Eyeball tracking and positioning method based on artificial intelligence large model and automatic high-risk area (blood vessel and like) avoidance

Abstract

An AI-driven ophthalmic treatment system (20) includes a laser radiation source (48), a controller (44) integrated with an artificial intelligence processing unit, a neural network processing unit (NPU), an electronic control unit, and a storage module. The controller is configured to perform the following artificial intelligence-based operations: automatically identify and calibrate a plurality of local target regions (84) within an eyeball region (25) of a patient (22) by a deep learning model, and assign an appropriate predetermined laser energy to each target region, respectively, using a machine learning algorithm; controlling the radiation source to perform laser irradiation on at least the first target area; after the irradiation of the first target area is completed, recognizing the change of an eyeball structure through a real-time image analysis and computer vision method; and based on the identified structural change, dynamically inhibiting or adjusting predetermined energy output corresponding to the non-irradiated second target area through an AI decision model. Other implementation modes adopting artificial intelligence and machine learning technologies are also covered.

Description

[0001] This invention utilizes 3D eye modeling and tracking, along with a large AI model, to automatically and in real-time determine high-risk / safe zones and automatically avoid areas such as blood vessels.

[0002] This invention utilizes artificial intelligence technology to track real-time eye movements and automatically identify areas such as the iris, pupil, and trabecular meshwork. Invention Field

[0003] This invention relates to novel ophthalmic devices and treatment methods for treating glaucoma, ocular hypertension and other eye diseases, especially those combining artificial intelligence, machine learning and deep learning technologies to achieve intelligent diagnosis and treatment. background

[0004] Existing technology discloses an optically pumped mid-infrared solid-state laser with a high pulse repetition frequency, suitable for laser surgery. This laser operates in the wavelength range of 1.7 micrometers to 4.0 micrometers and employs an optical pumping method.

[0005] Another existing medical laser unit includes a laser body composed of at least one laser medium. A first-type pump source is used to continuously excite the laser medium and generate continuous laser radiation; a second-type pump source is used for pulsed excitation and generates pulsed laser radiation. The emitting unit guides the continuous or pulsed laser radiation to the surgical site. Specifically, this unit has two operating modes: continuous laser for cutting and pulsed laser for high-power fragmentation.

[0006] Another optical system includes a seed source and a coupled optical amplifier, which controls the pulse output by adjusting the power of the seed signal. The seed signal can present one or more pulse bursts, each burst consisting of a single pulse or multiple pulses. During the pulse interval or burst interval, the seed signal power can be adjusted to an intermediate value between a minimum and a maximum value to control the amplifier gain, thereby optimizing the behavior of subsequent pulses or bursts.

[0007] A diode-pumped frequency doubling system is also disclosed, comprising a pump diode, a crystal, and a frequency multiplier within a laser cavity, operating in pulsed or low duty cycle pumping modes. The cavity length is relatively short to stabilize transient response and generate stable and controllable energy output in discontinuous pumping modes. In a preferred embodiment, the frequency multiplier is clamped along a noncritical axis, and the control system establishes an isothermal distribution within the active module volume. The controller regulates an independent heat source or heat sink to preheat the diode or frequency doubling crystal to maintain a directional thermal gradient and adjusts the isothermal migration rate or direction according to the pulse sequence to maintain laser operation stability. This thermal management system may include a heat sink, a heater, and control elements for preheating the laser diode.

[0008] Another laser system employs a nonlinear crystal for second harmonic generation, combined with a solid-state gain medium, and supports multiple pump power modes under the control of a data processor. In low-power mode, pump power modulation is used, while in high-power mode, continuous pumping is achieved using a Q-switch. Optionally, modulation can be used in both modes, with modulation parameters controlled by a program. Second harmonic generation without a Q-switch can also be achieved in high-power mode.

[0009] A method for operating an off-cavity frequency-converting solid-state laser is also disclosed, wherein the laser resonator includes an optically pumped gain medium and is configured to compensate for thermal lensing effects within a predetermined range. A nonlinear crystal located outside the resonator converts the fundamental frequency laser into frequency-converted radiation. Laser processing is performed by a sequence of frequency-converted pulses with sufficient power. The power of the frequency-converted radiation depends on the output parameters of the laser resonator. The laser is operated in a manner that ensures the resonator outputs the same average power of fundamental frequency radiation before and after the processing operation, thus keeping the thermal lensing effect of the gain medium within a predetermined range. By adjusting the output parameters before and after the processing operation, the frequency-converted radiation generated in the earlier stage is insufficient to perform laser processing.

[0010] In laser trabeculoplasty, a laser is used to irradiate the trabecular meshwork inside the patient's eye with one or more treatment beams, thereby reducing intraocular pressure. Invention Overview

[0011] According to some embodiments of the present invention, a system is provided, including a radiation source and a controller. The controller is configured to: mark multiple target regions within a patient's eye using a deep learning model, each region being pre-assigned a specific irradiation energy; control the radiation source to irradiate at least a first target region; identify intraocular changes after irradiation using an artificial intelligence-based image processing algorithm; and, in response to the identified changes, suppress irradiation of a second target region that has not yet been irradiated using a machine learning decision model.

[0012] In some embodiments, the controller suppresses illumination of the second target region by:

[0013] Designate a new target area and control the radiation source to irradiate the new area in place of the second target area.

[0014] In some embodiments, the controller achieves suppression by irradiating the second target region with a dose lower than the pre-allocated energy.

[0015] In some embodiments, the variation includes bleeding, which is identified via a convolutional neural network.

[0016] In some embodiments, the changes include tissue swelling, detected using a deep learning image segmentation model.

[0017] In some embodiments, the change includes a color change, which is identified by an artificial intelligence color analysis algorithm.

[0018] In some embodiments, the change includes bubble formation, which is detected by a machine learning classifier.

[0019] In some embodiments, the controller suppresses illumination of the second target region in response to the distance between the second target region and other regions of the eye, the distance being calculated by an AI spatial relationship model.

[0020] In some embodiments, the controller uses a deep learning model to identify anatomical features of the second target region and suppresses irradiation accordingly.

[0021] In some embodiments, the controller uses a machine learning algorithm to calculate a predicted overlap metric between the radiation beam irradiating the second target region and the anatomical feature, and suppresses irradiation accordingly.

[0022] In some embodiments, the anatomical features are specific to the second target region, and the controller also uses an artificial intelligence model to identify the anatomical features of the first target region and thereby suppress irradiation of the second target region.

[0023] In some embodiments, when the anatomical feature types of the first and second target regions are the same, the controller suppresses illumination of the second target region through a machine learning decision tree.

[0024] In some embodiments, the controller calculates an estimated overlap metric between the first radiation beam and the first anatomical feature, and a predicted overlap metric between the second radiation beam and the first anatomical feature, using an AI model, and suppresses irradiation accordingly.

[0025] In some embodiments, the controller uses a deep learning regression model to calculate the estimated energy of the first radiation beam delivered to the first anatomical feature and the predicted energy of the second radiation beam delivered to the second anatomical feature, and suppresses irradiation accordingly.

[0026] In some embodiments, the controller uses a machine learning algorithm to calculate a risk metric associated with irradiating the second target area and suppresses irradiation accordingly.

[0027] In some embodiments, risk metrics are calculated using an artificial intelligence predictive model based on the patient’s medical records.

[0028] In some embodiments, the controller uses a deep learning network to identify anatomical features of the second target region and calculates a risk metric based on the feature type using a machine learning classifier.

