An intelligent injection and aspiration control system for ophthalmic surgery and its usage method

Abstract

This invention relates to an intelligent injection and aspiration control system for ophthalmic surgery and its usage method, belonging to the field of medical device technology. It includes an intelligent host system, which integrates a main control circuit board, a main aspiration pump, and a high-speed servo-assisted compensating infusion pump; a dual-channel intelligent needle, with a main aspiration channel and a compensating infusion channel arranged in parallel within its handle, and its ends connected to an aspiration needle tube and a compensating infusion tube, respectively; and an image processing system, comprising an image acquisition unit and an image processing unit, the image processing unit including a multimodal sensor fusion module and a feedforward predictive control module. This invention can proactively prevent anterior chamber collapse and surge effects, realizing a shift from passive response to predictive intervention, effectively improving intraoperative anterior chamber stability and surgical safety.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, specifically to an intelligent injection and aspiration control system for ophthalmic surgery and its usage method. Background Technology

[0002] In phacoemulsification cataract surgery, the infusion and aspiration phase is one of the core steps for surgical success. During the procedure, the phacoemulsification needle breaks down the cloudy lens nucleus into an emulsified state, which is then removed along with the cortex using an infusion and aspiration system. Simultaneously, a balanced salt solution is continuously infused into the eye to maintain the anterior chamber's spatial morphology. As a critical operating area in intraocular surgery, the stability of the anterior chamber's depth and volume directly affects the integrity of corneal endothelial cells, the safety of the posterior capsule, and the precision of manipulation within the lens bag. Abnormal collapse or severe fluctuations in the anterior chamber during infusion and aspiration can easily lead to serious complications such as mechanical damage to corneal endothelial cells, accidental rupture of the posterior capsule, or even vitreous prolapse, causing irreversible damage to the patient's postoperative visual quality. Therefore, how to efficiently remove lens fragments while maintaining a stable anterior chamber morphology remains a core concern in the field of ophthalmic surgical instruments.

[0003] Currently, the widely used infusion-aspiration control technology in clinical practice is generally based on a combined architecture of a single-channel infusion-aspiration needle and an external negative pressure pump system. Its working principle is to simultaneously achieve passive inflow of the perfusion fluid and active aspiration of the fragment mixture through a single-channel needle, and to indirectly monitor changes in negative pressure within the aspiration tubing using a pressure sensor installed in the tubing, thereby inferring the pressure state within the anterior chamber. When the negative pressure within the tubing increases, the system assumes that the anterior chamber volume may have decreased due to excessive aspiration, and then passively adjusts by regulating the perfusion flow rate or reducing the negative pressure value. This technical solution has been widely used in phacoemulsification surgery in medical institutions at all levels due to its relatively simple structure and low operational threshold. However, this control method, which indirectly infers the anterior chamber state based on tubing pressure, has gradually revealed significant limitations in clinical practice. Because the pressure sensor can only monitor the fluid pressure inside the tubing, it cannot directly perceive the actual physical morphological parameters of the anterior chamber, such as real-time changes in anterior chamber depth, the trend of volume increase or decrease, and the rate of change, thus creating a fundamental "information blind spot" in the control chain. By the time changes in pipeline pressure are transmitted along the fluid path to the sensor and detected by the system, anterior chamber collapse or morphological abnormalities have often already occurred. The control system can only provide a remedial response after the fact, and cannot achieve proactive preventative intervention. This lag in response means that corneal endothelial cells may suffer short-term exposure or mechanical friction due to the instantaneous shallowing of the anterior chamber, and the posterior capsule may also face the risk of rupture due to insufficient support of the anterior chamber.

[0004] Furthermore, existing technologies have inherent mechanistic flaws in addressing aspiration port blockage. When lens nucleus fragments or cortical debris block the aspiration port of the infusion needle, the negative pressure within the aspiration tubing accumulates rapidly. To clear the blockage, conventional systems typically increase the negative pressure automatically or manually by the surgeon. However, once the blockage is instantly removed, the high negative pressure accumulated within the tubing is abruptly released, creating a violent "surge effect." This causes a large amount of fluid to be drawn from the anterior chamber in a very short time, resulting in a rapid decrease in anterior chamber depth or even instantaneous collapse. This sudden surge creates a strong impact suction on the iris, lens capsule, and anterior vitreous interface, significantly increasing the risk of posterior capsule rupture and vitreous prolapse. Although some improvements attempt to integrate miniature optical sensors into the infusion needle or tubing to directly monitor the anterior chamber status, such designs not only significantly increase the manufacturing cost of disposable surgical consumables, but the electronic components and optical parts in the sensors are also difficult to withstand the high-temperature, high-pressure sterilization processes required by clinical standards, failing to meet the stringent sterility requirements of ophthalmic surgical instruments, thus hindering their widespread application in clinical practice. How to achieve direct, real-time, and accurate perception of the anterior chamber status without compromising the existing aseptic operating system, and how to construct a predictive active intervention mechanism based on this perception, has become a key technical problem that urgently needs to be solved in this field, under the premise of continuously and significantly increasing the cost of disposable consumables. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose an intelligent injection and aspiration control system for ophthalmic surgery. This system comprises an intelligent host system consisting of a main aspiration pump, a high-speed servo-compensated irrigation pump, and a main control circuit board with an integrated NPU chip; a dual-channel intelligent needle consisting of parallel fixed aspiration and compensation irrigation tubes, and miniature force and pressure sensors; an image processing system consisting of an image analysis module, a multimodal sensing fusion module, a feedforward predictive control module, an adaptive learning module, and a surgical parameter library; and a safety redundancy mechanism that automatically switches to pressure control mode when image signals are abnormal. This addresses the problems in existing technologies, such as information blind spots leading to delayed response in anterior chamber stability control, surge effects causing impact damage to intraocular tissues when blockage occurs, the significant conflict between the cost of disposable consumables and integrated sensors and sterility requirements, and the lack of individualized adaptability and predictive intervention capabilities in control strategies. The system achieves multimodal real-time perception of anterior chamber status, feedforward predictive intervention based on volume change trends, individualized adaptive matching of control parameters, and continuous incremental optimization of postoperative models, thereby significantly improving anterior chamber stability and surgical safety during the injection and aspiration process.

[0006] This invention is achieved through the following technical solution:

[0007] An intelligent injection and aspiration control system for ophthalmic surgery includes an intelligent host system, a dual-channel intelligent needle, and an image processing system;

[0008] The intelligent host system includes a casing, a power module, a main control circuit board, a main suction pump, and a high-speed servo-compensated infusion pump. The power module is fixedly disposed inside the bottom of the casing, the main control circuit board is fixedly disposed above the power module, the main suction pump is fixedly disposed on the upper left side of the main control circuit board, and the high-speed servo-compensated infusion pump is fixedly disposed on the upper right side of the main control circuit board. A suction port is fixedly disposed on the left side of the front of the casing, and a compensation infusion port is fixedly disposed on the right side of the front of the casing. The output end of the main suction pump is sealed to the suction port through a fluid pipeline, and the output end of the high-speed servo-compensated infusion pump is sealed to the compensation infusion port through a fluid pipeline. Both the main suction pump and the high-speed servo-compensated infusion pump are connected to the main control circuit board through control signal lines, and the power module is connected to the main control circuit board, the main suction pump, and the high-speed servo-compensated infusion pump through power lines.

[0009] The dual-channel smart needle includes a needle handle, a Luer connector, and a needle assembly. The Luer connector is fixedly disposed at the tail of the needle handle and includes an aspiration Luer connector and a compensation infusion Luer connector. The aspiration Luer connector is detachably connected to the aspiration interface via a fluid conduit, and the compensation infusion Luer connector is detachably connected to the compensation infusion interface via a fluid conduit. A main aspiration channel and a compensation infusion channel are arranged parallel to each other along the length direction inside the needle handle. One end of the main aspiration channel is connected to the aspiration Luer connector, and the other end of the main aspiration channel is fixedly connected to an aspiration needle tube. One end of the compensation infusion channel is connected to the compensation infusion Luer connector, and the other end of the compensation infusion channel is fixedly connected to a compensation infusion tube. The proximal ends of the aspiration needle tube and the compensation infusion tube are fixedly formed by laser welding to form the needle assembly.

