Pulse diagnosis algorithm based on plane wave scanning
Through a pulse diagnosis algorithm based on plane wave strafing, combined with a linear ultrasound probe and Transformer model, the problems of slow data processing and insufficient diagnostic accuracy of the existing technology mid-pipe diagnosis instrument are solved, and more efficient and accurate pulse diagnosis results are achieved.
Patent Information
- Application Number
- CN202411860746.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-17
- Publication Date
- 2025-04-11
AI Technical Summary
The existing pulse diagnosis instruments have insufficient data processing and diagnostic accuracy, and the performance limitations of terminal equipment lead to excessive data volume and slow processing, and there are problems of one-sided diagnosis and integrity.
Using a pulse diagnosis algorithm based on plane wave strafing, a linear ultrasound probe emits plane waves, collects reflected signals, and establishes a data set of ultrasound signals corresponding to the ultrasound signal. The model is trained to convert the pulse ultrasound signal into IQ signal, and data enhancement is enhanced by combining the traditional Chinese medicine knowledge graph, and training is performed using the Transformer model.
It improves the standardization and efficiency of pulse signal acquisition, enhances the richness of data information, improves the accuracy and efficiency of diagnosis, and reduces the subjectivity of diagnosis results.
Smart Images
Figure CN119807924A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence, and in particular to a pulse diagnosis algorithm based on plane wave scanning. Background Art
[0002] Pulse diagnosis is one of the four most important diagnoses in traditional Chinese medicine. It is to infer the patient's physical condition by observing and touching the patient's pulse shape, rhythm, strength and other characteristics. As a unique diagnostic method, pulse diagnosis occupies an important position in traditional Chinese medicine. However, the traditional pulse diagnosis method is highly dependent on the doctor's experience and senses, and has defects such as strong subjectivity, low diagnostic accuracy, and difficult to quantify results. This requires a high level of experience and level of the doctor, and the diagnostic results are easily affected by personal subjective factors and are not accurate enough. In addition, traditional Chinese pulse diagnosis requires the doctor to manually palpate the patient's pulse, and each diagnosis takes a certain amount of time, the diagnostic efficiency is low, and the learning cost is high, which is not conducive to widespread promotion.
[0003] In order to solve the above problems, pulse diagnosis instruments came into being. Their main purpose is to realize the objectivity of traditional Chinese medicine pulse diagnosis. Pulse diagnosis instruments use the feeling of fingers to analyze the characteristics of pulse such as "position, number, shape, and potential" to judge the functional state of internal organs, thereby achieving the purpose of non-invasive pulse diagnosis. However, existing pulse diagnosis instruments usually process information on terminal devices to form a final diagnosis report. This method is prone to problems such as excessive data volume, slow processing and analysis, and inaccurate results. At the same time, due to the limitations of terminal equipment performance, patients are often required to enter the characteristics of the disease for analysis by themselves, which poses the risk of one-sided diagnosis and also brings about integrity issues.
[0004] At present, AI big models, with their ultra-large-scale parameters (usually more than one billion) and powerful computing resources, are able to process massive amounts of data and complete various complex tasks. Different types of big models can process different types of data, such as language big models for processing language data and visual big models for processing visual data. However, the application of big models to biomedical signal processing, especially pulse diagnosis using IQ signals obtained by plane wave scanning, still faces unique challenges.
[0005] Most large model applications are concentrated in the fields of language and image processing. The existing data-driven biomedical signal analysis methods lack the embedding of professional knowledge and experience, and the accuracy of the models needs to be improved. Although the methods based on case reasoning and recommendation algorithms can provide some references, they are often just a copy of historical solutions and cannot integrate current signal information with empirical knowledge to generate targeted diagnostic results.
[0006] At the same time, although large visual language models perform well in understanding and generating complex language descriptions related to visual content, they may face challenges when performing some simple signal classification tasks. These models are usually designed to handle more advanced language understanding and generation tasks rather than directly targeting the classification and analysis of biomedical signals.
