A periorbital aging multidimensional evaluation model and a construction method thereof

Abstract

The application discloses a periorbital aging multidimensional evaluation model and a construction method, and belongs to the technical field of medical artificial intelligence. The method comprises the following steps: multi-source data acquisition, multi-modal registration, six-dimensional feature extraction, converter-encoding-decoding network deep fusion modeling, treatment quantitative mapping and aging trend time series prediction. The application synchronously integrates static multi-modal images, high-speed dynamic expression sequences, biomechanics and individual baseline data, realizes unified quantitative evaluation of bone structure, soft tissue, skin, mechanics, dynamic function and individual information, completes cross-modal deep feature fusion through a converter-encoding-decoding network hybrid architecture, and realizes 1-3 year aging trajectory prediction by using a long short-term memory network. The application solves the problems of strong subjectivity, single dimension, dynamic and static split and lack of predictability in the prior art, and has the advantages of objective and comprehensive evaluation, high precision, direct guidance of individualized clinical treatment and suitability for medical cosmetology, plastic surgery and aging-related research.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence technology, specifically involving an orbital aging assessment model and its construction method based on the fusion of multimodal imaging, dynamic facial mechanics, biomechanics and individualized baseline information. It is particularly suitable for objective quantitative assessment of orbital aging, precise design of rejuvenation treatment plans and temporal prediction of aging trends. Background Technology

[0002] The periorbital region is the area with the thinnest facial skin, the most complex anatomical structure, and the earliest and most significant signs of aging. Its aging process involves multi-level coupled changes, including bone resorption, ligament laxity, fat displacement, muscle dysfunction, skin texture deterioration, and abnormal pigment and blood flow. Clinically and with current technology, periorbital aging assessment methods mainly include subjective grading (such as the facial aging local area grading method), two-dimensional image assessment, three-dimensional morphological measurement, and ultrasound soft tissue assessment. Among these, subjective grading relies entirely on the physician's clinical experience, leading to significant differences in assessment standards among physicians and low consistency. Two-dimensional image assessment only reflects the planar morphology of the periorbital region and cannot reflect three-dimensional structural changes. While three-dimensional morphological measurement can capture contour features, it does not consider tissue responses and biomechanical characteristics under dynamic facial expressions, failing to comprehensively reflect the true state of periorbital aging. Ultrasound assessment focuses only on subcutaneous soft tissue, with significant limitations in assessing bone structure, skin texture, and dynamic function.

[0003] The existing technology objectively has the following defects:

[0004] 1. The assessment dimensions are singular and lack systematic integration. It fails to uniformly quantify and model bone structure, dynamic facial biomechanics, biomechanical support characteristics, subcutaneous blood flow, and pigmentation. As a result, the assessment results are one-sided and cannot fully reflect the true mechanism of periorbital aging. Specifically, existing technologies often conduct assessments on a single dimension, such as focusing only on skin texture or soft tissue morphology. They fail to recognize that periorbital aging is the result of the coordinated degeneration of multiple structures, including bone, ligaments, fat, muscles, and skin. Consequently, the assessment results cannot provide comprehensive technical support for clinical treatment.

[0005] 2. It is highly subjective, with the consistency of assessments among different physicians usually below 65%, and lacks a unified, repeatable, and traceable objective quantitative standard. The existing subjective grading method does not have clear quantitative indicators and is graded solely by physicians' visual observation. It is easily affected by factors such as physicians' experience, aesthetic differences, and observation environment, resulting in poor repeatability of assessment results and failing to meet the requirements of precision medicine for objective and standardized assessments.

[0006] 3. The assessment is mainly static and does not collect data on tissue traction, displacement and stress distribution under facial expressions such as blinking and smiling. It is impossible to identify dynamic weak areas, which can easily lead to postoperative facial stiffness and unnatural results. The periorbital tissue will undergo obvious traction and displacement under dynamic facial expressions. The weak areas of tissue in the dynamic state are often the key causes of postoperative stiffness and poor results. The current static assessment cannot capture this kind of information, resulting in a lack of targeted treatment plan design.

[0007] 4. The evaluation accuracy is limited because the multi-source heterogeneous data are simply superimposed without deep feature fusion and adaptive weight allocation. The data sources, data types and dimensions of visible light, near-infrared, three-dimensional structured light, ultrasound, cone-beam computed tomography and biomechanical data are different and belong to heterogeneous data. Existing technology simply superimposes the evaluation results of various data without exploring the inherent relationship between different data, and cannot realize the synergistic use of multi-source information, which makes it difficult to improve the evaluation accuracy.

[0008] 5. The lack of aging trajectory prediction and quantitative mapping for individualized treatment leads to severe homogenization of treatment plans. Existing models can only assess the current periorbital aging status of subjects and cannot predict future aging trends. Moreover, most treatment plans are general and do not take into account the individual differences of subjects (such as age, skin type, and aging type) for quantitative design, resulting in inconsistent treatment effects and difficulty in meeting individualized anti-aging needs.

[0009] The aforementioned problems collectively lead to an incomplete, objective, and precise assessment of periorbital aging. For a long time, those skilled in the art have been unable to obtain a complete technical solution that integrates multimodal imaging, dynamic facial expression sequences, biomechanics, artificial intelligence, trend prediction, and quantitative mapping of treatment, thus hindering the development of precise and individualized periorbital rejuvenation treatment. Summary of the Invention

[0010] In view of this, the present invention provides a multi-dimensional assessment model and construction method for periorbital aging, which realizes a comprehensive assessment that is objective, quantitative, dynamic and individualized, and outputs treatment plans and aging trend predictions that can directly guide clinical practice, solving the technical pain points that existing technologies cannot overcome, and providing precise and comprehensive technical support for periorbital rejuvenation treatment.

