Model training method, device, electronic equipment, medium and bone mass detection system

Abstract

The disclosed embodiments disclose a model training method, device, electronic device, medium and bone mass detection system. The model training method includes: obtaining first sample data, the first sample data includes first multi-channel time series data collected from human bones by ultrasound; obtaining the bone mass label of the collection object corresponding to the first multi-channel time series data; training a multi-channel convolutional neural network model based on the first sample data and the bone mass label, wherein the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model, so that a model for obtaining bone mass information can be trained, and the model has good recognition ability in obtaining bone mass information.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing technology, and in particular to a model training method, device, electronic equipment, medium and bone mass detection system. Background Art

[0002] The lack of osteoporosis diagnostic equipment is one of the important reasons for the low diagnosis and treatment rate of osteoporosis and the high incidence of osteoporotic fractures. The "gold standard" diagnostic equipment for osteoporosis is dual-energy X-ray absorptiometry (DXA), which cannot be widely used due to its high cost, high testing costs and the presence of certain radiation, and is difficult to popularize in primary medical institutions.

[0003] Quantitative ultrasound (QUS) is one of the most widely used osteoporosis screening devices. It has the characteristics of low cost, low testing cost, short testing time and no radiation. Its working principle is to detect bone quality by using the different speeds and attenuation of ultrasound propagation in bones of different components. At present, the commonly used QUS equipment in clinical practice transmits and receives ultrasound radio-frequency (RF) signals to the bone, records the time difference between the transmitted wave and the received wave, and calculates the speed of sound (SOS) of ultrasound transmission in the bone as the output index. Since the SOS value has a certain correlation with the bone density value, QUS can theoretically evaluate bone strength. A large number of studies have confirmed that the current QUS technology using the SOS value is not consistent with DXA in the detection of osteoporosis. At present, there is still controversy about whether the SOS-based QUS can be used for osteoporosis screening and diagnosis and fragility fracture risk assessment. Summary of the invention

[0004] In order to solve the problems in the related art, the embodiments of the present disclosure provide a model training method, device, electronic device, medium and bone mass detection system.

[0005] In a first aspect, a model training method is provided in an embodiment of the present disclosure.

[0006] Specifically, the model training method includes:

[0007] Acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound;

[0008] Acquire a bone mass label of a collection object corresponding to the first multi-channel time series data;

[0009] Training a multi-channel convolutional neural network model based on the first sample data and the bone mass label,

[0010] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0011] In combination with the first aspect, in a first implementation of the first aspect of the present disclosure, obtaining the first sample data includes any one of the following:

[0012] Controlling a transmitting transducer to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal;

[0013] Control two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and a receiving transducer to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determine the first multi-channel time series data based on the received second ultrasonic signal; or

[0014] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining multi-channel time series data based on the received second ultrasonic signal,

[0015] Optionally, the first ultrasonic signal or the second ultrasonic signal is an ultrasonic radio frequency signal.

[0016] In combination with the first implementation method of the first aspect, in the second implementation method of the first aspect of the present disclosure, the obtaining of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signal, and the multi-channel convolutional neural network includes a convolutional neural network with four convolution paths.

[0017] In combination with the first aspect, in a third implementation of the first aspect of the present disclosure:

[0018] Each of the convolution paths includes 8 convolution layers;

[0019] Each of the convolutional layers is provided with 8 convolutional kernels with a length of 5;

[0020] Each of the convolutional layers uses a rectified linear unit as an activation function; and / or

[0021] A pooling layer with a length of 2 is set after each convolutional layer, and the pooling layer adopts a maximum pooling layer.

[0022] In combination with the first aspect and the first to third implementations of the first aspect, in a fourth implementation of the first aspect of the present disclosure, the multi-channel convolutional neural network model further includes a fully connected layer, wherein:

[0023] The fully connected layer is connected to the ends of the fused multiple convolution paths;

[0024] The fully connected layer has 700-750 neurons;

[0025] The fully connected layer uses a rectified linear unit as an activation function.

[0026] In combination with the fourth implementation manner of the first aspect, in a fifth implementation manner of the first aspect of the present disclosure, the multi-channel convolutional neural network model further includes an output layer, wherein:

[0027] The output layer is connected to the fully connected layer;

[0028] The output layer includes a flexible maximum function, which is used to output the first bone mass information;

[0029] The bone mass signature includes normal bone mass, osteopenia and osteoporosis.

[0030] In combination with the first aspect and the first to third implementations of the first aspect, in a sixth implementation of the first aspect of the present disclosure, the first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after the following preprocessing:

[0031] The first multi-channel time series data is processed using an eighth-order low-pass filter; and / or the first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

[0032] In combination with the first aspect and the first to third implementations of the first aspect, in the seventh implementation of the first aspect of the present disclosure, the multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

[0033] In combination with the first aspect and the first to third implementations of the first aspect, in an eighth implementation of the first aspect of the present disclosure, the human skeleton includes a radius.

[0034] In a second aspect, an embodiment of the present disclosure provides a method for acquiring bone mass information.

[0035] Specifically, the bone mass information acquisition method includes:

[0036] Acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound;

[0037] Based on the second sample data, second bone mass information is obtained through a multi-channel convolutional neural network model.

[0038] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0039] In combination with the second aspect, in a first implementation of the second aspect of the present disclosure, the human skeleton includes a radius.

[0040] In combination with the second aspect, in a second implementation of the second aspect of the present disclosure, the multi-channel convolutional neural network model is trained using a method as described in any one of the first aspect and the first to eighth implementations of the first aspect.

[0041] In a third aspect, a model training device is provided in an embodiment of the present disclosure.

[0042] Specifically, the model training device comprises:

[0043] A first acquisition module is configured to acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound;

[0044] A second acquisition module is configured to acquire a bone mass label of a collection object corresponding to the first multi-channel time series data;

[0045] A training module is configured to train a multi-channel convolutional neural network model based on the first sample data and the bone mass label,

[0046] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0047] In combination with the third aspect, in a first implementation of the third aspect of the present disclosure, obtaining the first sample data includes any one of the following:

[0048] Controlling a transmitting transducer to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal;

[0049] Control two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and a receiving transducer to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determine the first multi-channel time series data based on the received second ultrasonic signal; or

[0050] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining multi-channel time series data based on the received second ultrasonic signal,

[0051] Optionally, the first ultrasonic signal or the second ultrasonic signal is an ultrasonic radio frequency signal.

