prior learning model to determine the basic spatial location and morphological features of arteries in ultrasound images.

[0036] According to an embodiment of the present invention, the fine-tuning stage of the localization prior learning model includes: acquiring a second training set, wherein the second training set is a finely annotated training set; selecting a set of sample images from the second training set according to a preset ratio; generating a corresponding pseudo-label dataset from the sample image set based on a preset geometric algorithm; fine-tuning the localization prior learning model based on the pseudo-label dataset, and outputting a localization map containing the artery highlight region.

[0037] For example, the second training set here could also be a publicly available arterial ultrasound image dataset, but this dataset would be a target dataset with fine annotations.

[0038] Furthermore, a small subset of samples from the target dataset is selected according to a preset proportion (e.g., 10%). For each image data sample in this dataset, coarse pseudo-label data is generated based on its fine annotation using a preset geometric algorithm, thus obtaining the pseudo-label dataset.

[0039] It should be noted that this embodiment uses the circumcircle as the pseudo-label for the best effect, but other geometric shapes can also be used as alternatives, such as: the minimum circumcircle rectangle generated from the truth mask, the ellipse fitted by methods such as principal component analysis, or the region obtained by performing a simple morphological dilation operation on the truth mask.

[0040] Furthermore, the newly generated pseudo-label dataset is used to fine-tune the localization prior learning model that has completed the training phase, thereby completing the entire training process. It is understandable that the aforementioned fine-tuning based on few-shot pseudo-labels efficiently transfers and adapts the general localization knowledge acquired during the training phase to a specific target dataset.

[0041] Through the embodiments of the present invention, the proposed few-sample pseudo-label fine-tuning strategy has great advantages. It can quickly adapt general positioning knowledge to new fields with only a small amount of target data, greatly reducing the cost of data annotation and demonstrating strong domain adaptability.

[0042] Furthermore, for the localization prior learning model that completes the entire training process, given any input original ultrasound image I, the model can generate a high-quality, single-channel localization map M. g Location map M g The value is between [0, 1], highlighting the area where the artery is located.

[0043] Through embodiments of the present invention, this method fundamentally solves the problem of localization ambiguity by providing explicit localization guidance through a localization prior learning model. Even in challenging images with low contrast, high noise, and blurred boundaries, it can first ensure the correct localization of the target, thereby making subsequent segmentation more stable and accurate, and effectively avoiding missed and incorrect segmentation. In addition, through the visualization of the localization map, it is possible to intuitively understand "where the model first saw," providing interpretability for the model's decision-making.

[0044] In operation S130, the original ultrasound image and localization map are input into the fusion model to obtain the fusion feature map.

[0045] In an embodiment of the present invention, the fusion model is a guided adaptive fusion model.

[0046] For example, when the fusion model is a guided attention fusion model, the fusion model is a multilayer perceptron structure. The original ultrasound image and the localization map are input into the fusion model to obtain a fusion feature map, including: calculating the average value of the original ultrasound image and the average value of the localization map; inputting the average value of the original ultrasound image and the average value of the localization map into the multilayer perceptron structure to generate a dynamic weight map containing two channels; and performing channel stitching based on the dynamic weight map to obtain the fusion feature map.

[0047] Specifically, the original ultrasound image I and the localization map M g The inputs to the guided attention fusion model are the original ultrasound image I and the localization map M. g The average value of the original ultrasound image I and the localization map M. g The average values ​​are input into a multilayer perceptron (MLP) structure, and the output, after passing through an activation function, generates a two-channel dynamic weight map W = [I, M]. g The weights of these two channels are applied to the original ultrasound image I and the localization map M, respectively. g Each pixel position. Furthermore, based on the dynamic weight map, channel concatenation is performed to obtain an enhanced, dual-channel fused feature map I. fused .

[0048] For example, the expression for the fusion process is as follows:

[0049]

[0050]

[0051] in, Element-wise multiplication is represented by W, concat represents channel concatenation, softmax represents the normalized exponential function, and W represents the normalized exponential function. I Wg represents the fused image of the original ultrasound images, and Wg represents the fused image of the localization map.

