Intelligent hemodynamics simulation platform based on deep learning

Abstract

The invention discloses an intelligent hemodynamics simulation platform based on deep learning, and relates to the technical field of medical instruments. The platform comprises a blood flow simulation device, a control system and a secondary bypass device, the blood flow simulation device is used for simulating a left descending branch artery and comprises a blood vessel pipeline, and a constant-temperature water tank for simulating blood supply in a human body temperature environment, a first motor for simulating a bridge blood vessel extrusion scene and a first pressure regulating valve for simulating myocardial bridge far-end vascular resistance which are sequentially arranged on the blood vessel pipeline; the secondary bypass device is used for simulating a coronary artery bypass transplantation CABG branch so as to bypass a myocardial bridge extrusion part and recover blood supply of ischemic myocardium; and the second-stage bypass device comprises a first water pump, a first compliance chamber and a second pressure regulating valve which are connected in sequence. The control system is integrated with a deep learning model, and driving signals of all parts of the simulation platform are generated through the deep learning model. According to the simulation platform, accurate simulation of treatment intervention of coronary artery bypass grafting (CABG) is realized.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to an intelligent hemodynamics simulation platform based on deep learning. Background Technology

[0002] Currently, cardiovascular disease is the leading cause of death among urban and rural residents, and about 40% of deaths from cardiovascular disease are caused by coronary artery myocardial bridging disease.

[0003] For patients with myocardial bridging who do not respond to drug therapy and have severe symptoms, coronary artery bypass grafting (CABG) is an important treatment option: a "bypass channel" is created by transplanting blood vessels (such as the internal thoracic artery or great saphenous vein) to bypass the myocardial segment compressed by the myocardial bridging and directly supply blood to the ischemic myocardium. However, current CABG surgical protocols and postoperative efficacy assessments are significantly inadequate.

[0004] Therefore, constructing a hemodynamic simulation platform capable of simulating CABG treatment intervention has become a critical technological need that urgently needs to be addressed. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent hemodynamic simulation platform based on deep learning to address the above-mentioned technical problems. This simulation platform can accurately simulate the treatment intervention of coronary artery bypass grafting (CABG).

[0006] The present invention adopts the following technical solution: This invention provides an intelligent hemodynamics simulation platform based on deep learning, comprising: a blood flow simulation device, a control system installed on the blood flow simulation device, and a secondary bypass device; wherein, the blood flow simulation device is used to simulate the left descending branch artery, including a blood vessel conduit, and a constant temperature water tank for simulating blood supply under human body temperature environment, a first motor for simulating the compression scenario of the bypass vessel, and a first pressure regulating valve for simulating the resistance of the distal blood vessel of the myocardial bridge, which are sequentially installed on the blood vessel conduit. The secondary bypass device is used to simulate the CABG branch in coronary artery bypass grafting to bypass the myocardial bridge compression site and restore blood supply to the ischemic myocardium. The secondary bypass device includes a first water pump, a first compliance chamber and a second pressure regulating valve connected in sequence. The second pressure regulating valve is connected to a constant temperature water tank and the first water pump is connected between the first motor and the first pressure regulating valve. The first water pump is used to simulate the additional hemodynamic source provided by the transplanted blood vessel after CABG surgery. The second pressure regulating valve is used to simulate the resistance regulation at the proximal connection point of the CABG transplanted blood vessel. The first compliance chamber is used to simulate the vessel wall compliance of CABG transplanted vessels; The control system integrates a deep learning model and analyzes the patient's pulse waveform and 3D geometric features of blood vessels through the deep learning model to control the operation of the first pressure regulating valve, the second pressure regulating valve, the first water pump, the first motor, and the first compliant chamber.

[0007] Optionally, the blood flow simulation device also includes a second compliant chamber disposed on the blood vessel, a third pressure regulating valve, a first flow sensor, a second water pump, a second flow sensor, a second motor, a pressure sensor disposed on the second compliant chamber, and a third flow sensor disposed on the first motor; The second compliance chamber is used to simulate the periodic compression experienced by the coronary artery segment proximal to the myocardial bridge; The third pressure regulating valve is used to simulate the resistance regulation of the proximal vessels of the myocardial bridge; The first flow sensor is used to simulate and measure blood flow proximal to the myocardial bridge. The second water pump is used to simulate the initial blood flow supply of the left descending branch artery; The second flow sensor is used to simulate and measure blood flow after passing through the myocardial bridge; The second motor is used to simulate compression of the coronary artery terminal to indirectly affect blood flow to the distal end of the myocardial bridge; A pressure sensor is used to simulate and monitor pressure changes in the first compliance chamber; The third flow sensor is used to simulate and measure blood flow in the bypass graft. The control system is also used to analyze the patient's pulse waveform and 3D geometric features of blood vessels through a deep learning model in order to control the operation of each water pump, each motor, each compliance chamber, and each pressure regulating valve, as well as to collect signals from the first flow sensor, the second flow sensor, the third flow sensor, and the air pressure sensor.

[0008] Optionally, the simulation platform also includes an output system, which includes a camera, a laser, a display, and a particle image velocimetry module; the camera and laser are positioned above the branch where the second motor is located; the blood vessel channel includes tracer particles; The particle image velocimetry module is used to control the camera and laser to synchronously acquire the trajectory of tracer particles in the blood flow simulation device, calculate the velocity distribution of the flow field based on the tracer particle trajectory, and send the velocity distribution of the flow field to the display. A display is used to show the velocity distribution of the flow field and the signals collected by the control system.