[0052] In combination with the first implementation method of the third aspect, in the second implementation method of the third aspect of the present disclosure, the obtaining of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signal, and the multi-channel convolutional neural network includes a convolutional neural network with four convolution paths.

[0053] In conjunction with the third aspect, in a third implementation of the third aspect of the present disclosure:

[0054] Each of the convolution paths includes 8 convolution layers;

[0055] Each of the convolutional layers is provided with 8 convolutional kernels with a length of 5;

[0056] Each of the convolutional layers uses a rectified linear unit as an activation function; and / or

[0057] A pooling layer with a length of 2 is set after each convolutional layer, and the pooling layer adopts a maximum pooling layer.

[0058] In combination with the third aspect and the first to third implementations of the third aspect, in a fourth implementation of the third aspect of the present disclosure, the multi-channel convolutional neural network model further includes a fully connected layer, wherein:

[0059] The fully connected layer is connected to the ends of the fused multiple convolution paths;

[0060] The fully connected layer has 700-750 neurons; and / or

[0061] The fully connected layer uses a rectified linear unit as an activation function.

[0062] In combination with the fourth implementation of the third aspect, in a fifth implementation of the third aspect of the present disclosure, the multi-channel convolutional neural network model further includes an output layer, wherein:

[0063] The output layer is connected to the fully connected layer;

[0064] The output layer includes a flexible maximum function for outputting first bone mass information; and / or the bone mass label includes normal bone mass, osteopenia and osteoporosis.

[0065] In combination with the third aspect and the first to third implementations of the third aspect, in a sixth implementation of the third aspect of the present disclosure, the first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after the following preprocessing:

[0066] The first multi-channel time series data is processed using an eighth-order low-pass filter; and / or the first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

[0067] In combination with the third aspect and the first to third implementations of the third aspect, in the seventh implementation of the third aspect of the present disclosure, the multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

[0068] In combination with the third aspect and the first to third implementations of the third aspect, in an eighth implementation of the third aspect of the present disclosure, the human skeleton includes a radius.

[0069] In a fourth aspect, an embodiment of the present disclosure provides a device for acquiring bone mass information.

[0070] Specifically, the bone mass information acquisition device comprises:

[0071] A third acquisition module is configured to acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound;

[0072] a fourth acquisition module, configured to acquire second bone mass information through a multi-channel convolutional neural network model based on the second sample data,

[0073] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0074] In combination with the fourth aspect, in a first implementation of the fourth aspect of the present disclosure, the human skeleton includes a radius.

[0075] In combination with the fourth aspect, in a second implementation of the fourth aspect of the present disclosure, the multi-channel convolutional neural network model is trained using a method as described in any one of the first aspect and the first to eighth implementations of the first aspect.

[0076] In a fifth aspect, an embodiment of the present disclosure provides an electronic device, comprising a memory and a processor, wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement a method as described in any one of the first aspect, the first to eighth implementations of the first aspect, the second aspect, the first or second implementations of the second aspect.

[0077] In the sixth aspect, a computer-readable storage medium is provided in an embodiment of the present disclosure, on which computer instructions are stored. When the computer instructions are executed by a processor, a method as described in any one of the first aspect, the first to eighth implementations of the first aspect, the second aspect, the first or second implementation of the second aspect is implemented.

[0078] In a seventh aspect, a bone mass detection system is provided in an embodiment of the present disclosure.

[0079] Specifically, the bone mass detection system comprises:

[0080] A wearable component, used for collecting multi-channel time series data of human bones through ultrasound, and sending the multi-channel time series data to a data processing device;

[0081] A data processing device, comprising a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement a method as described in any one of the first aspect, the first to eighth implementations of the first aspect, the second aspect, or the first or second implementation of the second aspect.

[0082] In combination with the seventh aspect, in a first implementation of the seventh aspect of the present disclosure, the human skeleton includes a radius.

[0083] In combination with the seventh aspect, in a second implementation method of the seventh aspect of the present disclosure, the wearable component includes a fixed structure and a detection module including multiple ultrasonic transducers, wherein the fixed structure includes a receiving slot, the detection module is detachably installed in the receiving slot, and the ultrasonic transducer is used to generate an ultrasonic radio frequency signal.

[0084] In combination with the second implementation manner of the seventh aspect, in a third implementation manner of the seventh aspect of the present disclosure, the detection module includes four ultrasonic transducers arranged in a straight line with intervals of 10 mm, 15 mm, and 10 mm, respectively.

[0085] In combination with the second or third implementation of the seventh aspect, in a fourth implementation of the seventh aspect of the present disclosure, the multiple ultrasonic transducers are arranged in the detection module at an inclination angle of 35° to 45°.

[0086] In combination with the second implementation manner of the seventh aspect, in a fifth implementation manner of the seventh aspect of the present disclosure, the ultrasonic transducer includes a piezoelectric array element module, and the piezoelectric array element module is manufactured by the following method:

[0087] sputtering electrodes on the upper and lower surfaces of the PZT 1-3 composite material;

[0088] A matching layer is provided on one surface of the PZT 1-3 composite material having the sputtered electrode;

[0089] The processed PZT 1-3 composite material is segmented into array elements and divided into electrodes.

[0090] In combination with the fifth implementation of the seventh aspect, in a sixth implementation of the seventh aspect of the present disclosure, the PZT 1-3 composite material is prepared by the following method:

[0091] Cut PZT piezoelectric ceramics horizontally and vertically;

[0092] Fill the cut with epoxy resin;

[0093] After the epoxy resin is cured, double-sided grinding is performed to a target thickness.

[0094] Among them, the cutting spacing is 0.4mm, the cutting width is 0.1142mm, and the target thickness is 1.6mm.

[0095] According to the technical solution provided by the embodiment of the present disclosure, by acquiring first sample data, the first sample data includes first multi-channel time series data collected from human bones by ultrasound; acquiring a bone mass label of the collection object corresponding to the first multi-channel time series data; and training a multi-channel convolutional neural network model based on the first sample data and the bone mass label, wherein the multi-channel convolutional neural network model includes multiple convolution paths, and time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model, so that a model for obtaining bone mass information can be trained, and the model has good recognition ability in obtaining bone mass information.

