Real-time controlled light therapy method and system based on retinal safety

Abstract

This invention provides a real-time controlled phototherapy method and system based on retinal safety, belonging to the interdisciplinary technical field of biomedical engineering and ophthalmic medical devices. First, a finite element damage model of a guinea pig's eye is constructed. Phototherapy experiments are conducted on guinea pig eyes under different phototherapy conditions, and phototherapy images of the guinea pig eyes are collected. Then, under different phototherapy inputs, a spatiotemporal distribution dataset of damage is obtained. By replacing human eye parameters, guinea pig image data with equivalent human eye damage labels is obtained. Second, using a general human fundus image dataset as the source domain and guinea pig images with equivalent human eye damage labels as the target domain, transfer learning is performed, and an early identification model for minor human eye damage is obtained through adversarial training. Third, during the human eye phototherapy experiment, the collected human eye images are input into the early identification model to obtain damage values. Finally, by formulating a damage value level control scheme, the phototherapy equipment is controlled step-by-step.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of biomedical engineering and ophthalmic medical devices, and in particular to a real-time controlled phototherapy method and system based on retinal safety. Background Technology

[0002] Since the mid-20th century, laser technology has been widely used in ophthalmology. As early as the 1960s, ruby ​​lasers and other lasers were used for retinal photocoagulation to treat blinding eye diseases such as diabetic retinopathy, pioneering ophthalmic phototherapy. After decades of development, ophthalmic laser treatment now encompasses various types and applications, including retinal photocoagulation, laser iridectomy, laser treatment for secondary cataracts, and excimer lasers for refractive surgery. In these traditional applications, lasers primarily act at high energy levels, aiming to coagulate or ablate diseased tissue. In recent years, low-intensity repetitive light therapy has emerged for ophthalmic disease intervention, particularly in the control of myopia in adolescents. Repetitive low-intensity red light (RLRL) irradiation has become a hot topic in research and clinical practice. Studies have shown that compared to traditional methods (such as increased outdoor activity and low-concentration atropine), RLRL demonstrates good efficacy in controlling axial elongation. Therefore, the application of ophthalmic phototherapy is expanding from traditional high-energy damaging treatments to low-energy protective interventions, and its scope is extending from lesion treatment to functional control. However, while the application of low-intensity phototherapy in ophthalmology is gradually expanding, the potential risks and mechanisms of retinal photodamage have received widespread attention. The retina is a highly sensitive photoreceptor tissue, and under certain conditions, light exposure can cause both photochemical and photothermal damage. When the light intensity is too high or the exposure time is too long, the retina, especially the macula, may experience phototoxic reactions. For example, with red light (approximately 600–700 nm) irradiation, if the intensity exceeds the safety threshold or the single irradiation time is too long, the accumulated light energy will produce a thermal effect, leading to damage and apoptosis of retinal pigment epithelial cells (RPE), and in severe cases, irreversible damage to the macula. Animal experiments and histological studies have confirmed that continuous or excessively strong laser irradiation can cause damage to retinal structures. Therefore, when promoting phototherapy techniques (especially RLRL red light therapy) for chronic eye diseases such as myopia, the safety of phototherapy has become one of the most pressing concerns for clinicians and patients.

[0003] Currently, clinical assessments of phototherapy safety primarily rely on post-treatment examinations and empirical parameters. On one hand, manufacturers and clinicians mainly set irradiation parameters based on limits stipulated in laser safety standards. However, these thresholds are usually average safety levels calculated using general models or extrapolated from animal data, lacking consideration for individual differences. On the other hand, during treatment, doctors typically monitor for abnormal changes in the retina through regular follow-up examinations (such as changes in visual acuity, fundus photography, or OCT optical coherence tomography). For example, according to the latest expert consensus, children and adolescents receiving RLRL red light therapy should have fundus OCT and other imaging examinations every 3–6 months to detect structural changes in the macular region as early as possible. However, this post-treatment monitoring is lagging; by the time visible signs of retinal damage are detected, it often indicates that the retina has already been damaged, requiring immediate cessation of treatment. Therefore, existing phototherapy equipment lacks real-time safety monitoring and feedback mechanisms, failing to dynamically sense and adjust the retinal tissue's response to light during treatment. Especially in the context of long-term, high-frequency myopia phototherapy applications, the lack of proactive safety protection may result in the failure to intervene early in the cumulative light damage in a few patients. In summary, how to ensure retinal safety while guaranteeing therapeutic efficacy has become a critical issue that urgently needs to be addressed in ophthalmic phototherapy technology. Summary of the Invention

[0004] To address the problems in the prior art, this invention provides a real-time controlled phototherapy method and system based on retinal safety. The invention first constructs a finite element damage model of a guinea pig's eye, conducts phototherapy experiments on guinea pig eyes under different phototherapy conditions, and collects phototherapy images of the guinea pig eyes. Then, under different phototherapy inputs, a spatiotemporal distribution dataset of damage is obtained. By replacing human eye parameters, guinea pig image data with equivalent human eye damage labels is obtained. Second, using a general human fundus image dataset as the source domain and guinea pig images with equivalent human eye damage labels as the target domain, transfer learning is performed, and an early identification model for minor human eye damage is obtained through adversarial training. Third, during the human eye phototherapy experiment, the collected human eye images are input into the early identification model to obtain damage values. Finally, by formulating a damage value level control scheme, the phototherapy equipment is controlled step-by-step. To achieve the above objectives, the technical solution is as follows:

[0005] On one hand, the present invention provides a method for reconstructing the outer contour function of a projectile head shape based on machine vision, the method comprising:

[0006] S1. Based on the retinal parameters of guinea pigs, a finite element simulation model of laser-induced damage to guinea pigs was obtained through a physical model of laser radiation and heat conduction.