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] A method for constructing a multidimensional assessment model for periorbital aging includes the following steps:

[0013] 1) Simultaneously collect static multimodal imaging data, dynamic facial expression sequence data, biomechanical data, and clinical baseline data of the subject's periorbital region;

[0014] 2) Noise reduction and normalization were performed on the multi-source data. Rigid registration of multimodal images was achieved based on 36 anatomical landmarks around the orbit. The optical flow method was used to complete non-rigid registration of dynamic expression sequences and to calculate tissue displacement, traction rate and stress concentration area.

[0015] 3) Extract quantitative features from six dimensions: bone structure, soft tissue morphology, skin texture, biomechanics, dynamic function, and individualized baseline;

[0016] 4) The random forest algorithm is used to screen core features, and cross-modal feature deep fusion is achieved through a converter-encoder-decoder network hybrid architecture. The clinical standardized grading is used as a label to train the evaluation model, and the output of the periorbital aging comprehensive index, aging grade, dominant degeneration level and dynamic mechanical risk heat map are output.

[0017] 5) Based on the assessment results, establish a quantitative mapping relationship for treatment plans and use a long short-term memory network model to predict the 1-3 year periorbital aging development trend.

[0018] Furthermore, the static multimodal imaging data includes visible light images, near-infrared blood flow images, three-dimensional structured light point clouds, ultrasound soft tissue images, and cone-beam computed tomography (CBCT) images of the orbital bone; wherein, the resolution of the visible light images is not less than 300 dpi, the wavelength of the near-infrared blood flow images is 700-900 nm, the accuracy of the three-dimensional structured light point clouds is not less than 0.1 mm, the frequency of the ultrasound images is 15 MHz, and the resolution of the cone-beam computed tomography images is not less than 0.01 mm;

[0019] The dynamic facial expression sequence data is collected by a high-speed camera with a frame rate of no less than 150 frames per second, including four standard facial expressions: blinking, smiling, closing eyes, and frowning, with no less than 50 frames of images captured for each expression;

[0020] The biomechanical data includes skin elastic modulus, ligament support stiffness, and orbicularis oculi muscle electromyography (EMG) signals. The skin elastic modulus test points are the infraorbital region, outer canthus region, upper eyelid region, and tear trough region. The ligament support stiffness test objects are the orbital septum ligament and the orbicularis oculi muscle supporting ligament. The EMG signal sampling frequency is 1000Hz.

[0021] Furthermore, the six dimensions of quantification features in step 3) specifically include:

[0022] Bone structure dimensions: orbital rim bone resorption, orbitozygomatic angle, orbital floor depth;

[0023] Soft tissue morphology dimensions: tear trough depth, eyelid bag volume, lateral canthus ptosis angle, and fat compartment displacement distance;

[0024] Skin texture dimensions: wrinkle density, skin roughness, pigmentation concentration, and blood perfusion index;

[0025] Biomechanical dimensions: skin elasticity, ligament stiffness, and electromyographic signal amplitude;

[0026] Dynamic functional dimensions: maximum displacement of marker points, tension rate, and distribution of stress hotspots;

[0027] Individualized baseline dimensions: coded age, gender, skin type, past treatment history, and lifestyle habit score.

[0028] Furthermore, in step 4):

[0029] The feature selection criterion is that the feature importance score is not lower than 0.05;

[0030] In the transformer-encoder-decoder hybrid architecture, the encoder-decoder network is used to extract local spatial features of the image, including convolutional layers, pooling layers, and deconvolutional layers; the transformer is used to map multimodal features to the same feature space, and achieves heterogeneous feature alignment, correlation weight calculation and adaptive global fusion through a multi-head self-attention mechanism.

[0031] The model training uses an adaptive momentum optimization algorithm, with no fewer than 80 training iterations and a learning rate of 0.0005-0.001.

[0032] Furthermore, the prediction process of the Long Short-Term Memory network model in step 5) is as follows:

[0033] A time-series sample was constructed using baseline data and multiple longitudinal follow-up assessments of the subjects. The follow-up period was once a month, and the follow-up duration was no less than 12 months. The long short-term memory network was trained to learn the temporal evolution pattern of periorbital aging, and the number of neurons in the hidden layer of the network was no less than 64. Using the current multidimensional assessment results as the initial input, iterative reasoning was used to obtain the change curve of the comprehensive periorbital aging index, the rate of degeneration in each dimension, and the aging risk warning for the next 1-3 years.

[0034] Furthermore, the periorbital aging comprehensive index is 0-100 points, and the classification is as follows:

[0035] 0-30 points indicates mild aging, 31-70 points indicates moderate aging, and 71-100 points indicates severe aging.

[0036] Furthermore, the treatment plan mapping in step 5) includes quantitative recommendations for injection filler dosage, radiofrequency energy, and surgical scope, which correspond one-to-one with aging grade, dominant degeneration level, and dynamic mechanical risk hotspot; wherein, the filler dosage for mild aging is 0.3-0.5 ml, and the radiofrequency energy is 30-40 joules; the filler dosage for moderate aging is 0.5-1.0 ml; and a comprehensive treatment plan is adopted for severe aging.

[0037] The second aspect of this invention is implemented by the following scheme:

[0038] A multidimensional assessment model for periorbital aging, constructed using the aforementioned method, includes the following sequentially connected components:

[0039] The input layer is used to receive heterogeneous data from multiple sources.

[0040] The feature layer is used to complete registration, feature extraction, and filtering.

[0041] The fusion evaluation layer achieves deep fusion of multimodal features and aging quantization evaluation based on a converter-encoder-decoder network hybrid architecture;

[0042] The output layer is used to output assessment reports, dynamic risk heatmaps, quantitative treatment recommendations, and aging trend prediction curves.

