A Doppler missing data filling method based on deep learning

Abstract

The present invention discloses a Doppler missing data filling method based on deep learning, including: constructing a prediction model based on deep neural network technology; obtaining PW sampling data under synchronous imaging mode; the PW sampling data includes: valid data arranged in time sequence and data existing at intervals of missing data; inputting the PW sampling data into the prediction model, and outputting corresponding filling data for each section of missing data through the prediction of the prediction model; filling the filling data into the corresponding missing data to form continuous and complete PW data. PW signal data of different parts of the human body are collected to form training samples, and supervised training of deep learning models is performed. Training is performed through real collected data, and then based on the deep network obtained by training, the missing data is estimated using the signal characteristics before and / or after the missing data, effectively improving the accuracy of missing data filling.

Description

Technical Field

[0001] The present invention relates to the field of Doppler measurement technology, and in particular to a Doppler missing data filling method based on deep learning. Background Art

[0002] Medical ultrasound Doppler spectroscopy is a medical imaging technique used to measure the speed and direction of blood flow. It is based on the Doppler effect, which measures the speed and direction of blood flow by sending high-frequency sound waves and receiving echoes. Doppler spectroscopy can help doctors detect conditions such as vascular stenosis, hemangiomas, and abnormal heart function. In medicine, Doppler spectroscopy is also widely used in prenatal examinations, heart disease diagnosis, and blood flow monitoring during surgery.

[0003] B-mode is a two-dimensional ultrasound imaging mode that can display the internal tissue structure and organ morphology of the human body. B-mode imaging scans ultrasound waves inside the human body and converts the echo signals into grayscale images, thereby presenting the morphology and structure of the internal tissues and organs of the human body. PW mode is a pulsed Doppler mode that can measure blood flow speed and direction. In PW mode, the ultrasound probe sends a short pulse and then receives the reflected signal. Based on the Doppler effect, the blood flow speed and direction can be calculated by analyzing the changes in the signal frequency.

[0004] Synchronization means that during ultrasound imaging, multiple imaging modes, such as B mode, color mode, and PW mode, can be displayed simultaneously. In this way, doctors can observe the internal structure and organ morphology of the human body in B mode, and measure the blood flow speed and direction in PW mode, so as to have a more comprehensive understanding of the condition and make an accurate diagnosis. Due to hardware system limitations, B mode and PW mode cannot collect signals at the same time, and time division multiplexing must be used. This results in missing data in PW mode in some cases when B mode is collecting signals.

[0005] To address the problem of missing data, existing technologies generally use model-based methods such as data interpolation and model estimation. Traditional methods usually require manual design of features and rules, and professional knowledge to design and extract features. However, in some specific scenarios, traditional methods may be more accurate. However, the model is not very adaptable and cannot meet the needs of spectra collected under different conditions, different subjects, and different parts of the body.

[0006] For example, patent US5891036 uses two valid sequences before and after the missing data segment, and uses the real and imaginary parts of the signal to weightedly fill the missing data in positive and negative order. Patent US8206302 extracts the data before and after the missing signal, and uses frequency analysis and compensation methods to fill the missing signal. Patent CN104783836 uses a combination of time series compensation and frequency analysis compensation to interpolate and compensate for the signal gap between the two signal segments. Patent US5642732 uses the autoregressive model method to estimate the parameters of the two signals before and after the missing signal, and uses the estimated parameters to construct an IIR filter to estimate the missing signal.

[0007] The above methods use fixed models and parameters to estimate missing data, so the estimation of missing data is not smooth and continuous. It is difficult to make an accurate estimate when the frequency of missing data changes dramatically. Summary of the invention

[0008] The present invention provides a Doppler missing data filling method based on deep learning to solve the above problems existing in the prior art.

[0009] The present invention provides a Doppler missing data filling method based on deep learning, comprising:

[0010] S100, which builds prediction models based on deep neural network technology;

[0011] S200, obtaining PW sampling data in a synchronous imaging mode; the PW sampling data includes: data existing at intervals of valid data and missing data arranged in a time sequence;

[0012] S300, inputting the PW sampling data into the prediction model, and outputting corresponding filling data for each segment of missing data through prediction by the prediction model;

[0013] S400, filling the filling data into the corresponding missing data to form continuous and complete PW data.