[0007] Therefore, there is an urgent need for a large-model pulse diagnosis algorithm based on plane wave scanning IQ signals as input, so as to accurately analyze the signals and generate reliable results. Summary of the invention
[0008] The purpose of the present invention is to solve the problems existing in the prior art, provide a pulse diagnosis algorithm based on plane wave scanning, combine the experience of traditional Chinese medicine pulse diagnosis with modern computer technology, analyze the patient's pulse signal, give quantitative results, ensure the efficiency and security of data processing, and help improve the accuracy and efficiency of diagnosis.
[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions:
[0010] The pulse diagnosis algorithm based on plane wave scanning includes the following steps:
[0011] S1. Place the linear ultrasound probe close to the skin surface, and emit a plane wave without deflection from the direction in which the ultrasound probe faces perpendicular to the skin toward the inside of the skin;
[0012] S2, collecting the reflected signal as pulse information;
[0013] S3, establishing a disease data set corresponding to ultrasound signals;
[0014] S4. Based on the dataset training model, firstly, the pulse ultrasound signal is processed into an orthogonal IQ signal through linear processing; then the body state is converted into a multi-label distribution in the complete set of TCM symptoms, and all the symptoms present in the patient are set to 1, and the rest are set to 0; finally, the dataset is learned based on the language model containing time series information;
[0015] S5. Model-assisted acceleration of data collection.
[0016] Preferably, in step S1, the number of array elements of the linear ultrasonic probe is at least 64, and the center frequency of the probe is not less than 5 Mhz and not higher than 15 Mhz.
[0017] Preferably, in step S1, the total width of the linear ultrasonic probe is not less than 3 cm, and the length of the short side is not less than 5 mm.
[0018] Preferably, in step S1, the wide side of the linear ultrasound probe is parallel to the blood flow direction of the radial artery, and the linear ultrasound probe should cover the pulsation range of the radial artery at the wrist.
[0019] Preferably, in step S3, the patient's pulse ultrasonic signal and physical condition are collected simultaneously, the pulse ultrasonic signal is collected using steps S1-S2, and the patient's physical condition file is constructed using a medical interview method, thereby collecting the correspondence between the patient's pulse ultrasonic signal and physical condition at that moment and storing it as a piece of data in the data set.
[0020] Preferably, in step S3, the pulse ultrasound signal is enhanced by data enhancement, and the body state is randomly modified to expand the data set.
[0021] Preferably, in step S4, a randomly initialized Transformer is used as the initial model, the dimension with a length of the number of sampling times is regarded as a timestamp, and the two-dimensional matrix with a size of [number of array elements, 2] at each timestamp is converted into a one-dimensional vector with a length of [number of array elements*2] by direct concatenation, and this two-dimensional matrix with a shape of [number of sampling times, number of array elements*2] is used as the input of the Transformer model.
[0022] Preferably, in step S4, the data set is imported into the model for training in batches, the batch size is 16, the AdamW optimizer is used, the initial learning rate is 1e-4, and the learning rate is adjusted by CosineAnnealScheduler, and a total of 20 Epochs are trained.
[0023] Preferably, in step S5, the model continuously learns and aligns according to subsequently collected data and patient feedback.
[0024] Compared with the prior art, the beneficial effects of the present invention are:
[0025] 1. It is easy to standardize signal acquisition. The traditional pulse signal acquisition method is based on pressure sensors or photoelectric sensors. The problem with these sensors is that they are difficult to operate when collecting pulse signals. Taking the pressure sensor as an example, the radial artery is hidden under the radius, and a smaller pressure sensor is needed to avoid the interference of the radius and detect the fluctuations of the radial artery. At the same time, the pressure sensor is very sensitive to the force between the sensor and the skin surface. Any unexpected force will cause problems in signal acquisition. The same is true for photoelectric sensors. The advantage of an ultrasonic probe is that it does not need to be strictly fixed to the skin surface, because the detection range of the ultrasonic probe is an area where the probe extends, not just the place where the probe contacts. This means that when collecting signals, the ultrasonic probe only needs to ensure that the radial artery is within the detection area, which greatly reduces the difficulty of signal acquisition and facilitates the standardization of signal acquisition.
