Ultrasound contrast T-tube sinus tract identification method based on deep learning

Abstract

The invention relates to the technical field of ultrasound contrast, in particular to an ultrasound contrast T-tube sinus tract recognition method based on deep learning. The method comprises the following steps: firstly, collecting contrast agent parameters, ultrasonic equipment parameters and clinical information, adjusting a prediction model by utilizing the contrast agent parameters, and optimizing the contrast agent parameters in combination with the clinical information; and then, acquiring an ultrasonic contrast image of the T-tube sinus tract according to the updated contrast agent parameters and the ultrasonic equipment parameters. Thirdly, preprocessing the T-tube sinus tract ultrasonic contrast image, inputting the preprocessed T-tube sinus tract ultrasonic contrast image into the multi-task ultrasonic contrast image enhancement model, and performing image enhancement and adaptive adjustment in combination with ultrasonic equipment parameters; and finally, constructing an ultrasound contrast image multi-stage classification identification model, identifying the T-tube sinus tract ultrasound contrast enhanced image to obtain a T-tube sinus tract identification result, and combining clinical information to assist medical personnel in diagnosis. According to the invention, the T-tube sinus tract identification precision and accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of contrast-enhanced ultrasound, and particularly to a method for identifying T-tube sinus tracts based on deep learning using contrast-enhanced ultrasound. Background Art

[0002] Traditional methods for identifying T-tube sinus tracts use contrast-enhanced ultrasound technology and visual observation. Among them, contrast-enhanced ultrasound technology improves the imaging ability of ultrasound for sinus tract tissues by injecting contrast agents, and then the situation of the sinus tract is observed visually. However, the traditional identification process mainly relies on the naked eye, which is easily affected by the level and experience of image analysts. At the same time, the influence of external conditions such as the dosage of contrast agents, injection time, and injection speed on the identification of T-tube sinus tracts is not considered, thus reducing the accuracy and precision of T-tube sinus tract identification.

[0003] Therefore, a method for identifying T-tube sinus tracts based on deep learning using contrast-enhanced ultrasound is proposed. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying T-tube sinus tracts based on deep learning using contrast-enhanced ultrasound. First, collect contrast agent parameters, ultrasound equipment parameters, and clinical information, adjust the prediction model using the contrast agent parameters, and optimize the contrast agent parameters in combination with clinical information. Then, obtain contrast-enhanced ultrasound images of the T-tube sinus tract according to the updated contrast agent parameters and ultrasound equipment parameters. Next, preprocess the contrast-enhanced ultrasound images of the T-tube sinus tract and input them into a multi-task contrast-enhanced ultrasound image enhancement model to perform image enhancement and adaptive adjustment in combination with the ultrasound equipment parameters. Finally, construct a multi-stage classification and recognition model for contrast-enhanced ultrasound images to recognize the contrast-enhanced ultrasound images of the T-tube sinus tract, obtain the identification results of the T-tube sinus tract, and use them in combination with clinical information to assist medical staff in diagnosis. The present invention can effectively improve the accuracy and precision of T-tube sinus tract identification.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for identifying T-tube sinus tracts based on deep learning using contrast-enhanced ultrasound, comprising:

[0007] Obtain contrast agent parameters, ultrasound equipment parameters, and clinical information;

[0008] Input the contrast agent parameters and the clinical information into a contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters; obtain contrast-enhanced ultrasound images of the T-tube sinus tract according to the updated contrast agent parameters and the ultrasound equipment parameters;

[0009] Preprocess the contrast-enhanced ultrasound images of the T-tube sinus tract to obtain preprocessed contrast-enhanced ultrasound images of the T-tube sinus tract;

[0010] Input the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract and the ultrasound device parameters into a multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain a contrast-enhanced ultrasound image of the T-tube sinus tract;

[0011] Construct a multi-stage classification and recognition model for contrast-enhanced ultrasound images, and input the contrast-enhanced ultrasound image of the T-tube sinus tract into the multi-stage classification and recognition model for contrast-enhanced ultrasound images for recognition to obtain a recognition result of the T-tube sinus tract;

[0012] Use the recognition result of the T-tube sinus tract and the clinical information to assist medical staff in making a diagnosis.

[0013] Further, the contrast agent parameters include: contrast agent dosage, injection time window, and injection speed; the ultrasound device parameters include: frequency, gain, dynamic range, focus, and frame rate; the clinical information includes: imaging-related clinical information and imaging-unrelated clinical information.

[0014] Further, input the contrast agent parameters and the clinical information into a contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters; the process of obtaining a contrast-enhanced ultrasound image of the T-tube sinus tract according to the updated contrast agent parameters and the ultrasound device parameters includes:

[0015] Use historical contrast-enhanced ultrasound data to train the contrast agent parameter adjustment prediction model to obtain historical contrast agent parameter adjustment weights;

[0016] Input the contrast agent parameters, the clinical information, and the historical contrast agent parameter adjustment weights into the trained contrast agent parameter adjustment prediction model for parameter update to obtain current contrast agent parameter adjustment weights;

[0017] Use the current contrast agent parameter adjustment weights to update the contrast agent parameters to obtain the updated contrast agent parameters, and collect the contrast-enhanced ultrasound image of the T-tube sinus tract according to the updated contrast agent parameters and the ultrasound device parameters.