[0096] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Other features, objectives and advantages of the present disclosure will become more apparent through the following detailed description of non-limiting embodiments in conjunction with the accompanying drawings. In the accompanying drawings:

[0098] Figure 1 A flow chart showing a model training method according to an embodiment of the present disclosure;

[0099] Figure 2 A schematic diagram showing the acquisition of ultrasonic radio frequency signals according to an embodiment of the present disclosure;

[0100] Figure 3 and Figure 4 A schematic diagram showing a waveform drawn based on time series data according to an embodiment of the present disclosure;

[0101] Figure 5 A schematic diagram showing the structure of a multi-channel convolutional neural network model according to an embodiment of the present disclosure is shown;

[0102] Figure 6 A block diagram of a model training device according to an embodiment of the present disclosure is shown;

[0103] Figure 7 A block diagram showing a device for acquiring bone mass information according to an embodiment of the present disclosure;

[0104] Figure 8 A block diagram of an electronic device according to an embodiment of the present disclosure is shown;

[0105] Fig. 9 A schematic diagram showing the structure of a computer system suitable for implementing the method and apparatus according to an embodiment of the present disclosure;

[0106] Figure 10 to Figure 12 A schematic diagram showing a wearable component according to an embodiment of the present disclosure;

[0107] Fig.13 A schematic diagram showing the arrangement of an ultrasonic transducer in a wearable component according to an embodiment of the present disclosure; and

[0108] Fig.14 and Fig.15 A top view and a side view of a PZT 1-3 composite material according to an embodiment of the present disclosure are shown, respectively. DETAILED DESCRIPTION

[0109] Hereinafter, exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily implement them. In addition, for the sake of clarity, parts not related to the description of the exemplary embodiments are omitted in the accompanying drawings.

[0110] In the present disclosure, it should be understood that terms such as "include" or "have" are intended to indicate the presence of features, numbers, steps, behaviors, components, parts, or a combination thereof disclosed in the present specification, and are not intended to exclude the possibility that one or more other features, numbers, steps, behaviors, components, parts, or a combination thereof exist or are added.

[0111] It should also be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0112] The inventors learned that the current SOS-based QUS method for diagnosing osteoporosis has an overall accuracy of 0.435, a macro precision of 0.435, a macro recall of 0.438, a macro F1 score of 0.436, and a micro F1 score of 0.435. The accuracy is low and can only be used for preliminary screening of osteoporosis in women.

[0113] Ultrasonic radio frequency signals are the standard form of unprocessed raw data in ultrasound imaging. The application value of RF signals in the field of orthopedics needs to be further explored to improve detection efficiency.

[0114] As a popular branch in the field of artificial intelligence in recent years, deep learning has a powerful feature extraction capability and can extract recognizable features from data, analyze and utilize them. Inspired by the excellent results of deep learning in one-dimensional data fields such as electrocardiogram, electroencephalogram, and speech processing, this application proposes to use a multi-channel convolutional neural network (MCNN) to analyze the RF signal source data obtained by the QUS device and classify the bone mass of the subjects.

[0115] Unlike the traditional convolutional neural network (CNN) model based on multi-channel data jointly inputting stacked convolution blocks, although the traditional CNN can extract features from data and obtain feature maps or feature vectors of different channel data, CNN shares weights for different channels and cannot capture the differences between different channel data and the correlation between them well. In view of the above shortcomings, the inventors proposed an MCNN that incorporates a multi-channel mechanism and adopted a feature-level fusion strategy, that is, after inputting different channel signals into the convolution layer to extract features, the feature vectors output by each channel are spliced ​​and then input into the fully connected layer for classification processing. The model has good recognition ability in obtaining bone mass information.

[0116] Figure 1 FIG. 1 is a flow chart showing a model training method according to an embodiment of the present disclosure. Figure 1As shown, the model training method includes the following steps S110 to S130:

[0117] In step S110, first sample data is acquired, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound;

[0118] In step S120, a bone mass label of the acquisition object corresponding to the first multi-channel time series data is obtained;

[0119] In step S130, a multi-channel convolutional neural network model is trained based on the first sample data and the bone mass label.

[0120] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0121] According to the technical solution provided by the embodiment of the present disclosure, by acquiring first sample data, the first sample data includes first multi-channel time series data collected from human bones by ultrasound; obtaining a bone mass label of the collection object corresponding to the first multi-channel time series data; training a multi-channel convolutional neural network model based on the first sample data and the bone mass label, wherein the multi-channel convolutional neural network model includes multiple convolution paths, and time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model, so that a model for obtaining bone mass information can be trained, and the model has good recognition ability in obtaining bone mass information.

[0122] According to the embodiment of the present disclosure, the ultrasonic wave in the form of pulses emitted by the transmitting transducer propagates in human tissue and generates side waves and reflected waves, which are received by the receiving transducer and converted into electrical signals for output. This signal is an ultrasonic radio frequency signal. The ultrasonic radio frequency signal contains all amplitude, frequency and phase information, that is, a large amount of information on the interaction between the sound field and tissue and microstructural characteristics. The ultrasonic radio frequency signal is expressed as time series data, that is, a sequence data including the real-time signal values ​​received at each moment, for example, {128, 132, 140, 148, 157, 165, ...}.

[0123] According to an embodiment of the present disclosure, the human skeleton may include, for example, a radius, a tibia or a root bone, and preferably, the radius may be selected. The ultrasonic radio frequency signal may be collected at the distal 1 / 3 of the radius on the non-dominant side of the human body. The radius is an easily exposed part, and is more sensitive to bone metabolism, and is also one of the important osteoporotic fracture sites. It has the basic conditions for promoting it on a large scale as a diagnostic site. The distal 1 / 3 of the radius is used to assess the risk of brittle fractures, especially hip fractures, with good results. Since the non-dominant side is less used, it is less affected by other factors, which is conducive to reflecting the actual situation of the human skeleton as a whole. Of course, in the case of fractures and other conditions on the non-dominant side, the radius on the dominant side can also be used for measurement.

