Bone mineral density prediction method and device based on ultrasonic radio frequency signals and artificial intelligence ultrasonic bone mineral density instrument

Abstract

The invention relates to the technical field of medical examination and inspection, in particular to a bone mineral density prediction method and device based on ultrasonic radio-frequency signals and an artificial intelligence ultrasonic bone mineral densitometer, the method comprises the steps that detection data of a detected user is obtained, and the detection data comprises ultrasonic radio-frequency signal data collected by quantitative ultrasonic equipment from the detected user; based on the detection data, a trained bone mineral density prediction model is used for conducting bone mineral density prediction on the detected user, and the bone mineral density prediction model is obtained by training multiple pieces of sample data collected from multiple subjects and labels of the multiple pieces of sample data. According to the method, rich amplitude, frequency and phase information in the ultrasonic radio-frequency signal is fully utilized, the information reflects interaction between the ultrasonic sound field and the human tissue and microstructure characteristics of the human tissue, and compared with an existing mode of deriving the bone mineral density based on acoustic parameters in the ultrasonic signal, the method has higher intelligence and accuracy.

Description

Technical Field

[0001] The present disclosure relates to the field of medical examination and testing technology, and in particular to a bone density prediction method and device based on ultrasonic radio frequency signals and an artificial intelligence ultrasonic bone densitometer. Background Art

[0002] Osteoporosis is a systemic bone disease characterized by low bone mass and damaged bone tissue microstructure, which leads to increased bone brittleness and susceptibility to fractures. Osteoporotic fractures, also known as brittle fractures, are the most serious consequence of osteoporosis. They refer to fractures that occur when a person is subjected to a force equivalent to a fall from a standing height or below. Osteoporotic fractures have a high incidence, high disability and mortality rates, and high medical costs, placing a heavy burden on patients, families, and society. Preventing or reducing the occurrence of fractures has always been the main goal of osteoporosis prevention and treatment. However, osteoporosis prevention and treatment in my country still faces the dilemma of "high incidence" and "low awareness, judgment, and treatment rates". The fundamental reason is the lack of economical, portable, and accurate osteoporosis screening equipment.

[0003] Currently, the commonly used methods for screening and judging osteoporosis include dual energy X-ray absorptiometry (DXA), peripheral dual energy X-ray absorptiometry (pDXA), quantitative computed tomography (QCT), peripheral quantitative computed tomography (pQCT) and quantitative ultrasound (QUS).

[0004] Dual-energy X-ray absorptiometry (DXA) is a recognized osteoporosis detection method both at home and abroad, and is also the "gold standard" recommended by the World Health Organization (WHO) for diagnosing osteoporosis. Its working principle is to obtain bone mineral density (BMD) and the corresponding T value by counting the X-ray uptake rate. However, due to the high cost, large size, radiation, the need for special personnel to operate DXA equipment, and the lack of such equipment in community-based hospitals (the number of DXA bone densitometers per million people is less than 0.35), it cannot meet the needs of large-scale osteoporosis screening. Quantitative computed tomography (QCT) uses the imaging principle of X-CT to achieve quantitative bone density imaging within a unit volume, but it has problems such as large size, high price, high radiation dose, and algorithm defects. It is not currently recommended as an osteoporosis detection device and is still in the stage of animal experiments and clinical research.

[0005] In addition, quantitative ultrasound QUS is a commonly used osteoporosis screening technology. Compared with DXA, QUS has the advantages of low cost, small size, portability, less time consumption, and no radiation, making it more suitable for a wide range of medical application scenarios and is expected to solve the problem of DXA shortage in primary medical institutions.

[0006] The basic principle of QUS in the prior art is to extract acoustic parameters from ultrasonic signals based on prior knowledge of acoustics, and deduce the bone density of the measured part through these parameters. Different QUS manufacturers use different acoustic parameters, and their parameter indicators include: ultrasonic speed of sound (SOS) of bones, broadband ultrasonic attenuation (BUA), and composite parameters combining BUA and SOS, such as stiffness index (SI) and quantitative ultrasonic index (QUI). However, these parameters are based on acoustic principles and are converted through a simplified "ideal solid model", which only reflects some characteristics of sound wave propagation in bones, ignoring or losing a lot of information on the interaction between the sound field and tissues, limiting the accuracy of their judgment. Summary of the invention

[0007] In order to solve the problems in the related art, the embodiments of the present disclosure provide a bone density prediction method, device and artificial intelligence ultrasonic bone densitometer based on ultrasonic radio frequency signals.

[0008] A first aspect of the present disclosure provides a bone density prediction method based on ultrasonic radio frequency signals, characterized by comprising:

[0009] Acquire detection data of the user under test, the detection data including ultrasonic radio frequency signal data collected from the user under test by a quantitative ultrasonic device, the quantitative ultrasonic device including an ultrasonic transmitter and an ultrasonic receiver, the ultrasonic radio frequency signal data including original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body;

[0010] Based on the detection data, the bone density of the tested user is predicted using a trained bone density prediction model, wherein:

[0011] The bone density prediction model is trained using a plurality of sample data collected from a plurality of subjects and respective labels of the plurality of sample data, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, and the respective labels of the plurality of sample data are determined based on the bone density data of the plurality of subjects;

[0012] The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0013] According to an embodiment of the present disclosure, the designated detection position includes the radius; the label includes the axial bone density value and / or the axial bone mass category; for any subject, the sample data and the bone density data are collected on the same day.

[0014] According to an embodiment of the present disclosure, the quantitative ultrasound device includes M ultrasonic generators and N ultrasonic receivers, and the ultrasonic radio frequency signal data includes ultrasonic radio frequency signal data of M*N transmit-receive channels formed by the M ultrasonic generators and the N ultrasonic receivers; the bone density prediction model includes M*N multi-scale attention convolution channels, which are used to process the ultrasonic radio frequency signal data of the M*N transmit-receive channels respectively.

[0015] According to an embodiment of the present disclosure, the output layer includes a fully connected layer.

[0016] According to an embodiment of the present disclosure, the bone density prediction result includes a bone density category and / or a bone density value; the fully connected layer includes a first fully connected layer and / or a second fully connected layer, the first fully connected layer is used to obtain a bone density category based on the fusion feature map, and the second fully connected layer is used to obtain a bone density value based on the fusion feature map.

[0017] According to an embodiment of the present disclosure, when the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: a first splicing module, used to splice the outputs of the multiple multi-scale attention convolution channels; a first global average pooling layer, used to perform global average pooling on the output of the first splicing module, wherein the fully connected layer obtains the bone density prediction result based on the output of the first global average pooling layer.

[0018] According to an embodiment of the present disclosure, when the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: a second splicing module, used to splice the clinical risk factor obtained based on the clinical risk factor data with the output of the first global average pooling layer, wherein the bone density prediction result obtained based on the output of the one or more multi-scale attention convolution channels includes obtaining the bone density prediction result based on the output of the second splicing module.

[0019] According to an embodiment of the present disclosure, the bone density prediction model also includes a third fully connected layer, which is used to obtain the clinical risk factor based on the clinical risk factor data; the clinical risk factor includes at least one or more of the following: gender, age, weight, height, and fracture history.

[0020] According to an embodiment of the present disclosure, the multi-scale attention convolution channel includes one multi-scale attention unit or multiple multi-scale attention units connected in series, and the multi-scale attention unit includes one or multiple multi-scale attention modules connected in series.

[0021] According to an embodiment of the present disclosure, when the multi-scale attention convolution channel includes a plurality of multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a pooling layer.

[0022] According to an embodiment of the present disclosure, the pooling layer includes any one of the following: a maximum pooling layer, an average pooling layer, a global pooling layer, an adaptive pooling layer, a global average pooling layer, and a random pooling layer.

[0023] According to an embodiment of the present disclosure, the multi-scale attention convolution channel includes three multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a maximum pooling layer, and the multi-scale attention unit includes two multi-scale attention modules connected in series.

[0024] According to an embodiment of the present disclosure, the multi-scale attention module includes a multi-scale cross-attention module, which includes m convolutional layers, m-1 cross-attention modules, and a third splicing module, m≥2, wherein: the m convolutional layers have different convolution kernel sizes, which are used to extract temporal features of different scales from the input of the multi-scale cross-attention module; the m-1 cross-attention modules are used to perform step-by-step fusion of the outputs of the m convolutional layers in the order of convolution kernels from small to large based on the cross-attention mechanism; the third splicing module is used to splice the output of the convolutional layer with the smallest convolution kernel with the output of the m-1 cross-attention modules.

