Cerebral hemorrhage detection device with noninvasive radio frequency microwave fused with near-infrared light

Abstract

The invention relates to the technical field of medical detection, in particular to a non-invasive radio frequency microwave and near-infrared light fused cerebral hemorrhage detection device, which adopts non-ionizing radiation low-power microwave, near-infrared light and other multi-mode technologies, realizes non-invasive continuous monitoring, and can detect cerebral hemorrhage by integrating the advantages of various physical signals. And the defects of a single signal mode can be effectively overcome, high-precision early diagnosis of cerebral hemorrhage is realized, and precious treatment time is won for patients.

Description

Technical Field

[0001] The present disclosure relates to the field of medical detection equipment, and in particular to a device that utilizes non-invasive radio frequency microwave technology to fuse near-infrared light for rapid and accurate detection of acute cerebral hemorrhage. Background Art

[0002] Intracerebral hemorrhage (ICH) is a common cerebrovascular disease with an acute onset, rapid changes in the condition, and a high disability and mortality rate. Early and accurate diagnosis and assessment of the location and range of acute ICH are of vital importance for timely and effective treatment. However, the main means of detecting ICH (such as CT and MRI) currently have certain limitations: CT and MRI equipment are expensive and bulky, and are usually limited to use in medical institutions; it usually takes a certain amount of time for patients to be diagnosed from the onset of the disease, especially in remote areas or areas with scarce medical resources; CT involves ionizing radiation, and long-term use may affect the health of patients. Although CT imaging speed is relatively fast, it has radiation hazards, and its sensitivity to tiny hemorrhage foci may be limited in the early stage; although MRI has high resolution for soft tissue, the examination time is long and the cost is high. Some patients have metal implants in their bodies and cannot be used, so it is not suitable for emergency rapid screening.

[0003] At present, the determination of stroke types, especially the determination between stroke / non-stroke and ischemic / hemorrhagic stroke, is not diagnosed until the hospital's electronic computed tomography (CT) and nuclear magnetic resonance imaging (fMRI) are differentiated, which leads to poor pre-hospital triage and delayed treatment of patients, affecting their postoperative recovery. It can be seen that although the existing CT and fMRI technologies are the gold standard, they take a long time from onset to diagnosis and the equipment is not portable, which can easily cause patients to miss the golden window period for stroke treatment.

[0004] The shortcomings of existing stroke monitoring devices are as follows:

[0005] (1) The test takes a long time. On the one hand, the time includes the time from onset to treatment. For example, CT and fMRI are large and inconvenient to move, so patients can only undergo brain imaging diagnosis after admission to the hospital, which makes the test time-consuming. On the other hand, the time-consuming process itself takes a long time. For example, EEG identification of stroke type requires a long time because the electrodes themselves need to be moistened and placed.

[0006] (2) The detection range is not wide. For example, volume impedance phase-shift spectroscopy can only be used to identify patients with anterior circulation large vessel occlusion (LVO), but is not suitable for identifying patients with posterior circulation LVO and cerebral hemorrhage.

[0007] (3) There is a risk of secondary injury during the detection process. The existing microwave stroke monitoring method is circular, and patients need to wear the device on their heads. In reality, stroke patients should minimize movement and lie flat. At the same time, if the patient lies flat, the existing device needs to move the patient's head and press it on the device. If the stroke site is located in the cerebral cortex, there is a risk of secondary injury. Moreover, although microwaves can penetrate the skull and sensitively detect the electromagnetic property differences between different tissues (such as blood and brain tissue), a single microwave signal may be affected by the complexity of the head structure, resulting in insufficient spatial resolution. Near-infrared light has strong tissue penetration ability and can sensitively detect changes in blood oxygen saturation. However, it may be difficult to accurately locate the cerebral hemorrhage area relying solely on near-infrared light because the light scattering and absorption of the skull will affect the signal quality. Summary of the Invention

[0008] To solve the above problems existing in the prior art, the present disclosure provides a non-invasive cerebral hemorrhage detection device that fuses radio frequency microwaves and near-infrared light. The device adopts multimodal technology and can effectively overcome the deficiencies of a single signal mode by integrating the advantages of multiple physical signals.