[0043] The beneficial effects of this invention are as follows:

[0044] 1. For the first time, it unifies the modeling of six dimensions: bone structure, soft tissue, skin, biomechanics, dynamic function, and individualized baseline, covering the complete hierarchy of periorbital aging and solving the problem of single assessment dimensions in existing technologies. Through multi-dimensional collaborative assessment, it can fully reflect the real mechanism of periorbital aging, providing comprehensive technical support for clinical treatment. Compared with existing single-dimensional assessments, the comprehensiveness of the assessment is improved.

[0045] 2. The entire process adopts artificial intelligence for automatic quantitative processing, establishes unified quantitative standards, eliminates differences in physicians' subjective judgment, and increases the consistency among assessors from below 65% to above 92%. The assessment results are repeatable and traceable, fully meeting the requirements of precision medicine for objectivity and standardization in assessment.

[0046] 3. By collecting dynamic facial expression sequences, we can identify dynamic stress concentration areas and tissue weaknesses, making the treatment plan more targeted. This can effectively avoid the problems of stiff facial expressions and unnatural results after treatment, reducing the postoperative facial stiffness rate from 20% to below 5%, significantly improving the naturalness of treatment, and increasing the treatment satisfaction of subjects.

[0047] 4. It enables 1-3 year aging trajectory prediction, allowing for early forecasting of periorbital aging trends and providing forward-looking guidance for long-term anti-aging treatment; at the same time, it establishes a quantitative mapping relationship between assessment results and treatment plans, outputting quantitative treatment parameters that can directly guide clinical practice, addressing the pain point of homogenization in existing treatment plans, improving treatment effectiveness, and enhancing long-term anti-aging effects. Attached Figure Description

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0049] Figure 1 This is a block diagram of the overall structure of the multi-dimensional assessment model for periorbital aging of the present invention;

[0050] Figure 2 This is a schematic diagram of the evaluation model construction method of the present invention. Detailed Implementation

[0051] like Figure 1 As shown, this invention provides a method for constructing a multi-dimensional assessment model for periorbital aging, comprising the following steps:

[0052] Step 1: Synchronous Acquisition of Multi-Source Heterogeneous Data

[0053] Four types of data were collected from the periorbital region of the subjects. All data were collected simultaneously under the same standard environment to avoid the influence of environmental factors (such as light and temperature) on data accuracy and to ensure the spatiotemporal consistency of multi-source data, laying the foundation for subsequent registration, fusion, and evaluation. The specific collection content and operation standards are as follows:

[0054] 1) Static multimodal imaging data: including visible light images, near-infrared blood flow images, 3D structured light point clouds, high-frequency ultrasound images, and cone-beam computed tomography (CBCT) images of the orbital bone; among them, visible light images were acquired using a 50-megapixel high-definition camera under standard white light (color temperature 5500K, illuminance 1000 lux) at three angles: frontal, 45° lateral, and 90° lateral, with an image resolution of no less than 300 dpi, used to extract surface features such as skin texture, wrinkles, and pigmentation; near-infrared blood flow images were acquired using a near-infrared imaging device with a wavelength of 700-900nm and a sampling frequency of 10fps, used to capture the percutaneous blood flow perfusion of the orbit and differentiate the types of dark circles (vascular, pigmented, and structural). The three-dimensional structured light point cloud was acquired using a 0.1mm precision 3D structured light imaging device, with the scanning range covering the entire periorbital region (from the brow bone to the upper edge of the cheekbone, and from the temporal hairline to the sides), used to reconstruct the three-dimensional morphology of the periorbital region and calculate parameters such as volume and depth; high-frequency ultrasound images were acquired using a 15MHz ultrasound device, with the probe perpendicularly attached to the skin and coupling agent evenly applied to avoid air bubbles, with the scanning depth set to 0-10mm, used to observe the layered structure of periorbital soft tissue and the displacement of fat compartments; cone-beam computed tomography (CBCT) images of the orbital bone were acquired using a 0.01mm resolution CBCT device, with the scanning range covering the entire orbital bone region, used to extract orbital bone structural parameters and assess the degree of bone resorption.

[0055] 2) Dynamic facial expression sequence data: Four standard facial expression sequences—blinking, smiling, closing eyes, and frowning—were captured using a high-speed camera with a frame rate of at least 150 frames per second, with an optimal frame rate of 200 frames per second to ensure clear capture of tissue dynamic displacement. Specifically, the blinking expression required the subject to blink naturally three times, with a two-second interval between each blink; the smiling expression included a natural smile (slightly upturned corners of the mouth, no teeth showing) and a big laugh (widely upturned corners of the mouth, teeth showing), each smile held for 3 seconds; the closed-eye expression required the subject to completely close their eyes and hold for 3 seconds; and the frowning expression required the subject to frown forcefully and hold for 3 seconds. Fifty frames were captured for each expression to ensure coverage of the complete movement process, used to analyze tissue displacement, traction rate, and stress distribution under dynamic conditions.

[0056] 3) Biomechanical data: including skin elastic modulus, ligament support stiffness, and orbicularis oculi muscle electromyography (EMG) signals. Skin elastic modulus was collected using a skin mechanics testing instrument, with testing points in four key areas: the infraorbital region, the lateral canthus, the upper eyelid region, and the tear trough region. Each point was tested three times, with each test lasting 10 seconds, and the average value was used as the final data. The test pressure was set to 0.5-1.0 N, in kilopascals. Ligament support stiffness was collected using a ligament stiffness testing instrument, focusing on two core ligaments: the orbital septum ligament and the orbicularis oculi muscle supporting ligament. Each ligament was tested three times, and the average value was used, in Newtons per meter (N / m). Orbicularis oculi muscle EMG signals were collected using EMG sensors, which were attached to the surface of the orbicularis oculi muscle (one sensor each for the upper and lower eyelids). EMG signals were collected in both resting (subject relaxed, eyes looking straight ahead) and dynamic facial expression states, with a sampling frequency of 1000 Hz, to assess the functional state and tension of the orbicularis oculi muscle.

