Centralized positioning system and method for femtosecond laser ophthalmic corneal refraction operation

Abstract

The invention discloses a center positioning system and method for a femtosecond laser ophthalmic corneal refraction operation, and relates to the technical field of image processing. The method comprises the following steps of S1, performing multi-source data synchronous acquisition on eyeballs of a patient before an operation; s2, fusing and preprocessing the collected multi-source data; s3, establishing an eyeball dynamics method, outputting six-degree-of-freedom position data in real time, and predicting the position of a laser action point; s4, dynamically adjusting the dual coordinate system according to the cornea; s5, constructing a real-time feedback system, and realizing closed-loop control, safety boundary control and a fault-tolerant mechanism; and S6, generating parameters of an operation process, constructing a four-dimensional cutting model, and performing path planning of the laser model. According to the method, the OCT, the Placido ring and the iris vein composite features are combined, the positioning precision is improved, Kalman filtering and neural network prediction compensation are utilized, eyeball movement is tracked in real time, and the accuracy of operation pupil positioning is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a central positioning system and method for femtosecond laser ophthalmic refractive surgery. Background Art

[0002] With the wide application of femtosecond laser technology in ophthalmic refractive surgery, accurate positioning of the corneal ablation center has become the core factor determining the surgical effect. A slight deviation (>0.2 mm) of the corneal ablation center may lead to problems such as increased postoperative higher-order aberrations, night glare, and decreased best corrected visual acuity, severely restricting the safety and predictability of the surgery.

[0003] The existing laser ophthalmic refractive surgery has the following deficiencies:

[0004] (1) Anatomical marker-dependent positioning

[0005] Traditional methods are based on the corneal vertex or pupil centroid as the positioning reference, but they have significant limitations. For example, the pupil center may shift by 0.1 - 0.5 mm due to changes in light intensity (photodynamic pupil displacement effect); the corneal vertex may change its position due to eyelid pressure when the patient is in the supine position (average shift of 0.15 mm), etc.

[0006] (2) Optical tracking technology

[0007] Mainstream devices (such as IntraLase, Femto LDV) track eye markers through infrared cameras, but their accuracy is limited by: insufficient sampling frequency (usually ≤200 Hz), making it difficult to capture high-frequency microtremors during surgery (frequency can reach 500 Hz); interference from reflections at the liquid interface (tear film) on the corneal surface, resulting in a decrease in the signal-to-noise ratio of the tracking signal.

[0008] (3) Biomechanical modeling compensation:

[0009] Some studies have attempted to predict ablation deformation through preoperative corneal elastic modulus, but changes in corneal humidity and temperature during surgery can lead to model mismatch, and the actual compensation error is as high as 0.3 - 0.5 mm. Summary of the Invention

[0010] The purpose of the present invention is to provide a central positioning system and method for femtosecond laser ophthalmic refractive surgery, which solves the problems of difficult pupil positioning, untimely surgical path tracking, and untimely feedback correction through steps such as multi-modal data synchronous acquisition, fusion and preprocessing, dynamic tracking and prediction, intelligent center calculation, real-time feedback calibration, fault tolerance mechanism, and surgical parameter generation.

[0011] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0012] The present invention is a central positioning method for femtosecond laser ophthalmic refractive surgery, including the following steps:

[0013] Step S1: Synchronously collect multi-source data of the patient's eyeball before the surgery;

[0014] Step S2: Fuse and preprocess the collected multi-source data;

[0015] Step S3: Establish an eyeball dynamics method to output 6-degree-of-freedom position data in real time and predict the position of the laser action point;

[0016] Step S4: Dynamically adjust the dual coordinate system according to the cornea;

[0017] Step S5: Construct a real-time feedback system to achieve closed-loop control, safety boundary control, and fault tolerance mechanism;

[0018] Step S6: Generate the parameters of the surgical process, construct a four-dimensional cutting model, and perform path planning for the laser model.

[0019] As a preferred technical solution, in the step S1, the synchronous acquisition of multi-source data includes three-dimensional OCT scanning, dynamic Placido ring imaging, iris-scleral vein dual biometrics, and non-contact eye movement tracking; the three-dimensional OCT scanning is used to obtain the corneal tomographic topographic map; the dynamic Placido ring imaging is used to obtain the curvature distribution of the cornea in real time; the iris-scleral vein dual biometrics adopts a multi-spectral imaging algorithm, using near-infrared light with a wavelength of 850 nm and short-wave infrared light with a wavelength of 940 nm, and a polarizer can be configured to eliminate corneal reflection interference at the same time. A CMOS sensor (frame rate 1000 fps) is used to synchronously capture the iris texture and the fractal structure of the scleral vein, extract the iris crypt structure and the scleral vein branch characteristics respectively, and segment the biometric features through a U-Net network and fuse the residual attention mechanism; at the same time, during the surgical process, the iris image is collected in real time and feature point matching is performed with the SIFT-FREAK algorithm used in the preoperative database, and the random sample consensus algorithm is used to eliminate the false matching caused by blinking; the non-contact eye movement tracking is used to establish a 6-degree-of-freedom dynamics equation of eyeball movement, introduce a viscoelastic damping coefficient (μ = 0.12 N·s / m 2 ) and the dynamic parameters of the lateral rectus muscle traction force, and realize the prediction of the eyeball movement trajectory through the extended Kalman filter.

[0020] As a preferred technical solution, the specific process of the iris-scleral vein dual biometrics in the step S1 is as follows:

[0021] Step S11, Multispectral Imaging and Hardware Configuration: Select a dual-band light source, and use an 850nm (near-infrared) and 940nm (short-wave infrared) dual-band LED array, which are respectively used to penetrate the corneal epithelium layer (850nm near-infrared acquisition) and enhance the contrast of the scleral veins (940nm short-wave infrared acquisition); In order to eliminate the interference of corneal surface reflection, a polarizer (extinction ratio ≥ 100:1) can be installed between the light source and the camera;

[0022] The specific hardware system includes: a high-speed CMOS camera (resolution 1280×1024, frame rate 1000fps); a beam splitter prism separates the dual-band light into two independent imaging channels; In order to reduce errors, each sensor is time-synchronized through an FPGA controller to achieve a dual-band exposure time difference ≤ 10μs, ensuring spatio-temporal consistency.