[0010] The image processing system includes an image acquisition unit and an image processing unit. The image acquisition unit is connected to the image processing unit via a data cable, and the image processing unit is connected to the main control circuit board via a gigabit network cable. The image acquisition unit is fixed to the side arm of the surgical microscope via a universal bracket, and the lens axis of the image acquisition unit is coaxial with the optical path of the microscope and aligned with the anterior chamber area.

[0011] The image processing unit includes an image analysis module, a multimodal sensing fusion module, and a feedforward predictive control module. The multimodal sensing fusion module is used to fuse in real time the anterior chamber visual feature data acquired by the image acquisition unit, the tubing pressure data acquired by the pressure sensor located inside the aspiration interface, and the contact force data between the needle and intraocular tissue acquired by the miniature force sensor located on the dual-channel smart needle, and generate an anterior chamber state fusion vector. The feedforward predictive control module is used to receive the anterior chamber state fusion vector, generate a predicted value of the anterior chamber state at future times using a trained LSTM neural network model, and send a feedforward control signal to the main control circuit board before the predicted value of the anterior chamber state reaches a preset threshold. The main control circuit board adjusts the operating parameters of the main aspiration pump and the high-speed servo-compensated perfusion pump in advance based on the feedforward control signal.

[0012] Furthermore, the main control circuit board includes a central processing unit, a drive module, and an NPU chip. The central processing unit is connected to the drive module via control signal lines, and the drive module is connected to the main suction pump and the high-speed servo-compensated injection pump via control signal lines respectively. The NPU chip is used to run the forward inference calculation of the LSTM neural network model.

[0013] Furthermore, the image processing unit also includes a data conversion module and an adaptive learning module; the data conversion module is connected to the main control circuit board via the gigabit network cable; the adaptive learning module is used to extract multimodal sensor data sequences and control action sequences after surgery, and update the parameters of the LSTM neural network model through an incremental learning algorithm.

[0014] Furthermore, the outlet end of the compensating infusion tube is located obliquely above the suction needle tube, and the outlet axis of the compensating infusion tube forms an acute angle with the axis of the suction needle tube. The outlet of the compensating infusion tube faces the corneal endothelium. The micro force sensor and the micro pressure sensor are fixedly installed on the needle handle or the needle tip assembly. The micro pressure sensor is used to monitor the local infusion pressure of the anterior chamber in real time. Both the micro force sensor and the micro pressure sensor are connected to the main control circuit board via embedded circuitry.

[0015] Furthermore, a pressure sensor is fixedly installed inside the suction port. The detection end of the pressure sensor is connected to the fluid channel of the suction port, and the signal output end of the pressure sensor is connected to the main control circuit board via a control signal line. A solenoid valve is fixedly installed inside the compensation infusion port. The valve body of the solenoid valve is connected in series in the fluid channel of the compensation infusion port, and the control end of the solenoid valve is connected to the main control circuit board via a control signal line. An LCD touch screen is also fixedly installed on the front of the outer casing. The LCD touch screen is used to display the current anterior chamber depth, volume change rate, predicted trend curve, tubing pressure, and tissue contact force in real time.

[0016] Furthermore, the image processing unit also includes a surgical parameter library, which pre-stores optimized control parameter sets corresponding to different patients' anterior chamber characteristics; the image analysis module is used to extract the patient's anterior chamber characteristics before surgery and match them with the surgical parameter library, and the main control circuit board is used to load the matched optimized control parameter sets as initial control parameters.

[0017] Furthermore, the system control method includes the following steps:

[0018] S1: Preoperative calibration, the image acquisition unit acquires anterior chamber images, and the image processing unit calculates the camera intrinsic parameter matrix and determines the baseline anterior chamber depth and baseline volume;

[0019] S2: Intraoperative real-time perception and fusion, the image acquisition unit acquires anterior chamber images in real time, while the pressure sensor and the miniature force sensor acquire tubing pressure and needle contact force respectively; the multimodal sensing fusion module fuses anterior chamber visual feature data, tubing pressure data and needle contact force data to generate a fused state vector containing anterior chamber depth, volume change rate, pressure change trend and contact force.

[0020] S3: Feedforward prediction and intervention. The feedforward prediction control module receives the fused state vector and predicts the future trend of anterior chamber volume change through an LSTM neural network model. When the predicted value indicates that anterior chamber collapse or surge is about to occur, the feedforward prediction control module sends a feedforward control signal to the main control circuit board before the corresponding event actually occurs. The main control circuit board adjusts the negative pressure of the main suction pump and the infusion flow rate of the high-speed servo compensated infusion pump in advance according to the feedforward control signal.

[0021] S4: Segmented closed-loop control. When the predicted volume change rate is within the first preset range, the main control circuit board controls the high-speed servo compensated infusion pump to output a preventive infusion flow rate. When the predicted volume change rate drops to the second preset range, the main control circuit board increases the infusion flow rate of the high-speed servo compensated infusion pump according to the compensation flow rate formula. When the predicted volume change rate suddenly drops to the third preset range, the main control circuit board increases the negative pressure of the main suction pump in advance and controls the high-speed servo compensated infusion pump to perform pulse infusion.

[0022] S5: Safety redundancy and adaptive learning. When the image processing unit fails to extract the complete contour for multiple consecutive frames, the main control circuit board switches to the pressure control mode based on the pressure sensor. After surgery, the adaptive learning module extracts and analyzes the surgical data and incrementally updates the LSTM neural network model.

[0023] Furthermore, in S2, the image processing unit calculates the current anterior chamber depth and volume change rate in real time using a stereo vision algorithm and a frustum volume formula. Specifically, it performs stereo matching on the corrected left and right camera images using a semi-global block matching algorithm to generate a disparity map, calculates the anterior chamber depth based on the disparity map, and calculates the anterior chamber volume according to the frustum model.

[0024] Furthermore, in S4, the compensation flow formula is as follows: ;

[0025] in, To compensate for the injection flow rate, The predicted rate of change of volume. This represents the change in foreroom volume per unit time. The sampling time interval is represented by k, which is the compensation coefficient corresponding to the patient's anterior chamber characteristics matched according to the surgical parameter library.

[0026] Furthermore, in S5, when the contact force detected by the micro force sensor exceeds a preset force threshold and the duration exceeds a preset time threshold, the main control circuit board reduces the suction negative pressure of the main suction pump to a safe value and issues an alarm signal.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention employs an intelligent infusion and aspiration control system that integrates real-time anterior chamber image perception and multimodal data fusion, dual-channel precise infusion and aspiration execution and directional compensation, and feedforward predictive control and adaptive learning. Through an image processing unit, it directly monitors the anterior chamber morphology and fuses tubing pressure and tissue contact force information, eliminating information blind spots inherent in traditional indirect monitoring and enabling a shift from passive response to proactive prevention. By integrating a miniature force sensor and a directional infusion outlet on the dual-channel intelligent needle, it mitigates the surge effect of blockage relief and avoids mechanical damage to intraocular tissues. The synergy of feedforward predictive control and a safety redundancy mechanism allows for seamless switching to pressure control mode in case of image abnormalities, ensuring uninterrupted surgical procedures. This significantly improves anterior chamber stability, reduces the risk of intraoperative complications, greatly enhances surgical safety and postoperative visual quality, and alleviates the operational burden and psychological stress on medical staff. Attached Figure Description

[0029] Figure 1 For the overall assembly structure drawing;

[0030] Figure 2 This is a diagram of the internal structure of the intelligent host system;

[0031] Figure 3 This is a diagram of the internal structure of a dual-channel smart needle.

[0032] Figure 4 Overall architecture diagram of intelligent injection and aspiration control system for ophthalmic surgery;

[0033] Figure 5 This is a diagram showing the internal modules and data flow of the image processing unit.