[0124] According to an embodiment of the present disclosure, obtaining the first sample data includes any one of the following:

[0125] Control a transmitting transducer to transmit a first ultrasonic signal to a human skeleton, and two or more receiving transducers to receive a second ultrasonic signal from the human skeleton, and determine the first multi-channel time series data based on the received second ultrasonic signal. For example, the first ultrasonic radio frequency signal transmitted by the transmitting transducer A1 generates a second ultrasonic signal after propagating through the human skeleton, and is received by the receiving transducers B1, B2, and B3 respectively, to generate three signals, i.e., three-channel time series data;

[0126] Control two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and a receiving transducer to receive a second ultrasonic signal from the human skeleton, and determine the first multi-channel time series data based on the received second ultrasonic signal. For example, the first ultrasonic radio frequency signal transmitted by the transmitting transducers A1, A2, and A3 generates a second ultrasonic signal after propagating through the human skeleton, and is received by the receiving transducer B1 respectively, to generate three-channel time series data; or

[0127] Control two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two or more receiving transducers to receive a second ultrasonic signal from the human skeleton, and determine multi-channel time series data based on the received second ultrasonic signals. For example, there are transmitting transducers A1, A2, A3 and receiving transducers B1, B2, and six-channel time series data of A1-B1, A1-B2, A2-B1, A2-B2, A3-B1, and A3-B2 can be generated.

[0128] Through multi-channel time series data, richer bone conditions can be obtained, which is conducive to improving bone mass identification capabilities.

[0129] Alternatively, if Figure 2As shown, two transmitting transducers and two receiving transducers can be set on the detection device. The acquisition of the first sample data includes controlling the two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and the two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signal, and the multi-channel convolutional neural network includes a convolutional neural network with four convolution paths. Table 1 schematically shows the four-channel time series data in a unit time.

[0130] Table 1 Four-channel time series data

[0131]

[0132] According to an embodiment of the present disclosure, the first multi-channel time series data includes data collected within a time length of 5 to 50 seconds. Optionally, the collection time may be 30 to 40 seconds, for example 38 seconds.

[0133] According to the embodiment of the present disclosure, 200 unit time ultrasonic radio frequency data can be collected from a subject, and each unit time can include 1024 time instants of data. Figure 3 The waveform formed by the time series data of a channel of 200 time units totaling 204800 moments collected in about 38 seconds is schematically shown. In order to clearly show its detailed information, an exemplary waveform of one time unit is taken as Figure 4 It should be understood that Figure 3 and Figure 4 This is just to show the time series data more intuitively. In actual applications, the neural network model is not input in the form of images, but still in the form of numerical sequences.

[0134] According to an embodiment of the present disclosure, the bone mass label may include, for example, normal bone mass, osteopenia and osteoporosis, and the bone mass label of the first sample data for training may be obtained by a DXA method. The label of osteopenia is helpful for indicating the risk of bone mass. Of course, normal bone mass and osteopenia may also be combined into non-osteoporosis and processed as one category, and the combined label may directly distinguish between osteoporosis and non-osteoporosis.

[0135] According to the technical solution provided by the embodiment of the present disclosure, a model for obtaining bone mass information can be trained through bone mass labels including normal bone mass, decreased bone mass or osteoporosis, and the model has good recognition ability in obtaining bone mass information.

[0136] According to an embodiment of the present disclosure, the first multi-channel time series data is used for training after being preprocessed as follows:

[0137] The first multi-channel time series data is processed using an eighth-order low-pass filter; and / or the first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

[0138] According to the embodiment of the present disclosure, since the results of spectrum analysis of the data show that most of the signal information is concentrated in the low frequency band, a low-pass filter is used to minimize the impact of noise on the experiment. The embodiment of the present disclosure uses an eighth-order low-pass filter to process the RF signal sequence data.

[0139] According to the embodiment of the present disclosure, the data can be scaled to a distribution with a mean of 0 and a variance of 1 to improve the operational efficiency of the model.

[0140] Figure 5 A schematic diagram of the structure of a multi-channel convolutional neural network model according to an embodiment of the present disclosure is shown.

[0141] like Figure 5 As shown, multiple convolution paths in the multi-channel convolutional neural network model are used to correspondingly receive time series data of multiple channels. Four channels are shown here, but the embodiments of the present disclosure are not limited to this.

[0142] According to the embodiment of the present disclosure, the structure of each convolution path is similar or identical. For example, each convolution path may include 8 convolution layers; each convolution layer may be provided with 8 convolution kernels of length 5. The convolution operation of the convolution layer can realize feature extraction and obtain a feature vector.

[0143] For the data of the kth convolution path, the feature vector y output by the lth convolution layer l (m,j=k), which can be expressed as:

[0144]

[0145] Among them, I represents the total length of the data time dimension, represents input, w l and b l They represent the weight parameters and bias values ​​of the convolutional layer respectively.

[0146] According to the disclosed embodiment, each convolutional layer can use a rectified linear unit (Rectified Linear Unit, ReLU) as an activation function to form an activation layer. The ReLU activation function determines whether the output of each neuron reaches a threshold, that is, whether the feature strength of a certain part of the data reaches a certain standard. If it does not reach it, it is set to 0, indicating that the features extracted from this part of the data have no obvious effect on classification, so these features are not output to improve the sparsity of the network, reduce the mutual dependence between parameters, and alleviate the occurrence of overfitting to a certain extent; and ReLU can reduce the amount of computation of the network.

[0147] According to the embodiment of the present disclosure, a pooling layer with a length of 2 can be set after each convolution layer, and the pooling layer can adopt a maximum pooling layer, for example. The pooling layer can implement a downsampling operation and compress the input feature vector, which can shorten the feature vector and simplify the model calculation complexity on the one hand; on the other hand, it can compress the features and extract the main features.

[0148] According to the disclosed embodiment, the fully connected layer is connected to the end of the fused multiple convolution paths. At the end of the multiple convolution paths, such as the pooling layer of the last convolution connection, the results output by different convolution paths are spliced ​​and input to the fully connected layer, and an output layer is set after the fully connected layer.

[0149] According to an embodiment of the present disclosure, the fully connected layer may have 700-750 neurons. For example, the number of neurons in the fully connected layer may be set to 732. The fully connected layer may use a rectified linear unit as an activation function.

[0150] After extracting the features, the feature vector Z of each convolution path j Splicing along the channel axis and inputting a fully connected layer with 732 neurons and ReLU activation function can be expressed as:

[0151]

[0152] Z j ={y 8 (1,j),y 8 (2,j),…,y 8 (m,j),…,y 8 (M,j)} #(3)

[0153] M stands for Z j The length, f n represents the output of the nth neuron, w and b n Represent the weight parameters and bias values ​​of the fully connected layer respectively.