[0023] Step S12, Band-Segmented Feature Extraction and Deep Learning Segmentation: Input a dual-channel image into the U-Net network, and pass through the encoder and decoder in sequence to output the iris crypts and scleral veins;

[0024] The U-Net network is an improved network architecture, and the specific network architecture is as follows:

[0025] Input layer: Input a dual-channel image, 850nm iris image + 940nm scleral vein image;

[0026] Encoder: It includes 5 layers of residual convolution blocks, and each layer contains Conv3*3, BatchNorm, ReLU, and SE attention modules;

[0027] Among them, the calculation formula of the SE attention module is as follows:

[0028]

[0029] In the formula, s c represents the global feature descriptor of channel c, indicating the average response intensity of this channel in the spatial dimension; H and W respectively represent the height and width of the feature map, corresponding to the spatial resolution of the output of the convolutional layer; x c (i,j) represents the activation value at the spatial position (i,j) of the feature map in channel c; x c represents the feature map after channel attention adjustment; W1 and W2 represent the weight matrices of the fully connected layers; among them, W1 ∈ R C / r×C , used for dimensionality reduction processing, r is the compression ratio; W2 ∈ R C×C / r is used to restore the dimension; σ(·) represents the Sigmoid function, mapping the weight to the interval [0, 1], indicating the channel importance; δ(·) represents the non-linear activation function, enhancing the expression ability of the model;

[0030] Decoder: A 5-layer transposed convolutional block that fuses multi-scale features with skip connections. Output layer: Iris crypts (binary mask) and scleral veins (skeletonized vector map). The loss function for training the Net network is: Joint optimization of DiceLoss + Focal Loss:

[0031] L = 0.5 × L Dice + 0.5 × L Focal ;

[0032] where L represents the total loss value, which is used to guide the neural network to optimize the segmentation accuracy and the problem of class imbalance; L Dice represents the Dice loss term, which measures the regional overlap between the prediction and the true label; L Focal represents the Focal loss term, which solves the class imbalance problem between the foreground (such as iris / vein) and the background; 0.5 represents the weight coefficient;

[0033] Step S13, SIFT-FREAK dynamic registration algorithm: Extract feature points according to the iris texture characteristics, perform dynamic feature matching, and calculate the sub-pixel displacement using frequency domain analysis.

[0034] As a preferred technical solution, in the step S13, the specific process of the SIFT-FREAK dynamic registration algorithm is as follows:

[0035] Step S131: According to the texture characteristics of the iris, adjust the scale factor of the Gaussian pyramid (σ = 1.6 → 1.2), screen the key points, and only retain the feature points with the top 10% curvature response values;

[0036] Step S132: Enhance the FREAK descriptor, adopt a concentric hexagonal sampling network (radius increasing ratio 1.5), improve the rotational invariance, and perform binary coding compression. When compressing, reduce the 512-bit descriptor to 256 bits through PCA dimensionality reduction, and the matching speed is increased by 40%;

[0037] Step S133: Use the Hamming distance to screen the candidate matching pairs, set the threshold to 0.3, and use RANSAC robust registration;

[0038] During registration, evaluate the model and solve the similarity transformation matrix (rotation + translation + scaling):

[0039]

[0040] where represents the scaling factor, which is used for the uniform scaling ratio of the original coordinates, represents magnification, denotes reduction; θ denotes the rotation angle, in radians, representing the angle by which the coordinate system rotates counterclockwise around the origin, t x ,t y denotes the translation amount, representing the translation distances along the x-axis and y-axis respectively, used to translate the transformed coordinates to the target position, x, y denote the original coordinates, the two-dimensional coordinate values of the input point; x′, y′ denote the transformed coordinates, the target point coordinates after rotation, scaling, and translation;

[0041] Step S134: Based on the RANSAC coarse registration, use frequency domain analysis to calculate the sub-pixel displacement, and the specific formula is as follows:

[0042]

[0043] In the formula, F represents the Fourier transform, I pre , I intra respectively represent the reference image and the image to be registered; F * represents the conjugate complex number, the spectrum conjugate of I intra , used to calculate the cross-power spectrum of the two images, F -1 represents the inverse Fourier transform argmax represents peak localization, finding the position of the maximum peak in the phase correlation matrix, which is the sub-pixel displacement amount between the two images; the final output configuration error ≤ ±2um.

[0044] As a preferred technical solution, in step S2, the OCT scan data and the Placido ring data are fused to construct a corneal curvature-thickness mapping model. The corneal thickness distribution data obtained by OCT and the Placido ring curvature data are spatially registered, and then thin plate spline interpolation is used to generate a three-dimensional corneal morphological surface; and based on LSTM to identify motion artifacts, a corneal reflection point automatic compensation algorithm is used to cope with tear film interference.

[0045] As a preferred technical solution, in step S2, the fusion and preprocessing of multi-source data include spatio-temporal calibration and abnormal data filtering; the spatio-temporal calibration adopts the one-step mode of the IEEE 1588 standard, eliminates the protocol stack processing delay through the hardware timestamp, and dynamically compensates for the link asymmetry in combination with the master-slave clock architecture; the master clock sends a Sync message, directly embeds the transmission timestamp t1, the slave clock records the reception time t2, calculates the offset and adjusts the local clock to achieve sub-microsecond synchronization accuracy; based on the laser system coordinate system, the original data of each sensor is mapped to a unified coordinate system through a coordinate transformation matrix (including a rotation matrix and a translation vector) to eliminate spatial deviation;

[0046] The abnormal data performs sliding window segmentation on the multimodal data, extracts time series features (such as corneal curvature change rate, eye movement trajectory acceleration), constructs a bidirectional LSTM network, inputs the time series data, and outputs the artifact probability. If the output probability exceeds a threshold (such as 0.9), it is determined as a motion artifact and the rejection mechanism is triggered; automatically compensate for the corneal reflection point, use the dynamic Placido ring imaging technology, detect the tear film break-up area through the multi-ring reflection point distribution, use the Gaussian mixture model to distinguish normal reflection points from tear film interference points, and based on the historical reflection point spatial distribution data (such as iris vein characteristics), reconstruct the corneal curvature data of the missing area through the interpolation algorithm to ensure the integrity of the corneal morphology.