[0009] Optionally, the control system includes a deep learning workstation, a pump drive module, a data acquisition module, a motor driver, an air pump, and a relay; the deep learning workstation is connected to the pump drive module, the data acquisition module, the motor driver, the air pump, the relay, and the display, respectively; the deep learning model is built into the deep learning workstation; the deep learning workstation is equipped with an A6000 GPU; The deep learning workstation is used to generate periodic drive signals and input pulse waveforms and 3D geometric features of blood vessels into the deep learning model to generate drive signals for the first water pump, the second water pump, the first pressure regulating valve, the second pressure regulating valve, the third pressure regulating valve, the first compliant chamber, the air pump, and the relay, and send the drive signals to the corresponding components. The pump drive module is used to receive the pump drive signal sent by the deep learning workstation and adjust the speed of the first and second water pumps according to the pump drive signal. The motor driver is used to receive periodic drive signals and control valve drive signals sent by the deep learning workstation, and to drive the first motor and the second motor to run periodically through the periodic drive signals, and to drive the first pressure control valve, the second pressure control valve, the third pressure control valve and the first compliant chamber to run through the control valve drive signals. An air pump and a relay are used to receive drive signals from the air pump and relay sent by the deep learning workstation, and to adjust the air pressure in the second compliance chamber according to the drive signals from the air pump and relay. The data acquisition module is used to acquire signals sensed by the first flow sensor, the second flow sensor, the third flow sensor, and the air pressure sensor.

[0010] Optionally, the intelligent hemodynamics simulation platform based on deep learning also includes an input system, which includes a pulse waveform input terminal connected to a deep learning workstation. The pulse waveform input terminal is used to input the patient's pulse waveform and 3D geometric features of blood vessels into the deep learning workstation.

[0011] Optionally, the deep learning model includes a spectrum extraction module, a two-layer convolutional neural network, a Mamba encoder, a multimodal fusion module, a physical perception module, and a control decoder, which are connected in series. The multimodal fusion module is constructed by fusing temporal and spatial information based on cross-attention. The physical perception module is a PINNs network. The control decoder consists of multiple parallel fully connected layers. The deep learning workstation is used to input pulse waveforms and 3D geometric features of blood vessels into a deep learning model. The pulse waveform is converted into a Mel spectrum by a spectrum extraction module, then downsampled by a two-layer convolutional neural network and processed by a Mamba encoder for feature extraction, resulting in a high-dimensional temporal embedding. This high-dimensional temporal embedding and the 3D geometric features of the blood vessels are input to a multimodal fusion module to obtain a state vector. The state vector and the current state of the device are used as initial and boundary conditions, respectively, and input to a physical perception module to predict future states. A control decoder converts the future states into drive signals for multiple actuators, namely, the first water pump, the second water pump, the first pressure regulating valve, the second pressure regulating valve, the third pressure regulating valve, the first compliant chamber, the air pump, and the relay. The current state of the device includes the physical information of the pumps, regulating valves, and chambers; the future state represents the physical field variables at future moments.

[0012] Optionally, one fully connected layer in the control decoder corresponds to one actuator; Each fully connected layer is used to generate the corresponding actuator's drive signal based on the future state output by the physical sensing module.

[0013] Optionally, the loss function during the training of the deep learning model is constructed by weighting multiple losses; the multiple losses include waveform matching loss, physical feasibility loss, and smoothness efficiency loss; waveform matching loss represents the mean square error between the predicted waveform and the target waveform in the time dimension; physical feasibility loss includes the residuals of the Navier-Stokes equations calculated from the physical field variables predicted by the physical sensing module; and smoothness efficiency loss includes the first-order difference of the driving signal.

[0014] Optionally, the blood vessel is made of silicone material using 3D printing. The artificial blood flowing in the blood vessel is composed of a mixture of glycerol and water, with a viscosity range of 0.0035-0.004 Pa∙s and a density range of 1030-1060 kg / m3 at 37±1℃.

[0015] Optionally, the first water pump includes two water pumps connected in parallel.

[0016] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: The blood flow simulation device of this invention simulates the left anterior descending artery (including the myocardial bridging lesion segment). The first motor simulates the compression of the blood vessel during myocardial bridging contraction, corresponding to "myocardial systolic coronary artery stenosis" in clinical practice, which is the lesion site that CABG needs to bypass. In the secondary bypass device, the first water pump is connected between the first motor (compression site) and the first pressure regulating valve (distal end of the myocardial bridging), and finally connected to a constant temperature water tank, completely corresponding to clinical CABG. The bypass construction logic is as follows: the transplanted vessel is drawn from the "upstream" of the myocardial bridge compression site (after the first motor, i.e., the proximal end of the compressed segment of the myocardial wall coronary artery), bypassing the compression site, and directly connected to the "downstream" of the myocardial bridge (before the first pressure regulating valve, i.e., the distal end of the compressed segment of the myocardial wall coronary artery), achieving the treatment goal of "blood flow bypassing the lesion and directly reaching the ischemic myocardium". Furthermore, in the secondary bypass device, the first water pump provides additional blood flow dynamics to the ischemic myocardium, the first compliance chamber restores the compliance of the real transplanted vessel, and the second pressure regulating valve ensures that the simulated blood flow of the transplanted vessel is consistent with the real blood flow resistance after anastomosis in clinical practice, avoiding deviations in blood supply caused by the lack of simulation of proximal resistance, thereby improving the accuracy of the simulation platform in simulating the treatment intervention of coronary artery bypass grafting (CABG). Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0018] Figure 1 A schematic diagram of the process of an intelligent hemodynamics simulation platform based on deep learning provided by the present invention; Figure 2 A schematic diagram of another intelligent hemodynamics simulation platform based on deep learning provided by the present invention; Figure 3 A schematic diagram of another intelligent hemodynamics simulation platform based on deep learning provided by the present invention; Figure 4 This is a flowchart illustrating a deep learning model provided by the present invention.