[0014] Preferably, the S100 includes:

[0015] S101, obtaining a plurality of complete blood flow signal data collected by a pulsed Doppler ultrasound device as original training samples;

[0016] S102, deleting part of the complete blood flow signal data in a preset manner to form a missing segment, where there is no data in the missing segment, and using the data with the missing segment as an actual training sample;

[0017] S103, using a deep neural network model to extract the front segment data and / or the back segment data of the missing segment in the actual training sample for training prediction to obtain filling data;

[0018] S104: Using the filled data as output of the prediction model.

[0019] Preferably, the S103 further includes:

[0020] S1031, after deleting a segment of data in a random manner, storing the deleted segment of data as feedback data;

[0021] S1032, comparing the filling data obtained by training prediction with the feedback data to obtain an error evaluation value of the prediction model;

[0022] S1033: If the error evaluation value is greater than or equal to the set evaluation threshold, a correction parameter of the prediction model is obtained based on the feedback data by means of supervised training, and the prediction model is optimized based on the correction parameter.

[0023] Preferably, the S101 includes:

[0024] S1011, dividing the collection part into a plurality of collection parts, each collection part corresponding to a corresponding part identifier; dividing the prediction model into sub-prediction models according to the part identifiers;

[0025] S1012, for each acquisition site, multiple complete blood flow signal data acquired by a pulsed Doppler ultrasound device are used as local original training samples;

[0026] S1013, storing the local original training samples according to the part identifiers;

[0027] S1014, after each local original training sample is obtained, the output of the sub-prediction model is obtained according to steps S102 to S104.

[0028] Preferably, before step S300, the method further includes:

[0029] Determine the collection location to which the PW sampling data belongs;

[0030] The sub-prediction model corresponding to the PW sampling data is determined according to the site identifier corresponding to the collection site, and the PW sampling data is input into the sub-prediction model.

[0031] Preferably, in S102, the complete blood flow signal data is deleted in a preset manner, including:

[0032] S1021, deleting one or more segments of data by random selection, where the deleted segments of data are discontinuous;

[0033] S1022, setting the length of the deleted data and the corresponding time starting point; the length of the deleted data is greater than or equal to a first length threshold and less than or equal to a second length threshold, and the first length threshold is less than the second length threshold;

[0034] S1023, marking the deleted data segments according to a time sequence, and recording and storing the deleted data according to the same time sequence.

[0035] Preferably, the S100 further includes:

[0036] S105, constructing a recurrent neural network model RNN, and using the recurrent neural network model RNN as a prediction model; the recurrent neural network model RNN includes: a long short-term memory neural network LSTM or a gated recurrent neural network GRU;

[0037] S106, obtaining PW sampling data that needs to be filled with data, wherein the PW sampling data includes multiple sections of valid data and multiple sections of missing data, and there is a vacant part between two sections of valid data, which needs to be filled with data;

[0038] S107, input the PW sampling data into the recurrent neural network model RNN, analyze the valid data based on the recurrent neural network model RNN, obtain predicted data based on the valid data, use the predicted data as filling data, and complete the missing data.

[0039] Preferably, the S200 includes:

[0040] S201, setting different combination modes under the synchronous imaging mode, the combination modes including: B mode + PW mode, B mode + color mode + PW mode;

[0041] S202, the scanning sequence is scanned and collected according to different set combination modes; the combination mode is presented in a manner of alternating scanning with the PW mode;

[0042] S203, when the PW mode is used for scanning and acquisition, valid data is generated, and when the B mode or color mode is used for scanning and acquisition, data corresponding to the PW sampling data is missing, thus generating missing data.

[0043] Preferably, the S1033 includes:

[0044] S1033-1, obtaining the most appropriate data parameters according to the feedback data;

[0045] S1033-2, optimizing the model parameters of the optimizer, initial learning rate, sample quantity parameter, and epoch parameter in the prediction model to form multiple optimization models;

[0046] S1033-3, obtaining the parameters corresponding to the optimization model with the smallest error, and optimizing the prediction model based on the parameters.