[0047] As a preferred technical solution, in step S3, the specific implementation process is as follows:

[0048] Step S31: Define the state vector and create the dynamic equation; the definition of the state vector is as follows:

[0049] X = [x, y, z, θ x , θ y , θ z , v x , v y , v z , w x , w y , w z T ;

[0050] The state vector contains the 6 degrees of freedom position, the translations in the x, y, and z axis directions, θ x , θ y , θ z represent the rotations in the x, y, and z axis directions, and the corresponding v on the velocity axis x , v y , v z , and the angular velocity w on the corresponding axis x , w y , w z ;

[0051] The dynamic equation is:

[0052] X K|K-1 = F K|K-1 + W K ;

[0053] In the formula, F K is the non-linear Jacobian matrix containing blinks and microsaccades, and W K represents the process noise;

[0054] Step S32: Set the observation method for the observation data source and update it in real time according to the sensor confidence; the observation formula is:​

[0055] z k = H k X k + V k ;

[0056] Wherein, H k is the observation matrix, and V k is Gaussian observation noise; R is updated in real time according to the sensor confidence, k and the Kalman gain K is optimized k

[0057] Step S33: Output 6-degree-of-freedom position data, and perform abnormal motion detection through residuals. The 6-degree-of-freedom position data outputs the position and rotation amount every 0.5 ms, with an accuracy reaching the sub-micron level; the abnormal motion detection is performed through the formula: y k = z k - H K X K|K-1 , and determine whether to trigger the intervention of the expert rule base;

[0058] Step S34: Construct a prediction model of the neural network;

[0059] Pre-training data: The eye movement trajectories, laser action point offsets, and postoperative corneal shape data in 100,000 surgeries.

[0060] Network structure of the prediction model:

[0061] Input layer: The Kalman filter state X K at the current moment and the historical 3 ms motion trajectory window;

[0062] Hidden layer: The bidirectional LSTM captures the temporal dependencies, and the CNN extracts the spatial features (such as the corneal curvature change pattern);

[0063] Output layer: Predict the laser action point position Δp k+3 = [Δx, Δy, Δz] T .

[0064] During online fine-tuning, the network weights are updated in real time during the operation, based on the eye movement characteristics of the current patient (by matching the sliding window data with the pre-training library);

[0065] Step S35: Update the network weights during the operation in real time, perform predictions based on the eye movement characteristics of the current patient, and achieve online fine-tuning; superimpose the predicted offset Δp k+3 onto the X output by the Kalman filter K to generate the final laser action point coordinates;

[0066] If the prediction confidence is lower than the threshold (such as the network output variance is too large), fallback to the pure Kalman filter result.

[0067] As a preferred technical solution, in step S4, the dual coordinate system includes an anatomical center and a visual center; the anatomical center is used for calculating the intersection point of the corneal apex and the optical axis; the intersection point of the corneal apex and the optical axis is based on OCT tomographic data, and the maximum curvature point C of the anterior corneal surface is located by using the gradient ascent algorithm. V And the Placido ring reflection points are used to fit the corneal curvature center C. C , and the optical axis C V and the connection line of the pupil geometric center C p are defined. The anatomical center C anat is the intersection point of the optical axis and the posterior corneal surface; the visual center is used for optimizing the visual axis by combining wavefront aberration data; the visual axis is the connection line between the wavefront aberration center and the retinal macula center.

[0068] When the cornea is dynamically adjusted, the entropy value H of the curvature is calculated, and the specific formula is as follows:

[0069]

[0070] In the formula, k i,j represents the observed value of the i-th row and j-th column in the original data matrix; Σk i,j represents the sum of all elements in the data matrix to obtain the global normalization denominator.

[0071] The weight is dynamically adjusted according to different entropy values; when H < 2.5, it indicates that the cornea is regular, and when H > 2.5, it indicates that the cornea is irregular.

[0072] As a preferred technical solution, in step S5, the closed-loop control tracks the 6-degree-of-freedom position (x, y, z, θ x , θ y , θ z ) output by the Kalman filter in real time, predicts the offset in the next 3 ms through a neural network, and performs a position check every 0.5 ms to trigger an emergency compensation mode for data exceeding the threshold; the safety boundary control realizes the anatomical safety zone and updates the rules in real time, such as: the area where the cutting depth exceeds the safety threshold (such as the remaining corneal thickness < 250 μm) is automatically frozen; the cutting edge is dynamically fitted by the moving least squares method to generate a smooth transition forbidden zone; the fault tolerance mechanism sets a three-level fault tolerance strategy, including:

[0073] Level 1 software layer: switch to the standby control model (such as pure PID mode) when the error exceeds the limit;

[0074] Level 2 hardware layer: the FPGA directly takes over the galvanometer drive and bypasses the main control system;

[0075] Level 3 mechanical layer: the physical shutter cuts off the laser path within 0.1 ms;

[0076] The OCT data packets are compared through CRC32 checksum. When an error is found, historical data interpolation is enabled. At the same time, the dual-redundancy communication channels (optical fiber + wireless) ensure zero packet loss of control signals.

