Skin disease course management system based on artificial intelligence

Abstract

The invention relates to the technical field of medical artificial intelligence, and discloses a skin disease course management system based on artificial intelligence. The system comprises an image receiving module, a feature extraction module, a sequence construction module, a state recognition module, a path generation module and a state judgment module. The system firstly receives a skin disease area sequence image from an image acquisition device, preprocesses the image and extracts a morphological feature vector of a skin focus; arranging the feature vectors according to a time sequence to construct a skin disease course feature sequence; analyzing the sequence by adopting a state recognition model and outputting a disease course state sequence; generating a disease course evolution path according to the transition probability of adjacent state nodes in the state sequence; and finally, state attribution judgment is carried out on subsequently newly collected feature vectors according to the path. According to the invention, dynamic, quantitative and automatic tracking and trend analysis of skin conditions are realized, the limitation of single static evaluation is overcome, and the objectivity and predictability of disease course management are improved.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a skin disease management system based on artificial intelligence. Background Technology

[0002] Currently, significant progress has been made in computer vision-based assisted diagnosis of skin diseases. These technologies typically rely on lesion images acquired during a single patient visit, using deep learning models for feature extraction and classification to identify specific skin diseases or assess their severity. This single-point static analysis model can provide doctors with immediate diagnostic references, but its evaluation results only represent the disease's manifestation at a specific time slice and cannot reflect the entire dynamic development of the disease.

[0003] Existing technical solutions have shortcomings. Static analysis methods disrupt the continuity of the disease course, making it difficult to capture the subtle changes in lesion shape, size, and texture over time. Clinicians need to manually compare patients' past medical records and photographs, relying on experience to subjectively judge the disease's progression and treatment response. This method is inefficient, lacks objective quantitative standards, and cannot quantitatively predict the future direction of the disease. Medical practice urgently needs a technical means to perform full-cycle, automated, quantitative tracking and trend analysis of skin diseases.

[0004] The purpose of this invention is to solve the problem of how to conduct continuous dynamic monitoring and objective quantitative assessment of skin diseases. The key is to overcome the limitations of single static analysis and achieve a leap from isolated point assessment to continuous time-series analysis. This requires constructing a quantifiable disease progression feature sequence from a series of time-series images, identifying the changing patterns of disease states based on this sequence, and ultimately characterizing the disease progression path and inferring future states. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based skin disease management system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based skin disease progression management system, the system comprising:

[0007] The image receiving module is used to receive sequential images of skin disease areas from the image acquisition device;

[0008] The feature extraction module is used to preprocess the sequence images of the skin disease area and extract the morphological feature vectors of the skin lesions;

[0009] The sequence construction module is used to arrange the morphological feature vectors in chronological order to construct a feature sequence of the skin disease course.

[0010] The state recognition module is used to analyze the characteristic sequence of the skin disease course using a state recognition model and output the state sequence of the skin disease course.

[0011] The path generation module is used to generate a disease evolution path based on the transition probability of adjacent state nodes in the skin disease disease state sequence.

[0012] The state determination module is used to determine the state attribution of subsequently received morphological feature vectors based on the disease progression path.

[0013] Preferably, the construction of the skin disease course characteristic sequence includes:

[0014] The difference between the morphological feature vector at each time point and the morphological feature vector at the initial time point is calculated to obtain the relative feature change.

[0015] Arrange the relative feature changes in chronological order to form a relative feature change sequence;

[0016] The relative feature change sequence is spliced ​​with the skin disease course feature sequence to generate an enhanced skin disease course feature sequence.

[0017] Preferably, the analysis of the skin disease course feature sequence using a state recognition model includes:

[0018] The enhanced skin disease course feature sequence is input into a pre-trained state recognition model;

[0019] The state recognition model identifies time periods with stable features by analyzing the similarity of feature vectors at adjacent time points in the enhanced skin disease course feature sequence.

[0020] Each stable time period is marked as a disease course state node, thus outputting a skin disease course state sequence composed of multiple disease course state nodes.

[0021] Preferably, the generated pathogenesis evolution path includes:

[0022] Calculate the feature distance between any two disease state nodes in the skin disease disease course state sequence;

[0023] Based on the characteristic distance and the time interval between the occurrence of the state, the transition probability between any two disease state nodes is calculated;

[0024] Connect all state transition relationships whose transition probabilities exceed a preset threshold to form the disease progression path.