[0154] According to an embodiment of the present disclosure, the output layer may include a flexible maximum function (softmax) for outputting the first bone mass information; the output layer may have three neurons, corresponding to the bone mass labels of normal bone mass, bone loss and osteoporosis, respectively.

[0155] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model uses categorical cross-entropy as a loss function for model training. The categorical cross-entropy is shown in the following formula:

[0156]

[0157] in, represents the expected probability, p o represents the predicted probability of the model.

[0158] The model parameter optimization process includes feedforward propagation, backpropagation, and gradient update. First, the feature vector is calculated through feedforward propagation until the output is obtained. Secondly, when obtaining the output prediction probability, the error is calculated using the loss function, and then the chain rule of derivatives is used to realize the backpropagation of the error and calculate the derivatives of each model parameter. Finally, the stochastic gradient descent optimizer is used to update the derivatives of the model parameters in the backpropagation to minimize the loss function and obtain the best model performance.

[0159] The inventors have found through a large number of experiments that better classification effects can be achieved through the configuration of the above neural network structure and parameters.

[0160] The training and testing experiments of the disclosed embodiments used 70 cases of RF signal source data collected by the QUS device. This data contains 23 osteoporosis patients, 24 osteopenia patients and 23 healthy controls. All subjects have one or more osteoporosis risk factors after DXA examination. The ultrasonic bone densitometer used for data collection can collect RF signals of four channels at the same time. In this experiment, we took 200 consecutive time units of data from each subject, and the single-channel data collected in each time unit was 1024. The recorded data is the real-time signal value size, which is in the range of 0-255. Due to equipment reasons, the first bit of the single-channel data collected in each time unit is abnormal, so it is deleted by preprocessing.

[0161] According to the embodiment of the present disclosure, the data after various preprocessing described above can be divided into a training set, a validation set and a test set, and the ratio can be, for example, 5:1:1, wherein the training set is input into the model for training. The validation set is used to guide the training process, and the test set is used to evaluate the performance of the model. The final results are shown in Table 2.

[0162] Table 2

[0163]

[0164] As can be seen from Table 2, the radial QUS-RF signal and MCNN model of the embodiment of the present disclosure are superior to the radial QUS-SOS method in terms of the determination efficiency for osteoporosis.

[0165] Analyzing the reasons, the current QUS equipment can only use the ultrasonic RF signal to calculate the SOS value as the only characteristic value, resulting in poor judgment efficiency. The deep learning method based on MCNN proposed in the disclosed embodiment uses its feature extraction and learning capabilities to extract a large number of characteristic values ​​with biological information from the RF signal, resulting in better classification performance. Since a large number of characteristic values ​​of the ultrasonic RF signal are directly extracted for analysis, the results are less affected by other factors. This can also minimize the impact of operational factors such as the operator, the collection site, and the relative position of the bone and the ultrasonic probe on the judgment of osteoporosis. Furthermore, since good performance can be obtained by only collecting data from a single point on the radius, this technology can be combined with portable wearable devices and promoted to the masses to better achieve targeted monitoring and prevention of people with bone loss and osteoporosis.

[0166] In addition, the current QUS equipment parameter calculation methods are all based on fixed physical models. Different body shapes of subjects and different operating methods of examiners will lead to large differences in results. The embodiment of the present disclosure uses a deep learning method, and as more data is collected, the diagnostic model can be continuously optimized through semi-supervised learning.

[0167] The present disclosure also provides a method for obtaining bone mass information, including:

[0168] Acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound;

[0169] Based on the second sample data, second bone mass information is obtained through a multi-channel convolutional neural network model.

[0170] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0171] According to the embodiment of the present disclosure, the multi-channel convolutional neural network model can be Figure 1 to Figure 5 The model training method described is used for training.

[0172] According to an embodiment of the present disclosure, the second sample data and the first sample data are collected in the same manner and have the same data format, except that in the bone mass information acquisition method, the bone mass label is unknown, and the bone mass label is output through a trained multi-channel convolutional neural network model.

[0173] According to an embodiment of the present disclosure, the method may also include determining the confidence of the bone mass label output by the multi-channel convolutional neural network model, and when the confidence is higher than a specific threshold, further training the multi-channel convolutional neural network model using the second sample data and the output bone mass label, so that the multi-channel neural network model can be continuously optimized in the actual application process through a semi-supervised learning method.

[0174] Figure 6 A block diagram of a model training device according to an embodiment of the present disclosure is shown. The device can be implemented as part or all of an electronic device through software, hardware, or a combination of both.

[0175] like Figure 6 As shown, the model training device 600 includes a first acquisition module 610 , a second acquisition module 620 and a training module 630 .

[0176] A first acquisition module 610 is configured to acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound;

[0177] A second acquisition module 620 is configured to acquire a bone mass label of a collection object corresponding to the first multi-channel time series data;

[0178] The training module 630 is configured to train a multi-channel convolutional neural network model based on the first sample data and the bone mass label.

[0179] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0180] According to the technical solution provided by the embodiment of the present disclosure, a first acquisition module is configured to acquire first sample data, wherein the first sample data includes first multi-channel time series data acquired from human bones by ultrasound; a second acquisition module is configured to acquire a bone mass label of an acquisition object corresponding to the first multi-channel time series data; and a training module is configured to train a multi-channel convolutional neural network model based on the first sample data and the bone mass label, wherein the multi-channel convolutional neural network model includes multiple convolution paths, and time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model, so that a model for acquiring bone mass information can be trained, and the model for acquiring bone mass information has good sensitivity and specificity.

[0181] According to an embodiment of the present disclosure, obtaining the first sample data includes any one of the following:

[0182] Controlling a transmitting transducer to transmit a first ultrasonic signal to the human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal;

[0183] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and controlling a receiving transducer to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; or

[0184] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal,

[0185] Wherein, the first ultrasonic signal or the second ultrasonic signal is an ultrasonic radio frequency signal.

[0186] According to an embodiment of the present disclosure, the acquiring of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signals, wherein the multi-channel convolutional neural network includes a convolutional neural network having four channels.

[0187] According to an embodiment of the present disclosure, each of the convolution paths includes 8 convolution layers.

[0188] According to an embodiment of the present disclosure, each convolution layer is provided with 8 convolution kernels of length 5.

[0189] According to an embodiment of the present disclosure, each of the convolutional layers uses a rectified linear unit as an activation function.

