Joint cavity rotary cutting hemostasis device with lesion hardness identification

Abstract

The application discloses a joint cavity rotary cutting hemostasis device with lesion hardness identification, and belongs to the technical field of medical devices. The device comprises a handle, a working rod and a control system. The distal end of the working rod adopts a four-layer concentric sleeve structure of a telescopic limiting outer sleeve tube, a static outer sheath tube, a perfusion channel and a rotary inner core rod. The outer sleeve tube can be axially adjusted and locked to limit the exposure length of the suction cutting window, thereby playing a cutting depth limiting role. The outer sheath tube is fixed with a ring-shaped plasma radio frequency electrode at the front end. The electrode lead wire is routed through an independent path and does not move with the rotating part. The perfusion channel and the negative pressure suction channel are independent of each other in structure. The control system takes the brushless direct current motor driving the inner core rod as a mechanical sensing element, and performs double-level lesion hardness identification through torque current signals. The device can reduce the risk of important tissue mis-cutting and improve the operation safety in arthroscopic surgery.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, specifically to a joint cavity rotary cutting hemostasis device with lesion hardness recognition. Background Technology

[0002] In arthroscopic surgery, rotary shaving instruments are used to remove hypertrophic synovium, trim menisci, and remove loose bodies. Radiofrequency ablation is also necessary to stop bleeding. Traditional rotary shaving instruments lack mechanical depth control at the tip. If the advance force is not properly controlled during synovectomy, the blade may penetrate dense tissues such as the joint capsule, causing damage to extra-articular structures. Some existing instruments integrate radiofrequency ablation, but their electrode leads often move with the rotating components, making them prone to breakage or signal interference with prolonged use, affecting reliability. Furthermore, perfusion and suction channels are often shared or interdependent. High negative pressure suction flow may weaken the perfusion effect, affecting joint cavity filling and visual clarity.

[0003] In terms of tissue identification, most existing rotary cutting instruments rely on the operator's tactile and visual experience to determine tissue type. When facing deep or limited areas, it is difficult to accurately distinguish tissues of different hardness, such as synovium, cartilage, ligaments, and osteophytes. Accidentally cutting hyaline cartilage or ligaments can lead to irreversible joint damage. While solutions have emerged in recent years to add force or torque sensors to the instruments to detect cutting resistance, these additional sensors not only increase instrument size and cost but also require signal leads to be placed on the cutting head or drive shaft, making it difficult to balance sealing and reliability.

[0004] Therefore, it is necessary to provide a joint cavity rotary cutting hemostasis device that can achieve cutting depth limitation in physical structure, lead wire rotation-free structure and irrigation suction decoupling, and intelligently determine tissue hardness and adaptively adjust cutting speed and hemostasis strategy by utilizing the motor's own force feedback during operation, so as to reduce surgical risks and improve operational safety. Summary of the Invention

[0005] The purpose of this invention is to provide a joint cavity rotary cutting hemostasis device with lesion hardness recognition, which can reduce the risk of accidental cutting of important tissues and improve operational safety during arthroscopic surgery.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A joint cavity rotary cutting hemostasis device with lesion hardness recognition includes: a handle, a working rod extending forward from the handle, and a control system built into the handle, wherein: The distal end of the working rod is the functional execution part, which adopts a four-layer concentric sleeve structure, consisting of a telescopic limiting outer sleeve, a stationary outer sheath, an injection fluid channel, and a rotating inner core rod from the outside to the inside. The retractable limiting outer sleeve is slidably sleeved on the outside of the stationary outer sheath and is equipped with a locking mechanism to fix the retractable limiting outer sleeve at any position in the axial direction of the outer sheath. The outer surface of the retractable limiting outer sleeve has scale markings, and the front end face of the retractable limiting outer sleeve forms an annular stop surface to limit the exposed length of the suction and cutting window opened on the distal sidewall of the outer sheath. The front end of the stationary outer sheath is fixed with an annular plasma radio frequency electrode, which is connected to the radio frequency energy interface at the rear end of the handle through an insulated wire embedded in the wall of the outer sheath. The infusion fluid channel is formed in the annular gap between the inner wall of the stationary outer sheath and the outer wall of the rotating inner core rod. Its proximal end is connected to the infusion fluid connector, and its distal end is connected to the micro-infusion hole opened on the side wall of the outer sheath. The rotating inner core rod is driven to rotate by a motor, and its distal sidewall is provided with a cutting edge, and the central channel forms a negative pressure suction channel. The control system includes a brushless DC motor that drives the rotating inner core rod, a current sampling module for collecting the phase current of the motor, a position sensor for acquiring the motor speed, a main control chip, and an alarm device. The main control chip is configured to: calculate torque current in real time from motor phase current and execute a two-level lesion hardness identification algorithm; wherein, the first layer is a fast-response protection layer, which executes motor emergency stop protection based on the comparison result of torque current change rate and preset threshold; the second layer is an intelligent classification layer, which extracts torque current time series data and auxiliary physical features in a sliding window manner, sends them to a lightweight multi-scale time series feature fusion neural network deployed inside the main control chip, outputs probability distributions corresponding to multiple tissue types, and adaptively adjusts motor speed or triggers shutdown and alarm according to probability distribution.

[0007] Furthermore, the retractable limiting outer sheath is made of medical-grade 304 stainless steel, with its inner diameter fitting with the outer diameter of the stationary outer sheath, and its locking mechanism is a locking knob with a radial thread connection; the scale markings are axially equidistant scale lines etched by laser.

[0008] Furthermore, the stationary outer sheath is made of 316LVM stainless steel, and the suction and cutting window on its distal sidewall is an oblong or rounded rectangle with rounded edges. The annular plasma radiofrequency electrode is fixed on the hemispherical end face at the foremost end of the outer sheath. The electrode is made of platinum-iridium alloy and is covered with a ceramic insulating layer except for the surface facing the tissue. The insulated wire connecting the electrode is recessed into the axial groove on the outer wall of the outer sheath and extends proximally into the handle along the groove.

[0009] Furthermore, the micro-infusion holes are opened on the side wall of the stationary outer sheath and located upstream of the proximal end of the suction and cutting window, with a number of 4 to 6. The axis of each micro-infusion hole forms an angle of 30° to 45° with the axis of the outer sheath and points towards the window area. The infusion fluid channel is connected to the infusion fluid connector at the proximal end through a distribution cavity. The infusion fluid connector is used to connect to an external infusion pump.

[0010] Furthermore, the current sampling module is a three-phase lower bridge arm sampling resistor integrated on the motor drive board; the position sensor is a three-channel Hall sensor installed on the stator end face of the motor; the main control chip is a microcontroller based on the ARM Cortex-M7 core; and the alarm device includes a buzzer and a dual-color LED indicator.

[0011] Furthermore, the main control chip calculates the three-phase current in real time using Clarke and Park transformations. Axial torque current The torque current With motor electromagnetic torque They are directly proportional, and the relationship is: ,in, This represents the number of pole pairs of the motor. It is a permanent magnet flux linkage.

[0012] Furthermore, the fast-response protection layer is executed in the current acquisition interrupt service routine. When the absolute value of the difference between two adjacent torque current samples exceeds the preset hard change threshold, or the absolute value of the torque current exceeds the preset safety upper limit, the main control chip forcibly shuts down the motor drive signal and implements active short-circuit braking, while triggering the alarm device.