[0057] 4) Clinical baseline data: including age, gender, skin type, past treatment history, sun exposure and sleep schedule scores; among them, skin type is divided into four categories: dry, oily, combination, and sensitive, and is classified and coded using commonly used clinical skin type assessment standards (dry=1, oily=2, combination=3, sensitive=4); past treatment history includes whether or not the patient has received periorbital injections, surgery, radiofrequency treatments, etc., coded as yes=1, no=0; sun exposure and sleep schedule scores are quantified on a scale of 0-10, with the scoring criteria as follows: no sun exposure, regular sleep schedule (7-8 hours of sleep per day) 8-10 points; occasional sun exposure, generally regular sleep schedule 5-7 points; frequent sun exposure, irregular sleep schedule (less than 6 hours of sleep per day) 0-4 points. All scores are jointly assessed by two senior plastic surgeons, and the average value is taken as the final score.

[0058] Step 2: Data Preprocessing and Multimodal Registration

[0059] The collected multi-source data is preprocessed to eliminate noise, artifacts, and dimensional differences, ensuring data quality. Simultaneously, multimodal registration unifies image data of different types and coordinate systems into a single coordinate system, achieving spatial alignment of the multi-source data. The specific operations are as follows:

[0060] First, various types of data are preprocessed: For image data such as visible light, near-infrared, 3D structured light, ultrasound, and cone-beam computed tomography, Gaussian filtering algorithm is used for noise reduction. The standard deviation of Gaussian filtering is set to 0.5-1.0, which effectively removes random noise and artifacts in the images while preserving the detailed features of the images. For biomechanical data such as skin elastic modulus, ligament stiffness, and electromyography signals, Z-score normalization method is used to normalize the data, eliminating the dimensional differences between different parameters and making all data of the same order of magnitude. The normalization formula is: x'=(x-μ) / σ, where x is the original data, μ is the mean of the data, and σ is the standard deviation of the data. For dynamic expression sequence images, inter-frame deduplication and motion blur correction algorithms are used to remove blurry and invalid frames and retain clear and complete dynamic sequences.

[0061] Secondly, based on an improved encoder-decoder network, 36 anatomical landmarks around the orbit were automatically identified, and rigid registration of multimodal images was completed to establish a unified coordinate system. The 36 anatomical landmarks include key points of the orbital rim (8 on the upper orbital rim and 8 on the lower orbital rim), key points of the eyelids (6 on the upper eyelid margin and 6 on the lower eyelid margin), and attachment points of the periorbital muscles (4 origins and 4 insertions of the orbicularis oculi muscle). The improved encoder-decoder network adds an attention mechanism module to the traditional encoder-decoder network to improve the accuracy of landmark recognition, with the recognition error controlled within 0.5 mm. Rigid registration uses the line connecting the pupil centers as a reference and adopts an affine transformation algorithm to map all static multimodal image data to the same coordinate system, ensuring that the coordinates of the same anatomical position in different images are consistent, providing a spatial basis for subsequent feature extraction and fusion.

[0062] Finally, the Lucas-Cannard optical flow method was used to perform non-rigid registration of the dynamic facial expression sequence, calculating the displacement field, traction velocity, and stress concentration areas of the marker points. The Lucas-Cannard optical flow method extracts the inter-frame deformation field of the dynamic facial expression sequence by calculating the displacement of pixels in adjacent frames, achieving non-rigid registration with the registration error controlled within 0.3 mm. Based on the registration results, the displacement (in millimeters) and traction velocity (in millimeters / second) of each anatomical marker point under different facial expressions were calculated, and the stress concentration areas were visualized through heatmaps. The areas with a displacement ≥ 2 mm were defined as stress concentration areas, i.e., dynamic tissue weakness areas.

[0063] Step 3: Extraction and Quantization of Features in Six Dimensions

[0064] Based on the preprocessed and registered data, the system extracts quantitative features of six core dimensions of periorbital aging. For each dimension, representative, quantifiable, and repeatable feature parameters are selected to ensure that the features can comprehensively and accurately characterize the periorbital aging state. The specific extraction methods and quantification standards are as follows:

[0065] 1) Bone structure dimensions: Based on cone-beam computed tomography (CBCT) images of the orbital bone, three core parameters were extracted using image analysis software: orbital rim bone resorption, orbitozygomatic angle, and orbital floor depth. Orbital rim bone resorption was quantified by the change in orbital rim cortical thickness. Three measurement points were selected at the anterior, middle, and posterior orbital rim. The difference between the cortical thickness at each measurement point and the normal cortical thickness for the same age group was calculated, and the average value was taken as the orbital rim bone resorption. The orbitozygomatic angle is the angle between the line connecting the lateral point of the orbital rim and the apex of the zygomatic arch and the horizontal baseline. It was directly measured using an angle measurement tool, and the unit is degrees. Orbital floor depth is the vertical distance between the deepest point of the orbital floor and the orbital rim plane, and the unit is millimeters.

[0066] 2) Soft tissue morphology dimension: Based on the three-dimensional structured light point cloud reconstruction of the periorbital 3D model, four core parameters were extracted using 3D modeling software: tear trough depth, eyelid bag volume, lateral canthus ptosis angle, and fat compartment displacement distance. Among them, tear trough depth is the vertical distance between the deepest part of the tear trough and the surrounding skin plane, and the average value of three measurement points is taken, with the unit being millimeters; eyelid bag volume is calculated by segmenting the eyelid bag area and calculating its volume using a 3D model volume calculation tool, with the unit being cubic millimeters; lateral canthus ptosis angle is the angle between the line connecting the lateral canthus point and the pupil center and the horizontal baseline, with the unit being degrees; and fat compartment displacement distance is the distance between the center of the periorbital fat compartment and its normal anatomical position, with the unit being millimeters.