[0190] According to an embodiment of the present disclosure, a pooling layer with a length of 2 is arranged after each convolution layer, and the pooling layer adopts a maximum pooling layer.

[0191] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model also includes a fully connected layer, wherein:

[0192] The fully connected layer is connected to the ends of the fused multiple convolution paths;

[0193] The fully connected layer has 700-750 neurons; and / or

[0194] The fully connected layer uses a rectified linear unit as an activation function.

[0195] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model also includes an output layer, wherein:

[0196] The output layer is connected to the fully connected layer;

[0197] The output layer includes a flexible maximum function for outputting first bone mass information; and / or the bone mass label includes normal bone mass, osteopenia and osteoporosis.

[0198] According to an embodiment of the present disclosure, the first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after being pre-processed as follows:

[0199] The first multi-channel time series data is processed using an eighth-order low-pass filter; and / or the first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

[0200] According to an embodiment of the present disclosure, the first multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

[0201] According to an embodiment of the present disclosure, the human skeleton includes a radius.

[0202] Figure 7 A block diagram of a bone mass information acquisition device according to an embodiment of the present disclosure is shown. The device can be implemented as part or all of an electronic device through software, hardware, or a combination of both.

[0203] like Figure 7 As shown, the bone mass information acquisition device 700 includes a third acquisition module 710 and a fourth acquisition module 720 .

[0204] The third acquisition module 710 is configured to acquire second sample data, where the second sample data includes second multi-channel time series data collected from human bones by ultrasound;

[0205] The fourth acquisition module 720 is configured to acquire second bone mass information through a multi-channel convolutional neural network model based on the second sample data.

[0206] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0207] According to an embodiment of the present disclosure, the human skeleton includes a radius, and the multi-channel convolutional neural network model can be Figure 1 to Figure 5 The model training method described is used for training.

[0208] The present disclosure also discloses an electronic device, Figure 8 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0209] like Figure 8 As shown, the electronic device 800 includes a memory 801 and a processor 802, wherein the memory 801 is used to store a program that supports the electronic device to execute the model training method or bone mass information acquisition method in any of the above embodiments, and the processor 802 is configured to execute the program stored in the memory 801.

[0210] According to an embodiment of the present disclosure, the memory 801 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 802 to implement the following steps:

[0211] Acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound;

[0212] Acquire a bone mass label of a collection object corresponding to the first multi-channel time series data;

[0213] Training a multi-channel convolutional neural network model based on the first sample data and the bone mass label,

[0214] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the first multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0215] According to an embodiment of the present disclosure, obtaining the first sample data includes any one of the following:

[0216] Controlling a transmitting transducer to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal;

[0217] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and controlling a receiving transducer to receive a second ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; or

[0218] Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and controlling two or more receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining multi-channel time series data based on the received second ultrasonic signals,

[0219] Wherein, the first ultrasonic signal or the second ultrasonic signal is an ultrasonic radio frequency signal.

[0220] According to an embodiment of the present disclosure, the acquisition of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signals. The multi-channel convolutional neural network includes a convolutional neural network with four convolution paths.

[0221] According to an embodiment of the present disclosure, each of the convolution paths includes 8 convolution layers.

[0222] According to an embodiment of the present disclosure, each convolution layer is provided with 8 convolution kernels of length 5.

[0223] According to an embodiment of the present disclosure, each of the convolutional layers uses a rectified linear unit as an activation function.

[0224] According to an embodiment of the present disclosure, a pooling layer with a length of 2 is arranged after each convolution layer, and the pooling layer adopts a maximum pooling layer.

[0225] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model further includes a fully connected layer,

[0226] in:

[0227] The fully connected layer is connected to the ends of the fused multiple convolution paths;

[0228] The fully connected layer has 700-750 neurons;

[0229] The fully connected layer uses a rectified linear unit as an activation function.

[0230] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model also includes an output layer, wherein:

[0231] The output layer is connected to the fully connected layer;

[0232] The output layer includes a flexibility maximum as an activation function, and is used to output first bone mass information;

[0233] The bone mass signature includes normal bone mass, osteopenia and osteoporosis.

[0234] According to an embodiment of the present disclosure, the first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after being pre-processed as follows:

[0235] The first multi-channel time series data is processed using an eighth-order low-pass filter; and / or the first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

[0236] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

[0237] According to an embodiment of the present disclosure, the human skeleton includes a radius.

[0238] According to an embodiment of the present disclosure, the memory 801 is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor 802 to implement the following steps:

[0239] Acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound;

[0240] Based on the second sample data, second bone mass information is obtained through a multi-channel convolutional neural network model.

[0241] Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

[0242] According to an embodiment of the present disclosure, the human skeleton includes a radius.

[0243] According to an embodiment of the present disclosure, the multi-channel convolutional neural network model may be a model trained by the method or device described above.

[0244] Fig. 9 A schematic diagram showing the structure of a computer system suitable for implementing the method and apparatus according to an embodiment of the present disclosure is shown.

[0245] like Fig. 9 As shown, the computer system 900 includes a processing unit 901, which can perform various processes in the above-mentioned embodiments according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage part 908 into a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the system 900 are also stored. The processing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0246] The following components are connected to the I / O interface 905: an input section 906 including a keyboard, a mouse, etc.; an output section 907 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, a modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the I / O interface 905 as needed. A removable medium 911, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 910 as needed, so that a computer program read therefrom is installed into the storage section 908 as needed. Among them, the processing unit 901 can be implemented as a processing unit such as a CPU, a GPU, a TPU, an FPGA, an NPU, etc.

[0247] In particular, according to an embodiment of the present disclosure, the method described above can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program tangibly contained on a readable medium thereof, and the computer program includes a program code for executing the above method. In such an embodiment, the computer program can be downloaded and installed from a network through the communication part 909, and / or installed from the removable medium 911.

[0248] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0249] The units or modules involved in the embodiments described in the present disclosure may be implemented by software or programmable hardware. The units or modules described may also be set in a processor, and the names of these units or modules do not constitute limitations on the units or modules themselves in some cases.

[0250] As another aspect, the present disclosure further provides a computer-readable storage medium, which may be a computer-readable storage medium included in the electronic device or computer system in the above embodiment; or a computer-readable storage medium that exists independently and is not assembled into a device. The computer-readable storage medium stores one or more programs, and the programs are used by one or more processors to execute the method described in the present disclosure.