[0013] Furthermore, the lightweight multi-scale temporal feature fusion neural network includes a three-channel multi-scale convolutional encoder, a bidirectional long short-term memory network layer, a multi-head self-attention layer, a gated feature fusion layer, and a fully connected classification layer connected in sequence. The three-channel multi-scale convolutional encoder receives the normalized torque-current sequence, torque-current differential sequence, and motor speed sequence within the receiving window as three independent input channels. Each channel is processed by multiple parallel convolutional paths with different convolutional kernel sizes or dilation rates before being spliced ​​and dimension-reduced. The bidirectional long short-term memory network layer extracts bidirectional temporal dependencies from the encoded feature sequence; The multi-head self-attention layer performs time-weighted focusing on the bidirectional long short-term memory output, and obtains a sequence-level feature vector after global average pooling; The gated feature fusion layer adaptively fuses the sequence-level feature vectors with normalized auxiliary physical features using learnable gated vectors; The fully connected classification layer maps the fusion vector into probability vectors corresponding to five tissue categories: synovium and fat pad, meniscus fibrocartilage, articular hyaline cartilage, osteophytes and calcifications, and ligaments.

[0014] In summary, compared with the prior art, the joint cavity rotary cutting hemostasis device with lesion hardness recognition provided by the present invention has the following beneficial effects: Firstly, through the sliding cooperation and locking mechanism between the retractable limiting outer sheath and the stationary outer sheath, the surgeon can adjust the window exposure length according to the lesion range. The annular stop at the front end of the outer sheath forms a mechanical limit on the cutting depth, which can reduce the risk of accidental penetration due to excessive invasion of the tissue by the window, thus improving the safety of the operation.

[0015] Secondly, the annular plasma radiofrequency electrode is fixed on the foremost hemispherical surface of the stationary outer sheath. The lead wire is independently routed through the groove on the outer wall of the outer sheath. Neither the electrode nor the lead wire moves with the rotating inner core rod. Structurally, this avoids the problem of lead wire breakage due to rotational bending fatigue, resulting in more stable signal transmission and higher reliability of radiofrequency hemostasis.

[0016] Third, the perfusion fluid channel and the negative pressure suction channel are structurally independent. The perfusion fluid is output from the micro-perfusion hole arranged on the proximal side of the window. Even if the negative pressure suction is fully open, the perfusion flow can remain continuous and stable, which is conducive to maintaining the filling of the joint cavity and clear vision. At the same time, the perfusion flow can work together to guide the cutting debris to the window, reducing debris residue.

[0017] Fourth, the control system utilizes the brushless DC motor driving the rotating inner core rod as its own mechanical sensing element. By extracting torque current in real time, it eliminates the need for additional force or torque sensors, resulting in a simpler and more compact structure. In the dual-layer lesion hardness recognition algorithm, the rapid-response protective layer can immediately brake when torque current changes abruptly, effectively preventing instantaneous impact damage caused by accidental cutting of hard osteophytes or calcifications.

[0018] Fifth, the lightweight multi-scale temporal feature fusion neural network deployed in the main control chip can intelligently analyze the temporal patterns and auxiliary physical characteristics of torque current, achieving millisecond-level identification of five different tissue types. Based on the identification results, it can automatically adjust the cutting speed or trigger shutdown protection, reducing the surgeon's reliance on tissue judgment and lowering the risk of accidentally cutting important structures such as hyaline cartilage and ligaments. At the same time, it introduces an information entropy uncertainty rejection mechanism to adopt conservative speed control for working conditions with low classification confidence, further improving decision-making safety. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the external structure of the device of the present invention; Figure 2 This is a bar chart comparing the accuracy of the device of the present invention with that of the control scheme in various tissue categories; Figure 3 This is a line graph comparing the closed-loop response delay distribution of the device of the present invention with that of manual operation.

[0020] Reference numerals: 1. Telescopic limiting outer sheath; 2. Stationary outer sheath; 3. Annular plasma radio frequency electrode. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0022] like Figure 1 As shown, this invention provides a joint cavity rotary cutting hemostasis device with lesion hardness recognition, comprising: a handle, a working rod extending forward from the handle, and a control system built into the handle. The handle is the part held by the operator, made of medical-grade polycarbonate or polyetheretherketone injection molding, with ergonomic anti-slip texture on the surface, and internally contains a motor chamber, cable channel, and fluid pipeline interface. The working rod extends from the center of the front end of the handle along the longitudinal axis, and its distal section constitutes the functional actuator. The functional actuator adopts a four-layer concentric sleeve structure, from the outside to the inside: a retractable limiting outer sheath 1, a stationary outer sheath 2, an irrigation fluid channel, and a rotating inner core rod. Each layer is described below.

[0023] The first layer is a retractable limiting outer sheath 1. This outer sheath is made of medical-grade 304 stainless steel tubing, machined by turning and wire cutting, with a hardness of HRC 25 to 30. Both the inner and outer surfaces are electrolytically polished, with a surface roughness Ra of no more than 0.4 μm. The total length of the outer sheath is 40 mm to 60 mm, the inner diameter is 4.1 mm to 4.15 mm, and the wall thickness is 0.4 mm. The inner diameter is designed to form a clearance sliding fit with the outer diameter of the second layer, the stationary outer sheath 2, with the radial clearance controlled within the range of 0.03 mm to 0.05 mm, ensuring smooth sliding without radial wobble. On the outer cylindrical surface of the outer sheath, along the axial direction from the proximal end to the distal end, equally spaced graduation marks are formed using laser etching. The graduation lines cover a length of 20 mm to 30 mm, with each graduation spaced 1 mm apart. A long graduation line is set every 5 mm with a value marked, facilitating direct reading of the limiting depth by the surgeon during the operation. The proximal section of the outer sheath has a radially threaded through-hole with an M2 thread specification. A locking knob, made of the same stainless steel, with a 6mm outer diameter head and a knurled circumferential surface to increase turning friction, mates with the screw. The threaded section of the knob is long enough for its end face to press against the outer wall of the outer sheath when tightened, achieving axial locking through contact friction. The frontmost face of the outer sheath is a ring-shaped plane perpendicular to the axis, with a width equal to the wall thickness of 0.4mm. This plane forms a ring-shaped stop surface. After the surgeon adjusts and locks the position of the outer sheath, the front face protrudes beyond or is flush with the distal end of the suction and cutting window of the outer sheath. This protrusion limits the maximum axial length of the suction and cutting window that can be exposed within the joint cavity tissue, thus strictly limiting the cutting depth within safe boundaries. When treating the synovium at the joint capsule attachment margin, the surgeon can push the outer sheath forward to expose only 2mm of the distal end of the window. Even if accidentally pushed forward with excessive force, the stop surface of the outer sheath will first abut against the dense structure such as the capsule wall, preventing further penetration and providing mechanical protection against penetration.

[0024] The second layer is the stationary outer sheath 2. This outer sheath is the main structural component of the functional actuator, maintaining complete circumferential and axial stillness throughout the entire surgical procedure. The outer sheath is made of 316LVM medical-grade stainless steel, meeting ASTM F138 standards, exhibiting excellent biocompatibility and resistance to electrochemical corrosion, making it suitable for radiofrequency electrode carrier environments. The tube blank is formed through deep-hole drilling and cold drawing processes, with an outer diameter of 4.0 mm (tolerance ±0.01 mm), an inner diameter of 3.0 mm (tolerance ±0.01 mm), and a wall thickness of 0.5 mm. The total length of the outer sheath is 150 mm to 180 mm, and its proximal end is fixed to a metal base at the front end of the handle via laser welding or interference fit. This base also provides the electrical conduction path for the radiofrequency electrode.

[0025] Approximately 5 mm from the distal end of the outer sheath, a suction-cutting window is created on the sidewall using electrical discharge machining or precision laser cutting. The window is oblong in shape, meaning it is rectangular in the center and tapers to semi-circular ends, or a straight rectangle with rounded corners. The axial length of the window is 8 mm to 12 mm, and the circumferential opening angle is 110° to 130°, preferably 120°, measured against the outer wall of the outer sheath. All edges of the window are rounded with a radius of 0.3 mm to 0.5 mm to avoid mechanical cutting and scraping of soft tissue. The inner wall of the window, i.e., the inner surface of the outer sheath, serves as a cutting tool, engaging with the helical cutting edge of the fourth-layer rotating inner core rod to form a shearing pair.