[0067] 3) Skin texture dimension: Based on visible light images and near-infrared blood flow images, four core parameters were extracted: wrinkle density, skin roughness, relative pigment concentration, and blood perfusion index. Among them, wrinkle density was extracted using the gray-level co-occurrence matrix algorithm to calculate the number of wrinkles per unit area, with the unit being wrinkles / square centimeter; skin roughness was quantified using Ra / Rz values, where Ra is the arithmetic mean deviation of the contour and Rz is the maximum height of the contour, calculated using image gray-level analysis tools; relative pigment concentration was based on the gray-level values ​​of the visible light images, using the normal skin area as a benchmark, to calculate the relative gray-level values ​​of the periorbital pigmentation area and quantify the degree of pigmentation; blood perfusion index was based on the signal intensity of the near-infrared blood flow images to calculate the average intensity of the blood flow signal in the periorbital region and quantify the subcutaneous blood perfusion.

[0068] 4) Biomechanical dimension: Based on the preprocessed biomechanical data, three core parameters were extracted: skin elasticity, ligament stiffness, and electromyographic signal amplitude. Among them, skin elasticity was quantified by the average value of the skin elastic modulus, in kilopascals; ligament stiffness was quantified by the average value of the stiffness of the orbital septum ligament and the orbicularis oculi muscle supporting ligament, in Newtons per meter; and electromyographic signal amplitude was the maximum amplitude of the electromyographic signal in the resting state and dynamic facial expression state, in microvolts, used to assess the tension state of the orbicularis oculi muscle.

[0069] 5) Dynamic Functional Dimension: Based on the registration results of dynamic facial expression sequences, three core parameters are extracted: maximum displacement of anatomical landmarks, traction rate, and stress hotspot distribution. The maximum displacement of anatomical landmarks is the maximum displacement value of each anatomical landmark under four facial expressions, in millimeters. The traction rate is the ratio of the displacement of the landmark to time, in millimeters per second. The stress hotspot distribution is visualized through a heat map, marking the location, range, and stress intensity of stress concentration areas, and quantifying the degree of dynamic tissue weakness.

[0070] 6) Individualized baseline dimensions: The collected clinical baseline data were coded and quantified to form individualized baseline characteristics. Among them, age was directly taken as the subject's actual age (in years); gender was coded as male=1, female=0; skin type was coded as dry=1, oily=2, combination=3, sensitive=4; past treatment history was coded as yes=1, no=0; lifestyle habit score was taken as the sun exposure and rest score (0-10 points) collected in step 1. All coded data were standardized and used as individualized baseline characteristics.

[0071] Step 4: Multimodal Feature Fusion and Model Construction

[0072] The random forest algorithm was used to screen the importance of features, eliminating redundant features and retaining core features that significantly contribute to the assessment of periorbital aging, thereby reducing model complexity and improving model efficiency. The specific screening process was as follows: all quantitative features extracted from the six dimensions were used as input, and the standardized grading results of senior physicians were used as labels to train the random forest model. The importance score of each feature was calculated, and core features with an importance score of not less than 0.05 were selected to ensure that the core features can comprehensively represent the periorbital aging state, while avoiding model overfitting caused by feature redundancy.

[0073] A hybrid fusion network of converter-encoder-decoder network is constructed to achieve deep fusion of cross-modal features. This hybrid architecture combines the advantages of the strong local feature extraction capability of the encoder-decoder network and the strong global feature correlation capability of the converter network, realizing deep fusion of multi-source heterogeneous data. The specific implementation method is as follows:

[0074] Using an encoder-decoder network as a local feature extraction branch, visible light, near-infrared, 3D structured light, ultrasound, and cone-beam computed tomography images are encoded and decoded respectively. The encoding process extracts low-dimensional, mid-dimensional, and high-dimensional features of the images step by step through convolutional layers and pooling layers, capturing spatial local features such as skin texture, morphological contours, subcutaneous tissue structures, and bone resorption in the periorbital region. The decoding process restores the spatial resolution of the feature map step by step through deconvolutional layers, achieving refined extraction of local features, and finally outputting multi-scale image feature maps, with each scale feature map corresponding to different levels of image details.

[0075] Using a converter network as a global feature fusion branch, the local image features, biomechanical features, dynamic displacement features, and individualized baseline features output by the encoder-decoder network are mapped to the same high-dimensional feature space through a feature mapping layer, eliminating the heterogeneity of different modal features. Through a multi-head self-attention mechanism, the correlation weights between different modal features are calculated, and the contribution of each dimension of features to the assessment of periorbital aging is adaptively learned. For example, for subjects with significant bone resorption, the weight of bone structure dimension features is automatically increased, and for subjects with significant skin texture deterioration, the weight of skin texture dimension features is automatically increased. At the same time, through a global information interaction module, information complementarity between different modal features is achieved, the intrinsic correlation between multi-source data is explored, and the representational ability of features is improved.

[0076] The local fine features output by the encoder-decoder network are concatenated and fused with the global correlation features output by the converter network to obtain a high-dimensional fused feature vector that can comprehensively represent static structure, dynamic mechanics, biomechanics, and individual differences. The fusion process uses a feature concatenation algorithm to concatenate the local feature vector and the global feature vector by channel to form a unified fused feature vector. This vector contains multi-dimensional information about periorbital aging, providing comprehensive feature support for subsequent model training and evaluation.

[0077] The evaluation model was trained using standardized grading labels from senior physicians as supervised information. These standardized labels were jointly assigned by five plastic surgeons with the rank of associate chief physician or above, based on the subjects' periorbital aging status, according to a grading standard of mild (0-30 points), moderate (31-70 points), and severe (71-100 points). In case of discrepancies in the labeling results, a voting method was used to determine the final label. The model training employed an adaptive momentum optimization algorithm with 100 training iterations, a learning rate of 0.001, and a batch size of 16. Backpropagation was used to continuously adjust the model parameters and reduce prediction errors. After training, the model outputs: a comprehensive periorbital aging index (0-100 points), mild / moderate / severe grading, dominant degeneration level (i.e., the dimension contributing most to periorbital aging), and a dynamic biomechanical risk heatmap (visualizing stress concentration areas).