[0077] As a preferred technical solution, in the step S6, the specific process of constructing a four-dimensional cutting model and performing path planning for the laser model is as follows:

[0078] Step S61: Obtain the real-time data of the wavefront image (dynamic update of Zernike coefficients) and the OCT corneal humidity monitoring data (updated every 0.5 ms);

[0079] Step S62: Construct a dynamic energy model;

[0080] The dynamic energy model includes an energy mapping formula, and the specific formula is as follows:

[0081]

[0082] In the formula, Z n (t) is the high-order difference coefficient, β is the dynamic gain coefficient, f(H 湿度 ) represents the humidity compensation function;

[0083] Step S63: Optimize the non-uniform cutting surface shape, and the objective function is:

[0084] min(||S 切削 -S 目标 ||2 + λ × curvature smoothness);

[0085] In the formula, S 目标 is the ideal corneal surface shape based on wavefront aberration, and λ is the smooth constraint weight;

[0086] Step S64: Optimize the laser path; specifically including:

[0087] Step S641, basic path generation: Generate an Archimedean spiral with the corneal vertex as the center:

[0088] r(θ) = kθ, k = 0.1 - 0.3 μm / rad;

[0089] The initial density is determined by the cutting depth, and each μm depth corresponds to 0.2 turns / mm 2 ;

[0090] Step S642, dynamic density adjustment: For the region where the curvature change rate |ΔK| > 0.5 D / mm, the density is increased by 1.5 times; for the weight W coma > 0.3, the path direction is deflected by 5° - 15° to enhance the asymmetric cutting effect.

[0091] Step S643: Calculate the basic overlap rate: v represents the scanning speed, and d spot represents the spot diameter;

[0092] Step S644: Dynamic adjustment rule: When the predicted local temperature is greater than 50 degrees Celsius, reduce the overlap rate O base by -10%; for the high-order aberration region, increase the overlap rate to O base +15%;

[0093] Step S65: Combine the Kalman filter to predict the scanning position error in the next 3 ms, and dynamically adjust the galvanometer acceleration to match the target overlap rate.

[0094] The present invention is a central positioning system for femtosecond laser ophthalmic refractive surgery, including a multi-source data synchronous acquisition module, a data fusion and processing module, a dynamic tracking and prediction module, an intelligent center calculation module, a real-time feedback calibration module, a fault tolerance module, and a surgical parameter generation module; the multi-source data synchronous acquisition module is used to collect eye data by different devices; the data fusion and processing module is used to perform spatio-temporal calibration on the collected data and filter abnormal data; the dynamic tracking and prediction module is used to construct an eyeball movement dynamics equation for tracking and predict the position of the laser action point; the intelligent center calculation module is used to solve the dual coordinate system and adaptively allocate weights; the real-time feedback calibration module is used to construct a closed-loop control mechanism and a safety boundary control; the fault tolerance module is used to set a three-level fault tolerance strategy, compare the OCT data packet through CRC32 checksum, enable historical data interpolation when an error is found, and at the same time, the dual-redundant communication channel ensures zero packet loss of the control signal; the surgical parameter generation module is used to construct a four-dimensional cutting model and perform laser path planning.

[0095] The present invention has the following beneficial effects:

[0096] (1) By combining the composite features of OCT, Placido ring, and iris vein, the present invention improves the positioning accuracy, and uses the Kalman filter and neural network prediction compensation to track the eyeball movement in real time, improving the accuracy of surgical pupil positioning;

[0097] (2) Based on the personalized corneal shape prediction model of transfer learning, the present invention uses the transfer learning model to predict the corneal shape in advance and optimize the surgical parameters;

[0098] (3) The present invention is provided with a real-time feedback calibration system, which realizes sub-micron level accuracy calibration through a closed-loop control mechanism and a safety boundary control;

[0099] (4) By setting the surgical safety margin, based on the real-time measurement of the remaining corneal thickness, the range of the cutting prohibited area is automatically adjusted, and the laser scanning path is optimized by combining the central positioning data, so as to reduce the risk of ophthalmic surgery.

[0100] Of course, it is not necessary for any product implementing the present invention to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0102] Figure 1 It is a flowchart of a central positioning method for a femtosecond laser ophthalmic corneal refractive surgery according to the present invention;

[0103] Figure 2 It is a schematic structural diagram of a central positioning system for a femtosecond laser ophthalmic corneal refractive surgery according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0104] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0105] In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0106] In order to make the purpose, technical solutions and advantages of the present application clearer, the following Figure 1-2 and embodiments are used to further describe the present application in detail. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0107] Embodiment 1

[0108] Please refer to Figure 1 As shown, the present invention is a central positioning method for a femtosecond laser ophthalmic corneal refractive surgery, including the following steps:

[0109] Step S1: Synchronously collect multi-source data of the patient's eyeball before the surgery;

[0110] Step S2: Fuse and preprocess the collected multi-source data;

[0111] Step S3: Establish an eye movement dynamics method to output 6-degree-of-freedom position data in real time and predict the laser action point position;

[0112] Step S4: Dynamically adjust the dual coordinate system according to the cornea;

[0113] Step S5: Construct a real-time feedback system to achieve closed-loop control, safety boundary control, and fault tolerance mechanism;

[0114] Step S6: Generate the parameters of the surgical process, construct a four-dimensional cutting model, and perform path planning for the laser model.

[0115] In step S1, the multi-source data synchronous acquisition includes three-dimensional OCT scanning, dynamic Placido ring imaging, iris-scleral vein dual biometric identification, and non-contact eye movement tracking; three-dimensional OCT scanning is used to obtain corneal tomographic topographic maps; dynamic Placido ring imaging is used to obtain the curvature distribution of the cornea in real time; the iris-scleral vein dual biometric identification adopts a multi-spectral imaging algorithm, using a dual-band of 850nm / 940nm wavelengths, respectively extracting the iris crypt structure and scleral vein branch features, and segmenting the biometric features through a U-Net network and fusing the residual attention mechanism; at the same time, during the surgical process, iris images are collected in real time and feature point matching is performed with the SIFT-FREAK algorithm used in the preoperative database, and the random sample consensus algorithm is used to eliminate the false matching caused by blinking; non-contact eye movement tracking is used to establish a 6-degree-of-freedom dynamics equation of eye movement, introducing a viscoelastic damping coefficient (μ = 0.12 N·s / m 2 ) and the dynamic parameters of the lateral rectus muscle pulling force, and realizing the prediction of the eye movement trajectory through the extended Kalman filter.

