AI-Based Personalized Ultrasonic Skincare System

Abstract

The present invention relates to a customized ultrasonic skin care system that analyzes the characteristics of each skin layer using artificial intelligence and automatically calculates and corrects treatment conditions accordingly. The system comprises: a skin measurement unit that transmits and receives ultrasound to a user's skin to measure the depth distribution and tissue characteristics of multiple skin layers, including the epidermal layer, dermal layer, and superficial muscular aponeurotic layer (SMAS layer); a user server that stores the measured skin data and user input information; an AI analysis server that determines the target skin layer and the purpose of improvement based on the data, and calculates treatment parameters including the optimal frequency, output intensity, and irradiation time within a range of 1 MHz to 20 MHz; a feedback-based reinforcement learning module that updates the treatment policy using a reinforcement learning method based on feedback data collected during the treatment; and a skin treatment device that controls and outputs ultrasound according to the treatment parameters. According to the present invention, a closed-loop skin care structure is implemented in which optimal parameters for each user are automatically updated during repeated procedures by collecting skin response data in real time during the procedure and reflecting the results as reinforcement learning reward values. In addition, by separating data access rights between the user server and the AI ​​analysis server, the AI ​​analysis server uses only anonymized skin reaction data for training, and personal information is stored in a separate secure area within the user server, thereby simultaneously ensuring the protection of personal information and the stability of medical data management. Accordingly, the present invention provides an intelligent ultrasonic skin care technology that automatically controls treatment conditions based on ultrasonic response characteristics of each skin layer and is continuously optimized through feedback learning.

Description

Technology Field

[0001] The present invention relates to a technology for quantitatively measuring skin condition and automatically calculating treatment conditions using artificial intelligence (AI). More specifically, it relates to an AI-based customized ultrasound skin care system that measures multiple layers of the skin (epidermis, dermis, and superficial muscular aponeurotic system (SMAS)) using ultrasound to analyze the depth distribution and acoustic reflection characteristics of each layer, calculates treatment parameters including treatment frequency, output intensity, and irradiation time based on the results, and performs skin care treatments optimized for each user by reflecting skin change data measured after treatment based on reinforcement learning.

[0002] More specifically, the present invention relates to a closed-loop skin care control technology that transmits ultrasonic skin data collected through a skin measurement unit to a user server and an AI analysis server, calculates a combination suitable for a target skin layer from among a plurality of candidate frequencies and output intensities using feature values ​​for each skin layer and user input information (e.g., improvement purpose, skin type, sensitivity, etc.) in the AI ​​analysis server, and provides post-procedure feedback data to a reinforcement learning module to gradually improve the procedure parameter calculation policy.

[0003] Furthermore, the present invention relates to a security structure that simultaneously secures the protection of personal information and the independence of technical processing by separating data processing authority between a user server and an AI analysis server, storing personal information containing user identification information only in a secure area within the user server, and utilizing only anonymized skin response data as training data in the AI ​​analysis server.

[0004] Therefore, the present invention belongs to the field of AI healthcare and medical device technology, which fuses artificial intelligence and ultrasonic control technology to automatically calculate and optimize treatment parameters suitable for various cosmetic and skin treatment purposes, such as improving skin elasticity, lifting, and pore reduction. Background Technology

[0005] The human skin is generally divided into the epidermis, dermis, and the superficial musculoaponeurotic system (SMAS).

[0006] Each of these layers has a different cellular structure and physical properties. The epidermal layer is responsible for defense against the external environment and pigment formation, the dermal layer contains a fibrous matrix composed of collagen and elastin to maintain the elasticity and strength of the skin, and the SMAS layer is a fibrous fascial structure located between the lower dermis and the muscle layer, which directly affects skin sagging and facial contour formation.

[0007] Due to these anatomical characteristics, for non-invasive energy procedures (e.g., HIFU, RF, IR, etc.) aimed at improving skin elasticity or lifting, it is very important to adjust the focus of energy to match the depth and density of each skin layer.

[0008] In other words, if energy is applied at a fixed depth only, the stimulation may be excessively concentrated in the epidermal layer, causing side effects such as burns, or conversely, problems may arise where sufficient thermal coagulation effects are not delivered to the dermal or SMAS layers.

[0009] High-Intensity Focused Ultrasound (HIFU) technology is a representative non-invasive procedure introduced to solve these problems, forming a Thermal Injury Zone (TIZ) at a specific depth within the dermis and SMAS layers without damaging the skin surface.

[0010] When focused ultrasound energy reaches the target tissue, the temperature of the focal area rises to about 60–70°C, causing localized coagulation, which induces collagen denaturation and remodeling.

[0011] Subsequently, fibroblasts are activated by the wound healing response, leading to the formation of neocollagen (neocollagenesis), which in turn results in improved skin elasticity and a lifting effect.

[0012] This principle of operation has been proven in various studies.

[0013] Laubach et al. (2008) reported that micro-coagulation zones can be formed in the lower dermis using focused ultrasound, and White et al. (2007) experimentally demonstrated through cadaveric tissue experiments that tissue contraction and lifting effects occur when localized thermal damage is induced in the SMAS layer.

[0014] In addition, the U.S. Food and Drug Administration (FDA), through the approval of the Ulthera™ System (K072505), has officially acknowledged that a High-Intensity Focused Ultrasound (HIFU)-based device can induce a non-surgical eyebrow lifting effect by forming thermal coagulation points in subcutaneous tissues, such as the SMAS layer, without damaging the epidermis.

[0015] This is a representative example of regulatory confirmation that ultrasonic energy can selectively exert a lifting effect depending on the depth of each skin layer.

[0016] However, most existing HIFU or RF devices operate only under single focal depth or fixed energy conditions, so they have limitations in that they cannot reflect differences in skin layer thickness or individual tissue characteristics.

[0017] For example, at the same output intensity, users with thin skin may experience redness and pain due to overheating of the epidermal layer, while users with thick skin or deep SMAS layers may experience problems where the lifting effect is limited because the energy does not reach them sufficiently.

[0018] Furthermore, existing systems that manually adjust frequency and output intensity based on the operator's experience make it difficult to ensure consistency in treatment results due to unclear quantitative standards.

[0019] Meanwhile, with the development of artificial intelligence (AI) and deep learning technologies, research is actively underway to automatically calculate customized treatment parameters tailored to the characteristics of individual patients or users by analyzing various biosignals or image data; however, technology for automatically controlling treatment parameters by quantitatively modeling the multi-layered structure of the skin and the resulting thermal response is still in the early stages.

[0020] In particular, there are rare cases where a closed-loop reinforcement learning structure is applied to systematically collect and learn skin response data after a procedure to update optimal parameters for the next procedure.

[0021] Therefore, there is a need to develop an AI-based ultrasound skin care system capable of automatically calculating appropriate frequencies and output intensities based on the depth and tissue characteristics of the multi-layered structure (epidermis-dermis-SMAS) of human skin, and providing optimized treatments for each user by analyzing and updating actual post-procedure skin response data based on reinforcement learning.

[0022] Furthermore, since such systems process biometric data containing personal information, it is essential to implement a security structure that separates data access rights between the AI ​​analysis server and the user server to utilize only anonymized technical data for training. The problem to be solved

[0023] Conventional ultrasound or radiofrequency-based skin treatment devices irradiate energy to the skin based on a single frequency and output intensity, and thus fail to adequately reflect the varying skin layer thickness or tissue characteristics of each user.

[0024] Accordingly, there was a problem where some users experienced overheating or pain in the epidermal layer, while others did not form sufficient thermal coagulation in the dermal layer or SMAS layer, which limited the expected lifting or elasticity improvement effect.

[0025] Furthermore, most existing systems rely on the operator's experience to manually adjust frequency, intensity, and irradiation time, which has limitations such as unclear quantitative standards and low consistency in treatment results.

[0026] In particular, even when using the same equipment, it was common for variations in treatment effects and side effects to occur due to individual differences such as skin age, type, thickness, and elasticity.

[0027] To overcome such limitations, the first objective of the present invention is to provide a system capable of automatically calculating an appropriate frequency, output intensity, and irradiation time according to the physical characteristics of each skin layer by analyzing the depth distribution and acoustic reflection characteristics of multiple skin layers (epidermal layer, dermal layer, SMAS layer) obtained from an ultrasonic sensor.

[0028] Second, another objective of the present invention is to implement a closed-loop customized treatment system in which the level of optimization for each user improves as the treatment is repeated by measuring changes in the skin condition after the treatment in real time, quantifying the degree of improvement in the skin's thermal response and elasticity, and utilizing this as a reinforcement learning-based reward value to progressively improve the treatment parameter calculation policy.

[0029] Third, the present invention aims to simultaneously secure data efficiency and personalization precision by clustering procedure result data collected from multiple users, configuring separate policy networks for user groups with similar skin characteristics, and applying the learning results of the corresponding cluster via a transfer learning method during the initial procedure of a new user.

[0030] Fourth, the present invention aims to implement a safe data processing structure that complies with the Personal Information Protection Act and international security regulations by separating data access rights between a user server and an AI analysis server to safely manage biometric data containing personal information, thereby allowing the AI ​​analysis server to utilize only anonymized skin reaction data.

[0031] Fifth, the present invention aims to provide an AI-based ultrasonic skin care system capable of simultaneously realizing real-time analysis of skin condition, automatic optimization of treatment conditions, policy evolution through feedback learning, and personal information protection by integrating these technical configurations.

[0032] Therefore, the present invention can provide a scientific and data-based skin care technology that is distinguished from existing simple beauty devices by learning the user's actual skin reaction data and automatically suggesting optimal treatment parameters without relying on the operator's skill or experience. means of solving the problem

[0033] A customized skin care system according to one embodiment of the present invention comprises: a skin measurement unit that transmits and receives ultrasound to a user's skin to acquire skin data including depth distribution and acoustic reflection characteristics of a plurality of skin layers, including an epidermal layer, a dermal layer, and a SMAS layer; a user server that creates and maintains a skin profile database by storing user-specific skin data, treatment history, improvement results, and user input information in a time-series format; and an AI analysis server configured to receive skin data and treatment history information from the skin measurement unit and the user server, and (i) extract feature values ​​corresponding to a target skin layer and an improvement purpose among the plurality of skin layers, (ii) calculate treatment parameters suitable for the target skin layer among a plurality of candidate frequencies of 1 MHz or more and 20 MHz or less, output intensity, and irradiation time, (iii) evaluate the treatment effect by analyzing skin change data re-received from the skin measurement unit or a separate feedback sensor after the treatment, and then accumulate and store the evaluation results in the user server, and (iv) update a treatment parameter calculation algorithm by reflecting the accumulated data based on reinforcement learning to evolvely optimize the optimal treatment parameters for each user through repeated treatments. The device is characterized by comprising: an ultrasound-based skin treatment device that receives treatment parameters calculated from the AI ​​analysis server and is controlled to selectively output ultrasound corresponding to the frequency, output intensity, and irradiation time included in the treatment parameters; and a communication module that transmits and receives data between the skin measurement unit, user server, AI analysis server, and ultrasound-based skin treatment device via wired or wireless means.

