Intelligent evaluation system for scalp health based on multi-modal imaging and data fusion

Abstract

The application discloses a scalp health intelligent evaluation system based on multi-modal imaging and data fusion, relates to the technical field of hair quality detection and analysis, and comprises multi-source heterogeneous scalp data acquisition and fusion decision, abnormal image and sensor data are corrected through high-order PDE interpolation and multi-scale filtering, multi-modal features such as hair follicle morphology, inflammatory hot spots and grease fluctuations are extracted by using CNN / LSTM and adaptive attention, a weighted decision result is generated through multi-head attention and knowledge graph fusion, and the robustness of processing of missing modes is enhanced by means of interpretability visualization and dynamic completion mechanism; after data acquisition is completed, abnormal readings can be automatically detected, and a deep prediction model is used for local interpolation, so that the accuracy and stability of multi-modal decision are improved; meanwhile, a personalized model fine-tuning is introduced, longitudinal incremental training and risk prediction are performed on multiple detection records, and a fine management scheme is provided for long-term monitoring and nursing intervention of scalp health.

Description

Technical Field

[0001] This invention relates to the field of hair quality detection and analysis technology, specifically to an intelligent scalp health assessment system based on multimodal imaging and data fusion. Background Technology

[0002] In today's scalp health management field, with the increasing demand for personal care and medical aesthetics, technologies related to scalp care and the diagnosis of corresponding symptoms are gradually gaining attention. In particular, many professional institutions are beginning to use multimodal acquisition equipment to perform visible light microscopic imaging, infrared thermal imaging monitoring, and measurements using various sensors such as humidity and sebum to more comprehensively assess hair follicle status, inflammation distribution, and sebum secretion levels. Real-world scenarios typically occur in various beauty salons, professional scalp care centers, or home self-monitoring environments: in these settings, operators use visible light microscopes to observe hair follicle morphology and infrared imaging to detect scalp surface temperature distribution to capture early symptoms of inflammation; simultaneously, they collect information on the environmental and scalp surface water-oil balance using hygrometers and sebum sensors, combining this information with the user's lifestyle habits and washing history to help determine if there are problems such as scalp dryness, excessive sebum secretion, or hair follicle blockage. Because the scalp environment is complex and easily affected by external factors, such as temperature fluctuations, varying equipment performance, and individual differences, data collection may be incomplete or abnormally high, thus affecting the reliability of the final assessment conclusion. In response, the market is calling for a technological solution that can both cope with multimodal data noise and make accurate decisions by combining prior medical knowledge, in order to meet the needs of personalized care and long-term preventive management.

[0003] A search revealed a Chinese invention patent with publication number CN119048726A, which discloses a hair feature analysis method and system. This method includes the following steps: first, capturing high-magnification and low-magnification images of a scalp region; then, preprocessing these two images. Next, inputting the preprocessed high-magnification and low-magnification images into an artificial intelligence model, simultaneously detecting hair follicles in the scalp region and calculating hair width. Afterward, at least one hair feature can be calculated from the high-magnification and low-magnification images analyzed by the artificial intelligence model. The artificial intelligence model can employ an R-CNN model or a variant thereof. The beneficial effect of this invention is that it provides more accurate and consistent results in detecting hair follicles and calculating hair width.

[0004] However, considering the existing technologies and practical application scenarios mentioned above:

[0005] In the aforementioned multimodal scalp monitoring and assessment, analyzing only a single dimension of the collected visible light / infrared images and sensor outputs such as humidity and oil levels often fails to accurately grasp the multifaceted health status of the scalp, easily leading to assessment bias or overlooking potential symptoms. For example, some infrared thermal images may show momentary abnormal spikes, but the image may not exhibit significant inflammatory characteristics. In such cases, relying solely on infrared signals can easily misinterpret normal blood circulation as folliculitis. Similarly, a low hygrometer reading may only indicate a short-term fluctuation in a dry environment; without comparative information from the image side, it's difficult to determine whether a true risk of scalp dryness exists. This not only renders the assessment results inaccurate but may also cause users to spend more time and money on care strategies, even delaying the early detection of potential diseases. Especially in professional care institutions, providing care plans based on a single data source may lead to treatment bias or the use of inappropriate products, further causing serious consequences such as scalp irritation or seborrheic alopecia. It is evident that while acquiring multi-source heterogeneous data provides more comprehensive scalp information, it also brings a high degree of complexity to data fusion and decision-making analysis. This necessitates an intelligent technology that can accurately correct noisy or missing data and leverage prior knowledge to enhance the rationality of decisions.

[0006] To this end, the present invention provides an intelligent scalp health assessment system based on multimodal imaging and data fusion. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent scalp health assessment system based on multimodal imaging and data fusion. It corrects abnormal images and sensor data through high-order PDE interpolation and multi-scale filtering, and extracts multimodal features such as hair follicle morphology, inflammation hotspots, and sebum fluctuations using CNN / LSTM and adaptive attention. Weighted decision results are generated through multi-head attention and knowledge graph fusion, and robust handling of missing modalities is enhanced by interpretable visualization and dynamic completion mechanisms. After data acquisition, it automatically detects abnormal readings and performs local interpolation using a deep prediction model, thereby improving the accuracy and stability of multimodal decision-making. Simultaneously, it introduces personalized model fine-tuning, performing longitudinal incremental training and risk prediction on multiple detection records, providing a refined management solution for long-term monitoring and care intervention of scalp health; thus solving the technical problems described in the background art.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention is implemented through the following technical solution: a scalp health intelligent assessment system based on multimodal imaging and data fusion, comprising: after detecting the completion of scalp multimodal acquisition, extracting microscopic and infrared images and sensor sequences, calling high-order PDE interpolation, multi-scale filtering and contrast enhancement, and outputting high-quality data with anomaly correction and accompanying metadata;

[0011] After receiving the corrected image and sensor time-series data, the deep network structure is read and multimodal features are generated by combining adaptive attention.

[0012] After receiving multimodal features, the fusion network reads multi-head attention and knowledge graph, uses collaborative gain metric to measure feature conflicts and calls graph matching, and outputs fusion representation and decision results;

[0013] Load Grad-CAM and saliency map, call the missing prediction model to perform missing completion and dynamic weight update, and generate image thermal visualization and time series deviation description;

[0014] The system identifies multiple user detection records and categorizes them into a long-term profile vector. It reads user fine-tuning parameters and embeds them into a knowledge graph constraint function. Using incremental training and lifestyle habit mapping, it outputs longitudinal prediction and intervention plans.