[0251] The present disclosure also provides a bone mass detection system, including:

[0252] A wearable component, used for collecting multi-channel time series data of human bones through ultrasound, and sending the multi-channel time series data to a data processing device;

[0253] A data processing device comprises a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the model training method or bone mass information acquisition method as described above.

[0254] According to the embodiment of the present disclosure, the data processing device may refer to Figure 8 or Fig. 9 The introduction of will not be repeated here.

[0255] According to an embodiment of the present disclosure, the human skeleton includes a radius.

[0256] Figure 10 to Figure 12 A schematic diagram of a wearable component 1000 according to an embodiment of the present disclosure is shown.

[0257] like Figure 10 to Figure 12 As shown, the wearable component 1000 may include a fixed structure 1010 and a detection module 1020 including a plurality of ultrasonic transducers, wherein the fixed structure 1010 includes a receiving slot 1011, the detection module 1020 is detachably installed in the receiving slot, and the ultrasonic transducer is used to generate an ultrasonic radio frequency signal.

[0258] According to the embodiment of the present disclosure, the wearable component 1000 forms a bracelet when fixed to the arm, which fits tightly to the arm to achieve ultrasonic bone density measurement. If necessary, a coupling agent can be used when wearing it to improve the fit and enhance the detection effect.

[0259] According to an embodiment of the present disclosure, the detection module 1020 may include four ultrasonic transducers arranged along a straight line, and the spacings between the center positions of the ultrasonic transducers are d1, d2, and d3, respectively. For example, d1, d2, and d3 may be 10 mm, 15 mm, and 10 mm, respectively. Other sizes may also be used, and need to be specifically designed according to influencing factors such as distance depth during actual application. For example, it can be adjusted according to the size of the wearable device; multiple ultrasonic transducers are arranged in the detection module 1020 at an inclination angle θ, and the value of θ may be, for example, 35° to 45°, such as Fig.13 The inventor has determined through repeated experiments that the above optimal arrangement can reduce the situation where detection cannot be performed due to problems with bracelet wearing and improve the detection success rate, but it does not mean that other parameters cannot be implemented, and the parameters of the embodiments of the present disclosure are not limited to this.

[0260] According to an embodiment of the present disclosure, the ultrasonic transducer includes a piezoelectric array element module, and the piezoelectric array element module is manufactured by the following method:

[0261] sputtering electrodes on the upper and lower surfaces of the PZT 1-3 composite material;

[0262] A matching layer is provided on one surface of the PZT 1-3 composite material having the sputtered electrode;

[0263] The processed PZT 1-3 composite material is segmented into array elements and divided into electrodes.

[0264] According to an embodiment of the present disclosure, the PZT 1-3 composite material is prepared by the following method:

[0265] Cut PZT piezoelectric ceramics horizontally and vertically;

[0266] Fill the cut with epoxy resin;

[0267] After the epoxy resin is cured, double-sided grinding is performed to a target thickness.

[0268] Preferably, the cutting interval is 0.4 mm, the cutting width is 0.1142 mm, and the target thickness is 1.6 mm.

[0269] The final PZT 1-3 composite material is as follows Fig.14 and Fig.15 As shown, Fig.14 and Fig.15 A top view and a side view of a PZT 1-3 composite material according to an embodiment of the present disclosure are shown, respectively.

[0270] The electromechanical coupling coefficient of the PZT 1-3 composite material obtained according to the technical solution of the embodiment of the present disclosure reaches 0.69, which is significantly higher than that of the traditional lead zirconate titanate piezoelectric ceramics (i.e., PZT ceramics, whose electromechanical coupling coefficient is 0.51), and has a higher energy conversion efficiency. In addition, the acoustic impedance of the PZT 1-3 composite material is 15.23 MRayls, which is also lower than that of the traditional PZT ceramics (acoustic impedance 34.35 MRayls), and can better match human tissue. Therefore, while the bone density detection equipment is miniaturized, the bandwidth and sensitivity of the collected signal can be guaranteed.

[0271] Although the above embodiments describe a method of using PZT piezoelectric ceramics to make a composite material and further make a piezoelectric array element module, the embodiments of the present disclosure are not limited thereto, and other piezoelectric materials, such as PMN-PT piezoelectric single crystal, etc., may also be used.

[0272] According to the technical solution provided by the embodiment of the present disclosure, the wearable component forms a portable bone density detection device, which provides a strong guarantee for the early screening of osteoporosis and the establishment of a prevention and control system for osteoporosis with hierarchical diagnosis and treatment, two-way referral, acute and chronic treatment, and linkage between upper and lower levels. This will help medical personnel improve the diagnosis and treatment level of osteoporosis and the operating efficiency of the medical system, and improve the regional differences in medical resources.

[0273] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present disclosure is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the inventive concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other.

Claims

1. A model training method, characterized in that: include: Acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound; Acquire a bone mass label of a collection object corresponding to the first multi-channel time series data; Training a multi-channel convolutional neural network model based on the first sample data and the bone mass label, Wherein, the multi-channel convolutional neural network model includes multiple convolution paths, and time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model; The obtaining of the first sample data includes any one of the following: Controlling a transmitting transducer to transmit a first ultrasonic signal to a human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and controlling a receiving transducer to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; or Controlling two or more transmitting transducers to transmit a first ultrasonic signal to a human skeleton, and two or more receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; The first ultrasonic signal and the second ultrasonic signal are ultrasonic radio frequency signals.

2. The method according to claim 1, characterized in that The acquisition of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signal, wherein the multi-channel convolutional neural network includes a convolutional neural network with four convolution paths.

3. The method according to claim 1, wherein: Each of the convolution paths includes 8 convolution layers; Each of the convolutional layers is provided with 8 convolutional kernels with a length of 5; Each of the convolutional layers uses a rectified linear unit as an activation function; and / or A pooling layer with a length of 2 is set after each convolutional layer, and the pooling layer adopts a maximum pooling layer.

4. The method according to any one of claims 1 to 3, characterized in that The multi-channel convolutional neural network model also includes a fully connected layer, wherein: The fully connected layer is connected to the ends of the fused multiple convolution paths; The fully connected layer has 700-750 neurons; The fully connected layer uses a rectified linear unit as an activation function.