[0019] Explanation of reference numerals in the attached figures: 1. Blood vessel tubing; 2. Constant temperature water tank; 3. First motor; 4. First pressure regulating valve; 5. First water pump; 6. First compliant chamber; 7. Second pressure regulating valve; 8. Control system; 9. Second compliant chamber; 10. Third pressure regulating valve; 11. First flow sensor; 12. Second water pump; 13. Second flow sensor; 14. Second motor; 15. Air pressure sensor; 16. Third flow sensor; 17. Camera; 18. Laser; 19. Display; 20. Deep learning workstation; 21. Pump drive module; 22. Data acquisition module; 23. Motor driver; 24. Air pump; 25. Relay; 26. Pulse waveform input terminal. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the term "comprising" and any variations thereof in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion. In the description of embodiments of this application, technical terms such as "first," "second," "third," and "fourth" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of embodiments of this application, "a plurality of" means two or more, unless otherwise explicitly defined.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] In the description of the embodiments of this application, unless otherwise expressly specified and limited, the technical term "connection" and other terms should be interpreted broadly. For example, it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal connection of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application according to the specific circumstances.

[0024] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Figure 1 This is a schematic diagram of the structure of an intelligent hemodynamics simulation platform based on deep learning in this invention. Specifically, it includes: a blood flow simulation device, a control system installed on the blood flow simulation device, and a secondary bypass device; wherein, the blood flow simulation device includes a blood vessel pipeline 1, a constant temperature water tank 2 installed on the blood vessel pipeline 1 to simulate blood supply under human body temperature, a first motor 3 to simulate the compression scenario of the bridge blood vessel, and a first pressure regulating valve 4 to simulate the resistance of the distal blood vessel of the myocardial bridge.

[0026] The secondary bypass device is used to simulate the CABG branch in coronary artery bypass grafting to bypass the myocardial bridge compression site and restore blood supply to the ischemic myocardium. The secondary bypass device includes a first water pump 5, a first compliance chamber 6 and a second pressure regulating valve 7 connected in sequence. The second pressure regulating valve 7 is connected to a constant temperature water tank 2, and the first water pump 5 is connected between the first motor 3 and the first pressure regulating valve 4.

[0027] The first water pump 5 is used to simulate the additional hemodynamic source provided by the transplanted blood vessel after CABG surgery.

[0028] The second pressure regulating valve 7 is used to simulate the resistance regulation at the proximal connection point of the CABG transplanted blood vessel.

[0029] The first compliance chamber 6 is used to simulate the vessel wall compliance of CABG transplanted vessels.

[0030] The control system 8 is used to integrate a deep learning model and analyze the patient's pulse waveform and 3D geometric features of blood vessels through the deep learning model in order to control the operation of the first pressure regulating valve 4, the second pressure regulating valve 7, the first water pump 5, the first motor 3, and the first compliant chamber 6.

[0031] The blood flow simulation device also includes a second compliant chamber 9, a third pressure regulating valve 10, a first flow sensor 11, a second water pump 12, a second flow sensor 13, a second motor 14, a pressure sensor 15 on the second compliant chamber, and a third flow sensor 16 on the first motor 3.

[0032] The second compliance chamber 9, the third pressure regulating valve 10, the first flow sensor 11, the second water pump 12, the constant temperature water tank 2, the second flow sensor 13, the second motor 14, the first pressure regulating valve 4 and the first motor 3 on the blood flow simulation device are connected in sequence.

[0033] The second compliance chamber 9 is used to simulate the periodic compression of the coronary artery segment proximal to the myocardial bridge; the third pressure regulating valve 10 is used to simulate the resistance regulation of the proximal vessel of the myocardial bridge; the first flow sensor 11 is used to simulate the measurement of blood flow proximal to the myocardial bridge; the second water pump 12 is used to simulate the initial blood flow supply of the left descending branch artery; the second flow sensor 13 is used to simulate the measurement of blood flow after passing through the myocardial bridge; the second motor 14 is used to simulate the compression of the coronary artery terminal to indirectly affect the blood flow distal to the myocardial bridge; the pressure sensor 15 is used to simulate monitoring the pressure changes in the first compliance chamber; and the third flow sensor 16 is used to simulate the measurement of blood flow in the bypass vessel.

[0034] The control system 8 is also used to analyze the patient's pulse waveform and 3D geometric features of blood vessels through a deep learning model in order to control the operation of each water pump, each motor, each compliance chamber, and each pressure regulating valve, as well as to collect signals from the first flow sensor 11, the second flow sensor 13, the third flow sensor 16, and the air pressure sensor 15.

[0035] The simulation platform also includes an output system, which comprises a camera 17, a laser 18, a display 19, and a particle image velocimetry module. The display 19 is connected to the control system 8, the camera 17, and the laser 18. The camera 17 and the laser 18 are positioned above the branch where the first motor 3 is located. The blood vessel channel 1 includes tracer particles. The display 19 is connected to the camera 17 and the laser 18 via the particle image velocimetry module. It should be noted that the particle image velocimetry module is not shown in the figure.

[0036] The particle image velocimetry module is used to control the camera 17 and the laser 18 to synchronously acquire the trajectory of tracer particles in the blood flow simulation device, calculate the velocity distribution of the flow field based on the tracer particle trajectory, and send the velocity distribution of the flow field to the display.

[0037] Display 19 is used to display the velocity distribution of the flow field and the signals acquired by the control system.

[0038] The constant temperature water tank 2 is placed in the constant temperature water bath to simulate the internal temperature environment of the human body.