[0026] The outer sheath has a hemispherical closed end face at its front end. The hemisphere is integrally formed with the outer sheath or joined by precision laser welding, both made of 316LVM. The center of the top of the hemisphere is located on the tube axis, and the radius of the hemisphere is equal to half the outer diameter of the tube, i.e., 2.0 mm. A ring-shaped plasma radiofrequency electrode 3 is fixedly mounted on this hemispherical end face. The electrode body is made of platinum-iridium alloy with a platinum-iridium mass ratio of 90:10, exhibiting excellent conductivity, biocompatibility, and resistance to electro-erosion. The ring electrode is a closed-loop metal ring with a rectangular cross-section. The outer diameter of the ring is 3.0 mm, the width of the ring cross-section (radial thickness) is 0.4 mm, and the axial height (exposed thickness) is 0.25 mm. The outer surface of the electrode ring is basically flush with or slightly protrudes within 0.1 mm from the outer contour line of the hemisphere to ensure good contact with tissue without forming sharp protrusions that could cause accidental scratches. Except for the exposed areas facing outwards and in contact with tissue, the remaining inner and side surfaces of the electrodes are covered with a dense biocompatible ceramic insulating layer made of alumina or zirconium oxide, with a thickness of 0.05 mm to 0.08 mm, through physical vapor deposition or plasma spraying. This insulating layer has both good dielectric strength and wear resistance.

[0027] The trailing edge of the annular electrode is spot-welded or brazed to one end of an insulated wire. The insulated wire is made of 0.2mm diameter silver-plated copper stranded wire, with an outer polyimide insulation layer, and a total outer diameter of approximately 0.3mm. The wire extends axially backward along the outer wall of the outer sheath from the electrode connection point. An axial routing groove with a depth of 0.25mm and a width of 0.35mm is pre-machined on the outer wall of the outer sheath by laser etching or precision milling, extending from the root of the hemisphere to approximately 10mm from the front end of the handle. The wire is completely submerged in the groove, and its surface is then coated with a layer of medical-grade epoxy resin or UV-curable adhesive for coverage and fixation, restoring the smoothness of the outer cylindrical surface of the outer sheath and preventing snagging during tissue entry. After the proximal end of the wire enters the handle, it is soldered to the stator terminal of an annular conductive slip ring fixed inside the handle, and then connected to the radio frequency energy interface at the tail end of the handle via the slip ring's lead wire. The radio frequency energy interface is a standard pin or contact that can be plugged into an external low-temperature plasma radio frequency host. The host outputs low-frequency radio frequency energy with a frequency of 100kHz to 500kHz, forming a plasma vapor layer with a thickness of about 100μm to 200μm between the electrode and the tissue immersed in physiological saline, achieving precise low-temperature ablation and closure of small blood vessels.

[0028] The third layer is the infusion fluid channel. This channel is spatially located in the annular gap between the inner wall of the outer sheath and the outer wall of the rotating inner core rod. The radial width of the annular gap is determined by the difference in diameter between the two tubes: the inner diameter of the outer sheath is 3.0 mm, and the outer diameter of the rotating inner core rod is 2.4 mm, therefore the radial gap is 0.3 mm, forming a flow area with a circular cross-section. This channel starts from the fluid distribution chamber inside the handle and extends to the distal functional section; its total axial length is the length of the working rod. The flow area of ​​the channel is the area of ​​the inner bore of the outer sheath minus the area of ​​the outer circle of the inner core rod, i.e. mm 2 It can meet the typical perfusion flow rate requirements of 50ml / min to 150ml / min for arthroscopic surgery, and the flow rate is not too high to maintain laminar flow and reduce tissue impact.

[0029] The proximal interface of the perfusion fluid channel is located at the tail end or side of the handle, specifically a standard Luer connector or a hose connector with a reverse taper, connecting to the output line of an external peristaltic pump or constant pressure perfusion pump. Physiological saline or Ringer's solution is pumped in by the perfusion pump at a set pressure of 60 mmHg to 100 mmHg, entering the distribution chamber inside the handle via the connector. The distribution chamber is a machined annular channel, ensuring that the fluid enters the annular gap evenly around its perimeter, avoiding excessive flow velocity on one side. At the distal end, the perfusion fluid flows along the annular gap towards the distal region of the outer sheath. On the sidewall of the outer sheath, near the proximal end of the suction window, approximately 2 mm to 3 mm upstream of the proximal edge of the window, multiple micro-perfusion holes are created using a microporous laser drilling process. The number of micro-perfusion holes is 4 to 6, evenly arranged on the same circumferential cross-section of the outer sheath, with the hole axis forming a 30° to 45° angle with the tube axis, inclined towards the window. Each micropore has a diameter of 0.3 mm, with its outlet located on the outer wall of the outer sheath. The channel extends from the inner wall to the outer wall of the outer sheath, connecting the perfusion fluid channel to the outside. When the perfusion fluid is ejected from these micropores, it forms multiple fine jets that directly wash over the tissue surface and cutting area near the cutting window, promptly dispersing the cut debris and allowing it to enter the window with the negative pressure airflow. Simultaneously, it removes the frictional heat generated during cutting, reducing the local temperature. Because the perfusion channel and the negative pressure suction channel are structurally completely independent, they do not interfere with each other functionally. Even under fully open negative pressure conditions, the perfusion flow can continue to be output stably, maintaining the fullness of the joint cavity and a clear field of vision.

[0030] The fourth layer is the rotating inner core rod. This rotating inner core rod is the only component that performs high-speed rotation, directly driven by a brushless DC motor located inside the handle via a rigid coupling. The inner core rod is made of 17-4PH or 455 stainless steel, solution-treated and aged to achieve a hardness of HRC 40 to 45, possessing excellent wear resistance and toughness to withstand high-frequency shearing and alternating loads. The outer diameter of the rod is 2.4 mm, with a tolerance controlled within -0.01 mm, and the inner diameter is 1.6 mm, i.e., a wall thickness of 0.4 mm. The inner hole extends the entire length, forming a central negative pressure suction channel. The proximal end of the rod is a solid section, connected to the motor shaft via a coupling. The coupling uses a miniature stainless steel bellows coupling or a cross-slider coupling, allowing for minute axial and angular deviations. At the point behind the coupling, where the rod passes through the handle sealing partition, a combination of a rotating lip seal and a ceramic rolling bearing is installed. This ensures both axial and radial support and achieves an airtight seal, preventing leakage from the negative pressure suction channel and backflow of body fluid into the motor cavity.

[0031] At the distal end of the inner core rod, aligned with the axial section of the outer sheath's suction-cutting window, at least one helical cutting edge is machined. The cutting edge consists of a pair of circumferentially symmetrically arranged double helical blades with a helix angle of 15° to 25°, covering the entire axial length of the window, approximately 10mm to 12mm. The rake angle of the double helical blades is +5° to 10°, and the clearance angle is 3° to 5°. Microscopic examination reveals no nicks or burrs on the cutting edge. The cutting edge undergoes precision grinding and micro-arc oxidation treatment, increasing hardness and reducing the coefficient of friction, thus minimizing soft tissue charring and debris adhesion. When the inner core rod rotates at 3000rpm to 8000rpm, the helical cutting edge and the stationary edge of the outer sheath window create a continuous shearing action similar to rotating scissors.

[0032] The cross-sectional area of ​​the channel through which the central negative pressure suction channel passes is approximately 2.01 mm². 2 The proximal end of the channel is connected to an external negative pressure source via a rotary airtight connector, with the negative pressure value typically adjustable within the range of -150 mmHg to -400 mmHg. When the negative pressure is activated, the suction channel generates a directional airflow at the cutting window, drawing soft tissue floating outside the window into the inner wall of the window. This allows the tissue to be controlled and sheared between the cutting edge and the window edge. The cut debris is instantly entrained by the airflow into the inner hole and drawn along the channel into a collection bottle outside the body. Because the inner hole is straight-through and the inner wall is electrolytically polished, with a surface roughness Ra of less than 0.2 μm, the adhesion and clogging probability of debris on the tube wall are effectively reduced.