[0029] According to some embodiments of the present invention, a method is also provided, comprising: using artificial intelligence to mark multiple target regions in a patient's eye, each pre-assigned a specific irradiation energy; controlling a radiation source to irradiate at least a first target region; identifying intraocular changes after irradiation through deep learning image processing; and, in response to the identified changes, suppressing irradiation of a second unirradiated target region through machine learning decision-making.

[0030] According to some embodiments of the present invention, a system is also provided, including a radiation source and a controller. The controller is configured to: acquire eye images; identify multiple edge points using a deep learning edge detection algorithm, each point being at a different angle relative to an intraocular reference point and located at the edge of a blood vessel; define multiple target regions between the reference point and the edge points through AI geometric analysis; and control the radiation source to irradiate these target regions.

[0031] In some embodiments, the reference point is located at the center of the iris and is located using a convolutional neural network.

[0032] In some embodiments, the reference point is located at the center of the limbus and is determined by a deep learning segmentation model.

[0033] In some embodiments, the reference point is located at the center of the pupil and is identified using a machine learning algorithm.

[0034] In some embodiments, the blood vessel edge point corresponding to each angle is the blood vessel edge closest to the reference point at that angle, determined by AI nearest neighbor search.

[0035] In some embodiments, the controller defines the target area in the following ways:

[0036] At least one treatment path is defined between the edge point and the reference point using a machine learning path planning algorithm, and the target area is located on this path.

[0037] In some embodiments, the shortest distance between the treatment path and any edge point is not less than 0.001 mm, which is calculated by an AI spatial relationship model.

[0038] In some embodiments, the controller defines the treatment pathway in the following ways:

[0039] Artificial intelligence curve fitting is used to define at least one curve that passes through the edge point, and this curve is shifted towards the reference point. The treatment path is defined by a deep learning model based on the shifted curve.

[0040] In some embodiments, the treatment path is defined as a predetermined shape perimeter tangent to the offset curve, the shape being determined through machine learning optimization.

[0041] In some embodiments, the predetermined shape is an ellipse, the parameters of which are calculated by a deep learning regression model.

[0042] In some embodiments, the treatment path is defined as the perimeter of a predetermined shape with a maximum area tangent to the offset curve, which is solved by an AI optimization algorithm.

[0043] In some embodiments, the treatment path is defined as a predetermined shape perimeter centered on a reference point and inscribed within the offset curve, with the center located by a convolutional neural network.

[0044] In some embodiments, the treatment path is defined as a closed curve that is tangent to the offset curve and whose shape matches the limbus, the matching being achieved through deep learning shape analysis.

[0045] According to some embodiments of the present invention, a method is also provided, comprising: acquiring an eye image; identifying multiple edge points using artificial intelligence edge detection, each point being at a different angle relative to an intraocular reference point and located at the edge of a blood vessel; defining a target region between the reference point and the edge points through machine learning geometric analysis; and controlling a radiation source to irradiate these regions.

[0046] According to some embodiments of the present invention, a system is also provided, including a radiation source and a controller. The controller is configured to: acquire an eye image; identify multiple edge points using a deep learning algorithm, each point being at a different angle relative to an intraocular reference point and located at the edge of a blood vessel; define a curve passing through the edge points using AI curve fitting; offset the curve towards the reference point; display the offset curve to a user and receive the user's calibration of the intraocular target region; and control the radiation source to irradiate the target region.

[0047] According to some embodiments of the present invention, a method is also provided, comprising: acquiring an eye image; identifying multiple edge points using artificial intelligence, each point being at a different angle relative to an intraocular reference point and located at the edge of a blood vessel; defining a curve passing through the edge points by machine learning curve fitting; offsetting the curve toward the reference point; displaying the offset curve to a user and receiving the user's calibration of an intraocular target region; and controlling a radiation source to irradiate the target region.

[0048] According to some embodiments of the present invention, a method is also provided, comprising: administering an α2 agonist to a patient's eye and treating the eye with laser radiation within 40 minutes after administration.

[0049] In some embodiments, the treatment includes irradiating the trabecular mesh with laser radiation.

[0050] In some embodiments, treatment is performed within 30 minutes after administration of the α2 agonist.

[0051] In some embodiments, treatment is responded to an instruction from an AI controller that indicates, via a deep learning vasoconstriction analysis model, that ocular blood vessels have sufficiently constricted due to an α2 agonist.

[0052] According to some embodiments of the invention, an α2 agonist is also provided for use in a method of inducing ocular vasoconstriction, wherein the agonist is administered to the patient less than 40 minutes prior to laser treatment, the timing being determined by a machine learning optimization model.

[0053] According to some embodiments of the present invention, a system is provided, comprising a camera and a controller. The camera is configured to acquire images of the eye before laser radiation therapy is applied to the patient's eye. The controller is configured to calculate a contraction metric characterizing the degree of contraction of ocular blood vessels under the influence of an α2 agonist using a deep learning-based image processing algorithm. The controller is further configured to output an indication that the blood vessels have sufficiently contracted in response to the contraction metric exceeding a predetermined threshold, via an AI decision model.

[0054] According to some embodiments of the present invention, a system is provided comprising a laser and a controller. The laser includes a pump source and a laser medium. The controller is configured to mark multiple target regions on a patient's eye using a machine learning algorithm for sequential irradiation by the laser. The controller is further configured to initiate irradiation of the target regions by driving the pump source to initiate a pump sequence comprising a series of laser excitation pulses, each pulse configured to cause the laser medium to produce laser output. The controller is further configured to, after initiation of irradiation, utilize an AI control strategy to cause the pump source to replace one of the laser excitation pulses in the original sequence with one or more heating pulses, the heating pulses being configured to heat the laser medium without inducing laser emission.

[0055] In some embodiments, the controller is further configured to use a machine learning preheating strategy to heat the laser medium from the pump source without inducing laser emission before initiating irradiation of the target area.

[0056] In some embodiments, the total energy of the heating pulses is between 70% and 100% of the energy of a single laser excitation pulse, which is determined by a deep learning optimization model.

[0057] In some embodiments, the one or more heating pulses include N>1 heating pulses, the number of which is calculated by an AI sequence optimization algorithm.

[0058] In some embodiments, the duration of each heating pulse is D0 / N, where D0 is the duration of a single laser excitation pulse, and this allocation is achieved through a machine learning time allocation model.

[0059] In some embodiments, the peak power of each heating pulse is equal to the peak power of each laser excitation pulse, and this power matching is achieved through an AI power control algorithm.

[0060] In some embodiments, the pump sequence is a periodic sequence with a period of T, and the controller is configured to replace the original laser excitation pulse with a heating pulse at time point {k*T / N} (where k = 0...N-1), the timing being determined by deep learning timing optimization.

[0061] In some embodiments, the controller is configured to replace the pump source with a heating pulse in response to an error indication signal via an AI fault handling model.

[0062] In some embodiments, the controller is further configured to process one or more eye images acquired by the camera and, in response to the image processing, replace the pump source with heating pulses via machine learning image analysis.

[0063] In some embodiments, the controller is configured to identify obstructions in the eye using a deep learning-based object detection algorithm and, in response to the identification of an obstruction, replace the pump source with a heating pulse.

[0064] In some embodiments, the controller is configured to process the image via a convolutional neural network to identify changes in the eyes and, in response to identifying a change, to replace the pump source with a heating pulse.

[0065] In some embodiments, the variation includes the formation of one or more bubbles, identified by a machine learning bubble detection model.