[0039] Specifically, the blood vessel 1 is 3D printed using silicone material. The artificial blood flowing in the blood vessel 1 is composed of a mixture of glycerol and water, with a viscosity ranging from 0.0035 to 0.004 Pa∙s and a density ranging from 1030 to 1060 kg / m³ at 37±1℃. 3It is similar to blood.

[0040] The first flow sensor 11, the second flow sensor 13, and the third flow sensor 16 measure blood flow using ultrasonic non-contact technology and convert the blood flow into an electrical signal in real time, which is then transmitted to the deep learning workstation.

[0041] The laser 18 is a dual-pulse Nd:YAG laser (532nm, green light) with a cylindrical lens emitter, mounted below the side of the flow channel, at the distal end of the graft vessel and the native coronary artery. Camera 17 captures the behavior of tracer particles displayed by the laser emitted by laser 18, directly facing the laser beam direction and observing the cross-section perpendicularly. Both camera 17 and laser 18 are connected to a particle image velocimetry module, which controls the timing of laser 18 emission and the trigger time of camera 17 capture. This module typically provides TTL signals with nanosecond-level accuracy (camera 17 has microsecond-level accuracy) and is mounted externally on a support on one side of the blood flow simulation device.

[0042] Optionally, such as Figure 2 As shown, the control system 8 includes a deep learning workstation 20, a pump drive module 21, a data acquisition module 22, a motor driver 23, an air pump 24, and a relay 25. The deep learning workstation 20 is connected to the pump drive module 21, the data acquisition module 22, the motor driver 23, the air pump 24, the relay 25, and the display 19. The deep learning model is built into the deep learning workstation 20; the deep learning workstation 20 is equipped with an A6000 GPU.

[0043] The deep learning workstation 20 is used to generate periodic drive signals and input pulse waveforms and 3D geometric features of blood vessels into the deep learning model to generate drive signals for the first water pump 5, the second water pump 12, the first pressure regulating valve 4, the second pressure regulating valve 7, the third pressure regulating valve 10, the first compliance chamber 6, the air pump 24, and the relay 25, and sends the drive signals to the corresponding components.

[0044] The pump drive module 21 is used to receive the pump drive signal sent by the deep learning workstation 20, and adjust the speed of the first water pump 5 and the second water pump 12 according to the pump drive signal.

[0045] The motor driver 23 is used to receive the periodic drive signal and the control valve drive signal sent by the deep learning workstation 20, and drive the first motor 3 and the second motor 14 to run periodically through the periodic drive signal, and drive the first pressure control valve 4, the second pressure control valve 7, the third pressure control valve 10 and the first compliance chamber 6 to run through the control valve drive signal. The air pump 24 and the relay 25 are used to receive the drive signals of the air pump 24 and the relay 25 sent by the deep learning workstation 20, and to adjust the air pressure of the second compliance chamber 9 according to the drive signals of the air pump 24 and the relay 25.

[0046] The data acquisition module 22 is used to acquire signals sensed by the first flow sensor 11, the second flow sensor 13, the third flow sensor 16, and the air pressure sensor 15.

[0047] It should be noted that the first motor 3 and the second motor 14 are stepper motors, which respectively simulate the compression of the bypass vessels and the compression of the coronary artery terminal. This compression is periodic and pre-set, and is also driven by the deep learning workstation 20.

[0048] The pressure regulating valve is equipped with a motor (which can also be a stepper motor), but the motor stops after being compressed to a specified degree, simulating the resistance that blood flow encounters in the blood vessels. It does not cycle back and forth periodically, and the specific value of the specified degree is obtained by the deep learning model of the deep learning workstation 20.

[0049] exist Figure 1In this system, vascular conduit 1 simulates the main coronary artery, particularly the portion of the left descending artery (LAD) before compression of the myocardial bridge; the second compliance chamber 9 simulates the coronary artery segment proximal to the myocardial bridge, subjected to periodic compression; the third pressure regulating valve 10 simulates resistance regulation of the proximal vessels of the myocardial bridge; the first flow sensor 11 measures blood flow proximal to the myocardial bridge; the second water pump 12 serves as the main circuit, simulating the initial blood supply to the left descending artery (LAD); and the constant-temperature water tank 2 simulates a blood source maintained at body temperature (e.g., the heart or aorta). Second flow sensor 13: measures blood flow after passing through the myocardial bridge; Second motor 14: controls the first pressure regulating valve 4, indirectly affecting blood flow distal to the myocardial bridge; First pressure regulating valve 4: simulates resistance regulation of distal vessels of the myocardial bridge; The branch where the first motor 3 is located is the graft vessel, i.e., the sub-branch coronary artery, simulating the periodic compression scenario of the graft vessel (i.e., the transplanted vessel segment, such as the internal mammary artery or great saphenous vein), representing the distal connection point of the graft vessel and its physiological environment; Pressure sensor 15: monitors the second compliance chamber 9 ( Pressure changes at the proximal end of the myocardial bridge; third flow sensor 16: measures blood flow in the bypass vessel (branch where the first motor 3 is located); secondary bypass device, simulating the CABG branch; second pressure regulating valve 7: simulates resistance regulation at the proximal connection point of the CABG transplanted vessel (such as the para-aortic or internal mammary artery origin); first compliance chamber 6: simulates the vessel wall compliance of the CABG transplanted vessel; first pump 5: simulates an additional hemodynamic source, representing the direct blood flow provided by the transplanted vessel after CABG surgery, bypassing the myocardial bridge compression site, and delivering blood flow to the distal end of the bypass vessel or the distal end of the coronary artery. The branch where the first pump 5 is located is the CABG branch constructed through surgery. Through the transplanted vessel, blood flow from other blood supply sources (such as the aorta) is directly delivered to the distal end of the bypass vessel, thus requiring an additional pump to provide pressure; control system 8: simulates the heart's own neuroendocrine regulatory mechanism and the doctor's regulation of hemodynamics during and after surgery; output system (display 19): simulates the process by which the doctor obtains the patient's physiological parameters (such as blood pressure and blood flow) through monitoring equipment.