[0116] In step S1, the specific process of the iris-scleral vein dual biometric identification is as follows:

[0117] Step S11: Multi-spectral imaging and hardware configuration: Select a dual-band light source, adopt a dual-band LED array of 850nm (near-infrared) and 940nm (short-wave infrared), which are respectively used to penetrate the corneal epithelium layer (850nm near-infrared acquisition) and enhance the contrast of the scleral veins (940nm short-wave infrared acquisition); in order to eliminate the interference of corneal surface reflection, a polarizer (extinction ratio ≥ 100:1) can be installed between the light source and the camera;

[0118] The specific hardware system includes: a high-speed CMOS camera (resolution 1280×1024, frame rate 1000fps); a beam splitter prism that separates the dual-band light into two independent imaging channels; in order to reduce errors, each sensor is time-synchronized through an FPGA controller to achieve a dual-band exposure time difference ≤10μs, ensuring spatio-temporal consistency.

[0119] Step S12: Feature extraction in frequency bands and deep learning segmentation: Input the dual-channel image into the U-Net network, and pass through the encoder and decoder in sequence to output the iris crypts and scleral veins;

[0120] The U-Net network is an improved network architecture, and the specific network architecture is as follows:

[0121] Input layer: Input the dual-channel image 850nm iris image + 940nm scleral vein image;

[0122] Encoder: It includes 5 layers of residual convolution blocks, and each layer contains Conv3*3, BatchNorm, ReLU, and SE attention module;

[0123] Among them, the calculation formula of the SE attention module is as follows:

[0124]

[0125] In the formula, s c represents the global feature descriptor of channel c, indicating the average response intensity of this channel in the spatial dimension; H and W respectively represent the height and width of the feature map, corresponding to the spatial resolution of the output of the convolutional layer; x c (i,j) represents the activation value at the spatial position (i,j) of the feature map in channel c; x c represents the feature map after channel attention adjustment; W1 and W2 represent the weight matrices of the fully connected layers; among them, W1∈R C / r×C , used for dimensionality reduction processing, r is the compression ratio; W2∈R C×C / r is used to restore the dimension; σ(·) represents the Sigmoid function, mapping the weight to the interval [0, 1], indicating the channel importance; δ(·) represents the non-linear activation function, enhancing the expression ability of the model;

[0126] Decoder: 5 layers of transposed convolution blocks, fusing multi-scale features of skip connections (Skip Connection). Output layer: Iris crypts (binary mask) and scleral veins (skeletonized vector map). The loss function for training the Net network is: Joint optimization of DiceLoss + Focal Loss:

[0127] L = 0.5×L Dice + 0.5×L Focal ;

[0128] In the formula, L represents the total loss value, which is used to guide the neural network to optimize the segmentation accuracy and the class imbalance problem; L Dice represents the Dice loss term, which measures the regional overlap degree between the prediction and the true label; L Focal represents the Focal loss term, which solves the class imbalance problem between the foreground (such as iris / vein) and the background; 0.5 represents the weight coefficient;

[0129] Step S13, SIFT-FREAK dynamic registration algorithm: Extract feature points according to the iris texture characteristics, perform dynamic feature matching, and calculate the sub-pixel displacement using frequency domain analysis.

[0130] In step S13, the specific process of the SIFT-FREAK dynamic registration algorithm is as follows:

[0131] Step S131: According to the texture characteristics of the iris, adjust the scale factor of the Gaussian pyramid (σ = 1.6 → 1.2), screen the key points, and only retain the feature points whose curvature response values rank in the top 10%;

[0132] Step S132: Enhance the FREAK descriptor, adopt a concentric hexagonal sampling network (radius increasing ratio 1.5), improve the rotational invariance, and perform binary coding compression. When compressing, reduce the 512-bit descriptor to 256 bits through PCA dimensionality reduction, and the matching speed is increased by 40%;

[0133] Step S133: Use the Hamming distance to screen the candidate matching pairs, set the threshold to 0.3, and use RANSAC robust registration;

[0134] During registration, evaluate the model and solve the similarity transformation matrix (rotation + translation + scaling):

[0135]

[0136] In the formula, represents the scaling factor, which is used for the uniform scaling ratio of the original coordinates, represents magnification, represents reduction; θ represents the rotation angle, in radians, which represents the angle of counterclockwise rotation of the coordinate system around the origin, t x ,t y represents the translation amount, which respectively represents the translation distances along the x-axis and y-axis, and is used to translate the transformed coordinates to the target position. x, y represent the original coordinates, the two-dimensional coordinate values of the input point; x′, y′ represent the transformed coordinates, the target point coordinates after rotation, scaling and translation;

[0137] Step S134: Based on the RANSAC rough registration, calculate the sub-pixel displacement using frequency domain analysis. The specific formula is as follows:

[0138]

[0139] In the formula, F represents the Fourier transform, I pre , I intra respectively represent the reference image and the image to be registered; F * represents the conjugate complex number, which is the conjugate of the spectrum of I intra , and is used to calculate the cross-power spectrum of the two images. F -1 represents the inverse Fourier transform. argmax represents peak localization, which is to find the position of the maximum peak in the phase correlation matrix, that is, the sub-pixel displacement between the two images; the final output configuration error ≤ ±2um.

[0140] In step S2, fuse the OCT scan data and the Placido ring data to construct a corneal curvature-thickness mapping model. Perform spatial registration on the corneal thickness distribution data obtained by OCT and the Placido ring curvature data, and then use thin plate spline interpolation to generate a three-dimensional corneal morphological surface; and identify motion artifacts based on LSTM, and use the automatic corneal reflection point compensation algorithm to cope with the tear film interference.