[0066] In some embodiments, the controller is configured to process the position of the eye reference point in the image using a deep learning algorithm, and calculate the position of one of the target regions based on the reference point position using an AI geometric model; the controller is further configured to determine that the laser is not aligned with the target region position using machine learning alignment detection, and in response to this determination, replace the pump source with a heating pulse.

[0067] In some embodiments, the system further includes one or more motors; the controller is also configured to use the motors to align the laser and to determine the location of the laser misaligned target area based on a corresponding signal from the motor encoder using an AI alignment algorithm.

[0068] In some embodiments, the laser is a therapeutic laser; the system further includes an alignment laser; the controller is further configured to: cause the alignment laser to emit an alignment beam at the position aligned with the therapeutic laser, and identify the position of the alignment beam through deep learning image processing; the controller is configured to determine, based on the offset between the position of the alignment beam and the position of the target area, that the therapeutic laser is not aligned with the target area position through machine learning offset detection.

[0069] In some embodiments, the wavelength of the alignment beam is greater than 700 nm.

[0070] In some embodiments, the image includes a first image and a second image: the first image is acquired when the alignment beam is emitted and includes the alignment beam; the second image is acquired before or after the alignment beam is emitted and does not include the alignment beam; the controller is configured to identify the position of the alignment beam in the first image using a deep learning image recognition algorithm and to identify the position of a reference point in the second image using a machine learning model.

[0071] In some embodiments, the image is a single image comprising a first frame and a second frame: the alignment beam appears in the first frame and does not appear in the second frame; the controller is configured to identify the position of the alignment beam in the first frame based on artificial intelligence visual analysis technology, and to identify the position of the reference point in the second frame through a deep learning network.

[0072] According to some embodiments of the present invention, a method is provided, comprising: marking multiple target regions on a patient's eye using a machine learning algorithm for sequential irradiation by a laser; driving a pump source to initiate a pump sequence to begin irradiation via an AI control system, the sequence comprising a series of laser excitation pulses, each pulse configured to cause a laser medium to generate laser output; and, after initiation of irradiation, causing the pump source to replace one of the laser excitation pulses in the sequence with one or more heating pulses based on an artificial intelligence decision model, the heating pulses being configured to heat the laser medium without inducing laser emission. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of a system for performing trabeculoplasty according to some embodiments of the present invention; Figure 2 This is a schematic diagram of a trabeculectomy apparatus according to some embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the technique of marking a target area in the eye according to some embodiments of the present invention; Figure 4 This is a flowchart of a deep learning-based target region labeling algorithm according to some embodiments of the present invention; Figure 5 A flowchart of an AI-driven trabeculoplasty surgical execution algorithm according to some embodiments of the present invention; Figure 6 This is a flowchart of AI image processing steps according to some embodiments of the present invention; Figure 7 This is a flowchart of intelligent inspection steps according to some embodiments of the present invention; Figure 8The illustration shows an example execution process of the AI ​​inspection step and the target region shifting step according to some embodiments of the present invention; Figure 9 This is a schematic diagram of the timing of laser excitation pulses and heating pulses according to some embodiments of the present invention. Detailed Implementation

[0074] Overview

[0075] To reduce the risk of bleeding and adverse side effects during laser trabeculoplasty of the eye, it is necessary to avoid irradiating blood vessels and other sensitive anatomical structures.

[0076] To address this, embodiments of the present invention provide an artificial intelligence-based treatment path calibration technology that avoids blood vessels in the eye. After determining the treatment path, multiple target areas are calibrated along the path using a machine learning algorithm and then irradiated.

[0077] To define the treatment path, the controller first uses a deep learning vision algorithm to identify multiple points along the edge of the blood vessels on the inner side of the corneal limbus, and then generates a curve passing through these points through AI path planning. Subsequently, the controller shifts this curve inward toward the center of the eye and embeds the treatment path within the shifted curve.

[0078] However, in some cases, such as due to abnormal vascular distribution, the treatment pathway may not be able to be mapped as described above. Furthermore, irradiating sensitive areas outside of blood vessels (such as hyperplastic tissue) may also lead to adverse reactions.

[0079] While AI algorithms could be used to remove segments of the path that pass through blood vessels or other sensitive areas, the inventors recognized that it is often impossible to accurately determine the eye's sensitivity to radiation beforehand; for example, some patients do not experience bleeding even when the laser beam is directly applied to blood vessels. Therefore, for some patients, completely avoiding all sensitive areas may unnecessarily reduce the effectiveness of treatment.

[0080] To address this issue, embodiments of the present invention allow the treatment path to traverse sensitive areas, while continuously monitoring the eye's condition using deep learning-based image processing technology during treatment. If the AI ​​model observes any abnormal changes (such as hemorrhage), the controller will assess the risk of causing similar changes for each target area to be irradiated using a machine learning algorithm. If the risk is high, the AI ​​decision-making system can deviate from or skip the target area. Thus, the treatment path can be selectively adjusted, avoiding unnecessary loss of therapeutic efficacy.

[0081] For example, in response to observed changes in the irradiated area, the controller can use a machine learning model to calculate a risk metric that depends on: the type of sensitive anatomical structures in the irradiated area (if present), the estimated radiation energy received by those structures, the type of sensitive anatomical structures in the area to be irradiated (if present), and the estimated radiation energy they will receive. If the risk metric exceeds a predetermined threshold, the AI ​​control system can deviate from or skip subsequent target areas.

[0082] Optionally or additionally, to reduce the risk of bleeding (e.g., to completely avoid bleeding), the eye may be pretreated with a vasoconstrictive α2 agonist (such as aclomid or brimonidine) before irradiation (e.g., within 40 minutes pre-treatment). After administration, the controller processes the eye image using AI image analysis technology to monitor the degree of vasoconstriction and outputs an indication to begin irradiation once sufficient vasoconstriction is confirmed.

[0083] In addition to monitoring changes in the eye, the controller continuously uses deep learning image recognition technology to check for obstructions in the treatment path. If the AI ​​detects an obstruction, it can deviate from or skip one or more target areas.

[0084] Typically, before each irradiation, the controller verifies that the laser is aligned with the target location using an AI alignment check system (this check is independent of the aforementioned change or obstruction checks). If misalignment or the location cannot be calculated, the AI ​​system aborts the irradiation; it can be retried after acquiring and processing a new image, or the target area can be skipped entirely.

[0085] To verify alignment, the controller can use machine learning algorithms to process signals from the laser guide motor encoder and / or a deep learning vision system to process eye images to locate the alignment beam illuminating the ocular surface. This alignment beam is invisible to the patient (but visible to the camera) to avoid disturbing the patient; its wavelength can also be configured or filters can be used to avoid interfering with the AI ​​eye-tracking system. For example, the camera can be equipped with a filter matrix so that the image includes a first frame with the beam and a second frame without the beam; or images with / without the beam can be acquired separately. The controller can locate the beam based on an AI positioning algorithm in the first frame / image, and track eye movements using a deep learning network based on the second frame / image.

[0086] In some embodiments, if irradiation needs to be stopped due to obstructions, bleeding, alignment failure, or other reasons, the AI ​​control system can drive the pump source to heat the pulsed pump laser medium. Typically, the total energy of the heating pulses is approximately equal to the original irradiation energy, but it is distributed over a longer period to avoid laser emission. This advantageously maintains the laser's thermal equilibrium until the next irradiation. Optionally or additionally, the laser medium can be preheated by an AI temperature control system before treatment, allowing the system to reach thermal equilibrium before emitting the first treatment beam. Conversely, if the laser medium is not heated for an extended period, the energy, spot profile, or alignment stability of the subsequent first treatment beam may be lower than expected.