[0033] The spatial relative positions of the annular plasma radiofrequency electrode 3, the micro-perfusion port, the aspiration and cutting window, and the cutting edge are as follows: the window faces the side of the instrument, the annular electrode is located on the foremost hemisphere, the micro-perfusion port is located on the proximal side of the window and points towards the window area, and the cutting edge corresponds directly below the window. When the operator performs the planing operation, the window is pressed tightly against the target synovial area, and the four steps of negative pressure adsorption, cutting edge rotation, irrigation fluid flushing, and debris aspiration are performed simultaneously. When a bleeding point is encountered, the motor rotation is stopped, the annular electrode on the hemisphere of the instrument tip is gently pressed or touched to the bleeding point, the radiofrequency pedal is depressed, and the low-temperature plasma can achieve point-like or surface coagulation hemostasis under the immersion of physiological saline. At this time, the irrigation fluid can continue to flow to dissipate heat and maintain the field of vision, and the negative pressure simultaneously aspirates smoke mist and bloody fluid. The operator does not need to withdraw the scope or change instruments throughout the entire process, realizing a four-in-one operation of planing, aspiration, irrigation, and hemostasis.

[0034] The inner hole of the rotating inner core rod can also serve as an auxiliary channel, allowing the passage of auxiliary instruments such as guide wires or micro biopsy forceps when needed, but its primary function remains negative pressure suction. The maximum external diameter of the entire four-layer structure is the outer diameter of the retractable limiting outer sleeve 1, approximately 4.8mm to 5.0mm, suitable for the standard 5mm arthroscopic working channel, or can be reduced to 3.5mm or 3.0mm versions depending on different joint sizes. The straightness and coaxiality of each layer of tubing are controlled through precision fixture assembly, ensuring that the head end deflection after assembly is no more than 0.02mm, guaranteeing dynamic balance during high-speed rotation.

[0035] Through the above structural scheme, the present invention provides adjustable cutting depth protection, reliable connection of stationary electrodes without rotation, and fluid decoupling of injection and suction channels from a physical level, laying a precise mechanical foundation for subsequent intelligent hardness identification and adaptive control.

[0036] The core function of the control system is lesion hardness identification and adaptive control based on mechanical feedback. This device does not install force or torque sensors on the cutter head or drive shaft. Instead, it utilizes the brushless DC motor driving the rotating inner core rod as its own mechanical sensing element. Through sensorless torque estimation technology, it extracts the load torque from the motor's operating current in real time. Then, it uses a lightweight, multi-scale temporal feature fusion neural network to intelligently determine the hardness category of the contacted tissue at the millisecond level, automatically adjusting the cutting speed or initiating a protective shutdown accordingly. The complete design scheme of this network will be referred to as the lesion hardness identification network, which will be described in detail below: 1. Hardware composition of the control system The control system hardware includes: an inner rotor brushless DC motor that drives the rotating inner core rod, with a rated speed of 10,000 rpm, a rated torque of 15 mN·m, and a number of pole pairs. The motor driver board features a three-phase lower bridge arm current sampling resistor with a resistance of 0.8Ω and an inductance of 0.12mH. It also includes a 0.05Ω resistor with an accuracy of ±1% integrated on the motor driver board, used for real-time acquisition of three-phase current. Three Hall effect position sensors, installed on the stator end face of the motor at 120° electrical angle intervals, are used to acquire rotor sector information and calculate real-time speed. The main control chip, which executes algorithm calculations and outputs control signals, is an ARM Cortex-M7 core-based microcontroller with a 480MHz clock frequency, integrating a single-precision floating-point unit and digital signal processing instructions, with 2MB of on-chip Flash and 1MB of RAM. A MOSFET three-phase full-bridge inverter circuit drives the motor with a switching frequency of 20kHz, employing space vector pulse width modulation. Finally, an alarm device connected to the main control chip is included, comprising an active buzzer and a red-green dual-color LED indicator.

[0037] 2. Mechanical Feedback Principle In the dq rotating coordinate system, the electromagnetic torque generated by the permanent magnet synchronous motor With q-axis current component They are directly proportional, and the relationship is as follows: ; In the formula, This represents the number of pole pairs of the motor. The flux linkage of the permanent magnet is a constant after the motor design is determined. Pre-calibration using the back electromotive force method is performed in this embodiment. Wb. Ignoring rotor friction torque and wind resistance losses, electromagnetic torque Load torque applied to the cutting edge during cutting satisfy Therefore, through real-time calculation This allows for the characterization of the mechanical resistance encountered by the cutting head within the tissue without delay; the higher the tissue hardness, the greater the load torque. The larger the amplitude.

[0038] 3. Overall Architecture of the Two-Level Lesion Hardness Recognition Algorithm The lesion hardness identification algorithm adopts a two-layer architecture. The first layer is a fast-response protection layer, which executes hard protection logic based on physical thresholds with a microsecond delay within the current acquisition interrupt service routine. The second layer is an intelligent classification layer, which performs online feature extraction and neural network inference on torque and current time-series data using a sliding window approach, outputs tissue hardness category, and drives adaptive speed control.

[0039] Torque current is obtained by real-time sampling of three-phase current and performing Clarke and Park transforms. This signal enters the fast-response protection layer on one hand, and is continuously written to the circular buffer on the other to construct the timing window. When the fast-response layer does not trigger an emergency stop, the intelligent classification layer slides a step size at each interval. Window data is extracted once per millisecond, and after normalization preprocessing, it is sent to the lesion hardness recognition network deployed inside the main control chip for forward inference, outputting the posterior probability of five tissue categories.

[0040] 4. Input and output definitions of the lesion hardness recognition network The input for a single inference iteration of the lesion hardness recognition network is a three-dimensional tensor. and an auxiliary feature vector .

[0041] 3D tensor The three channels are as follows: Channel 1: Inside the window The torque current sequence of a series of consecutive sampling points is denoted as... The unit is amperes, and the sequence is min-max normalized to the interval [0, 1]. During both training and inference, the sequence within this window is independently min-max normalized, i.e., the sequence of the [0, 1]th [order / type] is [normalized / ... The normalized value of each sampling point is If the window contains Then the normalized value is uniformly set to 0.

[0042] Channel Two: Inside the Window The first-order difference sequence of torque current at consecutive sampling points is denoted as: ,in when At that time, the first element , indicating that the initial difference is filled with zeros. The sequence is also normalized to the [0, 1] interval within the window using the same normalization method as channel one.

[0043] Channel 3: Inside the window The real-time speed sequence of the motor at a series of consecutive sampling points is denoted as . The unit is rpm, and it is normalized to the [0, 1] interval by minimum-maximum within the window. The normalization method is the same as that of channel one.

[0044] Auxiliary feature vector The five components are calculated using the original data before normalization, and are defined as follows: , represents the average torque current within the window, reflecting the average tissue hardness, and is measured in amperes; This represents the rate of change of torque current, reflecting the trend of tissue stiffness within the window. The sampling period is in seconds. The unit is ampere per second; ,in , representing the torque current variance, reflects the degree of resistance fluctuation during the cutting process, and is expressed in ampere squared; ,in To the window Frequency of the sequence after performing a 256-point real-number fast Fourier transform The corresponding complex amplitude, The sum of squares of the amplitude across the entire frequency band, excluding the DC component. It is the sum of the squares of the amplitudes of each frequency component within the 20Hz to 50Hz frequency band. It is a dimensionless ratio that reflects the proportion of high-frequency vibration components during the cutting process; ,in , represents the standard deviation of rotational speed within the window, reflecting the instability of rotational speed caused by load fluctuations, and is expressed in rpm.