[0047] The present invention also provides a system of Doppler synchronous imaging mode, comprising: a beamforming control unit, a transmission beamforming module, a transmitting unit, an ultrasound probe, an A / D analog-to-digital conversion module, a receiving beamforming unit, a PW mode image forming unit, a missing data filling unit, a synchronous mode image forming unit and a display unit;

[0048] The beamforming control unit controls the transmitting beamforming module, excites the transmitting unit, and excites the ultrasonic probe to send an excitation pulse signal; after the ultrasonic probe is transmitted, it starts to receive echo data, which is converted into a digital signal through the A / D analog-to-digital conversion module, and after corresponding processing by the receiving beamforming unit, a synchronous mode image is formed through the synchronous mode image forming unit, and a PW mode image is formed through the PW mode image forming unit. The PW mode image is filled with missing data through the missing data filling unit to form complete continuous data, and the complete continuous data is displayed through the display unit; the synchronous mode image is directly displayed through the display unit;

[0049] The missing data filling unit is specifically used to complete the filling of missing data by executing the Doppler missing data filling method to form continuous and complete PW data.

[0050] Compared with the prior art, the present invention has the following advantages:

[0051] The present invention provides a Doppler missing data filling method based on deep learning, including: constructing a prediction model based on deep neural network technology; obtaining PW sampling data under synchronous imaging mode; the PW sampling data includes: valid data arranged in time sequence and data existing at intervals of missing data; inputting the PW sampling data into the prediction model, and outputting corresponding filling data for each section of missing data through the prediction of the prediction model; filling the filling data into the corresponding missing data to form continuous and complete PW data. PW signal data of different parts of the human body are collected and deep learning model training is performed. Using the trained deep network, the missing data is estimated by using the bidirectional signal features before and after the missing data, or only using the unilateral features before or after the missing signal, which can effectively improve the accuracy of missing data filling. With the further expansion of training data, the accuracy of the model method is also further improved.

[0052] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings.

[0053] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0055] Figure 1 This is a flow chart of a Doppler missing data filling method based on deep learning in an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of the structure of PW sampling data generation in an embodiment of the present invention;

[0057] Figure 3 The figure is a schematic diagram of the structure of a system of Doppler synchronous imaging mode in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0059] The embodiment of the present invention provides a Doppler missing data filling method based on deep learning, please refer to Figure 1 , the Doppler missing data filling method based on deep learning includes:

[0060] S100, which builds prediction models based on deep neural network technology;

[0061] S200, obtaining PW sampling data in a synchronous imaging mode; the PW sampling data includes: data existing at intervals of valid data and missing data arranged in a time sequence;

[0062] S300, inputting the PW sampling data into the prediction model, and outputting corresponding filling data for each segment of missing data through prediction by the prediction model;

[0063] S400, filling the filling data into the corresponding missing data to form continuous and complete PW data.

[0064] The working principle of the above technical solution is: the solution adopted in this embodiment is to build a prediction model based on deep neural network technology; obtain PW sampling data under synchronous imaging mode; the PW sampling data includes: valid data and data existing at intervals of missing data arranged in time sequence; the PW sampling data is input into the prediction model, and through the prediction of the prediction model, corresponding filling data is output for each segment of missing data; the filling data is filled into the corresponding missing data to form continuous and complete PW data.

[0065] The present invention adopts PW sampling data combined with deep learning methods to collect PW signal data from different parts of the human body and perform deep learning model training. Using the trained deep network, the missing data is estimated using the bidirectional signal features before and after the missing data, or only using the unilateral features before or after the missing signal, which can effectively improve the accuracy of missing data filling. With the further expansion of training data, the accuracy of the model method is further improved.

[0066] The beneficial effect of the above technical solution is: adopting the solution provided by this embodiment and using the deep learning technology method, in the ultrasound PW synchronization mode, the missing data is compensated and filled, so that the supplemented data is as close to the actual situation as possible, and the PW image noise and artifacts caused by missing data or poor compensation effect are reduced.

[0067] In another embodiment, the S200 includes:

[0068] S201, setting different combination modes in the synchronous imaging mode, the combination modes including: B mode + PW mode, B mode + color mode + and PW mode;

[0069] S202, the scanning sequence is scanned and collected according to different set combination modes; the combination mode is presented in a manner of alternating scanning with the PW mode;

[0070] S203, when the PW mode is used for scanning and acquisition, valid data is generated, and when the B mode or color mode is used for scanning and acquisition, data corresponding to the PW sampling data is missing, thus generating missing data.

[0071] The working principle of the above technical solution is: the solution adopted in this embodiment is that when the B mode, C mode and PW mode are used for data acquisition at the same time, the time division multiplexing method is intelligently adopted, and the B mode, C mode or PW mode is used in a certain time period, and the combination mode is presented in an alternating scanning method with the PW mode, for example, Figure 2As shown, the scanning order is: PW mode, B mode, PW mode, C mode, PW mode. Therefore, the generated PW sampling data is valid data generated when scanning in PW mode. When scanning in B mode and C mode, the PW sampling data has no data and is missing data.