[0034] A customized skin care system according to one embodiment of the present invention is disclosed, wherein the AI ​​analysis server comprises: a step of first determining a target skin layer by comparing depth distribution and tissue characteristic data of a plurality of skin layers included in skin data received from the skin measurement unit with reference data; a step of secondarily correcting the target skin layer by reflecting improvement objectives specified by a user (pore reduction, elasticity improvement, lifting, etc.); a step of creating a treatment candidate parameter space limited to a frequency range, output intensity range, and irradiation time range corresponding to the target skin layer and improvement objectives, and configuring a set of treatment candidate parameters by applying constraints according to device safety limits and user skin type; a step of, for each of the candidate parameters, calculating a skin reaction similarity score with a similar user cluster among a skin profile database stored in a user server and an effect prediction score based on past improvement rates, and calculating an overall suitability score by weighted summing these scores; and a step of sorting the candidate parameters in descending order according to a multi-criteria ranking rule that considers both the overall suitability score and the side effect risk score, and finally selecting one or more of the top-ranked parameters as treatment parameters.

[0035] In a customized skin care system according to one embodiment of the present invention, the AI ​​analysis server receives, as real-time feedback data, reflected ultrasound signals that are reflected or transmitted from the skin by ultrasound output to the skin by the ultrasound-based skin treatment device during the treatment, or bio-response data corresponding to changes in skin surface temperature and elasticity, and if the real-time feedback data deviates from a reference threshold value, it is configured to dynamically correct treatment parameters by adjusting at least one of the currently applied frequency, output intensity, or irradiation time in steps, or by switching to a parameter among the candidate parameters ranked in claim 2 that has a higher adaptability to the real-time feedback data.

[0036] In another embodiment of the present invention, the AI ​​analysis server,

[0037] A customized skin care system is disclosed, characterized by accumulating and storing skin change data received from the skin measurement unit or a separate feedback sensor during multiple treatment sessions as time-series data in a user's skin profile database, and from the accumulated data, precisely correcting a user-specific skin response model based on at least one indicator among (i) the rate of change in skin response over time for the same treatment parameter, (ii) the pattern of improvement in effect during repeated treatment, and (iii) the response deviation from a cluster of similar users with the same skin characteristics, and by reflecting the precisely corrected skin response model, sequentially updating the weight calculation or policy selection logic of candidate parameters in an episode-by-episode manner using a reinforcement learning method when calculating treatment parameters in the future.

[0038] In a customized skin care system according to one embodiment of the present invention, the AI ​​analysis server forms a plurality of user clusters based on skin characteristics and treatment response patterns stored in a user's skin profile database, and when calculating treatment parameters, it preferentially applies a policy network or policy weight set corresponding to the cluster with the closest characteristics of the user among the user clusters, and if, after performing the treatment, the deviation of the user response data from the representative response pattern of the cluster exceeds a reference value, the policy network or policy weight set for the user is configured to be individually updated by reflecting the deviation.

[0039] In a customized skin care system according to one embodiment of the present invention, the user server stores user identification information, skin data, and treatment history data in mutually separated secure areas, and the AI ​​analysis server receives only anonymized skin reaction data when transmitting and receiving data with the user server and utilizes it for calculating treatment parameters and reinforcement learning, wherein data access rights are separated so that original user personal information is not transmitted to the AI ​​analysis server, and treatment result data is configured to be attributed only to the user server after undergoing a re-identification process when necessary.

[0040] Other specific details of this invention are included in the detailed description and drawings. Effects of the invention

[0041] The present invention can automatically optimize treatment parameters (frequency, output intensity, irradiation time) according to physiological differences such as depth, density, and reflection characteristics of the skin layers by performing artificial intelligence analysis based on ultrasound measurement results for multiple layers of the skin (epidermis, dermis, SMAS layer).

[0042] Accordingly, problems such as energy deviation, side effects, and imbalance in treatment effects that occurred in conventional operator experience-dependent methods or single-focus depth-based devices can be fundamentally resolved.

[0043] First, the present invention can quantitatively determine the structure of a user's skin layers by having an ultrasonic sensor measure the actual physical depth distribution and acoustic reflection characteristics of the skin in real time.

[0044] The AI ​​analysis server uses this data to automatically determine which layer among the epidermis, dermis, and SMAS is suitable for improvement purposes (e.g., pore reduction, elasticity improvement, lifting, etc.), and calculates the corresponding frequency and output intensity.

[0045] As a result, physiological responses such as collagen regeneration in the dermis or fascial contraction in the SMAS layer can be precisely induced without damaging the epidermis, which can be considered a technical achievement that systematically implements the same principle as the layer-by-layer thermal coagulation mechanism demonstrated in studies by Laubach (2008) and White (2007).

[0046] Second, the present invention quantitatively measures changes such as the skin's thermal response, elasticity recovery, and pore reduction after the procedure, and utilizes these as reinforcement learning-based reward values ​​to gradually improve the procedure policy.

[0047] In other words, the system does not stop at a single procedure result but continuously learns from the result data to independently identify “more effective and safer conditions for the next procedure.”

[0048] This closed-loop reinforcement learning structure forms a self-evolving mechanism in which the system's accuracy and personalization naturally improve as the procedures are repeated.

[0049] As a result, the variation in effectiveness gradually decreases with repeated treatments on the same user, and parameters optimized for the individual user's skin characteristics are automatically accumulated.

[0050] Third, the present invention analyzes multiple user data to cluster them into groups exhibiting similar skin characteristics or reaction patterns, and configures a separate policy network for each cluster.

[0051] Accordingly, when a new user is registered in the system, the pre-trained policy of the cluster to which the user belongs can be applied immediately, ensuring high accuracy even in the initial procedure.

[0052] In addition, by using individual procedure data to fine-tune the cluster policy network and, conversely, reflecting the personalized policy results back into the clusters, it is possible to implement a transfer learning and federated learning structure in which the entire system is continuously enhanced.

[0053] This configuration provides practical “group-individual simultaneous optimization technology” that can preserve individual characteristics while utilizing large amounts of user data.

[0054] Fourth, the present invention can detect temperature rise, changes in reflectance, and changes in elasticity within the dermis that occur during the procedure through a real-time feedback sensor, and can immediately adjust the output or irradiation time.

[0055] For example, if the skin surface temperature exceeds a preset threshold (e.g., Δ2℃),

[0056] The AI ​​analysis server immediately recognizes the data and automatically reduces the output intensity or delays the investigation interval.

[0057] This dynamic control realizes an active safety management function capable of immediately responding to the user's skin reaction, and significantly reduces the probability of side effects compared to existing manual control or fixed intensity methods.

[0058] Fifth, the present invention fundamentally blocks the risk of personal information exposure during the AI ​​analysis stage by separating data processing authority between the user server and the AI ​​analysis server and physically separating and managing personally identifiable information and skin reaction data using anonymized tokens.

[0059] Thus, the present invention, as an artificial intelligence healthcare device, can satisfy all major international security standards, including the Personal Information Protection Act, GDPR, and HIPAA. Furthermore, this structure provides a technical foundation that enables the simultaneous securing of stable data management and legal reliability in various operating environments, such as clinical institutions, beauty clinics, and home care devices.

[0060] Sixth, according to the present invention, even with the same device, treatment conditions automatically change according to the skin characteristics of each user, so deviations due to the operator's proficiency or subjective judgment can be minimized.

[0061] In other words, since the AI ​​analysis server automatically calculates procedure parameters and updates policies through reinforcement learning, the operator only needs to perform the role of approving or monitoring the system's recommended conditions.

[0062] This structure significantly improves the consistency, efficiency, and reproducibility of the procedure.

[0063] Finally, since the present invention can predict trends in individual users' skin changes based on long-term treatment data, it can be extended to various healthcare services such as long-term management programs, optimization of treatment cycles, and recommendations for customized cosmetics.

[0064] In other words, the present invention is not merely a simple beauty device, but can be described as a next-generation intelligent skin care platform that provides scientifically verified skin improvement effects by having AI control thermal coagulation, collagen regeneration, and fascia contraction responses in different skin layers based on actual physiological data.

[0065] In summary, the present invention provides excellent effects that can advance existing passive and experiential beauty treatment systems into active and intelligent skin care technology by combining ① accurate ultrasonic measurement of skin layer structure, ② automatic calculation of treatment parameters based on artificial intelligence, ③ policy evolution through feedback learning, ④ transfer learning by user cluster, ⑤ real-time safety control, and ⑥ a security structure based on anonymization. Brief explanation of the drawing

[0066] FIG. 1 is a system block diagram according to an embodiment of the present invention. FIG. 2 is a block diagram of an AI analysis server according to an embodiment of the present invention. FIG. 3 is a block diagram showing a feedback-based reinforcement learning loop of an artificial intelligence-based customized ultrasonic skin care system according to an embodiment of the present invention. FIG. 4 illustrates an exemplary scene in which a user measures or performs a procedure on a skin layer of a face area using a portable ultrasound treatment device, as an embodiment of an ultrasound-based skin care system according to the present invention. The drawings above are provided as examples to ensure that the concept of the present invention is sufficiently conveyed to those skilled in the art. Accordingly, the present invention is not limited to the drawings presented below and may be embodied in other forms. In addition, the same reference numbers throughout the specification represent the same components. In addition, please note that in the drawings above, specific parts have been enlarged or reduced without proportion to the scale to aid understanding. Specific details for implementing the invention

[0067] Various embodiments are now described with reference to the drawings. In this specification, various descriptions are provided to facilitate understanding of the present invention. However, it is evident that these embodiments can be practiced without such specific descriptions.

