Intraoperative risk assessment system for cataract surgical robot

Abstract

The invention relates to the field of surgical robots, and particularly discloses an intraoperative risk assessment system for a cataract surgical robot, comprising: a patient biomechanical monitoring unit for collecting a contact force between an instrument and an eye tissue, and calculating an anterior chamber equivalent pressure value according to an axial force component of the contact force and a contact area between the instrument and a cornea; the patient optical imaging monitoring unit is used for collecting microstructure images of the cornea and the lens capsule and quantifying the energy release density of ultrasonic emulsification; the robot motion parameter monitoring unit is used for collecting joint motion resistance, instrument tail end three-dimensional coordinates, motion speed and attitude angle of the robot; the multi-mode joint reasoning unit is used for identifying a high-risk operation state when the contact force is higher than a first preset value and the energy release density is higher than a second preset value; by adopting the technical scheme of the invention, multi-dimensional real-time monitoring can be realized, and a dynamic risk prediction model can be established by fusing heterogeneous data through an intelligent algorithm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of surgical robots, in particular to an intraoperative risk assessment system for a cataract surgery robot. BACKGROUND

[0002] As the most common ophthalmic surgery worldwide, the operation precision and safety of cataract surgery directly affect the postoperative visual recovery of patients. At present, the prior art has realized the preoperative planning of cataract surgery, for example, CN107361737A realizes preoperative lens tilt measurement by laser phacoemulsification combined with multi-slice image reconstruction of the three-dimensional graphics of the lens, but lacks intraoperative dynamic monitoring capability, and cannot respond to real-time changes in tissue deformation and energy release during the operation, which is not convenient for the operating physician to make intraoperative adjustments.

[0003] In view of the safety problem during the operation, the probe sensing function of CN113813099A can only trigger passive obstacle avoidance through a contact force threshold (such as stopping movement when the distance from the corneal endothelium is <0.5mm), but does not establish a risk prediction model based on the fusion of multiple parameters, and cannot provide early warning of potential complications. Although the force-position hybrid control method of CN119655901A introduces a mobility controller to adjust the operation instructions, it only relies on force error values to optimize motion parameters, and relies more on the fine operation of the operating physician, and does not integrate the actual performance of the patient during the operation, resulting in a single evaluation dimension.

[0004] In order to improve the operation ability of the operating physician, AI video analysis such as CN115223071A is used to improve the operation standardization of the operating physician, but does not associate physiological parameters with mechanical arm motion trajectories, and ignores the risk representation value of real-time data during the operation.

[0005] These technical defects make it difficult for the operator (operating physician) to dynamically evaluate the risk in key steps such as phacoemulsification, and require reliance on experience to judge, resulting in high human operation errors, late complications prevention and control, and other problems. In the context of the operating physician using a cataract surgery robot, more assistance needs to be provided to the operating physician to reduce the likelihood of operating physician errors. Current cataract surgery robot technology has realized the combination of force feedback mechanical arms and microscopic imaging technology (such as CN113813099A, which realizes obstacle avoidance of the phacoemulsification probe through sound ranging and capacitive sensing, and CN119564348A, which realizes remote center positioning of surgical instruments using RCM mechanisms), but more in the context of completing the operation.

[0006] Therefore, on the basis of the existing cataract surgery robot, there is an urgent need for an intraoperative risk assessment system for a cataract surgery robot that can realize multi-dimensional real-time monitoring and establish a dynamic risk prediction model by fusing heterogeneous data through intelligent algorithms. SUMMARY

[0007] The application provides an intraoperative risk assessment system for a cataract surgery robot, which can realize multidimensional real-time monitoring and can establish a dynamic risk prediction model by intelligently fusing heterogeneous data.

[0008] To solve the above technical problems, the application provides the following technical solutions:

[0009] The intraoperative risk assessment system for the cataract surgery robot comprises:

[0010] A patient biomechanics monitoring unit is configured to collect contact force between the instrument and the ocular tissue, and to calculate an anterior chamber equivalent pressure value by using an axial force component of the contact force and a contact area between the instrument and the cornea;

[0011] A patient optical imaging monitoring unit is configured to collect microscopic structure images of the cornea and the lens capsule and to quantify energy release density of phacoemulsification;

[0012] A robot motion parameter monitoring unit is configured to collect joint motion resistance of the robot, three-dimensional coordinates of the instrument end, motion speed and attitude angle;

[0013] A multimodal joint inference unit is configured to identify a high-risk operation state when the contact force is higher than a first preset value, the energy release density is higher than a second preset value, and the anterior chamber equivalent pressure value is lower than a third preset value; and to obtain preliminary data by time-series correlation of the joint motion resistance, the three-dimensional coordinates of the instrument end, the motion speed and the attitude angle, the contact force, the energy release density and the microscopic structure images, and to judge a time delay of the preliminary data by using a pre-established historical data causal chain model, and to generate a potential risk transmission path warning if the time delay exceeds a threshold value, and to extract features from the preliminary data to obtain feature data if the time delay does not exceed the threshold value;