[0009] In a first aspect, the present disclosure provides a non-invasive cerebral hemorrhage detection device that fuses radio frequency microwaves and near-infrared light. The device includes a multimodal sensor, a signal processing module, and a fusion analysis module. Among them: The multimodal sensor is configured to synchronously acquire near-infrared spectral signals and radio frequency microwave signals during cerebral hemorrhage monitoring. The signal processing module is configured to construct voxels according to the Cartesian coordinate system set based on the average volume of human head nuclear magnetic resonance data, match the position information of the multimodal sensor with the voxels, and then obtain the near-infrared spectral signals and radio frequency microwave signals corresponding to each voxel. The light intensity attenuation rate, photon average flight time, and photon count distribution variance extracted based on the near-infrared spectral signals, and the S21 scattering parameter, phase shift, and time-domain pulse response peak obtained based on the radio frequency microwave signals are used as the signal feature vectors corresponding to each voxel. The fusion analysis module is configured to generate a three-dimensional reconstruction map of the cerebral hemorrhage area based on the signal feature vectors of the voxels and the spatial information of the voxels using a trained deep learning fusion model.

[0010] In an embodiment of the above technical solution, the deep learning fusion model includes a 3D convolutional neural network and a graph attention network. Among them: The 3D convolutional neural network is configured to extract the cerebral hemorrhage feature vectors of all voxels based on the signal feature vectors of all voxels. The graph attention network is configured to use the voxels as graph nodes based on the cerebral hemorrhage feature vectors of all voxels, learn the relationships between the cerebral hemorrhage features of the nodes using the attention mechanism, obtain the bleeding probabilities of each node, and then generate a three-dimensional reconstruction effect diagram of the cerebral hemorrhage area.

[0011] In an implementation of the above technical solution, the device further includes a reliability analysis module; the reliability analysis module is configured to calculate the radio frequency data confidence and the near-infrared data confidence of all multimodal sensors. If there is a confidence greater than the first set threshold among the two, resample the signal data corresponding to the confidence with a locally changed depth; if the sum of the two confidences is greater than the second set threshold, it is considered that the radio frequency microwave signals and the near-infrared spectral signals of all multimodal sensors have high confidence, and the fusion analysis module is used to generate a three-dimensional reconstruction map of the cerebral hemorrhage area based on the signal feature vectors.

[0012] In an implementation of the above technical solution, the radio frequency data confidence and the near-infrared data confidence of all multimodal sensors are calculated as follows:

[0013]

[0014] In the formula: C RF is the radio frequency data confidence of n multimodal sensors, C NIR is the near-infrared data confidence of n multimodal sensors, ΔΦ i is the phase shift of the i-th multimodal sensor, Φ th is the phase shift threshold; S21 drop，i represents the S21 parameter attenuation of the i-th channel, S21 th is the S21 scattering parameter threshold; ΔI j / I 0，j is the light intensity attenuation rate of the j-th multimodal sensor, I th is the light intensity attenuation threshold; TOF j is the average photon flight time of the j-th multimodal sensor, TOF th is the average photon flight time threshold.

[0015] In an implementation of the above technical solution, the voxel size is 16mm × 16mm × 16mm.

[0016] Second aspect, the present disclosure proposes a multimodal sensor for detecting cerebral hemorrhage. The multimodal sensor includes an MCU control module, a microwave signal transceiver module, a near-infrared light signal transceiver module, and a probe integrating near-infrared and microwave; wherein: the MCU control module includes STM32; the microwave transceiver module includes a radio frequency transmission link and a radio frequency reception link. In the radio frequency transmission link, radio frequency is generated by LMX2595, and after passing through QPA2211 and ADL5920, it is transmitted from the patch electrode. After the microwave signal is received by the patch electrode in the radio frequency reception link, it is transmitted to STM32 after passing through a low-noise amplifier, a mixer, and an analog-to-digital converter; the near-infrared signal transceiver module includes an optical signal generation optical path and an optical signal reception optical path. In the optical signal generation optical path, a multi-channel LP785-SAV50 laser diode is excited to generate near-infrared light through driving the LMH6521 high-speed constant current source and is emitted through the light source probe. The optical signal reception optical path is received by the silicon photomultiplier of S13360-3050CS of the light source probe, and is sent to STM32 after passing through a preamplifier OPA1612, a filter LTC1569, and an analog-to-digital converter; the microstrip of the microwave transceiver module and the silicon photomultiplier of the near-infrared signal transceiver module are coupled through the probe integrating near-infrared and microwave.