[0052] It should be noted that the above embodiment performs guided attention fusion at the input layer. In another embodiment, the localization map M... g Or its features can also be injected into different levels of the segmentation network (e.g., multiple stages of the encoder or multiple stages of the decoder) through different fusion methods (such as convolution after splicing, element-wise addition, or more complex attention gating mechanisms) to achieve adaptive fusion guidance.

[0053] Through the embodiments of the present invention, the guided attention fusion mechanism is more flexible than simple feature stitching or fixed shape priors, and can adaptively adjust the intensity of localization guidance and the weight of image details according to the specific content of each image.

[0054] In operation S140, based on the segmentation network model, the fused feature map is segmented to obtain the target segmented image.

[0055] In an embodiment of the present invention, the segmentation network model is pre-trained, and the training process of the segmentation network model includes: regularizing the segmentation network model by smoothing the label boundaries.

[0056] According to an embodiment of the present invention, regularizing a segmentation network model by smoothing label boundaries includes: obtaining a third training set, the third training set including multiple finely labeled binary ground truth masks; determining the boundary of each binary ground truth mask using morphological gradients; performing convolution processing on the boundary of each binary ground truth mask using a preset filter; integrating the convolution-processed boundary and the binary ground truth mask to generate soft label data; and regularizing the segmentation network model based on the soft label data.

[0057] Understandably, in order to address the problem of blurred boundaries in ultrasound images, the model is regularized by smoothing the label boundaries during the training of the segmentation network, enabling it to better handle the uncertainty of the boundaries.

[0058] Specifically, a third training set is obtained, which can be a set of finely labeled binary ground truth masks. For each finely labeled binary (0 or 1) ground truth mask Y, its boundary region B is first extracted using morphological gradient operations, as shown in the following formula:

[0059]

[0060] Where K is the structuring element and Y represents the truth mask. This indicates element-wise addition. This indicates element-wise subtraction.

[0061] Furthermore, for example, using a Gaussian filter The extracted boundary region B is convolved (blurred).

[0062] Furthermore, the blurred boundary is integrated with the original label (i.e., the binary truth mask) to generate the soft label Y. soft The value inside the lumen remains 1, the value outside remains 0, while the value at the boundary is a smooth transition value between 0 and 1. As shown in the following formula:

[0063]

[0064] Where Y represents the truth mask and B represents the boundary region. This represents the weight matrix calculated by the Gaussian filter.

[0065] Based on the obtained soft label Ysoft Implement regularization processing for the segmentation network model.

[0066] Furthermore, after obtaining the training data corresponding to the soft labels and the fused feature maps, the constructed segmentation network model is trained using the training data of the soft labels and the training data of the fused feature maps. Specifically, the training data corresponding to the fused feature maps is used as input, and the training data of the soft labels is used as the supervision signal. End-to-end training is performed using a composite loss function to complete the training of the segmentation network model.

[0067] It should be noted that the segmentation network model here can be a lightweight U-shaped structure or a segmentation network with an encoder-decoder structure. The composite loss function here can be a combination of binary cross-entropy loss (BCE) and Dice coefficient (Dice) loss, or a loss function that optimizes class imbalance or boundary segmentation.

[0068] In an embodiment of the present invention, the fused feature map is segmented based on a segmentation network model to obtain a target segmented image. This includes: inputting the fused feature map into a pre-trained segmentation network model, and using a composite loss function based on soft-label data to complete the segmentation of the fused feature map, thereby obtaining a predicted target segmented image. The target segmented image is a probability map. Further, based on a preset threshold, the target segmented image is judged to obtain a target binary segmentation mask.

[0069] Specifically, based on the trained segmentation network model, the fused feature map obtained by fusing any input original ultrasound image and localization map is input into the segmentation network model, and a composite loss function is used based on soft label data to complete the segmentation of the fused feature map and obtain the predicted target segmented image.

[0070] It should be noted that the target segmentation image is a probability map. Therefore, the target segmentation image is judged by a preset threshold (e.g., 0.5) to obtain the target binary segmentation mask.