[0050] In the secondary bypass device: the second pressure regulating valve 7 is controlled by a motor to open and close, and the first compliance chamber 6 is controlled by an electromagnetic pump to inflate. The motor on the second pressure regulating valve 7 and the electromagnetic pump on the first compliance chamber 6 are both connected to the motor driver 23. The first water pump 5 is connected to the pump drive mode. All of them are controlled by the drive signals output by the deep learning model inside the deep learning workstation 20. The function of the secondary bypass device is to simulate the CABG procedure, that is, to bypass the myocardial bridge compression site and restore blood supply.

[0051] Optionally, the deep learning-based intelligent hemodynamics simulation platform also includes an input system, which includes a pulse waveform input terminal 26 connected to the deep learning workstation 20. The pulse waveform input terminal 26 is used to input the patient's pulse waveform and 3D vascular geometric features into the deep learning workstation 20.

[0052] The pulse waveform input terminal 26 provides a touchscreen, sampled data import, and preset data selection to support multiple data input formats. Specifically, the pulse waveform input terminal 26 supports input of various data structures and is suitable for different situations. For example, during debugging, preset standard pulse waveforms and standard vascular 3D geometric features can be directly selected, and the output of the data acquisition module 22 can be used to determine whether the deep learning-based intelligent hemodynamics simulation platform is working properly. When it is necessary to simulate the personalized blood flow of patients with specific conditions, the patient's arterial pressure waveform can be input in comma-separated values ​​(CSV) data sampling format. The deep learning-based intelligent hemodynamics simulation platform can automatically process and quickly adjust to achieve the output. In addition, the pulse waveform input terminal 26 of the deep learning-based intelligent hemodynamics simulation platform also supports touchscreen and image format input for general medical testing and experiments.

[0053] Two branches, also made of silicone tubing, are led out from the top of the second compliant chamber 9. Before measurement, both branches can be filled with water using a water pump. One branch connects to an air pump 24, which draws air from the outside to pressurize the water back (i.e., inflates the system). The other branch connects to a relay 25 and a solenoid valve (leading out from the branch on the compliant chamber; the solenoid valve is directly connected to the branch, and the other end of the branch leads to the atmosphere, essentially acting as a venting switch controlled by the relay 25). The relay 25 receives a small signal to control the opening and closing of the solenoid valve for venting. Because there is gas inside the second compliant chamber 9 during venting (different from the initial state), there is no water in the upper tube. Due to the air pressure, the solenoid valve opens to vent the gas.

[0054] Optionally, the second water pump 12 includes two water pumps connected in parallel; the second motor 14 and the first motor 3 are stepper motors, such as... Figure 3 As shown.

[0055] Optionally, the deep learning model is built into the deep learning workstation 20; the deep learning model includes a spectrum extraction module, a two-layer convolutional neural network, a Mamba encoder, a multimodal fusion module, a physical perception module, and a control decoder connected in series; the multimodal fusion module is constructed by fusing temporal and spatial information based on cross-attention; the physical perception module is a PINNs network; and the control decoder consists of multiple parallel fully connected layers.

[0056] The deep learning workstation 20 is used to convert the pulse waveform into a Mel spectrum, which is then downsampled by two layers of convolutional neural network and feature extracted by a Mamba encoder to obtain a high-dimensional temporal embedding. The high-dimensional temporal embedding and the 3D geometric features of blood vessels are input into a multimodal fusion unit to obtain a state vector. The state vector and the current state of the device are used as initial and boundary conditions to be input into the physical perception module to predict the future state. The future state is converted into drive signals for multiple actuators, namely the first water pump 5, the second water pump 12, the first pressure regulating valve 4, the second pressure regulating valve 7, the third pressure regulating valve 10, the first compliant chamber 6, the air pump 24, and the relay 25, through the control decoder. The current state of the device includes the physical information of the pumps, regulating valves, and chambers. The future state is the physical field variables at future moments.

[0057] The multimodal fusion unit employs a cross-attention feature concatenation structure, whereby the Mamba-encoded pulse waveform features interact with the vascular geometric features extracted by the GNN through a cross-attention module. In practice, this interaction is achieved simply through cross-attention.

[0058] Specifically, the deep learning workstation 20 incorporates a deep learning model. The workstation inputs the pulse waveform into the deep learning model for encoding and performs multimodal data fusion by combining it with patient-specific 3D vascular geometry reconstructed from CT / IVUS. The fused unified state vector, along with the current physical states of pumps, valves, and chambers, is then fed into a physical information neural network (PINNs) for learning under the constraints of the physical model. Furthermore, leveraging the differentiability of the entire framework, waveform matching loss, physical feasibility loss, and smoothness / efficiency loss are designed to predict the drive signals for the first pump 5, the second pump 12, the first pressure regulating valve 4, the second pressure regulating valve 7, the third pressure regulating valve 10, the first compliance chamber 6, the air pump 24, and the relay 25. Finally, the output of the platform is monitored in real-time using an online adaptive strategy. If there are deviations in the output, the PINNs are incrementally fine-tuned online using new data.

[0059] The pulse waveforms (target and real-time) are first converted to the Mel spectrum and then encoded using the Mamba architecture, which learns the waveform’s intrinsic patterns through self-supervised pre-training.