[0014] The risk assessment unit is configured to dynamically weight and fuse the feature data, and to output short-term, medium-term and long-term risk prediction results according to a preset risk assessment model; the risk prediction results comprise parameters and prediction probabilities of capsule fold degree, anterior chamber equivalent pressure and energy cumulative value, and risk assessment results of posterior capsule rupture, anterior chamber collapse and complication results;

[0015] An operation suggestion generation unit is configured to generate safe operation guidelines and emergency braking strategies comprising energy segmentation modulation, spiral retreat path planning and liquid flow-energy collaborative regulation according to the risk assessment results.

[0016] The basic scheme principle and beneficial effects are as follows: the present application converts micro-mechanical signals into macro-physiological indexes by calculating the anterior chamber equivalent pressure through the axial component of contact force and the corneal contact area, and realizes real-time quantitative monitoring of intraocular pressure stability. By collecting microstructure images of the cornea and the capsule, combining energy release density quantification, a two-dimensional monitoring matrix of tissue morphology-energy distribution is constructed to capture the tissue changes during phacoemulsification. The joint resistance of the robot, the instrument coordinates and the attitude angle are collected to form a six-degree-of-freedom monitoring system of the instrument motion trail, and the dynamic tracking of the operation precision is realized. At the same time, the biological mechanics, optical structure, energy release and instrument movement are monitored to form a three-dimensional monitoring system of "tissue-energy-mechanical", which improves the risk identification accuracy compared with single-dimensional monitoring.

[0017] The present application judges the high-risk state through the triple threshold (contact force, energy density, anterior chamber pressure), simultaneously generates a delay by using time sequence correlation analysis data, identifies the risk transmission path (such as the potential link of "contact force mutation-energy overload-capsular stress concentration") based on the historical causal chain model, and adaptively weights and fuses multi-source feature data (such as increasing the weight of biological mechanics data when the anterior chamber pressure is abnormal), combines short-term (millisecond level), medium-term (second level) and long-term (operation stage) prediction models, and outputs a probabilistic evaluation result.

[0018] The present application adjusts the ultrasonic power output curve according to the real-time risk, realizes safe evacuation through the spiral retreat trail, and avoids the tissue pulling caused by linear retreat.

[0019] The present application predicts the risk of posterior capsule rupture in advance through the causal chain model (the traditional threshold trigger can only respond in real time), the prediction time is advanced, the intervention adjustment is carried out in advance, and the probability of complications is reduced. The short-term prediction guides the millisecond-level operation adjustment, the medium-term prediction optimizes the stage switching strategy, and the long-term prediction assists in the prevention and control of postoperative complications, forming a whole-cycle risk prevention and control system.

[0020] The present application establishes the causal relationship of "mechanics-optics-motion" data through time sequence delay analysis, breaks through the traditional parallel processing mode, realizes the dynamic tracking of the risk evolution process, constructs multi-time scale prediction models, and enables the system to respond to emergency risks (such as instrument collision) in milliseconds while optimizing the overall operation strategy in minutes.

[0021] In summary, the present application realizes multi-dimensional real-time monitoring, can establish a dynamic risk prediction model by fusing heterogeneous data through intelligent algorithms, reduces the operation difficulty of the operating doctor, improves the experience of doctors and patients, and reduces the medical risk.

[0022] Further, the patient biomechanics monitoring unit collects contact force through a thin film six-dimensional force sensor, and the anterior chamber equivalent pressure value is calculated by the formula P = F / A, where F is the axial force component, and A is the preset contact area between the device and the cornea. The pressure calculation resolution is ≤0.05 kPa, and the contact force collection accuracy is ≤0.05 N.

[0023] Further, the patient optical imaging monitoring unit comprises:

[0024] A miniature swept-source OCT probe with an axial resolution of ≤5 μm and a scanning frequency of ≥200 frames / s is used to extract the microstructure features of the capsule fold degree and the change rate of the posterior capsule curvature radius according to a preset microimage recognition model.

[0025] An infrared energy sensor is used to monitor the ultrasonic emulsification energy release density, with an accuracy of ≤5 mJ / pulse, and a second preset value of the energy release density is 80 mJ / pulse.

[0026] Further, the robot motion parameter monitoring unit comprises:

[0027] A joint torque sensor is used to collect joint motion resistance with an accuracy of ≤0.1 N·m;

[0028] A grating displacement sensor and an inertial measurement unit are used to solve the three-dimensional coordinates, motion speed and attitude angle of the device tip.

[0029] Further, the high-risk operation state identification logic of the multi-modal joint inference unit is: when the contact force > 1.2 N, the energy release density > 90 mJ / pulse, and the anterior chamber equivalent pressure value < 1.5 kPa, it is determined as a high-risk operation state and an orange pre-warning is triggered.