[0015] Furthermore, visible light microscopic image data, infrared thermal image data, and sensor data are collected; a meta-information tag set is established for each set of collected data, the tags are bound and stored with the data body, an integrated data management module is constructed, and a unique identifier is assigned to each piece of collected data.

[0016] Furthermore, the moments when there are abrupt changes or extreme readings in the sensor data are marked, and anomaly masks are established for significant noise pixels in visible light microscopic image data and infrared thermal image data.

[0017] After preprocessing, the corrected visible light microscopic image, corrected infrared thermal image, and corrected sensor data are written back to the data management module.

[0018] Furthermore, the corrected visible light microscopic image and the corrected infrared thermal image are fed into an image processing pipeline consisting of multi-layer convolution operators and nonlinear activations; multi-scale kernel combinations are introduced in each convolution layer and fused in the output feature map; the feature map obtained from the last convolution layer is globally pooled to obtain the image modality feature vector.

[0019] Furthermore, multi-channel sensor data is extracted from the corrected sensor sequence and input into the timing processing pipeline to capture the dynamic patterns of each channel in the time dimension.

[0020] An improved gated memory update function is adopted, combined with an exponential decay mechanism to enhance the parallel modeling of long-term trends and short-term fluctuations. After defining the update rules and iterating, the hidden state at the last moment is taken as the temporal modality feature vector.

[0021] Furthermore, the image modal feature vector and the temporal modal feature vector are concatenated or concatenated to obtain a preliminary combined vector. In multi-head mode, attention outputs are calculated separately after different heads are introduced, and the results are concatenated or weighted and merged to output a preliminary fused vector.

[0022] Furthermore, based on the modal feature information, potential related nodes and corresponding edges are retrieved in the pre-constructed scalp health knowledge graph. The potential symptom features learned in the network are matched with the medical knowledge in the graph at the keyword or vector level. The set of symptom nodes with the highest correlation to the preliminary fusion vector is found in the scalp health knowledge graph.

[0023] The matched nodes and their edges are converted into graph enhancement vectors, which are then concatenated or weighted with the initial fusion vectors to obtain the fusion representation.

[0024] Furthermore, the fused representation is input into the final decision module, which consists of a multilayer perceptron or an attention-based classification head. The decision module verifies the rationality of the reasoning results based on the graph prior information embedded in the graph enhancement vector.

[0025] Furthermore, a collaborative gain metric is introduced for conflict judgment and decision correction. If the collaborative gain metric is not greater than the corresponding threshold, the conflict handling logic is triggered, and the conflict handling or secondary verification logic is initiated; otherwise, the decision result output by the decision module is directly regarded as the scalp health assessment conclusion.

[0026] Furthermore, based on the decision results and fusion representation, the image channels are located and highlighted. After projecting the key areas of network interest onto the original image, the resulting attention distribution map is superimposed and rendered with the corrected visible light microscopic image or the corrected infrared thermal image.

[0027] On the output of the time-series sensor, by marking key points and distinguishing intervals, the peaks or valleys that may lead to a certain conclusion in diagnosis are marked significantly. Combined with the threshold information in the scalp health knowledge graph, the UI will indicate the high-risk interval for oily skin.

[0028] Furthermore, after dynamically updating the attention parameters, they are written to the system configuration library. If a path is missing or abnormal during subsequent use, the historical feature sequence of the temporal modality feature vector and the channel information in the image modality feature vector are called to supplement the missing data through the missing prediction model.

[0029] If the prediction reliability does not reach the threshold, the user will be prompted to re-collect data or perform manual verification.

[0030] The adjustment parameters, completion behavior, and user feedback information are written into the corresponding database. If the user confirms that the completion result is accurate, the positive feedback on the missing prediction model parameters is strengthened.

[0031] Furthermore, heatmaps are overlaid on the image area to mark key areas such as scalp inflammation and sebum buildup; key peaks and intervals are highlighted in the time-series curve; and current attention parameters and model prediction accuracy are displayed in pop-up windows or charts.

[0032] The model can be continuously iterated and evolved by globally retraining or periodically iterating on the missing prediction model and attention parameters.

[0033] If a model or modality performs poorly in a real-world scenario for an extended period, an automatic notification will be sent to the development or maintenance team, and an alert will be issued to external parties.

[0034] Furthermore, the corrected visible light microscopic images, corrected infrared thermal images, and corrected sensor data obtained by users at different time periods are collected and archived together with decision results, interpretability outputs, and user feedback records to establish a long-term archive vector, which also includes users' lifestyle habits and dietary preferences.

[0035] Based on the data management module and scalp health knowledge graph, a hierarchical index is created for the image modality feature vector, temporal modality feature vector, and fusion representation of each detection in the long-term archive vector.

[0036] Furthermore, when a user records multiple tests within a certain period, or when symptom indicators fluctuate significantly in a short period of time, a personalized training process is triggered, and a personalized model is obtained after fine-tuning.

[0037] When detecting input corresponding to the user, the personalized model is prioritized for inference. If a large range of modalities are missing or outliers occur, the system automatically reverts to the global base model.

[0038] Furthermore, based on long-term archival vectors and personalized models, scalp health trends are periodically assessed:

[0039] If the hair quality abnormality index shows signs of rising in the future assessment period, the corresponding risk factors will be matched according to the cause-symptom chain in the scalp health knowledge graph, and targeted intervention or care suggestions will be made to the user; when the assessment period is short and the risk level of the predicted result is high, an immediate reminder will be triggered.

[0040] (III) Beneficial Effects

[0041] This invention provides an intelligent scalp health assessment system based on multimodal imaging and data fusion, which has the following beneficial effects:

[0042] Establishing a unified acquisition protocol and data management framework, and using high-order PDE interpolation and multi-scale wavelets to adaptively correct abnormal images and sensor data, can ensure data integrity and consistency, reduce noise interference, and improve the usability of subsequent models.