[0141] In step S2, the fusion and preprocessing of multi-source data include spatio-temporal calibration and abnormal data filtering; the spatio-temporal calibration adopts the one-step mode of the IEEE 1588 standard, eliminates the protocol stack processing delay through the hardware timestamp, and dynamically compensates the link asymmetry in combination with the master-slave clock architecture; the master clock sends a Sync message, directly embeds the transmission timestamp t1, the slave clock records the reception time t2, calculates the offset and adjusts the local clock to achieve sub-microsecond synchronization accuracy; based on the laser system coordinate system, map the original data of each sensor to a unified coordinate system through a coordinate transformation matrix (including rotation matrix and translation vector) to eliminate spatial deviation;

[0142] For abnormal data, perform sliding window segmentation on multi-modal data, extract time series features (such as corneal curvature change rate, eye movement trajectory acceleration), construct a bidirectional LSTM network, input time series data, and output the artifact probability. If the output probability exceeds the threshold (such as 0.9), it is determined as a motion artifact and the rejection mechanism is triggered; perform automatic compensation for corneal reflection points. Use the dynamic Placido ring imaging technology to detect the tear film rupture area through the distribution of multi-ring reflection points, use the Gaussian mixture model to distinguish normal reflection points from tear film interference points, and based on the historical reflection point spatial distribution data (such as iris vein characteristics), reconstruct the corneal curvature data of the missing area through interpolation algorithm to ensure the integrity of the corneal morphology.

[0143] In step S3, the specific implementation process is as follows:

[0144] Step S31: Define and create a dynamic equation for the state vector; the definition of the state vector is as follows:

[0145] X = [x, y, z, θ x , θ y , θ z , v x , v y , v z , w x , w y , w z T ;

[0146] The state vector includes the position of 6 degrees of freedom, translations in the x, y, and z axis directions, θ x , θ y , θ z representing the rotations in the x, y, and z axis directions, and the corresponding v on the velocity axis x , v y , v z , and the corresponding angular velocity w on the axis x , w y , w z ;

[0147] The dynamic equation is:

[0148] X K|K-1 = F K|K-1 + W K ;

[0149] In the formula, F K is a non - linear Jacobian matrix including blinks and microsaccades, and W K represents the process noise;

[0150] Step S32: Set the observation method for the observation data source and update it in real - time according to the sensor confidence; the observation formula is:

[0151] z k = H k X k + V k ;

[0152] In the formula, H k is the observation matrix, and V k is the Gaussian observation noise; update R k in real - time according to the sensor confidence, and optimize the Kalman gain K k

[0153] ​Step S33: Output 6 - degree - of - freedom position data, and perform abnormal motion detection through residuals. The 6 - degree - of - freedom position data outputs the position and rotation amount every 0.5 ms, with an accuracy reaching the sub - micron level; the abnormal motion detection is through the formula: y k = z k - H K X K|K-1 , and determine whether to trigger the intervention of the expert rule base;

[0154] Step S34: Build a prediction model of the neural network;

[0155] Pre - training data: The eye movement trajectories, laser action point offsets, and postoperative corneal shape data in 100,000 surgeries.

[0156] Network structure of the prediction model:

[0157] Input layer: The Kalman filter state X K at the current moment and the historical 3 - ms motion trajectory window;

[0158] Hidden layer: The bidirectional LSTM captures temporal dependencies, and the CNN extracts spatial features (such as corneal curvature change patterns);

[0159] Output layer: Predict the laser action point position Δp k+3 = [Δx, Δy, Δz] T .

[0160] During online fine - tuning, the network weights are updated in real - time during the operation, based on the eye movement characteristics of the current patient (by matching the sliding window data with the pre - training library);

[0161] Step S35: Update the network weights during the operation in real - time, make predictions based on the eye movement characteristics of the current patient, and achieve online fine - tuning; superimpose the predicted offset Δp k+3 onto the X K output by the Kalman filter to generate the final laser action point coordinates;

[0162] If the prediction confidence is lower than the threshold (such as the network output variance is too large), fallback to the pure Kalman filter result.

[0163] In Step S4, the dual coordinate systems include the anatomical center and the visual center; the anatomical center is used for calculating the intersection point of the corneal vertex and the optical axis; the intersection point of the corneal vertex and the optical axis is based on the OCT tomographic data, and the maximum curvature point C V of the anterior corneal surface is located using the gradient ascent algorithm and the corneal curvature center C C is fitted by the Placido ring reflection points, define the line connecting the optical axis C V and the geometric center C p of the pupil, and the anatomical center C anatis the intersection point of the optical axis and the posterior surface of the cornea; the visual center is used for optimizing the visual axis by combining wavefront aberration data; the visual axis is the line connecting the wavefront aberration center and the center of the macula of the retina;

[0164] When the cornea is dynamically adjusted, the entropy value H of the curvature is calculated, and the specific formula is as follows:

[0165]

[0166] In the formula, k i,j represents the observed value of the i-th row and j-th column in the original data matrix; ∑k i,j represents the sum of all elements in the data matrix to obtain the global normalization denominator;

[0167] The weight is dynamically adjusted according to different entropy values:

[0168] When H < 2.5, it indicates that the cornea is regular; then the weight is set to a fixed value; the weight formula in this embodiment is as follows:

[0169] C final = 0.7C anat + 0.3C vis ; this weight is based on the clinically optimal ratio verified by statistics;

[0170] When H > 2.5, it indicates that the cornea is irregular, and a dynamic adjustment model needs to be set up. The specific process is as follows:

[0171] 1. Feature extraction: Let the OCT curvature distribution gradient be ΔK, the wavefront aberration RMS value be E, and the angle between the visual axis and the optical axis be α;

[0172] 2. The weight formula is:

[0173]

[0174] In the formula, a, b, and c are determined through preoperative database training and calculated in real time during the operation.