[0071] Based on the localization-driven ultrasound artery segmentation method of this embodiment, the method was tested on two publicly available arterial ultrasound datasets. The segmentation accuracy of this embodiment significantly surpasses that of many mainstream and state-of-the-art segmentation models. Through explicit localization guidance, the model training convergence speed is improved by approximately 40% compared to traditional end-to-end models. Furthermore, only 10% of the samples in the target dataset need to be fine-tuned to achieve efficient adaptation and superior performance to new datasets.

[0072] Through the embodiments of the present invention, based on the above-mentioned localization knowledge-driven segmentation framework, the two tasks of localization and segmentation are decoupled. By explicitly learning and utilizing localization knowledge to guide the segmentation process, the problem of the ambiguity in the localization of arterial lumens in ultrasound images, which limits the accuracy and robustness of segmentation, is fundamentally solved.

[0073] Based on the aforementioned location-knowledge-driven ultrasound artery segmentation method, this invention also provides a location-knowledge-driven ultrasound artery segmentation system. The following will combine... Figure 2 The device is described in detail.

[0074] Figure 2 A schematic diagram of a location-knowledge-driven ultrasonic artery segmentation system according to an embodiment of the present invention is shown.

[0075] like Figure 2 As shown, the ultrasound artery segmentation system 200 based on localization knowledge in this embodiment includes an original image acquisition module 210, a localization prior learning module 220, a fusion module 230, and a segmentation module 240.

[0076] The raw image acquisition module 210 is used to acquire raw ultrasound images, which are images containing the arterial lumen. In one embodiment, the raw image acquisition module 210 can be used to perform the operation S110 described above, which will not be repeated here.

[0077] The localization prior learning module 220 is used to perform coarse region mask mapping on the original ultrasound image based on the localization prior learning model to obtain a localization map containing the arterial highlight region. In one embodiment, the localization prior learning module 220 can be used to perform the operation S120 described above, which will not be repeated here.

[0078] The fusion module 230 is used to input the original ultrasound image and the localization map into the fusion model to obtain a fused feature map. In one embodiment, the fusion module 230 can be used to perform the operation S130 described above, which will not be repeated here.

[0079] The segmentation module 240 is used to segment the fused feature map based on a segmentation network model to obtain a target segmented image. In one embodiment, the segmentation module 240 can be used to perform the operation S140 described above, which will not be repeated here.

[0080] According to embodiments of the present invention, any multiple modules among the original image acquisition module 210, localization prior learning module 220, fusion module 230, segmentation module 240, basic information feature extraction module 250, and output module 260 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of the present invention, at least one of the original image acquisition module 210, localization prior learning module 220, fusion module 230, segmentation module 240, basic information feature extraction module 250, and output module 260 can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or any other reasonable means of integrating or packaging circuits, or implemented in hardware or firmware, or in any one of software, hardware, and firmware implementations, or in a suitable combination of any of these. Alternatively, at least one of the original image acquisition module 210, localization prior learning module 220, fusion module 230, segmentation module 240, basic information feature extraction module 250, and output module 260 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0081] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0082] Figure 3 A block diagram of an electronic device suitable for implementing a location knowledge-driven ultrasonic artery segmentation method according to an embodiment of the present invention is shown schematically.

[0083] like Figure 3As shown, an electronic device 300 according to an embodiment of the present invention includes a processor 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage portion 308 into a random access memory (RAM) 303. The processor 301 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 301 may also include onboard memory for caching purposes. The processor 301 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0084] RAM 303 stores various programs and data required for the operation of electronic device 300. Processor 301, ROM 302, and RAM 303 are interconnected via bus 304. Processor 301 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 302 and / or RAM 303. It should be noted that programs may also be stored in one or more memories other than ROM 302 and RAM 303. Processor 301 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in one or more memories.

[0085] According to an embodiment of the present invention, the electronic device 300 may further include an input / output (I / O) interface 305, which is also connected to a bus 304. The electronic device 300 may also include one or more of the following components connected to the input / output (I / O) interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN card, modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the input / output (I / O) interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 310 as needed so that computer programs read from it can be installed into the storage section 308 as needed.