[0025] Preferably, the determination of state attribution for subsequently received morphological feature vectors includes:

[0026] The newly extracted morphological feature vectors are mapped onto the disease progression path;

[0027] Calculate the feature similarity between the newly extracted morphological feature vector and the feature similarity between each disease state node on the disease evolution path;

[0028] The disease state node with the highest feature similarity is determined as the belonging state of the newly extracted morphological feature vector.

[0029] Preferably, the system further includes dynamically adjusting the disease progression path:

[0030] When the ownership status of newly extracted morphological feature vectors is determined to point to the same new disease state node for a preset number of consecutive times, the new disease state node is created in the disease evolution path.

[0031] Recalculate the transition probabilities between all disease state nodes in the disease evolution path, and update the disease evolution path.

[0032] Preferably, the dynamic adjustment of the disease progression path further includes:

[0033] Monitor the frequency at which newly extracted morphological feature vectors are assigned to each disease state node on the disease evolution path;

[0034] If a disease state node is not assigned within a preset long period, the weight of that disease state node in the disease evolution path will be reduced.

[0035] When the weight is lower than the failure threshold, the disease state node is removed from the disease evolution path.

[0036] Preferably, the system further includes abnormal disease course pattern detection:

[0037] Real-time calculation of the feature difference between the newly extracted morphological feature vector and the feature difference between the node to which it belongs in the disease course state;

[0038] If the feature difference degree continuously exceeds the abnormal threshold, then the disease data corresponding to the time interval in which the feature difference degree continuously exceeds the abnormal threshold is marked as an abnormal disease segment.

[0039] The feature patterns of the abnormal disease course segments are extracted, and the matching degree is calculated with the feature patterns of historical abnormal disease course segments.

[0040] Preferably, the abnormal disease course pattern detection further includes:

[0041] When the feature pattern of a newly emerging abnormal disease segment matches the feature pattern of any historical abnormal disease segment less than the preset match, the feature pattern of the newly emerging abnormal disease segment is stored as a new abnormal disease pattern.

[0042] When the feature pattern of a newly emerging abnormal disease segment matches the feature pattern of a certain historical abnormal disease segment better than the preset match, the feature pattern of the historical abnormal disease segment is updated, and the frequency of occurrence of the abnormal pattern corresponding to the historical abnormal disease segment is recorded.

[0043] Preferably, the steps for constructing the state recognition model include:

[0044] Collect sequential image data of historical skin disease cases, with each case containing images of the skin disease area at multiple time points and their corresponding disease status labels;

[0045] The images of the skin disease area at each time point were preprocessed, including image normalization, noise removal and lesion segmentation, and morphological feature vectors were extracted.

[0046] Arrange the morphological feature vectors of each case in chronological order to construct a sequence of skin disease course features for training.

[0047] Using the skin disease course feature sequence and corresponding course state labels used for training, a recurrent neural network model is trained. This model outputs a course state sequence by learning temporal dependencies.

[0048] During training, the model parameters are optimized using the cross-entropy loss function and gradient descent algorithm until the model's accuracy on the validation set reaches a preset threshold.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] By extracting quantitative features from sequential images and constructing a disease progression feature sequence along a timeline, discrete single-point observations are transformed into continuous dynamic trajectories. This technique enables the objective quantification and visual tracking of the morphological evolution of lesions, such as area, color, and texture, revealing slow changing trends and patterns that cannot be detected in a single examination. It elevates the qualitative assessment, which relies on subjective experience and memory comparison by clinicians, to a systematic, data-driven quantitative analysis, providing objective evidence throughout the entire medical cycle for judging treatment response and disease progression.

[0051] This technique uses a state recognition model to parse continuous feature sequences into discrete state sequences and generates evolutionary paths based on the transition probabilities between states. It models disease progression as a dynamic system with probabilistic characteristics, enabling the inference of possible future evolution directions based on historical patterns. This represents a leap from retrospective description to prospective prediction. The system not only reports the past progression of the disease but also calculates the probability of migration to different subsequent states based on the current state, providing probabilistic decision support for early intervention and personalized treatment adjustments. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based skin disease management system described in this invention.