[0025] According to an embodiment of the present disclosure, the input ends of the m convolutional layers are connected to the input ends of the multi-scale cross-attention module, the convolution kernel of the i-th convolutional layer is smaller than the convolution kernel of the i+1-th convolutional layer, 1≤i≤m-1; the first input end of the first cross-attention module is connected to the output end of the second convolutional layer, and the second input end of the first cross-attention module is connected to the output end of the first convolutional layer; the first input end of the j-th cross-attention module is connected to the output end of the j+1-th convolutional layer, and the second input end of the j-th cross-attention module is connected to the first output end of the j-1-th cross-attention module, wherein the first output end of the cross-attention module outputs the cross-attention fusion result of the first input data of the first input end of the cross-attention module and the second input data of the second input end, 2≤j≤m-1.

[0026] According to an embodiment of the present disclosure, the cross-attention module includes: a first projection module, used to perform linear projection on the first input data using 1*1 point-by-point convolution to obtain a first vector matrix Q; a second projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a second vector matrix K; a third projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a third vector matrix V; a first dot product module, used to calculate the dot product of the transpose of the first vector matrix Q and the second vector matrix K; a normalization module, used to normalize the output of the first dot product module; a second dot product module, used to calculate the dot product of the output of the normalization module and the third vector matrix V; a residual connection module, used to superimpose the first input data with the output of the second dot product module; a first output end, the first output end outputs the calculation result of the second dot product module; a second output end, the second output end outputs the calculation result of the residual connection module.

[0027] According to an embodiment of the present disclosure, the multi-scale cross-attention module also includes a batch normalization module and an activation function module which are sequentially connected in series to the output of the third splicing module.

[0028] According to an embodiment of the present disclosure, the multi-scale attention module also includes a compression-excitation attention module, which is used to scale the weights of features of different scales contained in the output of the multi-scale cross-attention module to remove redundant information.

[0029] According to an embodiment of the present disclosure, the compressed incentive attention module includes a second global average pooling layer, a fourth fully connected layer, a fifth fully connected layer, and a third dot multiplication module connected in series in sequence, and the third dot multiplication module is used to multiply the output of the fifth fully connected layer with the input of the compressed incentive attention module.

[0030] A second aspect of the present disclosure provides a bone density prediction model training method based on ultrasonic radio frequency signals, characterized by comprising:

[0031] Acquire a plurality of sample data collected from a plurality of subjects, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, wherein the quantitative ultrasound device includes an ultrasound transmitter and an ultrasound receiver, and the ultrasound radio frequency signal data includes original radio frequency signal data detected by the ultrasound receiver after the ultrasound signal emitted by the ultrasound transmitter passes through a designated detection position on the human body;

[0032] Based on the bone density data of the multiple subjects, determining labels of the multiple sample data respectively;

[0033] Using the plurality of sample data and the labels of the plurality of sample data, a bone density prediction model is trained, wherein:

[0034] The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, wherein the multi-scale attention convolution channels include: multiple convolution layers with different convolution kernel sizes, used to extract time features of multiple scales of the ultrasonic radio frequency signal data; a cross-attention module, used to weightedly fuse the time features of the multiple scales to obtain a fused feature map for bone density prediction, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0035] According to an embodiment of the present disclosure, the quantitative ultrasound device includes M ultrasonic generators and N ultrasonic receivers, and the ultrasonic radio frequency signal data includes ultrasonic radio frequency signal data of M*N transmit-receive channels formed by the M ultrasonic generators and the N ultrasonic receivers; the bone density prediction model includes M*N multi-scale attention convolution channels, which are used to process the ultrasonic radio frequency signal data of the M*N transmit-receive channels respectively.

[0036] According to an embodiment of the present disclosure, the multi-scale attention convolution channel includes one multi-scale attention unit or multiple multi-scale attention units connected in series, and the multi-scale attention unit includes one or multiple multi-scale attention modules connected in series.

[0037] According to an embodiment of the present disclosure, when the multi-scale attention convolution channel includes a plurality of multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a pooling layer.

[0038] According to an embodiment of the present disclosure, the multi-scale attention module includes a multi-scale cross-attention module and a compressed excitation attention module. The multi-scale cross-attention module includes m convolutional layers, m-1 cross-attention modules, and a third splicing module, where m≥2. It also includes a batch normalization module and an activation function module that are sequentially connected in series to the output of the third splicing module. The compressed excitation attention module is used to scale the weights of features of different scales contained in the output of the multi-scale cross-attention module to remove redundant information, wherein: the m convolutional layers have different convolution kernel sizes, which are used to extract temporal features of different scales from the input of the multi-scale cross-attention module; the m-1 cross-attention modules are used to perform step-by-step fusion of the outputs of the m convolutional layers in the order of convolution kernels from small to large based on the cross-attention mechanism; the third splicing module is used to splice the output of the convolutional layer with the smallest convolution kernel with the output of the m-1 cross-attention modules.

[0039] According to an embodiment of the present disclosure, the input ends of the m convolutional layers are connected to the input ends of the multi-scale cross-attention module, the convolution kernel of the i-th convolutional layer is smaller than the convolution kernel of the i+1-th convolutional layer, 1≤i≤m-1; the first input end of the first cross-attention module is connected to the output end of the second convolutional layer, and the second input end of the first cross-attention module is connected to the output end of the first convolutional layer; the first input end of the j-th cross-attention module is connected to the output end of the j+1-th convolutional layer, and the second input end of the j-th cross-attention module is connected to the first output end of the j-1-th cross-attention module, wherein the first output end of the cross-attention module outputs the cross-attention fusion result of the first input data of the first input end of the cross-attention module and the second input data of the second input end, 2≤j≤m-1; the compressed excitation attention module includes a second global average pooling layer, a fourth fully connected layer, a fifth fully connected layer, and a third dot multiplication module connected in series in sequence, and the third dot multiplication module is used to multiply the output of the fifth fully connected layer with the input of the compressed excitation attention module.

[0040] According to an embodiment of the present disclosure, the cross-attention module includes: a first projection module, used to perform linear projection on the first input data using 1*1 point-by-point convolution to obtain a first vector matrix Q; a second projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a second vector matrix K; a third projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a third vector matrix V; a first dot product module, used to calculate the dot product of the transpose of the first vector matrix Q and the second vector matrix K; a normalization module, used to normalize the output of the first dot product module; a second dot product module, used to calculate the dot product of the output of the normalization module and the third vector matrix V; a residual connection module, used to superimpose the first input data with the output of the second dot product module; a first output end, the first output end outputs the calculation result of the second dot product module; a second output end, the second output end outputs the calculation result of the residual connection module.

[0041] The third aspect of the present disclosure provides a bone density prediction device based on ultrasonic radio frequency signals, characterized in that it includes:

[0042] A first acquisition module is configured to acquire detection data of the user under test, wherein the detection data includes ultrasonic radio frequency signal data collected from the user under test by a quantitative ultrasonic device, wherein the quantitative ultrasonic device includes an ultrasonic transmitter and an ultrasonic receiver, and wherein the ultrasonic radio frequency signal data includes original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body;

[0043] The prediction module is configured to predict the bone density of the tested user using a trained bone density prediction model based on the detection data, wherein:

[0044] The bone density prediction model is trained using a plurality of sample data collected from a plurality of subjects and respective labels of the plurality of sample data, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, and the respective labels of the plurality of sample data are determined based on the bone density data of the plurality of subjects;

[0045] The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0046] A fourth aspect of the present disclosure provides a bone density prediction model training device based on ultrasonic radio frequency signals, characterized in that it includes:

[0047] a second acquisition module, configured to acquire a plurality of sample data collected from a plurality of subjects, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, wherein the quantitative ultrasound device includes an ultrasound transmitter and an ultrasound receiver, and the ultrasound radio frequency signal data includes original radio frequency signal data detected by the ultrasound receiver after the ultrasound signal emitted by the ultrasound transmitter passes through a designated detection position on the human body;

[0048] A determination module, configured to determine a label of each of the plurality of sample data based on the bone density data of the plurality of subjects;

[0049] The training module is configured to train a bone density prediction model using the plurality of sample data and the labels of the plurality of sample data, wherein:

[0050] The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0051] The fifth aspect of the present disclosure provides an electronic device, characterized in that it includes a processor and a memory, wherein the memory is used to store one or more computer instructions, and wherein the one or more computer instructions are executed by the processor to implement any of the method steps described above.

[0052] A sixth aspect of the present disclosure provides an artificial intelligence ultrasonic bone densitometer, characterized in that it includes:

[0053] straps;

[0054] One or more ultrasonic transmitters and one or more ultrasonic receivers are disposed on the strap;

[0055] A control host, the control host comprising a processor and a memory, 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 any of the method steps described above.

[0056] A seventh aspect of the present disclosure provides a computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions implement any of the method steps described above when executed by a processor.