[0007] S2. Based on the phototherapy experimental conditions and the finite element simulation model of laser-induced damage in guinea pigs, a phototherapy experiment and image acquisition were carried out in guinea pigs. Through simulation calculation with human eye parameter substitution, a guinea pig image dataset and spatiotemporal distribution dataset with equivalent human eye damage labels were obtained.

[0008] S3. Based on the general image dataset of the human fundus, a general low-level feature recognition model of the retina is obtained through source domain learning training of the damage detection network model;

[0009] S4. Based on the guinea pig image dataset with the equivalent human eye injury label, the target domain of the general retinal low-level feature recognition model is trained to obtain the early micro-damage recognition model of guinea pig images.

[0010] S5. Based on the spatiotemporal distribution dataset of the damage and the early minor damage recognition model of the guinea pig image, an early minor damage recognition model of human eye image is obtained through adversarial training of the adversarial learning network model.

[0011] S6. Using fundus color imaging and OCT image acquisition devices, collect multimodal image data of the retina during human eye phototherapy;

[0012] S7. Based on the multimodal image data, the early damage value of the human eye is obtained through the early micro-damage recognition model of human eye image;

[0013] S8. Based on the early damage value of the human eye, the control strategy of the phototherapy device is obtained by judging the damage level.

[0014] Optionally, in S2, based on the phototherapy experimental conditions and the finite element simulation model of laser-induced damage in guinea pigs, a phototherapy experiment and image acquisition are conducted in guinea pigs. Through simulation calculations with human eye parameter substitution, an equivalent human eye damage label guinea pig image dataset and damage spatiotemporal distribution dataset are obtained, including:

[0015] S21. Based on the experimental conditions for phototherapy, conduct a phototherapy experiment on guinea pigs and collect images of the phototherapy process to obtain a guinea pig phototherapy experiment image dataset, including: a fundus color image dataset and a retinal optical coherence tomography image dataset;

[0016] S22. Based on the guinea pig phototherapy experimental image dataset, the finite element simulation verification model of laser-induced damage in guinea pigs was obtained by modifying the finite element simulation model of laser-induced damage in guinea pigs.

[0017] S23. Based on the experimental conditions of phototherapy, the spatiotemporal distribution dataset of damage was obtained through simulation calculation of the finite element simulation verification model of laser-induced damage in guinea pigs.

[0018] S24. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, the experimental image dataset of phototherapy in guinea pigs, and the experimental conditions of phototherapy, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculation by replacing human eye parameters.

[0019] Optionally, in step S24, based on the finite element simulation verification model of laser-induced damage in guinea pigs, the guinea pig phototherapy experimental image dataset, and the phototherapy experimental conditions, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculations using human eye parameter substitution. This dataset includes:

[0020] S241. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, a finite element simulation verification model of laser-induced damage in human eyes is obtained by replacing human eye parameters.

[0021] S242. Based on the experimental conditions of phototherapy, the human eye damage dataset is obtained through simulation calculation of the finite element simulation verification model of laser-induced damage to the human eye;

[0022] S243. Based on the human eye injury dataset and the guinea pig phototherapy experimental image dataset, and using the phototherapy experimental conditions as an intermediate mapping relationship, obtain the correspondence between the human eye injury dataset and the guinea pig phototherapy experimental image dataset;

[0023] S244. Based on the guinea pig phototherapy experimental image dataset, a guinea pig image dataset with equivalent human eye injury labels is obtained through the labeling transformation of the corresponding relationship.

[0024] Optionally, in S3, a general low-level feature recognition model for the retina is obtained by training a damage detection network model through source domain learning based on a general human fundus image dataset, including:

[0025] S31. Based on the general human fundus image dataset, an enhanced general human fundus image dataset is obtained through preprocessing;

[0026] S32. Based on the enhanced general image dataset of human fundus, a general feature recognition model of the fundus layer is obtained by pre-training a shared backbone CNN;

[0027] S33. Based on the underlying general feature model of the fundus image, a general retinal structure recognition model is obtained by training a Fast R-CNN sub-network;

[0028] S34. Based on this general retinal structure recognition model, a general retinal low-level feature recognition model is obtained by training and recognizing the U-Net sub-network.