[0043] Construct image modality feature vectors for image and temporal features respectively. With modal eigenvectors By employing a learnable attention mechanism that balances fine-grained and global features, the model leverages multi-head attention and a scalp health knowledge graph to address issues such as scalp inflammation and abnormal sebum production. Deep fusion, generating fusion representation Furthermore, by leveraging collaborative gain metrics to detect multimodal conflicts or missing information, the accuracy of decision-making can be ensured while rapidly identifying potential anomalies. Knowledge graph priors can constrain personalized models through functions, maintaining a balance between medical common sense and individual differences.

[0044] At the interpretability level, tools such as Grad-CAM are used to analyze the corrected visible light micrographs. And corrected infrared thermal image The area of ​​interest is highlighted and visualized, and the deviation of sensor readings is displayed, so that users and professionals can understand the basis of the system's decision-making.

[0045] Simultaneously using the missing prediction model Addressing modal missingness and implementing dynamic weights Adaptive optimization builds a feedback loop, ensuring that adjustments are made at any time based on actual usage and user feedback.

[0046] In terms of population personalization and long-term tracking, long-term archival vectors are utilized. With personalized models Incremental fine-tuning, combined with external information such as diet and lifestyle habits, allows for vertical management. Once long-term trends in inflammation or sebum fluctuations are identified, a scalp health knowledge graph can be used to address these issues. Based on the internal trigger-symptom pathway, preventive interventions and nursing recommendations are proposed. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the scalp health intelligent assessment system based on multimodal imaging and data fusion according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 This invention provides an intelligent scalp health assessment system based on multimodal imaging and data fusion, comprising:

[0050] Step 1: After completing the multimodal scalp data acquisition, the data management framework extracts microscopic images. With infrared image and sensor sequence It also invokes high-order PDE interpolation, multi-scale filtering, and contrast enhancement to output a corrected visible light micrograph with anomaly correction and associated metadata labels. Corrected infrared thermal image and corrected sensor data ;

[0051] Step one includes the following:

[0052] Step 101: Unified and synchronized multimodal data acquisition and metadata annotation:

[0053] Acquisition of visible light microscopic image data Infrared thermal imaging data And data from sensors such as humidity and grease. Among them, visible light microscopic image data With infrared thermal imaging data Using a uniform resolution and frame rate, and comparing it with sensor data over time. Maintain synchronized data acquisition; use a unified acquisition clock. All modes were subjected to baseline correction to ensure that visible light microscopic image data were acquired at the same time stamp. Infrared thermal imaging data And data from sensors such as humidity and grease. It can represent the same scalp condition;

[0054] A metadata tag set is established for each set of collected data, and the tags are bound and stored with the data ontology. An integrated data management module is constructed, and a unique identifier is assigned to each piece of collected data to facilitate backtracking or indexing in the feature extraction and deep fusion stages.

[0055] When using it, a unified acquisition protocol and clock are employed. This ensures that multimodal data can correspond one-to-one over time; this is crucial for extracting cross-modal correlation features in subsequent steps. With the help of the Data Management Module (DMG) and a complete set of metadata tags, subsequent steps (such as multimodal feature extraction in step two) can quickly locate and match visible light microscopic image data from the same time point. Infrared thermal imaging data And data from sensors such as humidity and grease. This lays the foundation for data management in order to achieve deep integration.

[0056] Step 102, Adaptive Cleaning and Higher-Order Interpolation Completion:

[0057] Extract visible light microscopic image data from the same acquisition session from the data management module using metadata tag IDs. Infrared thermal imaging data and sensor data ;

[0058] Sensor data The moments of abrupt changes or extreme readings were marked, and the visible light microscopic image data were annotated. Infrared thermal imaging data Anomaly masks are created for significant noise pixels (such as overly dark or overly bright areas) that appear in the image. , anomaly mask Used to determine the range required for subsequent interpolation or correction, ensuring that operations are only performed on suspicious or missing areas without harming the overall information;

[0059] For visible light microscopic images For locally missing or extreme pixels, an image completion model based on partial differential equations (PDEs) is used for correction:

[0060]

[0061] in: Indicating evolutionary time The intermediate correction image below, Define the domain for the image;

[0062] This refers to the failure or noise region corresponding to the anomaly mask.

[0063] It is the correction rate coefficient (the range can be ). (The specific value can be optimized according to the level of image noise). and These represent the gradient and divergence operators, respectively, used to spatially smooth and progressively repair anomalous blocks;

[0064] Evolution time is updated iteratively. The intermediate correction image below ,exist Corrected visible light micrographs were obtained at that time. And complete the effective filling of locally missing or abnormal pixels;

[0065] For sensor data and infrared thermal imaging data For temporal or local anomalies present in the data, multi-scale wavelet transform is used for high-frequency noise suppression and missing data interpolation, where:

[0066] Sensor data Multi-scale wavelet decomposition is performed in the time dimension to extract high-frequency detail coefficients and low-frequency trend coefficients. The high-frequency detail coefficients corresponding to outlier segments are attenuated to remove resonance noise. Interpolation basis functions are used to locally reconstruct the low-frequency coefficients of the missing segments, and then an inverse transform is performed between these reconstructed coefficients and the attenuated high-frequency coefficients to obtain the completed sensor data. Similarly, infrared images Alternatively, abnormal hotspot regions can be located by performing wavelet decomposition on the rows and columns separately, and then performing multi-scale coefficient correction and reconstruction on these regions to obtain the corrected results. ;

[0067] Finally, the obtained corrected visible light micrographs Corrected infrared thermal image and corrected sensor data Write it back to the data management module, keeping its ID unchanged and only updating the status label for use in subsequent steps;

[0068] When using it, use the exception mask. The definition of PDE interpolation and wavelet transform correction is only applied to areas with anomalies or missing information, which reduces the adverse effects of over-repair on the original effective information. The combination of partial differential equations (PDE) and multi-scale wavelets can perform high-precision completion on the acquired images and time-series signals respectively, and has strong adaptive capabilities. On the one hand, PDE can smoothly repair the missing areas in the spatial dimension, and on the other hand, wavelets can suppress noise and reconstruct key waveforms in the time or frequency scale.