[0057] An eighth aspect of the present disclosure provides a computer program product, which includes computer instructions, and is characterized in that when the computer instructions are executed by a processor, any of the method steps described above are implemented.

[0058] The technical solution provided according to the embodiment of the present disclosure makes full use of the rich amplitude, frequency and phase information in the ultrasonic radio frequency signal. This information reflects the interaction between the ultrasonic sound field and human tissue and the microstructural characteristics of human tissue. The bone density prediction model is obtained by training a deep neural network model including an attention mechanism. The bone density prediction model is used to predict bone density based on the ultrasonic radio frequency signal of the user being tested. The key characteristic variables related to bone density prediction in the ultrasonic radio frequency signal can be automatically extracted through the attention mechanism. Compared with the existing method of extracting acoustic parameters from the ultrasonic signal and deriving bone density based on the acoustic parameters, it has higher intelligence and accuracy. At the same time, compared with DXA, QCT and other detection methods, the quantitative ultrasonic detection method according to the embodiment of the present disclosure has the advantages of low cost, small size, portability, less time consumption, no radiation, etc., does not require a special inspection site to shield radiation, and the measurement is more convenient, which is more in line with the needs of large-scale screening, clinical monitoring, epidemiological surveys, etc.

[0059] 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

[0060] 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:

[0061] Figure 1 A flowchart of a method for predicting bone density based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0062] Figure 2 Ultrasonic radio frequency data of three consecutive cycles of four channels of the quantitative ultrasound device are shown.

[0063] Figure 3 A structural diagram of a bone density prediction model according to an embodiment of the present disclosure is shown.

[0064] Figure 4 A specific example of a multi-scale attention module according to an embodiment of the present disclosure is shown.

[0065] Figure 5A specific example of a cross-attention module according to an embodiment of the present disclosure is shown.

[0066] Figure 6 A flowchart of a bone density prediction model training method based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0067] Figure 7 A structural block diagram of a bone density prediction device based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0068] Figure 8 A structural block diagram of a bone density prediction model training device based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0069] Fig. 9 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0070] Fig.10 A schematic diagram showing the structure of a computer system suitable for implementing the method according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0071] 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.

[0072] 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.

[0073] 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.

[0074] In the present disclosure, if it involves operations of obtaining user information or user data or displaying user information or user data to others, the operations are all authorized and confirmed by the user, or actively selected by the user.

[0075] As mentioned above, the bone density detection technologies commonly used in the prior art, such as DXA and QCT, all have defects such as large size, high price, and large radiation dose, which are not conducive to large-scale promotion and use. The traditional QUS bone densitometer derives bone density based on acoustic parameters extracted from ultrasonic radio frequency signals, ignoring or losing a large amount of information on the interaction between the sound field and human tissue, limiting the accuracy of its judgment.

[0076] The present disclosure provides a bone density prediction method based on ultrasonic radio frequency signals, comprising: obtaining detection data of a user under test, wherein the detection data comprises ultrasonic radio frequency signal data collected from the user under test by a quantitative ultrasonic device, wherein the quantitative ultrasonic device comprises an ultrasonic transmitter and an ultrasonic receiver, and wherein the ultrasonic radio frequency signal data comprises original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body; based on the detection data, using a trained bone density prediction model to predict the bone density of the user under test, wherein: the bone density prediction model is trained using a plurality of sample data collected from a plurality of subjects and the labels of the plurality of sample data The sample data obtained includes ultrasonic radio frequency signal data collected by a quantitative ultrasound device from corresponding subjects, and the labels of the multiple sample data are determined based on the bone density data of the multiple subjects; the bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channel includes a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0077] The ultrasonic radio frequency signal data includes the original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through the designated detection position on the human body, for example, it can be the amplitude time series of the original radio frequency signal. The technical solution provided according to the embodiment of the present disclosure makes full use of the rich amplitude, frequency and phase information in the ultrasonic radio frequency signal, which reflects the interaction between the ultrasonic sound field and the human tissue and the microstructural characteristics of the human tissue. The bone density prediction model is obtained by training a deep neural network model including an attention mechanism, and the bone density prediction model is used to predict the bone density based on the ultrasonic radio frequency signal of the measured user. The key characteristic variables related to the bone density prediction in the ultrasonic radio frequency signal can be automatically extracted through the attention mechanism. Compared with the existing method of extracting acoustic parameters from the ultrasonic signal and deriving bone density based on the acoustic parameters, it has higher intelligence and accuracy. At the same time, compared with the detection methods such as DXA and QCT, the quantitative ultrasonic detection method according to the embodiment of the present disclosure has the advantages of low cost, small size, portability, less time consumption, no radiation, etc., does not require a special inspection site to shield radiation, and the measurement is more convenient, which is more in line with the needs of large-scale screening, clinical monitoring, epidemiological surveys, etc.

[0078] Figure 1 A flowchart of a method for predicting bone density based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0079] like Figure 1 As shown, the bone density prediction method based on ultrasonic radio frequency signals includes the following steps S101-S102:

[0080] In step S101, detection data of the measured user is obtained, the detection data including ultrasonic radio frequency signal data collected from the measured user by a quantitative ultrasonic device, the quantitative ultrasonic device including an ultrasonic transmitter and an ultrasonic receiver, the ultrasonic radio frequency signal data including original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body.

[0081] In step S102, based on the detection data, a trained bone density prediction model is used to predict the bone density of the tested user, wherein: the bone density prediction model is trained using multiple sample data collected from multiple subjects and respective labels of the multiple sample data, the sample data includes ultrasonic radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, and the respective labels of the multiple sample data are determined based on the bone density data of the multiple subjects; the bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channel includes a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract multiple scales of time features of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the multiple scales of time features to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0082] According to the self-developed prototype host of the embodiment of the present disclosure, a commercial portable ultrasonic bone densitometer host (BMD-9) is used, and a self-developed bracelet probe is used to detect the position of the radius. The bracelet probe includes a bracelet strap and a probe assembly, and the probe assembly has a modular design that can be detached from the strap. The probe assembly includes a "dual-transmit and dual-receive" piezoelectric ceramic high-performance ultrasonic transducer, which are transducers A, B, C, and D, respectively, of which A and B are used as transmitting transducers E1 and E2 (used as ultrasonic transmitters), respectively, and C and D are used as receiving transducers R1 and R2 (used as ultrasonic receivers), respectively. The angle of each array element is about 30°, forming an axial transmission multi-channel ultrasonic transducer combination. In a specific embodiment, the probe center frequency is 1.25MHz, the bandwidth is 41.27%, the probe sensitivity is -46.69dB, the maximum crosstalk is -52.60dB, the capacitance is about 520pF, and the sound velocity accuracy and repeatability are both less than 50m / s sound velocity deviation (PM: 0.1797%; CU: 0.108%).

[0083] The strap can be made of fine frosted soft rubber material, designed with buckles for adjusting different tightness and grooves that fit the radial side, which can keep the probe assembly close to the radius and as parallel as possible, ensuring stable and reliable signals during detection.

[0084] The portable ultrasound bone densitometer host (BMD-9) can display the SOS value and T value on the host display interface after matching, connecting and improving the bracelet probe, and can save and extract the original ultrasound radio frequency signal data detected by the probe assembly. According to the host system's own algorithm, if the subject's T value is less than or equal to -2.5, it is classified as "osteoporosis", if its T value is higher than -2.5 but lower than -1.0, it is classified as "osteopenia", and if its T value is higher than or equal to -1.0, it is "normal".

[0085] According to the embodiments of the present disclosure, the designated detection site is the distal third of the radius of the non-dominant hand (generally the left hand). If there is any trauma to the forearm or wrist joint on that side, the contralateral forearm is tested. During the measurement process, the user to be tested takes a comfortable seat and places their forearm naturally on the examination table. After applying a sufficient amount of coupling agent on the probe assembly, the operator assists the user to be tested to wear the bracelet probe on the designated detection site, start the measurement and keep it in this position until the measurement is completed.

[0086] The advantage of using the radius as the designated detection site in the embodiment of the present disclosure is that the user does not need to take off clothes or make special preparations, and the detection site only needs to be equipped with tables and chairs. However, those skilled in the art will understand that the designated detection site may also be other sites besides the radius.

[0087] The probe assembly of the disclosed embodiment includes a "dual-transmit and dual-receive" piezoelectric ceramic high-performance ultrasonic transducer, so it can collect ultrasonic radio frequency signals of four channels (E1 to R1, E1 to R2, E2 to R1, E2 to R2). Each acquisition takes about 40 seconds, and the sampling rate is 40M. If the process of one emission and reception is set as a single cycle, data of 1180 time points can be collected in a single cycle. The ultrasonic radio frequency data of each cycle records the time series of the amplitude of the ultrasonic wave under the four channels, and after normalization, the mean of the amplitude is 0 and the variance is 1.