[0029] Optionally, in S4, based on the guinea pig image dataset with the equivalent human eye injury label and the retinal general low-level feature recognition model, an early minor injury recognition model for guinea pig images is obtained through target domain training, including:

[0030] S41. Based on this general retinal low-level feature recognition model, by freezing the low-level parameters, a recognition model of reusable general features is obtained;

[0031] S42. Based on the guinea pig image dataset with the equivalent human eye injury label and the recognition model with the reusable general features, a guinea pig image early minor injury recognition model is obtained by training by adjusting the top-level parameters.

[0032] Optionally, in S5, based on the spatiotemporal distribution dataset of the damage and the early minor damage recognition model for guinea pig images, an early minor damage recognition model for human eye images is obtained through adversarial training of an adversarial learning network model, including:

[0033] S51. Based on the spatiotemporal distribution dataset of the damage, constrain the adversarial learning network to obtain a physically constrained adversarial learning network model;

[0034] S52. Based on the guinea pig image early minor damage recognition model, an adversarial game training of a physical constraint adversarial learning network model is used to obtain a human eye image early minor damage recognition model.

[0035] On the other hand, the present invention provides a real-time controlled phototherapy system based on retinal safety, which is applied to a real-time controlled phototherapy method based on retinal safety. The system includes:

[0036] The finite element model acquisition module is used to obtain a finite element simulation model of laser-induced damage in guinea pigs based on the parameters of the guinea pig's retina and the physical model of heat caused by laser radiation and conduction.

[0037] The simulation and phototherapy experiment fusion module is used to carry out guinea pig phototherapy experiments and image acquisition based on the phototherapy experiment conditions and the finite element simulation model of the laser-induced damage to the guinea pig. Through simulation calculation with human eye parameter replacement, the equivalent human eye damage label guinea pig image dataset and damage spatiotemporal distribution dataset are obtained.

[0038] The universal human fundus source domain training module is used to obtain a universal retinal low-level feature recognition model by learning and training the source domain of the damage detection network model based on the universal human fundus image dataset.

[0039] The guinea pig image target domain training module is used to train the target domain of the guinea pig image dataset with equivalent human eye damage labels to obtain an early micro-damage recognition model for guinea pig images.

[0040] The adversarial training module is used to obtain an early minor damage recognition model for human eye images by adversarially training the adversarial learning network model based on the damage spatiotemporal distribution dataset and the guinea pig image early minor damage recognition model.

[0041] The human eye retina multimodal image acquisition module is used to acquire multimodal image data of the human eye retina during phototherapy using fundus color imaging and OCT image acquisition devices.

[0042] The human eye early damage value acquisition module is used to obtain the human eye early damage value based on the multimodal image data and the human eye image early micro-damage recognition model.

[0043] The phototherapy device control strategy generation module is used to obtain the control strategy of the phototherapy device by judging the damage level based on the early damage value of the human eye.

[0044] Compared with the prior art, the technical solution of the present invention has at least the following beneficial effects:

[0045] The aforementioned solution integrates real-time retinal multimodal imaging monitoring into phototherapy equipment, constructing a closed-loop "sensing-analysis-control" structure. This fills the gap in existing phototherapy equipment lacking a monitoring feedback mechanism, endowing the equipment with the ability to perceive the patient's retinal condition in real time and providing a fundamental guarantee for treatment safety. Secondly, through an early damage detection model based on transfer learning and adversarial training, which integrates theoretical simulation data and animal experiments, the solution enables the equipment to possess the core capability of intelligently judging retinal condition and identifying early signs of damage, overcoming the limitations of existing technologies that lack intelligent image analysis. Thirdly, by embedding retinal illumination safety thresholds and grading standards, the solution achieves adaptive adjustment of the light source output. First, it can dynamically adjust treatment parameters according to the safety status of the retina, solving the problem of inflexible fixed parameter settings in existing technologies and ensuring that each treatment is within a safe range. Second, it establishes a complete biological verification chain based on animal experiments, determines the damage pattern and reversibility boundary by combining histological gold standards, and transforms animal experimental results into early risk warnings for human eyes by leveraging the structural homology of mammalian retina, effectively overcoming the technical bottleneck of scarce clinical damage samples and providing reliable support for safety monitoring. Third, it integrates various functional modules to form a complete phototherapy system. The overall integrated solution broadens the safe application scope of phototherapy equipment and improves the safety and effectiveness of treatment. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention;

[0048] Figure 2 This is a flowchart illustrating the process of fusing simulation and phototherapy experiments to obtain an equivalent human eye injury label guinea pig image dataset and injury spatiotemporal distribution dataset, based on a real-time controlled phototherapy method embodiment of the present invention for retinal safety.

[0049] Figure 3 This is a flowchart of an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention, which obtains a guinea pig image dataset with equivalent human eye damage labels by replacing the guinea pig finite element simulation model with human eye parameters.

[0050] Figure 4 This is the process for obtaining the general underlying feature recognition model of the retina in an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention;

[0051] Figure 5 Yes, this is a flowchart of the target domain training process of an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention.

[0052] Figure 6 This is a flowchart of the adversarial learning training process in an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention.