[0045] Auxiliary feature vector Before being fed into the network, normalization is required: each of the five components is standardized using the mean and standard deviation obtained from the training set, i.e., ,in , and The training set The mean and standard deviation of each physical feature are stored in the main control chip after training and can be used directly during inference.

[0046] The output of the lesion hardness recognition network is a probability vector. ,in Corresponding to synovial membrane and fat pad categories, Corresponding to the fibrocartilage category of the meniscus, Corresponding to the hyaline cartilage category of the joint, Corresponding to the categories of osteophytes and calcifications, Corresponding ligament type, meets the requirements each The value range is [0, 1].

[0047] 5. Topology of the lesion hardness recognition network The lesion hardness recognition network adopts a five-stage cascaded architecture, consisting of three-channel multi-scale convolutional coding, bidirectional long short-term memory temporal modeling, multi-head self-attention focusing, gated prior feature fusion, and classification output. The specific configuration of each layer is explained below.

[0048] Phase 1: Three-channel multi-scale convolutional encoder Input tensor First, it is split into three independent sub-tensors based on the channel: , and The shape of each subtensor is These three subtensors are fed into three structurally identical but weight-independent multi-scale convolutional branches. For each subtensor, the multi-scale convolutional branch consists of five parallel convolutional paths: The first path is a standard one-dimensional convolution with a kernel size of 3, a stride of 1, a same padding strategy, and 32 filters. After the convolution, a batch normalization layer and a ReLU activation function are applied.

[0049] The second path is a standard one-dimensional convolution with a kernel size of 5, a stride of 1, a same padding strategy, and 32 filters. After the convolution, a batch normalization layer and a ReLU activation function are applied.

[0050] The third path is a standard one-dimensional convolution with a kernel size of 7, a stride of 1, a same padding strategy, and 32 filters. After the convolution, a batch normalization layer and a ReLU activation function are applied.

[0051] The fourth path is a dilated one-dimensional convolution with a kernel size of 3, a dilation rate of 2, a stride of 1, a same-pad edge padding strategy, and 32 filters. This convolution is followed by a batch normalization layer and a ReLU activation function. This dilated convolution can expand the receptive field to 5 time steps without increasing the number of parameters, capturing temporal correlations with certain intervals.

[0052] The fifth path is a dilated one-dimensional convolution with a kernel size of 3, a dilation rate of 4, a stride of 1, a same-pad edge padding strategy, and 32 filters. This convolution is followed by a batch normalization layer and a ReLU activation function. This dilated convolution extends the receptive field to 9 time steps, enabling the capture of local patterns over longer spans.

[0053] After each subtensor is processed through the above five paths, the shape of the feature map output by each path is as follows: The five outputs are spliced ​​along the channel dimension to obtain the shape as follows: The single-channel multi-scale feature map. After each of the three sub-tensors completes multi-scale encoding, the three shapes are... The feature maps are then concatenated along the channel dimension to obtain a shape of... The fusion feature tensor.

[0054] The fused feature tensor is then passed through a pointwise convolutional layer for channel dimensionality reduction and cross-channel information fusion. The pointwise convolutional kernel size is 1, the number of filters is 128, and after convolution, a batch normalization layer and a ReLU activation function are applied, resulting in an output shape of... Encoding feature tensor The size of each filter in this pointwise convolution is... The number of parameters is Each weight parameter.

[0055] Phase Two: Bidirectional Long Short-Term Memory Temporal Modeling Encoding feature tensor The data is fed into a single-layer bidirectional Long Short-Term Memory (LSTM) network. This layer is configured with 64 hidden units in each of the forward and backward LSM units, using tanh as the activation function and sigmoid as the recurrent activation function, returning a complete output sequence. The forward units process the sequence in forward chronological order, and the backward units process it in reverse chronological order. At each time step, the hidden states from both directions are concatenated to obtain a shape... output sequence This layer can simultaneously capture the gradual trend from front to back and the contextual information from back to front in the torque-current sequence, such as when cutting from the synovium to the cartilage. The pattern of continuous rise.

[0056] Each time step At this point, the calculation of forward long short-term memory units follows the formula: ; ; ; ; ; ; In the formula, The input vector at the current time step, i.e. The OK, This is the hidden state from the previous moment. This represents the cell state at the previous moment. , , These are the forget gate, input gate, and output gate vectors, respectively. Candidate cell state, , , , The input weight matrix has dimensions of [size missing]. , , , , This is a circular weight matrix, with dimensions of [missing information]. , , , , For bias vectors, For the sigmoid function, This is an element-wise product. The formula for calculating the backward unit is the same as that for the forward unit, but... The input order is to Final output The Behavior Forward Hidden State With backward hidden state The splicing has a dimension of 128.

[0057] Phase Three: Bulls Focus Their Attention Bidirectional Long Short-Term Memory Output Feed into the multi-head self-attention layer. The configuration of this layer is: number of attention heads. The key vector dimension of each attention head Value vector dimension The total output dimension is .

[0058] For each attention head, firstly, it is processed through three independent linear mappings. Transform into a query matrix Key matrix Sum matrix The mapping weight matrices are respectively , and Excluding bias terms. The output calculation of each attention head follows the scaled dot product attention formula: ; In the formula, This is a causal mask matrix, where the elements in the upper triangular region are negative infinity and the elements in the lower triangular region are zero. This causal mask ensures that each time step only considers the current position and previous time steps, thus guaranteeing that the computation results during online inference (before future data arrives) are completely consistent with those during training. This mask is enabled during both the training and inference phases.

[0059] The outputs of the four attention heads are concatenated along the last dimension, and then subjected to a linear mapping. Projecting back into 128-dimensional space, the resulting shape is Attention output Multi-head self-attention mechanisms enable networks to automatically increase attention weights at specific tissue transition points, such as the instant the blade tip contacts the cartilage boundary from the synovium. Variation patterns are particularly crucial for classification.

[0060] right Global average pooling is performed along the time dimension to obtain a sequence-level feature vector with a shape of 128 dimensions. The calculation of global average pooling is as follows: ; Phase 4: Gated Prior Feature Fusion sequence-level feature vectors With the auxiliary feature vector after Z-score standardization Adaptive fusion is achieved through a gated fusion mechanism. The motivation behind this mechanism is that the relative importance of data-driven deep features and mechanically based physical prior features differs under different cutting scenarios. Physical features are particularly important under no-load or extremely low-load conditions. and While the time-series patterns of depth features are more reliable, they are more discriminative under complex variable load cutting conditions. Therefore, the network needs to autonomously determine the fusion ratio of the two.

[0061] The specific calculation process of the gating fusion mechanism is as follows: First, and The vectors are concatenated into a 133-dimensional joint vector, and then passed through a gated network with 32 neurons and a sigmoid activation function to generate a 5-dimensional gated vector. Gating vector Each element controls the degree to which the corresponding physical feature component participates in the fusion. Simultaneously, the physical feature vector... Through a linear mapping Transformed into a 32-dimensional physical embedding vector. Final fused vector. The calculation follows the formula: ; In the formula, This represents a vector concatenation operation. This represents element-wise product. The dimension is .

[0062] Phase 5: Classification Header and Output Fusion Vector The classification is performed sequentially through two fully connected layers. The first fully connected layer contains 96 neurons, followed by a batch normalization layer and a ReLU activation function. Dropout regularization with a dropout rate of 0.3 is used during training, and dropout is disabled during inference. The second fully connected layer is the output layer, containing 5 neurons, followed by a Softmax activation function, and outputs a five-dimensional probability vector. The Softmax function is calculated as follows: ; In the formula, For the second fully connected layer The output logit value of each neuron.