[0018] Further, the process of preprocessing the contrast-enhanced ultrasound image of the T-tube sinus tract to obtain a preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract includes:

[0019] Denoise the contrast-enhanced ultrasound image of the T-tube sinus tract to obtain a denoised contrast-enhanced ultrasound image of the T-tube sinus tract;

[0020] Enhance the contrast of the denoised contrast-enhanced ultrasound image of the T-tube sinus tract to obtain a pre-enhanced contrast-enhanced ultrasound image of the T-tube sinus tract;

[0021] Extract the edges of the pre-enhanced contrast-enhanced ultrasound image of the T-tube sinus tract to obtain an edge image of the contrast-enhanced ultrasound image of the T-tube sinus tract;

[0022] Normalize the contrast-enhanced ultrasound image of the T-tube sinus tract and combine it with the edge image of the contrast-enhanced ultrasound of the T-tube sinus tract to obtain the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract.

[0023] Further, the multi-task contrast-enhanced ultrasound image enhancement model includes: an input layer, an adaptive adjustment layer, a motion compensation layer, a detail enhancement layer, and an output layer;

[0024] Among them, the input layer is used to perform feature transformation on the contrast-enhanced ultrasound image and device parameters;

[0025] The adaptive adjustment layer is used to generate adaptive parameters according to the ultrasound device parameters;

[0026] The motion compensation layer is used to learn motion information from consecutive contrast-enhanced ultrasound images of the T-tube sinus tract and perform motion compensation in combination with the adaptive parameters;

[0027] The detail enhancement layer is used to learn high-frequency details in the contrast-enhanced ultrasound image of the T-tube sinus tract in combination with image edge features and adjust the enhancement intensity through the adaptive parameters;

[0028] The output layer is used to output the contrast-enhanced ultrasound image of the T-tube sinus tract.

[0029] Further, the specific implementation process of inputting the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract and the ultrasound device parameters into the multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain the contrast-enhanced ultrasound image of the T-tube sinus tract includes:

[0030] Input the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract and the ultrasound device parameters into the input layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain the contrast-enhanced ultrasound image features of the T-tube sinus tract and the ultrasound device parameter features;

[0031] Use the adaptive adjustment layer of the multi-task contrast-enhanced ultrasound image enhancement model to process the ultrasound device parameter features to obtain adaptive parameters;

[0032] Use the optical flow features predicted by the motion compensation layer of the multi-task contrast-enhanced ultrasound image enhancement model and combine the adaptive parameters to process the contrast-enhanced ultrasound image features of the T-tube sinus tract to obtain the motion compensation features of the contrast-enhanced ultrasound of the T-tube sinus tract;

[0033] Input the motion compensation features of the contrast-enhanced ultrasound of the T-tube sinus tract and the adaptive parameters into the detail enhancement layer of the multi-task contrast-enhanced ultrasound image enhancement model for detail enhancement to obtain the detail enhancement features of the contrast-enhanced ultrasound of the T-tube sinus tract;

[0034] Using the output layer of the multi-task contrast-enhanced ultrasound image enhancement model to process the enhanced features of the T-tube sinus tract contrast-enhanced ultrasound details, the enhanced image of the T-tube sinus tract contrast-enhanced ultrasound is obtained.

[0035] Further, the process of inputting the enhanced image of the T-tube sinus tract contrast-enhanced ultrasound into the multi-stage classification and recognition model of the contrast-enhanced ultrasound image to obtain the recognition result of the T-tube sinus tract includes:

[0036] Using the multi-stage classification and recognition model of the contrast-enhanced ultrasound image to preliminarily classify the enhanced image of the T-tube sinus tract contrast-enhanced ultrasound to obtain T-tube features and non-T-tube features;

[0037] Inputting the non-T-tube features into the sinus tract morphology classification layer of the multi-stage classification and recognition model of the contrast-enhanced ultrasound image to obtain the sinus tract morphology classification result; wherein, the sinus tract morphology classification result includes the first morphology classification and the second morphology classification;

[0038] Selecting the corresponding sinus tract recognition layer according to the sinus tract morphology classification result, and inputting the T-tube features and the non-T-tube features into the sinus tract recognition layer for recognition to obtain the recognition result of the T-tube sinus tract.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. The present invention proposes a method for adjusting contrast agent parameters to adaptively adjust the parameters affecting ultrasound contrast imaging; this method trains the contrast agent parameter adjustment prediction model by using historical ultrasound contrast data to obtain the historical contrast agent parameter adjustment weights, and inputs the contrast agent parameters and clinical information into the trained contrast agent parameter adjustment prediction model, and updates the contrast agent parameters according to the obtained current contrast agent parameter adjustment weights; this method effectively adjusts the contrast agent parameters by combining patient clinical data and deep learning models to improve the imaging quality, thereby improving the recognition accuracy and accuracy of the T-tube sinus tract.

[0041] 2. The present invention proposes a method for enhancing contrast-enhanced ultrasound images to improve the quality of the acquired images of the T-tube sinus tract contrast-enhanced ultrasound; this method uses the multi-task contrast-enhanced ultrasound image enhancement model to perform motion compensation and detail enhancement on the preprocessed T-tube sinus tract contrast-enhanced ultrasound image, and at the same time uses the adaptive parameters obtained by the adaptive adjustment layer of the multi-task contrast-enhanced ultrasound image enhancement model to adaptively adjust the enhancement amplitude of the model; this method combines optical flow compensation and adaptive detail enhancement to effectively reduce the artifact phenomenon and detail blurring problem caused by motion acquisition, thereby improving the recognition accuracy and accuracy of the T-tube sinus tract.