[0068] As used herein, terms such as “component,” “module,” “system,” etc. refer to computer-related entities, hardware, firmware, software, combinations of software and hardware, or executions of software. For example, a component may be, but is not limited to, a procedure executed on a processor, a processor, an object, an execution thread, a program, and / or a computer. For example, both an application executed on an electronic device and the electronic device itself may be a component. One or more components may reside within a processor and / or an execution thread. A component may be localized within a single computer. A component may be distributed among two or more computers. Additionally, these components may be executed from various computer-readable media having various data structures stored therein. Components may communicate through local and / or remote processing, for example, according to signals having one or more data packets (e.g., data from a component interacting with another component in a local system or distributed system, and / or data transmitted through signals to other systems and networks such as the Internet).

[0069] Furthermore, the term "or" is intended to mean an implicit "or" rather than an exclusive "or." That is, unless otherwise specified or evident from the context, it is intended to mean one of the natural implicit substitutions "X uses A or B." In other words, if X uses A; if X uses B; or if X uses both A and B, "X uses A or B" may apply to any of these cases. Additionally, the term "and / or" as used herein should be understood to refer to and include all possible combinations of one or more of the enumerated related items.

[0070] Additionally, the terms “comprising” and / or “comprising” should be understood to mean that such features and / or components are present. However, the terms “comprising” and / or “comprising” should be understood not to exclude the presence or addition of one or more other features, components and / or groups thereof. Furthermore, unless otherwise specified or clearly evident from the context to indicate a singular form, the singular in this specification and claims should generally be interpreted to mean “one or more.”

[0071] And, the term "at least one of A or B" should be interpreted to mean "a case including only A," "a case including only B," or "a combination of A and B."

[0072] Those skilled in the art should recognize that the various exemplary logical blocks, configurations, modules, circuits, means, logics, and algorithmic steps described in connection with the embodiments disclosed herein may be implemented in electronic hardware, computer software, or a combination of both. To clearly exemplify the interchangeability of hardware and software, the various exemplary components, blocks, configurations, means, logics, modules, circuits, and steps have generally been described above in terms of their functionality. Whether such functionality is implemented in hardware or software depends on the specific application and design constraints imposed on the overall system. Skilled technicians may implement the described functionality in various ways for each specific application. However, such decisions regarding implementation should not be construed as going beyond the scope of this publication.

[0073] The description of the presented embodiments is provided to enable those skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein may be applied to other embodiments without departing from the scope of the present invention. Thus, the present invention is not limited to the embodiments presented herein. The present invention should be interpreted in the broadest possible scope consistent with the principles and novel features presented herein.

[0074] In this institution, network functions, artificial neural networks, and neural networks can be used interchangeably.

[0075] The various embodiments described herein may be implemented, for example, in recording media and storage media readable by a computer or similar device using software, hardware, or a combination thereof.

[0076] According to hardware implementation, the embodiments described herein may be implemented using at least one of ASICs (application specific integrated circuits), DSPs (digital signal processors), DSPDs (digital signal processing devices), PLDs (programmable logic devices), FPGAs (field programmable gate arrays), processors, controllers, microcontrollers, microprocessors, and other electrical units for performing functions. In some cases, the embodiments described herein may be implemented as the processor itself of an electronic device.

[0077] FIG. 1 is a system block diagram according to an embodiment of the present invention.

[0078] As illustrated in FIG. 1, the customized skin care system of the present invention includes a skin measurement unit (110), a user server (120), an AI analysis server (130), an ultrasound-based skin treatment device (14), and a communication module (150).

[0079] FIG. 2 is a block diagram of an AI analysis server according to an embodiment of the present invention.

[0080] The configuration of the AI ​​analysis server shown in Fig. 2 is merely a simplified example.

[0081] In one embodiment of the present invention, the AI ​​analysis server (130) may include other components for performing a computing environment, and only some of the disclosed components may constitute the AI ​​analysis server (130). The AI ​​analysis server (130) may include a processor (160) and memory (170).

[0082] The processor (160) may be composed of one or more cores and may include a processor (160) for data analysis and deep learning, such as a central processing unit (CPU), a general-purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU) of the AI ​​analysis server (130). The processor (160) may read a computer program stored in memory (170) and perform data processing for machine learning according to one embodiment of the present invention.

[0083] Additionally, the processor (160) controls the configuration of the AI ​​analysis server (130) to operate and can implement the overall operation of the system.

[0084] For example, the processor (160) can typically control the overall operation of the AI ​​analysis server (130). The processor (160) can provide or process appropriate information or functions to the user by processing signals, data, information, etc. that are input or output through the components described above, or by running applications stored in memory (170).

[0085] Additionally, the processor (160) can control at least some of the components of the AI ​​analysis server (130) to run an application stored in memory (170). Furthermore, the processor (160) can operate at least two or more of the components included in the AI ​​analysis server (130) in combination with each other to run the application.

[0086] According to one embodiment of the present invention, the processor (160) can perform operations for learning a neural network. The processor (160) can perform operations for learning a neural network, such as processing input data for learning in deep learning (DL), extracting features from input data, calculating errors, and updating the weights of the neural network using backpropagation. At least one of the CPU, GPGPU, and TPU of the processor (160) can process the learning of the network function. For example, the CPU and GPGPU can together process the learning of the network function and data classification using the network function.

[0087] According to one embodiment of the present invention, memory (170) is any form of information generated or determined by the processor (160) and any form received by the network unit Information can be stored. According to one embodiment of the present invention, the memory (170) may include at least one type of storage medium among a flash memory (170) type, a hard disk type, a multimedia card microtype, a card type memory (170) (e.g., SD or XD memory (170)), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory (170), a magnetic disk, and an optical disk. AI analysis The server (130) may operate in connection with web storage that performs the storage function of the memory (170) on the internet. The description of the memory (170) described above is merely an example and is not limited thereto.

[0088] A communication module (150) that transmits wireless or wired signals to another device can communicate with an external device, and in particular, the communication module (150) may include various communication chips such as a Wi-Fi chip, a Bluetooth chip, a wireless communication chip, an NFC chip, a low-power Bluetooth chip (BLE chip), etc., or may be composed of a communication circuit that performs such functions.

[0089] FIG. 3 is a block diagram showing a feedback-based reinforcement learning loop of an artificial intelligence-based customized ultrasound skin care system according to an embodiment of the present invention, and FIG. 4 is an exemplary illustration of a scene in which a user measures or performs a procedure on a skin layer of a face area using a portable ultrasound treatment device as an embodiment of an ultrasound-based skin care system according to the present invention.

[0090] The present invention relates to a customized skin care system in which ultrasound-based skin measurement, AI analysis, procedure control, and feedback learning form a closed loop.

[0091] A customized skin care system according to one embodiment of the present invention utilizes skin data including the depth distribution and acoustic reflection characteristics of multiple skin layers (epidermis, dermis, SMAS) obtained from a user's skin to calculate treatment parameters (frequency, output intensity, irradiation time) suitable for the target skin layer and improvement purpose (e.g., pore reduction, elasticity improvement, lifting), and analyzes post-treatment skin change data to update the calculation algorithm using a reinforcement learning method, thereby evolutionarily deriving optimal parameters for each user through repeated treatments.

[0092] The system (100) according to the present invention can be understood as follows without drawings.

[0093] The skin measuring unit (110) includes an ultrasonic transducer array (including a replaceable or variable type in the range of 1 to 20 MHz), a driving unit, a receiving and amplifying unit, an A / D converter, and a signal processing unit.

[0094] The skin measuring unit transmits and receives ultrasound to the user's skin to acquire reflection signals corresponding to each layer of the epidermis, dermis, and SMAS, and calculates acoustic reflection characteristics such as time delay, amplitude, and frequency spectrum.

[0095] The signal processing unit performs A-scan / B-mode-based time-of-flight (TOF) analysis, band-pass filtering, moving average / adaptive noise suppression, and spectrum estimation to output skin data (depth distribution + acoustic reflection characteristics) as a digital signal.

[0096] The signal processing unit according to the present invention quantifies the depth of the skin layer, reflection characteristics, frequency response characteristics, etc., based on the ultrasonic RF reception signal.

[0097] The received RF signal is first passed through a filter matching the probe's center frequency band to remove noise in non-corresponding bands, and then processed through envelope detection to clearly reveal the flow of signal changes on the time axis.

[0098] At this time, the attenuation phenomenon occurring differently depending on the frequency as it passes through skin tissue is corrected using an inverse estimation method based on a pre-established tissue-specific attenuation model, thereby standardizing it to enable a comparison of relative reflection intensities under the same conditions.

[0099] Once the envelope signal is secured, the point of reflection is searched based on the point where the peak or slope change of the reflected signal appears most distinct, and this point of reflection is converted into the physical distance traveled by the ultrasound through the time difference from the transmission point.

[0100] Subsequently, based on the structural characteristics of the skin, the combination with the highest signal quality and structural validity among multiple reflection candidates is determined as the layer boundary based on whether the order of the epidermis, dermis, and SMAS is maintained. The boundary location derived in this way and the relative distance information between the segments are quantified as the depth and thickness of each layer.

[0101] To analyze reflection intensity by layer boundary, the reflection intensity of the envelope extracted at each boundary point is compared with a reference reflection intensity obtained in advance using a phantom and quantified in the form of a relative reflection coefficient. In this process, surrounding area signals are referenced together to stabilize the reflection signal so that it is not over- or under-reflected by transient noise.

[0102] To more precisely reflect the tissue characteristics of each layer, the RF signal is divided into specific depth intervals into lengths and converted into the frequency domain to analyze the spectral characteristics.

[0103] At this point, the center value of the tissue's frequency response is derived based on the band with the highest concentration of energy in the spectral distribution, and the diffusivity corresponding to the bandwidth is evaluated through the degree to which the signal spreads relative to the center.

[0104] The tendency of reflection intensity to decrease as frequency increases is quantified in the form of the spectral slope, and the degree of non-uniformity or irregularity in the signal distribution is expressed by the spectral disorder index (entropy).

[0105] If necessary, stability can be ensured so as not to be sensitive to noise by storing data repeatedly measured in the same section or replacing it with statistically based representative values.

[0106] To assess the quality and reliability of the entire signal, the signal-to-noise ratio is calculated, and a boundary clarity index is derived based on how distinctly the change before and after reflection appears near the layer boundary.

