A method and system for detecting a bleeding location in a human body based on an ultrasound image

Abstract

This invention provides a method and system for detecting the location of internal bleeding in the human body based on ultrasound images, belonging to the field of ultrasound detection technology. The method includes the following steps: constructing an image dataset using acquired ultrasound images containing bleeding points; constructing and training an improved neighborhood attention converter model based on the image dataset; acquiring the ultrasound image to be detected, using the trained improved neighborhood attention converter model to detect the bleeding location, and obtaining the detection result; and visualizing the detection result using a visualization tool. This invention applies deep learning networks to the detection of bleeding points in ultrasound images. With the support of the dataset, the deep learning network can achieve higher detection accuracy and better noise robustness compared to traditional algorithms. For the small target detection task of bleeding points, specific optimizations have been made at the network design level, resulting in a better detection rate.

Description

Technical Field

[0001] This invention relates to the field of ultrasound detection technology, and more specifically, to a method and system for detecting the location of internal bleeding in the human body based on ultrasound images. Background Technology

[0002] Currently, ultrasound imaging technology is one of the most widely used technologies in the field of medical imaging, and the results of ultrasound imaging can provide important guidance in the clinical diagnosis process. For the detection of internal bleeding, other methods, such as endoscopy, are invasive and require stringent clinical conditions, making them unsuitable for field diagnosis and treatment of injured persons. Therefore, using non-invasive and convenient ultrasound imaging to locate the location of internal bleeding is of great practical significance for field diagnosis and timely treatment.

[0003] The task of detecting the location of bleeding using ultrasound imaging is usually performed by physicians with the assistance of ultrasound imaging equipment (systems). In actual clinical diagnosis, the accurate determination of whether internal bleeding exists and the location of the bleeding point is closely related to subjective factors such as the physician's clinical experience. Diagnosing and treating a large number of wounded soldiers with internal bleeding in the field places a heavy burden on physicians and is prone to misdiagnosis; therefore, there is an urgent need for intelligent diagnostic systems to assist physicians in providing diagnostic opinions.

[0004] In the task of bleeding point detection, current deep learning methods primarily focus on bleeding point detection based on endoscopic images and fundus images. However, there is a lack of relevant deep learning models for bleeding point detection in ultrasound images. Generally speaking, bleeding point detection can be considered a special type of object detection task, where object detection networks can be adopted and specifically improved upon. In 2023, Ali Hassani et al. proposed the Neighborhood Attention Transformer (NA Transformer) for computer vision tasks including object detection, achieving good results. Its model structure is as follows: Figures 3-5 As shown.

[0005] Currently, there is no network designed specifically for detecting bleeding points in ultrasound images. Compared to object detection in computer vision, bleeding point detection only requires the location of the bleeding point and does not require additional classification. Furthermore, considering the small sample size of current ultrasound image bleeding point detection datasets, this paper proposes to modify the NA Transformer to achieve intelligent detection of bleeding point locations in ultrasound images.

[0006] Traditional bleeding point detection methods employ machine learning approaches, comprising two parts: manual feature extraction and a feature classifier. Feature extraction typically focuses on texture features and color space features, utilizing statistical machine learning methods for feature statistical analysis. Traditional machine learning methods require manual feature design and classifier selection, a cumbersome process reliant on engineer experience and inherently prone to uncertainty. In contrast, this invention applies an improved neighborhood attention converter method to bleeding point detection in ultrasound images, aiming to provide bleeding point detection results that are less susceptible to subjective interference and have a simpler process.

[0007] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0008] In view of this, the present invention provides a method and system for detecting the location of internal bleeding in the human body based on ultrasound images, so as to solve the above-mentioned problems.

[0009] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0010] According to one aspect of the present invention, a method for detecting the location of internal bleeding in the human body based on ultrasound images is provided, the method comprising the following steps:

[0011] S1. Construct an image dataset using the acquired ultrasound images containing bleeding points;

[0012] S2. Based on the image dataset, construct and train an improved neighborhood attention converter model;

[0013] S3. Acquire the ultrasound image to be detected, and use the trained and improved neighborhood attention converter model to detect the bleeding location and obtain the detection result;

[0014] S4. Use visualization tools to visualize the test results.