[0030] Further, the historical data causal chain model is a third-order Markov chain model trained based on a number of clinical data, which is used to analyze the time sequence correlation of "contact force mutation → energy release anomaly → capsule deformation", and the preset delay threshold is: force mutation to energy anomaly ≤ 150 ms, and energy anomaly to capsule deformation ≤ 300 ms.

[0031] Further, the dynamic weighted fusion of the risk assessment unit adopts an adaptive gating mechanism, and the expression is:

[0032] F fusion =σ(FC([F 力 , F 运动 , F 图像 , F 能量 ]))⊙(F 力 +F 运动 +F 图像 +F 能量 )

[0033] Wherein, sigma is a Sigmoid activation function, FC is a fully connected layer, and the force feature weight is automatically increased to more than 60% when the current anterior chamber equivalent pressure value is less than 2 kPa by learning to automatically assign weights.

[0034] Further, the risk assessment model is a CNN-LSTM hybrid neural network:

[0035] The CNN branch is used for extracting spatial features of the capsule membrane wrinkle degree, the posterior capsule membrane curvature radius change rate, the capsule membrane edge roughness and the anterior chamber bubble volume in the microstructure image;

[0036] The LSTM branch is used for analyzing the time sequence fluctuation of the contact force, the motion speed and the energy release density, and outputting the risk probability of the posterior capsule rupture and the anterior chamber collapse.

[0037] The bubble volume does not directly represent the anterior chamber pressure, and the anterior chamber bubble is mainly generated by the gasification of the perfusion liquid caused by the energy release in the phacoemulsification process. The bubble volume is positively correlated with the ultrasonic energy density (when the energy density is greater than 80 mJ / pulse, the bubble volume growth rate significantly rises). A small amount of bubble (volume less than 0.1 mm 3 ) is allowed to exist in the normal operation, but excessive bubble (>0.5 mm 3 ) will cause the problem of visual field obstruction and anterior chamber stability destruction. The anterior chamber bubble volume is a key indicator reflecting the ultrasonic energy management, and forms a complementary monitoring with the anterior chamber pressure, and together constitutes an energy and fluid double-dimensional risk assessment system, which improves the early warning ability of the system to energy-related complications.

[0038] Further, the safety operation guide of the operation suggestion generation unit comprises:

[0039] Energy segmented modulation: limit the single pulse energy to be less than or equal to 60 mJ when the risk is medium, and forcibly reduce the ultrasonic power to 50% of the basic power when the risk is high;

[0040] Spiral retreat path planning: generate a spiral path based on the RCM mechanism, and the instrument end is withdrawn at a radius of 0.1 mm and a speed of 0.2 mm / s, and the deflection angle adjustment is less than or equal to 10°;

[0041] Liquid flow-energy collaborative regulation: automatically increase the perfusion flow by 5 mL / min and reduce the energy release density by 5 mJ / pulse for every 0.5 kPa decrease of the anterior chamber equivalent pressure value, until the pressure returns to the safe interval of 6-8 kPa.

[0042] Further, the emergency braking strategy comprises the following contents:

[0043] Start the electromagnetic brake of the motor and lock the mechanical limiting structure;

[0044] Through the voice module, "stop operation immediately, execute capsule membrane integrity check" is broadcasted, and a red warning box is superimposed on the surgical field to mark the risk area. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 A simple logic block diagram of an intraoperative risk assessment system for a cataract surgery robot;

[0046] Figure 2 A complete logic block diagram of an intraoperative risk assessment system for a cataract surgery robot. DETAILED DESCRIPTION

[0047] Further details are described below through specific embodiments:

[0048] An intraoperative risk assessment system for a cataract surgery robot (as shown in Figure 1 , Figure 2 includes:

[0049] A patient biomechanics monitoring unit for collecting the contact force of the instrument and the eye tissue, and calculating the anterior chamber equivalent pressure value through the axial force component of the contact force and the contact area of the instrument and the cornea;

[0050] A patient optical imaging monitoring unit for collecting microstructure images of the cornea and the lens capsule and quantifying the energy release density of phacoemulsification;

[0051] A robot motion parameter monitoring unit for collecting the joint motion resistance of the robot, the three-dimensional coordinates of the instrument tip, the motion speed and the attitude angle;

[0052] A multi-modal joint inference unit for identifying a high-risk operation state when the contact force is higher than a first preset value, the energy release density is higher than a second preset value, and the anterior chamber equivalent pressure value is lower than a third preset value; and for performing time sequence correlation on the joint motion resistance, the three-dimensional coordinates of the instrument tip, the motion speed and the attitude angle, the contact force, the energy release density, and the microstructure images to obtain preliminary data, and for using a pre-established historical data causal chain model to judge the delay of the generation time in the preliminary data, and if the delay exceeds a threshold, a potential risk transmission path warning is generated; if the delay does not exceed the threshold, feature extraction is performed on the preliminary data to obtain feature data

[0053] The risk assessment unit is configured to dynamically weight and fuse the feature data, and output short-term, medium-term and long-term risk prediction results according to a preset risk assessment model; the risk prediction results include parameters and prediction probabilities of capsule membrane wrinkle degree, anterior chamber equivalent pressure, and energy cumulative value, as well as risk assessment results of posterior capsule rupture, anterior chamber collapse, and complications;

[0054] The operation suggestion generation unit is configured to generate, for the risk assessment result, a safe operation instruction containing energy segmentation modulation, spiral retreat path planning, liquid flow-energy collaborative regulation, and an emergency braking strategy.