[0072] In another embodiment, the S100 includes:

[0073] S101, obtaining a plurality of complete blood flow signal data collected by a pulsed Doppler ultrasound device as original training samples;

[0074] S102, deleting part of the complete blood flow signal data in a preset manner to form a missing segment, where there is no data in the missing segment, and using the data with the missing segment as an actual training sample;

[0075] S103, using a deep neural network model to extract the front segment data and / or the back segment data of the missing segment in the actual training sample for training prediction to obtain filling data;

[0076] S104: Using the filled data as output of the prediction model.

[0077] The working principle of the above technical solution is as follows:

[0078] The specific schemes for data sample collection and expansion are as follows:

[0079] We use existing ultrasound equipment to collect complete human blood flow signals (without gaps) as our training samples. We randomly erase a segment of data from the collected one-dimensional signal to simulate the missing data when a gap occurs. The erased data is used as the output of the model to supervise the training model. In this way, we can simply and accurately simulate the situation of gaps of different lengths at different signal time points, reducing the difficulty of collecting and labeling data.

[0080] The model and training steps are implemented as follows:

[0081] Using mature deep learning-based time series signal prediction technology, from the perspective of model selection, we can use the RNN model of the recurrent neural network, including but not limited to the long short-term memory neural network LSTM (Long-short term memory) and the gated recurrent neural network GRU (gated recurrent neural network), etc., and use the data before and after the gap to comprehensively predict the missing signal in the middle. Specifically, the use of LSTM or GRU is to better capture the dependency between time series data, predict the data of the vacant part, obtain the predicted data to fill the vacant part, and use it to fill the corresponding vacant part, and connect with the data before and after the vacant part to form complete data. Another feasible solution is to use the popular Transformer structure (attention mechanism), which can directly realize long-range connection. Considering the way we collect and generate training data, the large amount of data required for Transformer training can be easily obtained.

[0082] The scheme adopted for model optimization and implementation is as follows:

[0083] Considering that the model needs to generate gap data in real time, we first adopt a lightweight design when designing the model to reduce the amount of calculation in the reasoning process, such as reducing the number of neurons in the middle layer and the number of heads in the multi-head attention mechanism. On the other hand, we use the deep learning framework and the deployment tools provided by Nvidia to deploy the model to the machine, further speeding up the model reasoning speed and meeting the hardware requirements on our equipment.

[0084] In another embodiment, the S103 further includes:

[0085] S1031, after deleting a segment of data in a random manner, storing the deleted segment of data as feedback data;

[0086] S1032, comparing the filling data obtained by training prediction with the feedback data to obtain an error evaluation value of the prediction model;

[0087] S1033: If the error evaluation value is greater than or equal to the set evaluation threshold, a correction parameter of the prediction model is obtained based on the feedback data by means of supervised training, and the prediction model is optimized based on the correction parameter.

[0088] The working principle of the above technical solution is: the solution adopted in this embodiment is that S103 also includes: S1031, after deleting a section of data in a random manner, the deleted section of data is stored as feedback data; S1032, the filling data obtained by training prediction is compared with the feedback data to obtain an error evaluation value of the prediction model; S1033, if the error evaluation value is greater than or equal to the set evaluation threshold, a supervised training method is used to obtain a correction parameter of the prediction model based on the feedback data, and the prediction model is optimized based on the correction parameter.

[0089] The beneficial effect of the above technical solution is: the solution provided in this embodiment uses feedback data, that is, the actual collected data (feedback data) as the true value in the supervised training process. By using such a supervised training method, the obtained filling data is closer to the true value, and therefore, the estimation of missing data is more accurate.

[0090] In another embodiment, the S101 includes:

[0091] S1011, dividing the collection part into a plurality of collection parts, each collection part corresponding to a corresponding part identifier; dividing the prediction model into sub-prediction models according to the part identifiers;

[0092] S1012, for each acquisition site, multiple complete blood flow signal data acquired by a pulsed Doppler ultrasound device are used as local original training samples;

[0093] S1013, storing the local original training samples according to the part identifiers;

[0094] S1014, after each local original training sample is obtained, the output of the sub-prediction model is obtained according to steps S102 to S104.