[0107] These reliability and clarity indicators are reflected in the form of weights when interpreting previously calculated depth, reflection, and spectrum-related indicators, adjusting to ensure that indicators with low signal quality do not have an excessive influence in subsequent analyses.

[0108] Through this process, the skin condition of each user is converted into a feature vector composed of various elements such as layer depth, layer-specific reflectance characteristics, spectral center band, spectral diffusion, frequency attenuation tendency, spectral disorder, signal-to-noise ratio, and boundary clarity.

[0109] This feature vector is statistically normalized based on the data distribution of past similar groups and input into an AI analysis server in its normalized form, where it is directly utilized in the processes of target layer identification, parameter candidate generation, candidate group evaluation, and procedure parameter optimization.

[0110] The user server (120) can be implemented on a user terminal or in the cloud and creates and maintains a skin profile database (SPD) by storing user-specific skin data, treatment history, improvement results, and user input information (e.g., sensitive skin, contraindications, preferences) in a time-series format.

[0111] Personal information is stored in a secure area and can be linked based on anonymized tokens when linked with an AI analysis server (detailed security may be further extended in a separate clause, but structural separation is valid even at the stage of Clause 1).

[0112] The AI ​​analysis server organizes layer-by-layer depth information, reflection characteristics, and spectrum-based tissue indicators transmitted from the skin measurement unit into a specific feature vector structure through the procedure parameter calculation module. Based on this feature vector, it first determines the target skin layer and finally calculates procedure parameters consisting of ultrasound frequency, output intensity, and irradiation time.

[0113] In this case, the procedure parameters function not as simple cosmetic values, but as control signals converted into actual device control inputs, and are automatically determined according to the following step-by-step procedure.

[0114] First, the analysis server uses the layer-by-layer reflection coefficient pattern, center band of the spectral response, tissue absorption tendency, and layer thickness information included in the feature vector as comparison criteria, and analyzes the proximity to a predefined layer-by-layer reference table to estimate the primary target layer of the procedure.

[0115] Subsequently, the primary target layer may be adjusted toward a shallower or deeper layer depending on the improvement objective entered by the user (e.g., improvement of elasticity, reduction of pores, lifting, etc.), and in this process, information on skin type or contraindications is also taken into account to ensure that the target layer is not set in a direction that could cause excessive penetration.

[0116] Once the target layer is determined, a frequency range known to be suitable for that layer, a output intensity range based on past history where tissue responsiveness was good, and an appropriate irradiation time range are searched through a reference table within the system, and a set of multiple parameter candidates is generated by subdividing this range into regular intervals.

[0117] After parameter candidates are generated, the suitability of tissue penetration depth for each candidate, estimated temperature rise results, and average effect cases from past user groups are compared and evaluated to analyze the expected efficacy and risk associated with each candidate.

[0118] At this time, the allowable output upper limit, energy density limit, skin overheating threshold condition, and attenuation condition for users with sensitive skin recorded in the safety constraint table are automatically applied, and candidates that violate the criteria or fall under the risk level are immediately excluded from the search.

[0119] For candidates that pass the constraints, the level of effect prediction is determined based on the similarity between the current user response characteristics estimated from the feature vector and the response patterns of similar user groups accumulated in the skin profile database.

[0120] Next, the improvement rate and the frequency of side effects occurring at the same candidate parameter among past treatment cases are considered together, and each candidate is organized into an evaluation score that simultaneously reflects improvement efficiency and the possibility of risk occurrence.

[0121] In this evaluation process, internal system weights are adjusted based on whether the user's objective is effect-oriented or stability-oriented, and as a result, candidates are sorted in order of highest priority starting from those that achieve the most appropriate balance of effect versus risk.

[0122] In this way, one or more combinations from the relevant top candidate groups are finally selected as the procedure parameters, and the selected parameter values ​​are automatically converted into frequency, output intensity, and irradiation time that the ultrasound-based procedure device can actually operate in, and transmitted to the device.

[0123] Consequently, the procedure parameter calculation module of the present invention is distinguished from existing simple recommendation-type systems in that it does not merely provide a simple skin condition classification function, but rather implements a technical processing process that derives final parameters at the device control level after performing multi-stage constraint filtering and effect / risk evaluation starting from actual measurement-based statistics of body tissues.

[0124] Then, by referring to user input (for improvement purposes) and SPD, treatment parameters suitable for the target skin layer are calculated among multiple candidate frequencies, output intensity, and irradiation time within the range of 1 to 20 MHz.

[0125] After the procedure, skin change data received again is analyzed to calculate effect indicators such as improvement rate, change in elasticity, and change in pores, and risk indicators such as excessive temperature rise and pain indicators (procedure effect evaluation).

[0126] By interpreting the above evaluation results as a reward signal and updating the parameter calculation policy, prediction and recommendation performance for identical and similar users is improved. In summary, the policy is gradually updated in a direction that 'increases effectiveness and lowers risk.'

[0127] The ultrasound-based skin treatment device (140) includes a drive controller, an output amplifier, a transducer head, and a cooling / temperature monitoring auxiliary unit (optional).

[0128] Energy focusing corresponding to the target skin layer is implemented in accordance with the treatment parameters (frequency, output, irradiation time) received from the AI ​​analysis server.

[0129] It can be integrated with feedback sensors (built-in or external) to measure temperature, reflection signals, elasticity changes, etc., in real-time or retrospectively during and after the procedure.

[0130] The communication module (150) exchanges data between each component through a wired and wireless communication network.

[0131] For example, BLE (short-range between the operator and the measurement unit) and Wi-Fi / Ethernet (analysis server to user server) can be used, and encryption and integrity verification are performed during data transmission.

[0133] The system of the present invention operates in the following order.

[0134] (1) Pre-setting step

[0135] Register the user profile (skin sensitivity, contraindications, past history), improvement purpose (pores / elasticity / lifting, etc.), and procedure history structure (date, area, parameters, response) on the user server.

[0136] The skin measurement unit performs calibration (transducer sensitivity correction, standard phantom-based depth scale correction).

[0137] (2) Measurement step

[0138] The skin measurement unit performs a multi-point scan on the measurement target area and calculates the depth distribution and acoustic reflection characteristics for each point.

[0139] The results are packaged as digital skin data and transmitted to the AI ​​analysis server and the user server.

[0140] (3) Analysis and calculation stage

[0141] The procedure parameter calculation module of the AI ​​analysis server reconstructs the skin condition into quantified feature values ​​based on layer boundary locations, layer thickness, spectral change patterns, and reflection response characteristics transmitted from the skin measurement unit.

[0142] Subsequently, the target group and treatment direction for the current procedure are determined first by comparing the corresponding feature values ​​with the user history and improvement target candidates stored in the skin profile database.

[0143] Once the target layer is determined, a frequency band suitable for that layer, an output intensity range in which a stable tissue response was observed, and an irradiation time range that was clinically effective in past cases are searched through an internal reference table, and these are combined according to a set interval or step condition to generate multiple candidate treatment parameters.

[0144] Safety review is immediately applied to the candidate group generated at this time.

[0145] In other words, peak power limits, allowable energy density per tissue unit, tissue temperature rise threshold conditions, and output attenuation conditions for users with sensitive skin are simultaneously reflected, and combinations that deviate from the standard values ​​are automatically excluded.

[0146] For the candidates filtered in this way, the response trends of the current user and similar users within the skin profile database are compared to predict the level of expected effect, and risk factors such as the possibility of side effects are evaluated.

[0147] In this process, each candidate is evaluated in a manner where the score increases as the effect is greater and decreases as the risk increases, and finally, the combination with the most appropriate balance of effect versus risk is selected as the procedure parameter.

[0148] This parameter calculation process is not a simple sensory selection method based on user manuals, but is performed as a technical process that converts into physical output values ​​suitable for device control commands through feature analysis based on measured biosignal values ​​and comparative evaluation based on past response data.

[0149] (4) Procedure step

[0150] Ultrasound-based skin treatment devices are driven at the corresponding frequency according to received parameters and deliver energy during the irradiation time. For example, contact cooling or pulse duty cycle control can be performed in parallel for epidermal protection.

[0151] (5) Feedback and learning stage

[0152] Immediately after the procedure or after a certain period of time, skin change data (e.g., elasticity index, pore index, change in reflection signal, surface temperature) is recollected through a skin measurement unit or feedback sensor.

[0153] The AI ​​analysis server evaluates re-collected data to calculate treatment effectiveness indicators (improvement rate) and risk indicators (estimation of overheating and pain).

[0154] These evaluations are interpreted as reward signals, and the reinforcement learning module (132) updates the procedure parameter calculation policy.

[0155] The update results are reflected in the SPD, enabling more precise personalization during the next procedure.

[0156] In the case of ultrasound, frequencies of 3-10 MHz (center of the epidermis), 10-20 MHz (epidermis / superficial dermis), and 1-3 MHz (deep penetration into the SMAS) are used, and the output intensity is managed with upper limits of peak output / energy density in a pre-table in accordance with device safety standards, and pulse / burst mode options may be used for the irradiation time, and the maximum continuous irradiation time may be limited depending on the skin type and area.

[0157] Signal processing can suppress noise using adaptive filters, moving averages, and simple median filters for speckle reduction, and boundary candidates can be calculated by combining the peak positions of reflected signals and changes in reflection coefficients, and then refined using physical consistency constraints (e.g., epidermal-dermal-SMAS order) to detect layer boundaries.

[0158] In managing mechanical indicators, the system maintains a temperature rise estimation model and an output limit table. When a feedback sensor detects a real-time temperature change and exceeds a threshold (e.g., baseline skin temperature + Δ2°C), the device performs irradiation suspension or stepwise output attenuation. These constraints ensure both industrial applicability and safety.

[0159] Once the procedure is completed, skin changes immediately after the procedure and after a certain period of time are measured again through a skin measurement unit or feedback sensor. Positive improvement effects, such as the degree of elasticity enhancement, pore reduction, and lifting improvement, and negative reactions, such as the level of temperature rise during the procedure, user discomfort or pain response, and the persistence of erythema after the procedure, are evaluated as separate items.

[0160] These results are grouped into effect and risk categories; for effect categories, a higher value indicates a positive outcome, while for risk categories, a higher value is considered to indicate a negative outcome.

[0161] The AI ​​analysis server determines the relative improvement or risk level by comparing each item to a preset standard range, and a positive evaluation is made as effective items increase above the standard level and risk items remain within the acceptable range.