[0078] Step 5: Treatment Plan Mapping and Aging Trajectory Prediction

[0079] Based on the periorbital aging comprehensive index, dominant degeneration type, and stress hotspots output by the model, a quantitative treatment plan mapping relationship is established to achieve automatic recommendation of individualized treatment plans. The mapping relationship is constructed based on a large amount of clinical case data to ensure the scientific validity and practicality of the treatment plan; the specific mapping rules are as follows:

[0080] Mild aging (0-30 points): If the dominant degeneration level is skin texture, radiofrequency tightening treatment is recommended. The radiofrequency energy is set according to skin roughness and wrinkle density, ranging from 30-40 joules. If the dominant degeneration level is dynamic function, botulinum toxin injection is recommended. The injection dosage is set according to the amplitude of electromyography signals, ranging from 2-5 units. If the dominant degeneration level is soft tissue morphology, a small amount of hyaluronic acid filler is recommended. The filler dosage is set according to the depth of the tear trough and the volume of the eyelid bag, ranging from 0.3-0.5 ml.

[0081] Moderate aging (31-70 points): If the dominant degeneration level is soft tissue morphology (depression of the tear trough, protrusion of the eyelid bag), injection filling + orbital fat repositioning is recommended. The filling dosage is calculated based on the volume of the eyelid bag and the depth of the tear trough, ranging from 0.5 to 1.0 ml. The orbital fat repositioning range is determined based on the displacement distance of the fat compartment. If the dominant degeneration level is biomechanics (ligament laxity), ligament repair surgery is recommended. The repair range is determined based on the ligament stiffness value. If the dominant degeneration level is bone structure, bone augmentation treatment is recommended. The bone augmentation dosage is calculated based on the amount of bone resorption at the orbital rim.

[0082] Severe aging (71-100 points): If the dominant degenerative level is bone structure + ligament laxity, bone augmentation surgery + ligament repair surgery is recommended. The range of bone augmentation is determined based on the orbital floor depth and the amount of bone resorption at the orbital rim, and the range of ligament repair is determined based on ligament stiffness and displacement. If the dominant degenerative level is multi-dimensional synergistic degeneration, a comprehensive treatment plan is recommended, combining multiple treatment methods such as injection, radiofrequency ablation, and surgery, and quantifying the parameters and range of each treatment method.

[0083] A long short-term memory network was used to construct a temporal prediction model for periorbital aging, enabling quantitative prediction of the periorbital aging trend over the next 1-3 years. This model can capture the temporal evolution of periorbital aging and, combined with individual differences among subjects, provide forward-looking anti-aging guidance. The specific implementation process is as follows:

[0084] First, a time-series training sample set was constructed: 50 subjects were selected for a 12-month longitudinal follow-up, and periorbital multidimensional assessment data (including comprehensive index, six-dimensional sub-item scores, and dynamic biomechanical parameters) were collected monthly. Combined with baseline assessment data, time-series input sequences were constructed in chronological order (baseline, 3 months, 6 months, 9 months, and 12 months). Each subject's time-series sequence served as a training sample, resulting in a total of 50 training samples. At the same time, the aging grading at each time point was used as a label for model training.

[0085] Secondly, the Long Short-Term Memory (LSTM) network was trained: a time-series training sample set was input into the LTM network, which consists of an input layer, a hidden layer, and an output layer. The number of neurons in the hidden layer was set to 64, and the ReLU activation function was used. The output layer used a linear activation function to output the aging comprehensive index and scores for each dimension at future time points. The training process used an adaptive momentum optimization algorithm with 80 training iterations and a learning rate of 0.0005. The network parameters were adjusted through backpropagation so that the model could accurately learn the degeneration patterns and individual differences of periorbital tissues over time.

[0086] Finally, aging trend prediction: After training, the subjects' current multi-dimensional assessment results (comprehensive index, scores of six major dimensions, dynamic biomechanical parameters, and individualized baseline data) are used as initial input. Through the trained long short-term memory network, iterative reasoning is performed to output the change curve of the periorbital aging comprehensive index, the predicted value of the degeneration rate of each dimension (unit: points / year), and the key aging nodes (i.e. the time nodes when the aging rate is significantly accelerated) for the next 1-3 years (each 3 months is a time node). This provides a predictive basis for long-term individualized anti-aging intervention and helps physicians formulate phased treatment plans.

[0087] At the same time, such as Figure 2 As shown, this invention also protects the multi-dimensional assessment model for periorbital aging constructed by the above method. Its structure consists of an input layer, a feature layer, a fusion assessment layer, and an output layer. Each layer works collaboratively to achieve integrated assessment, visualization, and report export. The specific structure and functions are as follows:

[0088] Input layer: Used to receive multi-source heterogeneous data, including static multimodal image data, dynamic facial expression sequence data, biomechanical data and clinical baseline data, to achieve unified input and format conversion of multi-source data, and to convert different types of data into standardized data formats that the model can recognize.

[0089] Feature layer: Used to complete data preprocessing, landmark localization, feature extraction and filtering, specifically including image denoising, data normalization, multimodal registration, six-dimensional feature extraction, random forest feature filtering and other functions, outputting the filtered core feature vector to support subsequent fusion evaluation.

[0090] Fusion Assessment Layer: Based on a converter-encoder-decoder network hybrid architecture, it realizes deep fusion of multimodal features and aging quantification assessment. It deeply fuses the core features output by the feature layer, combines them with the trained assessment model, and completes the calculation of the periorbital aging comprehensive index, aging classification, identification of the dominant degeneration level, and dynamic mechanical risk analysis, outputting intermediate assessment results.