[0042] 3. The present invention proposes a multi-stage classification and recognition method for contrast-enhanced ultrasound images to improve the accuracy of identifying T-tube sinus tracts. The method first uses a multi-stage classification and recognition model for contrast-enhanced ultrasound images to preliminarily classify the input image, obtaining T-tube features and non-T-tube features. Among them, the T-tube features contain the spatial position information of the T-tube, which can be used to assist in identifying the sinus tract. Then, the model is used to classify the morphology of the non-T-tube features to obtain the corresponding sinus tract morphology classification results and select the corresponding sinus tract recognition layer. The T-tube features and non-T-tube features are input into the sinus tract recognition layer for recognition to obtain the T-tube sinus tract recognition result. This method combines multi-layer classification and recognition with T-tube feature-assisted recognition to effectively improve the recognition accuracy and precision of T-tube sinus tracts. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 FIG. is a schematic flow chart of a method for identifying a T-tube sinus tract in contrast-enhanced ultrasound based on deep learning according to the present invention;

[0044] Figure 2 FIG. is a schematic structural diagram of a multi-task contrast-enhanced ultrasound image enhancement model according to the present invention;

[0045] Figure 3 FIG. is a schematic flow chart for obtaining the T-tube sinus tract recognition result according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0047] Please refer to Figures 1 to 3 , the present invention provides a method for identifying a T-tube sinus tract in contrast-enhanced ultrasound based on deep learning, and the technical solution is as follows:

[0048] Embodiment 1

[0049] In order to improve the recognition accuracy and precision of the T-tube sinus tract in contrast-enhanced ultrasound, a certain hospital uses a method for identifying a T-tube sinus tract in contrast-enhanced ultrasound based on deep learning proposed by the present invention. The flow chart of the method is as Figure 1 shown, specifically as follows:

[0050] Obtain contrast agent parameters, ultrasound equipment parameters, and clinical information;

[0051] Furthermore, the contrast agent parameters, ultrasound equipment parameters, and clinical information can be obtained by extracting from the system log;

[0052] Further, the contrast agent parameters include: the dosage of the contrast agent, the injection time window, and the injection speed; the ultrasound device parameters include: frequency, gain, dynamic range, focus, and frame rate; the clinical information includes: imaging-related clinical information and imaging-unrelated clinical information; the injection time window is the duration calculated starting from the start of the scan;

[0053] By introducing the contrast agent parameters, ultrasound device parameters, and clinical information, reliable data support can be provided for subsequent enhancement and recognition of T-tube sinus tract ultrasound contrast images, thereby improving the recognition accuracy and precision of the T-tube sinus tract.

[0054] Further, the imaging-related clinical information includes: history information related to the gallbladder, T-tube blockage and slippage, physical examination information, bilirubin test information, etc.; the imaging-unrelated clinical information includes: other medical histories and living habits;

[0055] Further, the contrast examination process of the T-tube sinus tract is mainly divided into the following three parts: the T-tube - bile duct continuous segment, the abdominal segment, and the peritoneum - T-tube continuous segment; since in actual situations, the patient will be injected multiple times to better display the T-tube sinus tract at different observation stages, for the convenience of tabular display, one injection record of the same examination part at the same observation stage is selected from three groups of contrast agent parameter samples from different patients as a reference, and are respectively denoted as Sample One, Sample Two, and Sample Three; the contrast agent parameter information is shown in Table 1.

[0056] Table 1 Contrast Agent Parameter Information

[0057] Reference data Contrast agent dosage Injection time window Injection speed Sample 1 2.4 mL 15 s - 45 s 0.08 mL / s Sample 2 1.6 mL 15 s - 75 s 0.02 mL / s Sample 3 2.0 mL 15 s - 55 s 0.05 mL / s

[0058] Input the contrast agent parameters and clinical information into the contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters; according to the updated contrast agent parameters and ultrasound device parameters, obtain the T-tube sinus tract ultrasound contrast image;

[0059] Further, the process of inputting the contrast agent parameters and clinical information into the contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters, and obtaining the T-tube sinus tract ultrasound contrast image according to the updated contrast agent parameters and ultrasound device parameters includes:

[0060] Use historical ultrasound contrast data to train the contrast agent parameter adjustment prediction model to obtain historical contrast agent parameter adjustment weights;

[0061] Input the contrast agent parameters, clinical information, and historical contrast agent parameter adjustment weights into the trained contrast agent parameter adjustment prediction model for parameter update to obtain the current contrast agent parameter adjustment weights;

[0062] Update the contrast agent parameters by using the current contrast agent parameter adjustment weights to obtain updated contrast agent parameters, and acquire T-tube sinus tract contrast-enhanced ultrasound images based on the updated contrast agent parameters and ultrasound device parameters;

[0063] Furthermore, the contrast agent parameter adjustment prediction model includes: a clinical information encoder, a contrast agent parameter encoder, a feature vector fusion device, and a parameter predictor;

[0064] Furthermore, the clinical information encoder and the contrast agent parameter encoder are respectively used to extract the clinical information feature vector and the contrast agent parameter feature vector. The encoder uses a CNN network; the feature vector fusion device uses a Transformer network to fuse the clinical information feature vector and the contrast agent parameter feature vector; the parameter predictor uses a multi-layer perceptron to realize the prediction of the contrast agent parameters;

[0065] Furthermore, the current contrast agent parameter adjustment weights include: the current contrast agent dosage adjustment weight, the current injection time window adjustment weight, and the current injection speed adjustment weight;

[0066] Furthermore, the T-tube sinus tract contrast-enhanced ultrasound images are acquired by reading from an imaging device.