[0053] Figure 7 This is a system block diagram of an embodiment of the real-time controlled phototherapy system based on retinal safety of the present invention. Detailed Implementation

[0054] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0055] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] like Figure 1 The flowchart shown is an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention. The present invention provides a real-time controlled phototherapy method based on retinal safety, which is implemented by a real-time controlled phototherapy system based on retinal safety. The method includes:

[0058] S1. Based on the retinal parameters of guinea pigs, a finite element simulation model of laser-induced damage to guinea pigs was obtained through a physical model of laser radiation and heat conduction.

[0059] Because ocular tissues are sensitive to thermal effects, accurate study of intraocular temperature distribution under laser irradiation is crucial. This module, based on COMSOL Multiphysics finite element software, constructs detailed human and guinea pig eye models for analyzing laser photothermal effects. Simulations of guinea pig retinal damage are performed using COMSOL, with comparisons made to guinea pig animal experiments during the simulation process. The model of laser-induced retinal damage is continuously modified, and finally, the validated model is used to simulate human retinal damage, effectively translating this research from animal models to human applications.

[0060] The core of eye simulation model design lies in accurately reflecting the complex geometry, multi-layered biological tissue structure, and heterogeneous physical and optical properties of the eyeball. The simulation model is based on a reasonable simplification of the real eyeball anatomy.

[0061] The study employs a three-dimensional model to more comprehensively capture the complex heat transfer and laser irradiation processes. The human eye model is subdivided into eight main tissues: cornea, aqueous humor, iris, lens, vitreous body, retina, choroid, and sclera. Guinea pig eye structures are fundamentally similar to human structures, both containing refractive media such as the cornea, aqueous humor, lens, and vitreous body, as well as tissues like the retina and sclera, but differing slightly in the proportions and microstructures of each tissue. The fundus is a crucial area for laser treatment. Based on the team's previous research, a precise mathematical model was established to predict the actual irradiance received by the retina under different conditions and to determine safety thresholds. Therefore, the posterior segment of the eye was refined in the human eye simulation model. Since the retinal pigment epithelium, containing melanin, is the primary energy absorption area, the retina is further subdivided into the neuroretina and the pigment epithelium. The model independently simulates the neuroretina, retinal pigment epithelium, choroid, and sclera to improve accuracy. Simulation accuracy depends on the accurate definition of thermophysical and optical parameters. Thermophysical parameters include density, specific heat capacity, and thermal conductivity. Optical parameters describe a tissue's ability to absorb and scatter laser energy, including absorption coefficient and scattering coefficient.

[0062] The damage index calculated by COMSOL is correlated with the microscopic changes obtained from OCT and fundus photography in the experiment, and used as a label to improve the accuracy of model training. The temporal and spatial patterns of damage occurrence are further calculated in the simulation, and these patterns are incorporated into the network as constraints during model training to standardize the model. Since there are few samples of human retinal optical damage, the modified model is modified by replacing parameters with those of the human eye, and calibrated using the physical difference coefficients calculated by COMSOL, thus obtaining numerical information on human retinal optical damage. COMSOL provides conversion logic from guinea pigs to the human eye, thereby theoretically guiding and verifying the safety and accuracy of the system in human applications. The model can study the sensitivity of key variables such as laser power and pupil diameter to temperature peaks and damage range, providing a fast, easy, and non-invasive assessment method that can effectively and safely verify laser damage.

[0063] S2. Based on the phototherapy experimental conditions and the finite element simulation model of laser-induced damage in guinea pigs, a phototherapy experiment and image acquisition were carried out in guinea pigs. Through simulation calculation with human eye parameter substitution, a guinea pig image dataset and spatiotemporal distribution dataset with equivalent human eye damage labels were obtained.

[0064] Specifically, such as Figure 2 The flowchart shown is a process for obtaining an equivalent human eye injury label guinea pig image dataset and injury spatiotemporal distribution dataset by fusing simulation and phototherapy experiments in an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention. In step S2, based on the phototherapy experimental conditions and the finite element simulation model of the laser-induced damage in the guinea pig, a guinea pig phototherapy experiment and image acquisition are carried out. Through simulation calculations using human eye parameter substitution, an equivalent human eye injury label guinea pig image dataset and injury spatiotemporal distribution dataset are obtained, including:

[0065] S21. Based on the experimental conditions for phototherapy, conduct a phototherapy experiment on guinea pigs and collect images of the phototherapy process to obtain a guinea pig phototherapy experiment image dataset, including: a fundus color image dataset and a retinal optical coherence tomography image dataset;

[0066] Healthy adult guinea pigs (wild type) were selected as experimental subjects because their retinal structure is comparable to that of humans. The guinea pigs were randomly divided into several groups, including a control group (no phototherapy) and an experimental group (receiving phototherapy with different parameters). The experimental group was further subdivided according to the total exposure time or dose, such as: a low-dose group (simulating conventional safe doses), a medium-dose group, and a high-dose group (approaching or slightly exceeding existing safety limits to assess the damage threshold). The sample size for each group was determined according to statistical requirements, generally with no fewer than 6 guinea pigs per group to ensure statistical significance.