[0017] In an implementation manner of the above technical solution, STM32 generates periodic TTL pulses, and at the same time, the microwave transceiver module starts to transmit radio frequency signals and the near-infrared light source of the near-infrared signal transceiver module is modulated. The analog-to-digital converter of the radio frequency reception link and the analog-to-digital converter of the near-infrared share the same trigger signal.

[0018] In an implementation manner of the above technical solution, the substrate of the probe integrating near-infrared and microwave is in a shape of a Chinese character "hui". The size of the "hui"-shaped substrate is 10mm×10mm×0.6mm. The outer circle of the "hui" shape is a ring-shaped ground wire GND with a line width of 1mm, which is used to constrain radio frequency microwave signals. The side length of the middle circle is 7mm and the line width is 1mm, which is connected to the feeding position Feed. Openings are respectively made below the transmitting end Optical for passing through the multimode optical fiber for light conduction, and openings are made below the receiving end SiPM for passing through the light source probe SiPM. An IPEX socket is used on the back of the probe, and coaxial cables are used for radio frequency microwave signal transmission.

[0019] In an implementation manner of the above technical solution, the wavelength range during the operation of the near-infrared signal transceiver module is 650 - 950nm.

[0020] In an implementation manner of the above technical solution, the frequency band range during the operation of the microwave transceiver module is 0.5 - 10GHz.

[0021] Advantages of the present disclosure: It can be applied to the field of acute cerebral hemorrhage detection. Since microwave / radio frequency signals are highly sensitive to the electromagnetic properties of brain tissue (such as dielectric constant and conductivity), they can detect the nature and scope of cerebral hemorrhage, while near-infrared light is sensitive to changes in the concentrations of oxyhemoglobin and deoxyhemoglobin, and can evaluate the metabolic and physiological changes related to bleeding. Therefore, the signals of radio frequency microwave and near-infrared light can be used complementary to achieve measurement from shallow to deep. Since low-power microwave and infrared light are both non-ionizing radiation and harmless to human tissue, the device design can achieve non-invasive, continuous and safe monitoring, avoiding the pain of patients. The device can be designed as a portable device, suitable for patient screening and monitoring in pre-hospital first aid, primary hospitals or remote areas. The signal processing algorithm and intelligent analysis module can achieve real-time diagnosis. Moreover, the device of the present disclosure realizes early diagnosis of cerebral hemorrhage with low cost and high precision through multimodal fusion of radio frequency microwave and near-infrared light, winning precious treatment time for patients. Description of the Drawings

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0023] Figure 1 、 One Schematic diagram of the multimode sensor structure in a certain implementation manner.

[0024] Figure 2 、 One Schematic diagram of the probe in a certain implementation manner.

[0025] Figure 3 、 One Schematic diagram of the data fusion method flow in a certain implementation manner.

[0026] Figure 4 、 One Schematic diagram of the three-dimensional reconstruction of the bleeding area in a certain implementation manner. Detailed Implementation Manner

[0027] The following will clearly and completely describe how to implement the technical solutions of this case in combination with the drawings. Obviously, the described implementation manners are only some implementation manners of this case, rather than all implementation manners.

[0028] (I) Multimode Sensor

[0029] A cerebral hemorrhage detection device integrating near-infrared and microwave in implementation combines near-infrared spectroscopy (NIRS) and microwave detection technology, and utilizes the sensitivity of near-infrared light to hemoglobin and the response characteristics of microwave to tissue conductivity changes to achieve highly sensitive and non-invasive detection of cerebral hemorrhage. Specifically, it cooperatively collects near-infrared spectroscopy (NIRS) and microwave signals through a multi-modal sensor array, and combines machine learning algorithms to achieve high-precision detection and localization of cerebral hemorrhage.

[0030] The multi-modal sensor includes an MCU main control module, a microwave signal transceiver module, and a near-infrared light signal transceiver module.

[0031] See Figure 1 , the MCU main control module includes STM32, a power supply unit, RX07050, and LMK00304. Among them, the power supply unit is used to supply the electricity required for the operation of STM32. The external crystal oscillator RX07050 provides a stable reference clock of 10 MHz, and this clock signal is used for data distribution and level conversion through the low-jitter buffer LMK00304.

[0032] The microwave signal transceiver module includes a radio frequency transmission link and a radio frequency reception link. Among them, the radio frequency transmission link generates radio frequency by LMX2595, and after passing through QPA2211 and ADL5920, it is emitted from the patch electrode. After the radio frequency reception link receives the microwave signal from the patch electrode, it is transmitted to STM32 after passing through a low-noise amplifier, a mixer, and an analog-to-digital converter.