[0055] In specific use, the patient biomechanics monitoring unit is a thin film six-axis force sensor (model FSR-600, accuracy 0.05N) integrated at the end of the phacoemulsification probe, and the contact area of the sensor with the cornea is preset to 0.8mm 2 (adjusted dynamically according to the model of the instrument), the anterior chamber equivalent pressure is calculated by the formula P=F / A, and the pressure resolution is 0.05kPa. The X / Y / Z axis force and torque are collected at a sampling frequency of 1000Hz, and when the axial force (Z axis) fluctuates more than 0.1N, the pressure calculation module triggers to update the value of the anterior chamber equivalent pressure.

[0056] Specifically, in the patient biomechanics monitoring unit, the end of the phacoemulsification probe and the thin film six-axis force sensor are integrated. The specific connection relationship is that in the shell at the end of the phacoemulsification probe, a dedicated sensor mounting chamber is provided, and the thin film six-axis force sensor is fixed at the center position of the chamber through a high-precision positioning clamp. The sensing surface of the sensor is parallel and rigidly connected to the surgical contact surface of the probe end, ensuring that the contact force between the instrument and the eye tissue during the operation can be transmitted to the sensor sensing area without loss. At the same time, the data line of the sensor is connected to the main control system through the shielded wire slot reserved in the probe, realizing stable transmission of data.

[0057] Because the transducer at the end of the phacoemulsification probe generates high-frequency ultrasonic vibration when working, this vibration will interfere with the measurement of the thin film six-axis force sensor, causing the reading to deviate. To solve this problem, the thin film six-axis force sensor adopts a multi-layer composite filtering processing method. On the hardware level, the sensor integrates a low-pass filter circuit inside, which can effectively suppress the high-frequency vibration noise generated by the transducer, and attenuate the interference signal above 1000Hz by more than 80%; on the software level, the main control system uses an adaptive Kalman filter algorithm, combined with a preset surgical operation force model, to process the data collected by the sensor in real time. By establishing a state space model of the force signal, the filter parameters are dynamically adjusted, which can further filter out residual periodic interference noise while preserving the true force signal, so that the contact force measurement accuracy is stably maintained at ≤0.05N, ensuring the accuracy and reliability of the biomechanics monitoring data.

[0058] Specifically, a thin film six-axis force sensor (such as the FSR-600 series) is used, which has an axial force (Z axis) measurement accuracy of 0.05N, combined with a preset instrument and corneal contact area A=0.8mm 2 (the contact area of a commonly used phacoemulsification probe is estimated as πr 2 ≈0.8mm 2The pressure is calculated by the formula P = F / A:

[0059]

[0060] 0.05 kPa resolution can capture 0.375 mmHg pressure fluctuations, while clinical studies have shown that the intraocular pressure (intraocular equivalent pressure) drops > 0.5 kPa (3.75 mmHg) need intervention.

[0061] The contact force acquisition accuracy is ≤ 0.05 N. The technical implementation is that the standard range of the FSR-600 sensor is 0-1 N, the non-linear error is ≤ 5%, and in the 0.1-0.5 N force range commonly used in surgery, the accuracy can reach 0.05 N (5% x 1 N = 0.05 N).

[0062] The safe contact force threshold for capsulorhexis operation in cataract surgery (although it is completed by using a capsulorhexis forceps, only different operation ends, it can also be detected by a thin film six-dimensional force sensor) is usually set to 0.8-1.2 N, and the 0.05 N accuracy can ensure that the warning is within 5% of the safe threshold, avoiding tissue damage caused by force perception errors.

[0063] The patient optical imaging monitoring unit is a miniature swept-source OCT probe (central wavelength 1050 nm, axial resolution 5 μm, scanning frequency 200 frames / s) and an infrared energy sensor (accuracy 5 mJ / pulse) integrated in the optical path of the surgical microscope.

[0064] Image analysis process: OCT images are segmented in real time by the U-Net++ network, and features such as capsular fold degree (threshold 15°) and posterior capsule curvature radius change rate (threshold 0.2 mm / s) are extracted; the infrared sensor synchronously monitors the energy release density, and when the density > 80 mJ / pulse, it is marked as an energy anomaly.