[0034] Figure 6 Here is a flowchart of feedforward prediction and piecewise closed-loop control;

[0035] Figure 7 Diagram of safety redundancy and adaptive learning mechanism.

[0036] Explanation of reference numerals in the attached figures:

[0037] 1. Intelligent host system; 2. Dual-channel intelligent needle; 3. Image processing system; 11. Housing; 12. Suction interface; 14. Compensation infusion interface; 15. Main suction pump; 16. High-speed servo compensation infusion pump; 17. Main control circuit board; 171. NPU chip; 18. Power module; 19. LCD touch screen; 21. Needle assembly; 211. Suction needle tube; 212. Compensation infusion tube; 22. Needle handle; 23. Main suction channel; 24. Compensation infusion channel; 25. Luer connector; 251. Suction Luer connector; 252. Compensation infusion Luer connector; 26. Miniature force sensor; 27. Miniature pressure sensor; 31. Image acquisition unit; 32. Image processing unit; 33. Gigabit network cable; 34. Data cable. Detailed Implementation

[0038] like Figures 1 to 7 As shown, this embodiment provides an intelligent injection and aspiration control system for ophthalmic surgery, including an intelligent host system 1, a dual-channel intelligent needle 2, and an image processing system 3. The intelligent host system 1 is detachably connected to the dual-channel intelligent needle 2 via a fluid pipeline and is signal-connected to the image processing system 3 via a gigabit network cable 33. The image processing system 3 is connected to its internal image acquisition unit 31 via a data cable 34. The three subsystems are arranged independently in physical space but are synergistically coupled in the signal and fluid pathways, jointly constructing a complete closed loop of anterior chamber multimodal real-time perception, feedforward predictive control, and adaptive learning.

[0039] The outer casing 11 of the intelligent host system 1 is a box made of medical-grade metal sheet through laser cutting and bending processes. Inside the casing 11, the power module 18 is fixed to the base plate at the bottom. The power module 18 is a switching power supply structure, with its input connected to AC power and its output providing multiple DC regulated outputs, supplying power to the upper-level electrical components via power lines. A main control circuit board 17 is horizontally suspended directly above the power module 18, and is fixedly connected to the casing 11, forming a heat dissipation duct below it. The main control circuit board 17 is a six-layer printed circuit board, integrating a central processing unit, a driver module, and an NPU chip 171. The central processing unit adopts an embedded microprocessor architecture and is connected to the control signal input of the driver module via internal wiring. The driver module is an H-bridge driver circuit array, and its output is connected to the motor terminals of the main suction pump 15 and the high-speed servo-compensated infusion pump 16 via shielded control signal lines. The NPU chip 171 is a dedicated acceleration unit for neural network inference. It is interconnected with the central processing unit via an internal high-speed bus and is dedicated to the forward inference calculation of LSTM neural network models, ensuring that the latency of a single prediction inference does not exceed 5 milliseconds. The main suction pump 15 is mechanically fixed to the upper left side of the main control circuit board 17 via a pump body mounting bracket. The main suction pump 15 is a peristaltic pump structure driven by a brushless DC motor, with a silicone tubing embedded in the pump head as the suction channel. The high-speed servo-compensated infusion pump 16 is mechanically fixed to the upper right side of the main control circuit board 17 via a pump body mounting bracket of the same specifications. The high-speed servo-compensated infusion pump 16 is a miniature plunger pump structure driven by a servo motor, possessing high dynamic response characteristics.

[0040] A circular through-hole is provided on the front side of the outer casing 11. The suction port 12 passes through this through-hole and is fixed to the panel of the outer casing 11 by a locking nut. The outward end of the suction port 12 has a standard Luer female head structure, and the inward end is sealed and plugged into the outlet end of the main suction pump 15 through a medical PVC fluid tube and secured with a cable tie. A circular through-hole of the same specification is provided on the right side of the front of the outer casing 11. The compensation infusion port 14 passes through and is fixed in the same manner. The inward end of the compensation infusion port 14 is sealed and connected to the outlet end of the high-speed servo compensation infusion pump 16 through a medical PVC fluid tube. A pressure sensor is embedded in the internal fluid channel of the suction port 12. The detection surface of the silicon piezoresistive chip of the pressure sensor is in direct contact with the liquid in the flow channel. The signal output end of the pressure sensor is soldered to the analog signal acquisition port of the main control circuit board 17 through a three-core shielded control signal line. A solenoid valve is connected in series in the internal fluid channel of the compensation injection interface 14. The solenoid valve body is a two-position, normally closed miniature solenoid valve. The coil control terminal of the solenoid valve is connected to the digital output port of the main control circuit board 17 via a two-core control signal line. An emergency stop button is also fixedly installed in the center of the front of the housing 11. The emergency stop button is a self-locking large push-button switch. Its normally closed contact is connected in series in the DC output main circuit of the power module 18, and its auxiliary contact is connected to the interrupt input pin of the main control circuit board 17 via a control signal line. A rectangular window is opened on the upper right side of the front of the housing 11. The LCD touch screen 19 is embedded in the window and fixed with four corner screws. The display interface of the LCD touch screen 19 is connected to the graphics output port of the main control circuit board 17 via a ribbon cable. The touch control interface interacts with the main control circuit board 17 via a serial bus. A communication interface is fixedly installed on the top of the housing 11. The communication interface is an RJ45 socket with a shielded shell. The internal signal lines are soldered to the network communication module of the main control circuit board 17.

[0041] The dual-channel intelligent needle 2 is a disposable sterile surgical instrument. Its main structure is the needle handle 22, which is injection-molded from medical-grade polycarbonate material and has an elongated elliptical cylindrical shape for easy gripping by the surgeon. A non-slip grip sleeve is coated on the center of the outer surface of the needle handle 22 using a secondary injection molding process. The non-slip grip sleeve is made of thermoplastic elastomer with a micro-textured surface. Two control buttons are symmetrically embedded on the left and right sides of the non-slip grip sleeve. These control buttons are membrane switches, connected to contact springs at the tail of the needle handle 22 via embedded circuitry within the needle handle 22 wall. When the dual-channel intelligent needle 2 is connected to the intelligent host system 1 via the Luer connector 25, the contact springs contact the corresponding contacts in the host interface, thereby transmitting the control button press signal to the main control circuit board 17 through the internal circuitry of the host. The needle handle 22 has an integrally formed Luer connector 25 at its tail. The Luer connector 25 adopts a standard medical Luer taper structure and consists of two parallel aspiration Luer connectors 251 and a compensation infusion Luer connector 252. The aspiration Luer connector 251 is detachably connected to the aspiration interface 12 of the intelligent host system 1 via a medical PVC fluid line. When connected, the Luer lock is tightened clockwise to achieve a seal. The compensation infusion Luer connector 252 is detachably connected to the compensation infusion interface 14 via another fluid line of the same specification.

[0042] The needle handle 22 has two independent fluid channels running parallel to each other along its length: a main aspiration channel 23 on the left and a compensation perfusion channel 24 on the right. These two channels are formed in one piece during injection molding of the needle handle 22 using a pre-set core, resulting in smooth, seamless inner walls. The tail end of the main aspiration channel 23 communicates with the inner hole of the aspiration Luer connector 251. The front end of the main aspiration channel 23 is fixedly connected to an aspiration needle tube 211 using medical-grade adhesive. The aspiration needle tube 211 is a 304 stainless steel capillary tube with an inner diameter sufficient to allow lens nucleus fragments to pass through. The tail end of the compensation perfusion channel 24 communicates with the inner hole of the compensation perfusion Luer connector 252. The front end of the compensation perfusion channel 24 is fixedly connected to a compensation perfusion tube 212, which is also a stainless steel capillary tube of the same material. The proximal portions of the aspiration needle tube 211 and the compensation perfusion tube 212 are welded together along the contact line of the tube walls using laser welding to form the needle assembly 21. In the needle assembly 21, the outlet end of the compensating infusion tube 212 is located obliquely above and behind the tip of the suction needle tube 211. The outlet axis of the compensating infusion tube 212 and the axis of the suction needle tube 211 form an acute angle in space. This acute angle range allows the infusion fluid jet ejected from the compensating infusion tube 212 to be directed directly towards the corneal endothelial region at the top of the anterior chamber, thus providing directional flushing protection for the corneal endothelium during the infusion and aspiration process.