[0033] The microwave signal transceiver module can generate, adjust the frequency, and transmit and receive microwave signals, and sense the change of the dielectric constant in the cerebral hemorrhage area through the ultra-wideband (UWB) signal in the 0.5 - 10 GHz frequency band.

[0034] The near-infrared light signal transceiver module includes an optical signal generation optical path and an optical signal reception optical path. Among them, the optical signal generation optical path drives the LMH6521 high-speed constant current source to excite the multi-channel LP785-SAV50 laser diode to generate near-infrared light and emit it through the light source probe. The optical signal reception optical path is received by the silicon photomultiplier (SiPM) of S13360-3050CS of the light source probe, and is sent to STM32 after passing through the preamplifier OPA1612, the filter LTC1569, and the analog-to-digital converter.

[0035] The near-infrared module obtains the blood oxygen parameters of the brain tissue by emitting and receiving light signals of a specific wavelength through the transmitting / receiving unit, and the wavelength range during the operation of the near-infrared signal transceiver module is 650 - 950 nm.

[0036] The microstrip of the microwave transceiver module and the silicon photomultiplier of the near-infrared signal transceiver module are coupled through a probe integrating near-infrared and microwave rectangular microstrip.

[0037] See Figure 2 , the substrate of the probe integrating near-infrared and microwave rectangular microstrips is in a figure-eight shape. The size of the figure-eight substrate is 10 mm × 10 mm × 0.6 mm. The outer ring of the figure-eight is a circular ground wire GND with a line width of 1 mm, which is used to constrain the radio frequency and microwave signals. The side length of the middle ring is 7 mm and the line width is 1 mm, which is connected to the feeding position Feed. Open holes are respectively made at the lower parts of the feeding positions, with a hole diameter of 2 mm, that is, holes are made below the transmitting end Optical and below the receiving end SiPM. Among them, the hole below the transmitting end Optical is used to penetrate the multimode optical fiber for light conduction, and the hole below the receiving end SiPM is used to penetrate the light source probe SiPM. Since the optoelectronic signals do not interfere with each other at the physical layer, the effect of integrated arrangement can be achieved. Figure 2 This is the front of the probe. An IPEX socket is used on the back of the probe, and a coaxial cable is used for radio frequency and microwave signal transmission.

[0038] The material of the probe substrate is R04350B, with a dielectric constant of 3.66 and a loss tangent of 0.0037.

[0039] (2) Cerebral hemorrhage detection

[0040] Taking the use of 8 multimodal sensors as an example. Arrange the 8 multimodal sensors in a circular array to cover the frontal lobe, temporal lobe, and parietal lobe. When detecting cerebral hemorrhage, synchronously acquire near-infrared spectral signals and radio frequency and microwave signals.

[0041] (2.1) Dielectric constant change in the cerebral hemorrhage area

[0042] Enable the multimodal sensor to acquire the radio frequency and microwave signal that senses the dielectric constant change in the cerebral hemorrhage area.

[0043] Specifically, the STM32 MCU module controls the LMX2595 broadband frequency synthesizer for parameter writing and control. The external crystal oscillator RX07050 provides a stable reference clock signal of 10 MHz. This clock signal passes through the low-jitter buffer LMK00304 and is distributed to the REFIN pin of the LMX2595 for phase-locked loop frequency synthesis based on the reference clock input. Configure the divider parameters, and set the integer division ratio (N) and fractional division ratio (F) of the PLL through the SPI interface. The calculation formula is as follows:

[0044] f out = f ref × (N + F / 2 24 )

[0045] In the formula: f out is the output frequency, and f ref is the reference clock frequency.