[0065] The robot motion parameter monitoring unit includes joint torque sensors (accuracy 0.1 N·m) installed on each drive joint of the robot, grating displacement sensors (linear accuracy 0.02 mm) and inertial measurement units (IMU) integrated in the instrument end connector.

[0066] The motion solving algorithm is to fuse joint angle, grating displacement and IMU data through Kalman filtering, and to solve the three-dimensional coordinates of the instrument end (error ≤ 0.05 mm), motion speed (resolution 0.1 mm / s) and attitude angle (deflection angle ≤ 0.5°) in real time.

[0067] When the contact force = 1.3N (> first preset value 1.2N), the energy release density = 95mJ / pulse (> second preset value 90mJ / pulse), and the anterior chamber equivalent pressure = 1.4kPa (< third preset value 1.5kPa) are monitored, the multi-modal joint inference unit triggers an orange warning, and simultaneously starts a causal chain analysis.

[0068] Based on a third-order Markov chain model trained on several clinical data (specifically determined according to the data situation), the time sequence correlation is analyzed: if the force mutation (amplitude > 30%) to energy anomaly (amplitude > 20%) delay = 140ms (≤ threshold 150ms), and the energy anomaly to the capsule membrane deformation (wrinkle degree > 15°) delay = 280ms (≤ threshold 300ms), it is determined as "high-risk conduction path activation", and a potential risk warning is generated.

[0069] When the anterior chamber equivalent pressure = 1.8kPa (< 2kPa) is detected, the adaptive gating mechanism automatically increases the force feature weight to 65%, and the fusion formula is:

[0070] F fusion = σ (FC([F 力 , F 运动 , F 图像 , F 能量 ])) ⊙ (F 力 + F 运动 + F 图像 + F 能量 )

[0071] The CNN-LSTM model outputs hierarchical prediction results:

[0072] Short-term (0-500ms): posterior capsule rupture probability = 28% (threshold 25%, trigger yellow warning);

[0073] Medium-term (500ms-2s): anterior chamber collapse probability = 32% (threshold 30%, trigger orange warning);

[0074] Long-term (> 2s): postoperative corneal edema probability = 15% (auxiliary postoperative treatment planning).

[0075] When a medium-risk state (posterior capsule rupture probability = 35%) occurs, the ultrasonic power is limited to: basic power 60W + single pulse energy ≤ 60mJ (original pulse energy 80mJ), and the pulse interval is extended to 200ms (original interval 150ms).

[0076] Then generate a spiral trajectory based on the RCM mechanism, with the center axis as the remote center of motion, the radius 0.1mm, and the evacuation speed 0.2mm / s; the deflection angle is limited within 8°, and the instrument end height is adjusted by 0.05mm every 180° rotation to avoid the pulling of the capsule membrane by linear retreat.

[0077] When the anterior chamber equivalent pressure is monitored to be 1 kPa (0.5 kPa decrease from the lower limit of the safe range of 1.5 kPa): the perfusion flow is automatically increased by 5 mL / min (original flow of 20 mL / min to 25 mL / min); the energy release density is reduced by 5 mJ / pulse (original density of 80 mJ / pulse to 75 mJ / pulse), until the pressure recovers to 6.2 kPa.

[0078] When the red warning (anterior chamber collapse probability = 40%) is triggered, the electromagnetic brake (braking distance 0.02 mm) and mechanical limit locking are started within 0.08 s; the surgical field (display content within the screen) is superimposed with a red translucent mask, marking the risk area coordinates (X = 1.2 mm, Y = 0.8 mm, Z = 3.5 mm), and the voice broadcast is "stop energy output immediately, perform capsule membrane integrity check".

[0079] In the ultrasonic emulsification nucleus fragmentation stage, the following data is collected:

[0080] Contact force Z axis = 1, 1N, anterior chamber equivalent pressure = 1.8 kPa;

[0081] Capsule membrane wrinkle degree = 12°, energy release density = 75 mJ / pulse;

[0082] Instrument tip speed = 0.3 mm / s, deflection angle = 3°.

[0083] Multi-modal joint reasoning: contact force, energy density, and anterior chamber pressure are within the safe threshold, and the causal chain delay is <100 ms;

[0084] The risk prediction is short-term posterior capsule rupture probability = 12%, medium-term anterior chamber collapse probability = 8%. The operation suggestion is to maintain the current energy output (75 mJ / pulse), and the instrument movement speed is suggested to be reduced to 0.2 mm / s to avoid force mutations caused by rapid movement; the liquid flow system maintains a perfusion flow of 20 mL / min, and the anterior chamber pressure fluctuation is monitored in real time.