[0067] A red laser source (wavelength approximately 650 nm) similar to that used in clinical myopia phototherapy was employed. Experimental parameters for each group were precisely controlled by adjusting the light power density and irradiation time. The low-dose group used a power density (2 mW at eye level) that met human eye safety standards, irradiating the guinea pigs' eyes for 3 minutes daily for 8 consecutive days. The medium-dose group increased the daily irradiation time or power (5 or 10 mW at eye level), irradiating the guinea pigs' eyes for 3 minutes daily for 8 consecutive days. The high-dose group (20 mW at eye level) irradiated the guinea pigs' eyes for 3 minutes daily for 8 consecutive days, with the total cumulative dose approaching or exceeding the estimated retinal safety threshold. During irradiation, the guinea pigs were anesthetized and their eyes were kept directly facing the light source by a fixation device to ensure consistent irradiation conditions each time. Ocular characteristics (such as corneal transparency and pupillary response) were monitored in real time during the experiment to detect signs of acute damage; irradiation of the animal was immediately terminated if any abnormalities were observed.

[0068] Throughout the experiment, fundus color imaging (RGB images) was performed on each group of guinea pigs to visually observe changes in the appearance of the retina. Specifically, fundus color images were taken at baseline (before irradiation), during irradiation (e.g., after a certain number of days), and at the end of irradiation. High-resolution color photographs of the retina were obtained using a small animal fundus imaging device, focusing on changes in the optic disc, macula (guinea pigs lack a macula, but the posterior pole can be observed), and retinal vessels. Data acquisition time points included, for example, 1 day before irradiation, 3 days during irradiation, 7 days during irradiation, and 1 day after the final irradiation. Illumination and camera parameters were kept consistent throughout all image acquisition processes to ensure image comparability.

[0069] Layered imaging of the guinea pig retina using optical coherence tomography (OCT) equipment specifically designed for small animals is a crucial method for obtaining subtle structural changes in the retina. OCT scans were performed at similar time points to RGB imaging (before irradiation, multiple times during irradiation, and after irradiation), focusing on acquiring cross-sectional images of the macular region (the central area of ​​the guinea pig retina). Acquisition criteria included: maintaining the same scanning position and depth for each eye to obtain clear images of the retinal layered structure (including the retinal nerve fiber layer, nuclear layer, photoreceptor layer, and RPE layer, etc.).

[0070] S22. Based on the guinea pig phototherapy experimental image dataset, the finite element simulation verification model of laser-induced damage in guinea pigs was obtained by modifying the finite element simulation model of laser-induced damage in guinea pigs.

[0071] S23. Based on the experimental conditions of phototherapy, the spatiotemporal distribution dataset of damage was obtained through simulation calculation of the finite element simulation verification model of laser-induced damage in guinea pigs.

[0072] S24. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, the experimental image dataset of phototherapy in guinea pigs, and the experimental conditions of phototherapy, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculation by replacing human eye parameters.

[0073] Furthermore, such as Figure 3 The flowchart shown in this embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention illustrates the process of obtaining a guinea pig image dataset with equivalent human eye damage labels by replacing human eye parameters in a guinea pig finite element simulation model. In step S24, based on the finite element simulation verification model of laser-induced damage in guinea pigs, the guinea pig phototherapy experimental image dataset, and the phototherapy experimental conditions, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculation by replacing human eye parameters. This dataset includes:

[0074] S241. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, a finite element simulation verification model of laser-induced damage in human eyes is obtained by replacing human eye parameters.

[0075] S242. Based on the experimental conditions of phototherapy, the human eye damage dataset is obtained through simulation calculation of the finite element simulation verification model of laser-induced damage to the human eye;

[0076] S243. Based on the human eye injury dataset and the guinea pig phototherapy experimental image dataset, and using the phototherapy experimental conditions as an intermediate mapping relationship, obtain the correspondence between the human eye injury dataset and the guinea pig phototherapy experimental image dataset;

[0077] S244. Based on the guinea pig phototherapy experimental image dataset, a guinea pig image dataset with equivalent human eye injury labels is obtained through the labeling transformation of the corresponding relationship.

[0078] S3. Based on the general image dataset of the human fundus, a general low-level feature recognition model of the retina is obtained through source domain learning training of the damage detection network model;

[0079] Specifically, such as Figure 4 The illustrated embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention shows the process for obtaining the retinal universal low-level feature recognition model. In step S3, based on a general human fundus image dataset, the retinal universal low-level feature recognition model is obtained through source domain learning training of a damage detection network model, including:

[0080] S31. Based on the general human fundus image dataset, an enhanced general human fundus image dataset is obtained through preprocessing;

[0081] S32. Based on the enhanced general image dataset of human fundus, a general feature recognition model of the fundus layer is obtained by pre-training a shared backbone CNN;

[0082] S33. Based on the underlying general feature model of the fundus image, a general retinal structure recognition model is obtained by training a Fast R-CNN sub-network;

[0083] S34. Based on this general retinal structure recognition model, a general retinal low-level feature recognition model is obtained by training and recognizing the U-Net sub-network.