[0088] As an example, Figure 2 The ultrasonic radio frequency data of three consecutive cycles of four channels of the quantitative ultrasonic device are shown. The data of 100 single cycles with stable and effective transmission signals are selected, that is, a total of 118,000 time points are recorded, and the data shape is (118000, 4).

[0089] According to an embodiment of the present disclosure, when training a bone density prediction model, ultrasonic radio frequency signals of multiple subjects are obtained as samples. For each subject, its bone density data is obtained and the label of the corresponding sample is determined based on the bone density data. For example, since the DXA method is the "gold standard" recommended by the World Health Organization WHO for diagnosing osteoporosis, the bone density data used to determine the label can be the axial bone density value obtained by the DXA device. The axial bone includes, for example, the lumbar spine, hip, femoral neck, etc. The label may include the axial bone density value and / or the axial bone mass category. According to an embodiment of the present disclosure, the T value of the subject can be calculated based on the axial bone density value. In this example, healthy adults aged 25-35 are used as the reference population, and the T value = (axial bone density value-mean bone density of the same-sex reference population) / standard deviation of bone density of the same-sex reference population. According to the comparison result of the T value and the threshold, the axial bone mass can be classified into three categories: normal bone mass, osteopenia, and osteoporosis. The sample data of each subject and the bone density data used to determine the label were collected on the same day to ensure a good correspondence between the two.

[0090] According to an embodiment of the present disclosure, before the ultrasound radio frequency data is input into the bone density prediction model, it is first preprocessed, including data splicing, data noise reduction and data normalization.

[0091] Data splicing: Based on the observation and analysis of ultrasound radio frequency signal data, it was found that the ultrasound radio frequency signal data of each subject contained tens of thousands of cycles of data, which was very large, and the difference between each cycle of data was small. In order to enable the bone density prediction model to more effectively learn the characteristics of the ultrasound radio frequency signal data of the subject and improve the efficiency of model training, a stable and effective transmission signal was intercepted from the ultrasound radio frequency signal data of each subject, and a specified number (for example, 100) of single-cycle data were selected and spliced ​​to obtain the sample of the subject.

[0092] Data denoising: The disclosed embodiment uses an 8th-order Butterworth low-pass filter with a cutoff frequency of 20 MHz to remove the high-frequency noise present in the samples after data splicing, thereby avoiding the influence of noise on the model.

[0093] Data normalization: The signal amplitude of ultrasonic radio frequency signal data varies between -40000 and 40000. The signal amplitude is too large and the magnitude difference between signals is large, which is not conducive to the convergence of the model and the learning of intrinsic features. Therefore, after data denoising, each sample is normalized to have a specified amplitude mean and equation, for example, the amplitude mean is 0 and the variance is 1. After unifying its numerical range, on the one hand, it helps to avoid gradient explosion in the model calculation process and facilitates the parameter update of the model. On the other hand, it can avoid the model from being overly sensitive to certain features with a large numerical range, which helps the model learn the intrinsic characteristics of the data, thereby improving the numerical stability of the model processing process and the effectiveness of model training.

[0094] After the above processing, each sample represents a subject.

[0095] Figure 3 A structural diagram of a bone density prediction model according to an embodiment of the present disclosure is shown.

[0096] like Figure 3 As shown, the bone density prediction model according to the embodiment of the present disclosure includes an output layer and one or more multi-scale attention convolution channels, and the multi-scale attention convolution channel includes: multiple convolution layers with different convolution kernel sizes, used to extract the time features of multiple scales of the ultrasonic radio frequency signal data; a cross-attention module, used to weightedly fuse the time features of the multiple scales to obtain a fused feature map for bone density prediction, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0097] According to an embodiment of the present disclosure, a quantitative ultrasound device for acquiring timed radio frequency signal data includes M ultrasound generators and N ultrasound receivers, and therefore, the ultrasound radio frequency signal data includes ultrasound radio frequency signal data of M*N transmit-receive channels formed by the M ultrasound generators and the N ultrasound receivers. Accordingly, the bone density prediction model includes M*N multi-scale attention convolution channels, which are respectively used to process the ultrasound radio frequency signal data of the M*N transmit-receive channels.

[0098] exist Figure 3 In the example of , the quantitative ultrasound device includes two ultrasound generators and two ultrasound receivers, and the ultrasound radio frequency signal data includes ultrasound radio frequency signal data of four transmit-receive channels formed by the two ultrasound generators and the two ultrasound receivers. Accordingly, Figure 3 The bone density prediction model shown includes 4 multi-scale attention convolution channels, which are used to process the ultrasonic radio frequency signal data of the 4 transmit-receive channels respectively.

[0099] According to an embodiment of the present disclosure, the quantitative ultrasound device may have more or fewer ultrasound transmitters and / or ultrasound receivers, for example, one ultrasound transmitter and one ultrasound receiver. In this case, the bone density prediction model includes one multi-scale attention convolution channel for processing the ultrasound radio frequency signal data of one transmit-receive channel.

[0100] According to an embodiment of the present disclosure, each multi-scale attention convolution channel processes the ultrasonic radio frequency signal data of the corresponding channel, extracts the time features of multiple scales of the ultrasonic radio frequency signal data through multiple convolution layers with different convolution kernel sizes, and performs weighted fusion of the time features of the multiple scales through a cross-attention module to obtain a fused feature map for bone density prediction.

[0101] According to an embodiment of the present disclosure, the output layer includes a fully connected layer, and the bone density prediction result includes a bone density category and / or a bone density value. For example, the fully connected layer may include a first fully connected layer and / or a second fully connected layer, wherein the first fully connected layer is used to obtain a bone density category based on the fused feature map, and the second fully connected layer is used to obtain a bone density value based on the fused feature map.

[0102] Specifically, Figure 3 As shown, the output layer may include a first fully connected layer FC1 and a second fully connected layer FC2, wherein the first fully connected layer is used to obtain a bone density category based on the fused feature map, and the second fully connected layer is used to obtain a bone density value based on the fused feature map. Those skilled in the art may understand that the bone density prediction model may also include only one of the first fully connected layer FC1 and the second fully connected layer FC2.

[0103] According to the embodiment of the present disclosure, in the model training, the bone density value prediction task adopts the mean square error loss function L pre :

[0104]

[0105] Among them, m i is the label bone density value of the i-th sample, b i is the bone density value predicted by the model, and n is the number of samples.

[0106] According to an embodiment of the present disclosure, in model training, the bone density classification task adopts the cross entropy loss function:

[0107]

[0108] in, is a symbolic function. If the label category of the i-th sample is category c, it takes 1, otherwise it takes 0. It represents the probability that the model predicts that the i-th sample is of category c, M represents the number of categories, and n is the number of samples.

[0109] According to an embodiment of the present disclosure, when the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: a first splicing module, used to splice the outputs of the multiple multi-scale attention convolution channels; a first global average pooling layer, used to perform global average pooling on the output of the first splicing module, wherein the fully connected layer obtains the bone density prediction result based on the output of the first global average pooling layer. According to an embodiment of the present disclosure, the output of the multi-scale attention convolution channel is the fused feature map of the output of the multi-scale attention convolution channel.

[0110] For example, Figure 3 As shown, the bone density prediction model includes 4 multi-scale attention convolution channels, and also includes a first splicing module C1 and a first global average pooling layer GAP1. The first splicing module C1 is used to splice the outputs of the 4 multi-scale attention convolution channels, so as to cascade the features of the 4 channels in the channel direction. The first global average pooling layer GAP1 is used to perform global average pooling on the output of the first splicing module C1 to reduce the number of parameters and the risk of overfitting. The fully connected layers FC1 and FC2 make predictions based on the output of the first global average pooling layer GAP1.

[0111] According to an embodiment of the present disclosure, when the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: a second splicing module, which is used to splice the clinical risk factor obtained based on the clinical risk factor data with the output of the first global average pooling layer, wherein the bone density prediction result obtained based on the output of the one or more multi-scale attention convolution channels includes the bone density prediction result obtained based on the output of the second splicing module. The bone density prediction model also includes a third fully connected layer, which is used to obtain the clinical risk factor based on the clinical risk factor data; the clinical risk factor includes at least one or more of the following: gender, age, weight, height, and fracture history.

[0112] For example, Figure 3 As shown, the bone density prediction model also includes a third fully connected layer FC3, which is used to obtain clinical risk factors based on clinical risk factors. The second splicing module C2 splices the clinical risk factor with the output of the first global average pooling layer GAP1, and the first fully connected layer FC1 and the second fully connected layer FC2 make predictions based on the second splicing module C2. By introducing the consideration of clinical risk factors in bone density prediction through the third fully connected layer and the second splicing module, the accuracy of the prediction results can be further improved.