[0015] Preferably, constructing an image dataset using acquired ultrasound images containing bleeding points includes the following steps:

[0016] S11. Acquire multiple consecutive ultrasound images containing bleeding points and extract semantic features through feature convolution processing;

[0017] S12. Perform dilated convolution processing on the extracted semantic features using different dilation rates, and perform fusion analysis on the processing results to obtain fused features;

[0018] S13. Upsample the fusion features to obtain a cleared image of the current frame ultrasound image;

[0019] S14. Use annotation tools to annotate the bleeding points in the cleared ultrasound images to obtain an image dataset.

[0020] Preferably, acquiring multiple consecutive ultrasound images containing bleeding points and extracting semantic features through feature convolution processing includes the following steps:

[0021] S111. Perform image preprocessing on the acquired continuous multi-frame ultrasound images containing bleeding points. The preprocessing includes noise removal, contrast adjustment and image normalization.

[0022] S112. Convert the format of the preprocessed ultrasound image to obtain a standard ultrasound image;

[0023] S113. Using the data processing layer in the pre-defined neural network model, convolve the ultrasound standard image of the current frame to extract the semantic features corresponding to the ultrasound standard image of the current frame.

[0024] Preferably, the extracted semantic features are subjected to dilated convolution processing using different dilation rates, and the processing results are fused and analyzed to obtain fused features, including the following steps:

[0025] S121. Use dilated convolution with different dilation rates to process semantic features and obtain a semantic feature dilation map;

[0026] S122. Perform dimensionality reduction and activation function mapping on the semantic feature expansion map to obtain the attention feature map;

[0027] S123. Concatenate all attention feature maps to obtain a secondary attention feature map;

[0028] S124. Multiply the secondary attention feature map and the semantic features of the ultrasound standard image to obtain the fused features.

[0029] Preferably, upsampling the fused features to obtain a sharpened image of the current frame ultrasound image includes the following steps:

[0030] S131. Upsample the fused features using the nearest neighbor interpolation method to obtain a primary upsampled image;

[0031] S132. Perform a convolution operation on the primary upsampled image to obtain a convolution-enhanced image;

[0032] S133. Perform upsampling processing on the convolutional enhanced image to obtain a secondary upsampled image;

[0033] S134. Perform a convolution operation on the secondary upsampled image to obtain a cleared image of the current frame ultrasound image.

[0034] Preferably, building and training an improved neighborhood attention converter model based on an image dataset includes the following steps:

[0035] S21. Divide the image dataset into a training set and a test set;

[0036] S22. Based on the improved neighborhood attention converter structure, construct an improved neighborhood attention converter model;

[0037] S23. The improved neighborhood attention converter model is trained using the training set, and the model parameters are continuously adjusted using the backpropagation algorithm and gradient descent method to minimize the loss function.

[0038] S24. After training, the performance of the improved neighborhood attention converter model is evaluated using the test set. Based on the evaluation results, the model structure is adjusted to obtain the final improved neighborhood attention converter model.

[0039] The preferred and improved neighborhood attention converter structure includes: an input layer, an overlapping word segmenter, a neighborhood attention block, an overlapping downsampler, a fully connected layer, and an output layer;

[0040] The size feature maps of the neighborhood attention block output are H / 4×W / 4 and H / 8×W / 8.

[0041] Preferably, training the improved neighborhood attention converter model using the training set and continuously adjusting the model parameters through backpropagation and gradient descent includes the following steps:

[0042] S231. Based on the improved neighborhood attention converter structure, initialize the weights and biases of all structural layers;

[0043] S232. Based on forward propagation, the training set is input into the improved neighborhood attention converter model, and the prediction results of bleeding points are generated through the processing of each layer of the structure.

[0044] S233. Use the loss function to evaluate the difference between the model prediction results and the true labels, and obtain the loss gradient;

[0045] S234. Based on the loss gradient, update the weights and biases of the improved neighborhood attention converter model using gradient descent based on backpropagation.

[0046] Preferably, the loss function is the normalized Vickers distance loss function.

[0047] According to another aspect of the present invention, a system for detecting the location of internal bleeding in the human body based on ultrasound images is provided. The system includes: a dataset construction module, a model construction module, a bleeding location detection module, and a visualization module.