[0095] The working principle of the above technical solution is: the solution adopted in this embodiment is that the S101 includes: S1011, dividing the collection site into multiple collection sites, each collection site corresponds to a corresponding site identification; dividing the prediction model into sub-prediction models according to the site identification; S1012, for each collection site, multiple complete blood flow signal data collected by the pulse Doppler ultrasound device are used as local original training samples; S1013, the local original training samples are stored according to the site identification; S1014, after each local original training sample is obtained, according to steps S102 to S104, the output of the sub-prediction model is obtained.

[0096] In another embodiment, before step S300, the method further includes:

[0097] Determine the collection location to which the PW sampling data belongs;

[0098] The sub-prediction model corresponding to the PW sampling data is determined according to the site identifier corresponding to the collection site, and the PW sampling data is input into the sub-prediction model.

[0099] The working principle of the above technical solution is: the solution adopted in this embodiment is before the S300 step, and also includes: judging the collection part to which the PW sampling data belongs; determining the sub-prediction model corresponding to the PW sampling data according to the part identifier corresponding to the collection part, and inputting the PW sampling data into the sub-prediction model.

[0100] The beneficial effects of the above technical solution are: the solution provided in this embodiment adopts a method of dividing the acquisition parts, and different sub-prediction models can be used to obtain filling data for different parts. Since the parameters of the prediction model of each part may be slightly different, in order to minimize the impact of different parts, sub-prediction models with different parameters can be set for different acquisition parts to improve the accuracy of model estimation.

[0101] In another embodiment, in S102, deleting part of the complete blood flow signal data in a preset manner includes:

[0102] S1021, deleting one or more segments of data by random selection, where the deleted segments of data are discontinuous;

[0103] S1022, setting the length of the deleted data and the corresponding time starting point; the length of the deleted data is greater than or equal to a first length threshold and less than or equal to a second length threshold, and the first length threshold is less than the second length threshold;

[0104] S1023, marking the deleted data segments according to a time sequence, and recording and storing the deleted data according to the same time sequence.

[0105] The working principle of the above technical solution is: the solution adopted in this embodiment is that in S102, part of the data of the complete blood flow signal data is deleted in a preset manner, including: S1021, deleting one or more segments of data in a random selection manner, and the deleted multiple segments of data are discontinuous; S1022, setting the length of the deleted data and the corresponding time starting point; the length of the deleted data is greater than or equal to the first length threshold, and less than or equal to the second length threshold, and the first length threshold is less than the second length threshold; S1023, marking the deleted multiple segments of data according to the time series, and recording and storing the deleted data according to the same time series.

[0106] The beneficial effect of the above technical solution is: the solution provided by this embodiment is used to record the time period corresponding to the deleted data, to ensure the accurate connection between the filled data and the valid data, and to ensure the integrity and accuracy of the data.

[0107] In another embodiment, the S100 further includes:

[0108] S105, constructing a recurrent neural network model RNN, and using the recurrent neural network model RNN as a prediction model; the recurrent neural network model RNN includes: a long short-term memory neural network LSTM or a gated recurrent neural network GRU;

[0109] S106, obtaining PW sampling data that needs to be filled with data, wherein the PW sampling data includes multiple sections of valid data and multiple sections of missing data, and there is a vacant part between two sections of valid data, which needs to be filled with data;

[0110] S107, input the PW sampling data into the recurrent neural network model RNN, analyze the valid data based on the recurrent neural network model RNN, obtain predicted data based on the valid data, use the predicted data as filling data, and complete the missing data.

[0111] The working principle of the above technical solution is as follows: the solution adopted in this embodiment is that the S100 also includes: S105, constructing a recurrent neural network model RNN, and using the recurrent neural network model RNN as a prediction model; the recurrent neural network model RNN includes: a long short-term memory neural network LSTM or a gated recurrent neural network GRU; S106, obtaining PW sampling data that needs to be filled with data, the PW sampling data includes multiple sections of valid data and multiple sections of missing data, and there is a vacant part between two sections of valid data, which needs to be filled with data; S107, inputting the PW sampling data into the recurrent neural network model RNN, analyzing the valid data based on the recurrent neural network model RNN, obtaining prediction data based on the valid data, using the prediction data as filling data, and completing the data of the vacant part. The structures using the two networks of LSTM and GRU do not have encoding and decoding structures, but are more suitable for time series signals. In this embodiment, the PW sampling data is time series data. Therefore, the use of LSTM and GRU can make more accurate predictions on the missing data in the PW sampling data, and the predicted filling data is more accurate and more consistent with the actual data.