[0085] In other embodiments, in the multi-modal joint reasoning unit, the preliminary data of time sequence correlation is aligned through a dynamic time warping (DTW) algorithm, and the historical data causal chain model is a risk transmission probability graph constructed based on a Bayesian network, and the conditional probability expression is:

[0086]

[0087] Wherein, Rj represents the jth risk transmission path, X1, X2,..., X n is the preliminary data feature vector of time sequence correlation, and m is the total number of preset risk paths; the delay judgment of the generated time is determined by a threshold function Implementation, when f (Δt) = 1 triggers potential risk conduction path warning.

[0088] DTW algorithm is used to solve the problem of inconsistent sampling frequency of multi-modal data (such as force sensor 1000Hz, OCT image 200Hz), through dynamic programming to calculate the optimal time mapping path, ensure the time sequence correlation accuracy of contact force mutation and energy anomaly ≤5ms, lower error than traditional linear interpolation method.

[0089] Using several clinical data to train Bayesian network, construct the conditional probability table of "contact force mutation→energy overload→capsular membrane stress concentration" risk conduction path, for example, the conditional probability of posterior capsule rupture risk path P (posterior capsule rupture | force mutation∩energy overload) =0.82, the risk conduction recognition accuracy is improved by 15% compared with Markov chain model.

[0090] Through threshold t thres (For example, the threshold of force mutation to energy anomaly is 150ms) to quantify the timeliness of risk conduction, when the measured delay Δt = 160ms > t thres , f (Δt) = 1 triggers warning, avoid risk evolution to irreversible stage.

[0091] The joint resistance, instrument coordinate and other 6 types of data are normalized, and the DTW algorithm is used to map different sampling rate data to a unified time axis, for example, the 1000Hz force data and the 200Hz OCT image are aligned according to the action cycle, to ensure the accuracy of time sequence correlation.

[0092] Input the time sequence aligned feature vector X = [contact force change rate, energy density, anterior chamber pressure, instrument speed], calculate the posterior probability P (Rj | X) of each risk path through Bayesian network, when the maximum posterior probability path P (Rj | X) > 0.7 and delay Δt > t thres , generate visual risk conduction path graph (such as red arrow marked "force mutation→capsular membrane wrinkle" conduction link).

[0093] Threshold decision: through piecewise function f (Δt) to realize binary decision, when Δt = 180ms (threshold of force mutation to capsular membrane deformation 300ms), f (Δt) = 0, only feature extraction; when Δt = 320ms > 300ms, f (Δt) = 1, immediately trigger yellow warning and mark the risk conduction node.

[0094] DTW algorithm solves the problem of asynchronous multi-modal data, makes the time sequence correlation error of force-energy-image ≤5ms, improves the risk conduction recognition efficiency by 40% compared with traditional interpolation method, avoids misjudgment caused by asynchronous data.

[0095] Bayesian network quantifies the conditional probability of each path, for example, when a contact force mutation (40% increase) and energy density = 95 mJ / pulse are detected, the P(Rj|X) of the posterior capsule rupture risk path is calculated = 0.85, 2.5 s early warning, 1.5 s earlier than threshold trigger technology. By adaptive adjustment of the warning strategy through piecewise function f(Δt), personalized risk monitoring is achieved in elderly patients (poor tissue elasticity, delay threshold can be relaxed to 180 ms) and young patients (good tissue elasticity, threshold is strictly set to 150 ms), and the warning accuracy is improved.

[0096] Generate risk transmission probability heat map to mark the risk probability of each node with different colors (such as red for contact force mutation node and orange for energy overload node), which helps the operator understand the risk evolution process and shorten the decision-making time. By introducing DTW algorithm, Bayesian network and threshold function, the deficiencies of multi-modal data time alignment accuracy and weak risk transmission quantification ability are filled, realizing the technical leap from "data monitoring" to "causal reasoning", and providing a solution for intraoperative risk prediction.

[0097] In another embodiment, in the risk assessment unit, the dynamic weighted fusion adopts a spatio-temporal feature fusion network (ST-FusionNet) based on attention mechanism, and the weight calculation expression is:

[0098]

[0099] Wherein, α i is the weight of the i-th feature (force / motion / optical / energy), Attention(F i ) is the attention score of feature F i , which is calculated by a double-layer perception machine (MLP):

[0100] Attention(F i )=o(W2·tanh(W1·F i +b1)+b2)

[0101] The short-term, medium-term and long-term risk prediction is realized by a three-dimensional convolution long short-term memory network (3D-CLSTM), and the state transition equation is:

[0102]

[0103] Wherein, f t / i t / o t is the forget gate / input gate / output gate, c t / h t is the cell state / hidden state, x tFor the fused feature vector, the time window is set to short-term (t = 1-50), medium-term (t = 51-200), and long-term (t = 201-500).

[0104] The feature importance score is calculated by MLP. When the current intraocular pressure is less than 1.5 kPa, the attention score of the force feature increases to 0.72 (optical feature decreases to 0.28), and the weight α i is automatically adjusted to 65%, reducing the risk assessment error compared to fixed weight algorithms (such as force / optical 50% each).