[0069] Step 2: Upon receiving the corrected data, the feature extraction module utilizes deep structures such as CNN and LSTM, and injects adaptive attention. Generate image modal feature vectors that combine spatial texture and temporal dynamics. and temporal modal feature vectors ;

[0070] Step two includes the following:

[0071] Step 201: Depth feature extraction of image modalities:

[0072] The corrected visible light micrograph output from step one And corrected infrared thermal image The data is uniformly fed into an image processing pipeline consisting of multi-layer convolution operators and non-linear activations. ;

[0073] Unlike traditional models with fixed kernel sizes, this model introduces multi-scale kernel combinations in each convolutional layer. The features are then fused in the output feature map to accommodate the spatial distribution of scalp features of various sizes (such as hair follicles, inflammatory spots, etc.).

[0074] Define image feature mapping For the first Layer output:

[0075]

[0076] in: This indicates an input image from visible light or infrared light. Indicates the convolution operation; This indicates a Hadamard multiplication per channel; Learnable attention gating coefficients (range of values) ), used to dynamically allocate weights in the output of multi-scale convolution kernels; It is a non-linear activation function (higher-order activation forms such as Swish or GELU can be selected);

[0077]

[0078] Through learnable attention gating coefficients The network can adaptively amplify or suppress convolutional channels of different scales, thereby better capturing multiple forms such as minor inflammation and large-area scalp damage;

[0079] The feature maps obtained from the last convolutional layer are then globally pooled to obtain the image feature vector. ,in The image feature dimension is denoted as and the final image modal feature vector is denoted as . ,in and The aggregation results correspond to the visible light / microscopic channel and the infrared channel, respectively;

[0080] When in use, by introducing multi-scale kernel combination and learnable attention gating, the network can extract more refined features for scalp symptoms of different shapes and sizes, improving the ability to perceive hair follicle blockage and inflammation distribution.

[0081] Compared to traditional single-kernel convolution methods, this approach significantly enhances the simultaneous detection capability of local lesions and global abnormalities, and can carry more expressive image representations in subsequent fusion networks; multi-scale kernel combination and learnable gating coefficients are also included. The joint design enables the network to dynamically adjust its dependence on convolution channels of different scales, breaking through the bottleneck of the previous fixed scale that could not take into account both the microscopic hair follicle structure and the macroscopic scalp texture.

[0082] Step 202: Deep feature extraction of temporal modalities:

[0083] The corrected sensor sequence output from step one In the process, humidity data is extracted. Data on oils Different sensor channels are uniformly represented as multi-channel sensor data. ,in Number of sensor channels The sequence length is given.

[0084] Multi-channel sensor data Input to timing processing pipeline It is used to capture the dynamic patterns of each channel in the time dimension;

[0085] An improved gated memory update function is employed, combined with an exponential decay mechanism, to enhance the parallel modeling of long-term trends and short-term fluctuations; let and These are the hidden state and memory unit state from the previous time step, respectively. The input at the current time step is... The update rules are defined as follows:

[0086] ,

[0087]

[0088]

[0089] in: For optional higher-order activation functions (such as SiLU or GELU); For the natural exponential function, an adaptive memory decay or memory enhancement mechanism is introduced for the state of the previous time step; , , , The matrix is ​​a learnable parameter matrix (its range can be converged through training iterations after random initialization). To update the gate vector, For exponential memory gate vectors, through... Element-wise multiplication allows the network to flexibly retain or attenuate historical information;

[0090] go through After iteration, take the last time step. As a feature vector of temporal modes ,in This represents the temporal feature dimension; if multiple sensor sub-sequences exist (such as humidity, grease, and temperature), the final state vectors of each sub-sequence can be concatenated into a temporal composite vector. Specifically, image modal feature vectors and temporal modal feature vectors The output is an integrated output of images and sensor sequences, which includes in-depth characterization of core elements such as scalp inflammation, hair follicle blockage, and sebum fluctuations.

[0091] In use, compared to traditional LSTM or GRU structures, it utilizes exponential memory gates. The system achieves weighted adjustment of historical states, which can prevent the disappearance of long-term dependencies and quickly complete state switching when there are sudden data fluctuations. The combination of multi-channel decomposition strategy and adaptive gating memory enables the network to capture key fluctuations of a single channel while taking into account the coupling relationship between channels when facing multiple types of sensors such as humidity and oil. This provides rich temporal clues for scalp health assessment. By introducing an exponential memory gate in the recurrent structure, the network can adaptively amplify or decay the hidden state and memory unit of the previous moment, avoiding the rigid processing of long-term and short-term information by traditional temporal networks and achieving a more flexible temporal learning capability.

[0092] Step 3: Obtain image modal feature vectors and temporal modal feature vectors Subsequently, the fusion network employs multi-head attention and graph embedding. and utilize the synergistic gain metric Determine multimodal collaborative consistency and output a fusion representation integrating image and temporal features. and decision results ;

[0093] Step three includes the following:

[0094] Step 301: Multi-head attention alignment and initial fusion:

[0095] The image modal feature vector output in step two With temporal modal feature vectors Perform concatenation or cascading operations to obtain preliminary combined vectors. ,in:

[0096]

[0097] First map to query ,key Sum Vector space for multi-head attention fusion:

[0098]

[0099] , and These are the learnable projection matrices, and their range can be set according to the network structure (e.g., ...). ); , , All in Abstract features representing multimodalities in space;

[0100] Under single-head attention, the attention output can be expressed as:

[0101]

[0102] Softmax is a commonly used normalized activation function used to map any real vector to a function whose elements sum to 1 and whose values ​​are all equal to 1. The probability distribution between, where This is a learnable amplification factor (which can be used to amplify or smooth attention scores), with a value range of... ;

[0103] In multi-head mode, different heads are introduced. Calculate attention output separately And then concatenate or weighted merge the results;

[0104]

[0105] in, Indicates vector concatenation. For learnable head weights, the final output is This is the initial fusion vector;

[0106] In practice, by initially concatenating the image with temporal features and applying multi-head attention, adaptive alignment between multimodal features is achieved. When certain modalities contribute more to the current context, the attention score automatically increases, and vice versa. Compared to simple feature concatenation or weighted averaging, multi-head attention can capture finer-grained correlations between channels, while also incorporating learnable amplification coefficients. With head weight The optimal fusion mode can be dynamically determined during training; by introducing multi-head attention after cross-modal splicing, it can not only handle dimensionality asymmetry. The problem can also be solved by automatically adjusting the focus on image features and temporal features within the same network structure.