[0048] The dataset building module is used to construct an image dataset from acquired ultrasound images containing bleeding points;

[0049] The model building module is used to build and train an improved neighborhood attention converter model based on an image dataset;

[0050] The bleeding location detection module is used to acquire the ultrasound image to be detected, and to detect the bleeding location using a trained and improved neighborhood attention converter model to obtain the detection result.

[0051] The visualization module is used to visualize the test results using visualization tools.

[0052] The beneficial effects of this invention are as follows:

[0053] 1. This invention applies deep learning networks to the detection of bleeding points in ultrasound images. With the support of the dataset, deep learning networks can achieve higher detection accuracy and better noise robustness compared to traditional algorithms. Traditional detection methods often can only extract a few features and are subject to strong human interference. In contrast, this invention uses deep learning networks, which eliminate the need for manual feature selection. The network can automatically extract features at multiple different levels. For the small target detection task of bleeding points, the network design has been specifically optimized to achieve a better detection rate.

[0054] 2. This invention can effectively improve the quality of ultrasound images by using feature convolution processing, dilated convolution processing, and upsampling processing, making them clearer and easier to identify bleeding points. By using different dilation rates to perform dilated convolution processing on semantic features and fusing and analyzing the processing results, richer feature information can be obtained, which helps to accurately detect bleeding points in the future. Bleeding point annotation processing is performed on the cleared ultrasound images. Since the image quality has been improved, the accuracy of the annotation will also be improved accordingly. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0056] Figure 1 This is a flowchart of a method for detecting the location of internal bleeding in the human body based on ultrasound images, according to an embodiment of the present invention.

[0057] Figure 2This is a schematic diagram of a human internal bleeding location detection system based on ultrasound images according to an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the principle of the neighborhood attention block before improvement in a method for detecting the location of internal bleeding in the human body based on ultrasound images according to an embodiment of the present invention;

[0059] Figure 4 This is a structural diagram of the neighborhood attention converter before improvement in a method for detecting the location of internal bleeding in the human body based on ultrasound images, according to an embodiment of the present invention.

[0060] Figure 5 This is a structural diagram of the neighborhood attention block before improvement in a method for detecting the location of internal bleeding in the human body based on ultrasound images, according to an embodiment of the present invention.

[0061] Figure 6 This is a schematic diagram of the structure of an improved neighborhood attention converter in a method for detecting the location of internal bleeding in the human body based on ultrasound images, according to an embodiment of the present invention.

[0062] In the picture:

[0063] 1. Dataset building module; 2. Model building module; 3. Bleeding location detection module; 4. Visualization module. Detailed Implementation

[0064] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0065] According to an embodiment of the present invention, a method and system for detecting the location of internal bleeding in the human body based on ultrasound images are provided.

[0066] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for detecting the location of internal bleeding in the human body based on ultrasound images is provided. This method includes the following steps:

[0067] S1. Construct an image dataset using the acquired ultrasound images containing bleeding points;

[0068] As a preferred embodiment, constructing an image dataset using acquired ultrasound images containing bleeding points includes the following steps:

[0069] S11. Acquire multiple consecutive ultrasound images containing bleeding points and extract semantic features through feature convolution processing;

[0070] As a preferred embodiment, acquiring multiple consecutive ultrasound images containing bleeding points and extracting semantic features through feature convolution processing includes the following steps:

[0071] S111. Perform image preprocessing on the acquired continuous multi-frame ultrasound images containing bleeding points. The preprocessing includes noise removal, contrast adjustment and image normalization.

[0072] It should be noted that noise removal of consecutive ultrasound images containing bleeding points reduces interference information in the images. Contrast adjustment enhances the contrast between bleeding points and other tissues, facilitating subsequent feature extraction. Image normalization scales the image data to a uniform numerical range, improving the model's stability and generalization ability.

[0073] S112. Convert the format of the preprocessed ultrasound image to obtain a standard ultrasound image;

[0074] The preprocessed ultrasound images are converted to a new format to ensure that the image data meets the input requirements of subsequent processing steps, resulting in standard ultrasound images.

[0075] S113. Using the data processing layer in the pre-defined neural network model, convolve the ultrasound standard image of the current frame to extract the semantic features corresponding to the ultrasound standard image of the current frame.