[0087] System Description

[0088] First refer to Figure 1 This is a schematic diagram of a system 20 for performing trabeculoplasty according to some embodiments of the present invention, including a trabeculoplasty apparatus 21. See also... Figure 2 This is a schematic diagram of the device 21.

[0089] The trabeculoplasty device 21 includes an optical unit 30 and an AI controller 44. The optical unit 30 includes a radiation source 48 configured to irradiate the patient's eye 25 (e.g., trabecular mesh) with one or more therapeutic beams 52.

[0090] Typically, the radiation source 48 includes a therapeutic laser 43, such as Ekspla. TM NL204-0.5K-SH laser. The therapeutic laser 43 includes a laser medium 45, such as a semiconductor, glass, or crystal (e.g., Nd:YAG crystal). The laser medium 45 can be pumped by any suitable pump source 47 (e.g., a laser diode), and the pumping method can be optical, electrical, or other types. The laser 43 also includes multiple mirrors 49, forming a stable or unstable resonant cavity; each mirror 49 can be an independent element or a coating layer. The laser 43 also includes a heat storage device 57 for dissipating heat from the laser medium 45.

[0091] In some embodiments, the laser also includes a Q-switch 63. Optionally or additionally, a second harmonic generation crystal 51 may be included to convert the wavelength emitted by the laser medium 45 into a wavelength suitable for treatment; the SHG crystal 51 may be placed inside or outside the cavity. The laser may also be equipped with external components such as attenuators, power meters, and / or mechanical shutters.

[0092] Typically, the radiation source also includes an AI driver 53, through which the AI ​​controller 44 drives the pump source 47 to pump the laser medium 45. Specifically, the AI ​​driver 53 receives the intelligent control signal 55 from the AI ​​controller 44 via a cable 59, and accordingly outputs an electrical drive signal to the pump source 47 to trigger the pumping action. In other embodiments, the AI ​​controller may directly output a drive signal to the pump source 47.

[0093] The optical unit 30 further includes one or more beam guiding elements, such as one or more galvanometers 50, collectively referred to as "scanning galvanometers". Before emitting each treatment beam 52, the AI ​​controller 44 controls the beam guiding elements to align the treatment laser 43 to a target area on the eyeball 25 via an AI orientation algorithm, thereby guiding the beam towards that area. (Given that each treatment beam forms a non-infinitely small spot on the eyeball, this application describes it as an "area" incident on the eyeball, rather than a single "point".) For example, the beam may be deflected by the galvanometer 50 to the beam combiner 56, and then deflected again by the beam combiner before finally incident on the target area. Thus, the beam propagates along path 92, which extends from the downstream optical component (such as the beam combiner 56) of the optical unit 30 to the eyeball 25.

[0094] Typically, the AI ​​controller achieves orientation (thus completing laser alignment) by sending an AI control signal 39 (e.g., via cable 23) to the corresponding calibration motor 61 of the beam guide element. Simultaneously, the AI ​​controller (e.g., via cable 23 or other cables) receives a feedback signal 37 from the encoder 67 of the calibration motor 61, which is processed by a machine learning algorithm to indicate the real-time orientation of the beam guide element.

[0095] In some embodiments, the treatment beam comprises visible light. Optionally or additionally, the treatment beam may comprise invisible electromagnetic radiation, such as microwave radiation, infrared radiation, X-ray radiation, gamma radiation, or ultraviolet radiation. Typically, the wavelength of the treatment beam is between 200 nm and 11000 nm, for example, 500 nm–850 nm (including 520 nm–540 nm, such as 532 nm). The spatial profile of each treatment beam 52 on the eyeball may be elliptical (e.g., circular), square, or other suitable shape.

[0096] In some embodiments, radiation source 48 further includes an alignment laser for emitting a visible or invisible (e.g., infrared) alignment beam. This alignment laser is fully coaxial with the treatment laser 43, ensuring that, under any beam guiding element orientation, the alignment beam and the treatment beam are guided to the same location. Therefore, as described below... Figure 6 The alignment laser can be used in an AI calibration system to assist in the treatment of lasers.

[0097] In addition to lasers, radiation sources may also include any other suitable therapeutic beam or aligning beam emitter.

[0098] The optical unit 30 also includes a camera 54, which the AI ​​controller 44 uses to acquire images of the eye and process them using deep learning algorithms. For example... Figure 2As shown, camera 54 is generally aligned with path 92; for example, the angle between path 92 and the hypothetical line extending from eyeball 25 to camera may be less than 15 degrees. In some embodiments, the camera is located behind beam combiner 56 and receives light through the beam combiner; in other embodiments, the camera is offset relative to the beam combiner.

[0099] Before the surgery, camera 54 acquires at least one image of the eyeball 25. Based on this image, AI controller 44 can define the target area to be irradiated using a deep learning model, as described below. Figure 3 – Figure 4 The aforementioned; and / or identification of blood vessels or other anatomical features of the eyeball through machine learning algorithms, such as Figure 4 – Figure 5 As stated above.

[0100] During the surgery, camera 54 continuously acquires multiple images of the eye at a high frequency. AI controller 44 processes these images using real-time AI image processing technology and intelligently adjusts radiation source 48 and beam guiding elements accordingly to ensure that the target area is irradiated while avoiding obstructions and sensitive anatomical structures, as detailed below. Figure 6 and Figure 7 .

[0101] Camera 54 may include any suitable imaging sensor, such as a charge-coupled device (CCD), complementary metal-oxide-semiconductor (CMOS), optical coherence tomography (OCT) sensor, and / or hyperspectral image sensor. Based on these sensors, the camera can acquire two-dimensional or three-dimensional images, including monochrome images, color images (such as those based on three-color frames), multispectral images, hyperspectral images, OCT images, or images generated by fusing multiple types of images, for use in AI algorithm analysis.

[0102] In some embodiments, the optical unit 30 further includes a light source 66 that is substantially aligned with the path 92. For example, the angle between the hypothetical line connecting the end of the path 92 and the light source 66 and the path 92 may be less than 20 degrees (e.g., less than 10 degrees). The light source 66 acts as a fixed target 64 by emitting visible fixed light 68, which helps the AI ​​eye-tracking system stabilize the eye position.

[0103] Specifically, before the surgery, the patient 22 is instructed to fix their eyeball 25 in the direction of the light source 66. During the surgery, the eyeball remains fixed with the help of the fixing light 68 emitted by the light source 66, ensuring that the line of sight is basically aligned with the path 92, thereby maintaining eye stability through an AI positioning algorithm. In this state, the radiation source irradiates the eyeball with a treatment beam 52.

[0104] In some embodiments, the light source 66 includes a light emitter (such as a light-emitting diode, LED); in other embodiments, the light source includes a reflector for reflecting the light beam emitted by the light emitter.

[0105] Typically, the wavelength of the stationary light 68 (which may be higher or lower than the wavelength of the treatment beam) is between 350 nm and 850 nm. For example, the stationary light may be orange or red (600 nm–750 nm), while the treatment beam is green (527 nm–537 nm).

[0106] An optical unit typically includes an optical bench on which some or all of the aforementioned components (such as a radiation source, galvanometer, and beam combiner) are integrated. The unit also has a front surface 33 through which the therapeutic beam and the fixation beam are emitted. For example, the optical unit 30 may include a housing 31 that partially surrounds the optical bench and forms the front surface 33. (The housing 31 may be made of plastic, metal, or other suitable materials.) Optionally, the front surface 33 may be fixed to the optical bench or be an integral part of it.