[0067] To illustrate the contrast agent parameter update scheme proposed by the present invention, one injection record is randomly selected from each of the two groups of contrast agent parameters for update testing, and they are respectively denoted as Data One (before update), Data One (after update), Data Two (before update), and Data Two (after update); the two groups of contrast agent parameters and the corresponding patient clinical information are input into the contrast agent parameter adjustment prediction model, and the contrast agent parameters are updated by using the obtained current contrast agent parameter adjustment weights. The contrast agent parameter update test results are shown in Table 2.

[0068] Table 2 Contrast Agent Parameter Update Test Results

[0069] Test data Contrast agent dosage Injection time window Injection speed Data 1 (before update) 2.1 mL 15 s - 45 s 0.07 mL / s Data 1 (after update) 2.4 mL 15 s - 50 s 0.08 mL / s Data 2 (before update) 2.7 mL 15 s - 45 s 0.09 mL / s Data 2 (after update) 2.0 mL 15 s - 40 s 0.08 mL / s

[0070] In this embodiment, the contrast agent parameters are adaptively adjusted by using the contrast agent parameter adjustment prediction model and combining with clinical information to improve the imaging quality of the T-tube sinus tract contrast-enhanced ultrasound images, thereby further improving the recognition accuracy and precision of the T-tube sinus tract.

[0071] Preprocess the T-tube sinus tract contrast-enhanced ultrasound images to obtain preprocessed T-tube sinus tract contrast-enhanced ultrasound images;

[0072] Furthermore, the process of preprocessing the T-tube sinus tract contrast-enhanced ultrasound images to obtain preprocessed T-tube sinus tract contrast-enhanced ultrasound images includes:

[0073] Denoise the T-tube sinus tract contrast-enhanced ultrasound images to obtain T-tube sinus tract contrast-enhanced ultrasound denoised images;

[0074] Perform contrast enhancement on the denoised image of the T-tube sinus tract contrast-enhanced ultrasound to obtain a pre-enhanced image of the T-tube sinus tract contrast-enhanced ultrasound;

[0075] Perform edge extraction on the pre-enhanced image of the T-tube sinus tract contrast-enhanced ultrasound to obtain an edge image of the T-tube sinus tract contrast-enhanced ultrasound;

[0076] Perform normalization on the enhanced image of the T-tube sinus tract contrast-enhanced ultrasound and combine it with the edge image of the T-tube sinus tract contrast-enhanced ultrasound to obtain a preprocessed image of the T-tube sinus tract contrast-enhanced ultrasound.

[0077] Furthermore, median filtering, non-local means filtering, Gaussian filtering, wavelet threshold filtering, etc. can be used in the denoising process; gamma correction, histogram equalization, homomorphic filtering, etc. can be used in the contrast enhancement process; Sobel operator, Laplace operator, Canny operator, etc. can be used in the edge extraction process.

[0078] By performing denoising and contrast enhancement on the T-tube sinus tract contrast-enhanced ultrasound image, the interference of noise on model recognition can be reduced and the accuracy of model feature recognition can be improved. Edge extraction can be used to assist the model in restoring high-frequency edge details, and normalization can reduce the model calculation time; this preprocessing process can improve the feature recognition ability of the subsequent enhancement model, thereby improving the recognition accuracy and accuracy of the T-tube sinus tract.

[0079] Input the preprocessed T-tube sinus tract contrast-enhanced ultrasound image and ultrasound device parameters into a multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain an enhanced image of the T-tube sinus tract contrast-enhanced ultrasound;

[0080] Furthermore, the structure of the multi-task contrast-enhanced ultrasound image enhancement model can refer to Figure 2 , including: an input layer, an adaptive adjustment layer, a motion compensation layer, a detail enhancement layer, and an output layer;

[0081] Among them, the input layer is used to perform feature transformation on the contrast-enhanced ultrasound image and device parameters using a convolutional layer to obtain the T-tube sinus tract contrast-enhanced ultrasound image features and ultrasound device parameter features;

[0082] Furthermore, the T-tube sinus tract contrast-enhanced ultrasound image features include non-edge image features of the T-tube sinus tract contrast-enhanced ultrasound and edge image features of the T-tube sinus tract contrast-enhanced ultrasound;

[0083] The adaptive adjustment layer is used to generate adaptive parameters according to the ultrasound device parameter features;

[0084] Furthermore, the adaptive adjustment layer is an MLP network;

[0085] The motion compensation layer is used to learn motion information from consecutive T-tube sinus tract contrast-enhanced ultrasound images and perform motion compensation in combination with adaptive parameters;