[0107] Step 302, Knowledge Graph-Assisted Decision Mapping:

[0108] Build a scalp health knowledge graph in advance The nodes include various scalp symptoms and potential causes (such as abnormal oil production, clogged hair follicles, scalp allergies, etc.), and the edges represent the association, intensity, or temporal relationship between symptoms and causes.

[0109] Based on the modal characteristic information obtained in step two (such as high temperature hotspots or low humidity, visible light / microscopic image modality, infrared thermal imaging modality, environmental / scalp sensor modality, and user feedback / questionnaire modality, etc.), a scalp health knowledge graph can be created. The search retrieved potential related nodes and their corresponding sets of edges. The retrieval process is mainly based on graph indexing and concept mapping algorithms, which match potential symptom features learned in the network with medical knowledge in the graph at the keyword or vector level.

[0110] For the initial fusion vector In the scalp health knowledge graph Find the set of symptom nodes that are most closely related to it. And through the following mapping function accomplish:

[0111]

[0112] in: Represents a node Vectorized representation in knowledge graphs (obtained by training a graph embedding model, with the value range dynamically adjusted according to the graph size):

[0113] The mapping relevance calculation function (not a simple inner product or cosine similarity) can be implemented based on higher-order projection or curve fitting, and is used to measure the initial fusion vector. With node vectors The degree of matching between them, where:

[0114]

[0115] in: For multi-head attention fusion vectors, Embedded into knowledge graph nodes; , , The learnable weight matrix is ​​applied to... , and their element-wise products ;

[0116] It is the bias vector; Hyperbolic tangent activation is used to introduce nonlinearity; The projection vector maps the hidden layer output to a scalar.

[0117] Ensure final relevance It is positive and differentiable.

[0118] Matched nodes and its connecting edges Converted into graph enhancement vectors and the initial fusion vector Perform splicing or weighted fusion to obtain the fusion representation. :

[0119]

[0120] in, It can be a vector concatenation or higher-order combination operation, fully injecting prior knowledge of the graph into the model representation.

[0121] In practice, by introducing knowledge graph mapping after multi-head attention fusion, medical prior knowledge is superimposed on data-driven features, which helps to make reasonable inferences even in cases of symptom conflict or modality loss; fusion representation After merging the initial fusion vector Feature information and map enhancement vector After associating knowledge, the interpretability and stability of subsequent decision-making can be significantly improved; by employing higher-order mapping relevance... Matching graph nodes can avoid the limitations of simple inner product or cosine similarity, and can continuously iterate and optimize the fit between graph embedding and feature fusion during model training.

[0122] Step 303, Comprehensive Reasoning and Conflict Resolution:

[0123] Fusion representation Input to the final decision module Decision module Composed of a multilayer perceptron or an attention-based classification head, it is used to perform qualitative or quantitative assessments of scalp health (such as determining whether there is excessive oil, dry scalp, or localized folliculitis).

[0124] Meanwhile, the decision-making module Vector enhancement based on the graph The built-in prior information of the graph is used to verify the rationality of the reasoning results. For example, when there is a contradiction between obvious abnormality in infrared thermal image and severely low humidity, the probability can be assessed by the inflammation model or dryness symptom node in the graph, and then the conflicting results can be balanced or iteratively corrected.

[0125] For conflict assessment and decision correction, a collaborative gain metric is introduced. This can be understood as a consistency index of fused information (the larger the value, the more supportive the modalities and the knowledge graph are to each other; the smaller the value, the more likely there is a conflict). Assuming the current decision outcome is... It can be defined as follows:

[0126]

[0127] in: Multi-scale or multi-stage mapping functions are used to project vector inputs onto different resolutions or feature subspaces. These functions can be wavelet transforms, morphological multi-scale analysis, or other operators that reflect local-to-global differences for the same input. In other words, It can generate multiple different feature representations as it changes from 0;

[0128] It is an exponential decay factor. ; For continuous parameters in the multi-scale mapping process, ; These are reference symptom vectors that can be obtained from the atlas, corresponding to known pathological patterns;

[0129] The selected higher-order norm form ( (To avoid being similar to the conventional L1 or L2 norm).

[0130] like This indicates that the current fusion result deviates significantly from the reference map, requiring the triggering of conflict resolution logic, such as readjusting the initial fusion vector. It may prompt for modality missing completion; conflict handling or secondary verification logic can be initiated, such as prompting that data needs to be collected again or manually reviewed.

[0131] like This indicates that the fusion representation It shows good consistency with the knowledge graph, and the decision-making module... Output decision results This can be directly regarded as a scalp health assessment conclusion; if the multimodal conflict is still not resolved, the result will be marked as pending review and accompanied by explanatory information (such as contradiction between infrared thermal imaging and humidity sensor, it is recommended to re-collect data or further manual verification), and the user or professional will be prompted through a visual interface;

[0132] When using it, combine it with fusion representation With decision-making module The system incorporates prior medical knowledge during automated evaluation and provides timely alerts or corrections when multimodal conflicts occur, along with synergistic gain measurement. This study provides a quantitative method for detecting conflicts and consistency, enabling the fusion network not only to output results but also to identify when supplementary data collection or manual verification is needed, significantly improving the system's security and reliability. On one hand, the synergy between multi-head attention mechanisms and knowledge graphs enhances reasoning accuracy and transparency when dealing with complex scalp symptoms; on the other hand, synergistic gain measurement and conflict handling ensure high robustness even in cases of modality loss or contradictory information. By combining deep fusion with prior medical knowledge, the limitations of single-modality or simple voting fusion are significantly overcome, laying the foundation for core decision-making capabilities in building an intelligent and reliable scalp health assessment system.

[0133] Step 4: When the fusion network outputs the decision results and fusion characterization It can interpret the module's calls to Grad-CAM and time-series visualization, and leverage the missing data prediction model. Dynamically complete missing modes, generate visual thermal and index deviation reports, and promptly correct anomalies;

[0134] Step four includes the following:

[0135] Step 401, Multimodal Visualization and Interpretable Presentation:

[0136] Based on the decision results obtained in step three and fusion characterization In image channels (such as corrected visible light micrographs) And corrected infrared thermal image To locate and highlight regions of interest on the original image, interpretable methods based on gradient activation (such as Grad-CAM) or techniques based on saliency maps can be used.