[0076] Specifically, the pre-defined neural network model is designed and trained for a specific task (such as detecting the location of internal bleeding in ultrasound images). These models typically contain multiple layers to automatically extract useful features from the input data and perform complex pattern recognition.

[0077] S12. Perform dilated convolution processing on the extracted semantic features using different dilation rates, and perform fusion analysis on the processing results to obtain fused features;

[0078] As a preferred embodiment, the extracted semantic features are subjected to dilated convolution processing using different dilation rates, and the processing results are fused and analyzed to obtain fused features, including the following steps:

[0079] S121. Use dilated convolution with different dilation rates to process semantic features and obtain a semantic feature dilation map;

[0080] It should be noted that dilated convolution increases the receptive field by inserting holes in the convolution kernel (i.e., skipping some input features) without increasing the number of parameters or computational cost.

[0081] Using multiple convolutional kernels with different dilation rates can capture contextual information at different scales.

[0082] The outputs of each dilated convolution are stacked or concatenated to obtain a semantic feature dilation map containing multi-scale information.

[0083] S122. Perform dimensionality reduction and activation function mapping on the semantic feature expansion map to obtain the attention feature map;

[0084] Semantic feature inflation maps can contain a large number of feature channels and high dimensionality. Directly processing this high-dimensional data increases computational burden and may lead to overfitting. Dimensionality reduction operations, such as using 1x1 convolutional layers or pooling layers, can reduce the number of channels in the feature map, thereby reducing its dimensionality.

[0085] Activation functions are a crucial component of deep learning, introducing non-linearity and enabling models to learn more complex mappings. Applying activation functions (such as ReLU, Sigmoid, or Tanh) to the dimensionality-reduced feature maps can enhance the model's non-linear expressive power.

[0086] S123. Concatenate all attention feature maps to obtain a secondary attention feature map;

[0087] Specifically, the processing involves merging all attention feature maps along the channel dimension.

[0088] Suppose there are N attention feature maps, each with dimensions [H, W, Ci], where H and W are the height and width of the feature map, and Ci is the number of channels of the i-th feature map.

[0089] The size of the concatenated second-level attention feature map will be [H, W, C], where C is the sum of the number of channels in all attention feature maps, i.e., C = C1 + C2 + ... + CN.

[0090] The feature map obtained by splicing is called the secondary attention feature map.

[0091] S124. Multiply the secondary attention feature map and the semantic features of the ultrasound standard image to obtain the fused features.

[0092] It should be noted that the operation involves element-wise multiplication of the secondary attention feature map and the semantic features of the ultrasound standard image. This operation is equivalent to weighting the semantic features of the ultrasound standard image, with the weights provided by the secondary attention feature map. This multiplication operation ensures that the fused feature includes information from both the original semantic features and the weighted effect of the attention mechanism, thus enhancing the feature's representational power. The feature map obtained through this multiplication operation is called the fused feature. It integrates information from the secondary attention feature map and the semantic features of the ultrasound standard image, resulting in stronger representational power and better task adaptability.

[0093] S13. Upsample the fusion features to obtain a cleared image of the current frame ultrasound image;

[0094] As a preferred embodiment, upsampling the fused features to obtain a cleared image of the current frame ultrasound image includes the following steps:

[0095] S131. Upsample the fused features using the nearest neighbor interpolation method to obtain a primary upsampled image;

[0096] S132. Perform a convolution operation on the primary upsampled image to obtain a convolution-enhanced image;

[0097] S133. Perform upsampling processing on the convolutional enhanced image to obtain a secondary upsampled image;

[0098] S134. Perform a convolution operation on the secondary upsampled image to obtain a cleared image of the current frame ultrasound image.

[0099] It should be noted that the nearest neighbor interpolation method is used to upsample the fused features to obtain a primary upsampled image. The core idea is to assign the nearest pixel value to the new pixel, thereby increasing the image size. The specific steps are as follows:

[0100] Determine the upsampling ratio: First, you need to determine the upsampling ratio, that is, how many times you want to enlarge the image size. This ratio will determine which region in the original image each pixel in the new image corresponds to.