[0063] The total number of parameters in the lesion hardness recognition network is approximately 0.9M, including approximately 0.35M of convolutional encoder, approximately 0.20M of bidirectional long short-term memory layer, approximately 0.15M of self-attention layer, and approximately 0.20M of gated fusion and classification head. On a Cortex-M7 microcontroller, the storage footprint after 8-bit integer quantization is approximately 230KB, and the time for a single forward inference is approximately 1.8ms, meeting the latency requirement for real-time classification with a 50ms sliding step.

[0064] 6. Construction of the training dataset The training dataset was constructed through ex vivo tissue cutting experiments. Five fresh porcine knee joint specimens were used in the experiments. A senior surgeon with over ten years of arthroscopic surgery experience operated the prototype device under direct arthroscopic visualization, sequentially cutting and collecting data from the following five target tissues: synovium and infrapatellar fat pad, meniscus fibrocartilage, hyaline cartilage of the femoral condyle, subchondral bone sclerosis and marginal osteophytes, anterior cruciate ligament, and posterior cruciate ligament. At least 200 effective cutting fragments were collected from each tissue type, with each fragment lasting from 3 to 10 seconds.

[0065] During the cutting process, the main control chip transmits the original three-phase current values ​​at a frequency of 1kHz. , , The Hall effect signal sector values ​​are recorded in real time and output to the host computer via serial port. The host computer runs data acquisition software to synchronously record the values ​​after Clarke-Park transformation. Value and speed The values ​​are then aligned with the arthroscopic video frame using timestamps. The start and end times of each cutting segment are marked by the physician via a foot switch. After the procedure, the physician also labels each segment with tissue type based on the arthroscopic playback video. The labeling is in a one-out-of-five unique thermal encoding format.

[0066] The collected raw data is used to construct training samples using the following pipeline: First, the continuous data stream is divided into groups according to the window length. sampling points, sliding step size A sliding window is used to extract samples from each point, resulting in a large number of time-series windows. For each window, the three-channel input tensor is calculated according to the definition in Part IV. and five-dimensional auxiliary feature vectors The tissue category label of a window is determined by the majority of physician labels within that window. If a window spans two tissues, it is classified into the category that accounts for more than 60% of the sampling points; otherwise, the window is discarded.

[0067] To improve the robustness of the model under no-load conditions, additional motor idling data was collected to construct an no-load window. The labels of the no-load window were uniformly set to the categories of slicker and fat pad, because the load during slicker cutting is closest to no-load.

[0068] The final training dataset contains approximately 12,000 labeled window samples, including about 3,500 for synovium and fat pads, about 2,500 for menisci, about 2,500 for hyaline cartilage, about 2,000 for osteophytes and calcifications, and about 1,500 for ligaments. The dataset was randomly divided into training, validation, and test sets in a 7:1.5:1.5 ratio, ensuring that all window samples of the same cut segment were grouped into the same subset to prevent data leakage.

[0069] Data augmentation strategies include: applying an amplitude of 0.5 to each window of samples in the training set. Gaussian noise perturbation to The sequence was designed to simulate intraoperative electromagnetic interference and individual differences; with a probability of 0.3, the sequence was used to... The applied amplitude of the sequence is A random bias is applied to simulate different motor operating conditions; the window is randomly cropped in the time dimension with a probability of 0.2 and linearly interpolated to restore it to its original state. The length was used to simulate the temporal stretching effect at different cutting speeds.

[0070] Before training begins, statistically analyze five physical characteristics of all window samples in the training set. to mean and standard deviation ,in Used for Perform Z-score standardization. and After training, the parameters are stored as fixed parameters in the main control chip and used directly during inference.

[0071] 7. Model Training Scheme The training of the lesion hardness recognition network is divided into three stages: teacher model pre-training stage, knowledge distillation training stage, and quantitative perception fine-tuning stage.

[0072] Phase 1: Teacher Model Pre-training First, a teacher model was constructed that has the exact same network structure as the lesion hardness recognition network, but with the number of channels and hidden units increased fourfold. In the teacher model, the number of filters in the multi-scale convolutional branches is 128, the number of pointwise convolutional filters is 512, the number of bidirectional long short-term memory hidden units is 256, and the number of self-attention heads is 8, with each head... The first fully connected layer has 384 neurons. The total number of parameters in the teacher model is approximately 56M.

[0073] The teacher model is trained on a GPU server using the training set. The loss function is the standard classification cross-entropy loss. ; In the formula, The first of the real labels Unique thermal coding components, The first output of the teacher model Class probability value.

[0074] The optimizer uses AdamW with an initial learning rate of Weight decay coefficient The learning rate scheduling strategy employs cosine annealing with preheating: the learning rate for the first 10 cycles starts from... Linear preheating to It then decays according to the cosine curve to Total number of training cycles The batch size is 128. After each training round, the classification accuracy is calculated on the validation set, and the weight of the model with the highest validation accuracy is recorded as the final teacher model. The teacher model achieves approximately 94.3% five-class classification accuracy on the test set.

[0075] Phase Two: Knowledge Distillation Training Using the teacher model as the source of soft labels, the lesion hardness recognition network (i.e., the student model) is trained through knowledge distillation. The joint loss function used for distillation training is: ; In the formula, The probability vector output by the student model. and These are the logits vectors of the teacher model and the student model before Softmax, respectively. The equilibrium coefficient is set to 0.3. Distillation temperature. Set to 3. The specific calculation formula is as follows: ; This formula represents temperature. The Kullback-Leibler divergence under the following conditions is multiplied by The subsequent unfolding form, in which Indicates the Softmax term after the Softmax term. The components of a dimension.

[0076] Distillation training uses the same training and validation sets as the teacher model training. The optimizer is AdamW, with an initial learning rate of... Weight decay coefficient The learning rate uses cosine annealing with no preheating scheduling, from... decay to Total number of cycles The batch size is 256. The weights of the model with the highest validation accuracy are recorded as the distilled student model. This stage does not involve updating the teacher model parameters; the teacher model is only used for forward inference to generate soft labels. The distilled student model achieves approximately 92.1% five-class classification accuracy on the test set.

[0077] Phase Three: Quantitative Perception Fine-tuning The distilled student model is quantized using 8-bit integer quantization. The weight quantization scheme is channel-by-channel symmetric quantization, with the zero point fixed at 0 after quantization. The quantization range is [-127, 127], and the quantization formula is: ; In the formula, and These are floating-point weights and integer quantization weights, respectively. For the weight quantization step size, This represents the maximum absolute value of the weights within that channel.

[0078] The activation value quantization scheme is layer-by-layer asymmetric quantization, with a quantization range of [0, 255], and the quantization formula is: ; In the formula, and These are the floating-point activation value and the integer quantization activation value, respectively. Quantize the step size for the activation value. Quantize the zero point for the activation value.

[0079] Quantization-aware fine-tuning continues training for 20 epochs on the training set, using a pass-through estimator to approximate the gradient of the quantization operation during backpropagation. The optimizer employs SGD with a fixed learning rate. The batch size is 512, and weight decay is not used. After fine-tuning, the quantization model is exported as a C language header file in TensorFlow Lite Micro format and burned into the internal Flash of the main control chip.

[0080] The quantized, fine-tuned model achieves a five-class classification accuracy of approximately 91.5% on the test set, a decrease of only about 0.6 percentage points compared to the floating-point model. The quantized model requires approximately 230KB of storage on the Cortex-M7 microcontroller and has a single inference execution time of 1.8ms, meeting the real-time requirement of a 50ms sliding step.

[0081] 8. Complete calculation process of forward inference During actual intraoperative operation, the forward inference process of the lesion stiffness recognition network is integrated into the main loop of the main control chip, triggered by the sliding window update event. The complete single inference calculation process is as follows: Step 1: Window Data Reading and Normalization. Read the most recent data from the circular buffer. sampling points sequence, sequence sum sequence. The sequence is The adjacent differences of the sequences are set, with the first element filled with 0. Min-max normalization is performed on each of the three sequences within a window: each sequence is subtracted from its minimum value, then divided by the difference between the maximum and minimum values; if the maximum value equals the minimum value, it is set to zero. The three normalized sequences are stacked in the order of channel 1, channel 2, and channel 3 as the input tensor. This input tensor will be fed into the quantization convolution encoder in step three.