[0060] like Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of a deep learning model.

[0061] Optionally, in the deep learning model, the pulse waveform is interpolated before generating the Mel spectrum, and the Mel spectrum is generated based on the interpolated pulse waveform. The deep learning model samples the pulse waveform at a frequency of 1 kHz and uses cubic spline interpolation to fit the pulse waveform, supplementing the sampling point value by 16 kHz. This approach is consistent with traditional models, facilitating model transfer and construction. Furthermore, the interpolated pulse waveform provides more information for constructing the Mel spectrum, thus helping the model capture the internal structure of the data waveform.

[0062] Downsampling the Mel spectrum in the frequency domain using a convolutional neural network includes downsampling the Mel spectrum using two convolutional neural network layers.

[0063] In this control decoder, one fully connected layer corresponds to one actuator; each fully connected layer is used to generate the corresponding actuator's drive signal based on the future state output by the physical sensing module.

[0064] Optionally, the loss function during the training of the deep learning model is constructed by weighting multiple losses; the multiple losses include waveform matching loss, physical feasibility loss, and smoothness efficiency loss; waveform matching loss represents the mean square error between the predicted waveform and the target waveform in the time dimension; physical feasibility loss includes the residuals of the Navier-Stokes equations calculated from the physical field variables predicted by the physical sensing module; and smoothness efficiency loss includes the first-order difference of the driving signal.

[0065] The predicted waveform is the measured waveform (i.e., the signal collected by the flow sensor or pressure sensor) generated by the actual platform, driven by the predicted drive signal. The error loss between this waveform and the target waveform (human-set and expected to be achieved) is calculated.

[0066] The residuals of the Navier-Stokes equations are calculated by substituting the physical field variables predicted by the physical sensing module into the Navier-Stokes equations and then using automatic differentiation to calculate the residuals. The smaller the residuals, the more closely the predictions conform to physical laws. This method does not rely on global sensor measurements but is purely based on the physical constraints within the model.

[0067] Optionally, the physical feasibility loss includes the residuals of the Windkessel model calculated from the physical field variables predicted by the physical sensing module.

[0068] Optionally, the current differentiable physical states of the pump, valve, chamber, etc., can be directly used as input parameters or boundary conditions for PINNs.

[0069] The deep learning workstation 20 utilizes the pulse waveform and 3D geometric features of blood vessels emitted by the input system to adjust the drive signals of the second water pump 12, the third pressure regulating valve 10, the first pressure regulating valve 4, the air pump 24, and the relay 25 through a deep learning model. Based on the hemodynamic information (flow rate, pressure, and velocity, etc.) fed back by the data acquisition module 22, it optimizes the parameters in the deep learning model in real time. It should be noted that: firstly, the parameters here refer to the parameters of each component in the system; secondly, the optimization is usually completed very quickly (on the order of seconds in actual measurements), after which the behavior of each regulating component remains unchanged, and measurement work continues. It is not continuously optimized; it adjusts and measures simultaneously.

[0070] The entire deep learning-based intelligent hemodynamics simulation platform is adjusted using negative feedback regulation, with a mechanism of coarse adjustment followed by fine adjustment. The deep learning workstation 20 recalculates the parameters of each component and adjusts them rapidly based on the waveform data obtained in the previous step (pressure and flow change over time; they are waves, and waveform refers to the waveforms of these two). This process is repeated multiple times to approximate the desired waveform.

[0071] Optionally, the output system also includes a particle image velocimetry module, which is connected to the camera 17 and the laser 18 respectively; the particle image velocimetry module is used to control the camera 17 and the laser 18 to achieve synchronous acquisition.

[0072] Specifically, secondary bypass devices can simulate different lesions and surgical procedures, such as myocardial bridging and CABG procedures, to provide reference for physician training and surgical simulation. Secondary bypass devices include proximal and distal anastomoses to simulate CABG procedures, and parallel multi-motor connections at the drive site to achieve a biphasic pulse waveform. This involves two types of secondary devices. Proximal and distal anastomoses in CABG procedures constitute secondary bypass devices, but parallel multi-motor connections are only considered secondary devices, not bypasses. Firstly, proximal and distal anastomoses in CABG procedures are two crucial points connecting the transplanted vessel to the coronary artery system; the CABG procedure is achieved through their connection. Proximal anastomosis refers to the connection between the transplanted vessel and the blood supply source, i.e. Figure 1 The branch line leading out from the upper end of the constant temperature water tank 2; distal anastomosis refers to the connection between the transplanted blood vessel and the compressed part of the coronary artery, i.e. Figure 1 The branch circuit extends from the right side of the first motor 3, below the camera 17. Secondly, the parallel connection of multiple motors refers to the possibility of replacing the second motor 14 with multiple motors in parallel during actual platform use. This is because, in the actual measurements of this invention, due to the lag in motor response, the simulation of the two waveforms of the pulse (high-pressure peak and low-pressure peak, also known as the contraction peak and the dilation peak) using the signal superposition of a single motor is not accurate enough. Therefore, accurate simulation can be achieved by simulating different peaks separately using multiple motors in parallel. This is an alternative solution, hence the name "secondary device."

[0073] This invention proposes using multiple pumps connected in parallel as a secondary device. On the one hand, waveform decomposition of the pulse waveform allows each pump to process a simpler waveform output, reducing response delays and accuracy losses caused by complex changes in rotation speed. On the other hand, the pulse wave itself has two peaks, a systolic and a diastolic phase, within a cycle, which is naturally beneficial for superposition processing.