[0084] S4. Based on the guinea pig image dataset with the equivalent human eye injury label, the target domain of the general retinal low-level feature recognition model is trained to obtain the early micro-damage recognition model of guinea pig images.

[0085] Specifically, such as Figure 5 The flowchart shown is a target domain training flowchart of an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention. In S4, based on the guinea pig image dataset with equivalent human eye injury labels, the target domain is trained through the retinal general low-level feature recognition model to obtain an early minor injury recognition model for guinea pig images, including:

[0086] S41. Based on this general retinal low-level feature recognition model, by freezing the low-level parameters, a recognition model of reusable general features is obtained;

[0087] S42. Based on the guinea pig image dataset with the equivalent human eye injury label and the recognition model with the reusable general features, a guinea pig image early minor injury recognition model is obtained by training by adjusting the top-level parameters.

[0088] S5. Based on the spatiotemporal distribution dataset of the damage and the early minor damage recognition model of the guinea pig image, an early minor damage recognition model of human eye image is obtained through adversarial training of the adversarial learning network model.

[0089] Specifically, such as Figure 6 The flowchart shown is an adversarial learning training process for an embodiment of the real-time controlled phototherapy method based on retinal safety of the present invention. In step S5, based on the spatiotemporal distribution dataset of the injury and the guinea pig image early minor injury recognition model, an adversarial learning network model is used to obtain a human eye image early minor injury recognition model through adversarial training, including:

[0090] S51. Based on the spatiotemporal distribution dataset of the damage, constrain the adversarial learning network to obtain a physically constrained adversarial learning network model;

[0091] S52. Based on the guinea pig image early minor damage recognition model, an adversarial game training of a physical constraint adversarial learning network model is used to obtain a human eye image early minor damage recognition model.

[0092] S6. Using fundus color imaging and OCT image acquisition devices, collect multimodal image data of the retina during human eye phototherapy;

[0093] S7. Based on the multimodal image data, the early damage value of the human eye is obtained through the early micro-damage recognition model of human eye image;

[0094] S8. Based on the early damage value of the human eye, the control strategy of the phototherapy device is obtained by judging the damage level.

[0095] Specifically, in order to achieve intelligent control of laser parameters, this study, based on previous experiments on the evolution of optical damage to the guinea pig retina, analyzed the evolution of OCT microscopic features during the damage process and summarized the "four-level early warning standard for retinal optical damage" as follows:

[0096] Grade 0 (Completely Safe Zone): OCT shows clear retinal structure, no change in ONL thickness, and continuous and sharp EZ bands. Phototherapy equipment control strategy: Maintain current treatment power.

[0097] Grade I (Subclinical Alert Zone): OCT shows a slight decrease in the reflectivity of the EZ zone (IS / OS), or scattered punctate high-reflectivity signals on the RPE layer surface, indicating an increased cellular metabolic load in a reversible phase. Phototherapy equipment control strategy: The system automatically and dynamically reduces laser power (e.g., by 30%) and shortens the remaining treatment time per session.

[0098] Grade II (Reversible Damage Zone): OCT shows typical "dome-shaped hyperreflective" lesions (located between the RPE and EZ zones), or local retinal thickening due to microedema >5%. Phototherapy device control strategy: Immediately shut down (stop) the current treatment and lock the device for at least 3 days until the OCT re-examination indicators return to Grade 0.

[0099] Grade III (Irreversible Damage Area): OCT shows significant thinning of the ONL layer (>10%), indicating photoreceptor apoptosis; or extensive EZ band rupture. Phototherapy device control strategy: The system is permanently locked and a high-risk alarm is issued, prompting immediate medical attention.

[0100] The standard is based on the existence of a "reversible window" for photochemical damage. In Grade I, non-invasive treatment can be achieved through adaptive fine-tuning of the PD algorithm; in Grade II, irreversible damage is avoided through forced circuit breaking. This system adopts a hierarchical closed-loop control strategy of "graded early warning (upper layer) + PID / PD continuous control (lower layer)": the upper layer is responsible for safety boundaries and interlocking, and the lower layer is responsible for power tracking within the safety boundaries, thus balancing efficacy and safety.

[0101] The phototherapy system employs a closed-loop feedback system of "perception-decision-execution." Ultimately, the system achieves the judgment, identification, and feedback of early damage, realizing a paradigm shift from "post-damage assessment" to "pre-damage warning." This system aims to intervene before irreversible damage is caused by lasers, serving as a proactive, feedforward safety mechanism. It prevents laser damage during treatment, further ensuring the safety of laser ophthalmology treatments and achieving a transformation from "experience-based medicine" to "data-driven, precise, and automated treatment."

[0102] like Figure 7 The diagram shown is a system block diagram of an embodiment of the real-time controlled phototherapy system based on retinal safety of the present invention. The present invention provides a real-time controlled phototherapy system based on retinal safety, which is applied to a real-time controlled phototherapy method based on retinal safety. The system includes: a finite element model acquisition module, a simulation and phototherapy experiment fusion module, a human eye universal fundus source domain training module, a guinea pig image target domain training module, an adversarial training module, a human eye retinal multimodal image acquisition module, a human eye early damage value acquisition module, and a phototherapy equipment control strategy generation module. Specifically:

[0103] The finite element model acquisition module is used to obtain a finite element simulation model of laser-induced damage in guinea pigs based on the parameters of the guinea pig's retina and the physical model of heat caused by laser radiation and conduction.