[0112] The beneficial effect of the above technical solution is: the solution provided by this embodiment is similar to the structure of the encoder and decoder, which reduces the data volume requirement of the U-net model. Even in the case of small samples, the U-net model can still obtain high-precision prediction results.

[0113] In another embodiment, the S1033 includes:

[0114] S1033-1, obtaining the most appropriate data parameters according to the feedback data;

[0115] S1033-2, optimizing the model parameters of the optimizer, initial learning rate, sample quantity parameter, and epoch parameter in the prediction model to form multiple optimization models;

[0116] S1033-3, obtaining the parameters corresponding to the optimization model with the smallest error, and optimizing the prediction model based on the parameters.

[0117] The working principle of the above technical solution is: the solution adopted in this embodiment is that S1033 includes: S1033-1, obtaining the most appropriate data parameters according to the feedback data; S1033-2, optimizing the model parameters of the optimizer and the initial learning rate, sample quantity parameters, and epoch parameters in the prediction model to form multiple optimization models; S1033-3, obtaining the parameters corresponding to the optimization model with the smallest error, and optimizing the prediction model based on the parameters.

[0118] The beneficial effect of the above technical solution is: adopting the solution provided by this embodiment ensures that the optimized prediction model can estimate more accurate filling data.

[0119] In another embodiment, this embodiment also provides a system of Doppler synchronous imaging mode, please refer to Figure 3 The system includes: a beamforming control unit, a transmit beamforming module, a transmit unit, an ultrasound probe, an A / D analog-to-digital conversion module, a receive beamforming unit, a PW mode image forming unit, a missing data filling unit, a synchronous mode image forming unit and a display unit;

[0120] The beamforming control unit controls the transmitting beamforming module, excites the transmitting unit, and excites the ultrasonic probe to send an excitation pulse signal; after the ultrasonic probe is transmitted, it starts to receive echo data, which is converted into a digital signal through the A / D analog-to-digital conversion module, and after corresponding processing by the receiving beamforming unit, a synchronous mode image is formed through the synchronous mode image forming unit, and a PW mode image is formed through the PW mode image forming unit. The PW mode image is filled with missing data through the missing data filling unit to form complete continuous data, and the complete continuous data is displayed through the display unit; the synchronous mode image is directly displayed through the display unit;

[0121] The missing data filling unit is specifically configured to complete the filling of missing data by executing a Doppler missing data filling method to form continuous and complete PW data.

[0122] The working principle of the above technical solution is as follows: The solution adopted in this embodiment is to use deep learning to learn high-order features of data through multi-layer neural networks, thereby improving the characterization ability of data. Compared with traditional methods, deep learning can more accurately describe the complex characteristics of data, thereby improving the processing effect of data. Deep learning can adaptively adjust model parameters according to the characteristics of the data, so as to better adapt to different data distributions and task requirements. Compared with traditional methods, deep learning can better adapt to different application scenarios and improve the generalization ability of the model.

[0123] The beneficial effects of the above technical solution are as follows: the solution provided by this embodiment uses PW sampling data combined with a deep learning method to collect PW signal data from different parts of the human body for deep learning model training. Using the trained deep network, using the bidirectional signal features before and after the missing data, or only using the unilateral features before or after the missing signal, the missing data is estimated, which can effectively improve the accuracy of missing data filling. With the further expansion of training data, the accuracy of the model method is further improved.