[0086] Further, the motion compensation layer includes an optical flow estimation network G fl and a motion compensation network G mc ;

[0087] Further, the optical flow estimation network takes two sets of features of T-tube sinus tract contrast-enhanced ultrasound images as inputs to obtain two corresponding predicted optical flow features; the motion compensation network takes the two sets of non-edge image features of T-tube sinus tract contrast-enhanced ultrasound after optical flow compensation and the feature vector of the adaptive parameter as inputs, outputs a single non-edge image feature of T-tube sinus tract contrast-enhanced ultrasound with motion compensation, and merges it with the single edge image feature of T-tube sinus tract contrast-enhanced ultrasound with motion compensation obtained after channel fusion to obtain the motion compensation feature of T-tube sinus tract contrast-enhanced ultrasound; the optical flow estimation network uses the FlowNet network, and the motion compensation network is a UNet network;

[0088] The detail enhancement layer is used to take the motion compensation feature of T-tube sinus tract contrast-enhanced ultrasound as an input, combine the image edge feature to learn the high-frequency details in the T-tube sinus tract contrast-enhanced ultrasound image and adjust the enhancement intensity through the adaptive parameter;

[0089] Further, the detail enhancement layer includes a multi-scale detail learning network G md and an adaptive adjustment module M ad ;

[0090] Further, the multi-scale detail learning network uses a spatial pyramid network to extract multi-scale features, and there is an edge feature guidance module on each scale branch to fuse the edge feature and the multi-scale feature by using the spatial attention layer and output the detail feature;

[0091] Further, the adaptive adjustment module weights the detail feature and the non-edge image feature input to the detail enhancement layer by using the feature vector of the adaptive parameter to obtain the detail enhancement feature of T-tube sinus tract contrast-enhanced ultrasound.

[0092] The output layer is used to output the enhanced image of T-tube sinus tract contrast-enhanced ultrasound.

[0093] By using the motion compensation layer and the detail enhancement layer in the multi-task contrast-enhanced ultrasound image enhancement model to perform motion compensation and detail enhancement on the preprocessed T-tube sinus tract contrast-enhanced ultrasound image, and at the same time using the adaptive parameter obtained by the adaptive adjustment layer of the multi-task contrast-enhanced ultrasound image enhancement model to adaptively adjust the enhancement amplitude of the model, the generation quality of the T-tube sinus tract contrast-enhanced ultrasound image can be effectively improved.

[0094] Further, the specific implementation process of inputting the preprocessed T-tube sinus tract contrast-enhanced ultrasound image and the ultrasound device parameters into the multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain the enhanced image of T-tube sinus tract contrast-enhanced ultrasound includes:

[0095] Input the preprocessed T-tube sinus tract contrast-enhanced ultrasound images and ultrasound device parameters into the input layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain T-tube sinus tract contrast-enhanced ultrasound image features and ultrasound device parameter features;

[0096] Process the ultrasound device parameter features using the adaptive adjustment layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain adaptive parameters;

[0097] Furthermore, the feature vector of the adaptive parameter can be expressed as:

[0098] f ad = MLP(f e );

[0099] where f ad represents the feature vector of the adaptive parameter; MLP represents a multi-layer perceptron network; f e represents the feature vector of the ultrasound device parameters.

[0100] Process the T-tube sinus tract contrast-enhanced ultrasound image features using the optical flow features predicted by the motion compensation layer of the multi-task contrast-enhanced ultrasound image enhancement model and combine with the adaptive parameters to obtain T-tube sinus tract contrast-enhanced ultrasound motion compensation features;

[0101] Furthermore, the T-tube sinus tract contrast-enhanced ultrasound motion compensation features can be expressed as

[0102]

[0103] where F mc is the T-tube sinus tract contrast-enhanced ultrasound motion compensation feature; is the T-tube sinus tract contrast-enhanced ultrasound motion compensation non-edge image feature; is the T-tube sinus tract contrast-enhanced ultrasound motion compensation edge image feature; Concat is a channel fusion operation; O t1→t2 is the predicted optical flow feature from the acquisition time t1 to the acquisition time t2; is an element-wise product operation; is the T-tube sinus tract contrast-enhanced ultrasound edge image feature acquired at time t1; O t2→t1 is the predicted optical flow feature from the acquisition time t2 to the acquisition time t1; is the T-tube sinus tract contrast-enhanced ultrasound edge image feature acquired at time t2; is the T-tube sinus tract contrast-enhanced ultrasound non-edge image feature acquired at time t1; is the T-tube sinus tract contrast-enhanced ultrasound non-edge image feature acquired at time t2;

[0104] Further, the ultrasonic contrast imaging features of the T-tube sinus tract collected at two time points t1 and t2 are continuous in time, including the non-edge ultrasonic contrast imaging features and the edge ultrasonic contrast imaging features of the T-tube sinus tract.

[0105] Input the motion compensation features and adaptive parameters of the ultrasonic contrast imaging of the T-tube sinus tract into the detail enhancement layer of the multi-task ultrasonic contrast imaging enhancement model for detail enhancement, and obtain the detail enhancement features of the ultrasonic contrast imaging of the T-tube sinus tract;

[0106] Further, the detail enhancement features of the ultrasonic contrast imaging of the T-tube sinus tract can be expressed as

[0107]

[0108] where F de is the detail enhancement feature of the ultrasonic contrast imaging of the T-tube sinus tract; F d is the detail feature of the ultrasonic contrast imaging of the T-tube sinus tract.