[0091] Output layer: Used to output assessment reports, dynamic risk heatmaps, quantitative treatment recommendations, and aging trend prediction curves. The assessment report includes information such as comprehensive index, grade, and dominant degeneration level; the dynamic risk heatmap visualizes stress concentration areas; the quantitative treatment recommendations specify treatment methods, parameters, and ranges; and the aging trend prediction curves intuitively display the aging development trend over the next 1-3 years. It also supports the export of assessment reports for convenient clinical archiving and use.

[0092] Based on the above scheme, the following specific embodiments are proposed:

[0093] Example 1: Model Construction and Training

[0094] 1. Data Collection

[0095] One hundred subjects aged 25-65 years were selected, including 50 males and 50 females, with an even age distribution (20 cases aged 25-35, 20 cases aged 36-45, 30 cases aged 46-55, and 30 cases aged 56-65). The subjects covered four skin types: dry, oily, combination, and sensitive. Thirty subjects had a history of periorbital treatment, and 70 subjects had no such history. Following the data collection criteria in step 1, the following data were collected simultaneously: visible light, near-infrared, three-dimensional structured light, ultrasound, and cone-beam computed tomography (CBCT) images; a high-speed dynamic facial expression sequence at 200 frames per second (covering blinking, smiling, eye closing, and frowning); biomechanical testing of four periorbital areas (infraorbital region, lateral canthus, upper eyelid region, and tear trough region) and electromyographic signals of the orbicularis oculi muscle; and clinical baseline information (age, gender, skin type, previous treatment history, sun exposure, and sleep-wake cycle scores).

[0096] 2. Registration and Feature Extraction

[0097] An improved encoder-decoder network was used to automatically identify 36 anatomical landmarks around the orbit, with the identification error controlled within 0.4 mm. Using the line connecting the pupil centers as a reference, an affine transformation algorithm was used to complete rigid registration of multimodal images and establish a unified coordinate system. The Lucas-Cannard optical flow method was used to perform non-rigid registration of dynamic expression sequences, with the registration error controlled within 0.2 mm. The displacement field, traction velocity, and stress concentration area of ​​each landmark were calculated. Based on the registered data, 31 quantitative features in six dimensions were extracted, including 3 bone structure dimensions, 4 soft tissue morphology dimensions, 4 skin texture dimensions, 3 biomechanical dimensions, 3 dynamic function dimensions, and 14 individualized baseline dimensions (including coded clinical baseline data).

[0098] 3. Feature selection and fusion modeling

[0099] A random forest algorithm was used to screen features based on their importance score, with a score no lower than 0.05. This resulted in 28 core features (3 from bone structure, 4 from soft tissue morphology, 4 from skin texture, 3 from biomechanics, 3 from dynamic function, and 11 from individualized baseline), with 3 redundant features removed. A hybrid converter-encoder-decoder network was constructed. The encoder-decoder branch contains 4 convolutional layers, 4 pooling layers, and 4 deconvolutional layers for extracting local image features. The converter branch contains 6 attention heads for cross-modal global feature fusion. The model was trained using an adaptive momentum optimization algorithm with labels uniformly annotated by 5 associate chief physicians or above. The training iterations were 100 times, with a learning rate of 0.001 and a batch size of 16.

[0100] 4. Model Output and Validation

[0101] After model training, the model outputs a comprehensive aging index, grading, dominant degeneration layer, and dynamic biomechanical risk heatmap. Twenty subjects who did not participate in training were selected as validation samples for external validation of the model. The model achieved an accuracy of 91.2%, a recall of 89.7%, an F1 score of 0.90, and inter-rater consistency of 92.6%, validating the model's accuracy and stability. Furthermore, the model's evaluation results were compared with those of five senior physicians, showing a consistency rate of 93.5%, significantly higher than the consistency rate of existing technologies (below 65%).

[0102] Example 2 Clinical Assessment Application

[0103] A 35-year-old female subject, with no history of periorbital treatment, combination skin, frequently stayed up late (sleep 5-6 hours a day), and had irregular sun protection, with a sun exposure and sleep schedule score of 4; according to the construction method of this invention, the subject underwent periorbital aging assessment and treatment guidance.

[0104] 1. Data Acquisition and Preprocessing: In accordance with the acquisition standards of Example 1, static multimodal images, dynamic facial expression sequences, biomechanical and clinical baseline data of the subject were acquired. The data were denoised and normalized to complete multimodal registration.

[0105] 2. Feature Extraction and Model Evaluation: Six core features were extracted and input into the trained evaluation model. The model output results were: Aging Comprehensive Index of 45 points (moderate aging), with the dominant degeneration level being soft tissue morphology (tear trough depression, eyelid bag protrusion) + skin texture (periocular fine lines, mild pigmentation); Dynamic biomechanical risk heat map showed that the outer canthus and infraorbital region were stress concentration areas (maximum displacement of landmark point was 2.3mm); The scores for each dimension were: bone structure 7 points, soft tissue morphology 8 points, skin texture 6 points, biomechanics 5 points, dynamic function 4 points, and individualized baseline 5 points.

[0106] 3. Recommended Treatment Plan: Based on the model output and the quantitative treatment plan mapping rules, the model provides an individualized treatment plan: 0.8 ml of hyaluronic acid filler under the eyelids (calculated based on tear trough depth and eyelid bag volume) + 35 joules of radiofrequency tightening around the eyes (set according to skin roughness and wrinkle density). It is also recommended to adjust lifestyle habits, maintain a regular schedule (7-8 hours of sleep per day), and strengthen sun protection (apply sunscreen daily and avoid prolonged sun exposure).