[0162] Conversely, a negative evaluation is assigned if the improvement effect is negligible or risk items exceed the allowable limit. This evaluation result is integrated into a single reward value, which is reflected in such a way that the reward value increases as the improvement effect is greater and decreases as the risk increases.

[0163] In this case, the reward value is restricted to change gradually within a certain range to prevent it from being excessively influenced by short-term reaction fluctuations, and if the same trend is repeatedly confirmed, a progressively greater weight is assigned and accumulated to exert a high influence on policy updates.

[0164] This compensation structure is designed to prevent specific parameter combinations from being prioritized simply because of their high effectiveness, and to ensure that only combinations capable of maintaining a balance between effectiveness and safety above a certain level can grow into policies of high reliability in the long term.

[0165] The reinforcement learning module recognizes the parameter combination applied during the previous procedure as a single action, interprets the skin condition input at the time of the procedure as the state in which the action occurred, and adjusts the fit between the state and the action based on the reward value.

[0166] Parameter combinations that are repeatedly granted positive rewards have a gradually increasing probability of being selected again under identical or similar skin conditions, while combinations that repeatedly incur risks have their selection priority automatically lowered. Through this process, the system progressively optimizes its policy by reinforcing effective combinations and excluding risky ones, based on repeated treatment experience.

[0167] Example 1: Pore improvement (nose area)

[0168] In this embodiment, a procedure was performed on a user in their late 20s with oily skin for the purpose of improving pores in the nose area.

[0169] In the initial measurement phase, the skin measurement area had a relatively thin epidermal thickness, and an acoustic reflection peak was predominantly observed near the boundary between the epidermis and the shallow dermis layer.

[0170] Based on this data, the AI ​​analysis server determined that a low-to-medium output, short burst mode in the frequency band of 12 to 18 MHz would be suitable, and calculated a combination of 15 MHz, medium output, 0.8 seconds × 3 times, and an irradiation interval of 2 mm as the optimal condition among the candidate parameters.

[0171] The ultrasound-based treatment device was operated according to the above parameters, and as a result of re-measurement immediately after the procedure, an initial response was observed in which the scattering pattern of the reflected signal decreased and the pore structure contracted. One week after the procedure, the pore index re-measured by the skin measurement unit decreased by approximately 18%, and no side effects such as heat damage or erythema were confirmed.

[0172] The feedback-based reinforcement learning module of the AI ​​analysis server evaluated the result as a positive reward signal, and accordingly, the policy was updated so that weights in the range of 15–17 MHz were increased for user clusters with the same skin characteristics and treatment purpose.

[0173] This resulted in the same parameter combination being recommended to similar users with a higher probability in the future.

[0175] Example 2: Lifting (SMAS target, cheekbone line)

[0176] This example was performed on subjects in their early 40s with combination skin for the purpose of cheekbone line lifting.

[0177] Initial measurement results confirmed that the depth distribution of the skin measurement area up to the SMAS layer boundary was uniform, and that the deep reflection component relative to the dermal layer reflection was maintained uniformly.

[0178] The AI ​​analysis server analyzed this data and set the SMAS layer as the target layer, and determined the medium-to-high power pulse mode in the 1.5~3 MHz band as a candidate region.

[0179] To minimize thermal stimulation of the treatment site, the use of a cooling device was recommended, and finally, the procedure was performed under conditions of 2 MHz, medium power, and pulses of 100 ms × 10 times.

[0180] According to real-time data from the feedback sensor, the temperature rise of the skin surface immediately after the procedure was limited to Δ1.2°C, allowing the procedure to be performed stably within the safety constraint range.

[0181] Lifting indicators measured 4 weeks after the procedure improved by approximately 23%, and users did not complain of side effects such as pain or overheating.

[0182] Based on the procedure data, the AI ​​analysis server evaluated that the efficacy indicators were high and the risk indicators were low, and promoted the policy priority of the corresponding parameter combination to a higher grade.

[0183] Accordingly, in the policy network of the same cluster (SMAS-centered lifting type user group), output parameters in the 2MHz range were reflected as key reference values.

[0184] Example 3: Improvement of elasticity (cheeks and jawline)

[0185] This embodiment was performed on a user in their late 30s with a history of sensitive skin for the purpose of improving elasticity in the cheek and jawline areas.

[0186] Based on the information indicating 'sensitive skin' in the user profile, the AI ​​analysis server automatically applied a conservative output upper limit table, shortened the investigation time, and strengthened constraints to select burst mode instead of continuous investigation.

[0187] Accordingly, the procedure parameter calculation module searched for a candidate group within a lower output range than before, and the procedure proceeded while receiving real-time feedback from the skin measurement unit and the feedback sensor.

[0188] Immediately after the procedure, the sensor detected a temperature rise of Δ2.3°C, and the AI ​​analysis server immediately determined this to be a signal exceeding the threshold value, automatically reduced the output by 15%, and delayed the next irradiation cycle for a certain period of time.

[0189] As a result, no localized overheating or pain occurred in the skin after the procedure, and the elasticity index measured after 2 weeks improved by about 12%.

[0190] Based on these results, the AI ​​analysis server evaluated that the “attenuation-delay” strategy maintained stable effectiveness while minimizing risk.

[0191] Accordingly, the reinforcement learning module adopted the strategy as a safety-prioritized policy, and the algorithm was reflected as the default behavioral pattern within the policy network of the sensitive skin user group.

[0192] The above embodiments specifically demonstrate the process by which the system of the present invention ① accurately analyzes skin characteristics by layer to calculate optimal ultrasound parameters, ② evaluates treatment results in real time to perform feedback learning, and ③ evolves cluster- and individual policies through accumulated data.

[0193] In addition, the treatment effects confirmed in each embodiment (18% reduction in pores, 23% improvement in lifting, and 12% improvement in elasticity) are all reflected in the policy updates of the AI ​​analysis server, allowing the system to learn on its own so that it can provide customized treatment parameters with improved accuracy and safety to the same or similar users in the future.

[0194] The ultrasonic head can be selected from linear or curved arrays, and coupling efficiency can be increased with a contact gel or membrane interface.

[0195] The communication module supports flexible protocols (BLE / Wi-Fi / Ethernet) tailored to hospital network or home router environments.

[0196] The user server can be implemented on-premises or in the cloud, and it is linked with the AI ​​analysis server using a de-identification token to perform analysis without transmitting personal information.

[0197] In a customized skin care system according to one embodiment of the present invention, the AI ​​analysis server (130) takes skin data received from the skin measurement unit (110) and the skin profile database (SPD) stored in the user server (120) as inputs and calculates treatment parameters (frequency, output intensity, irradiation time) optimized for the user's skin condition and treatment purpose.

[0198] This function is performed by a procedure parameter calculation module (131), and the entire algorithm consists of a ‘first determination and second correction step of the target skin layer’, a ‘parameter search space limit and constraint application step’, a ‘candidate parameter evaluation and similarity-based score calculation step’, and a ‘multi-criteria ranking and final selection step’.

[0199] The primary determination and secondary correction stages of the target skin layer are divided into the primary determination stage and the secondary correction stage.

[0200] The AI ​​analysis server compares the skin layer depth distribution and tissue acoustic reflection characteristics included in the input skin data with the layer-specific reference tables (epidermis, dermis, SMAS) to primarily determine the target skin layer where the focus will be formed.

[0201] The floor-by-floor standard table can be configured as shown in the following example.

[0202] skin layer Average depth (mm) Acoustic reflection characteristics Frequency response band Key organizational structure epidermis ０．１～０．５ High reflection coefficient 10–20 MHz Keratin, melanin layer dermis １．０～３．０ Intermediate reflection coefficient 3~10 MHz Collagen, Elastin SMAS layer 4.0 or higher Low reflection coefficient, wide band 1~3 MHz fascia, fibrous tissue

[0203] The target layer is determined by searching for the item whose skin layer depth and reflection coefficient patterns are closest to the above criteria.

[0204] Subsequently, the improvement objective specified by the user (e.g., pore reduction, elasticity improvement, lifting, etc.) acts as a correction factor in determining the target layer. For example, for the purpose of pore reduction → secondary correction is applied mainly to the epidermis and shallow dermis layer, for elasticity improvement → middle and lower dermis layer, and for lifting → SMAS layer.

[0205] In this process, the AI ​​analysis server analyzes the user's skin type (dry, sensitive, oily) and historical data in parallel to automatically calculate the depth of focus correction value (Δd) for the target layer. Δd is adjusted, for example, within a range of ±0.2 to 0.5 mm, and is a value used to control the treatment focus so that it is accurately formed on the target tissue.

[0207] Once the target layer and improvement objective are determined, the AI ​​analysis server generates a three-dimensional parameter space for frequency, output intensity, and irradiation time. For example, if the dermis layer is targeted, the frequency candidates are set to 3–10 MHz, and if the SMAS layer is targeted, they are set to 1–3 MHz.

[0208] At this time, the following set of constraints is applied depending on the device specifications and user characteristics.

[0209] division Limitation criteria Example values safety limits output of power 2049830W / 20499, energy density 204981.2J / 20501 Refer to the hardware limit table Skin type correction For sensitive skin, output is reduced by 15% and irradiation time is shortened by 10%. User input-based Forbidden information Exclude the wound area and implant area. User DB reference

[0210] The reference values ​​listed in the table above are not used as simple reference data; instead, they are converted into standard levels for quantitative comparison of each item and then scored.

[0211] For example, the closer the target layer is to the epidermis, the higher the score is assigned to the high frequency range, and the closer it is to the dermis or SMAS layer, the higher the score is assigned to the mid-to-low frequency range.

[0212] In addition, output limits, temperature thresholds, and contraindication conditions listed in the constraint table are reflected as deduction factors in the goodness-of-fit score of the corresponding parameter combination, or act as automatic exclusion criteria if the threshold is exceeded.

[0213] The scores transformed in this way are weighted by parameter domain and accumulated to be converted into a single score representing the suitability of each candidate combination, which is then reflected in the final evaluation along with similarity-based and risk-based scores in subsequent stages.

[0214] As a result of applying constraints, disallowed parameter combinations are automatically eliminated, leaving only the Candidate Set of procedure parameters. This set limits the AI ​​model's search range to a realistic and safe area.

[0215] Next is the candidate parameter evaluation and similarity-based scoring step, where the AI ​​analysis server calculates the following two types of evaluation scores for each combination of candidate procedure parameters:

[0216] The first similarity score (S₁) compares the current user profile with the average skin response data of similar user clusters within the SPD.