[0137]

[0138] in: To be at the image location With fusion characterization Associated activation mapping function; These are trainable convolutional or attention weights; A function that maps gradients to a visual heatmap (may include normalization or thresholding operations). This can be viewed as a distribution map of attention;

[0139] Attention distribution map Compared with the corrected visible light micrograph Or corrected infrared thermal image Overlay rendering allows users to visually see specific areas with potential abnormalities (such as clogged hair follicles or spreading inflammation);

[0140] On the outputs of time-series sensors such as humidity and oil sensors, peaks or troughs that might lead to a biased diagnosis are clearly marked by key point annotations and interval differentiation. For example, when the oil sensor reading is abnormally high during a specific time period, it is highlighted in a visual curve with a striking color or symbol. Combined with threshold information from the scalp health knowledge graph (such as exceeding a certain value may indicate excessive oil secretion), the UI displays high-risk oil zones, allowing users or professionals to quickly understand why the system diagnoses oil-related conditions.

[0141] In practice, by explicitly showing which regions or segments the network assigns high weight to at the image and sensor levels, users can intuitively verify the rationality of the system's judgments. This process, combined with prior information from the knowledge graph (such as certain numerical ranges corresponding to specific symptoms), provides a dual-level explanation of medical and data aspects. This increases user confidence in the model's output and contributes to a positive feedback loop in subsequent stages. Furthermore, it extends the traditional Grad-CAM approach to multimodal scenarios, meaning it can handle not only image convolutional features but also fused representations. Visualize time-series features to cover a wider range of interpretations.

[0142] Step 402, Learning Dynamic Weight Self-Adjustment and Data Completion:

[0143] The head weights used in multi-head attention fusion are dynamically updated based on user and expert feedback. and the growth rate coefficient This allows for a better match to the actual usage environment in the next analysis. For example, if users repeatedly report that a certain modality has a lot of noise, the initial weight of the corresponding modality in multi-head attention can be reduced to increase the influence of other modalities. Specific methods include incremental learning and adaptive adjustment, feedback correction mechanism, expert-guided fine-tuning, and incremental learning and adaptive adjustment.

[0144] Updated attention parameters This information is written to the system configuration library for use in the next round of multimodal fusion, forming an adaptive evolution mechanism based on usage feedback. If a certain path (such as infrared thermal imaging or grease sensor) is missing or abnormal in subsequent use, the previously mentioned time-series modal feature vector will be invoked. Historical feature sequences and image modal feature vectors The relevant channel information is used to predict missing information using a specialized model. Estimate and supplement missing data :

[0145]

[0146] in: It is a function that can be trained based on graph neural networks or sequence generation models, used to infer missing modes from existing modal features;

[0147] If the prediction confidence level does not reach the threshold, the user is prompted to re-collect data or perform manual verification; for example, the prediction confidence level can be evaluated by the maximum class probability or regression error output by the model.

[0148] The adjustment parameters, completion actions, and user feedback information generated during this phase will be written into the corresponding database for further optimization in subsequent version iterations. If users confirm the accuracy of the completion results, the missing data prediction model will be strengthened. The model parameters provide positive feedback, while negative feedback penalizes and encourages further exploration and updates.

[0149] When in use, by combining interpretable output with actual user / expert feedback, it continuously optimizes itself during use, reducing the accuracy drop caused by fixed weights or missing modalities; it directly exposes the internal weights of multi-head attention and allows for manual or automatic adjustment, forming a human-machine collaborative optimization mechanism, rather than the traditional black-box self-learning approach, introducing cross-modal inference ideas; when a certain signal is missing, it can dynamically use image and other sensor information to complete it with high reliability.

[0150] Step 403, Visual Interaction and Closed-Loop Management:

[0151] The output results and dynamic update strategy of steps 401 and 402 are integrated to provide an interactive visualization interface, which includes: overlaying heat maps in the image area to mark key areas such as scalp inflammation and sebum accumulation; highlighting key peaks and intervals in the time series curve; and displaying in-depth information such as the system's current attention parameters and the accuracy of the supplementary model prediction in the form of pop-ups or charts, so that experts can review and correct the entire process.

[0152] After multiple uses by users, the system will accumulate enough feedback data and completion records, which will then be used to improve the missing data prediction model. and attention parameters , Perform global retraining or periodic iterations to continuously evolve;

[0153] If a model (such as a multimodal fusion model, a time-series processing model, etc.) or modality (such as sensor data, image data, etc.) performs poorly in real-world scenarios for an extended period, an automatic notification will be sent to the R&D personnel or maintenance team. Hardware replacement or algorithm upgrades can then be performed to maintain the long-term high reliability of the scalp health assessment platform.

[0154] In use, through visual interaction and closed-loop management, the system not only possesses interpretability and adaptive completion capabilities during single use, but also continuously upgrades itself based on actual usage over a long period, providing users and professionals with transparent and controllable diagnostic tools. Through multimodal visual interpretation in step 401, dynamic weight self-adjustment and data completion in step 402, and the interactive interface and closed-loop management in step 403, a complete feedback path of self-learning, human-computer collaboration, and continuous iteration is constructed while ensuring the transparency and credibility of the diagnostic process. Combined with the multimodal fusion and knowledge graph-assisted decision-making mentioned earlier, a rigorous yet flexible chain is formed from data acquisition to decision output to interpretability and adaptive optimization. Ultimately, users and professionals can not only obtain accurate and understandable scalp health assessments, but also participate in the continuous improvement of the model in real time, providing higher credibility and practical value for intelligent scalp management.

[0155] Step 5: When the user's multiple test results are accumulated into the long-term archive vector In fine-tuning parameters The system introduces knowledge graph constraint functions and combines them with information on lifestyle habits and nursing cycles to output longitudinal individualized intervention and risk prediction plans, thereby intervening in lipid imbalance or inflammatory trends in advance and continuously optimizing individualized nursing pathways.