[0104] The simulation and phototherapy experiment fusion module is used to carry out guinea pig phototherapy experiments and image acquisition based on the phototherapy experiment conditions and the finite element simulation model of the laser-induced damage to the guinea pig. Through simulation calculation with human eye parameter replacement, the equivalent human eye damage label guinea pig image dataset and damage spatiotemporal distribution dataset are obtained.

[0105] The universal human fundus source domain training module is used to obtain a universal retinal low-level feature recognition model by learning and training the source domain of the damage detection network model based on the universal human fundus image dataset.

[0106] The guinea pig image target domain training module is used to train the target domain of the guinea pig image dataset with equivalent human eye damage labels to obtain an early micro-damage recognition model for guinea pig images.

[0107] The adversarial training module is used to obtain an early minor damage recognition model for human eye images by adversarially training the adversarial learning network model based on the damage spatiotemporal distribution dataset and the guinea pig image early minor damage recognition model.

[0108] The human eye retina multimodal image acquisition module is used to acquire multimodal image data of the human eye retina during phototherapy using fundus color imaging and OCT image acquisition devices.

[0109] The human eye early damage value acquisition module is used to obtain the human eye early damage value based on the multimodal image data and the human eye image early micro-damage recognition model.

[0110] The phototherapy device control strategy generation module is used to obtain the control strategy of the phototherapy device by judging the damage level based on the early damage value of the human eye.

[0111] This invention provides a real-time controlled phototherapy method and system based on retinal safety. First, a finite element model of guinea pig eye damage is constructed. Phototherapy experiments are conducted on guinea pig eyes under different phototherapy conditions, and phototherapy images of the guinea pig eyes are collected. Then, under different phototherapy inputs, a spatiotemporal distribution dataset of damage is obtained. By replacing human eye parameters, guinea pig image data with equivalent human eye damage labels is obtained. Second, a general human fundus image dataset is used as the source domain, and guinea pig images with equivalent human eye damage labels are used as the target domain. Transfer learning is performed, and an early identification model for minor human eye damage is obtained through adversarial training. Third, during the human eye phototherapy experiment, the collected human eye images are input into the early identification model to obtain damage values. Finally, by formulating a damage value level control scheme, the phototherapy equipment is controlled step-by-step.

[0112] It is understood that the present invention has been described through the above embodiments and should not be construed as limiting the implementation and scope of the present invention. Those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A real-time controlled phototherapy method based on retinal safety, characterized in that, The method includes: S1. Based on the retinal parameters of guinea pigs, a finite element simulation model of laser-induced damage to guinea pigs was obtained through a physical model of laser radiation and heat conduction. S2. Based on the phototherapy experimental conditions and the finite element simulation model of laser-induced damage in guinea pigs, conduct phototherapy experiments and image acquisition in guinea pigs. Through simulation calculations with human eye parameter substitution, obtain guinea pig image datasets and spatiotemporal distribution datasets with equivalent human eye damage labels. S3. Based on the general image dataset of the human fundus, a general low-level feature recognition model of the retina is obtained through source domain learning training of the damage detection network model; S4. Based on the guinea pig image dataset with equivalent human eye injury labels, the target domain of the retinal general low-level feature recognition model is trained to obtain an early micro-damage recognition model for guinea pig images. S5. Based on the damage spatiotemporal distribution dataset and the guinea pig image early minor damage recognition model, an adversarial training of the adversarial learning network model is used to obtain a human eye image early minor damage recognition model. S6. Using fundus color imaging and OCT image acquisition devices, collect multimodal image data of the retina during human eye phototherapy; S7. Based on the multimodal image data, the early damage value of the human eye is obtained through the early micro-damage recognition model of human eye images; S8. Based on the early damage value of the human eye, the control strategy of the phototherapy device is obtained by judging the damage level.

2. The real-time controlled phototherapy method based on retinal safety according to claim 1, characterized in that, In step S2, based on the phototherapy experimental conditions and the finite element simulation model of laser-induced damage in guinea pigs, a phototherapy experiment and image acquisition are conducted in guinea pigs. Through simulation calculations using human eye parameter substitution, an equivalent human eye damage label guinea pig image dataset and a spatiotemporal distribution dataset of damage are obtained, including: S21. Based on the phototherapy experimental conditions, conduct a guinea pig phototherapy experiment and collect images of the guinea pig phototherapy experimental process to obtain a guinea pig phototherapy experimental image dataset, including: a fundus color image dataset and a retinal optical coherence tomography image dataset; S22. Based on the guinea pig phototherapy experimental image dataset, a finite element simulation verification model for guinea pig laser-induced damage is obtained by modifying the finite element simulation model of guinea pig laser-induced damage. S23. Based on the phototherapy experimental conditions, the spatiotemporal distribution dataset of the damage is obtained through simulation calculation of the finite element simulation verification model of laser-induced damage in guinea pigs; S24. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, the guinea pig phototherapy experimental image dataset, and the phototherapy experimental conditions, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculation by replacing human eye parameters.