[0217] The AI ​​analysis server evaluates the similarity of the current user's skin response pattern based on the treatment results of similar users recorded in the Skin Profile Database (SPD), and utilizes this value as the first evaluation criterion.

[0218] This similarity evaluation functions as an indicator to estimate the likelihood that a specific parameter combination is effective for current users, based on how users with identical or similar skin layer structures and improvement goals in the past responded to that combination.

[0219] In addition, for the same candidate combination, the level of improvement predictable by referencing quantitative effects measured in past procedures, such as pore reduction rate, elasticity improvement, and lifting enhancement, is evaluated as a separate item. In this evaluation, combinations with superior past average effects are reflected with higher scores, and this constitutes the second evaluation item.

[0220] The relative importance of the calculated similarity and predictive effect items is adjusted according to the type of treatment purpose and the characteristics of the current user profile, and items with higher importance have a greater influence on the final judgment.

[0221] For example, combinations that have consistently shown stable improvement among the same cluster of users are evaluated with relatively high reliability, whereas combinations with high predictive effects but few similar cases may be reflected at a conservative level initially.

[0222] The AI ​​analysis server integrates the results of the two items to stepwise compare and evaluate the overall fit of each candidate parameter combination, and classifies combinations with an integrated fit exceeding a certain standard as top-tier candidates.

[0223] Through this, a comprehensive judgment is made that balances similarity-based stability and effect-based expectations, rather than relying on simple average values.

[0224] The final stage is the multi-criteria ranking and final selection stage.

[0225] For each treatment candidate combination, the AI ​​analysis server separately calculates a risk score that reflects the likelihood of side effects, such as thermal stimulation, excessive output, and the possibility of pain, in addition to the overall suitability score.

[0226] This risk score is quantified by classifying potential safety events during the procedure into categories such as temperature changes, output levels, and user discomfort, and then combining this with the frequency of adverse events from similar cases recorded in the historical Skin Profile Database (SPD).

[0227] In this case, parameter combinations that maintain a consistently low risk level are evaluated as having high stability, while conversely, combinations where temperature increases are frequent or user discomfort is reported to be high are considered to have high risk.

[0228] The final evaluation is performed by integrating the effectiveness-based and risk-based scores, and combinations with high effectiveness and below-acceptable risk are classified as high priority. Conversely, combinations that exceed risk thresholds, even if they offer somewhat higher effectiveness, are automatically excluded from consideration or classified as lower grades during the ranking comparison stage.

[0229] This integrated score is calculated in a manner that reflects the balance between effectiveness and risk, and weighted adjustments are made to relatively amplify the impact ratio of risk items when safety is prioritized.

[0230] The AI ​​analysis server selects one or more combinations with the best integrated evaluation results, and the selection criteria are based on whether the integrated score is above a preset threshold and by performing progressive ranking comparisons within the same grade.

[0231] This multi-criteria scoring method is intended to prevent the excessive selection of combinations that are merely highly effective, and to ensure both safety and efficacy simultaneously during actual clinical application.

[0232] Therefore, the procedure parameter calculation structure of the present invention is not a simple guideline-based selection, but rather corresponds to a technical decision-making process based on the balance of effectiveness and safety.

[0233] Unlike the existing simple “manual intensity adjustment by skin type,” this module determines treatment parameters based on multi-layered judgment by combining multiple physical and statistical factors, such as layer depth information, improvement objectives, similarity, and risk.

[0234] It minimizes the risk of skin damage based on data by automatically limiting excessive output or irradiation time through constraint sets and risk score evaluations.

[0235] By referencing similar user response data accumulated in the Skin Profile Database (SPD), it is possible to present pre-learned safe combinations to new users, that is, to recommend a personalized optimized structure.

[0236] The AI ​​analysis server is integrated with a feedback-based reinforcement learning module, enabling continuous accuracy improvement and enhanced personalization by immediately reflecting feedback data accumulated after the procedure.

[0237] As a result of measuring the user's face using the skin measuring unit, the average thickness of the dermis layer was found to be approximately 2.3 mm, and the acoustic reflection coefficient was approximately 0.42.

[0238] The AI ​​analysis server recognized from the input user profile that the user's improvement goal was 'improvement of elasticity' and determined the dermis layer to be the target skin layer. Accordingly, multiple treatment candidate parameters were generated centered on the frequency band of 3 to 10 MHz.

[0239] Since the user's skin type was classified as sensitive, the output intensity was automatically limited to 25 W / cm² or less, and a record was referenced showing an average elasticity improvement rate of 17% observed in a user group of the same age group (40s) and similar skin type within the SPD.

[0240] The AI ​​analysis server calculated the skin reaction pattern similarity between this user and the corresponding cluster as 0.86, and the effect prediction score was evaluated as 0.74. The overall goodness of fit score (S-=0.8) was calculated as the weighted sum of the two indicators, and the final evaluation indicator F was calculated as 0.725 by considering the side effect risk score (R-=0.15).

[0241] Accordingly, among the top three candidates, the 7 MHz, medium output, and 0.8-second irradiation condition recorded the highest ranking, and the procedure parameters were transmitted to an ultrasound-based procedure device to perform the procedure.

[0242] Skin elasticity indicators measured immediately after the procedure improved by approximately 13%, and results showing an 18% improvement were confirmed after 2 weeks.

[0243] In other subjects, the boundary of the SMAS layer was detected at an average depth of 4.8 mm from the skin measurement site.

[0244] The improvement goal set by the user was 'lifting,' and the AI ​​analysis server determined the target skin layer to be the SMAS layer. At this time, a frequency range of 1–3 MHz was designated as a candidate area, and as a result of analyzing the treatment history of a similar user cluster (women in their 30s and 40s with combination skin) stored in the SPD, an average lifting improvement effect of more than 20% was reported under conditions of 2 MHz frequency, 80% output intensity, and 2.0 seconds of irradiation.

[0245] Based on this data, the AI ​​analysis server evaluated the similarity and effect prediction scores of the candidate parameters, and selected them as the final procedure parameters after determining that the aforementioned conditions were superior in terms of both overall suitability score and safety.

[0246] Subsequently, the ultrasound-based procedure device operated according to the parameters, and results were confirmed showing an improvement of approximately 22% in the lifting index at 4 weeks post-procedure.

[0247] These embodiments demonstrate that the procedure parameter calculation module of the present invention can calculate frequency, output, and irradiation time optimized for each skin layer and improvement purpose based on actual user data, and prove that the accuracy of the skin profile database improves and personalized treatment effects are continuously improved through repeated procedures.

[0248] Accordingly, the procedure parameter calculation module of the present invention provides a technical optimization system that is AI-based and ensures physical and medical safety by ① quantifying skin layer depth and acoustic characteristics to identify the target layer, ② configuring a search space that reflects improvement objectives and safety constraints, and ③ selecting parameters by weighted combination of similarity, effect prediction, and risk scores in a multi-criteria ranking.

[0249] This structure is distinguished from simple cosmetic recommendation type BM and clearly has the character of a technical invention that automatically calculates actual device control parameters.

[0251] In a customized skin care system according to one embodiment of the present invention, the feedback-based reinforcement learning module (132) of the present invention is configured to analyze skin change data collected from a skin measurement unit or a separately installed feedback sensor after an ultrasound-based skin procedure is performed, quantitatively evaluate the effectiveness and safety of the procedure, convert the result into a compensation value, and gradually improve the procedure parameter calculation policy.

[0252] Once the procedure is completed, the skin measurement unit transmits and receives ultrasound to the treatment area again and acquires quantitative data such as changes in skin layer depth, changes in elasticity, temperature increase, pore size fluctuations, and changes in skin surface reflectance by comparing the reflection signals before and after the procedure.

[0253] In some embodiments, a feedback sensor may also measure the skin surface temperature or micro-vibration response in real time to collect risk data corresponding to overheating or pain during the procedure.

[0254] Based on the collected data, the feedback-based reinforcement learning module of the AI ​​analysis server first calculates a procedure performance score.

[0255] The evaluation criteria are broadly divided into two categories.

[0256] First, as an effect indicator, it includes positive outcome variables such as the rate of change in elasticity, lifting height, pore reduction rate, and increase rate in skin brightness.

[0257] Second, as risk indicators, it includes negative response variables such as the amount of temperature increase immediately after the procedure, user-reported pain scores, duration of mild erythema, and the incidence of microburns.

[0258] At this point, the AI ​​analysis server assigns weights to each item and calculates the integrated score.

[0259] For example, the procedure performance evaluation score (P) can be calculated by offsetting the effectiveness indicator by multiplying it by a weight of +1 and the risk indicator by -1. This score is not a simple average, but reflects the relative improvement compared to historical user data accumulated in the Skin Profile Database (SPD). For instance, if the improvement rate is higher than the historical average under the same conditions, the reward value is calculated as a positive number, and if the risk is greater than the historical average, the reward value is calculated as a negative number.

[0260] The feedback-based reinforcement learning module converts the procedure performance evaluation score into a reward value (R) and utilizes it in the reinforcement learning process.

[0261] Specifically, the parameter combination used in the previous procedure is considered as the state, and the procedure parameters selected by the AI ​​are considered as the action; Q-learning or policy-based reinforcement learning (policy gradient) techniques are performed using the reward value obtained as a result.

[0262] In the case of the Q-learning approach, the module updates the value function of the current state-action pair and modifies the policy to predict a higher expected reward when applying the same parameters to users under the same or similar conditions in the future.

[0263] In the policy-based approach, the parameter selection probability distribution is directly updated to gradually strengthen the selection probability of highly effective actions.

[0264] These learning results are stored in the Skin Profile Database (SPD) and are subsequently referenced when calculating treatment parameters for the same or similar users.

[0265] For example, if a user's treatment result shows that the 7 MHz·0.8 second irradiation condition yielded a high reward, then in the next treatment, the weight of that combination is reflected highly not only to the same user but also to other users with similar skin characteristics.

[0266] Conversely, parameter combinations in which negative reactions such as overheating or pain are observed yield low reward values ​​and are automatically downgraded in priority within the policy.

[0267] Furthermore, since this reinforcement learning process is not limited to a single user but operates based on accumulated data from multiple users, the entire system forms a self-learning structure that evolves its policies over time.