5. The method according to claim 4, characterized in that The multi-channel convolutional neural network model also includes an output layer, wherein: The output layer is connected to the fully connected layer; The output layer includes a flexible maximum function, which is used to output the first bone mass information; The bone mass signature includes normal bone mass, osteopenia and osteoporosis.

6. The method according to any one of claims 1 to 3, characterized in that: The first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after being pre-processed as follows: Processing the first multi-channel time series data using an eighth-order low-pass filter; and / or The first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

7. The method according to any one of claims 1 to 3, characterized in that The multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

8. The method according to any one of claims 1 to 3, characterized in that: The human skeleton includes a radius.

9. A method for obtaining bone mass information, characterized in that: include: Acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound; Acquiring second bone mass information through a multi-channel convolutional neural network model based on the second sample data, wherein the multi-channel convolutional neural network model is trained using the method described in any one of claims 1 to 8; Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

10. The method according to claim 9, characterized in that The human skeleton includes a radius.

11. A model training device, characterized in that: include: A first acquisition module is configured to acquire first sample data, wherein the first sample data includes first multi-channel time series data collected from human bones by ultrasound; A second acquisition module is configured to acquire a bone mass label of a collection object corresponding to the first multi-channel time series data; A training module is configured to train a multi-channel convolutional neural network model based on the first sample data and the bone mass label, Wherein, the multi-channel convolutional neural network model includes multiple convolution paths, and time series data of different channels in the first multi-channel time series data enter different convolution paths of the multi-channel convolutional neural network model; The obtaining of the first sample data includes any one of the following: Controlling a transmitting transducer to transmit a first ultrasonic signal to the human skeleton, controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; Controlling two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and controlling a receiving transducer to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; or Controlling two or more transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and controlling two or more receiving transducers to receive a second ultrasonic signal generated by the first ultrasonic signal from the human skeleton, and determining the first multi-channel time series data based on the received second ultrasonic signal; The first ultrasonic signal and the second ultrasonic signal are ultrasonic radio frequency signals.

12. The device according to claim 11, characterized in that The obtaining of the first sample data includes controlling two transmitting transducers to transmit a first ultrasonic signal to the human skeleton, and two receiving transducers to receive a second ultrasonic signal from the human skeleton, and determining four-channel time series data based on the received second ultrasonic signal, wherein the multi-channel convolutional neural network includes a convolutional neural network having four channels.

13. The device according to claim 11, wherein: Each of the convolution paths includes 8 convolution layers; Each of the convolutional layers is provided with 8 convolutional kernels with a length of 5; Each of the convolutional layers uses a rectified linear unit as an activation function; and / or A pooling layer with a length of 2 is set after each convolutional layer, and the pooling layer adopts a maximum pooling layer.

14. The device according to any one of claims 11 to 13, characterized in that: The multi-channel convolutional neural network model also includes a fully connected layer, wherein: The fully connected layer is connected to the ends of the fused multiple convolution paths; The fully connected layer has 700-750 neurons; and / or The fully connected layer uses a rectified linear unit as an activation function.

15. The device according to claim 14, characterized in that The multi-channel convolutional neural network model also includes an output layer, wherein: The output layer is connected to the fully connected layer; The output layer includes a flexible maximum function, which is used to output the first bone mass information; and / or The bone mass signature includes normal bone mass, osteopenia and osteoporosis.

16. The device according to any one of claims 11 to 13, characterized in that: The first multi-channel time series data includes data collected within a time length of 5 to 50 seconds, and the first multi-channel time series data is used for training after being pre-processed as follows: Processing the first multi-channel time series data using an eighth-order low-pass filter; and / or The first multi-channel time series data is scaled to a distribution with a mean of 0 and a variance of 1.

17. The device according to any one of claims 11 to 13, characterized in that: The multi-channel convolutional neural network model uses multi-classification cross entropy as a loss function for model training.

18. The device according to any one of claims 11 to 13, characterized in that: The human skeleton includes a radius.

19. A bone mass information acquisition device, characterized in that: include: A third acquisition module is configured to acquire second sample data, wherein the second sample data includes second multi-channel time series data collected from human bones by ultrasound; a fourth acquisition module, configured to acquire second bone mass information through a multi-channel convolutional neural network model based on the second sample data, wherein the multi-channel convolutional neural network model is trained using the method described in any one of claims 1 to 8; Among them, the multi-channel convolutional neural network model includes multiple convolution paths, and the time series data of different channels in the second multi-channel time series data enters different convolution paths of the multi-channel convolutional neural network model.

20. The device according to claim 19, characterized in that The human skeleton includes a radius.

21. An electronic device, characterized in that: The method comprises a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1 to 10.

22. A readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the method according to any one of claims 1 to 10 is implemented.

23. A bone mass detection system, comprising: A wearable component, used for collecting multi-channel time series data of human bones through ultrasound, and sending the multi-channel time series data to a data processing device; A data processing device, comprising a memory and a processor; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1-10.

24. The system according to claim 23, characterized in that The human skeleton includes a radius.

25. The system according to claim 24, characterized in that The wearable component includes a fixed structure and a detection module including a plurality of ultrasonic transducers, wherein the fixed structure includes a receiving slot, and the detection module is detachably installed in the receiving slot.

26. The system according to claim 25, characterized in that The detection module includes four ultrasonic transducers arranged along a straight line with intervals of 10 mm, 15 mm, and 10 mm respectively.

27. The system according to claim 25 or 26, characterized in that The multiple ultrasonic transducers are arranged in the detection module at an inclination angle of 35° to 45°.

28. The system according to claim 25, characterized in that The ultrasonic transducer includes a piezoelectric array element module, and the manufacturing method of the piezoelectric array element module includes: sputtering electrodes on the upper and lower surfaces of the PZT 1-3 composite material; A matching layer is provided on one surface of the PZT 1-3 composite material having the sputtered electrode; The processed PZT 1-3 composite material is segmented into array elements and divided into electrodes.

29. The system according to claim 28, characterized in that The method for preparing the PZT 1-3 composite material comprises: Cut PZT piezoelectric ceramics horizontally and vertically; Fill the cut with epoxy resin; After the epoxy resin is cured, double-sided grinding is performed to a target thickness. Among them, the cutting spacing is 0.4mm, the cutting width is 0.1142mm, and the target thickness is 1.6mm.

30. The system according to any one of claims 25 to 29, characterized in that: The ultrasonic transducer is used to generate an ultrasonic radio frequency signal.

Images