[0101] Calculate the position of the new pixel: Based on the upsampling ratio, calculate the corresponding position of each pixel in the new image within the original image. This position may be a non-integer coordinate because the size of the upsampled image is usually not an integer multiple of the original image size.

[0102] Assigning pixel values: For each pixel in the new image, find its nearest neighbor pixel in the original image and assign that pixel's value to the new pixel. This is the meaning of "nearest neighbor".

[0103] Generate the initial upsampled image: Repeat the above steps until all pixels of the new image have been assigned values. This gives you the initial upsampled image.

[0104] S14. Use annotation tools to annotate the bleeding points in the cleared ultrasound images to obtain an image dataset.

[0105] Specifically, a dataset was created and labeled based on the acquired ultrasound images containing hemorrhage points using labelimg. Labelimg is an open-source image annotation tool based on Qt, and the labels can be used for classification and object detection. In the labelimg toolbox, the data type was selected as "PascalVOC," and the annotation method was "Create RectBox." A rectangular box was used to mark the location of the hemorrhage point in each ultrasound image; the rectangle should be as small as possible while ensuring it covers the hemorrhage point. Under the guidance of an ultrasound physician, the images were labeled one by one using the labelimg toolbox in the above manner. The toolbox automatically generates a .txt annotation file with the same name as the image after each annotation. The annotation file is in VOC format; only datasets annotated in VOC format can be read correctly by the network. Based on this, a dataset was constructed, where each data point consists of an ultrasound image and its corresponding annotation. After obtaining the total dataset, it was divided into a training set, a validation set, and a test set, with the label format being the VOC format commonly used in object detection.

[0106] S2. Based on the image dataset, construct and train an improved neighborhood attention converter model;

[0107] As a preferred implementation, constructing and training an improved neighborhood attention converter model based on an image dataset includes the following steps:

[0108] S21. Divide the image dataset into a training set and a test set;

[0109] Specifically, the ratio of training set to test set is set. Common ratios include 70% / 30% and 80% / 20%, with the training set usually having a higher proportion.

[0110] S22. Based on the improved neighborhood attention converter structure, construct an improved neighborhood attention converter model;

[0111] As a preferred embodiment, the improved neighborhood attention converter structure includes: an input layer, an overlapping word segmenter, a neighborhood attention block, an overlapping downsampler, a fully connected layer, and an output layer;

[0112] S23. The improved neighborhood attention converter model is trained using the training set, and the model parameters are continuously adjusted using the backpropagation algorithm and gradient descent method to minimize the loss function.

[0113] As a preferred implementation, the improved neighborhood attention converter model is trained using a training set, and the model parameters are continuously adjusted using the backpropagation algorithm and gradient descent method, including the following steps:

[0114] S231. Based on the improved neighborhood attention converter structure, initialize the weights and biases of all structural layers;

[0115] S232. Based on forward propagation, the training set is input into the improved neighborhood attention converter model, and the prediction results of bleeding points are generated through the processing of each layer of the structure.

[0116] S233. Use the loss function to evaluate the difference between the model prediction results and the true labels, and obtain the loss gradient;

[0117] S234. Based on the loss gradient, update the weights and biases of the improved neighborhood attention converter model using gradient descent based on backpropagation.

[0118] S24. After training, the performance of the improved neighborhood attention converter model is evaluated using the test set. Based on the evaluation results, the model structure is adjusted to obtain the final improved neighborhood attention converter model.