[0268] As a result, when calculating treatment parameters for new users, combinations with verified safety and efficacy can be recommended first by referring to numerous feedback cases accumulated in the past.

[0269] Through such a feedback-based reinforcement learning procedure, the AI ​​analysis server of the present invention functions as a dynamic optimization system that reflects both real-time feedback and long-term accumulated data, rather than simply selecting parameters based on static rules.

[0270] Accordingly, as the procedure is repeated, the accuracy and stability of the treatment policy improve, and more precise customized treatments can be provided tailored to each individual's skin characteristics and reaction patterns.

[0271] In summary, this embodiment specifically demonstrates a technical procedure for collecting skin change data obtained after a procedure in real time, interpreting it as a reward signal for a reinforcement learning structure, and updating a procedure policy.

[0272] This provides the effect of fundamentally improving the limitations of existing fixed surgical devices by implementing an autonomous closed-loop system in which AI learns in a direction that “maximizes effectiveness and minimizes risk,” rather than relying on feedback using simple statistical averages.

[0273] In a customized skin care system according to one embodiment of the present invention, the AI ​​analysis server (130) does not merely accumulate and store feedback data obtained during multiple treatment processes, but manages it as time-series data by aligning it based on a time axis.

[0274] In other words, the treatment date, treatment parameters (frequency, output, irradiation time), measured effect indicators (elasticity, pores, lifting, etc.), and risk indicators such as side effects or pain are recorded in chronological order in each user's Skin Profile Database (SPD) and stored as a single continuous response history.

[0275] The feedback-based reinforcement learning module (132) of the AI ​​analysis server performs a precision correction routine based on this time series data, which reflects not only short-term effects but also long-term response trends.

[0276] For example, when the procedure is repeated at one-month intervals with the same parameters, the elasticity index improves rapidly in the beginning, but after a certain point, the rate of improvement slows down or a plateau phenomenon occurs.

[0277] The AI ​​analysis server calculates the rate of change in response over time (ΔE / Δt) and evaluates short-term and long-term effects separately.

[0278] Through this, if a certain parameter combination has excellent short-term effects but low sustainability, the policy weight of that combination is gradually reduced; conversely, combinations with stable long-term effectiveness are strengthened by cumulatively adding reward values.

[0279] During this correction process, the AI ​​analysis server periodically retrains the skin response model for each user.

[0280] This model can be implemented as a regression structure (deep learning-based regression network or time series LSTM structure, etc.) with input variables such as skin layer depth, reflective characteristics, treatment intensity, and irradiation time, and output variables such as the rate of change of effect and risk indicators.

[0281] The AI ​​analysis server calculates the “predicted effect at day n after the procedure” for a specific user through this model, defines the error between the prediction and the actual measurement as the model error (loss), and incorporates it as one of the reward items in reinforcement learning.

[0282] In other words, learning proceeds in a direction that strengthens the corresponding policy “the higher the prediction accuracy and the more stable the long-term trend.”

[0283] In addition, the AI ​​analysis server compares and analyzes long-term data from multiple users within the SPD to detect cases where significant response variations occur between individual users, even when the same parameters are used.

[0284] This deviation analysis is performed by groups such as similar skin type, age, and gender, and if the deviation exceeds a certain threshold (e.g., ±15%), the policy confidence score for the corresponding parameter combination is down-corrected.

[0285] Through this, the AI ​​model finely tunes policies to accurately reflect individual user response characteristics without over-relying on cluster mean effects.

[0286] For example, if the improvement in effectiveness slows down after the third session in a user's repeated treatment record, the AI ​​analysis server recognizes this as a fatigue effect and reduces the probability of applying the same parameters in the future.

[0287] On the other hand, for other users showing a steady improvement trend with the same parameters, the parameter combination is adjusted by slightly increasing the investigation time or slightly changing the frequency while maintaining the same combination.

[0288] As such, the present invention evolves the reinforcement learning policy over the long term not only through simple feedback rewards but also through pattern analysis of temporally accumulated data.

[0289] In addition, such precise correction can be performed periodically (e.g., every 10 procedures), and the system corrects the policy by calculating at least three key indicators from the entire time-series data accumulated in the SPD, such as (i) the rate of change in effect over time for the same parameter, (ii) the pattern of effect accumulation during repeated procedures, and (iii) the response deviation from similar user clusters.

[0290] The AI ​​analysis server resets the internal weights or value functions of the procedure parameter calculation module based on one or more of the above indicators, and this process can be implemented using a reinforcement learning algorithm with a Q-learning or Actor-Critic structure.

[0291] For example, a Critic network that predicts long-term rewards learns the cumulative sum of improvement rates over time as a reward, and an Actor network continuously updates a policy that selects the parameters best suited to the current skin condition.

[0292] As a result, the AI ​​analysis server of the present invention does not simply rely on feedback from a single procedure, but utilizes the entire accumulated temporal response data to model the user's skin response characteristics with increasing precision.

[0293] Therefore, in subsequent procedures, the skin's response can be predicted more accurately than in the initial stage, the consistency of the treatment effect is improved, and the possibility of side effects caused by excessive energy irradiation is significantly reduced.

[0294] In summary, the present invention according to claim 4 analyzes skin change data accumulated after a procedure in a time series form and implements an “AI skin response model that becomes more sophisticated over time” by precisely correcting a reinforcement learning policy based on the rate of change and the deviation between clusters.

[0295] Unlike conventional skin treatment devices based on fixed models, this features a continuous self-learning structure where the quality of the policy improves as treatment experience accumulates, providing the effect of simultaneously enhancing the reliability and safety of personalized skin care technology.

[0296] In a customized skin care system according to one embodiment of the present invention, an AI analysis server (130) statistically analyzes the skin characteristics, treatment response patterns, and reinforcement learning results of multiple users stored in a skin profile database (SPD) and classifies users exhibiting similar characteristics into one or more user clusters.

[0297] This classification process can be performed using unsupervised learning-based clustering algorithms (K-means, Gaussian Mixture Model, Hierarchical Clustering, etc.), and input variables include skin layer depth, skin reflectance coefficient, age, gender, skin type, purpose of treatment, and past treatment response indicators.

[0298] For example, users who have ① a dermal layer thickness of 2.0 to 2.5 mm, ② a skin reflection coefficient in the range of 0.4 to 0.5, and ③ show a high elasticity improvement rate in the frequency band of 7 to 9 MHz are classified into one group (Group D1), while users who are lifting-focused, with a thick SMAS layer and a good response to low frequencies of 1 to 3 MHz, are classified into another group (Group S1).

[0299] The AI ​​analysis server maintains an independent policy network or policy weight set for each cluster formed in this way.

[0300] Each policy network is a reinforcement learning model trained to reflect the characteristics of the cluster, representing the average response pattern of the cluster members.

[0301] In other words, the policy network of cluster D1 is optimized for dermal-centered elasticity improvement procedures, and the policy network of cluster S1 is optimized for SMAS-centered lifting procedures.

[0302] When a new user registers in the system or an existing user requests a treatment again, the AI ​​analysis server compares the user's skin data and SPD profile with existing clusters to calculate a cluster matching score based on cosine similarity or Mahalanobis distance.

[0303] The policy network of the cluster with the highest fitness score is selected as the base policy network to be applied preferentially to that user.

[0304] Subsequently, as the procedure is repeated and response data specialized for the user is accumulated, the AI ​​analysis server uses this data to perform transfer learning from the cluster's representative policy network.

[0305] The transfer learning process is carried out as follows.

[0306] The weights of the existing cluster policy network are loaded as initial values, and fine-tuning is performed using actual user procedure feedback data.

[0307] In this case, instead of retraining the entire network, learning stability is maintained and only individual characteristics are reflected by updating only some parameters of the upper layer.

[0308] As a result, a personalized policy network optimized only for that user is formed.

[0309] After the procedure is performed, the AI ​​analysis server compares the newly formed individual policy network with the average response pattern of the cluster to which it belongs to determine whether the deviation in terms of effectiveness or risk indicators exceeds a threshold value (e.g., ±10%).

[0310] If the user's response variance is large, it implies that the user possesses characteristics different from the general response of the existing cluster; therefore, the AI ​​analysis server partially separates the user from the cluster or forms a new sub-cluster, and updates the personalized policy network to maintain it independently.

[0311] Conversely, if the deviation is stable below the threshold value, the learning results of the individual policy network are integrated back into the cluster policy network to perform retraining in a direction that improves the policy quality of the entire cluster.

[0312] This mutual correction between clusters and individuals combines the transfer learning concept of reinforcement learning with the structure of federated learning, improving overall system performance by sharing each individual's learning outcomes at the cluster level without the entire server directly exchanging data from multiple users.

[0313] Furthermore, since the policy networks generated for each cluster are utilized as "pre-trained models" for calculating initial parameters for new users, the system enables very rapid initial customization for new users as well.

[0314] For example, when a new user is registered and characteristics such as a dermal depth of 2.4 mm, a reflection coefficient of 0.45, sensitive skin, and the purpose of improving elasticity are detected from the skin measurement unit, the AI ​​analysis server determines that the user is matched to cluster D1 (dermal-centered improvement type) with a fitness of 0.91, and loads the policy network of cluster D1 as the initial policy to immediately calculate the treatment parameters.

[0315] Afterward, as feedback data accumulates through several procedures, fine-tuning tailored to individual characteristics is performed, and as a result, the individual policy network achieves an average improvement rate 7% higher than the cluster policy network.

[0316] At this point, the server reverse transfers some of the weights from the individual policy network to the cluster policy network to improve the model quality of the entire cluster.

[0317] Therefore, unlike the conventional method of applying all users collectively to a single model, the cluster-based policy network structure according to the present invention finely reflects physiological and physical differences among users, thereby ① increasing the efficiency of group learning among similar users, ② implementing an adaptive model that rapidly reflects individual specificities, and ③ providing the effect of increasing personalization precision while maintaining the generalization performance of the entire system.

[0318] Ultimately, the present invention implements a self-learning customized skin care system capable of calculating treatment parameters with increasing precision and reliability over time through a bidirectional evolutionary structure in which the learning results of each cluster are transferred back to a personalized model, and the experience of the personalized model is fed back to the cluster mean.

[0319] In summary, claim 5 defines a process in which an AI analysis server forms a cluster-based policy network using multiple user data, and transfers, corrects, and individualizes the policy network based on actual feedback results from new or repeat users.