[0053] Figure 2 A flowchart for constructing enhanced sequences of skin disease progression characteristics;

[0054] Figure 3 This is a flowchart for determining the state attribution of subsequent morphological feature vectors;

[0055] Figure 4 A heatmap of the probability matrix of disease progression states in skin diseases;

[0056] Figure 5 Line graph for monitoring the differences in abnormal characteristics of the course of skin diseases. Detailed Implementation

[0057] 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.

[0058] Please see Figure 1 This invention provides an artificial intelligence-based skin disease progression management system. The system includes: an image receiving module that receives sequence images of skin disease areas from an image acquisition device; a feature extraction module that preprocesses the sequence images and extracts morphological feature vectors of skin lesions; a sequence construction module that arranges the morphological feature vectors in chronological order to construct a skin disease progression feature sequence; a state recognition module that analyzes the skin disease progression feature sequence using a state recognition model and outputs a skin disease progression state sequence; a path generation module that generates a progression evolution path based on the transition probability of adjacent state nodes in the skin disease progression state sequence; and a state determination module that determines the state attribution of subsequently received morphological feature vectors based on the progression evolution path.

[0059] Example 1: See Figure 2In specific implementation, when constructing the dermatology course feature sequence, the sequence construction module calculates the difference between the morphological feature vector at each time point and the morphological feature vector at the initial time point to obtain the relative feature change. The relative feature changes are then arranged chronologically to form a relative feature change sequence. This relative feature change sequence is then concatenated with the dermatology course feature sequence to generate an enhanced dermatology course feature sequence. The construction of the enhanced dermatology course feature sequence relies on vector subtraction operations to ensure that all feature dimensions participate in the calculation. In some embodiments, the calculation of relative feature changes uses an element-wise differencing method. The arrangement of the relative feature change sequence is based on the image acquisition time index, and the concatenation process merges the original sequence and the change sequence by adding feature dimensions. The sequence construction module uses an element-wise differencing method to calculate the relative feature change. This method performs independent differencing operations on each feature dimension of the morphological feature vector. Specifically, for each element in the vector, the system calculates the arithmetic difference between its value at the current time point and its value at the initial time point, thereby generating a relative change vector with the same dimensions as the original feature vector. The arrangement of the relative feature change sequence strictly follows the time index recorded by the image acquisition device, that is, the relative change vectors at each time point are organized into a sequence according to the order of the timestamps; the splicing process is achieved by expanding the feature dimension, that is, connecting the original skin disease course feature sequence and the relative feature change sequence in the feature dimension, thereby forming an enhanced sequence with doubled feature dimension but consistent time points.

[0060] In practical implementation, the state recognition module uses a state recognition model to analyze the enhanced skin disease course feature sequence. The enhanced skin disease course feature sequence is input into the pre-trained state recognition model, which identifies stable time periods by analyzing the similarity of feature vectors at adjacent time points in the enhanced skin disease course feature sequence. The similarity calculation uses the following formula:

[0061] ;

[0062] in: Indicates a point in time With time point The cosine similarity value, Indicates a point in time Morphological feature vectors, Indicates a point in time Morphological feature vectors, point operations Dot product of vectors, norm operation This represents the Euclidean norm of the vector. When the cosine similarity values ​​at multiple consecutive time points are higher than a preset stability threshold, the state recognition model determines that the interval is a stable feature time period. Optionally, the stable feature time periods are labeled using a fixed-length sliding window method, with the window size determined according to the image acquisition interval. Each stable feature time period is mapped to a disease progression state node, and the state recognition model outputs a sequence of skin disease progression states composed of these disease progression state nodes. In some embodiments, the state recognition model is designed based on a long short-term memory network architecture, with the network hidden states capturing temporal patterns, and the output layer using a softmax function to generate a state probability distribution. It can be understood that by evaluating feature stability through cosine similarity, the state recognition model transforms continuous sequences into discrete state sequences, supporting disease progression analysis.

[0063] In practical implementation, the generation and state recognition processes of enhanced skin disease course feature sequences are executed sequentially. After the sequence construction module completes the splicing, it directly transmits the data to the state recognition module, where the state recognition model performs forward inference to calculate similarity values. Optionally, the extraction of morphological feature vectors includes color, texture, and shape features. Preprocessing steps ensure uniformity of feature dimensions, and cosine similarity calculation reduces the influence of vector amplitude, focusing on directional changes.