[0107] In some embodiments, the front surface 33 has an opening 58 through which the therapeutic beam and the fixation beam exit; in other embodiments, the front surface includes an exit window instead of the opening 58 through which the beam exits. The exit window may be made of plastic, glass, or other suitable materials.

[0108] Typically, the optical unit 30 is also equipped with one or more illumination sources 60 (such as an LED ring composed of white LEDs or infrared LEDs). The AI ​​controller 44 enables the illumination sources 60 to intermittently flash light towards the eye. This flashing aids in camera imaging and assists in pupil constriction through AI light control algorithms, without damaging the eye or causing discomfort. (For simplified illustration, ...) Figure 2 The electrical connection between the AI ​​controller 44 and the lighting source 60 is not explicitly shown. In some embodiments, the lighting source 60 is mounted on the front 33, such as... Figure 2 As shown.

[0109] To facilitate the positioning of the optical unit by the AI ​​positioning system, the unit may be equipped with multiple beam emitters 62 (such as laser diodes) for projecting triangulation beams onto the eyeball. In some embodiments, the beam emitters 62 are mounted on the front 33; in other embodiments, they are directly fixed to the optical bench.

[0110] The optical unit 30 is mounted on the XYZ motion stage unit 32 and is controlled by an AI control mechanism 36 (such as a smart joystick). Users (such as ophthalmologists) can adjust the position of the optical unit (e.g., its distance from the eyeball) before treatment via the AI ​​control mechanism 36. In some embodiments, the XYZ motion stage unit 32 is equipped with an AI locking element, which prevents movement after positioning using a smart algorithm.

[0111] In some embodiments, the XYZ mobile station unit 32 includes one or more motors 34, and the control mechanism 36 is connected to the controller 44 via an interface circuit 46. When the user operates the control mechanism, the interface circuit converts the action into an electrical signal and transmits it to the controller 44. The controller uses an AI-based decision-making algorithm to regulate the motors in the XYZ mobile station unit.

[0112] In other embodiments, the XYZ motion stage unit 32 can be operated by a manual control mechanism, and such a unit may be equipped with a gear set instead of the motor 34.

[0113] System 20 also includes a head frame 24, which has a forehead rest 26 and a chin rest 28. During trabeculoplasty, the patient 22 places their forehead against the forehead rest 26 and their chin on the chin rest 28. In some embodiments, the head frame 24 is also equipped with a fixation strap 27 for securing the head from behind to keep it in close contact with the head frame.

[0114] like Figure 1 As shown, in some embodiments, both the headframe 24 and the XYZ moving stage unit 32 are mounted on a surface 38 (such as a tray or tabletop). In other embodiments, the XYZ moving stage unit is mounted on the surface 38, and the headframe is attached to the XYZ moving stage unit.

[0115] Usually, such as Figure 1 As shown, when the eye is illuminated, the optical unit is angled upwards towards the eye, while the eye looks downwards at the optical unit, forming an inclined path 92. For example, the angle θ between the path and the horizontal line can be between 5 and 20 degrees. This orientation helps reduce the obstruction of the eyeball by the patient's upper eyelid and related anatomical structures.

[0116] In some embodiments, the tilting of path 92 is achieved by mounting the optical unit to the wedge 40 (fixed to the XYZ moving stage unit). That is, the optical unit is mounted to the XYZ moving stage unit via the wedge 40. Figure 2 (Wedge 40 is not shown.)

[0117] In addition to adjusting the angle of the optical unit, the patient's head can also be tilted back to reduce obstruction of the eyes.

[0118] System 20 is also equipped with a monitor 42 for displaying eye images acquired by the camera. The monitor 42 may be attached to the optical unit 30 or placed in other suitable locations. In some embodiments, the monitor 42 is a touchscreen through which the user inputs commands; the system may also be equipped with other input devices such as a keyboard and mouse.

[0119] In some embodiments, the monitor 42 is directly connected to the controller 44 via a wired or wireless interface; in other embodiments, it is connected to the controller 44 via an external processor (such as a standard desktop computer processor).

[0120] In some embodiments, the controller 44 is located inside the XYZ motion stage unit 32. Figure 2 In other embodiments, it is located externally. The controller may also work in conjunction with an external processor to perform some functions.

[0121] In some embodiments, before irradiating the eyeball 25, an α2 agonist 29 is administered to the eyeball. Figure 1(e.g., azithromycin or brimonidine.) It is usually administered via eye drops at 35°C.

[0122] The α2 agonist 29 causes vasoconstriction in the eye. After sufficient vasoconstriction, the eye is irradiated with a therapeutic beam 52. Irradiation is usually performed within 40 minutes after administration (e.g., within 30 minutes, 20 minutes, 10 minutes, or 5 minutes).

[0123] In some embodiments, the α2 agonist is administered at least twice: the first administration is given 45–60 minutes before the scheduled irradiation to relieve the peak intraocular pressure; the second administration is given within 40 minutes before the scheduled irradiation (e.g., within 30, 20, 10, or 5 minutes). If insufficient contraction occurs within the scheduled time, the drug can be administered again, and treatment can begin within 40 minutes.

[0124] In some embodiments, the user judges the degree of vasoconstriction themselves; in other embodiments, the controller 44 uses a deep learning model to assist in assessing the degree of vasoconstriction based on the images captured by the camera 54. The controller calculates vasoconstriction indices using image processing algorithms (e.g., using a convolutional neural network (CNN) to analyze green / blue frame images). When the index output by the model exceeds a preset threshold, the system automatically prompts that irradiation can proceed (e.g., displaying confirmation information on the monitor 42); if the threshold is not reached, it prompts for re-administration.

[0125] Shrinkage index can be based on limbal 86 ( Figure 3 The percentage of pixels with grayscale values ​​greater than a threshold (e.g., 230) within a predetermined distance; alternatively, it can be based on statistical features such as the number, density, average width, and maximum width of blood vessels detected by a machine learning model. See below for details on blood vessel detection technology. Figure 4 .

[0126] In some embodiments, the functionality of controller 44 is implemented in hardware (such as an application-specific integrated circuit (ASIC) or a field-programmable gate array (FPGA); it can also be implemented by executing software / firmware code that integrates machine learning components, for example, a programmable processor configured with a central processing unit (CPU) and / or a graphics processing unit (GPU). The program code and data can be downloaded to the controller via a network or stored in a non-transitory tangible medium (such as magnetic, optical, or electronic memory). Such code and data enable the controller to become a dedicated machine for performing the AI ​​tasks.

[0127] In some embodiments, the controller employs a modular system (SOM), such as Variste. TM DART-MX 8M to support high-performance AI inference. **Target Area Definition**

[0128] For reference Figure 3It illustrates a technical schematic diagram of defining a target region 84 on an eyeball 25 using AI technology according to some embodiments.

[0129] Typically, the camera 54 ( Figure 2 Prior to treatment, at least one image 70 of the eyeball 25 is acquired. Based on this image, the controller 44 uses a machine learning model to automatically define a target region 84, ensuring it avoids visible blood vessels 72 in the image. (However, the target region may cover small or deep blood vessels that are not visible in the image.)

[0130] To define the target region, the controller first identifies multiple edge points 76 in image 70 using a deep learning segmentation model. Each point is located at the edge of a corresponding blood vessel 72. These edge points 76 are at different angles relative to a reference point 74 on the eyeball. Distribution. Typically, the edge points at each angle are the vessel edges closest to the reference point 74 at that angle. (At least 50 edge points 76 should be identified; for simplified illustration, ...) Figure 3 Only three points are displayed.