[0320] Through this, the present invention moves away from the conventional "uniform treatment intensity setting" method and provides an evolutionary personalized treatment algorithm tailored to each user's skin characteristics, which serves as a key technical means to continuously improve both clinical safety and efficacy.

[0321] In a customized skin care system according to one embodiment of the present invention, the system has a structure in which an AI analysis server calculates treatment parameters based on a user's skin data and performs reinforcement learning, but includes a security design that strictly separates data access rights between a user server (120) and an AI analysis server (130) to prevent personal identification information from being exposed to the outside during this process.

[0322] The user server (120) stores user-specific identification information (ID, name, contact information, etc.), skin data, and treatment history data in mutually separated areas.

[0323] For example, the internal database is dual-separated into a “personal information table” and a “procedure data table,” and the two tables are connected only through an encrypted matching token or anonymous key.

[0324] This structure forms a connectionless security architecture where the entire personal information cannot be restored even from a single data leak.

[0325] Before transmitting data to the AI ​​analysis server via the communication module (150), the user server removes the user's direct identification elements (name, phone number, date of birth, etc.) from the procedure data table and inserts a randomly generated non-identification token (UID) instead.

[0326] This UID is generated internally using a hash-based random number generator (e.g., SHA-256, UUID v4, etc.) and is designed so that the user's identity cannot be traced back even if the AI ​​analysis server receives it.

[0327] The AI ​​analysis server (130) performs procedure parameter calculation and reinforcement learning using only the de-identified skin reaction data among the data received from the user server.

[0328] In other words, the analysis server processes only pure technical data (feature data) such as UID, skin layer depth, reflection characteristics, effectiveness indicators, and risk indicators, and does not access direct identity information such as the user's name, location, body image, or contact information.

[0329] This process is a kind of “data masking” procedure that maintains an anonymous state in which personal identification is impossible in the computational environment of the AI ​​analysis server.

[0330] When the AI ​​analysis server performs reinforcement learning, it recognizes the procedure results for each UID stored in the SPD as a “virtual agent.”

[0331] In other words, each training sample of the reinforcement learning model is represented as an anonymous user corresponding to a UID, and the model updates the policy using only performance and risk feedback at the UID level.

[0332] As such, the AI ​​analysis server uses skin reaction data solely for technical purposes, and is not granted any access rights to the original data, including identification information.

[0333] In the system according to the present invention, re-identification is allowed only within the user server (120).

[0334] In other words, if it is necessary to match procedure result data or AI analysis results to a user account, the result file output by the AI ​​analysis server is assigned a UID, and the user server restores the actual user account using this UID through an internal matching table.

[0335] This restoration procedure is an encryption key-based re-identification process, where only the user server possesses the corresponding key.

[0336] For example, if the AI ​​analysis server returns an analysis result stating that “the combination of procedure parameters corresponding to UID 8F29A-XX21 is optimal,” the user server refers to the UID-User ID mapping table and attributes the result to the actual user account.

[0337] The AI ​​analysis server immediately deletes session data associated with the UID after transmitting results, thereby preventing the possibility of personally identifiable information being leaked outside the system during data exchange between servers.

[0338] In addition, the present invention applies TLS (Transport Layer Security)-based encrypted communication in the data transmission and reception path, and the communication module (150) prevents data tampering by performing an electronic signature and integrity verification procedure for each packet.

[0339] This structure is technically significant in that it is not merely a software-based personal information protection measure, but rather forms a physically separated dual-domain structure in which personal information data and technical data are physically separated at the level of actual system architecture.

[0340] According to the data security design of the present invention, the AI ​​analysis server does not access any personally identifiable information while performing procedure parameter calculation and reinforcement learning, and utilizes only anonymized skin response data. Therefore, the risk of personal information infringement is fundamentally blocked during the AI ​​model training process.

[0341] In addition, since procedure result data can only be re-identified through the user server,

[0342] As a result, ownership and management rights of user data are retained by the user.

[0343] This structure satisfies the technical requirements of major personal information regulatory frameworks, such as the European Union's GDPR (General Data Protection Regulation), the Republic of Korea's Personal Information Protection Act, and the U.S. HIPAA (Health Insurance Portability and Accountability Act), and also complies with the principles of data minimization and the principle of purpose limitation required in AI medical and healthcare systems.

[0344] Furthermore, since user identity information and technical data are stored and processed independently, the present invention can be safely operated in various environments such as medical institutions, clinics, or home care devices, and only non-identifiable data can be shared during collaborative data analysis with external research institutions, thereby simultaneously ensuring data security and technical usability.

[0345] In summary, the present invention according to claim 6 provides an advanced security structure that simultaneously secures the technical utility and legal safety of an AI-based skin care system by using a data separation structure between a user server and an AI analysis server to ① store personal information separately, ② utilize only de-identified technical data in the AI ​​analysis stage, and ③ perform re-identification only within the user server in a limited manner.

[0346] Thus, the present invention completes the technical foundation that satisfies both domestic and international regulatory compliance and reliability as an AI system handling “biometric data including personal information.”

[0347] Those skilled in the art will understand that the various exemplary logic blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented by electronic hardware, various forms of programs or design code (referred to herein as software for convenience), or a combination of all such. To clearly illustrate this interoperability between hardware and software, various exemplary components, blocks, modules, circuits, and steps have been generally described above in relation to their functions. Whether such functions are implemented as hardware or software depends on the design constraints imposed on the specific application and the overall system. Those skilled in the art may implement the functions described in various ways for each specific application, but such implementation decisions should not be construed as being outside the scope of this invention.

[0348] The various embodiments presented herein may be implemented as methods, devices, or articles manufactured using standard programming and / or engineering techniques. The term "article manufactured" includes a computer program, a carrier, or a medium accessible from any computer-readable storage device. For example, computer-readable storage media include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips, etc.), optical discs (e.g., CDs, DVDs, etc.), smart cards, and flash memory devices (e.g., EEPROMs, cards, sticks, key drives, etc.). Additionally, the various storage media presented herein include one or more devices and / or other machine-readable media for storing information.

[0349] It should be understood that the specific order or hierarchy of steps in the presented processes is merely an example of exemplary approaches. It should be understood that the specific order or hierarchy of steps in the processes may be rearranged within the scope of this application based on design priorities. The appended method claims provide various step elements in a sample order, but do not imply limitation to the specific order or hierarchy presented.

[0350] The description of the presented embodiments is provided to enable any person skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments without departing from the scope of the present invention. Thus, the present invention is not limited to the embodiments presented herein, but should be interpreted in the broadest possible scope consistent with the principles and novel features presented herein. Explanation of the symbols

[0352] 100: User-customized skincare system

Claims

Claim 1 In a customized skin care system, a skin measurement unit that transmits and receives ultrasound to a user's skin to acquire skin data including the depth distribution and acoustic reflection characteristics of multiple skin layers, including the epidermal layer, dermal layer, and SMAS layer; a user server that creates and maintains a skin profile database by storing user-specific skin data, treatment history, improvement results, and user input information in a time-series format; an AI analysis server configured to receive skin data and treatment history information from the skin measurement unit and the user server, and (i) extract feature values ​​corresponding to the target skin layer and improvement purpose among the multiple skin layers, (ii) calculate treatment parameters suitable for the target skin layer among multiple candidate frequencies of 1 MHz or more and 20 MHz or less, output intensity, and irradiation time, (iii) evaluate the treatment effect by analyzing skin change data re-received from the skin measurement unit or a separate feedback sensor after the treatment, and then accumulate and store the evaluation results in the user server, and (iv) update a treatment parameter calculation algorithm by reflecting the accumulated data based on reinforcement learning to evolvely optimize the optimal treatment parameters for each user through repeated treatments; and receiving the treatment parameters calculated from the AI ​​analysis server, The device is characterized by comprising: an ultrasound-based skin treatment device controlled to selectively output ultrasound corresponding to the frequency, output intensity, and irradiation time included in the above treatment parameters; and a communication module that transmits and receives data between the skin measurement unit, user server, AI analysis server, and ultrasound-based skin treatment device in a wired or wireless manner, wherein the AI ​​analysis server comprises the steps of: determining a target skin layer in the first step by comparing the depth distribution and tissue characteristic data of a plurality of skin layers included in the skin data received from the skin measurement unit with reference data; and correcting the target skin layer in the second step by reflecting an improvement objective specified by a user (pore reduction, elasticity improvement, lifting, etc.).A customized skin care system characterized by comprising: a step of generating a space of treatment candidate parameters limited to a frequency range, output intensity range, and irradiation time range corresponding to the target skin layer and improvement purpose, and configuring a set of treatment candidate parameters by applying constraints according to device safety limits and user skin type; a step of, for each of the candidate parameters, calculating a skin reaction similarity score with a cluster of similar users in a skin profile database stored on a user server and an effect prediction score based on past improvement rates, and calculating an overall suitability score by weighted summing these scores; and a step of sorting the candidate parameters in descending order according to a multi-criteria ranking rule that considers both the overall suitability score and the side effect risk score, and finally selecting one or more of the top-ranked parameters as treatment parameters. Claim 2 delete Claim 3 A customized skin care system according to claim 1, wherein the AI ​​analysis server receives, as real-time feedback data, reflected ultrasound signals reflected or transmitted from the skin, or bio-response data corresponding to changes in skin surface temperature and elasticity, which are ultrasounds output to the skin by the ultrasound-based skin treatment device during the treatment, and if the real-time feedback data deviates from a reference threshold value, the system is configured to dynamically correct treatment parameters by stepwise adjusting at least one of the currently applied frequency, output intensity, or irradiation time, or by switching to a parameter among the candidate parameters ranked in claim 1 that has a higher adaptability to the real-time feedback data. Claim 4 A customized skin care system according to claim 3, wherein the AI ​​analysis server accumulates and stores skin change data received from the skin measurement unit or a separate feedback sensor during multiple treatment processes as time-series data in the user's skin profile database, and from the accumulated data, precisely corrects a user-specific skin response model based on at least one indicator among (i) the rate of change in skin response over time for the same treatment parameter, (ii) the pattern of improvement in effect during repeated treatment, and (iii) the response deviation from a cluster of similar users with the same skin characteristics, and is configured to sequentially update the weight calculation or policy selection logic of candidate parameters in an episode-by-episode manner using a reinforcement learning method when calculating treatment parameters in the future by reflecting the precisely corrected skin response model. Claim 5 delete Claim 6 delete

Images