[0105] The three-dimensional convolution kernel (3x3x3) captures both spatial features (such as capsular fold degree) and time dimension (500 frame sequence), the short-term memory window (50 frames) captures millisecond-level force mutations, and the long-term window (500 frames) analyzes energy accumulation effects, improving the prediction accuracy of multiple time scales compared to traditional LSTM.

[0106] Short-term prediction focuses on instrument collision (time window 1-50 ms), medium-term prediction focuses on energy accumulation damage (51-200 ms), and long-term prediction evaluates postoperative complications (201-500 ms), forming full-time axis risk coverage.

[0107] ST-FusionNet aligns the time dimension of force features (1000 Hz) and optical features (200 Hz) and dynamically allocates weights through attention mechanism. For example, when detecting energy density = 90 mJ / pulse and intraocular pressure = 5.5 kPa, the attention score of the force feature is calculated by MLP as 0.81, and the weight α i = 73%, giving priority to the fusion of contact force fluctuation features.

[0108] Short-term prediction (1-50 frames): input force mutation features (such as Z-axis force increase of 40%), extract spatial gradients through 3D convolution, LSTM unit remembers force change trend in the past 50 ms, output posterior capsule rupture probability (probability at time step t = 50 is 28%).

[0109] Medium-term prediction (51-200 frames): cumulative energy features (such as total energy of 100 pulses = 800 mJ) are input, 3D-CLSTM captures the time sequence correlation between energy and tissue damage, and the probability of anterior chamber collapse at t = 200 is 35%.

[0110] Long-term prediction (201-500 frames): combined with instrument motion trajectory (such as spiral path completion), predicts the probability of postoperative corneal edema (probability at t = 500 is 12%), and assists in postoperative treatment planning.

[0111] Attention mechanism makes the system automatically enhance the weight of key features under complex conditions (such as energy fluctuation ± 25%), and the risk assessment error is stable at ≤4%, which improves the robustness compared with fixed weight algorithm.

[0112] 3D-CLSTM simultaneously processes millisecond-level force mutations and minute-level energy accumulation, and the prediction time of posterior capsule rupture and anterior chamber collapse is advanced, providing more sufficient intervention time for the operator.

[0113] 3×3×3 convolution kernel extracts spatial dimension (capsular fold direction) and time dimension (wrinkle rate of change) at the same time, and the risk identification accuracy of nonlinear motion (such as instrument rotation 7°) is improved. Through Attention mechanism, the force feature weight of elderly patients (poor tissue elasticity) is automatically increased to 70%, and the optical feature weight of young patients (good elasticity) is increased to 60%, realizing individualized risk assessment, and improving the warning coincidence rate.

[0114] By introducing ST-FusionNet attention fusion and 3D-CLSTM time series model, the problems of fixed multi-modal feature weight and single time scale prediction are solved, realizing the technical upgrade from "data fusion" to "intelligent inference", and providing support for multi-dimensional and full-period intraoperative risk prediction.

[0115] The above is only an embodiment of the present application, which is not limited to this embodiment. The specific structure and characteristics of the scheme known in the art are not described in detail, and the ordinary skilled person in the art knows all the ordinary technical knowledge in the art before the filing date or the priority date, can know all the prior art in the art, and has the ability to apply conventional experimental means before the date. The ordinary skilled person in the art can improve and implement the scheme under the guidance of this application, and some typical known structures or known methods should not be an obstacle for the ordinary skilled person in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should be considered as the protection scope of the present application, which will not affect the effect and practicality of the patent. The protection scope of this application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. An intraoperative risk assessment system for cataract surgery robots, comprising, characterized in that, The patient biomechanical monitoring unit is used to collect the contact force between the instrument and the eye tissue, and to calculate the anterior chamber equivalent pressure value by the axial force component of the contact force and the contact area between the instrument and the cornea. The patient optical imaging monitoring unit is used to acquire microscopic structural images of the cornea and lens capsule and quantify the energy release density of phacoemulsification. The robot motion parameter monitoring unit is used to collect data on the robot's joint motion resistance, the three-dimensional coordinates of the end effector, motion speed, and attitude angle. The multimodal joint reasoning unit is used to identify high-risk operation states when the contact force is higher than a first preset value, the energy release density is higher than a second preset value, and the anterior chamber equivalent pressure is lower than a third preset value. It also performs time-series correlation on joint movement resistance, three-dimensional coordinates of the instrument end, movement speed and posture angle, contact force, energy release density, and microstructure images to obtain preliminary data. It uses a pre-established historical data causal chain model to judge the generation time delay in the preliminary data. If it exceeds a threshold, it generates a potential risk transmission path warning. If the threshold is not exceeded, feature extraction is performed on the preliminary data to obtain feature data; The risk assessment unit is used to dynamically weight and fuse feature data, and output short-term, medium-term and long-term risk prediction results according to the preset risk assessment model. The risk prediction results include parameters and prediction probabilities of posterior capsule wrinkle degree, anterior chamber equivalent pressure, and cumulative energy value, as well as risk assessment results of posterior capsule rupture, anterior chamber collapse risk and complication results. The operation suggestion generation unit is used to generate safe operation guidelines and emergency braking strategies based on the risk assessment results, including energy segmented modulation, spiral retreat path planning, and fluid-energy coordinated regulation.