[0046] According to the above formula and the fact that the LMX2595 internally integrates a multi-core voltage-controlled oscillator (VCO), it can provide a frequency output from 300 MHz to 10 GHz. At the same time, the low-noise mode of the LMX2595 is enabled to optimize the loop filter bandwidth and suppress high-frequency spurs. Furthermore, through power amplification and filtering, the target power and waveform output are achieved. This device is designed for pre-stage drive, final-stage amplification, and band-pass filtering. In the pre-stage drive of the LMX2595, after drive amplification, the power can be increased to +20 dBm. As the QPA2211 microwave power amplifier receives the drive signal, the gain can be amplified to +45 dBm, and an isolator is connected to the output terminal to prevent damage caused by reflection. A cavity band-pass filter is used to filter out spurious signals outside the frequency band. At the same time, the device can switch the working frequency band according to the scanning depth. Its working logic is that the LMX2595 controls the path selection through a radio frequency switch. When the detection depth > 3 cm, direct frequency division of the LMX2595 is selected; when the detection depth ≤ 3 cm, the high-frequency path is enabled. The radio frequency switch is controlled by the TTL level output by the STM32 MCU.

[0047] In addition, this device also includes standing wave ratio detection and temperature protection. The ADL5920 monitors the forward / reflected power in real time. When the standing wave ratio (VSWR) is too large, the STM32 cuts off the power supply of the power amplifier. In addition, by monitoring the heat dissipation temperature of the power amplifier, when the temperature is too high, the power is reduced until the power is cut off.

[0048] The received signal link includes low-noise amplification, down-conversion processing, and digital demodulation. The received signal first passes through the QPL9547 low-noise amplifier, and the HMC8193 mixer mixes the signal with the local oscillator L0 to output an intermediate frequency IF signal. The AD7768 analog-to-digital converter (ADC) samples, and the STM32 MCU demodulates the I / O components and calculates the amplitude and phase.

[0049] (2.2) Detection of brain tissue blood oxygen parameters (HbO2, Hb)

[0050] The multi-modal sensor is used to obtain the near-infrared spectral signals for detecting brain tissue blood oxygen parameters (Hb02, Hb).

[0051] Specifically, using the pulse generator of the STM32 MCU to drive the LMH6521 high-speed constant current source to excite the multi-channel LP785-SAV50 laser diode to emit near-infrared light with a wavelength of 785 nm and a power of 50 mW, realizing the emission of optical signals controlled by the STM32 MCU as follows.

[0052] After the laser diode is powered on, the temperature is stabilized at 25 ± 0.1 °C through the TCLDM9 temperature control module to ensure that the wavelength drift < 0.1 nm. An internal photodiode (PD) is used to feedback and regulate the drive current to stabilize the output optical power at 50 mW. By adopting time-division multiplexing and frequency encoding methods, the multi-channel light source is sequentially lit at 1 ms intervals to avoid multi-source crosstalk. Each light source is loaded with a 1 kHz sine wave modulation through the voltage-controlled input of LMH6521 for subsequent frequency-domain separation, and the output parameters are a pulse width duty cycle of 20% and an instantaneous peak power of 250 mW. The light source is conducted to the probe that integrates near-infrared and microwave rectangular microstrip through an optical fiber bundle.

[0053] At the probe, the light source reception is completed through the photoelectric probe S13360-3050CS (SiPM). The optical signal then passes through the preamplifier OPA1612 → filter LTC1569 → analog-to-digital converter AD7768. After the optical signal undergoes multiple scatterings in the brain tissue, part of it reaches the photoelectric detection module S13360-3050CS.

[0054] The multi-modal sensor front-end integrated spherical lens focuses the scattered light onto the SiPM photosensitive surface of S13360-3050CS, and then suppresses and eliminates the ambient light through the 785BP10 filter.

[0055] A DE1 reference signal synchronized with the light source modulation frequency is generated by STM32, and the lock-in amplification technology is used to extract the effective signal. This signal is input to the preamplifier OPA1612 to convert the SiPM output current into a voltage signal. AD7768 acquires the signal at a rate of 256 kSPS, sets a multi-bit resolution, and synchronously captures all multi-detector channels. The signal is stored in the cache for optical signal analysis. The optical signal analysis includes time-domain analysis and frequency-domain analysis. The time-domain analysis records the photon arrival time through time-correlated single photon counting, which is characterized by an increase in the optical path length and a decrease in the photon counting rate at the bleeding position. The frequency-domain analysis performs a fast Fourier transform on the 1 kHz modulation signal to extract the fundamental wave amplitude reflecting the light intensity attenuation and the phase shift reflecting the optical path change. Finally, based on the Monte Carlo simulation, a pre-built brain tissue light transmission model is used to invert the absorption coefficient distribution of the bleeding area through an iterative optimization algorithm.