3. The real-time controlled phototherapy method based on retinal safety according to claim 2, characterized in that, In step S24, based on the finite element simulation verification model of laser-induced damage in guinea pigs, the guinea pig phototherapy experimental image dataset, and the phototherapy experimental conditions, a guinea pig image dataset with equivalent human eye damage labels is obtained through simulation calculations using human eye parameter substitution. This dataset includes: S241. Based on the finite element simulation verification model of laser-induced damage in guinea pigs, a finite element simulation verification model of laser-induced damage in human eyes is obtained by replacing human eye parameters. S242. Based on the phototherapy experimental conditions, the human eye damage dataset is obtained through simulation calculations of the finite element simulation verification model of laser-induced damage to the human eye; S243. Based on the human eye injury dataset and the guinea pig phototherapy experimental image dataset, and using the phototherapy experimental conditions as an intermediate mapping relationship, obtain the correspondence between the human eye injury dataset and the guinea pig phototherapy experimental image dataset; S244. Based on the guinea pig phototherapy experimental image dataset, a guinea pig image dataset with equivalent human eye injury labels is obtained through the labeling transformation of the corresponding relationship.

4. The real-time controlled phototherapy method based on retinal safety according to claim 1, characterized in that, In step S3, based on a general human fundus image dataset, a general low-level feature recognition model for the retina is obtained through source domain learning training of an injury detection network model, including: S31. Based on the general human fundus image dataset, an enhanced general human fundus image dataset is obtained through preprocessing; S32. Based on the enhanced general image dataset of human fundus, a general feature recognition model of the fundus layer is obtained by pre-training a shared backbone CNN; S33. Based on the underlying general feature model of the fundus image, a general retinal structure recognition model is obtained by training a Fast R-CNN sub-network; S34. Based on the aforementioned general retinal structure recognition model, a general retinal low-level feature recognition model is obtained by training and recognizing the U-Net sub-network.

5. The real-time controlled phototherapy method based on retinal safety according to claim 1, characterized in that, In step S4, based on the guinea pig image dataset with the equivalent human eye injury label and the retinal general low-level feature recognition model, an early minor injury recognition model for guinea pig images is obtained through target domain training, including: S41. Based on the aforementioned general retinal low-level feature recognition model, by freezing the low-level parameters, a recognition model with reusable general features is obtained; S42. Based on the guinea pig image dataset with the equivalent human eye injury label and the recognition model with the reusable general features, a guinea pig image early minor injury recognition model is obtained by training by adjusting the top-level parameters.

6. The real-time controlled phototherapy method based on retinal safety according to claim 1, characterized in that, In step S5, based on the spatiotemporal distribution dataset of the damage and the early minor damage recognition model of the guinea pig image, an early minor damage recognition model of the human eye image is obtained through adversarial training of the adversarial learning network model, including: S51. Based on the damage spatiotemporal distribution dataset, constrain the adversarial learning network to obtain a physically constrained adversarial learning network model; S52. Based on the guinea pig image early minor damage recognition model, an adversarial game training of a physical constraint adversarial learning network model is used to obtain a human eye image early minor damage recognition model.

7. A real-time controlled phototherapy system based on retinal safety, used to implement the real-time controlled phototherapy method based on retinal safety as described in any one of claims 1-6, characterized in that, The system includes: The finite element model acquisition module is used to obtain a finite element simulation model of laser-induced damage in guinea pigs based on the parameters of the guinea pig's retina and the physical model of heat caused by laser radiation and conduction. The simulation and phototherapy experiment fusion module is used to carry out guinea pig phototherapy experiments and image acquisition based on the phototherapy experiment conditions and the finite element simulation model of laser-induced damage in guinea pigs. Through simulation calculation with human eye parameter replacement, it obtains guinea pig image dataset and damage spatiotemporal distribution dataset with equivalent human eye damage labels. The universal human fundus source domain training module is used to obtain a universal retinal low-level feature recognition model by learning and training the source domain of the damage detection network model based on the universal human fundus image dataset. The guinea pig image target domain training module is used to obtain an early micro-damage recognition model for guinea pig images by training the target domain of the retinal general low-level feature recognition model based on the guinea pig image dataset with equivalent human eye damage labels. The adversarial training module is used to obtain an early minor damage recognition model for human eye images by adversarial training of the adversarial learning network model based on the damage spatiotemporal distribution dataset and the guinea pig image early minor damage recognition model. The human eye retina multimodal image acquisition module is used to acquire multimodal image data of the human eye retina during phototherapy using fundus color imaging and OCT image acquisition devices. The human eye early damage value acquisition module is used to obtain the human eye early damage value based on the multimodal image data and the human eye image early micro-damage recognition model. The phototherapy device control strategy generation module is used to obtain the control strategy of the phototherapy device by judging the damage level based on the early damage value of the human eye.

Images