[0109] Use the output layer of the multi-task ultrasonic contrast imaging enhancement model to process the detail enhancement features of the ultrasonic contrast imaging of the T-tube sinus tract, and obtain the enhanced image of the ultrasonic contrast imaging of the T-tube sinus tract.

[0110] This embodiment uses a multi-task ultrasonic contrast imaging enhancement model and combines optical flow compensation and adaptive detail enhancement, which can effectively reduce the artifact phenomenon and detail blurring problem caused by motion acquisition, thereby improving the recognition accuracy and accuracy of the subsequent model for the T-tube sinus tract.

[0111] Construct a multi-stage classification and recognition model for ultrasonic contrast imaging, input the enhanced image of the ultrasonic contrast imaging of the T-tube sinus tract into the multi-stage classification and recognition model for ultrasonic contrast imaging for recognition, and obtain the recognition result of the T-tube sinus tract;

[0112] Further, the process schematic of inputting the enhanced image of the ultrasonic contrast imaging of the T-tube sinus tract into the multi-stage classification and recognition model for ultrasonic contrast imaging for recognition to obtain the recognition result of the T-tube sinus tract is as Figure 3 shown, and the specific process is as follows:

[0113] Use the T-tube recognition layer of the multi-stage classification and recognition model for ultrasonic contrast imaging to preliminarily classify the enhanced image of the ultrasonic contrast imaging of the T-tube sinus tract, and obtain the T-tube features and non-T-tube features;

[0114] Further, the YOLOv8 model is used in the T-tube recognition layer.

[0115] Input the non-T-tube features into the sinus tract morphology classification layer of the multi-stage classification and recognition model for ultrasonic contrast imaging to obtain the sinus tract morphology classification result; among them, the sinus tract morphology classification result includes the first morphology classification and the second morphology classification

[0116] Further, the sinus tract morphology classification layer uses the AlexNet network model;

[0117] Further, the first morphology classification is the linear morphology of the sinus tract, and the second morphology classification is the non-linear morphology of the sinus tract.

[0118] Select the corresponding sinus tract recognition layer according to the sinus tract morphology classification result, and input the T-tube features and non-T-tube features into the sinus tract recognition layer for recognition to obtain the T-tube sinus tract recognition result;

[0119] Further, the first morphology classification corresponds to the first sinus tract recognition layer, and the second morphology classification corresponds to the second sinus tract recognition layer;

[0120] Further, the first sinus tract recognition layer uses a lightweight CNN model, such as: MobileNet, ShuffleNet, SqueezeNet, etc.; the second sinus tract recognition layer uses the Swim-Transformer model;

[0121] Further, both the first sinus tract recognition layer and the second sinus tract recognition layer have a T-tube feature guidance module for fusing the T-tube spatial position feature and the model feature to improve the recognition accuracy of the sinus tract.

[0122] This embodiment uses a multi-stage classification and recognition model for contrast-enhanced ultrasound images of the T-tube sinus tract to perform multi-layer classification and recognition on the contrast-enhanced ultrasound images of the T-tube sinus tract and uses the spatial position information in the recognized T-tube features to assist in recognizing the sinus tract, which can effectively improve the recognition accuracy and accuracy of the T-tube sinus tract.

[0123] Use the T-tube sinus tract recognition result and clinical information to assist medical staff in diagnosis.

[0124] This embodiment proposes a method for recognizing the T-tube sinus tract in contrast-enhanced ultrasound based on deep learning; first, collect contrast agent parameters, ultrasound device parameters and clinical information, use the contrast agent parameters to adjust the prediction model, and optimize the contrast agent parameters in combination with clinical information. Then, obtain the contrast-enhanced ultrasound image of the T-tube sinus tract according to the updated contrast agent parameters and ultrasound device parameters. Next, preprocess the contrast-enhanced ultrasound image of the T-tube sinus tract and input it into a multi-task contrast-enhanced ultrasound image enhancement model, and perform image enhancement and adaptive adjustment in combination with the ultrasound device parameters. Finally, construct a multi-stage classification and recognition model for contrast-enhanced ultrasound images to recognize the contrast-enhanced ultrasound image of the T-tube sinus tract to obtain the T-tube sinus tract recognition result, and use the clinical information to assist medical staff in diagnosis. This method can effectively improve the recognition accuracy and accuracy of the T-tube sinus tract.

[0125] Embodiment 2

[0126] The present invention proposes a method for identifying T-tube sinus tracts based on deep learning. In order to further verify the effectiveness of the multi-task contrast-enhanced ultrasound image enhancement model and the multi-stage classification and recognition model of contrast-enhanced ultrasound images proposed by the present invention, the present invention conducts an effectiveness test of the enhancement model and an accuracy test of the recognition model for different model selections. The present invention selects two hospitals, A and B, to conduct the above two groups of ablation tests respectively.

[0127] The present invention selects the historical data of Hospital A in the past 5 years as the dataset of the model, where the data in the first to third years are used as the training set of the model, and the data in the fourth to fifth years are used as the validation set; the sampling of the model dataset refers to the following rules: taking a week as a unit, randomly selecting 5 days of data per week; and extracting 3 groups of data in the morning and afternoon each day.