[0156] Step five includes the following:

[0157] Step 501: Establishment of Individual Historical Records and Cross-Time Period Data Management:

[0158] Collect corrected visible light microscopic images obtained by the user at different time periods (corresponding to multiple acquisition batches in step one). And corrected infrared thermal image Compared with the corrected sensor data Compare it with the decision results generated in step four. Interpretable outputs (such as saliency plots and abnormal oil ranges) and user feedback records should be archived uniformly and a system named […]. Long-term archive vector, long-term archive vector It also includes fields such as user lifestyle habits and dietary preferences;

[0159] Based on data management module and scalp health knowledge graph For long-term archive vectors Image modality feature vectors detected in each step Temporal modal feature vectors and fusion characterization Create a hierarchical index:

[0160] The first layer is the timeline (reflecting the user's longitudinal evolution);

[0161] The second layer consists of symptom categories or focal points (linking to nodes in the knowledge graph such as hair follicle blockage and sebum overflow).

[0162] The third layer consists of personalized tags (such as user's diet and care habits).

[0163] This hierarchical index allows for the rapid retrieval of a target user's long-term performance over a specific time period or under a particular symptom, laying the foundation for data management for personalized decision-making.

[0164] When in use, the longitudinal information from multiple tests is fully utilized to construct an individual-level archive, providing reliable data support for subsequent personalized model fine-tuning and preventive intervention. The combination of hierarchical index and knowledge graph can quickly link symptoms, time periods, and user habits, and can quickly compare historical records when the system makes conflicting judgments or issues with warnings. The user's multimodal characteristics and diagnostic results are stored in an integrated manner with external information such as lifestyle. Through the linkage between hierarchical index and knowledge graph, the system can more flexibly retrieve and analyze individual differences, rather than being limited to a single data collection.

[0165] Step 502: Personalized Model Fine-tuning and Incremental Training

[0166] When a user accumulates more than M test records within a certain period, or when symptom indicators fluctuate significantly in a short period of time (such as an abnormal increase in lipid levels), the personalized training process is triggered, as follows:

[0167] From long-term archive vectors Extract the most recent L high-quality samples and combine them with the global base model. Fine-tune the training (for the network weights used in steps two and three);

[0168] Let the current user's fine-tuning parameters be... In the original loss function An additional knowledge graph consistency term is added to prevent overfitting of personal data:

[0169]

[0170] in: The learning rate (which can be automatically adjusted and has a range of values ​​within a certain range) (Select from within);

[0171] This is the knowledge constraint coefficient. This allows personalized models to stay aligned with the scalp health knowledge graph. A reasonable connection;

[0172] This is a bias penalty function based on the atlas prior (the penalty is increased if the personalized model's inference results differ significantly from common pathological pathways).

[0173] After fine-tuning, The personalized model version marked for the current user (e.g.) Indicates the first (Next update), stored in the system's model management module;

[0174] In subsequent analysis, when input from this user is detected, the personalized model is prioritized. During inference, if a large range of modalities are missing or outliers occur, the system will automatically revert to the global base model. ;

[0175] In use, without altering the overall model structure, incremental training and knowledge graph consistency constraints enable the system to adapt to individual differences without deviating from medical common sense. This supports segmented populations or special cases (such as users with sensitive scalps), making diagnostic results closer to the individual's true condition. Personalized training is combined with the knowledge graph through… The function introduces constraints to make the model take into account user differences and medical priors, avoiding extreme fitting or erroneous learning during personalized parameter tuning.

[0176] Step 503: Preventive intervention and full-cycle monitoring:

[0177] In long-term archive vectors and personalized models Based on this, periodically assess scalp health trends:

[0178] If indicators such as inflammation and abnormal lipid levels are predicted in future assessment cycles If there are signs of an increase, then refer to the scalp health knowledge graph. The trigger-symptom chain is matched with corresponding risk factors, and targeted interventions or care suggestions are made to users, such as adjusting the frequency of washing and care or using specific products;

[0179] When the evaluation cycle When the prediction result is short and the risk level is high, an immediate reminder is triggered, suggesting that the user conduct a retest or seek professional diagnosis as soon as possible; the risk level can be obtained by comprehensively evaluating factors such as prediction reliability, data quality, and environmental changes after training the neural network.

[0180] Combining the dynamic feedback mechanism in step four, a retrospective analysis is conducted on the periods when users experience abnormalities to identify the key causes or lifestyle habits that cause the abnormalities. Multi-level alarm rules (such as low-risk alerts, medium-level alarms, and high-risk warnings) are set according to the persistence or frequency of the abnormalities, and the alarm information is synchronized to the backend database to further improve the personalized profile.

[0181] In use, by building prediction and intervention modules at the individual level, the system can not only perform diagnosis after a single test, but also proactively manage users' scalp health in the medium to long term. The risk grading mechanism allows users and professionals to understand the evolution of their own or group's scalp condition in a timely manner, taking early intervention measures to prevent small problems from developing into major risks. This is achieved by utilizing a scalp health knowledge graph. By identifying the correlation between triggers and symptoms, and linking personalized model predictions with prior medical knowledge, intervention plans can be proactively pushed out, transforming the system from passive diagnosis to proactive protection.

[0182] Through individual historical profile construction in step 501, personalized model fine-tuning in step 502, and preventative intervention and full-cycle monitoring in step 503, the solution not only achieves single-time detection and analysis of scalp health, but also provides the ability for continuous evolution and precise care in a vertical dimension. At the data level, it collects and indexes user information from multiple time periods; at the model level, it performs incremental training and introduces knowledge constraints; and at the decision-making level, it combines graph priors for early intervention. This allows the system to advance from collection-analysis-interpretation-feedback to a comprehensive health management framework that is personalized, long-term, and preventative. Ultimately, users can obtain scalp health assessments and intervention plans that are closer to their own physical condition and habits, enabling the intelligent system to play a greater practical role and provide more humanistic care in the field of scalp care.