[0082] Step two, auxiliary feature calculation and normalization. Using the original data before normalization... sequence sum The sequence, calculated according to the formula defined in Part IV, comprises five physical characteristics. to .in The FFT calculation uses the 256-point real-valued FFT function provided by the CMSIS-DSP library. This function performs floating-point operations and takes 200 points as input. The sequence is zero-padded to 256 points, resulting in 129 complex frequency components. The square of the modulus of each frequency component is taken as its energy value. The frequency band from 20Hz to 50Hz is calculated (corresponding to frequency bin numbers approximately 5 to 13 in a 256-point FFT, due to the frequency resolution). The ratio of the energy sum of Hz to the energy sum of the entire frequency band (excluding DC component, i.e., bin 0) is obtained. After the calculation is complete, the mean value fixed during the training phase is used. and standard deviation Z-score standardization was performed on the five features to obtain . It will be sent to the gating fusion module in step six.

[0083] Step 3, multi-scale convolutional encoding. Convert the input tensor... The three channels are fed into the corresponding three multi-scale convolution branches. Each channel independently performs convolution, batch normalization, and ReLU activation operations in five parallel convolution paths. The convolution operation uses the 8-bit quantized convolution function provided by the CMSIS-NN library. or The outputs of each path are concatenated along the channel dimension, and the outputs of the three channels are concatenated again along the channel dimension. After pointwise convolution for dimensionality reduction, the encoded feature tensor is obtained. This step involves approximately 12.3M multiplication and addition operations.

[0084] Step four, bidirectional long short-term memory reasoning. The 200 time steps are sequentially fed into the bidirectional long short-term memory layer. The forward unit is processed in time steps. to The forget gate, input gate, candidate cell state, cell state, output gate, and hidden state are updated sequentially according to the six formulas given in Phase 2. The backward unit is updated step by step according to time steps. to The same formula is used to update in reverse. At each time step, the hidden states in both directions are concatenated into a 128-dimensional vector, ultimately yielding... This step involves approximately 3.3M multiplication and addition operations.

[0085] Step 5, multi-head self-attention reasoning. The 200 time step vectors are processed by the weight matrices of the four attention heads. , , Parallel computation of query, key-value mapping. For each attention head, calculate the attention score matrix. And apply a causal mask After Softmax normalization, and Weighted summation. The four header outputs are concatenated and then mapped using a mapping matrix. Obtain attention output Then, the sequence-level feature vector is obtained by averaging along the time dimension. This step involves approximately 2.1M multiplication and addition operations.

[0086] Step Six, Gating Fusion and Classification. The results obtained in Step Five... Compared with the result obtained in step two The vectors are concatenated into a 133-dimensional joint vector, and then a gating vector is generated through a gating network. ,at the same time via linear mapping The physical embedding vector is obtained, and the fusion vector is calculated according to the gating fusion formula. The fusion vector is passed sequentially through a 96-neuron fully connected layer, batch normalization, ReLU activation, a 5-neuron fully connected layer, and Softmax, outputting a five-dimensional probability vector. The computational workload of this step is approximately 0.4M multiplication-addition operations.

[0087] The total computational load for a single inference run is approximately 18.1M multiply-accumulate operations. On a 480MHz Cortex-M7 microcontroller using the CMSIS-NN quantization inference library, the measured execution time is 1.8ms, meeting the real-time constraint of a 50ms sliding step. The FFT calculation in step two is a floating-point operation, taking approximately 0.05ms. Because it is independent of network inference and has a small computational load, it does not affect the overall real-time performance.

[0088] 9. Classification Decision Making and Adaptive Control After inference is completed, the main control chip outputs the probability vector. Classification decision-making and adaptive control are performed according to the following rules.

[0089] set up This is the category index corresponding to the highest probability. If For the corresponding synovial fat pad type, maintain the preset surgical speed, with an allowable range of 5000 rpm to 8000 rpm, and use a sharp cutting mode. If For the meniscus fibrocartilage category, the target speed is adjusted to 3000 rpm, with the cutting edge in a sharp cutting mode. This speed has a clear transition boundary with the lower speed range of the synovial fat pad category. If For the corresponding hyaline cartilage type of joint, the main control chip automatically reduces the target speed to 400 rpm via a PID speed controller, with a speed error... Drive PID calculation, proportional coefficient Integral coefficient Differential coefficients At the same time, the buzzer sounds briefly every 2 seconds, and the green LED flashes slowly. For the corresponding type of osteophyte calcification, the motor will immediately stop and lock; the buzzer will sound continuously, and the LED red light will remain on. The operator must manually press the reset button on the handle to release the lock. For the corresponding ligament type, the target speed will be automatically reduced to 800 rpm and a prompt will be issued, with a short beep from the buzzer.

[0090] Furthermore, an uncertainty rejection mechanism based on probability entropy is introduced. The confidence entropy of the classification result is defined as: ; like and The number of bits indicates that the network's classification uncertainty for this window is too high; at this point, regardless of... For each value, a conservative strategy is adopted: the speed is automatically reduced to 800 rpm with a prompt, and the corresponding strategy is switched only after the classification confidence of the next window recovers. This mechanism can effectively filter out momentary misclassifications caused by the window crossing the boundary between two tissues or electromagnetic interference.

[0091] Through the aforementioned complete lesion hardness recognition and adaptive control system, this device achieves millisecond-level full-link automatic response from real-time motor current sampling, torque analysis, timing window construction, neural network forward inference, tissue category decision-making to motor speed closed-loop control. The sensing delay is no more than 2ms, the single inference decision delay is no more than 2ms, and the total closed-loop delay from the blade contacting the tissue to the motor completing speed adjustment is no more than 5ms, providing reliable intelligent safety assurance for arthroscopic surgery.

[0092] The technical effects of the present invention will be further explained below with reference to specific experimental data.

[0093] Example 1: Comparative Experiment on the Accuracy of Ex vivo Tissue Classification To verify the tissue identification capability of the lesion hardness recognition network described in this invention, a cutting and classification experiment of isolated porcine knee joint tissue was conducted. The experiment used the engineering prototype of the device of this invention and a control scheme based solely on a traditional torque-current threshold discrimination method for simultaneous data acquisition and classification. The control scheme will... The average value is compared sequentially with five preset fixed thresholds. If the value falls within a certain threshold range, the corresponding tissue category is output. There is no sliding window or neural network inference.

[0094] Three fresh porcine knee joint specimens were used as experimental materials. Under direct vision, a senior arthroscopy surgeon sequentially dissected 60 fragments from each of five tissue types: synovium and fat pad, meniscus fibrocartilage, articular hyaline cartilage, osteophytes and calcifications, and ligaments. Each fragment lasted approximately 5 seconds, resulting in 300 valid fragments. The collected data were randomly divided into a training set and a test set at a 7:3 ratio. The test set contained 90 fragments, approximately 4500 temporal windows. Using the tissue categories labeled by the surgeon as the gold standard, the classification accuracy of the device and the control scheme was calculated. Accuracy was defined as the percentage of correctly classified windows out of the total number of windows.

[0095] The experimental results are shown in Table 1.