[0131] After identifying edge points, the controller defines a target region 84 between the reference point and the edge points based on a machine learning algorithm. (Typically, at least 50 target regions are identified; for simplicity, ...) Figure 3 (Only one example is shown.) The controller can first define at least one treatment path 82 between the edge point and the reference point, and then define target regions such that each target region is located on the treatment path (e.g., the center of the target region is located on the treatment path). Consecutive target regions can be spaced at any suitable angle, such as 2° to 4°. For example, with a 360° treatment path, the controller can define 90 to 180 target regions. (Note: Depending on the size of the target region and the spacing angle, adjacent target regions may overlap.)

[0132] In some embodiments, the controller defines a corresponding edge point and target region for each angle in a predefined set of angles. For angles where the blood vessel edge cannot be identified, the controller defines a synthetic edge point, which is not the actual blood vessel edge, but is located at a predefined distance from the reference point.

[0133] Treatment paths are typically defined such that the shortest distance from any edge point to the treatment path is not less than 0.1 mm (e.g., 0.1 mm to 1 mm) to ensure sufficient spacing between the target area and the blood vessel.

[0134] To define the treatment path, the controller first defines a curve 78 passing through edge point 76 using AI-driven spline interpolation and other curve fitting methods. Then, the controller offsets curve 78 towards the reference point by a certain distance (e.g., 0.001mm to 1mm), generating an offset curve 80. The treatment path is ultimately defined based on this offset curve.

[0135] At least a portion of the treatment path may coincide with the offset curve. Alternatively, the treatment path may be defined as a predetermined shape (such as an ellipse or circle) tangent to the offset curve, with its center arbitrarily set. For example, it may be defined as the perimeter of a predetermined shape with the largest area tangent to the offset curve, or the perimeter of a shape with the largest tangent centered at reference point 74. Another alternative is to define the treatment path as a closed curve having the shape of the limbus 86 and tangent to the offset curve. The treatment path can also be defined by applying AI smoothing to the offset curve 80 or by further offsetting it towards the reference point.

[0136] like Figure 3 As shown, if a patient's eyelid obstructs a blood vessel within a specific angle range, the controller typically uses computer vision algorithms to define multiple curves 78 (and corresponding offset curves 80) that traverse different exposure angle ranges. Based on these offset curves, the controller can define a closed treatment path as described above, but will use an AI model to suppress the definition of the target area within the obstruction angle range (and adjacent safety margin range). For example, in Figure 3 In the middle, the controller will suppress the 83-degree occlusion of the eyelid. to Define the target area within the angular range.

[0137] If the density of edge points within a certain exposure angle range is lower than the threshold required by the AI ​​curve fitting algorithm (e.g., insufficient number of recognitions due to vasoconstriction), the controller can define supplementary points on the limbus 86 within that range to reach the threshold density, and then make the curve 78 pass through both the edge points and the supplementary points simultaneously.

[0138] After defining the target area, the controller typically overlays a target area marker (optionally, a treatment path marker) onto the image 70. The user can then adjust the target area accordingly and confirm the final plan with the controller.

[0139] The controller can also perform target area illumination simulation: by directing the aiming motor 61 ( Figure 2 ) sends control signals 39 based on AI trajectory planning to sequentially align the treatment laser with each target area; via encoder 67 ( Figure 2 The alignment accuracy can be verified by providing feedback signals; or by superimposing laser alignment marks onto a real-time eye-tracking image sequence for user verification. For embodiments containing an alignment laser, the controller can sequentially emit alignment beams in each target area, verify the spot position using a deep learning model, or directly display a real-time image sequence containing visible spots for user confirmation.

[0140] After the user confirms the target area and / or completes AI-assisted alignment verification, the controller initiates the treatment laser irradiation of the target area.

[0141] Figure 4 The flowchart shows the algorithm 88 for defining the target region 84.

[0142] Algorithm 88 begins with blood vessel recognition step 90: The controller identifies blood vessels in image 70 using deep learning-based image segmentation, edge detection, feature enhancement, and pattern recognition techniques.

[0143] In the reference point definition step 91, the controller uses a machine learning model to define reference point 74: it can be automatically located by AI recognition of the center of the iris 85 / pupil 87 or the center of the limbus 86, or it can be specified by the user through an interactive interface. The controller can simultaneously calculate and store the offset of this point relative to other anatomical landmarks.

[0144] The controller then iteratively processes multiple angles relative to the reference point. After selecting an angle in angle selection step 94, check step 96 determines whether there are edge points at that angle based on the AI ​​vessel recognition results. If so, the edge point marking step 98 marks the edge points. Check step 100 then iterates through the remaining angles until all are processed.

[0145] Angle selection can use fixed intervals (such as 0.5°, 1°), and the 360° range can be covered by the number of iterations M.

[0146] After the edge point marking is completed, curve definition step 102 uses an AI interpolation algorithm to generate curve 78, and curve offset step 104 offsets it towards the reference point to generate offset curve 80. Treatment path definition step 106 generates a treatment path based on the offset curve, and finally, in target area definition step 108, the coordinates of the target area are determined by an AI positioning model (usually using the reference point 74 as a reference system, using radial or Cartesian coordinates).

[0147] In an alternative embodiment, the controller only displays the offset curve 80 superimposed on the image 70 to the user, who directly defines the target area location through an interactive device (such as a mouse click), and the controller uses a machine learning algorithm to specify the irradiation sequence accordingly. **Treatment Execution**

[0148] Figure 5 A flowchart of an AI-integrated automated trabeculae forming algorithm 110 is shown.

[0149] Algorithm 110 begins with the target area specification step 112: The controller uses a machine learning model to specify multiple target areas to be irradiated within the eye and the corresponding energy dose for each area (which can be set uniformly or differently).

[0150] The target area can be accessed through the aforementioned Figure 3-4 The AI ​​method definition is specified after user confirmation. Alternatively, users can directly specify location parameters (such as number and distance from the limbus) by referring to anatomical structures such as the limbus. The controller calculates the specific coordinates using the AI ​​algorithm, and the user confirms and the changes take effect.

[0151] After specifying the target area, the controller uses a deep learning model to scan the eye region in anatomical feature recognition step 114, identifying anatomical features with enhanced radiation sensitivity (in addition to unidentified blood vessels, this includes hyperpigmented limbal areas, lesions such as trachoma / pemphigoid / scleritis / burns, or proliferative tissues such as pterygium / panoptics / arcus senilis / dermoid tumors / limbal tumors). The recognition method can utilize AI-based blood vessel recognition technology.

[0152] The search scope for sensitivity features can be limited to treatment pathway 82 ( Figure 3 It can also cover the entire eye area at a predefined distance near the limbus (such as 1.5mm / 3mm / 5mm).

[0153] After selecting the first target region in step 116, the controller enters the iterative treatment process. Each iteration begins with image processing step 117: the controller uses an AI model to process the eye image (illumination can be provided during acquisition). Figure 6 The process of this step is described in detail.

[0154] Image processing step 117 begins with image acquisition step 118, through camera 54 ( Figure 2 (Acquire eye images.)

[0155] Subsequently, in the benchmark point identification step 119, an AI positioning algorithm is used to attempt to locate benchmark point 74 in the image. Figure 3 If successful, position calculation step 120 calculates the current target area position based on the reference point coordinates using a machine learning model (e.g., when the reference point is located at (x0, y0), the target area is (x0+dx, y0+dy)), thereby compensating for eye movements. If reference point localization fails, proceed directly to decision step 128.