[0043] In the transition area near the needle tip assembly 21 at the proximal end of the needle handle 22, a micro force sensor 26 and a micro pressure sensor 27 are fixedly installed via a micro-grooving embedding method. The micro force sensor 26 is a metal foil strain gauge full-bridge structure, attached to the thin wall of the elastic body at the connection between the root of the needle tip assembly 21 and the needle handle 22, used to monitor in real time the minute contact forces generated between the suction needle tube 211 and tissues such as the lens capsule and iris during intraocular manipulation. The micro pressure sensor 27 is a MEMS capacitive pressure sensing chip, whose pressure-sensing surface is connected to the near-outlet of the compensation perfusion channel 24 through a micro pressure-applying hole opened inside the needle handle 22, used to measure the local perfusion pressure of the anterior chamber in real time. The leads of the micro force sensor 26 and the micro pressure sensor 27 are all connected to the contact array at the tail of the needle handle 22 through the embedded circuitry. When the dual-channel smart needle 2 is connected to the smart host system 1, the sensing signal is transmitted through the host interface to the analog front end of the main control circuit board 17 for amplification and analog-to-digital conversion.

[0044] The image processing system 3 comprises two physical entities: an image acquisition unit 31 and an image processing unit 32. The image acquisition unit 31 consists of a binocular industrial camera module, an infrared ring light source, and a gimbal. The binocular industrial camera module contains two global shutter CMOS image sensors. The optical axes of the two lenses are parallel and the baseline distance is fixed. The camera housing is held and fixed by the ball joint of the gimbal, and the base of the gimbal is fixed to the side arm beam of the surgical microscope using C-clamps. Adjusting the joints of the gimbal ensures that the lens axis of the image acquisition unit 31 is coaxial with the observation optical path of the surgical microscope, and that the focusing plane is accurately positioned in the anterior chamber region of the patient. The camera data output port of the image acquisition unit 31 is connected to the corresponding input port of the image processing unit 32 via a data cable 34. The data cable 34 is a highly flexible USB 3.0 or CameraLink cable to meet the bandwidth requirements for high frame rate image transmission.

[0045] The image processing unit 32 is a high-performance computer workstation, internally equipped with an image acquisition card, a GPU accelerator card, and a large-capacity solid-state storage hard drive. The software architecture of the image processing unit 32 includes six functional modules: an image analysis module, a multimodal sensing fusion module, a feedforward predictive control module, an adaptive learning module, a data conversion module, and a surgical parameter library. The image analysis module receives the dual anterior chamber image stream from the image acquisition unit 31 via data line 34, and performs image preprocessing, anterior chamber region segmentation, edge contour extraction, and stereo vision 3D reconstruction algorithms on each frame. The multimodal sensing fusion module establishes connections with the image analysis module, the pressure sensor data stream from the intelligent host system 1, and the data streams from the miniature force sensor 26 and the miniature pressure sensor 27 via data interfaces. After timestamp synchronization and alignment, it runs a Kalman filter fusion algorithm to output a multidimensional anterior chamber state fusion vector. The feedforward predictive control module internally loads a pre-trained LSTM neural network model file. This module receives the fusion vector sequence from the multimodal sensor fusion module, performs forward inference on the NPU chip 171, generates predicted values ​​for the volume change rate at future times, and generates feedforward control commands based on the prediction results. The data conversion module is responsible for packaging the feedforward control commands and various state parameters according to a predetermined communication protocol and transmitting them to the main control circuit board 17 via a gigabit network cable 33 through the communication interface on the top of the casing 11. The adaptive learning module is automatically triggered after each surgery, reads the entire surgical sensor data record and control action log stored in the solid-state storage hard drive, fine-tunes the parameters of the LSTM neural network model through an incremental learning algorithm, and saves the updated model file. The surgical parameter library is a relational database that pre-stores optimized control parameter sets corresponding to four typical patient anterior chamber characteristics. The image analysis module can retrieve the surgical parameter library for feature matching and retrieval during the preoperative stage.

[0046] In practical clinical use, this embodiment includes four stages: preoperative calibration and parameter matching, intraoperative multimodal sensing and feedforward control, anomaly handling and safety redundancy, and postoperative adaptive learning. The operation process of each stage and the collaborative working mode of each component are described in detail below.

[0047] Preoperative calibration and parameter matching phase. Before the surgery begins, medical staff move the intelligent host system 1 to a designated position next to the operating table, connect the power cord, and press the power switch on the side of the casing 11. The power module 18 starts working, providing rated DC voltage to the main control circuit board 17, the main suction pump 15, the high-speed servo-compensated infusion pump 16, and the LCD touch screen 19. After the main control circuit board 17 is powered on, it executes a self-test program, sequentially checking the zero point of the pressure sensor, the action status of the solenoid valve, the motor winding resistance of the main suction pump 15 and the high-speed servo-compensated infusion pump 16, and the working status of the NPU chip 171. The self-test results are displayed on the LCD touch screen 19. After the self-test passes, the image processing unit 32 is placed on top of or near the intelligent host system 1. A gigabit network cable 33 is taken, one end is inserted into the network port of the image processing unit 32, and the other end is inserted into the communication interface on the top of the casing 11 until a locking sound is heard. Clamp the universal bracket base of the image acquisition unit 31 to the side arm of the surgical microscope, loosen the knobs of each joint of the universal bracket, manually adjust the camera posture to make the lens axis coaxial with the optical path of the microscope, observe the preview image on the monitor of the image processing unit 32, ensure that the anterior chamber area is in the center of the image and is in clear focus, and lock all joints.

[0048] Initiate the system calibration procedure. Take a standard checkerboard calibration board with a square side length of 2 mm and place it on the surgical microscope stage to simulate the eye position. Image acquisition unit 31 acquires 20 images of the calibration board from multiple different angles and distances. The image analysis module runs Zhang Zhengyou's calibration algorithm to extract the checkerboard corner points from each image, solves for the intrinsic parameter matrix and distortion coefficients of the binocular cameras through maximum likelihood estimation, and calculates the rotation matrix and translation vector between the left and right cameras. The intrinsic parameter matrix contains the equivalent focal length and principal point coordinates expressed in pixels, and the distortion coefficients contain radial and tangential distortion parameters. After calibration, the camera intrinsic parameter matrix data is saved to the configuration file of image processing unit 32.

[0049] Next, anterior chamber baseline calibration is performed. A simulated eye with known anterior chamber depth parameters is placed in the surgical position, and image acquisition unit 31 acquires images of the simulated anterior chamber. The image analysis module first converts the color image to a grayscale image, and uses a grayscale threshold segmentation algorithm to separate the pixel set of the anterior chamber region. The grayscale threshold is selected between 120 and 200, automatically adapted based on the grayscale difference between the cornea and aqueous humor. Then, the Canny edge detection algorithm is run, setting a low threshold of 30 and a high threshold of 100 to extract two continuous contour lines: the corneal endothelial edge and the anterior capsule edge of the lens. Based on the principle of binocular stereo vision, epipolar correction is performed on the images simultaneously acquired by the left and right cameras to ensure horizontal alignment of corresponding epipolar lines in the left and right images. Then, a semi-global block matching algorithm is used to calculate the disparity map, with the search window size set to an odd number of pixels wide, and the disparity range preset according to the anterior chamber depth range. For each effective pixel on the corneal endothelial contour line, based on the disparity value and the calibrated camera parameters,

[0050] Using trigonometric formulas:

[0051]

[0052] The formula for calculating the depth at this point is as follows:

[0053] Z represents the calculated anterior chamber depth, in millimeters (mm).

[0054] B is the baseline distance between the left and right optical centers of the binocular camera, obtained by camera calibration, and the unit is millimeters (mm).