[0113] According to an embodiment of the present disclosure, the multi-scale attention convolution channel includes one multi-scale attention unit or multiple multi-scale attention units connected in series. When the multi-scale attention convolution channel includes multiple multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a pooling layer. The pooling layer includes any one of the following: a maximum pooling layer, an average pooling layer, a global pooling layer, an adaptive pooling layer, a global average pooling layer, and a random pooling layer. For example, Figure 3 As shown in the figure, the multi-scale attention convolution channel includes three multi-scale attention units MU connected in series. Adjacent multi-scale attention units are connected through a pooling layer PL. The pooling layer PL can be a maximum pooling layer. The maximum pooling layer can reduce the amount of parameters on the one hand, and improve the model performance by selecting more recognizable features on the other hand.

[0114] When a multi-scale attention convolution channel includes one multi-scale attention unit, the input of the multi-scale attention convolution channel is input into the multi-scale attention unit. When a multi-scale attention convolution channel includes multiple multi-scale attention units connected in series, the input of the multi-scale attention convolution channel is input into the first multi-scale attention unit; the output of the first multi-scale attention unit is input into the second multi-scale attention unit, or is input into the second multi-scale attention unit after passing through the pooling layer, and so on.

[0115] According to an embodiment of the present disclosure, the multi-scale attention unit includes one or multiple multi-scale attention modules connected in series. When the multi-scale attention unit includes one multi-scale attention module, the input of the multi-scale attention unit is input into the multi-scale attention module. When the multi-scale attention unit includes multiple multi-scale attention modules connected in series, the input of the multi-scale attention unit is input into the first multi-scale attention module; the output of the first multi-scale attention module is input into the second multi-scale attention module, and so on.

[0116] According to an embodiment of the present disclosure, the multi-scale attention module includes a multi-scale cross-attention module, and the input of the multi-scale attention module is input into the multi-scale cross-attention module.

[0117] The multi-scale cross-attention module includes m convolutional layers, m-1 cross-attention modules, and a third splicing module, m≥2, wherein: the m convolutional layers have different convolution kernel sizes, and are used to extract temporal features of different scales from the input of the multi-scale cross-attention module; the m-1 cross-attention modules are used to perform step-by-step fusion of the outputs of the m convolutional layers in the order of convolution kernels from small to large based on the cross-attention mechanism; the third splicing module is used to splice the output of the convolutional layer with the smallest convolution kernel with the output of the m-1 cross-attention modules.

[0118] According to an embodiment of the present disclosure, the input ends of the m convolutional layers are connected to the input ends of the multi-scale cross-attention module, the convolution kernel of the i-th convolutional layer is smaller than the convolution kernel of the i+1-th convolutional layer, 1≤i≤m-1; the first input end of the first cross-attention module is connected to the output end of the second convolutional layer, and the second input end of the first cross-attention module is connected to the output end of the first convolutional layer; the first input end of the j-th cross-attention module is connected to the output end of the j+1-th convolutional layer, and the second input end of the j-th cross-attention module is connected to the first output end of the j-1-th cross-attention module, wherein the first output end of the cross-attention module outputs the cross-attention fusion result of the first input data of the first input end of the cross-attention module and the second input data of the second input end, 2≤j≤m-1.

[0119] According to an embodiment of the present disclosure, the cross-attention module includes: a first projection module, used to perform linear projection on the first input data using 1*1 point-by-point convolution to obtain a first vector matrix Q; a second projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a second vector matrix K; a third projection module, used to perform linear projection on the second input data using 1*1 point-by-point convolution to obtain a third vector matrix V; a first dot product module, used to calculate the dot product of the transpose of the first vector matrix Q and the second vector matrix K; a normalization module, used to normalize the output of the first dot product module; a second dot product module, used to calculate the dot product of the output of the normalization module and the third vector matrix V; a residual connection module, used to superimpose the first input data with the output of the second dot product module; a first output end, the first output end outputs the calculation result of the second dot product module; a second output end, the second output end outputs the calculation result of the residual connection module.

[0120] According to an embodiment of the present disclosure, the multi-scale cross-attention module also includes a batch normalization module and an activation function module which are sequentially connected in series to the output of the third splicing module.

[0121] According to an embodiment of the present disclosure, the multi-scale attention module also includes a compression-incentive attention module, which is used to scale the weights of features of different scales contained in the output of the multi-scale cross-attention module to remove redundant information. The compression-incentive attention module includes a second global average pooling layer, a fourth fully connected layer, a fifth fully connected layer, and a third dot product module connected in series in sequence, and the fourth dot product module is used to multiply the output of the fifth fully connected layer by the input of the compression-incentive attention module.

[0122] Reference below Figure 4 and Figure 5 The multi-scale attention module according to the embodiment of the present disclosure is described in detail.

[0123] Figure 4 A specific example of a multi-scale attention module according to an embodiment of the present disclosure is shown.

[0124] Figure 4 The multi-scale attention module MA shown includes a multi-scale cross-attention module MSCA and a squeeze-and-excitation (SE) attention module SE, and the input of the multi-scale attention module MA is input into the multi-scale cross-attention module MSCA.

[0125] The multi-scale cross attention module MSCA includes three convolutional layers CN1, CN2, CN3, two cross attention modules CA1, CA2, the third splicing module C3, the batch normalization module BN and the activation function module ReLU, X k is the input of the multi-scale cross-attention module MSCA, F′ m It is the output of the multi-scale cross attention module MSCA.

[0126] According to an embodiment of the present disclosure, the three parallel convolutional layers CN1, CN2, and CN3 of the multi-scale cross attention module MSCA have different convolution kernel sizes, generating receptive fields of different sizes to capture the time series X k The short-term, medium-term and long-term temporal dependencies of the image are obtained by multi-scale time information to extract features and generate feature maps. The advantage of multi-scale convolution over single-scale convolution is that it can produce receptive fields of different sizes, thereby capturing more information. The ultrasonic RF signal data input into the four multi-scale attention convolution channels is represented as a time series T k,L =(t k,1 ,t k,2 ,…,t k,L-1 ,t k,L ), where k represents the kth multi-scale attention convolution channel, and L represents the length of the time series. The ultrasonic RF signal data is input into the corresponding multi-scale attention convolution channel.

[0127] As an example, the convolution kernel sizes of the three convolutional layers CN1, CN2, and CN3 in the multi-scale cross-attention module MSCA are 3, 7, and 15 respectively, and the step size is 1. Their inputs are all time series X k , the output feature maps can be used respectively express.

[0128] The feature maps generated by the three convolutional layers CN1, CN2, and CN3 are input into the cross-attention module to calculate the attention weights between time features of different scales, and to obtain the contextual information of small-scale time features and the rich detail information contained in large-scale time features.

[0129] The structure of the first cross-attention module CA1 is shown in the figure Figure 5 As shown, the feature map output by the first convolutional layer CN1 And the feature map output by the second convolutional layer CN2 As the two inputs of the first cross attention module CA1, three projection modules P1, P2, and P3 are first used to perform feature linear projection on the two inputs through 1×1 point-by-point convolution to obtain the corresponding vector matrix Among them, W Q , W k , W V are the learnable parameters of three 1×1 point-by-point convolutions. Then, the cross-attention mechanism is used to fuse them and obtain The calculation process can be expressed as:

[0130]

[0131] Among them, the first point multiplication module M1 is used to calculate Q and K T Then use the Softmax activation function to transform Q and K T Then, the second dot product module M2 is used to perform dot product weighting on the normalized feature map and the matrix V, thereby modeling the small-scale temporal feature context information. Finally, the residual connection module ADD is used to superimpose the attention map to obtain the feature map because and The dimensions are inconsistent. Before adding, you need to The dimensions are adjusted to Consistent, It can be expressed as:

[0132]

[0133] Similarly, for the second cross attention module CA2 in the MSCA module, follow the same steps to convert the feature map obtained above Convolutional layer output feature map with the third scale As two inputs, the feature map Input projection module P1, feature map Input projection modules P2 and P3, and obtain feature maps after fusion through cross attention mechanism Then, the feature maps obtained above are concatenated in the channel dimension to obtain the feature map F m , which can be expressed as:

[0134]

[0135] It is understandable that the structure of the above-mentioned cross-attention module and the connection method with the convolution layer are only one implementation method of the embodiment of the present disclosure. The cross-attention module according to the embodiment of the present disclosure may also adopt other structures and connection methods with the convolution layer, as long as it can fuse the outputs of multiple convolution layers based on the attention mechanism.