[0124] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A Doppler missing data filling method based on deep learning, characterized in that: include: S100, which builds prediction models based on deep neural network technology; S200, obtaining PW sampling data in a synchronous imaging mode; The PW sampling data includes: data existing at intervals of valid data and missing data arranged in time sequence; S300, inputting the PW sampling data into the prediction model, and outputting corresponding filling data for each segment of missing data through prediction by the prediction model; S400, filling the filling data into the corresponding missing data to form continuous and complete PW data; The S200 includes: S201, setting different combination modes under the synchronous imaging mode, the combination modes including: B mode + PW mode, B mode + color mode + PW mode; S202, the scanning sequence is scanned and collected according to different set combination modes; the combination mode is presented in a manner of alternating scanning with the PW mode; S203, when the PW mode is used for scanning and acquisition, valid data is generated, and when the B mode or color mode is used for scanning and acquisition, data corresponding to the PW sampling data is missing, thus generating missing data; Wherein, the S100 includes: S101, obtaining a plurality of complete blood flow signal data collected by a pulsed Doppler ultrasound device as original training samples; S102, deleting part of the complete blood flow signal data in a preset manner to form a missing segment, where there is no data in the missing segment, and using the data with the missing segment as an actual training sample; S103, using a deep neural network model to extract the front segment data and / or the back segment data of the missing segment in the actual training sample for training prediction to obtain filling data; S104, using the filled data as the output of the prediction model; S105, constructing a recurrent neural network model RNN, and using the recurrent neural network model RNN as a prediction model; the recurrent neural network model RNN includes: a long short-term memory neural network LSTM or a gated recurrent neural network GRU; S106, obtaining PW sampling data that needs to be filled with data, wherein the PW sampling data includes multiple sections of valid data and multiple sections of missing data, and there is a vacant part between two sections of valid data, which needs to be filled with data; S107, inputting the PW sampling data into the recurrent neural network model RNN, analyzing the valid data based on the recurrent neural network model RNN to obtain prediction data based on the valid data, and using the prediction data as filling data to complete the missing data; Wherein, the S101 includes: S1011, dividing the collection part into a plurality of collection parts, each collection part corresponding to a corresponding part identifier; dividing the prediction model into sub-prediction models according to the part identifiers; S1012, for each acquisition site, multiple complete blood flow signal data acquired by a pulsed Doppler ultrasound device are used as local original training samples; S1013, storing the local original training samples according to the part identifiers; S1014, after each local original training sample is obtained, the output of the sub-prediction model is obtained according to steps S102 to S104; Among them, in said S102, deleting part of the complete blood flow signal data in a preset manner includes: S1021, deleting one or more segments of data by random selection, where the deleted segments of data are discontinuous; S1022, setting the length of the deleted data and the corresponding time starting point; the length of the deleted data is greater than or equal to a first length threshold and less than or equal to a second length threshold, and the first length threshold is less than the second length threshold; S1023, marking the deleted data segments according to a time sequence, and recording and storing the deleted data according to the same time sequence.

2. The Doppler missing data filling method based on deep learning according to claim 1, characterized in that: The S103 further includes: S1031, after deleting part of the data in a random manner, storing the deleted part of the data as feedback data; S1032, comparing the filling data obtained by training prediction with the feedback data to obtain an error evaluation value of the prediction model; S1033: If the error evaluation value is greater than or equal to the set evaluation threshold, a correction parameter of the prediction model is obtained based on the feedback data by means of supervised training, and the prediction model is optimized based on the correction parameter.

3. The Doppler missing data filling method based on deep learning according to claim 1, characterized in that: Before the step S300, the method further includes: Determine the collection location to which the PW sampling data belongs; The sub-prediction model corresponding to the PW sampling data is determined according to the site identifier corresponding to the collection site, and the PW sampling data is input into the sub-prediction model.

4. The Doppler missing data filling method based on deep learning according to claim 2, characterized in that: The S1033 includes: S1033-1, obtaining optimal data parameters according to feedback data; S1033-2, optimizing the model parameters of the optimizer, initial learning rate, sample quantity parameter, and epoch parameter in the prediction model to form multiple optimization models; S1033-3, obtaining the parameters corresponding to the optimization model with the smallest error, and optimizing the prediction model based on the parameters.

5. A system of Doppler synchronous imaging mode, characterized in that: include: A beamforming control unit, a transmit beamforming module, a transmit unit, an ultrasound probe, an A / D analog-to-digital conversion module, a receive beamforming unit, a PW mode image forming unit, a missing data filling unit, a synchronous mode image forming unit, and a display unit; The beamforming control unit controls the transmitting beamforming module, excites the transmitting unit, and excites the ultrasonic probe to send an excitation pulse signal; after the ultrasonic probe is transmitted, it starts to receive echo data, which is converted into a digital signal through the A / D analog-to-digital conversion module, and after corresponding processing by the receiving beamforming unit, a synchronous mode image is formed through the synchronous mode image forming unit, and a PW mode image is formed through the PW mode image forming unit. The PW mode image is filled with missing data through the missing data filling unit to form complete continuous data, and the complete continuous data is displayed through the display unit; the synchronous mode image is directly displayed through the display unit; The missing data filling unit is specifically used to complete the filling of missing data by executing the Doppler missing data filling method according to any one of claims 1 to 4, so as to form continuous and complete PW data.

Images