2. The intraoperative risk assessment system for cataract surgery robots according to claim 1, characterized in that, The patient biomechanical monitoring unit collects contact force through a thin-film six-dimensional force sensor. The anterior chamber equivalent pressure value is calculated using the formula P = F / A, where F is the axial force component and A is the preset contact area between the instrument and the cornea. The pressure calculation resolution is ≤0.05 kPa and the contact force acquisition accuracy is ≤0.05 N.

3. The intraoperative risk assessment system for cataract surgery robots according to claim 2, characterized in that, The patient optical imaging monitoring unit includes: A miniature swept-frequency OCT probe with an axial resolution ≤5μm and a scanning frequency ≥200 frames / second is used to extract microstructural features such as capsule wrinkle degree and posterior capsule curvature radius change rate based on a preset microscopic image recognition model. An infrared energy sensor is used to monitor the energy release density of ultrasonic emulsification with an accuracy of ≤5mJ / pulse. The second preset value for the energy release density is 80mJ / pulse.

4. The intraoperative risk assessment system for cataract surgery robots according to claim 3, characterized in that, The robot motion parameter monitoring unit includes: Joint torque sensor, used to collect joint motion resistance, with an accuracy ≤0.1N·m; The grating displacement sensor and inertial measurement unit are used to calculate the three-dimensional coordinates, motion speed and attitude angle of the instrument end effector.

5. The intraoperative risk assessment system for cataract surgery robots according to claim 4, characterized in that, The high-risk operation status identification logic of the multimodal joint reasoning unit is as follows: when the contact force is greater than 1.2N, the energy release density is greater than 90mJ / pulse, and the anterior chamber equivalent pressure is less than 1.5kPa, it is determined to be a high-risk operation status and an orange warning is triggered.

6. The intraoperative risk assessment system for cataract surgery robots according to claim 5, characterized in that, The historical data causal chain model is a third-order Markov chain model trained based on several clinical data cases. It is used to analyze the temporal correlation of "contact force mutation → abnormal energy release → capsule deformation". The preset delay thresholds are: force mutation to energy abnormality ≤150ms, energy abnormality to capsule deformation ≤300ms.

7. The intraoperative risk assessment system for cataract surgery robots according to claim 6, characterized in that, The dynamic weighted fusion of the risk assessment unit adopts an adaptive gating mechanism, and the expression is: F fusion =σ(FC([F 力 ,F 运动 ,F 图像 ,F 能量 ]))⊙(F 力 +F 运动 +F 图像 +F 能量 ) Where σ is the Sigmoid activation function, FC is a fully connected layer, and weights are automatically assigned through learning. When the equivalent pressure value of the current room is <2kPa, the weight of the force feature is automatically increased to more than 60%.

8. The intraoperative risk assessment system for cataract surgery robots according to claim 7, characterized in that, The risk assessment model is a hybrid CNN-LSTM neural network. The CNN branch is used to extract spatial features of capsule wrinkle degree, posterior capsule radius of curvature change rate, capsule edge roughness, and anterior chamber bubble volume from microstructure images; The LSTM branch is used to analyze the temporal fluctuations of contact force, motion velocity, and energy release density, and outputs the risk probability of posterior capsule rupture and anterior chamber collapse.

9. The intraoperative risk assessment system for cataract surgery robots according to claim 8, characterized in that, The safe operation guidelines of the operation suggestion generation unit include: Energy segmented modulation: During medium-risk conditions, the energy of a single pulse is limited to ≤60mJ; during high-risk conditions, the ultrasound power is forcibly reduced to 50% of the base power. Spiral retreat path planning: A spiral path is generated based on the RCM mechanism. The instrument end retracts with a radius of 0.1 mm and a speed of 0.2 mm / s, and the deflection angle is adjusted to ≤10°. Fluid flow-energy coordinated regulation: For every 0.5 kPa decrease in the anterior chamber equivalent pressure, the perfusion flow rate is automatically increased by 5 mL / min and the energy release density is reduced by 5 mJ / pulse until the pressure is restored to the safe range of 6-8 kPa.

10. The intraoperative risk assessment system for cataract surgery robots according to claim 9, characterized in that, The emergency braking strategy includes the following: The starting motor is locked by an electromagnetic brake and a mechanical limit structure. The system announces "Stop the operation immediately and perform a capsule integrity check" via voice module, and overlays a red warning box to mark the risk area in the surgical field.

Images