[0096] Table 1 Comparison of Classification Accuracy of Ex vivo Tissues

[0097] As shown in Table 1, the device of this invention achieves an average classification accuracy of 91.5% for the five tissue types, which is approximately 19.8 percentage points higher than the traditional thresholding scheme. Particularly in the ligament and meniscus categories, whose mechanical characteristics are similar to soft tissues, the traditional thresholding scheme achieves an accuracy of only 60%–65%, while this invention, by utilizing temporal features and an attention mechanism, improves the accuracy to over 90%, indicating that the dual-level lesion stiffness identification algorithm of this invention has a significant improvement in classification performance. Figure 2 This is a bar chart comparing the accuracy of the device of the present invention and the control scheme in various tissue categories. The horizontal axis represents tissue category, and the vertical axis represents the percentage of classification accuracy. Dark bars represent the device of the present invention, and light bars represent the control scheme.

[0098] Example 2: Closed-loop response performance test To verify the closed-loop response performance of the device of this invention, from the contact of the cutting head with the tissue to the adaptive speed adjustment of the motor, a simulated tissue membrane was used to measure the response time. The membrane contained four regions with different hardnesses, simulating the synovium, meniscus, cartilage, and ligament, respectively. A robotic arm propelled the device head through each region sequentially at a constant speed. During the experiment, the main control chip recorded the response time from... The time from the initial exceedance of the no-load baseline to the motor's actual speed stabilizing at the target speed is defined as the closed-loop response delay. Simultaneously, three physicians with arthroscopic experience, unknowingly using the device to manually cut the body membrane, reported the delay in sensing changes in tissue stiffness and manually adjusting the speed via a foot switch. Each physician repeated the experiment 10 times, recording a total of 30 sets of manual response delay data. The statistical results of the closed-loop response delay and manual response delay of the device of this invention are shown in Table 2.

[0099] Table 2 Comparison of Closed-Loop Response Delay

[0100] As shown in Table 2, the average closed-loop response delay of the device of this invention is only 4.8 ms, with minimal fluctuation and a standard deviation of only 0.4 ms; while the average reaction delay of manual operation by a physician is 287.3 ms, with a maximum delay exceeding half a second, and significant individual differences. This indicates that the device of this invention not only has a response speed far superior to that of a human operator, but also possesses excellent stability and repeatability, and can complete protective actions almost instantaneously in the event of accidental contact, effectively avoiding damage to important cartilage and ligaments. Figure 3 This is a line graph comparing the closed-loop response delay distribution of the device of the present invention and manual operation. The horizontal axis represents the number of experiments, and the vertical axis represents the delay time from tissue contact recognition to speed adjustment completion, in milliseconds. The blue circular dotted line represents the device of the present invention, and the red square dotted line represents manual operation by the physician.

[0101] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A joint cavity rotary cutting hemostasis device with lesion hardness recognition, characterized in that, include: The handle, the working lever extending forward from the handle, and the control system built into the handle, wherein: The distal end of the working rod is the functional execution part, which adopts a four-layer concentric sleeve structure, consisting of a telescopic limiting outer sleeve, a stationary outer sheath, an injection fluid channel, and a rotating inner core rod from the outside to the inside. The retractable limiting outer sleeve is slidably sleeved on the outside of the stationary outer sheath and is equipped with a locking mechanism to fix the retractable limiting outer sleeve at any position in the axial direction of the outer sheath. The outer surface of the retractable limiting outer sleeve has scale markings, and the front end face of the retractable limiting outer sleeve forms an annular stop surface to limit the exposed length of the suction and cutting window opened on the distal sidewall of the outer sheath. The front end of the stationary outer sheath is fixed with an annular plasma radio frequency electrode, which is connected to the radio frequency energy interface at the rear end of the handle through an insulated wire embedded in the wall of the outer sheath. The infusion fluid channel is formed in the annular gap between the inner wall of the stationary outer sheath and the outer wall of the rotating inner core rod. Its proximal end is connected to the infusion fluid connector, and its distal end is connected to the micro-infusion hole opened on the side wall of the outer sheath. The rotating inner core rod is driven to rotate by a motor, and its distal sidewall is provided with a cutting edge, and the central channel forms a negative pressure suction channel. The control system includes a brushless DC motor that drives the rotating inner core rod, a current sampling module for collecting the phase current of the motor, a position sensor for acquiring the motor speed, a main control chip, and an alarm device. The main control chip is configured to: calculate torque current in real time from motor phase current and execute a two-level lesion hardness identification algorithm; wherein, the first layer is a fast-response protection layer, which executes motor emergency stop protection based on the comparison result of torque current change rate and preset threshold; the second layer is an intelligent classification layer, which extracts torque current time series data and auxiliary physical features in a sliding window manner, sends them to a lightweight multi-scale time series feature fusion neural network deployed inside the main control chip, outputs probability distributions corresponding to multiple tissue types, and adaptively adjusts motor speed or triggers shutdown and alarm according to probability distribution.

2. The joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The retractable limiting outer sheath is made of medical-grade 304 stainless steel, with its inner diameter fitting with the outer diameter of the stationary outer sheath. The locking mechanism on it is a locking knob with a radial thread connection. The scale markings are axially equidistant scale lines etched by laser.

3. The joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The stationary outer sheath is made of 316LVM stainless steel, and the suction and cutting window on its distal sidewall is an oblong or rounded rectangle with rounded edges. The annular plasma radiofrequency electrode is fixed on the hemispherical end face at the foremost end of the outer sheath. The electrode is made of platinum-iridium alloy and is covered with a ceramic insulating layer except for the surface facing the tissue. The insulated wire connecting the electrode is recessed into the axial groove on the outer wall of the outer sheath and extends proximally into the handle along the groove.

4. The joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The micro-infusion holes are opened on the side wall of the stationary outer sheath and are located upstream of the proximal end of the suction and cutting window. There are 4 to 6 of them. The axis of each micro-infusion hole forms an angle of 30° to 45° with the axis of the outer sheath and points towards the window area. The infusion fluid channel is connected to the infusion fluid connector at the proximal end through the distribution cavity. The infusion fluid connector is used to connect to an external infusion pump.

5. A joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The current sampling module is a three-phase lower bridge arm sampling resistor integrated on the motor drive board; the position sensor is a three-channel Hall sensor installed on the stator end face of the motor; the main control chip is a microcontroller based on the ARM Cortex-M7 core; the alarm device includes a buzzer and a dual-color LED indicator.

6. The joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The main control chip calculates the three-phase current in real time using Clarke and Park transformations. Axial torque current The torque current With motor electromagnetic torque They are directly proportional, and the relationship is: ,in, This represents the number of pole pairs of the motor. It is a permanent magnet flux linkage.

7. A joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The fast-response protection layer is executed in the current acquisition interrupt service program. When the absolute value of the difference between two adjacent torque current samples exceeds the preset hard change threshold, or the absolute value of the torque current exceeds the preset safety upper limit, the main control chip forcibly shuts down the motor drive signal and implements active short-circuit braking, while triggering the alarm device.

8. A joint cavity rotary cutting hemostasis device with lesion hardness identification according to claim 1, characterized in that, The lightweight multi-scale temporal feature fusion neural network includes a three-channel multi-scale convolutional encoder, a bidirectional long short-term memory network layer, a multi-head self-attention layer, a gated feature fusion layer, and a fully connected classification layer connected in sequence. The three-channel multi-scale convolutional encoder receives the normalized torque-current sequence, torque-current differential sequence, and motor speed sequence within the receiving window as three independent input channels. Each channel is processed by multiple parallel convolutional paths with different convolutional kernel sizes or dilation rates before being spliced ​​and dimension-reduced. The bidirectional long short-term memory network layer extracts bidirectional temporal dependencies from the encoded feature sequence; The multi-head self-attention layer performs time-weighted focusing on the bidirectional long short-term memory output, and obtains a sequence-level feature vector after global average pooling; The gated feature fusion layer adaptively fuses the sequence-level feature vectors with normalized auxiliary physical features using learnable gated vectors; The fully connected classification layer maps the fusion vector into probability vectors corresponding to five tissue categories: synovium and fat pad, meniscus fibrocartilage, articular hyaline cartilage, osteophytes and calcifications, and ligaments.

Images