[0183] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0184] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0185] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0186] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0187] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A scalp health intelligent assessment system based on multimodal imaging and data fusion, characterized in that: include, After the scalp multimodal acquisition is completed, microscopic and infrared images and sensor sequences are extracted. Interpolation, multi-scale filtering and contrast enhancement are performed on the high-order PDE. High-quality data with anomaly correction and metadata are output, which are denoted as the corrected image and the corrected time series, respectively. After receiving the corrected image and sensor time-series data, the corrected image and time-series data are input into a deep network, and adaptive attention is used to obtain image modality feature vectors and time-series modality feature vectors, respectively. Specifically, the process includes: feeding the corrected visible light microscopic image and the corrected infrared thermal image into an image processing pipeline composed of multi-layer convolution operators and nonlinear activation; introducing multi-scale kernel combinations in each convolution layer and fusing them in the output feature map; performing global pooling on the feature map obtained from the last convolution layer to obtain the image modality feature vector; extracting multi-channel sensor data from the corrected sensor sequence and inputting it into the time-series processing pipeline to capture the dynamic patterns of each channel in the time dimension; using an improved gated memory update function, combined with an exponential decay mechanism to enhance the parallel modeling of long-term trends and short-term fluctuations; defining update rules and iterating; and taking the hidden state at the last moment as the time-series modality feature vector. Using the image modal feature vector and temporal modal feature vector as input, the fusion network reads multi-head attention and knowledge graph, uses collaborative gain metric to measure feature conflicts and calls graph matching to output a fused representation; the fused representation is then input into the decision module to obtain the decision result; specifically, this includes the following: concatenating or cascading the image modal feature vector and temporal modal feature vector to obtain a preliminary combined vector; in multi-head mode, after introducing different heads, calculating the attention output separately and concatenating or weighting the results to output a preliminary fused vector; based on the feature information of each modality, and within the pre-constructed scalp health... Potentially relevant nodes and corresponding edges are retrieved from the knowledge graph. The potential symptom features learned from the network are matched with medical knowledge in the graph using keywords or vector-level matching. The set of symptom nodes most relevant to the initial fusion vector is found in the scalp health knowledge graph. The matched nodes and their edges are transformed into graph augmentation vectors, which are then concatenated or weighted with the initial fusion vector to obtain a fusion representation. This fusion representation is input into a final decision module composed of a multilayer perceptron or an attention-based classification head. The decision module verifies the reasonableness of the inference results based on the graph prior information embedded in the graph augmentation vector. Furthermore, a collaborative gain metric is introduced for conflict judgment and decision correction. If the collaborative gain metric is not greater than the corresponding threshold, the conflict handling logic is triggered, and the conflict handling or secondary verification logic is initiated. Conversely, the decision result output by the decision module is directly regarded as the scalp health assessment conclusion; Based on the decision results and fusion representation, localization and highlighting are performed on the image channel. After projecting the key areas of network attention onto the original image, the resulting attention distribution map is superimposed and rendered with the corrected visible light microscopic image or the corrected infrared thermographic image. On the temporal sensor output, through key point annotation and interval differentiation, the peaks or valleys that may lead to a certain diagnostic conclusion are significantly marked. Combined with the threshold information in the scalp health knowledge graph, the high-risk interval for oiliness is indicated at the UI level. The system identifies multiple user detection records and categorizes them into a long-term archive vector. It reads user fine-tuning parameters and embeds knowledge graph constraint functions. Using incremental training and lifestyle habit mapping, it outputs longitudinal predictions and intervention plans, and writes the aforementioned decision results and interpretable outputs back to the long-term archive vector.

2. The intelligent scalp health assessment system according to claim 1, characterized in that: Acquire visible light microscopic image data, infrared thermal image data, and sensor data; A metadata tag set is established for each set of collected data, the tags are bound to the data body for storage, an integrated data management module is built, and a unique identifier is assigned to each piece of collected data.

3. The intelligent scalp health assessment system according to claim 2, characterized in that: The moments when there are sudden changes or extreme readings in the sensor data are marked, and anomaly masks are created for significant noise pixels in visible light microscopic image data and infrared thermal image data. After preprocessing, the acquired corrected visible light microscopic image, corrected infrared thermograph, and corrected sensor data are written back to the data management module.

4. The intelligent scalp health assessment system according to claim 3, characterized in that: After dynamically updating the attention parameters, they are written to the system configuration library. If a path is missing or abnormal during subsequent use, the historical feature sequence of the temporal modality feature vector and the channel information in the image modality feature vector are called to supplement the missing data through the missing prediction model. If the prediction reliability does not meet expectations, the user will be prompted to re-collect data or perform manual verification. The adjustment parameters, completion behavior, and user feedback information are written into the corresponding database. If the completion result is confirmed to be accurate, the positive feedback to the missing prediction model parameters is strengthened.

5. The intelligent scalp health assessment system according to claim 4, characterized in that: Overlay heatmaps over the image area to mark key parts; highlight key peaks and intervals in the time series curve; display current attention parameters and complete model prediction accuracy in the form of pop-ups or charts; The model can be continuously iterated and evolved by globally retraining or periodically iterating on the missing prediction model and attention parameters. If a model or modality performs poorly in a real-world scenario for an extended period, an automatic notification will be sent to the development or maintenance team, and an alert will be issued to external parties.

6. The intelligent scalp health assessment system according to claim 5, characterized in that: Collect corrected visible light microscopic images, corrected infrared thermal images, and corrected sensor data obtained by users at different time periods, and archive them together with decision results, interpretable outputs, and user feedback records to establish a long-term archive vector. The long-term archive vector also includes users' lifestyle habits and dietary preferences. Based on the data management module and scalp health knowledge graph, a hierarchical index is created for the image modality feature vector, temporal modality feature vector, and fusion representation of each detection in the long-term archive vector.

7. The intelligent scalp health assessment system according to claim 6, characterized in that: When a user accumulates multiple test records within a certain period, or when symptom indicators fluctuate significantly in a short period of time, a personalized training process is triggered, and a personalized model is obtained after fine-tuning. When the input of the corresponding user is detected, the personalized model is scheduled to perform inference first. If there is a large range of missing modalities or outliers, it will automatically fall back to the global base model.

8. The intelligent scalp health assessment system according to claim 7, characterized in that: Based on long-term archival vectors and personalized models, scalp health trends are periodically assessed: An immediate alert is triggered when the assessment period is short and the predicted risk level is high. If the hair quality abnormality indicators show signs of rising in future assessment cycles, corresponding risk factors will be matched according to the cause-symptom chain in the scalp health knowledge graph, and targeted intervention or care suggestions will be provided to the user.

Images