[0055] f is the equivalent focal length after epipolar correction, obtained from the camera intrinsic parameter matrix, and the unit is pixels (px).

[0056] d represents the disparity value of the corresponding pixel in the disparity map, in pixels (px).

[0057] The median of all effective depth values ​​within the anterior chamber region is taken as the baseline anterior chamber depth d. The anterior chamber morphology is approximately a frustum. R is obtained from the semi-major axis of the fitted ellipse of the corneal endothelial contour at the iris plane height, and r is obtained from the semi-major axis of the fitted ellipse of the anterior lens capsule contour at the same plane. These values ​​are then substituted into the formula for the volume of a frustum.

[0058]

[0059] Calculate the baseline anterior chamber volume V, in the formula:

[0060] The calculated anterior chamber volume is expressed in microliters (μL).

[0061] h represents the anterior chamber depth, which corresponds to the distance of the anterior chamber region along the optical axis, in millimeters (mm).

[0062] R is the radius of the top surface of the frustum, which is the equivalent radius of the ellipse fitted by the corneal endothelial contour at the height of the iris plane, and is expressed in millimeters (mm).

[0063] r is the radius of the lower base of the frustum, which is the equivalent radius of the ellipse fitted by the contour of the anterior capsule of the lens at the same plane height, and the unit is millimeters (mm).

[0064] π is set to 3.1416.

[0065] This baseline data serves as a reference for calculating the volume change rate during surgery. It is packaged by the data conversion module and sent to the main control circuit board 17 via gigabit network cable 33 for storage.

[0066] Simultaneously, the image analysis module extracts multiple feature values ​​from the patient's actual anterior chamber image, including anterior chamber depth, corneal endothelial curvature radius, and iris root position feature vectors. These features are then compared with four preset parameter groups in the surgical parameter library to calculate Euclidean distances, and the group with the smallest distance is selected as the matching result. The four preset parameter groups correspond to shallow anterior chamber, deep anterior chamber, high myopia, and pseudoexfoliation syndrome, respectively. The compensation coefficient k for the shallow anterior chamber parameter group is 0.8, with a 20% reduction in threshold values ​​for each segment; the compensation coefficient k for the deep anterior chamber parameter group is 0.5, with a 15% increase in threshold values ​​for each segment; the compensation coefficient k for the high myopia parameter group is 0.7, improving occlusion detection sensitivity by 30%; and the perfusion flow limit for the pseudoexfoliation syndrome parameter group is reduced by 25%, and the contact force alarm threshold is halved. The main control circuit board 17 loads the matched optimized control parameter group and writes it into the running memory as the initial control parameters.

[0067] After calibration, perform tubing venting and flow tests. Take a set of sterilized dual-channel smart needles 2, connect and tighten the aspiration Luer connector 251 to the aspiration interface 12, and connect and tighten the compensation perfusion Luer connector 252 to the compensation perfusion interface 14. Hang the perfusion bottle on the infusion stand, connect the tubing to the input of the high-speed servo compensation perfusion pump 16, and connect the aspiration tubing to the waste collection bottle. Press the test button on the outer casing 11. The main control circuit board 17 starts the high-speed servo compensation perfusion pump 16 at a low speed. The perfusion fluid flows out from the outlet of the compensation perfusion tube 212 through the compensation perfusion interface 14, the compensation perfusion Luer connector 252, and the compensation perfusion channel 24, purging air bubbles in the tubing. At the same time, the main aspiration pump 15 starts running at low negative pressure to confirm that the aspiration needle tube 211 is unobstructed. After the test is passed, the system enters standby mode.

[0068] Intraoperative multimodal perception and feedforward control stage. After making a clear corneal incision at the limbus, the surgeon holds the handle 22 of the dual-channel intelligent needle 2, with the thumb and forefinger naturally resting on the control buttons on both sides of the anti-slip grip, and slowly inserts the needle assembly 21 into the anterior chamber through the incision. At this time, the image acquisition unit 31 continuously acquires bilateral anterior chamber images at a rate of 100 frames per second, and each frame is transmitted in real time to the image processing unit 32 via the data cable 34. The image analysis module performs the same image processing procedure as the preoperative calibration on each frame of image, extracts the continuous contour of the corneal endothelium and the anterior capsule of the lens in real time through grayscale threshold segmentation and Canny edge detection, calculates the current anterior chamber depth d in real time based on the principle of stereoscopic vision triangulation, and calculates the current anterior chamber volume V in real time by substituting it into the frustum volume formula. At a fixed time interval of 10 milliseconds, the main control circuit board 17 receives two consecutive frames of volume data V(t) and V(t-0.01), and calculates the volume change rate difference formula.

[0069]

[0070] Calculate the rate of change of volume, where:

[0071] The volume change rate is expressed in microliters per second (μL / s).

[0072] The volume of the anterior chamber at the current time t is expressed in microliters (μL).

[0073] The volume of the anterior chamber at the previous moment is expressed in microliters (μL).

[0074] The sampling time interval is fixed at 0.01 seconds (i.e., 10 milliseconds) in this system.

[0075] The calculation results are output after being smoothed by a 5-point moving average filter. The filtering formula is as follows:

[0076]

[0077] In the formula:

[0078] This is the output value of the volume change rate at the nth sampling point after filtering;

[0079] The original volume change rate measurement value for the nth sampling point;

[0080] n is the sampling point number.

[0081] Meanwhile, the pressure sensor inside the suction port 12 continuously monitors the fluid pressure in the suction pipeline. The sensor output voltage signal is converted from analog to digital by the main control circuit board 17 to obtain the pipeline pressure value. The pressure change trend is obtained by calculating its first-order difference. The miniature force sensor 26 monitors the contact force between the suction needle 211 and the intraocular tissue in real time. Its Wheatstone bridge outputs a weak differential voltage, which is amplified and filtered by the instrumentation amplifier on the main control circuit board 17 to obtain the contact force value. The unit is Newton. The miniature pressure sensor 27 measures the local infusion pressure at the outlet of the compensation infusion channel 24 in real time, and the result is processed by the signal conditioning circuit. The unit is millimeters of mercury. The above four sensor data; anterior chamber depth d and volume change rate. From the visual channel, pipeline pressure Its changing trend originates from the suction pressure channel and contact force. Local pressure is infused from the force perception channel. The data originates from the injection pressure channel and is synchronously transmitted to the multimodal sensing fusion module.

[0082] The multimodal sensor fusion module internally runs a Kalman filter fusion algorithm. First, a state-space model of the anterior chamber state is established. The state vector X contains five components: anterior chamber depth, volume change rate, tubing pressure change trend, contact force, and corneal endothelial displacement. The observation vector Z consists of real-time measurements from each sensor. The fusion module executes a prediction step and an update step every 10 milliseconds: the prediction step calculates the prior estimate of the current state based on the previous state estimate and the system dynamic matrix. and error covariance The update step calculates the Kalman gain using the current measurements from each sensor. The prior estimate is then corrected to obtain the posterior state estimate. That is, the fused anterior chamber state fusion vector The fused vector is output in structured data format, with each component marked with a timestamp.

[0083] The mathematical expression of the Kalman filter state-space model is as follows:

[0084] State vector definition:

[0085]

[0086] Prediction step equation:

[0087]

[0088]

[0089] Update step equation:

[0090]

[0091]

[0092]

[0093] In the formula:

[0094] X is the system state vector, containing the anterior chamber depth. Volume change rate Pipeline pressure change trend Contact force and corneal endothelial displacement ;

[0095] This is the prior state estimate at time k;

[0096] Let be the state transition matrix at time k;

[0097] Let be the prior error covariance matrix at time k;

[0098] Let k be the process noise covariance matrix at time k.

[0099] Let K be the Kalman gain matrix at time k;

[0100] Let be the observation matrix at time k;

[0101] Let be the observation noise covariance matrix at time k;

[0102] Let be the observation vector at time k;

[0103] It is the identity matrix;

[0104] This is the posterior state estimate at time k, i.e., the fused anterior chamber state fusion vector.