[0156] After position calculation, align the starting step 125 to the alignment motor 61 ( Figure 2 Sending AI-based decision-making control signals 39 to drive the therapeutic laser 43. Figure 2 Align the calculated target position. Then, in inspection step 122, a deep learning model is used to detect whether there are static / dynamic obstructions in the target area. Static obstructions (fixed relative to eye position) include blood vessels, proliferating tissue, etc.; dynamic obstructions (position can change during treatment) include eyelids, eyelashes, fingers, or eyelid openers.

[0157] It should be noted that, in the context of this application (including the claims), "obstruction" refers to any object other than tissue that the user deems irradiable, and its specific scope varies depending on the type of surgery (for example, blood vessels can be irradiated in some surgeries).

[0158] Occlusion identification can combine image technologies such as AI template matching, edge detection, and sequence image difference analysis, and can also integrate user-pre-annotated high-risk area information. Inspection step 122 can also be extended to use machine learning models to identify potential occlusions that meet pre-judgment criteria (such as exceeding the size threshold or moving towards the target area), regardless of whether they directly occlude the target area.

[0159] If an obstruction is detected by the deep learning-based image recognition model, the system will immediately proceed to the decision step 128; otherwise, it will proceed to the inspection step 124, where a machine learning algorithm will be used to assess the radiation sensitivity of the target area to a specified dose (see paragraph 7 for details). If the AI ​​model determines that the sensitivity is too high, it will proceed to the decision step 128; otherwise, in the inspection step 127, the laser alignment accuracy will be verified by an AI-driven analysis module (e.g., intelligent analysis based on the encoder 67 feedback signal 37).

[0160] If the AI ​​system detects an alignment anomaly, it will proceed to decision step 128; otherwise, it may optionally execute AI-supported verification step 135 for final verification (this step can be performed earlier in the image processing stage). For example, it may use a computer vision model to verify that the target area does not involve static "no-light zones" in the camera's field of view; or use a machine learning algorithm to verify that the distance between the target area and the previous target area is within the tolerance range (reflecting eye movement stability). After completing the AI ​​verification, it proceeds to decision step 128.

[0161] In decision step 128, the controller, based on the intelligent processing results of the eye image in AI image analysis step 117, uses a machine learning decision model to determine whether to irradiate the selected target area. If the AI ​​system determines that the following conditions are met: the reference point can be accurately located, there are no obstructions, the target area sensitivity meets the requirements, the laser has been aligned, and AI verification step 135 has been successfully executed, then the controller determines to irradiate the target area.

[0162] After confirming the irradiation, the target area is irradiated with a therapeutic beam in irradiation step 130. Subsequently, the controller intelligently determines whether there are other unprocessed target areas through AI-driven inspection step 132. If so, the controller recommends and selects the next target area based on machine learning in target area selection step 133, and returns to AI image processing step 117 to start a new round of treatment iteration; otherwise, the treatment process is terminated.

[0163] If the AI ​​decision model determines that irradiation should not be performed, then in the heating start-up step 129, a control signal 55 is sent to the driver 53. Figure 2 ), initiate the heating process of the laser medium 45 (e.g. Figure 9As shown, the laser medium does not generate laser output during the heating process. Subsequently, in AI determination step 131, a machine learning algorithm is used to determine whether to skip the current target area (i.e., prohibit irradiation in subsequent iterations). For example, when AI image recognition detects an obstruction or excessive sensitivity in the target area, the system can decide to skip that area.

[0164] If the AI ​​system determines to skip the target area, the controller proceeds to check step 132; otherwise, in AI decision step 137, a deep learning model determines whether the target area position needs to be adjusted. For example, when there are obstructions or sensitive areas in the target area, the AI ​​algorithm may decide to move the target area (usually moving it away from the pupil).

[0165] If the AI ​​system determines that the target area needs to be moved, the controller performs a position adjustment based on machine learning recommendations in the target area shifting step 126, and then returns to the AI ​​image processing step 117. Thus, in subsequent iterations, the radiation source will illuminate the new, AI-optimized position instead of the original position.

[0166] If the AI ​​determines that the target area should not be moved, the controller directly returns to AI image processing step 117, preserving the original position for subsequent iterative illumination.

[0167] In other embodiments, when the AI ​​identifies an obstruction or a target area with high sensitivity, the controller, after executing alignment initiation step 125, inspection step 127, and verification step 135, may use an AI decision model to determine to use energy lower than the standard set value for irradiation. At this time, the controller uses an AI adjustment algorithm to control the pump to the laser medium 45 (…). Figure 2 The energy of the treatment beam 52 is used to intelligently control its output energy.

[0168] In the optional solution, if the AI ​​system detects an obstruction or that the target area is too sensitive, the controller can directly terminate the treatment procedure.

[0169] To improve efficiency, some embodiments perform AI calculation step 120 and alignment start step 125 based on the reference point position of the previous iteration before AI reference point recognition step 119. When the AI ​​detects that the reference point displacement does not exceed a preset threshold, the controller continues to execute the subsequent process of AI image processing step 117.

[0170] Figure 6 The AI ​​image processing step 117 flow is shown in other embodiments:

[0171] When the optical unit 30 is equipped with an alignment laser ( Figure 2After completing AI calculation step 120, the controller activates the dual lasers in alignment initiation step 162 to calculate the position of the target area. Subsequently, in emission step 164, the alignment lasers are triggered to emit a beam (optionally continuous emission). In some embodiments, the alignment beam uses an infrared wavelength greater than 700 nm to avoid visual interference with the patient.

[0172] The controller then acquires an image of the eye. In an exemplary embodiment, a single image can be acquired while the alignment beam is being emitted, comprising a first frame containing the alignment beam and a second frame with the alignment beam filtered out (achieved via a filter matrix of camera 54). For example, infrared wavelengths can be filtered out using a specific filter, or a red wavelength can be selected so that it appears only in the red channel frame.

[0173] In the AI ​​beam alignment and positioning step 166, the controller uses a deep learning model to attempt to identify the alignment beam in the first frame. If AI identification fails, it proceeds directly to the judgment step 128; if successful, it identifies the reference point in the second frame in the AI ​​reference point positioning step 168. If AI reference point identification fails, it proceeds to the judgment step 128; if successful, it proceeds to the AI ​​inspection step 127 to verify: (i) whether the alignment beam hits the target area to calculate the position; (ii) whether the offset between the current position of the reference point and the position of the previous iteration is less than a preset threshold. If any condition is not met, it proceeds to the judgment step 128; otherwise, it continues to execute the AI ​​inspection steps 122 and 124 and the verification step 135.

[0174] In other embodiments, the controller acquires a first image containing the aligned beam and a second image without the aligned beam, and processes them respectively using a deep learning model. Another embodiment uses an AI algorithm to simultaneously identify the aligned beam and the reference point in a single frame image.

[0175] Figure 7 This demonstrates an example flow for AI inspection step 124:

[0176] In the first AI assessment step 134, the controller analyzes current and historical images using a deep learning model to detect any abnormal changes in the eye (such as bleeding, swelling, changes in blood vessel density, bubble formation, or color changes). If the AI ​​does not identify any abnormalities, it determines that the target area is safe to irradiate; otherwise, it proceeds to the next AI assessment.