[0119] It should be understood that, such as Figure 6 An improved neighborhood attention converter is constructed as shown. It is trained on a pre-built dataset and continuously adjusts the network weights and updates the structural parameters using the gradient backpropagation of the loss function with respect to different network parameters. (Each training epoch follows the chain rule, starting from the loss function and iterating backwards through the gradients of the network weights with respect to the loss function, updating the corresponding parameters to minimize the loss function). The optimal global parameter matrix is ​​obtained and saved. The improved neighborhood attention converter structure is shown below. Figure 6As shown, considering the characteristics of ultrasound image bleeding point detection—which involves small target detection and a small dataset—the improved network abandons the two neighborhood attention blocks that generate H / 16×W / 16 and H / 32×W / 32 feature maps, and only uses H / 4×W / 4 and H / 8×W / 8 neighborhood attention blocks. This approach limits the number of parameters and prevents overfitting. More importantly, larger feature maps lose relatively less semantic information about small targets; excessively high downsampling rates would lead to a significant loss of semantic information about small targets, ultimately making detection difficult. Another important improvement is the use of NWD (Normalized Wasserstein Distance Loss) for network training (i.e., using the normalized Wasserstein distance loss function as the loss function to update network weights during training). NWD is a loss function used in object detection for small object detection tasks. Compared to general IoU (Intersection over Union) based loss functions, it offers advantages such as scale invariance, smoothness against positional biases, and superior ability to measure the similarity between non-overlapping or mutually containing bounding boxes. In each training round, the network calculates the specific value of the loss function for the validation set images based on the current weights. The network records the loss function values ​​on the validation set throughout the training process, saving the network weights corresponding to the round with the minimum loss function value as the best-performing weights on the validation set.

[0120] By loading the saved weights into the network and inputting the test data into the network, the corresponding test set images are processed by the network under these weights to obtain the detected image results (the test result image includes a rectangle that outlines the location of the bleed point and the confidence probability that the location is a bleed point) and index results (including precision, detection rate, etc.). The bleed point detection effect is quantitatively evaluated based on the precision and detection rate. The larger the index, the better the detection effect.

[0121] S3. Acquire the ultrasound image to be detected, and use the trained and improved neighborhood attention converter model to detect the bleeding location and obtain the detection result;

[0122] It should be noted that the ultrasound images to be detected need to be preprocessed before being input into the model, including adjusting the image size and normalizing the pixel values, to ensure that they match the data format used during model training.

[0123] The preprocessed ultrasound image is input into the model, and the model is run to detect the location of the bleeding. The model will output a prediction of the location of the bleeding in the image.

[0124] S4. Use visualization tools to visualize the test results.

[0125] Specifically, commonly used visualization tools include Matplotlib, OpenCV, and PIL (Python Imaging Library). These tools allow you to write code to mark bleeding locations on the original image. Different colors, shapes, and labels can be used to highlight the detection results. Visualization parameters can be adjusted as needed to improve the visualization effect. For example, the size, color, and transparency of the markers can be adjusted to ensure they are clearly visible on the image.

[0126] like Figure 2 As shown, according to an embodiment of the present invention, a system for detecting the location of internal bleeding in the human body based on ultrasound images is provided. The system for detecting the location of internal bleeding in the human body based on ultrasound images includes: a dataset construction module 1, a model construction module 2, a bleeding location detection module 3, and a visualization display module 4.

[0127] Dataset construction module 1 is used to construct an image dataset using acquired ultrasound images containing bleeding points;

[0128] Model building module 2 is used to build and train an improved neighborhood attention converter model based on an image dataset;

[0129] Bleeding location detection module 3 is used to acquire the ultrasound image to be detected, and to detect the bleeding location using the trained improved neighborhood attention converter model to obtain the detection result;

[0130] The visualization module 4 is used to visualize the detection results using visualization tools.

[0131] In summary, by utilizing the above-mentioned technical solution of this invention, deep learning networks are applied to the detection of bleeding points in ultrasound images. With the support of a dataset, deep learning networks can achieve higher detection accuracy and better noise robustness compared to traditional algorithms. Traditional detection methods often only extract a few features and are subject to strong human interference. In contrast, this invention uses deep learning networks, eliminating the need for manual feature selection. The network can automatically extract features at multiple different levels. Specifically optimized at the network design level for the small target detection task of bleeding points, it achieves a better detection rate. This invention effectively improves the quality of ultrasound images through feature convolution processing, dilated convolution processing, and upsampling processing, making them clearer and easier to identify bleeding points. Dilated convolution processing of semantic features using different dilation rates, followed by fusion analysis of the processing results, yields richer feature information, which helps in the accurate detection of subsequent bleeding points. Bleeding point annotation processing is performed on the clarified ultrasound images; due to the improved image quality, the accuracy of the annotation is also correspondingly improved.