[0026] 2. More data information. Traditional pulse diagnosis algorithms are also limited to the collection method (pressure sensor, photoelectric sensor), and their data can often only reflect a few characteristics of the pulse. For example, the pressure sensor can reflect the pulse frequency and pulsation strength, but cannot reflect the depth of the radial artery and blood flow rate. Ultrasonic signals are very high in dimensionality and can be restored to multiple modal data through signal processing, such as pulsation frequency, pulsation strength, blood flow rate, radial artery depth, vasodilation, etc. These rich data can enable the model to understand more information about pulse diagnosis and improve the theoretical upper limit of model performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0028] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.
[0029] Plane wave: A wave whose wavefront (i.e., the surface formed by points with the same phase of the wave) is flat and whose wavefront propagates in a certain direction in space.
[0030] Large models: Machine learning models with very large parameters and computing power.
[0031] The present invention proposes to represent pulse diagnosis information based on ultrasonic signals, and solves the problem of unstable pulse signal acquisition by establishing a standard acquisition process for pulse ultrasonic signals. In addition, compared with the current pressure signals and photoelectric signals, the information dimension of ultrasonic signals is higher, which enables the model to learn deeper information, have a deeper understanding of the signal, and improve the accuracy and robustness of the algorithm. At the same time, the present invention performs data enhancement on the model based on the TCM knowledge graph, aligns the model with the TCM knowledge graph, and improves the performance of the model on TCM specialized tasks.
[0032] Example: As attached Figure 1 As shown, the present invention is a pulse diagnosis algorithm based on plane wave scanning, which specifically includes the steps of:
[0033] S1. Ultrasonic probe transmits plane wave:
[0034] Use a linear ultrasound probe to transmit plane waves. The number of array elements of the linear ultrasound probe should be at least 64, and generally 128 array elements should be used. The probe center frequency should be no less than 5Mhz and no higher than 15Mhz, and generally 10Mhz. The array element width and spacing are arbitrary, but the total width of the probe should be no less than 3cm, and the short side length should be no less than 5mm. When collecting signals, place the linear ultrasound probe close to the skin surface (a small amount of ultrasound coupling agent is used to fill the space between the probe and the skin surface to ensure that the quality of the ultrasound signal will not be affected by air). The wide side should be parallel to the blood flow direction of the radial artery, and the probe should cover the range of the radial artery pulsation at the wrist. After the ultrasound probe is correctly placed and fixed, an undeflected plane wave is emitted from the ultrasound probe in a direction perpendicular to the skin surface toward the inside of the skin.
[0035] S2. Collecting reflected signals as pulse information:
[0036] After the plane wave is transmitted, the ultrasonic signal reflected on the array element is collected. The signal contains the information of the radial artery at that moment. The received reflected signal is sampled using a sampling frequency 4 times the probe center frequency, and the sampling time should be at least 5 seconds. The sampled signal does not undergo signal processing such as Fourier transform, but directly stores the original signal collected by the probe to represent the pulse information.
[0037] S3. Establishing a dataset of diseases corresponding to ultrasound signals:
[0038] The pulse ultrasound signal and physical condition (a collection of all the symptoms the patient suffers from at the moment) of the patient are collected simultaneously in the form of pulse diagnosis and medical questioning. For the pulse ultrasound signal, it is collected in the manner of steps S1-S2, and for the physical condition, the physical condition file of the patient is constructed using the doctor's medical questioning method, thereby collecting the corresponding relationship between the pulse ultrasound signal and the physical condition of the patient at that moment, and storing it as a piece of data in the data set.
[0039] For a piece of data, this embodiment uses data enhancement to enhance the pulse ultrasound signal of the patient, and performs covering operations on part of the ultrasound signal, including covering the full-time signals of certain array elements, or covering the full-time signals of certain array elements. At the same time, some unimportant content in the patient's physical state is randomly modified, thereby expanding a single piece of data into multiple pieces of data, and constructing a data set containing more than 100,000 pieces of data on the correspondence between pulse ultrasound signals and symptoms.