[0107] 4. Efficacy Verification: Three months after treatment, the subject was retested using the assessment model of this invention. The retest results showed that the comprehensive aging index decreased to 32 points (mild aging), the volume of eyelid bags decreased by 40%, the depth of tear troughs decreased by 50%, the density of fine lines around the eyes decreased by 35%, pigmentation was significantly reduced, and dynamic expressions were natural without stiffness. The subject's satisfaction with the treatment effect was 95%, which verified the accuracy of the assessment model of this invention and the effectiveness of the treatment guidance.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a multidimensional assessment model for periorbital aging, characterized in that, Includes the following steps: 1) Simultaneously collect static multimodal imaging data, dynamic facial expression sequence data, biomechanical data, and clinical baseline data of the subject's periorbital region; 2) Noise reduction and normalization were performed on the multi-source data. Rigid registration of multimodal images was achieved based on 36 anatomical landmarks around the orbit. The optical flow method was used to complete non-rigid registration of dynamic expression sequences and to calculate tissue displacement, traction rate and stress concentration area. 3) Extract quantitative features from six dimensions: bone structure, soft tissue morphology, skin texture, biomechanics, dynamic function, and individualized baseline; 4) The random forest algorithm is used to screen core features, and cross-modal feature deep fusion is achieved through a converter-encoder-decoder network hybrid architecture. The clinical standardized grading is used as a label to train the evaluation model, and the output of the periorbital aging comprehensive index, aging grade, dominant degeneration level and dynamic mechanical risk heat map are output. 5) Based on the assessment results, establish a quantitative mapping relationship for treatment plans and use a long short-term memory network model to predict the 1-3 year periorbital aging development trend.

2. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, The static multimodal imaging data includes visible light images, near-infrared blood flow images, three-dimensional structured light point clouds, ultrasound soft tissue images, and cone-beam computed tomography (CBCT) images of the orbital bone; wherein, the resolution of the visible light images is not less than 300 dpi, the wavelength of the near-infrared blood flow images is 700-900 nm, the accuracy of the three-dimensional structured light point clouds is not less than 0.1 mm, the frequency of the ultrasound images is 15 MHz, and the resolution of the cone-beam computed tomography images is not less than 0.01 mm. The dynamic facial expression sequence data is collected by a high-speed camera with a frame rate of no less than 150 frames per second, including four standard facial expressions: blinking, smiling, closing eyes, and frowning, with no less than 50 frames of images captured for each expression; The biomechanical data includes skin elastic modulus, ligament support stiffness, and orbicularis oculi muscle electromyography (EMG) signals. The skin elastic modulus test points are the infraorbital region, outer canthus region, upper eyelid region, and tear trough region. The ligament support stiffness test objects are the orbital septum ligament and the orbicularis oculi muscle supporting ligament. The EMG signal sampling frequency is 1000Hz.

3. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, Step 3) includes the following six dimensions of quantitative features: Bone structure dimensions: orbital rim bone resorption, orbitozygomatic angle, orbital floor depth; Soft tissue morphology dimensions: tear trough depth, eyelid bag volume, lateral canthus ptosis angle, and fat compartment displacement distance; Skin texture dimensions: wrinkle density, skin roughness, pigmentation concentration, and blood perfusion index; Biomechanical dimensions: skin elasticity, ligament stiffness, and electromyographic signal amplitude; Dynamic functional dimensions: maximum displacement of marker points, tension rate, and distribution of stress hotspots; Individualized baseline dimensions: coded age, gender, skin type, past treatment history, and lifestyle habit score.

4. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, In step 4): The feature selection criterion is that the feature importance score is not lower than 0.05; In the transformer-encoder-decoder hybrid architecture, the encoder-decoder network is used to extract local spatial features of the image, including convolutional layers, pooling layers, and deconvolutional layers; the transformer is used to map multimodal features to the same feature space, and achieves heterogeneous feature alignment, correlation weight calculation and adaptive global fusion through a multi-head self-attention mechanism. The model training uses an adaptive momentum optimization algorithm, with no fewer than 80 training iterations and a learning rate of 0.0005-0.

001.

5. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, Step 5) The prediction process of the medium- and long-term short-term memory network model is as follows: A time-series sample was constructed using baseline data and multiple longitudinal follow-up assessments of the subjects. The follow-up period was once a month, and the follow-up duration was no less than 12 months. The long short-term memory network was trained to learn the temporal evolution pattern of periorbital aging, and the number of neurons in the hidden layer of the network was no less than 64. Using the current multidimensional assessment results as the initial input, iterative reasoning was used to obtain the change curve of the comprehensive periorbital aging index, the rate of degeneration in each dimension, and the aging risk warning for the next 1-3 years.

6. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, The periorbital aging comprehensive index is 0-100 points, and the classification is as follows: 0-30 points indicates mild aging, 31-70 points indicates moderate aging, and 71-100 points indicates severe aging.

7. The method for constructing a multi-dimensional assessment model for periorbital aging according to claim 1, characterized in that, Step 5) The treatment plan mapping includes quantitative recommendations for injection filling dosage, radiofrequency energy, and surgical scope, which correspond one-to-one with aging grade, dominant degeneration level, and dynamic mechanical risk hotspot; among them, the filling dosage for mild aging is 0.3-0.5 ml, and the radiofrequency energy is 30-40 joules; the filling dosage for moderate aging is 0.5-1.0 ml; and a comprehensive treatment plan is adopted for severe aging.

8. A multi-dimensional assessment model for periorbital aging, characterized in that, Constructed by the method of any one of claims 1-7, comprising the following sequentially connected components: The input layer is used to receive heterogeneous data from multiple sources. The feature layer is used to complete registration, feature extraction, and filtering. The fusion evaluation layer achieves deep fusion of multimodal features and aging quantization evaluation based on a converter-encoder-decoder network hybrid architecture; The output layer is used to output assessment reports, dynamic risk heatmaps, quantitative treatment recommendations, and aging trend prediction curves.

Images