[0132] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0133] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An ultrasonographic image-based method of detecting a bleeding site in a human body, characterized by, Includes the following steps: S1. Construct an image dataset using the acquired ultrasound images containing bleeding points; S2. Divide the image dataset into a training set and a test set; Based on the improved neighborhood attention converter structure, an improved neighborhood attention converter model is constructed. The improved neighborhood attention converter model is trained using the training set, and the model parameters are continuously adjusted using the backpropagation algorithm and gradient descent method to minimize the loss function. After training, the performance of the improved neighborhood attention converter model is evaluated using a test set. Based on the evaluation results, the model structure is adjusted to obtain the final improved neighborhood attention converter model. S3. Acquire the ultrasound image to be detected, and use the trained and improved neighborhood attention converter model to detect the bleeding location and obtain the detection result; S4. Use visualization tools to display the test results visually; S1 includes the following steps: S11. Acquire multiple consecutive ultrasound images containing bleeding points and extract semantic features through feature convolution processing; S12. Perform dilated convolution processing on the extracted semantic features using different dilation rates, and perform fusion analysis on the processing results to obtain fused features; S13. Upsample the fusion features to obtain a cleared image of the current frame ultrasound image; S14. Use annotation tools to annotate the bleeding points in the cleared ultrasound images to obtain an image dataset; S12 includes the following steps: S121. Use dilated convolution with different dilation rates to process semantic features and obtain a semantic feature dilation map; S122. Perform dimensionality reduction and activation function mapping on the semantic feature expansion map to obtain the attention feature map; S123. Concatenate all attention feature maps to obtain a secondary attention feature map; S124. Multiply the secondary attention feature map and the semantic features of the ultrasound standard image to obtain the fused features; S13 includes the following steps: S131. Upsample the fused features using the nearest neighbor interpolation method to obtain a primary upsampled image; S132. Perform a convolution operation on the primary upsampled image to obtain a convolution-enhanced image; S133. Perform upsampling processing on the convolutional enhanced image to obtain a secondary upsampled image; S134. Perform a convolution operation on the secondary upsampled image to obtain a cleared image of the current frame ultrasound image; The improved neighborhood attention converter structure includes: an input layer, an overlapping word segmenter, a neighborhood attention block, an overlapping downsampler, a fully connected layer, and an output layer; The size feature maps output by the neighborhood attention block are H / 4×W / 4 and H / 8×W / 8.

2. The method of claim 1, wherein The acquisition of multiple consecutive ultrasound images containing bleeding points, and the extraction of semantic features through feature convolution processing, includes the following steps: S111. Perform image preprocessing on the acquired continuous multi-frame ultrasound images containing bleeding points. The preprocessing includes noise removal, contrast adjustment, and image normalization. S112. Convert the format of the preprocessed ultrasound image to obtain a standard ultrasound image; S113. Using the data processing layer in the pre-defined neural network model, convolve the ultrasound standard image of the current frame to extract the semantic features corresponding to the ultrasound standard image of the current frame.

3. The method of claim 2, wherein The process of training the improved neighborhood attention converter model using a training set and continuously adjusting the model parameters through backpropagation and gradient descent includes the following steps: S231. Based on the improved neighborhood attention converter structure, initialize the weights and biases of all structural layers; S232. Based on forward propagation, the training set is input into the improved neighborhood attention converter model, and the prediction results of bleeding points are generated through the processing of each layer of the structure. S233. Use the loss function to evaluate the difference between the model prediction results and the true labels, and obtain the loss gradient; S234. Based on the loss gradient, update the weights and biases of the improved neighborhood attention converter model using gradient descent based on backpropagation.

4. The method of claim 2, wherein, The loss function is the normalized Vickers distance loss function.

5. A system for detecting the location of internal bleeding in the human body based on ultrasound images, used to implement the method for detecting the location of internal bleeding in the human body based on ultrasound images according to any one of claims 1-4, characterized in that, The ultrasound image-based system for detecting the location of internal bleeding in the human body includes: a dataset construction module, a model construction module, a bleeding location detection module, and a visualization module. The dataset construction module is used to construct an image dataset using the acquired ultrasound images containing bleeding points. The model building module is used to build and train an improved neighborhood attention converter model based on an image dataset. The bleeding location detection module is used to acquire the ultrasound image to be detected, and to detect the bleeding location using the trained improved neighborhood attention converter model to obtain the detection result. The visualization module is used to visualize the detection results using visualization tools.

Images