[0105] The feedforward prediction control module continuously receives the fused vector sequence using a sliding time window. The module is pre-loaded with a pre-trained LSTM neural network model, which employs a three-layer LSTM hidden layer structure, with each layer containing 128 hidden units. The input sequence length is 50 time steps, with each step corresponding to 10 milliseconds. That is, the model takes the fused vector sequence from the past 500 milliseconds as input and outputs predicted volume change rates for the next four time points: 50 milliseconds, 100 milliseconds, 150 milliseconds, and 200 milliseconds.

[0106] The mathematical expressions for the input and output of the LSTM neural network model's forward inference are as follows:

[0107] Input sequence expression:

[0108]

[0109] in:

[0110] The input sequence for the LSTM neural network model consists of a fused state vector for 50 consecutive time steps (500 milliseconds in total);

[0111] Let be the multimodal fusion state vector at time t.

[0112] Predicted output expression:

[0113]

[0114] in:

[0115] For the future The predicted rate of volume change at any given time, in microliters per second (μL / s).

[0116] The prediction lead time can be set to 50 milliseconds, 100 milliseconds, 150 milliseconds, or 200 milliseconds.

[0117] This is the forward inference function of the LSTM neural network model.

[0118] The model's offline training is based on multimodal sensor data collected from over 2000 real cataract surgeries. Mean squared error is used as the loss function during training, and the Adam optimizer is used for parameter updates. The formulas used include:

[0119] Adam optimizer formula

[0120]

[0121] That is, the formula for updating the parameters of the Adam optimizer used during model training; in the formula:

[0122] These are the model parameters, specifically the weights and biases to be optimized in the LSTM neural network model.

[0123] The learning rate is a hyperparameter that controls the step size of parameter updates, and determines the magnitude by which the parameters move along the gradient direction in each iteration.

[0124] The bias-corrected first moment estimate is the estimate of the exponential moving average of the first moment (mean) of the gradient after bias correction, which reflects the average direction and strength of the gradient.

[0125] The bias-corrected second moment estimate is the bias-corrected estimate of the exponential moving average of the gradient second moment (uncentered variance), which is used to adaptively adjust the learning rate scale of each parameter.

[0126] The smoothing constant is a small constant added to prevent the denominator from being zero, thus ensuring numerical stability.

[0127] During online inference, upon receiving each new fused vector frame, the prediction module sends the updated sequence to the NPU chip 171 for forward computation. The parallel computing array of the NPU chip 171 completes all matrix multiplications and activation function operations within 5 milliseconds, outputting four predicted values. The prediction module then retrieves the predicted values ​​for the next 50 milliseconds. It serves as the primary control basis and simultaneously monitors the predicted trend over the next 100 to 200 milliseconds.

[0128] The main control circuit board 17 implements a feedforward segmented control strategy based on the predicted value, and the specific decision logic is as follows.

[0129] When the predicted volume change rate When the volume is between -5 μL / s and -10 μL / s, it indicates a slight decreasing trend in the anterior chamber volume. Although it has not reached the level of collapse risk, it shows early signs. At this time, the main control circuit board 17 uses feedforward control to send instructions to the drive module in advance, controlling the servo motor of the high-speed servo compensated infusion pump 16 to run at 1.2 times the base speed, outputting a preventive infusion flow rate, and adding a small amount of infusion fluid in advance to curb the inertia of volume decrease and keep the anterior chamber depth stable.

[0130] When the predicted volume change rate When the volume drops to between -15 and -20 microliters per second, it indicates that the foreroom volume is decreasing rapidly, and without intervention, significant collapse will occur within tens of milliseconds. The main control circuit board 17 immediately initiates active compensation control, based on the compensation flow formula...

[0131]

[0132] Calculate the required compensating perfusion flow rate. In the formula, Q represents the compensating perfusion flow rate, in microliters per second. To predict the rate of change of volume, the unit is microliters per second; This represents the change in foreroom volume per unit time. This represents the sampling time interval; k is the compensation coefficient, dimensionless, obtained by matching from the preoperative parameter database, with 0.8 for shallow anterior chamber patients, 0.5 for deep anterior chamber patients, and 0.7 for highly myopic patients. For example, when... When the flow rate is -18 μL / s and k is 0.6, Q = 10.8 μL / s. The main control circuit board 17 adjusts the speed servo loop of the high-speed servo-compensated infusion pump 16 through the drive module, so that the actual output flow rate tracks the target value within 10 milliseconds. Since the system is controlled based on predicted values ​​rather than measured values, the infusion compensation action is completed before the volume actually drops to the dangerous threshold, achieving zero-delay preventive intervention.

[0133] When the predicted volume change rate When the pressure drops sharply to between -25 μL / s and -35 μL / s, and the inter-frame differential algorithm detects a change in the anterior chamber projection area exceeding 80%, the system determines that the suction port is severely blocked and about to be breached. At the moment the blockage is breached, the high negative pressure accumulated in the pipeline will be released abruptly, generating a strong surge effect. To counteract this surge, the main control circuit board 17 executes a dual strategy before the blockage is completely breached: First, it increases the drive voltage of the main suction pump 15, raising the suction negative pressure from the conventional 300 mmHg to 450 mmHg to actively breach the blockage; second, it simultaneously controls the high-speed servo-compensated infusion pump 16 to enter a pulsed infusion mode with a pulse frequency of 10 Hz and a duty cycle of 30%, meaning that the infusion pump operates at high speed for 30 milliseconds and stops for 70 milliseconds within each pulse cycle, forming an intermittent high-pressure infusion flow. The moment the blockage is broken, the negative pressure in the suction line drops sharply, and the anterior chamber is at risk of being over-suctioned. At this moment, the compensating fluid released by the pulse perfusion precisely offsets the surge suction force. The two are precisely coordinated in timing, so that the anterior chamber volume change rate returns to the normal range in a very short time, avoiding the impact traction on the iris and lens capsule.

[0134] During the surgery, the LCD touchscreen display 19 graphically presents several key parameters in real time: the main display area shows a dynamic curve of the anterior chamber depth changing over time, with the current value accurate to 0.01 mm; the secondary display area shows a real-time bar chart of the volume change rate, with a dotted line indicating the predicted trend line for the next 50 to 200 milliseconds; the right-hand information bar displays the current perfusion and aspiration flow rates as progress bars, and shows the tubing pressure and tissue contact force in numerical and unit form; the top status bar indicates the current operating mode, with possible states including "normal mode," "predictive mode," "pressure control mode," and "emergency stop mode." The surgeon can manually fine-tune the compensation coefficient k, the upper and lower limits of each segment threshold, and the occlusion detection sensitivity using virtual buttons on the touchscreen. The adjusted parameters take effect immediately and are simultaneously recorded by the adaptive learning module for parameter optimization in subsequent surgeries.

[0135] Anomaly handling and safety redundancy phase. The system is designed with multi-layered safety protection mechanisms to cope with various non-ideal working conditions that may occur during the operation.

[0136] When intraoperative anterior chamber image clarity decreases due to lens cortical debris, microbubbles, or anterior chamber hemorrhage, and the image analysis module fails to extract the complete corneal endothelium or anterior lens capsule edge contour for five consecutive frames, the system automatically triggers an image abnormality switching mechanism. The image processing unit 32 sends an "image quality substandard" status code to the main control circuit board 17, and the main control circuit board 17 immediately switches the control mode from predictive control mode to pressure control mode. In pressure control mode, the system abandons its reliance on visual sensor data and instead relies entirely on the readings of the pressure sensor within the suction port 12 for traditional closed-loop control. The control algorithm uses a pressure PID regulator, setting the target tubing pressure to 250 mmHg, and the actual pressure value... The deviation 'e' from the target value is calculated using proportional, integral, and derivative operations to output the infusion flow control quantity. The control law is as follows:

[0137]

[0138] In the formula:

[0139] This is the perfusion flow rate output value under pressure control mode, in microliters per second (μL / s).

[0140] The actual pipeline pressure value is measured by the pressure sensor inside the suction port, and the unit is millimeters of mercury (mmHg).