[0074] The second compliant chamber 9 of this invention is connected to an air pump 24 (electromagnetic air pump), a pressure sensor 15, and a relay 25. The air pump 24 enables precise inflation, while the relay 25 controls the opening of the inflation / deflation port. The pressure sensor 15 connected to the second compliant chamber 9 can monitor the working status of the second compliant chamber 9 in real time. When the internal pressure is too high, it inputs a signal to the deep learning workstation 20 to control the relay 25 to force deflation, thus preventing excessive pressure from damaging the deep learning-based intelligent hemodynamics simulation platform and improving the safety factor of the deep learning-based intelligent hemodynamics simulation platform.

[0075] In this invention, after the input system acquires the pulse waveform and 3D geometric features of the blood vessels, it transmits them to the deep learning workstation 20. The deep learning workstation 20 quickly obtains the drive signals of each component in the blood flow simulation device through its built-in trained deep learning model, and transmits the drive signals to the pump drive module 21, motor driver 23, air pump 24, and relay 25 respectively. The pump drive module 21 adjusts the speed of the water pump, and the motor driver 23 drives the second motor 14 and the first motor 3 to adjust the opening and closing degrees of the first and second pressure sensors respectively. The air pump 24 pumps air or the relay 25 switches the solenoid valve to change the air pressure in the second compliance chamber 9, thereby implementing adjustments to different parts of the blood flow simulation device. At the same time, the data acquisition module 22 collects the parameters of each part in real time and returns them to the deep learning workstation 20 for negative feedback adjustment. After the adjustment is completed, a simulation under a specific scenario can be performed, and the blood dynamic parameters are displayed on the display 19 of the output system. After the adjustment is completed, the output of each sensor can be directly measured. This is because the completion of adjustment means that the flow waveform measured by the first flow sensor 11 is consistent with the patient's waveform. The pressure sensor 15 is directly connected to the second compliance chamber 9 to measure the air pressure above the simulated blood inside the second compliance chamber 9.

[0076] In this invention, the surgical / lesion site can be designed and adjusted according to specific circumstances, as referred to... Figure 1The lesion site is simulated (myocardial bridging), that is, the branch where the first motor 3 is located, which is physiologically a bypass vessel. The periodic compression of the myocardial bridging is simulated by the first motor 3, and the flow rate of the bypass vessel is measured by the parallel flow sensor. The flow rate of the myocardial bridging branch is obtained by subtracting the flow rate of the bypass vessel (the flow rate collected by the third flow sensor 16) from the total flow rate (the flow rate collected by the first flow sensor 11).

[0077] The intelligent hemodynamic simulation platform based on deep learning provided by this invention processes time-series pulse waveform data using a deep learning model, enabling analysis of data features in both the time and frequency domains. The trained model is integrated into a deep learning workstation 20, allowing for real-time digital control of motors and pumps to simulate coronary blood flow for specified pulse data or individual patient characteristics. Furthermore, this platform utilizes a multimodal deep learning model to construct an end-to-end differentiable system, integrating a physical model into deep learning. The desired output is obtained through goal-driven inverse optimization and online adaptive rapid adjustment of the platform.

[0078] This invention can arbitrarily preset ideal physiological waveforms and quickly adjust various parameters of the platform through a pre-trained deep learning model to achieve the expected output; it can also simulate different surgical interventions and lesion sites through the control system 8 and provide real-time feedback on the hemodynamic parameters in the platform, providing a reference for physicians in treatment and surgery.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. An intelligent hemodynamics simulation platform based on deep learning, characterized in that, include: Blood flow simulation device, control system installed on blood flow simulation device, and secondary bypass device; wherein, blood flow simulation device is used to simulate left descending branch artery, including blood vessel conduit, and a constant temperature water tank for simulating blood supply under human body temperature environment, a first motor for simulating the compression scenario of bypass blood vessel, and a first pressure regulating valve for simulating the resistance of distal blood vessel of myocardial bridge, which are sequentially installed on blood vessel conduit. The secondary bypass device is used to simulate the CABG branch in coronary artery bypass grafting to bypass the myocardial bridge compression site and restore blood supply to the ischemic myocardium; The secondary bypass device includes a first water pump, a first compliant chamber, and a second pressure regulating valve connected in sequence; the second pressure regulating valve is connected to a constant temperature water tank, and the first water pump is connected between a first motor and the first pressure regulating valve; The first water pump is used to simulate the additional hemodynamic source provided by the transplanted blood vessel after CABG surgery. The second pressure regulating valve is used to simulate the resistance regulation at the proximal connection point of the CABG transplanted blood vessel. The first compliance chamber is used to simulate the vessel wall compliance of CABG transplanted vessels; The control system integrates a deep learning model and analyzes the patient's pulse waveform and 3D geometric features of blood vessels through the deep learning model to control the operation of the first pressure regulating valve, the second pressure regulating valve, the first water pump, the first motor, and the first compliant chamber.

2. The intelligent hemodynamics simulation platform based on deep learning according to claim 1, characterized in that, The blood flow simulation device also includes a second compliant chamber, a third pressure regulating valve, a first flow sensor, a second water pump, a second flow sensor, a second motor, a pressure sensor installed on the second compliant chamber, and a third flow sensor installed on the first motor. The second compliance chamber is used to simulate the periodic compression experienced by the coronary artery segment proximal to the myocardial bridge; The third pressure regulating valve is used to simulate the resistance regulation of the proximal vessels of the myocardial bridge; The first flow sensor is used to simulate and measure blood flow proximal to the myocardial bridge. The second water pump is used to simulate the initial blood flow supply of the left descending branch artery; The second flow sensor is used to simulate and measure blood flow after passing through the myocardial bridge; The second motor is used to simulate compression of the coronary artery terminal to indirectly affect blood flow to the distal end of the myocardial bridge; A pressure sensor is used to simulate and monitor pressure changes in the first compliance chamber; The third flow sensor is used to simulate and measure blood flow in the bypass graft. The control system is also used to analyze the patient's pulse waveform and 3D geometric features of blood vessels through a deep learning model in order to control the operation of each water pump, each motor, each compliance chamber, and each pressure regulating valve, as well as to collect signals from the first flow sensor, the second flow sensor, the third flow sensor, and the air pressure sensor.