[0136] Then, the feature map F is normalized by the batch normalization module BN. m Perform batch normalization, which ensures that the input distribution remains unchanged during network training by normalizing the feature map. This operation helps to increase the amplitude of the gradient and effectively avoids the gradient vanishing problem. At the same time, a larger gradient means a faster convergence speed, which significantly improves the training efficiency. The last layer of the MSCA module is the activation function module, which uses the ReLU activation function. The main advantage of the ReLU activation function is that it can enhance the nonlinear connection between the layers of the network and help reduce overfitting by generating sparse activation features. More importantly, the combination of the batch normalization module BN and the activation function ReLU can make the model learning process more robust, thereby improving the model performance. Finally, the output of the MSCA module can be expressed as:

[0137] F′ m =ReLU(BN(F m ))

[0138] According to an embodiment of the present disclosure, the multi-scale attention module MA may further include a Squeeze-and-Excitation (SE) attention module SE.

[0139] like Figure 4 As shown, the compression-excitation attention module SE includes a second global average pooling layer GAP2, a fourth fully connected layer FC4, a fifth fully connected layer FC5, and a third dot product module M3, which are connected in series. The third dot product module M3 is used to combine the output of the fifth fully connected layer FC5 with the input F′ of the compression-excitation attention module. m The compression-excitation attention module SE selects important information and removes redundant information by rescaling the weight of each multi-scale temporal feature, thereby enhancing the recognition ability of the network. The compression-excitation attention module SE first converts the output F′ of the multi-scale cross-attention module MSCA into m The input is sent to the second global average pooling layer GAP2, which compresses the global information to the channel for representation, which is the "compression" operation of the SE attention mechanism. The channel feature response is generated by calculating the average value of all channels, and its output F q It can be expressed as:

[0140]

[0141] Among them, L f is the input vector X of the multi-scale cross-attention module MSCA k The length of F′ m (i) is F′ m The second global average pooling layer GAP2 is connected in series with the fourth fully connected layer FC4 and the fifth fully connected layer FC5, which is the "excitation" operation of the SE attention mechanism. After the excitation operation, the channel weight F e It can be expressed as:

[0142] F e =σ(W2ReLU(W1F q ))

[0143] Among them, W1 and W2 are the connection weight parameters of the fourth fully connected layer FC4 and the fifth fully connected layer FC5 respectively, and σ is the Sigmoid activation function. Finally, the generated channel weight is applied to the input feature map F′ m , and get the final output F of the MA module ma :

[0144] F ma =F e F′ m

[0145] It can be understood that the compressed incentive attention module SE is not necessary, but including the compressed incentive attention module SE can enhance the recognition ability of the network.

[0146] Figure 6 A flowchart of a bone density prediction model training method based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0147] like Figure 6 As shown, the bone density prediction model training method based on ultrasonic radio frequency signals according to an embodiment of the present disclosure includes steps S201-S203.

[0148] In step S201, a plurality of sample data collected from a plurality of subjects are acquired, wherein the sample data include ultrasonic radio frequency signal data collected from corresponding subjects by a quantitative ultrasonic device, wherein the quantitative ultrasonic device includes an ultrasonic transmitter and an ultrasonic receiver, and the ultrasonic radio frequency signal data include original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on a human body.

[0149] In step S202, labels of the plurality of sample data are determined based on the bone density data of the plurality of subjects.

[0150] In step S203, a bone density prediction model is trained using the multiple sample data and respective labels of the multiple sample data, wherein: the bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channel includes: multiple convolution layers with different convolution kernel sizes, used to extract time features of multiple scales of the ultrasonic radio frequency signal data; a cross-attention module, used to weightedly fuse the time features of the multiple scales to obtain a fused feature map for bone density prediction, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0151] According to an embodiment of the present disclosure, the bone density prediction model includes the above reference Figure 3-Figure 5 A bone density prediction model was described.

[0152] According to an embodiment of the present disclosure, the designated detection position includes the radius, and the label includes the axial bone density value and / or the axial bone mass category determined based on the bone density data (for example, the axial bone density value) obtained by detection by the DXA device. For any subject, the sample data and the bone density data are collected on the same day.

[0153] According to the bone density prediction model based on ultrasonic radio frequency signals in the embodiment of the present disclosure, the axial bone density value and bone density category can be accurately predicted. Compared with the T value corresponding to the bone density calculated by calculating the SOS value, BUA value and other quantitative parameters of the ultrasonic radio frequency signal, this method is more intuitive and is conducive to the establishment of an artificial intelligence osteoporosis diagnosis database based on ultrasonic radio frequency signals and the formulation of new thresholds for bone quantitative ultrasonic diagnosis. Secondly, the measurement site of the embodiment of the present disclosure is the distal 1 / 3 of the radius on the non-dominant side. Compared with the classic axial bone density measurement method (such as lumbar spine and hip DXA), it can be completed in a shorter time, and usually does not require the patient to take off clothes or make special preparations. In addition, the embodiment of the present disclosure uses the ultrasonic radio frequency signal of the quantitative ultrasonic device to obtain bone mass information. Compared with the common peripheral bone density methods (such as pDXA, pQCT, etc.), there is no radiation exposure, no special inspection site is required to shield radiation, and the measurement is more convenient, which is more suitable for large-scale screening, clinical monitoring, and epidemiological surveys.

[0154] In order to verify the performance of the bone density prediction method according to the embodiment of the present disclosure, the measurement results of the DXA device were used as a benchmark, and the prediction results of the bone density prediction method of the embodiment of the present disclosure were compared with the measurement results of the original host system of BMD-9. The classification performance and prediction performance of the bone density prediction method of the embodiment of the present disclosure achieved good results under multiple evaluation indicators, and the accuracy (ACC), recall (Recall), positive predictive value (PPV) and area under the curve (AUC) values ​​were increased by 27.70%, 25.86%, 29.23% and 22.90% respectively. In the bone density prediction task, the minimum mean square error of the bone density prediction method of the embodiment of the present disclosure is 0.011, the determination coefficient is 0.457, and the Pearson correlation coefficient is 0.689, showing good prediction performance, indicating that the model has good sensitivity to the changes in bone density of different subjects and has a strong recognition ability for osteoporosis.

[0155] Figure 7 A structural block diagram of a bone density prediction device based on ultrasonic radio frequency signals according to an embodiment of the present disclosure is shown.

[0156] The device may be implemented as part or all of an electronic device through software, hardware, or a combination of both.

[0157] like Figure 7 As shown, the bone density prediction device 700 includes a first acquisition module 710 and a prediction module 720 .

[0158] The first acquisition module 710 is configured to acquire detection data of the measured user, wherein the detection data includes ultrasonic radio frequency signal data collected from the measured user by a quantitative ultrasonic device, wherein the quantitative ultrasonic device includes an ultrasonic transmitter and an ultrasonic receiver, and the ultrasonic radio frequency signal data includes original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body.

[0159] The prediction module 720 is configured to perform bone density prediction on the tested user based on the detection data using a trained bone density prediction model, wherein: the bone density prediction model is trained using multiple sample data collected from multiple subjects and the labels of each of the multiple sample data, the sample data includes ultrasonic radio frequency signal data collected from the corresponding subjects by a quantitative ultrasound device, and the labels of each of the multiple sample data are determined based on the bone density data of the multiple subjects; the bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channel includes a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract multiple scales of time features of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the multiple scales of time features to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0160] Figure 8 The structural block diagram of the bone density prediction model training device based on ultrasonic radio frequency signals 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.

[0161] like Figure 8 As shown, the bone density prediction model training device 800 includes a second acquisition module 810 , a determination module 820 , and a training module 830 .

[0162] The second acquisition module 810 is configured to acquire a plurality of sample data collected from a plurality of subjects, wherein the sample data include ultrasonic radio frequency signal data collected from corresponding subjects by a quantitative ultrasonic device, wherein the quantitative ultrasonic device includes an ultrasonic transmitter and an ultrasonic receiver, and the ultrasonic radio frequency signal data include original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body.

[0163] The determination module 820 is configured to determine labels of the plurality of sample data based on the bone density data of the plurality of subjects.

[0164] The training module 830 is configured to use the multiple sample data and the respective labels of the multiple sample data to train a bone density prediction model, wherein: the bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channel includes a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic RF signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

[0165] The bone density prediction model in the bone density prediction device and the bone density prediction model training device according to the embodiment of the present disclosure can be the bone density prediction model in the above reference Figure 3-Figure 5 A bone density prediction model was described.

[0166] The present disclosure also discloses an electronic device, Fig. 9 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0167] like Fig. 9 As shown, the electronic device includes 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 bone density prediction method and / or bone density prediction model training method according to an embodiment of the present disclosure.

[0168] Fig.10 A schematic diagram showing the structure of a computer system suitable for implementing the method according to an embodiment of the present disclosure is shown.