[0177] AI evaluation processes can employ techniques such as deep learning-based optical flow, pattern recognition, edge detection, image segmentation, differential analysis, or color detection. For example, after aligning current and historical images using a deep learning model, the AI-generated differential image and edge detection can be used to identify abnormal features.

[0178] After AI identifies an anomaly, in the second AI evaluation step 136, it is determined whether the sensitive anatomical feature (determined in the AI ​​anatomical feature recognition step 114) is located in the target area (defined as any part of the feature being within a preset threshold distance of the target area). This threshold is usually set based on the eye movement range and the laser calibration accuracy; in some embodiments, the threshold is less than 3 mm.

[0179] When the AI ​​detects the presence of sensitive anatomical features, the AI ​​overlap prediction step 138 calculates the predicted overlap metric (such as the predicted overlap area) between the treatment beam and the anatomical features. The calculation may assume no beam offset or be based on a preset deviation probability distribution (such as maximum / average / median deviation) generated by the AI ​​for intelligent estimation.

[0180] Subsequently, in the third AI evaluation step 140, a machine learning model is used to determine whether the identified changes originate from the illumination of the sensitive feature by the illuminated target area (by analyzing the relative position of the changed area and the sensitive feature using AI). If a correlation exists, the estimated overlap metric in the actual illumination is calculated in the AI ​​overlap estimation step 142 (e.g., based on the alignment beam position or laser alignment data identified by AI).

[0181] In AI risk measurement calculation step 144, the controller calculates the risk measurement based on the following parameters using a machine learning model: - The presence of sensitive features in the target area (the presence of these features increases the risk). - Predicted overlap metric / Expected energy transfer (positive correlation) -Estimated overlap metric / actual transferred energy (negative correlation) -Patient medical records (age, gender, medication history, contact lens use, intraocular pressure, etc.) - Anatomical features (e.g., higher risk due to large blood vessels) - Similarity between features (type / color / size) - Type of change (higher risk of bleeding / swelling)

[0182] In the fourth AI assessment step 146, if the risk metric calculated by the AI ​​exceeds a preset threshold, it is determined that the target area is too sensitive, and irradiation can be prohibited or the irradiation energy can be reduced through the AI ​​algorithm.

[0183] When multiple sensitive features may trigger changes, AI risk measurement needs to comprehensively consider the types of features and their corresponding overlap measurements.

[0184] Other AI implementation methods include: (i) Evaluate the distance factor between the target area and sensitive areas (such as the pupil or areas with dense blood vessels) using machine learning models. (ii) AI identifies anomalies and verifies and monitors them using additional images. (iii) Simplified judgment logic based on AI-determined features such as presence of sensitive features, overlap measurement exceeding threshold, or feature type matching.

[0185] Figure 8 An example of AI inspection step 124 and target area relocation step 126 is shown: Through AI image analysis 148, the controller identifies a blood pool 150 near the irradiated target area 84a, indicating that irradiation may cause bleeding in blood vessel 72a. Therefore, before irradiating the new target area 84b, the AI ​​system moves it away from blood vessel 72b to reduce the risk.

[0186] AI-optimized heating pulse control Figure 9 ):

[0187] Before irradiation begins, the controller controls the pump source 47 to preheat the laser medium 45 to thermal equilibrium (without generating laser) using a heating pulse sequence 156 via an AI-controlled driver 53.

[0188] At the start of irradiation, the controller initiates a laser-inducing pulse sequence 158 with a cycle of T. Each pulse 152 causes the laser medium to generate a therapeutic beam 52 at the end of the pulse (corresponding to irradiation step 130).

[0189] During irradiation, the controller can replace part of the laser pulse with AI-optimized heating pulse 154 based on AI image analysis results (such as identifying occlusion, abnormal changes, or alignment abnormalities) or system error signals. The total energy of each heating pulse is E2≈E0-E1 (E0 is the laser pulse energy, and E1 is the energy loss during laser generation, which is usually 0-30% of E0).

[0190] Alternatives to AI optimization include: -E2 energy transferred by a single long-duration heating pulse -N low-power heating pulses with a duration of D0 (each transferring E2 / N energy). -N peak power P0 heating pulses with a duration of D0 / N (e.g.) Figure 9 (Example of N=2) The AI ​​system can intelligently select alternative time points, such as t0+k*T / N (k=0…N-1).

[0191] Although the example uses trabeculoplasty, this AI approach is also applicable to other ocular irradiation surgeries (such as transscleral cyclophotocoagulation or tissue shrinkage).

[0192] Those skilled in the art will understand that this invention is not limited to the specific embodiments shown, but also includes combinations and sub-combinations of the above-described AI features, as well as any AI variations and modifications not found in the prior art that can be conceived by those skilled in the art after reading the specification. References incorporated herein by reference are considered part of this application, and in the event of any conflict between their terminology and the express or implied definitions in this specification, the definitions in this specification shall prevail.

Claims

1. An eye-tracking and localization method based on a large artificial intelligence model and an automatic high-risk area (blood vessel, etc.) avoidance system, comprising: Laser radiation source; as well as The AI ​​controller is configured to: Multiple local target regions are identified and designated on images of an ophthalmology patient's eye using deep learning algorithms for irradiation with corresponding predetermined laser energy. The radiation source is controlled to irradiate at least a first target region in the target region according to an irradiation sequence optimized by machine learning; After irradiating the first target area, the real-time image of the eye is processed by a convolutional neural network to identify the microscopic changes in the eye tissue. as well as In response to the changes identified by the AI, the irradiation of a second target region that has not yet been irradiated in the target region with the original energy is suppressed by a decision tree model.

2. The system according to claim 1, characterized in that, The method of suppressing irradiation is as follows: a new target region is redesignated through a reinforcement learning algorithm, and the radiation source is controlled to irradiate the new target region to replace the second target region.

3. The system according to claim 1, characterized in that, The method of suppressing irradiation is as follows: the optimal energy parameters are predicted by a machine learning model, and the radiation source is controlled to irradiate the second target area with AI-optimized energy that is lower than the original energy.

4. The system according to claim 1, characterized in that, The microscopic changes include at least one of bleeding, swelling, color change patterns, or microbubble formation detected by a deep learning image recognition algorithm.

5. The system according to claim 1, characterized in that, The AI ​​controller is also configured to identify anatomical features at the second target region via a deep learning segmentation network and suppress illumination of the second target region in response to the anatomical features.

6. The system according to claim 5, characterized in that, The AI ​​controller is also configured to calculate a predicted overlap metric or predicted transfer energy between the irradiation beam and the anatomical feature, and to suppress irradiation in response to the metric or energy.

7. The system according to claim 1, characterized in that, The AI ​​controller is also configured to calculate a risk metric associated with irradiating the second target region using a risk assessment machine learning model, and to suppress irradiation in response to the risk metric.

8. A target area planning method based on artificial intelligence, characterized in that, include: Multiple edge points are identified in an eye image using computer vision algorithms. Each point is located at a different angle relative to a radially inward reference point and is situated at the edge of a blood vessel. A machine learning path planning algorithm is used to define the target area between the reference point and the edge points. The radiation source is then controlled to irradiate these areas.

9. A laser control method, characterized in that, include: In the process of irradiating the target area by pumping the laser medium with a pulse sequence, the pump source is controlled based on a machine learning algorithm to replace the original laser-generating pulses with one or more heating pulses that are only used to heat the laser medium without generating laser light.

10. A treatment method, characterized in that, include: An α2 agonist was administered to the patient's eye, and within 40 minutes of administration, the eye was treated with an artificial intelligence-controlled laser radiation system.

Images