[0128] The present invention inputs the training sets collected from Hospital A into different models for training respectively to obtain their respective pre-trained models; then, the validation sets are input into each pre-trained model to obtain the contrast-enhanced ultrasound images of the T-tube sinus tracts output by the models; then, the average peak signal-to-noise ratio and the average structural similarity of the contrast-enhanced ultrasound images of the T-tube sinus tracts output by all models in the validation set are calculated, and they are used for comparison in the effectiveness test of the enhancement model to obtain the test results of the effectiveness test of the enhancement model; the models corresponding to each test are respectively: the multi-task contrast-enhanced ultrasound image enhancement model proposed by the present invention, denoted as Model 1; the multi-task contrast-enhanced ultrasound image enhancement model without an adaptive adjustment layer, denoted as Model 2; the multi-task contrast-enhanced ultrasound image enhancement model without a motion compensation layer, denoted as Model 3; the multi-task contrast-enhanced ultrasound image enhancement model without a detail enhancement layer, denoted as Model 4.

[0129] The present invention uses a preset threshold to compare the model prediction data with the real data, and obtains the accuracy test results under different models according to the proportion within a reasonable range; the model accuracy test results are shown in Table 3.

[0130] Table 3 Test results of the effectiveness of the enhancement model

[0131] Model Average peak signal-to-noise ratio Average structural similarity Model 1 30.1 dB 0.87 Model 2 28.9 dB 0.83 Model 3 26.4 dB 0.74 Model 4 27.2 dB 0.78

[0132] It can be seen from the results in Table 3 that the model proposed by the present invention has better test results in the effectiveness test of the enhancement model than those of other models; thus, it can be explained that the multi-task contrast-enhanced ultrasound image enhancement model proposed by the present invention can obtain contrast-enhanced ultrasound image data of better quality for the T-tube sinus tracts, further improving the recognition accuracy and precision of the T-tube sinus tracts;

[0133] In order to further test the accuracy of the method process, in this embodiment, historical contrast-enhanced ultrasound image data of T-tube sinus tracts in Hospital B in the past 5 years was collected. According to the same dataset division method and data sampling rules, the training set was then input into different recognition models for training respectively to obtain their respective pre-trained recognition models; then the validation set was input into each pre-trained recognition model to obtain the recognition results of T-tube sinus tracts; the recognition accuracy of each group was obtained through manual verification; the recognition models of each group were respectively denoted as recognition model one, recognition model two, and recognition model three; among them, recognition model one is the multi-stage classification recognition model for contrast-enhanced ultrasound images proposed by the present invention; recognition model two removes the T-tube feature guiding module in the sinus tract recognition layer; recognition model three removes the sinus tract morphology classification layer in the multi-stage classification recognition model for contrast-enhanced ultrasound images and only uses one sinus tract recognition layer for recognition; the test results of the recognition model accuracy are shown in Table 4;

[0134] Table 4 Test Results of Recognition Model Accuracy

[0135] Recognition model Recognition accuracy Recognition model 1 90.35％ Recognition model 2 88.61％ Recognition model 3 85.97％

[0136] As can be seen from the results in Table 4, using recognition model one, that is, the multi-stage classification recognition model for contrast-enhanced ultrasound images proposed by the present invention, the test results of the recognition model accuracy obtained are better than those using other models. This shows that it is necessary to use the sinus tract morphology classification layer for targeted recognition and the T-tube feature guiding module, which can significantly improve the accuracy of T-tube sinus tract recognition.

[0137] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An ultrasound contrast T-tube sinus identification method based on deep learning, characterized in that, Including: Obtaining contrast agent parameters, ultrasound device parameters, and clinical information; Inputting the contrast agent parameters and the clinical information into a contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters; obtaining a T-tube sinus tract contrast-enhanced ultrasound image according to the updated contrast agent parameters and the ultrasound device parameters; Preprocessing the T-tube sinus tract contrast-enhanced ultrasound image to obtain a preprocessed T-tube sinus tract contrast-enhanced ultrasound image; Inputting the preprocessed T-tube sinus tract contrast-enhanced ultrasound image and the ultrasound device parameters into a multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain a T-tube sinus tract contrast-enhanced ultrasound enhanced image; Constructing a multi-stage classification and recognition model for contrast-enhanced ultrasound images, inputting the T-tube sinus tract contrast-enhanced ultrasound enhanced image into the multi-stage classification and recognition model for contrast-enhanced ultrasound images for recognition to obtain a T-tube sinus tract recognition result; Using the T-tube sinus tract recognition result and the clinical information to assist medical staff in making a diagnosis.

2. The method for identifying the T-tube sinus tract by contrast-enhanced ultrasound based on deep learning according to claim 1, wherein The contrast agent parameters include: contrast agent dosage, injection time window, and injection speed; the ultrasound device parameters include: frequency, gain, dynamic range, focus, and frame rate; the clinical information includes: imaging-related clinical information and imaging-unrelated clinical information.