[0056] (III) Signal Processing

[0057] The average volume of the human head nuclear magnetic resonance data is obtained, a Cartesian coordinate system is set according to the average volume, and it is divided into a uniform grid of 16×16×16. Each grid cell is called a voxel. The position information of the multi-modal sensor is matched with the voxel to obtain the near-infrared spectral signal and radio frequency microwave signal corresponding to each voxel.

[0058] Next, data preprocessing is performed, including:

[0059] Extract the S21 scattering parameter (amplitude attenuation), phase shift (Δφ), and peak value of the time-domain impulse response (TDR) from the radio frequency microwave signal, and use the fourth-order wavelet transform to remove high-frequency noise.

[0060] The light intensity attenuation rate, average photon flight time, and variance of photon count distribution extracted from the near-infrared spectral signal.

[0061] Complete the above signal processing by constructing a signal processing module.

[0062] (IV) Reliability analysis

[0063] Based on the above, the near-infrared module monitors the dynamic changes of Hb02 / Hb, and the microwave module captures the dielectric constant abnormality caused by bleeding.

[0064] First, synchronize the multimodal data. Since the synchronization mechanism includes hardware time synchronization and software time calibration, a low-jitter temperature-compensated crystal oscillator RX07050 is first used to provide a synchronous clock for LMX2595, the light source, and the ADC at the same time. In addition, it is distributed to each subsystem through a clock buffer (LMK00304) to ensure that the global clock deviation < 1 ns. In addition, this device generates periodic TTL pulses through STM32 to start the radio frequency signal transmission and NIR light source modulation at the same time. The ADC of the radio frequency receiving link and the ADC of the NIR share the same trigger signal, and the data frame header marks the synchronous timestamp, with the error accuracy controlled within ±10 ns. Software-wise, the on-board transmission delay of the radio frequency signal is compensated in the MCU through a pre-calibrated fixed delay (≈50 ns). The optical signal transmission delay is dynamically corrected based on the photon flight time (TOF) model. The spatial alignment criterion is based on the radio frequency, patch light source (Optical), and probe (SiPM) for coordinate calibration.

[0065] Secondly, perform data reliability judgment by calculating the radio frequency data confidence and near-infrared data confidence of all multimodal sensors. The calculation is as follows:

[0066]

[0067] Where: C RF is the radio frequency data confidence of n multimodal sensors, C NIR is the near-infrared data confidence of n multimodal sensors, ΔΦ i is the phase shift of the i-th multimodal sensor, Φ th is the phase shift threshold; S21 drop，i represents the S21 parameter attenuation amount of the i-th channel, S21 th is the S21 scattering parameter threshold; ΔI j / I 0，jis the light intensity attenuation rate of the j-th multimodal sensor, I th is the light intensity attenuation threshold; TOF j is the average photon flight time of the j-th multimodal sensor, TOF th is the average photon flight time threshold.

[0068] If the confidence of the RF data is greater than the first set threshold, or the confidence of the near-infrared data is greater than the first set threshold, then resampling for local change detection depth is performed. For example, if the initial detection depth is greater than 3 cm, high-frequency resampling with a value less than or equal to 3 cm is used at this time.

[0069] If C RF +C NIR > the second set threshold, it is determined as high confidence, and a signal feature vector is constructed and input into the fusion analysis module to output the bleeding location.

[0070] (V) Fusion Analysis Module

[0071] The fusion analysis module extracts and fuses features from multi-source data through a convolutional neural network (CNN) to generate a bleeding probability heat map and a three-dimensional reconstruction image. Since multi-source data is used, data fusion can overcome the limitations of a single modality.

[0072] In this part, through the fusion analysis module, the spatio-temporal registration and feature-level fusion of the brain tissue blood oxygen parameters obtained from the near-infrared module and the dielectric constant changes in the bleeding area obtained from the microwave module are performed to achieve high-precision bleeding detection and localization.

[0073] First, a feature vector is constructed. The feature dimension of the RF data is:

[0074] 3(S21, Δφ, TDR peak) × 8-channel antenna = 24 dimensions

[0075] The feature dimension of the near-infrared data is:

[0076] 3(Δl / l0, TOF, variance) × 8-channel optical path = 24 dimensions

[0077] The RF data features and the near-infrared data features are concatenated and fused into a signal feature vector (48 dimensions), and spatial grid information is added (see Figure 3 ), and the feature dimension after fusion:

[0078] 48-dimensional vector (24 + 24) + spatial grid information (X, Y, Z) = 51 dimensions

[0079] Among them, the spatial grid information is the central position information of a voxel.