[0141] 250 is the target pipeline pressure setting value, in millimeters of mercury (mmHg).

[0142] This is a pressure deviation signal;

[0143] 0.03 is the proportional gain coefficient;

[0144] 0.01 is the integral gain coefficient;

[0145] This is the integral term of the pressure deviation over time.

[0146] The flow rate is measured in microliters per second, and the pressure is measured in millimeters of mercury. Simultaneously, the LED warning light on the front of the casing flashes yellow, reminding the operator that they are currently in pressure control mode and need to clear their field of vision as soon as possible to restore visual perception.

[0147] When the micro force sensor 26 detects the contact force between the needle and the intraocular tissue If the force exceeds 0.1 Newtons and lasts for more than 20 milliseconds, the main control circuit board 17 determines that there is a potential risk of mechanical damage to the capsule or iris. The system immediately executes a three-level protection action: First, the motor voltage of the main suction pump 15 is rapidly reduced to idle level via the drive module, so that the suction negative pressure drops to a safe value below 100 mmHg; Second, the buzzer built into the housing 11 is triggered to emit a rapid intermittent alarm sound, and a red warning dialog box pops up in the center of the LCD touch screen 19, displaying the text "Contact force too high, please adjust the needle position"; Third, the event, along with a snapshot of the sensor data for 10 seconds before and after, is saved to a log file for postoperative analysis. When the contact force drops below 0.05 Newtons and is manually confirmed, the system returns to normal control mode.

[0148] In the event of an emergency requiring immediate interruption of infusion and aspiration, such as posterior capsule rupture or vitreous prolapse during the procedure, the surgeon or assistant can decisively press the emergency stop button located in the center of the front of the outer casing 11. The normally closed contact of the emergency stop button opens, physically cutting off the power supply circuit from the power module 18 to the main aspiration pump 15, the high-speed servo-assisted compensating infusion pump 16, and the solenoid valve, causing all power components to stop operating within 2 milliseconds. Simultaneously, the auxiliary contact sends a high-priority interrupt signal to the main control circuit board 17, which executes the emergency state saving procedure, writing 200 frames of binocular images before and after the current moment, all sensor data caches, and the history of control commands into non-volatile memory, generating an encrypted emergency event log file. After power loss, the solenoid valve returns to its normally closed state via an internal spring, cutting off the fluid passage of the compensating infusion interface 14 and preventing the infusion fluid from continuing to enter the eye due to gravity or siphon effect.

[0149] Postoperative adaptive learning phase. The surgery was successfully completed. The surgeon withdrew the dual-channel smart needle 2 from the patient's eye, pressed the Luer lock release mechanism, and disconnected the aspiration Luer connector 251 and the compensating infusion Luer connector 252 from their corresponding interfaces on the intelligent host system 1. The dual-channel smart needle 2, as a single-use consumable, was disposed of in the sharps container according to medical waste management regulations. The surgeon clicked the "Surgery Ended" button on the LCD touchscreen display 19, and the system automatically entered the postoperative processing flow.

[0150] The adaptive learning module is activated, and it first reads the entire surgical record file from the solid-state storage hard drive. The record file contains three parts of data: the first part is a multimodal sensor data sequence, which records the entire fused state vector from entry into the anterior chamber to the end of the surgery with a sampling period of 10 milliseconds. The fields include timestamp, anterior chamber depth, volume change rate, tubing pressure, contact force, local irrigation pressure, and corneal endothelial displacement. The second part is a control action sequence, which records each adjustment command from the main control circuit board 17 to the main suction pump 15 and the high-speed servo-compensated irrigation pump 16 in an event-triggered manner. The fields include timestamp, command type, target value, and actual feedback value. The third part is the surgical outcome evaluation index, which is input by the surgeon through the scoring interface on the LCD touch screen 19 after surgery. This includes the subjective score of anterior chamber stability, the number of collapse events, the number of surge events, and whether complications such as posterior capsule rupture occurred. The module packages these three parts of data and uploads them to the dedicated database of the image processing unit 32 through the hospital's internal network, or saves them to a local encrypted partition for offline processing.

[0151] The adaptive learning module runs an incremental learning algorithm to update the parameters of the LSTM neural network model. The algorithm employs an online gradient descent strategy, using the actual volume change rate sequence during the surgery as the supervision label and the predicted value output by the model under the given conditions as the prediction output, calculating the mean squared error loss function:

[0152] Adaptive learning loss function (mean squared error loss function)

[0153]

[0154] That is, the loss function value, used to measure the deviation between the model's predicted value and the true value; in the formula:

[0155] N is the total number of samples involved in the calculation;

[0156] This is the actual measured value of the volume change rate of the i-th sample, in microliters per second (μL / s).

[0157] This is the volume change rate model prediction for the i-th sample, expressed in microliters per second (μL / s).

[0158] The backpropagation process only fine-tuned the weights of the last two fully connected layers of the model, with a learning rate set to 0.0001 to prevent catastrophic forgetting of the converged model. The updated model weight file overwrote the original model file and included a version number. After dozens of surgeries, the model's root mean square error in predicting volume change rate significantly decreased, and its accuracy in predicting anterior chamber collapse events continued to improve, demonstrating the system's ability to adaptively evolve with the increase in clinical use.

[0159] After data processing is completed, medical staff wipe the surface of the outer shell 11, the suction port 12, and the outside of the compensation infusion port 14 of the intelligent host system 1 with a non-woven cloth soaked in 75% ethanol. They then use ultraviolet disinfection lamps to sterilize the lens surface of the image acquisition unit 31 and the operating table of the image processing unit 32 for 30 minutes. Finally, they turn off the main power of the system, reset all mechanical and electrical components to their initial state, and prepare to continue with the next surgery.

[0160] This embodiment achieves a fundamental leap from indirect inference to direct measurement of anterior chamber status by setting up an intelligent injection and aspiration control system that simultaneously possesses a real-time visual perception and multimodal sensing fusion structure for the anterior chamber, a dual-channel independent injection and aspiration and directional compensation perfusion structure, and a feedforward predictive control and adaptive learning structure. The multimodal sensing fusion module organically integrates visual depth information, tubing pressure trends, and needle contact force, eliminating the information blind spots of traditional single pressure monitoring; the feedforward predictive control module predicts the anterior chamber volume change trend in advance based on fusion vectors and LSTM neural networks, actively intervening before collapse or surge actually occurs, shifting the control response mode from passive remediation to active prevention; the introduction of miniature force sensors and miniature pressure sensors provides a force-sensory safety barrier for the system, effectively avoiding mechanical damage to intraocular tissues; the synergy between the adaptive learning module and the surgical parameter library enables the control strategy to match individual patient characteristics and continuously optimize with the increase in the number of surgeries. While maintaining the low cost and aseptic requirements of disposable needles, the entire system significantly improves the anterior chamber stability of the phacoemulsification injection and aspiration process, greatly reducing the risk of intraoperative corneal endothelial damage, posterior capsule rupture, and surge-related complications, effectively ensuring surgical safety and postoperative visual quality, while also reducing the surgeon's mental burden and operational intensity.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent injection and aspiration control system for ophthalmic surgery, characterized in that: Includes an intelligent host system, a dual-channel intelligent needle, and an image processing system; The intelligent host system includes a casing, a power module, a main control circuit board, a main suction pump, and a high-speed servo-compensated infusion pump. The power module is fixedly disposed inside the bottom of the casing, the main control circuit board is fixedly disposed above the power module, the main suction pump is fixedly disposed on the upper left side of the main control circuit board, and the high-speed servo-compensated infusion pump is fixedly disposed on the upper right side of the main control circuit board. A suction port is fixedly disposed on the left side of the front of the casing, and a compensation infusion port is fixedly disposed on the right side of the front of the casing. The output end of the main suction pump is sealed to the suction port through a fluid pipeline, and the output end of the high-speed servo-compensated infusion pump is sealed to the compensation infusion port through a fluid pipeline. Both the main suction pump and the high-speed servo-compensated infusion pump are connected to the main control circuit board through control signal lines, and the power module is connected to the main control circuit board, the main suction pump, and the high-speed servo-compensated infusion pump through power lines. The dual-channel smart needle includes a needle handle, a Luer connector, and a needle assembly. The Luer connector is fixedly disposed at the tail of the needle handle and includes an aspiration Luer connector and a compensation infusion Luer connector. The aspiration Luer connector is detachably connected to the aspiration interface via a fluid conduit, and the compensation infusion Luer connector is detachably connected to the compensation infusion interface via a fluid conduit. A main aspiration channel and a compensation infusion channel are arranged parallel to each other along the length direction inside the needle handle. One end of the main aspiration channel is connected to the aspiration Luer connector, and the other end of the main aspiration channel is fixedly connected to an aspiration needle tube. One end of the compensation infusion channel is connected to the compensation infusion Luer connector, and the other end of the compensation infusion channel is fixedly connected to a compensation infusion tube. The proximal ends of the aspiration needle tube and the compensation infusion tube are fixedly formed by laser welding to form the needle assembly. The image processing system includes an image acquisition unit and an image processing unit. The image acquisition unit is connected to the image processing unit via a data cable, and the image processing unit is connected to the main control circuit board via a gigabit network cable. The image acquisition unit is fixed to the side arm of the surgical microscope via a universal bracket, and the lens axis of the image acquisition unit is coaxial with the optical path of the microscope and aligned with the anterior chamber area. The image processing unit includes an image analysis module, a multimodal sensing fusion module, and a feedforward predictive control module. The multimodal sensing fusion module is used to fuse in real time the anterior chamber visual feature data acquired by the image acquisition unit, the tubing pressure data acquired by the pressure sensor located inside the aspiration interface, and the contact force data between the needle and intraocular tissue acquired by the miniature force sensor located on the dual-channel smart needle, and generate an anterior chamber state fusion vector. The feedforward predictive control module is used to receive the anterior chamber state fusion vector, generate a predicted value of the anterior chamber state at future times using a trained LSTM neural network model, and send a feedforward control signal to the main control circuit board before the predicted value of the anterior chamber state reaches a preset threshold. The main control circuit board adjusts the operating parameters of the main aspiration pump and the high-speed servo-compensated perfusion pump in advance based on the feedforward control signal.

2. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 1, characterized in that: The main control circuit board includes a central processing unit, a drive module, and an NPU chip. The central processing unit is connected to the drive module via control signal lines, and the drive module is connected to the main suction pump and the high-speed servo-compensated injection pump via control signal lines respectively. The NPU chip is used to run the forward inference calculation of the LSTM neural network model.

3. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 2, characterized in that: The image processing unit further includes a data conversion module and an adaptive learning module; the data conversion module is connected to the main control circuit board via the gigabit network cable; the adaptive learning module is used to extract multimodal sensor data sequences and control action sequences after surgery, and update the parameters of the LSTM neural network model through an online gradient descent incremental learning algorithm.

4. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 1, characterized in that: The outlet end of the compensating infusion tube is located obliquely above the suction needle tube, and the outlet axis of the compensating infusion tube forms an acute angle with the axis of the suction needle tube. The outlet of the compensating infusion tube faces the corneal endothelium. The micro force sensor and the micro pressure sensor are fixedly installed on the needle handle or the needle tip assembly. The micro pressure sensor is used to monitor the local infusion pressure of the anterior chamber in real time. Both the micro force sensor and the micro pressure sensor are connected to the main control circuit board via embedded circuitry.

5. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 1, characterized in that: The pressure sensor is fixedly installed inside the suction port. The detection end of the pressure sensor is connected to the fluid channel of the suction port, and the signal output end of the pressure sensor is connected to the main control circuit board through a control signal line. The solenoid valve is fixedly installed inside the compensation infusion port. The valve body of the solenoid valve is connected in series in the fluid channel of the compensation infusion port, and the control end of the solenoid valve is connected to the main control circuit board through a control signal line. An LCD touch screen is also fixedly installed on the front of the housing. The LCD touch screen is used to display the current anterior chamber depth, volume change rate, predicted trend curve, tubing pressure, and tissue contact force in real time.

6. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 1, characterized in that: The image processing unit also includes a surgical parameter library, which pre-stores optimized control parameter sets corresponding to different patients' anterior chamber characteristics; the image analysis module is used to extract the patient's anterior chamber characteristics before surgery and match them with the surgical parameter library, and the main control circuit board is used to load the matched optimized control parameter sets as initial control parameters.

7. An intelligent injection and aspiration control system for ophthalmic surgery according to any one of claims 1 to 6, characterized in that: The control method of the system includes the following steps: S1: Preoperative calibration, the image acquisition unit acquires anterior chamber images, and the image processing unit calculates the camera intrinsic parameter matrix and determines the baseline anterior chamber depth and baseline volume; S2: Intraoperative real-time perception and fusion, the image acquisition unit acquires anterior chamber images in real time, while the pressure sensor and the miniature force sensor acquire tubing pressure and needle contact force respectively; the multimodal sensing fusion module fuses anterior chamber visual feature data, tubing pressure data and needle contact force data to generate a fused state vector containing anterior chamber depth, volume change rate, pressure change trend and contact force. S3: Feedforward prediction and intervention. The feedforward prediction control module receives the fused state vector and predicts the future trend of anterior chamber volume change through an LSTM neural network model. When the predicted value indicates that anterior chamber collapse or surge is about to occur, the feedforward prediction control module sends a feedforward control signal to the main control circuit board before the corresponding event actually occurs. The main control circuit board adjusts the negative pressure of the main suction pump and the infusion flow rate of the high-speed servo compensated infusion pump in advance according to the feedforward control signal. S4: Segmented closed-loop control. When the predicted volume change rate is within the first preset range (-10μL / s to -5μL / s), the main control circuit board controls the high-speed servo compensated infusion pump to output a preventive infusion flow rate. When the predicted volume change rate drops to the second preset range (-20μL / s to -15μL / s), the main control circuit board increases the infusion flow rate of the high-speed servo compensated infusion pump according to the compensation flow rate formula. When the predicted volume change rate suddenly drops to the third preset range (-35μL / s to -25μL / s), the main control circuit board increases the negative pressure of the main suction pump in advance and controls the high-speed servo compensated infusion pump to perform pulsed infusion. S5: Safety redundancy and adaptive learning. When the image processing unit fails to extract a complete contour for 5 consecutive frames, the main control circuit board switches to the pressure control mode based on the pressure sensor. After surgery, the adaptive learning module extracts and analyzes the surgical data and incrementally updates the LSTM neural network model.

8. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 7, characterized in that: In S2 of the control method, the image processing unit calculates the current anterior chamber depth and volume change rate in real time using a stereo vision algorithm and a frustum volume formula. Specifically, it performs stereo matching on the corrected left and right camera images using a semi-global block matching algorithm to generate a disparity map, and then uses a triangulation formula based on the disparity map. Calculate the anterior chamber depth (where Z is the depth value in mm; B is the distance between the optical centers of the left and right cameras; f is the corrected equivalent focal length; d is the parallax value), and based on the frustum model... Calculate the anterior chamber volume (where h is the anterior chamber depth, R is the radius of the top of the frustum, r is the radius of the bottom of the frustum, and π is taken as 3.1416).

9. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 7, characterized in that: In S4 of the control method, the compensation flow formula is: in, To compensate for the injection flow rate, The predicted rate of change of volume. This represents the change in foreroom volume per unit time. The sampling time interval is represented by k, which is the compensation coefficient corresponding to the patient's anterior chamber characteristics matched according to the surgical parameter library.

10. The intelligent injection and aspiration control system for ophthalmic surgery according to claim 7, characterized in that: In S5 of the control method, when the contact force detected by the micro force sensor exceeds the preset force threshold (0.1N) and the duration exceeds the preset time threshold (20ms), the main control circuit board reduces the suction negative pressure of the main suction pump to a safe value (≤100mmHg) and issues an alarm signal.

Images