3. The method for identifying the T-tube sinus tract by contrast-enhanced ultrasound based on deep learning according to claim 1, wherein The process of inputting the contrast agent parameters and the clinical information into a contrast agent parameter adjustment prediction model to obtain updated contrast agent parameters; obtaining a T-tube sinus tract contrast-enhanced ultrasound image according to the updated contrast agent parameters and the ultrasound device parameters includes: Training the contrast agent parameter adjustment prediction model using historical contrast-enhanced ultrasound data to obtain historical contrast agent parameter adjustment weights; Inputting the contrast agent parameters, the clinical information, and the historical contrast agent parameter adjustment weights into the trained contrast agent parameter adjustment prediction model for parameter update to obtain current contrast agent parameter adjustment weights; Updating the contrast agent parameters using the current contrast agent parameter adjustment weights to obtain the updated contrast agent parameters, and acquiring the T-tube sinus tract contrast-enhanced ultrasound image according to the updated contrast agent parameters and the ultrasound device parameters.

4. A method for identifying T-tube sinus based on contrast-enhanced ultrasound using deep learning according to claim 1, characterized in that, The process of preprocessing the T-tube sinus tract contrast-enhanced ultrasound image to obtain a preprocessed T-tube sinus tract contrast-enhanced ultrasound image includes: Denosing the T-tube sinus tract contrast-enhanced ultrasound image to obtain a denoised T-tube sinus tract contrast-enhanced ultrasound image; Enhancing the contrast of the denoised T-tube sinus tract contrast-enhanced ultrasound image to obtain a pre-enhanced T-tube sinus tract contrast-enhanced ultrasound image; Extracting edges from the T-tube sinus tract contrast-enhanced ultrasound enhanced image to obtain a T-tube sinus tract contrast-enhanced ultrasound edge image; Normalizing the pre-enhanced T-tube sinus tract contrast-enhanced ultrasound image and combining it with the T-tube sinus tract contrast-enhanced ultrasound edge image to obtain the preprocessed T-tube sinus tract contrast-enhanced ultrasound image.

5. The method for identifying the T-tube sinus tract by contrast-enhanced ultrasound based on deep learning according to claim 1, wherein The multi-task contrast-enhanced ultrasound image enhancement model includes: an input layer, an adaptive adjustment layer, a motion compensation layer, a detail enhancement layer, and an output layer; Among them, the input layer is used to perform feature transformation on the contrast-enhanced ultrasound image and device parameters; The adaptive adjustment layer is used to generate adaptive parameters according to the ultrasound device parameters; The motion compensation layer is used to learn motion information from consecutive contrast-enhanced ultrasound images of the T-tube sinus tract and perform motion compensation in combination with the adaptive parameters; The detail enhancement layer is used to learn high-frequency details in the contrast-enhanced ultrasound image of the T-tube sinus tract in combination with image edge features and adjust the enhancement intensity through the adaptive parameters; The output layer is used to output the contrast-enhanced ultrasound image of the T-tube sinus tract.

6. The method for identifying the T-tube sinus tract by contrast-enhanced ultrasound based on deep learning according to claim 1, characterized in that, The specific implementation process of inputting the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract and the ultrasound device parameters into the multi-task contrast-enhanced ultrasound image enhancement model for image enhancement and adaptive adjustment to obtain the contrast-enhanced ultrasound image of the T-tube sinus tract includes: Inputting the preprocessed contrast-enhanced ultrasound image of the T-tube sinus tract and the ultrasound device parameters into the input layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain the contrast-enhanced ultrasound image features of the T-tube sinus tract and the ultrasound device parameter features; Processing the ultrasound device parameter features by using the adaptive adjustment layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain adaptive parameters; Processing the contrast-enhanced ultrasound image features of the T-tube sinus tract by using the optical flow features predicted by the motion compensation layer of the multi-task contrast-enhanced ultrasound image enhancement model and in combination with the adaptive parameters to obtain the motion compensation features of the contrast-enhanced ultrasound image of the T-tube sinus tract; Inputting the motion compensation features of the contrast-enhanced ultrasound image of the T-tube sinus tract and the adaptive parameters into the detail enhancement layer of the multi-task contrast-enhanced ultrasound image enhancement model for detail enhancement to obtain the detail enhancement features of the contrast-enhanced ultrasound image of the T-tube sinus tract; Processing the detail enhancement features of the contrast-enhanced ultrasound image of the T-tube sinus tract by using the output layer of the multi-task contrast-enhanced ultrasound image enhancement model to obtain the contrast-enhanced ultrasound image of the T-tube sinus tract.

7. A method for identifying T-tube sinus tracts in contrast-enhanced ultrasound based on deep learning according to claim 1, wherein, The process of inputting the contrast-enhanced ultrasound image of the T-tube sinus tract into the multi-stage classification and recognition model of the contrast-enhanced ultrasound image for recognition to obtain the T-tube sinus tract recognition result includes: Using the multi-stage classification and recognition model of the contrast-enhanced ultrasound image to perform preliminary classification on the contrast-enhanced ultrasound image of the T-tube sinus tract to obtain T-tube features and non-T-tube features; Inputting the non-T-tube features into the sinus tract morphology classification layer of the multi-stage classification and recognition model of the contrast-enhanced ultrasound image to obtain the sinus tract morphology classification result; wherein, the sinus tract morphology classification result includes the first morphology classification and the second morphology classification; Selecting the corresponding sinus tract recognition layer according to the sinus tract morphology classification result, and inputting the T-tube features and the non-T-tube features into the sinus tract recognition layer for recognition to obtain the T-tube sinus tract recognition result.

Images