[0080] Then, normalization processing is performed using the Z-score model, and each feature dimension is Processing. Where x is the parameter value, μ is the mean, and σ is the standard deviation.

[0081] Next, a 3D convolutional neural network and a graph attention network GAT fusion model are set up to fuse the 51-dimensional feature vector with the spatial grid. Specifically, the 3D convolutional layers are as follows: The first layer has a convolutional kernel of 3×3×3, the input channels are from 1 to 32, the stride is 1, and the padding is 1. The second layer has a convolutional kernel of 3×3×3, the input channels are from 32 to 64, the stride is 2, and downsampling is performed. The third layer has a convolutional kernel of 3×3×3, the input channels are from 64 to 128, and the stride is 2. In the graph attention layer GAT, the voxels are mapped to graph nodes, and the node features include radio frequency and near-infrared parameters. The graph attention weights are calculated as follows:

[0082] α ij = softmax(LeakyReLU(W[h i ||h j ))

[0083] where h i 、h j are the node features and W is the learnable weight matrix.

[0084] The output layer is a fully connected layer (128→64→1) and a Sigmoid activation function, which outputs a hemorrhage probability heat map.

[0085] In one implementation, the three-dimensional reconstruction effect diagram of the hemorrhage area is as Figure 4 shown. It can be seen from this figure that the fusion model extracts and fuses the features of multi-source data, generates a three-dimensional reconstruction image of the cerebral hemorrhage area, and can achieve high-precision hemorrhage detection and localization.

[0086] The resolution of the three-dimensional reconstruction image of the cerebral hemorrhage area is related to voxel segmentation. Exemplarily, if the size of a voxel is 16mm×16mm×16mm and the brain area of the hemorrhage site is 100mm, then the resolution of the hemorrhage site is 6mm 3 .

[0087] (VI) Summary

[0088] In view of the problems of long detection time, limited detection range, and the risk of secondary injury existing in the existing stroke monitoring devices, the present disclosure proposes a non-invasive cerebral hemorrhage detection device that fuses radio frequency microwave and near-infrared light. This device utilizes the multi-modal fusion of radio frequency microwave and near-infrared light to achieve low-cost and high-precision early diagnosis of cerebral hemorrhage, and strive for precious treatment time for patients.

[0089] Through the description of the above embodiments, those skilled in the art can clearly understand that the disclosed device can be implemented by means of software plus necessary general hardware. Of course, it can also be implemented by dedicated hardware including application specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. Generally, functions accomplished by computer programs can easily be implemented by corresponding hardware, and there can be various specific hardware structures for implementing the same function, such as analog circuits, digital circuits or dedicated circuits. However, in most cases for the present disclosure, implementation by software programs is a better embodiment.

[0090] Although the embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, the present disclosure is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present disclosure, and these all fall within the scope of protection of the present disclosure.

Claims

1. A non-invasive intracerebral hemorrhage detection device integrating radio frequency microwave and near-infrared light, characterized in that, The device includes a multi-modal sensor, a signal processing module, and a fusion analysis module; wherein: The multi-modal sensor is configured to synchronously acquire near-infrared spectroscopy signals and radio frequency microwave signals during intracerebral hemorrhage monitoring; The signal processing module is configured to construct voxels according to the Cartesian coordinate system set based on the average volume of human head nuclear magnetic resonance data, match the position information of the multi-modal sensor with the voxels, and then obtain the near-infrared spectroscopy signals and radio frequency microwave signals corresponding to each voxel, and use the light intensity attenuation rate, average photon flight time, and variance of photon count distribution extracted based on the near-infrared spectroscopy signals, and the S21 scattering parameter, phase shift, and peak value of time-domain impulse response obtained based on the radio frequency microwave signals as the signal feature vector corresponding to each voxel; The fusion analysis module is configured to generate a three-dimensional reconstruction map of the intracerebral hemorrhage area based on the signal feature vector of the voxel and the spatial information of the voxel by using a trained deep learning fusion model.

2. The device according to claim 1, wherein The deep learning fusion model includes a 3D convolutional neural network and a graph attention network; wherein: the 3D convolutional neural network is configured to extract the intracerebral hemorrhage feature vectors of all voxels based on the signal feature vectors of all voxels; the graph attention network is configured to use the voxels as graph nodes based on the intracerebral hemorrhage feature vectors of all voxels, and use the attention mechanism to learn the relationships between the intracerebral hemorrhage features of the nodes, obtain the bleeding probability of each node, and then generate a three-dimensional reconstruction effect diagram of the intracerebral hemorrhage area.