[0040] S4. Train the model based on the dataset:
[0041] For a corresponding data of "pulse ultrasound signal" and "physical state" in the data set, this embodiment converts the pulse ultrasound signal (RF signal is stored here) into an orthogonal IQ signal through linear processing, thereby converting the pulse ultrasound signal into a three-dimensional matrix with three-dimensional sizes of [number of samplings, number of array elements, 2], where the number of array elements is the number of array elements of the ultrasound probe, and the number of samplings is the sampling time * sampling frequency; for the physical state, it is converted into a multi-label distribution in the complete set of TCM symptoms, and all the symptoms existing in the patient are set to 1, and the rest are set to 0. In this way, the data set can be converted into a classic multi-label classification problem data set.
[0042] Then, this dataset is learned based on the language model containing timing information. Specifically, a randomly initialized Transformer is used as the initial model, and the dimension with a length of the number of sampling times is regarded as a timestamp. The two-dimensional matrix of size [number of array elements, 2] at each timestamp is directly concatenated into a one-dimensional vector with a length of [number of array elements*2], and this two-dimensional matrix with a shape of [number of sampling times, number of array elements*2] is used as the input of the Transformer model.
[0043] The data set is imported into the model for training in batches, with a batch size of 16. The AdamW optimizer is used, the initial learning rate is 1e-4, and the learning rate is adjusted using the CosineAnnealScheduler. A total of 20 Epochs are trained.
[0044] S5. Model-assisted data collection:
[0045] After the model is trained, it is deployed in the data collection process. During the consultation process, the doctor can see the symptoms that the model predicts based on the patient's pulse ultrasound signal, which improves the doctor's consultation efficiency and obtains feedback from the patient on the model's recommended results. The model will continue to learn and align based on the subsequent collected data and patient feedback.
Claims
1. The pulse diagnosis algorithm based on plane wave scanning is characterized by: Includes steps: S1. Place the linear ultrasound probe close to the skin surface, and emit a plane wave without deflection from the direction in which the ultrasound probe faces perpendicular to the skin toward the inside of the skin; S2, collecting the reflected signal as pulse information; S3, establishing a disease data set corresponding to ultrasound signals; S4. Based on the dataset training model, firstly, the pulse ultrasound signal is processed into an orthogonal IQ signal through linear processing; then the body state is converted into a multi-label distribution in the complete set of TCM symptoms, and all the symptoms present in the patient are set to 1, and the rest are set to 0; finally, the dataset is learned based on the language model containing time series information; S5. Model-assisted acceleration of data collection.
2. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S1, the number of array elements of the linear ultrasonic probe is at least 64, and the probe center frequency is not less than 5Mhz and not higher than 15Mhz.
3. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S1, the total width of the linear ultrasonic probe is not less than 3 cm, and the short side length is not less than 5 mm.
4. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S1, the wide side of the linear ultrasound probe is parallel to the blood flow direction of the radial artery, and the linear ultrasound probe should cover the pulsation range of the radial artery at the wrist.
5. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S3, the patient's pulse ultrasonic signal and physical condition are collected at the same time. The pulse ultrasonic signal is collected using steps S1-S2, and the patient's physical condition file is constructed using a medical interview method. In this way, the correspondence between the patient's pulse ultrasonic signal and physical condition at that moment can be collected and stored as a piece of data in the data set.
6. The pulse diagnosis algorithm based on plane wave scanning according to claim 5, characterized in that: In step S3, the pulse ultrasound signal is enhanced by data enhancement, and the body state is randomly modified to expand the data set.
7. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S4, a randomly initialized Transformer is used as the initial model, and the dimension with a length of the number of sampling times is regarded as a timestamp. The two-dimensional matrix of size [number of array elements, 2] at each timestamp is converted into a one-dimensional vector with a length of [number of array elements*2] by direct concatenation, and this two-dimensional matrix with a shape of [number of sampling times, number of array elements*2] is used as the input of the Transformer model.
8. The pulse diagnosis algorithm based on plane wave scanning according to claim 7, characterized in that: In step S4, the data set is imported into the model for training in batches, with a batch size of 16, using the AdamW optimizer, an initial learning rate of 1e-4, and the learning rate is adjusted using the CosineAnnealScheduler, and a total of 20 epochs are trained.
9. The pulse diagnosis algorithm based on plane wave scanning according to claim 1, characterized in that: In step S5, the model continuously learns and aligns based on the subsequently collected data and patient feedback.