[0086] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0088] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A location-knowledge-driven ultrasound artery segmentation method, characterized in that, The method includes: Acquire raw ultrasound images, which are images containing the lumen of an artery; Based on the localization prior learning model, a coarse region mask mapping is performed on the original ultrasound image to obtain a localization map containing the artery highlight region. The original ultrasound image and localization map are input into the fusion model to obtain the fusion feature map; Based on the segmentation network model, the fused feature map is segmented to obtain the target segmented image; The fusion model is a guided adaptive fusion model; The localization prior learning model is pre-trained, and the training process of the localization prior learning model includes: Obtain a first training set, which is a coarsely labeled training set, and the first training set includes multiple training images containing arterial lumens; The first training set is input into the localization prior learning network, and a coarse region mask is mapped to each image to be trained through a composite loss function. The training process of the localization prior learning model also includes: Obtain the second training set, which is a finely labeled training set; According to a preset ratio, select the sample image set from the second training set; Based on a preset geometric algorithm, the sample image set is used to generate a corresponding pseudo-label dataset; The localization prior learning model is fine-tuned based on the pseudo-label dataset, and a localization map containing the highlighted artery region is output.

2. The ultrasound artery segmentation method based on localization knowledge as described in claim 1, characterized in that, The fusion model is a multilayer perceptron structure. The process of inputting the original ultrasound image and localization map into the fusion model to obtain a fused feature map includes: The average value of the original ultrasound image and the localization map is calculated; The average values ​​of the original ultrasound image and the localization map are input into the multilayer perceptron structure to generate a dynamic weight map containing two channels. Based on the dynamic weight map, channel splicing is performed to obtain a fused feature map.

3. The ultrasound artery segmentation method based on localization knowledge as described in claim 1, characterized in that, The segmentation network model is pre-trained, and the training process of the segmentation network model includes: The segmentation network model is regularized by smoothing the label boundaries.

4. The ultrasound artery segmentation method based on localization knowledge as described in claim 3, characterized in that, The regularization of the segmentation network model by smoothing label boundaries includes: Obtain a third training set, which includes multiple finely labeled binary ground truth masks; The boundaries of each binary truth mask are determined using morphological gradients. The boundaries of each binary true value mask are convolved using a preset filter. The boundary and binary ground truth mask after convolution are integrated to generate soft label data; Based on the soft-label data, the segmentation network model is regularized.

5. The ultrasound artery segmentation method based on localization knowledge as described in claim 4, characterized in that, The step of segmenting the fused feature map based on the segmentation network model to obtain the target segmented image includes: The fused feature map is input into a pre-trained segmentation network model, and a composite loss function is used based on the soft-label data to complete the segmentation of the fused feature map and obtain the predicted target segmentation image.

6. The ultrasound artery segmentation method based on localization knowledge as described in claim 1, characterized in that, The target segmentation image is a probability map, and the method further includes: Based on a preset threshold, the target segmented image is judged to obtain a target binary segmentation mask.

7. A localization-based knowledge-driven ultrasound arterial segmentation system, characterized in that, include: The original image acquisition module is used to acquire the original ultrasound image, which is an image containing the arterial lumen; The localization prior learning module is used to perform coarse region mask mapping on the original ultrasound image based on the localization prior learning model, so as to obtain a localization map containing the artery highlight region. The fusion module is used to input the original ultrasound image and localization map into the fusion model to obtain a fusion feature map; The segmentation module is used to segment the fused feature map based on a segmentation network model to obtain a target segmented image; The fusion model is a guided adaptive fusion model; The localization prior learning model is pre-trained, and the training process of the localization prior learning model includes: Obtain a first training set, which is a coarsely labeled training set, and the first training set includes multiple training images containing arterial lumens; The first training set is input into the localization prior learning network, and a coarse region mask is mapped to each image to be trained through a composite loss function. The training process of the localization prior learning model also includes: Obtain the second training set, which is a finely labeled training set; According to a preset ratio, select the sample image set from the second training set; Based on a preset geometric algorithm, the sample image set is used to generate a corresponding pseudo-label dataset; The localization prior learning model is fine-tuned based on the pseudo-label dataset, and a localization map containing the highlighted artery region is output.

8. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 6.

Images