3. The intelligent hemodynamics simulation platform based on deep learning according to claim 2, characterized in that, The simulation platform also includes an output system, which includes a camera, a laser, a display, and a particle image velocimetry module; the camera and laser are located above the branch where the second motor is located; the blood vessel channel includes tracer particles; The particle image velocimetry module is used to control the camera and laser to synchronously acquire the trajectory of tracer particles in the blood flow simulation device, calculate the velocity distribution of the flow field based on the tracer particle trajectory, and send the velocity distribution of the flow field to the display. A display is used to show the velocity distribution of the flow field and the signals collected by the control system.

4. The intelligent hemodynamics simulation platform based on deep learning according to claim 3, characterized in that, The control system includes a deep learning workstation, a pump drive module, a data acquisition module, a motor driver, an air pump, and a relay; the deep learning workstation is connected to the pump drive module, the data acquisition module, the motor driver, the air pump, the relay, and the display; the deep learning model is built into the deep learning workstation. The deep learning workstation is equipped with an A6000 GPU; The deep learning workstation is used to generate periodic drive signals and input pulse waveforms and 3D geometric features of blood vessels into the deep learning model to generate drive signals for the first water pump, the second water pump, the first pressure regulating valve, the second pressure regulating valve, the third pressure regulating valve, the first compliant chamber, the air pump, and the relay, and send the drive signals to the corresponding components. The pump drive module is used to receive the pump drive signal sent by the deep learning workstation and adjust the speed of the first and second water pumps. The motor driver is used to receive periodic drive signals from the motor and drive signals from the regulating valve sent by the deep learning workstation, and to drive the first motor and the second motor to run periodically through the periodic drive signals, and to drive the first pressure regulating valve, the second pressure regulating valve, the third pressure regulating valve and the first compliant chamber to run through the drive signals from the regulating valve. An air pump and a relay are used to receive drive signals from the air pump and relay sent by the deep learning workstation, and to adjust the air pressure in the second compliance chamber according to the drive signals from the air pump and relay. The data acquisition module is used to acquire signals sensed by the first flow sensor, the second flow sensor, the third flow sensor, and the air pressure sensor.

5. The intelligent hemodynamics simulation platform based on deep learning according to claim 4, characterized in that, The intelligent hemodynamics simulation platform based on deep learning also includes an input system, which includes a pulse waveform input terminal connected to a deep learning workstation. The pulse waveform input terminal is used to input the patient's pulse waveform and 3D geometric features of blood vessels into the deep learning workstation.

6. The intelligent hemodynamics simulation platform based on deep learning according to claim 4, characterized in that, The deep learning model consists of a spectrum extraction module, a two-layer convolutional neural network, a Mamba encoder, a multimodal fusion module, a physical perception module, and a control decoder, all connected in series. The encoder uses a Mamba architecture. The multimodal fusion module is constructed by fusing temporal and spatial information based on cross-attention. The physical perception module is a PINNs network. The control decoder consists of multiple parallel fully connected layers. The deep learning workstation is used to input pulse waveforms and 3D geometric features of blood vessels into a deep learning model. The pulse waveform is converted into a Mel spectrum by a spectrum extraction module, then downsampled by a two-layer convolutional neural network and processed by a Mamba encoder for feature extraction, resulting in a high-dimensional temporal embedding. This high-dimensional temporal embedding and the 3D geometric features of the blood vessels are input to a multimodal fusion module to obtain a state vector. The state vector and the current state of the device are used as initial and boundary conditions, respectively, and input to a physical perception module to predict future states. A control decoder converts the future states into drive signals for multiple actuators, namely, the first water pump, the second water pump, the first pressure regulating valve, the second pressure regulating valve, the third pressure regulating valve, the first compliant chamber, the air pump, and the relay. The current state of the device includes the physical information of the pumps, regulating valves, and chambers; the future state represents the physical field variables at future moments.

7. The intelligent hemodynamics simulation platform based on deep learning according to claim 4, characterized in that, One fully connected layer in the control decoder corresponds to one actuator; Each fully connected layer is used to generate the corresponding actuator's drive signal based on the future state output by the physical sensing module.

8. The intelligent hemodynamics simulation platform based on deep learning according to claim 4, characterized in that, The loss function in the deep learning model training process is constructed by weighting multiple losses; the multiple losses include waveform matching loss, physical feasibility loss and smoothness efficiency loss; waveform matching loss represents the mean square error between the predicted waveform and the target waveform in the time dimension; physical feasibility loss includes the residual of the Navier-Stokes equation calculated from the physical field variables predicted by the physical sensing module; smoothness efficiency loss includes the first-order difference of the driving signal.

9. The intelligent hemodynamics simulation platform based on deep learning according to claim 1, characterized in that, The blood vessel tubes are 3D printed using silicone material. The artificial blood flowing in the blood vessel tubes is composed of a mixture of glycerol and water. At 37±1℃, the viscosity of the mixture ranges from 0.0035 to 0.004 Pa∙s, and the density ranges from 1030 to 1060 kg / m3.

10. The intelligent hemodynamics simulation platform based on deep learning according to claim 1, characterized in that, The first water pump comprises two water pumps connected in parallel.

Images