[0175] In step S5, the closed-loop control tracks the 6-degree-of-freedom position (x, y, z, θ x , θ y , θ z ) output by the Kalman filter in real time, predicts the offset in the next 3 ms through the neural network, and performs a position check every 0.5 ms to trigger the emergency compensation mode for data exceeding the threshold; the safety boundary control realizes the anatomical safety zone and updates the rules in real time, such as: the area where the cutting depth exceeds the safety threshold (such as the remaining corneal thickness < 250 μm) is automatically frozen; the cutting edge is dynamically fitted by the moving least squares method to generate a smooth transition forbidden zone; the fault tolerance mechanism sets a three-level fault tolerance strategy, including:

[0176] Level 1 software layer: Switch to the standby control model (such as pure PID mode) when the error exceeds the limit;

[0177] Level 2 hardware layer: The FPGA directly takes over the galvanometer drive and bypasses the main control system;

[0178] Level 3 mechanical layer: The physical shutter cuts off the laser path within 0.1 ms;

[0179] Compare the OCT data packet through CRC32 checksum. When an error is found, enable historical data interpolation. At the same time, the dual-redundancy communication channel (optical fiber + wireless) ensures zero packet loss of the control signal.

[0180] In step S6, the specific process of constructing the four-dimensional cutting model and performing the path planning of the laser model is as follows:

[0181] Step S61: Obtain the real-time data of the wavefront image (dynamic update of Zernike coefficients) and the OCT corneal humidity monitoring data (updated every 0.5 ms);

[0182] Step S62: Construct a dynamic energy model;

[0183] The dynamic energy model includes an energy mapping formula, and the specific formula is as follows:

[0184]

[0185] In the formula, Z n (t) is the high-order difference coefficient, β is the dynamic gain coefficient, f(H 湿度 ) represents the humidity compensation function;

[0186] Step S63: Optimize the non-accumulative cutting surface shape, and the objective function is:

[0187] min(||S 切削 -S 目标 ||2 + λ × curvature smoothness);

[0188] In the formula, S 目标 is the ideal corneal surface shape based on wavefront aberration, and λ is the smooth constraint weight;

[0189] Step S64: Optimize the laser path; specifically including:

[0190] Step S641, base path generation: Generate an Archimedean spiral with the corneal vertex as the center:

[0191] r(θ) = kθ, k = 0.1 - 0.3 μm / rad;

[0192] The initial density is determined by the cutting depth, and each μm depth corresponds to 0.2 turns / mm 2 ;

[0193] Step S642, dynamic density adjustment: For the area where the curvature change rate |ΔK| > 0.5 D / mm, the density is increased by 1.5 times; for the weight W coma > 0.3, the path direction deflects by 5° - 15° to enhance the asymmetric cutting effect;

[0194] Step S643, calculate the basic overlap rate: v represents the scanning speed, d spot represents the spot diameter;

[0195] Step S644, dynamic adjustment rule: When the local temperature prediction is greater than 50 degrees Celsius, the overlap rate O is reduced base by -10%; for the high-order aberration area, the overlap rate is increased to O base +15%;

[0196] Step S65: Combine the Kalman filter to predict the scanning position error in the next 3 ms, and dynamically adjust the galvanometer acceleration to match the target overlap rate.

[0197] Embodiment 2

[0198] Refer to Figure 2 As shown, the present invention is a central positioning system for femtosecond laser ophthalmic refractive surgery, which can be used to execute the method content of Embodiment 1 of the present invention, including: a multi-source data synchronous acquisition module, a data fusion and processing module, a dynamic tracking and prediction module, an intelligent center calculation module, a real-time feedback calibration module, a fault tolerance module, and a surgical parameter generation module; the multi-source data synchronous acquisition module is used to collect eye data by different devices; the data fusion and processing module is used to perform spatio-temporal calibration on the collected data and filter abnormal data; the dynamic tracking and prediction module is used to construct an eyeball movement dynamics equation for tracking and predict the laser action point position; the intelligent center calculation module is used to solve the dual coordinate system and perform adaptive weight allocation; the real-time feedback calibration module is used to construct a closed-loop control mechanism and safety boundary control; the fault tolerance module is used to set a three-level fault tolerance strategy, compare OCT data packets through CRC32 checksum, enable historical data interpolation when an error is found, and at the same time the dual-redundant communication channel ensures zero packet loss of control signals; the surgical parameter generation module is used to construct a four-dimensional cutting model and perform laser path planning.

[0199] It should be noted that in the above system embodiment, the included units are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of the functional units are only for the convenience of mutual distinction and do not limit the protection scope of the present invention.

[0200] In addition, those of ordinary skill in the art can understand that all or part of the steps in the methods of the above embodiments can be completed by instructing relevant hardware through a program, and the corresponding program can be stored in a computer-readable storage medium.

[0201] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A center positioning method for femtosecond laser ophthalmic corneal refractive surgery, characterized in that: The steps include: Step S1: synchronously collect multi-source data of the patient's eyeball before surgery; Step S2: fusing and preprocessing the collected multi-source data; Step S3: Establish an eyeball dynamics method, output 6-DOF position data in real time, and predict the position of the laser action point; Step S4: The dual coordinate system is dynamically adjusted according to the cornea; Step S5: construct a real-time feedback system to achieve closed-loop control, safety boundary control and fault tolerance mechanism; Step S6: Generate surgical process parameters, construct a four-dimensional cutting model and perform path planning of the laser model.

2. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In the step S1, the synchronous acquisition of multi-source data includes three-dimensional OCT scanning, dynamic Placido ring imaging, iris-scleral vein dual biometric identification and non-contact eye tracking; the three-dimensional OCT scanning is used to obtain corneal tomographic topography; the dynamic Placido ring imaging is used to obtain the curvature distribution of the cornea in real time; the iris-scleral vein dual biometric identification adopts a multispectral imaging algorithm to extract the iris crypt structure and scleral vein branch features respectively, and the biometric features are segmented through the U-Net network; during the operation, the iris image is collected in real time and the feature points are matched with the SIFT-FREAK algorithm adopted in the preoperative database, and the mismatch caused by blinking is eliminated by the random sampling consistency algorithm; the non-contact eye tracking is used to establish the 6-DOF dynamic equation of eye movement, introduce the viscoelastic damping coefficient and the dynamic parameters of the lateral rectus muscle pulling force, and realize the prediction of eye movement trajectory through the extended Kalman filter.

3. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 2, characterized in that: In step S1, the specific process of iris-sclera vein dual biometric identification is as follows: Step S11, multispectral imaging and hardware configuration: select a dual-band light source, which is used to penetrate the corneal epithelium and enhance the contrast of scleral veins respectively; Step S12, frequency band feature extraction and deep learning segmentation: input the dual-channel image to the U-Net network, and output the iris crypt and scleral vein through the encoder and decoder in sequence; Step S13, SIFT-FREAK dynamic registration algorithm: extract feature points based on iris texture characteristics, perform dynamic feature matching, and calculate sub-pixel displacement using frequency domain analysis.

4. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 3, characterized in that: In step S13, the specific process of the SIFT-FREAK dynamic registration algorithm is as follows: Step S131: adjusting the scale factor of the Gaussian pyramid according to the texture features of the iris; Step S132: enhancing the FREAK descriptor by using a concentric hexagonal sampling network and performing binary coding compression; Step S133: Use the Hamming distance to screen candidate matching pairs, and use RANSAC robust registration; Step S134: Based on the RANSAC coarse registration, sub-pixel displacement is calculated using frequency domain analysis.

5. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In step S2, the fusion and preprocessing of multi-source data include time-space calibration and abnormal data filtering; the time-space calibration adopts the one-step mode of the IEEE 1588 standard, eliminates the protocol stack processing delay through hardware timestamps, and dynamically compensates for link asymmetry in combination with the master-slave clock architecture; the master clock sends a Sync message, directly embeds the sending timestamp t1, and the slave clock records the receiving time t2, calculates the offset and adjusts the local clock to achieve sub-microsecond synchronization accuracy; Based on the laser system coordinate system, the raw data of each sensor is mapped to a unified coordinate system through a coordinate transformation matrix; The abnormal data performs sliding window segmentation on the multimodal data, extracts time series features, constructs a bidirectional LSTM network, inputs time series data, and outputs artifact probability; automatically compensates for corneal reflection points, uses dynamic Placido ring imaging technology, detects tear film rupture areas through multi-ring reflection point distribution, and uses a Gaussian mixture model to distinguish normal reflection points from tear film interference points.

6. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In step S3, the specific implementation process is as follows: Step S31: define and create a dynamic equation for the state vector; Step S32: setting the observation method for the observation data source and updating it in real time according to the sensor confidence; Step S33: output 6-DOF position data and perform abnormal motion detection through residuals; Step S34: constructing a prediction model of a neural network; Step S35: Update the intraoperative network weights in real time, make predictions based on the current patient's eye movement characteristics, and implement online fine-tuning.

7. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In step S4, the dual coordinate system includes an anatomical center and a visual center; the anatomical center is used to calculate the intersection of the corneal vertex and the optical axis; the intersection of the corneal vertex and the optical axis is based on OCT tomographic data, and a gradient ascent algorithm is used to locate the maximum point of the anterior corneal surface curvature and the Placido ring reflection point to fit the corneal curvature center, and a line connecting the optical axis and the pupil geometric center is defined, and the anatomical center is the intersection of the optical axis and the posterior surface of the cornea; The visual center is used for optimizing the visual axis in combination with the wavefront aberration data; the visual axis is the line connecting the wavefront aberration center and the macula center of the retina; When the cornea is dynamically adjusted, the entropy value of the curvature is calculated, and the weight is dynamically adjusted according to different entropy values.

8. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In step S5, the closed-loop control tracks the 6-DOF position output by the Kalman filter in real time, predicts the offset through a neural network, performs a position check every 0.5 ms, and triggers an emergency compensation mode for data exceeding the threshold; the safety boundary control realizes the anatomical safety zone, and the real-time update rule dynamically fits the cutting edge through the moving least squares method to generate a smooth transition restricted area; The fault-tolerant mechanism sets a three-level fault-tolerant strategy, checks and compares OCT data packets through CRC32, enables historical data interpolation when errors are found, and dual redundant communication channels ensure zero packet loss of control signals.

9. The center positioning method for femtosecond laser ophthalmic corneal refractive surgery according to claim 1, characterized in that: In step S6, the specific process of constructing a four-dimensional cutting model and performing path planning of the laser model is as follows: Step S61: Acquire real-time wavefront image data and OCT corneal humidity monitoring data; Step S62: constructing a dynamic energy model; Step S63: Optimizing the non-stacking cutting surface shape; Step S64: Optimizing the laser path; Step S65: Combine Kalman filtering to predict the scanning position error in the next 3 ms, and dynamically adjust the galvanometer acceleration to match the target overlap rate.

10. A center positioning system for femtosecond laser ophthalmic corneal refractive surgery, characterized in that: It includes a multi-source data synchronization acquisition module, a data fusion and processing module, a dynamic tracking and prediction module, an intelligent center calculation module, a real-time feedback calibration module, a fault-tolerant module and a surgical parameter generation module; the multi-source data synchronization acquisition module is used to collect eye data from different devices; the data fusion and processing module is used to perform spatiotemporal correction on the collected data and filter abnormal data; the dynamic tracking and prediction module is used to construct the eye movement dynamics equation for tracking and predict the position of the laser action point; the intelligent center calculation module is used to solve the dual coordinate system and adaptively allocate weights; the real-time feedback calibration module is used to build a closed-loop control mechanism and safety boundary control; the fault-tolerant module is used to set a three-level fault-tolerant strategy, and through CRC32 check and comparison of OCT data packets, historical data interpolation is enabled when errors are found, and dual redundant communication channels ensure zero packet loss of control signals; the surgical parameter generation module is used to build a four-dimensional cutting model and perform laser path planning.

Images