[0169] like Fig.10 As shown, the computer system includes a processing unit, which can perform the various methods in the above-mentioned embodiments according to the program stored in the read-only memory (ROM) or the program loaded from the storage part into the random access memory (RAM). In the RAM, various programs and data required for the operation of the computer system are also stored. The processing unit, ROM and RAM are connected to each other through a bus. The input / output (I / O) interface is also connected to the bus.

[0170] The following components are connected to the I / O interface: an input part including a keyboard, a mouse, etc.; an output part including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage part including a hard disk, etc.; and a communication part including a network interface card such as a LAN card, a modem, etc. The communication part performs a communication process via a network such as the Internet. The drive is also connected to the I / O interface as needed. Removable media, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive as needed so that the computer program read therefrom is installed into the storage part as needed. Among them, the processing unit can be implemented as a processing unit such as a CPU, a GPU, a TPU, an FPGA, an NPU, etc.

[0171] 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 machine-readable medium, 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 a communication part, and / or installed from a removable medium.

[0172] The embodiment of the present disclosure also provides an artificial intelligence ultrasonic bone densitometer. The artificial ultrasonic bone densitometer according to the embodiment of the present disclosure includes: a strap, one or more ultrasonic transmitters and one or more ultrasonic receivers arranged on the strap, and a control host.

[0173] The control host includes a processor and a memory, 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 bone density prediction method according to the embodiment of the present disclosure.

[0174] According to the embodiment of the present disclosure, the artificial ultrasonic bone densitometer adopts a ring probe to detect the position of the radius. The bracelet probe includes a bracelet strap and a probe assembly, and the probe assembly has a modular design that can be detached from the strap. In a specific example, the probe assembly includes a "dual-transmit and dual-receive" piezoelectric ceramic high-performance ultrasonic transducer, which are transducers A, B, C, and D, respectively, wherein A and B are respectively used as transmitting transducers E1 and E2 (used as ultrasonic transmitters), and C and D are respectively used as receiving transducers R1 and R2 (used as ultrasonic receivers), and each array element has an angle of about 30°, forming an axial transmission multi-channel ultrasonic transducer combination.

[0175] The strap can be made of fine frosted soft rubber material, designed with buckles for adjusting different tightness and grooves that fit the radial side, which can keep the probe assembly close to the radius and as parallel as possible, ensuring stable and reliable signals during detection.

[0176] The designated test site is the distal third of the radius of the non-dominant hand (usually the left hand). If there is any trauma to the forearm or wrist joint on that side, the contralateral forearm is tested. During the measurement, the user takes a comfortable seat and places their forearm naturally on the examination table. After applying sufficient coupling agent on the probe assembly, the operator assists the user to wear the bracelet probe on the measured part, starts the measurement and remains in this position until the measurement is completed.

[0177] The technical solution provided according to the embodiment of the present disclosure makes full use of the rich amplitude, frequency and phase information in the ultrasonic radio frequency signal. This information reflects the interaction between the ultrasonic sound field and human tissue and the microstructural characteristics of human tissue. The bone density prediction model is obtained by training a deep neural network model including an attention mechanism. The bone density prediction model is used to predict bone density based on the ultrasonic radio frequency signal of the user being tested. The key characteristic variables related to bone density prediction in the ultrasonic radio frequency signal can be automatically extracted through the attention mechanism. Compared with the existing method of extracting acoustic parameters from the ultrasonic signal and deriving bone density based on the acoustic parameters, it has higher intelligence and accuracy. At the same time, compared with DXA, QCT and other detection methods, the quantitative ultrasonic detection method according to the embodiment of the present disclosure has the advantages of low cost, small size, portability, less time consumption, no radiation, etc., does not require a special inspection site to shield radiation, and the measurement is more convenient, which is more in line with the needs of large-scale screening, clinical monitoring, epidemiological surveys, etc.

[0178] The flow charts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each box in the flow chart 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 a 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 flow chart, and the combination of the boxes in the block diagram and / or flow chart 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.

[0179] 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.

[0180] 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.

[0181] 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 bone density prediction method based on ultrasonic radio frequency signals, characterized in that: include: Acquire detection data of the user under test, the detection data including ultrasonic radio frequency signal data collected from the user under test by a quantitative ultrasonic device, the quantitative ultrasonic device including an ultrasonic transmitter and an ultrasonic receiver, the ultrasonic radio frequency signal data including original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body; Based on the detection data, the bone density of the tested user is predicted using a trained bone density prediction model, wherein: The bone density prediction model is trained using a plurality of sample data collected from a plurality of subjects and respective labels of the plurality of sample data, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, and the respective labels of the plurality of sample data are determined based on the bone density data of the plurality of subjects; The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

2. The method according to claim 1, characterized in that: The designated detection location includes the radius; The label includes an axial bone density value and / or an axial bone mass category; For any subject, the sample data and bone density data are collected on the same day.

3. The method according to claim 1, characterized in that: The quantitative ultrasound device includes M ultrasound generators and N ultrasound receivers, and the ultrasound radio frequency signal data includes ultrasound radio frequency signal data of M*N transmission-reception channels formed by the M ultrasound generators and the N ultrasound receivers; The bone density prediction model includes M*N multi-scale attention convolution channels, which are used to process the ultrasonic radio frequency signal data of the M*N transmit-receive channels respectively.

4. The method according to claim 1, characterized in that: The output layer includes a fully connected layer.

5. The method according to claim 4, characterized in that: The bone density prediction result includes a bone density category and / or a bone density value; The fully connected layer includes a first fully connected layer and / or a second fully connected layer, the first fully connected layer is used to obtain a bone density category based on the fused feature map, and the second fully connected layer is used to obtain a bone density value based on the fused feature map.

6. The method according to claim 4, characterized in that When the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: A first splicing module, used to splice the outputs of the multiple multi-scale attention convolution channels; A first global average pooling layer is used to perform global average pooling on the output of the first concatenation module. Among them, the fully connected layer obtains the bone density prediction result based on the output of the first global average pooling layer.

7. The method according to claim 6, characterized in that When the bone density prediction model includes multiple multi-scale attention convolution channels, the bone density prediction model also includes: A second concatenation module is used to concatenate the clinical risk factor obtained based on the clinical risk factor data with the output of the first global average pooling layer, Among them, the bone density prediction result is obtained based on the output of the one or more multi-scale attention convolution channels, including obtaining the bone density prediction result based on the output of the second splicing module.

8. The method according to claim 7, characterized in that: The bone density prediction model further includes a third fully connected layer, for obtaining the clinical risk factor based on the clinical risk factor data; The clinical risk factors include at least one or more of the following: gender, age, weight, height, and fracture history.

9. The method according to claim 1, characterized in that: The multi-scale attention convolution channel includes a multi-scale attention unit or multiple multi-scale attention units connected in series, and the multi-scale attention unit includes one or multiple multi-scale attention modules connected in series.

10. The method according to claim 9, characterized in that: When the multi-scale attention convolution channel includes multiple multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a pooling layer.

11. The method according to claim 10, characterized in that The pooling layer includes any one of the following: a maximum pooling layer, an average pooling layer, a global pooling layer, an adaptive pooling layer, a global average pooling layer, and a random pooling layer.

12. The method according to claim 9, characterized in that The multi-scale attention convolution channel includes three multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a maximum pooling layer, and the multi-scale attention unit includes two multi-scale attention modules connected in series.

13. The method according to claim 9, characterized in that The multi-scale attention module includes a multi-scale cross attention module, and the multi-scale cross attention module includes m convolutional layers, m-1 cross attention modules, and a third splicing module, where m≥2, wherein: The m convolutional layers have different convolution kernel sizes and are used to extract temporal features of different scales from the input of the multi-scale cross attention module; The m-1 cross attention modules are used to fuse the outputs of the m convolutional layers step by step in the order of convolution kernels from small to large based on the cross attention mechanism; The third splicing module is used to splice the output of the convolution layer with the smallest convolution kernel with the outputs of the m-1 cross-attention modules.

14. The method according to claim 13, characterized in that: The input ends of the m convolutional layers are connected to the input end of the multi-scale cross attention module, the convolution kernel of the i-th convolutional layer is smaller than the convolution kernel of the i+1-th convolutional layer, 1≤i≤m-1; The first input of the first criss-cross attention module is connected to the output of the second convolutional layer, and the second input of the first criss-cross attention module is connected to the output of the first convolutional layer; The first input end of the j-th cross-attention module is connected to the output end of the j+1-th convolutional layer, and the second input end of the j-th cross-attention module is connected to the first output end of the j-1-th cross-attention module, wherein the first output end of the cross-attention module outputs the cross-attention fusion result of the first input data of the first input end and the second input data of the second input end of the cross-attention module, 2≤j≤m-1.