3. The device according to claim 1, characterized in that, The device further includes a reliability analysis module; the reliability analysis module is configured to calculate the radio frequency data confidence and near-infrared data confidence of all multi-modal sensors. If the confidence of either of them is greater than the first set threshold, resample the signal data corresponding to the confidence at a local change depth; if the sum of the two confidences is greater than the second set threshold, it is considered that the radio frequency microwave signals and near-infrared spectroscopy signals of all multi-modal sensors have high confidence, and use the fusion analysis module to generate a three-dimensional reconstruction map of the intracerebral hemorrhage area based on the signal feature vector.

4. The device according to claim 3, characterized in that, The radio frequency data confidence and near-infrared data confidence of all multi-modal sensors are calculated as follows: Where: C RF is the RF data confidence of n multimodal sensors, C NIR is the near-infrared data confidence of n multimodal sensors, ΔΦ i is the phase shift of the i-th multimodal sensor, Φ th is the phase shift threshold; S21 drop，i represents the S21 parameter attenuation of the i-th channel, S21 th is the S21 scattering parameter threshold; ΔI j / I 0，j is the light intensity attenuation rate of the j-th multimodal sensor, and I th is the light intensity attenuation threshold; TOF j is the average photon flight time of the j-th multimodal sensor, and TOF th is the average photon flight time threshold.

5. The device according to claim 1, characterized in that, The voxel size is 16mm × 16mm × 16mm.

6. A multimodal sensor for detecting cerebral hemorrhage, characterized in that, The multi-modal sensor includes an MCU control module, a microwave signal transceiver module, a near-infrared light signal transceiver module, and a probe for fusing near-infrared and microwave; wherein: The MCU control module includes STM32; The microwave transceiver module includes a radio frequency transmission link and a radio frequency reception link. The radio frequency transmission link generates radio frequency by LMX2595, and after passing through QPA2211 and ADL5920, it is emitted from the patch electrode. After the radio frequency reception link receives the microwave signal from the patch electrode, it is transmitted to STM32 after passing through a low-noise amplifier, a mixer, and an analog-to-digital converter; The near-infrared signal transceiver module includes an optical signal generation optical path and an optical signal reception optical path. Among them, the optical signal generation optical path drives an LMH6521 high-speed constant current source to excite a multi-channel LP785-SAV50 laser diode to generate near-infrared light, which is emitted through a light source probe. The optical signal reception optical path is received by a silicon photomultiplier of S13360-3050CS of the light source probe, and is sent to the STM32 through a preamplifier OPA1612, a filter LTC1569, and an analog-to-digital converter. The microstrip of the microwave transceiver module and the silicon photomultiplier of the near-infrared signal transceiver module are coupled through a probe that combines near-infrared and microwave.

7. The multimodal sensor according to claim 6, characterized in that, The STM32 generates periodic TTL pulses. At the same time, the microwave transceiver module starts the radio frequency signal transmission and the near-infrared light source modulation of the near-infrared signal transceiver module. The analog-to-digital converter of the radio frequency reception link and the analog-to-digital converter of the near-infrared share the same trigger signal.

8. The multimodal sensor according to claim 6, characterized in that, The substrate of the probe that combines near-infrared and microwave is in a figure-eight shape. The size of the figure-eight substrate is 10mm×10mm×0.6mm. The outer ring of the figure-eight is a circular ground wire GND with a line width of 1mm, which is used to constrain the radio frequency and microwave signals. The side length of the middle ring is 7mm and the line width is 1mm, which is connected to the feeding position Feed. Openings are respectively made below the optical transmitter Opti cal for inserting a multimode optical fiber for light conduction, and openings are made below the SiPM at the receiving end for inserting the light source probe SiPM. An lPEX socket is used on the back of the probe, and a coaxial cable is used for the transmission of radio frequency and microwave signals.

9. The multimodal sensor according to claim 6, wherein, The wavelength range of the near-infrared signal transceiver module during operation is 650 - 950nm.

10. The multimodal sensor according to claim 6, wherein, The frequency band range of the microwave transceiver module during operation is 0.5 - 10GHz.

Images