15. The method according to claim 14, characterized in that The cross-attention module includes: A first projection module, configured to perform linear projection on the first input data using 1*1 point-by-point convolution to obtain a first vector matrix Q; A second projection module, used for performing linear projection on the second input data using 1*1 point-by-point convolution to obtain a second vector matrix K; A third projection module, used for performing linear projection on the second input data using 1*1 point-by-point convolution to obtain a third vector matrix V; A first dot product module, used to calculate the dot product of the first vector matrix Q and the transposed second vector matrix K; A normalization module, used for normalizing the output of the first dot product module; A second dot product module, used to calculate the dot product of the output of the normalization module and the third vector matrix V; A residual connection module, used for superimposing the first input data with the output of the second dot product module; a first output terminal, wherein the first output terminal outputs a calculation result of the second dot product module; A second output terminal, wherein the second output terminal outputs the calculation result of the residual connection module.

16. The method according to claim 13, characterized in that The multi-scale cross-attention module also includes a batch normalization module and an activation function module which are sequentially connected in series to the output of the third splicing module.

17. The method according to claim 13, characterized in that The multi-scale attention module also includes a compression-excitation attention module, which is used to scale the weights of features of different scales contained in the output of the multi-scale cross-attention module to remove redundant information.

18. The method according to claim 17, characterized in that The compressed incentive attention module includes a second global average pooling layer, a fourth fully connected layer, a fifth fully connected layer, and a third dot multiplication module connected in series in sequence, and the third dot multiplication module is used to multiply the output of the fifth fully connected layer with the input of the compressed incentive attention module.

19. A bone density prediction model training method based on ultrasonic radio frequency signals, characterized in that: include: Acquire a plurality of sample data collected from a plurality of subjects, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, wherein the quantitative ultrasound device includes an ultrasound transmitter and an ultrasound receiver, and the ultrasound radio frequency signal data includes original radio frequency signal data detected by the ultrasound receiver after the ultrasound signal emitted by the ultrasound transmitter passes through a designated detection position on the human body; Based on the bone density data of the multiple subjects, determining labels of the multiple sample data respectively; Using the plurality of sample data and the labels of the plurality of sample data, a bone density prediction model is trained, wherein: The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, wherein the multi-scale attention convolution channels include: multiple convolution layers with different convolution kernel sizes, used to extract time features of multiple scales of the ultrasonic radio frequency signal data; a cross-attention module, used to weightedly fuse the time features of the multiple scales to obtain a fused feature map for bone density prediction, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

20. The method according to claim 19, characterized in that: The quantitative ultrasound device includes M ultrasound generators and N ultrasound receivers, and the ultrasound radio frequency signal data includes ultrasound radio frequency signal data of M*N transmission-reception channels formed by the M ultrasound generators and the N ultrasound receivers; The bone density prediction model includes M*N multi-scale attention convolution channels, which are used to process the ultrasonic radio frequency signal data of the M*N transmit-receive channels respectively.

21. The method according to claim 19, characterized in that: The multi-scale attention convolution channel includes a multi-scale attention unit or multiple multi-scale attention units connected in series, and the multi-scale attention unit includes one or multiple multi-scale attention modules connected in series.

22. The method according to claim 21, characterized in that: When the multi-scale attention convolution channel includes multiple multi-scale attention units connected in series, adjacent multi-scale attention units are connected through a pooling layer.

23. The method according to claim 21, characterized in that The multi-scale attention module includes a multi-scale cross-attention module and a compression-excitation attention module, the multi-scale cross-attention module includes m convolutional layers, m-1 cross-attention modules, a third splicing module, m≥2, and also includes a batch normalization module and an activation function module connected in series to the output of the third splicing module, the compression-excitation attention module is used to scale the weights of features of different scales contained in the output of the multi-scale cross-attention module to remove redundant information, wherein: The m convolutional layers have different convolution kernel sizes and are used to extract temporal features of different scales from the input of the multi-scale cross attention module; The m-1 cross attention modules are used to fuse the outputs of the m convolutional layers step by step in the order of convolution kernels from small to large based on the cross attention mechanism; The third splicing module is used to splice the output of the convolution layer with the smallest convolution kernel with the outputs of the m-1 cross-attention modules.

24. The method according to claim 23, characterized in that: The input ends of the m convolutional layers are connected to the input end of the multi-scale cross attention module, the convolution kernel of the i-th convolutional layer is smaller than the convolution kernel of the i+1-th convolutional layer, 1≤i≤m-1; The first input of the first criss-cross attention module is connected to the output of the second convolutional layer, and the second input of the first criss-cross attention module is connected to the output of the first convolutional layer; The first input end of the j-th cross attention module is connected to the output end of the j+1-th convolutional layer, and the second input end of the j-th cross attention module is connected to the first output end of the j-1-th cross attention module, wherein the first output end of the cross attention module outputs the cross attention fusion result of the first input data of the first input end and the second input data of the second input end of the cross attention module, 2≤j≤m-1; The compressed incentive attention module includes a second global average pooling layer, a fourth fully connected layer, a fifth fully connected layer, and a third dot multiplication module connected in series in sequence, and the third dot multiplication module is used to multiply the output of the fifth fully connected layer with the input of the compressed incentive attention module.

25. The method according to claim 24, characterized in that The cross-attention module includes: A first projection module, configured to perform linear projection on the first input data using 1*1 point-by-point convolution to obtain a first vector matrix Q; A second projection module, used for performing linear projection on the second input data using 1*1 point-by-point convolution to obtain a second vector matrix K; A third projection module, used for performing linear projection on the second input data using 1*1 point-by-point convolution to obtain a third vector matrix V; A first dot product module, used to calculate the dot product of the first vector matrix Q and the transposed second vector matrix K; A normalization module, used for normalizing the output of the first dot product module; A second dot product module, used to calculate the dot product of the output of the normalization module and the third vector matrix V; A residual connection module, used for superimposing the first input data with the output of the second dot product module; a first output terminal, wherein the first output terminal outputs a calculation result of the second dot product module; A second output terminal, wherein the second output terminal outputs the calculation result of the residual connection module.

26. A bone density prediction device based on ultrasonic radio frequency signals, characterized in that: include: A first acquisition module is configured to acquire detection data of the user under test, wherein the detection data includes ultrasonic radio frequency signal data collected from the user under test by a quantitative ultrasonic device, wherein the quantitative ultrasonic device includes an ultrasonic transmitter and an ultrasonic receiver, and wherein the ultrasonic radio frequency signal data includes original radio frequency signal data detected by the ultrasonic receiver after the ultrasonic signal emitted by the ultrasonic transmitter passes through a designated detection position on the human body; The prediction module is configured to predict the bone density of the tested user using a trained bone density prediction model based on the detection data, wherein: The bone density prediction model is trained using a plurality of sample data collected from a plurality of subjects and respective labels of the plurality of sample data, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, and the respective labels of the plurality of sample data are determined based on the bone density data of the plurality of subjects; The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

27. A bone density prediction model training device based on ultrasonic radio frequency signals, characterized in that: include: a second acquisition module, configured to acquire a plurality of sample data collected from a plurality of subjects, wherein the sample data includes ultrasound radio frequency signal data collected from corresponding subjects by a quantitative ultrasound device, wherein the quantitative ultrasound device includes an ultrasound transmitter and an ultrasound receiver, and the ultrasound radio frequency signal data includes original radio frequency signal data detected by the ultrasound receiver after the ultrasound signal emitted by the ultrasound transmitter passes through a designated detection position on the human body; A determination module, configured to determine a label of each of the plurality of sample data based on the bone density data of the plurality of subjects; The training module is configured to train a bone density prediction model using the plurality of sample data and the labels of the plurality of sample data, wherein: The bone density prediction model includes an output layer and one or more multi-scale attention convolution channels, the multi-scale attention convolution channels include a cross-attention module and multiple convolution layers, the multiple convolution layers have different convolution kernel sizes, and are used to extract the time features of multiple scales of the ultrasonic radio frequency signal data, the cross-attention module is used to weightedly fuse the time features of the multiple scales to obtain a fused feature map, and the output layer is used to obtain a bone density prediction result based on the fused feature map.

28. An electronic device, characterized in that: The method comprises a processor and a memory, 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 25.

29. An artificial intelligence ultrasonic bone densitometer, characterized in that: include: straps; One or more ultrasonic transmitters and one or more ultrasonic receivers are disposed on the strap; A control host, the control host comprising a processor and a memory, 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 described in any one of claims 1-25.

30. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the method described in any one of claims 1 to 25 is implemented.

31. A computer program product, comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the method described in any one of claims 1 to 25 is implemented.

Images