[0064] Example 2: See Figure 3 In specific implementation, when the path generation module generates the disease evolution path, it calculates the feature distance between any two disease state nodes in the skin disease disease state sequence. Based on the feature distance and the time interval between state occurrences, it calculates the transition probability between any two disease state nodes. All state transition relationships with transition probabilities exceeding a preset threshold are connected to form the disease evolution path. The feature distance is calculated using the Euclidean distance method, and the time interval is determined based on the difference in timestamps corresponding to the state nodes. In some embodiments, the calculation of the transition probability integrates feature proximity and time accessibility factors; the smaller the feature distance, the higher the transition probability; the shorter the time interval, the higher the transition probability. The final disease evolution path is a directed graph structure, where nodes represent disease states and edges represent possible transition relationships between states.

[0065] In practical implementation, the state determination module determines the state attribution of subsequently received morphological feature vectors, maps the newly extracted morphological feature vectors onto the disease evolution path, calculates the feature similarity between the newly extracted morphological feature vectors and each disease state node on the disease evolution path, and determines the disease state node with the highest feature similarity as the attribution state of the newly extracted morphological feature vector; the feature similarity is calculated using the following formula:

[0066] ;

[0067] in: Represents the feature similarity value. This represents the newly extracted morphological feature vector. This represents the feature vector corresponding to a certain disease stage node on the disease progression path, represented by double vertical lines. This represents the Euclidean norm of the vector. Optionally, the mapping process involves alignment operations in the feature space, similarity comparisons cover all existing state nodes, and the node corresponding to the maximum value is selected for attribution determination. In some embodiments, feature similarity calculations are normalized to ensure that the result ranges between zero and one, and a minimum similarity threshold is set for attribution determination; values ​​below the threshold are marked as unknown states.

[0068] In practice, the path generation module and the state determination module work together. The disease evolution path output by the path generation module serves as a reference framework for the state determination module, and the attribution results from the state determination module can be fed back to the path generation module for path updates. Optionally, the feature distance calculation uses the average feature vector of the state nodes, the time interval calculation uses natural days as the unit, and the preset threshold is set based on clinical experience and supports dynamic adjustment.

[0069] In practice, the generation of disease progression paths relies on the statistical regularities of existing state sequences, while state attribution determination depends on real-time feature comparison; both processes are based on vector space models. It can be understood that by calculating quantified distance and similarity, the problem of disease progression is transformed into a computable spatial relationship problem, providing a quantitative basis for the management of skin disease progression.

[0070] See Figure 4 This is a matrix heatmap, with the current disease stage as the rows and the next stage as the columns. Through color depth and numerical labels, it visually displays the transition probabilities between the four disease stage nodes. The value of this graph lies in quantifying the dynamic migration patterns of disease stages: it reflects both the stability of stages such as S1 and S4 and clarifies the transition trends between stages. This provides an objective probabilistic basis for constructing disease evolution paths, helping clinicians predict the direction of disease progression and assisting in adjusting personalized treatment plans.

[0071] Example 3: In specific implementation, the system dynamically adjusts the disease progression path. When the ownership status of newly extracted morphological feature vectors all points to the same new disease progression state node after a preset number of consecutive determinations, a new disease progression state node is created in the disease progression path. The transition probability between all disease progression state nodes in the disease progression path is recalculated, and the disease progression path is updated. The threshold for the preset number of consecutive determinations is configured according to the application scenario. The feature vector of the new disease progression state node is obtained by calculating the mean of the morphological feature vectors corresponding to the ownership status. In some embodiments, the recalculation of the transition probability is based on a Bayesian update rule, integrating the occurrence frequency and time interval factors of the new node. The update operation is executed asynchronously in the background to avoid blocking the real-time status determination process. Optionally, the determination of the preset number of consecutive determinations adopts a sliding window mechanism, and the window size can be adjusted to adapt to different disease progression speeds. The creation of the new node triggers a visual update of the path, making it easy for users to observe changes.

[0072] In practice, dynamically adjusting the disease progression path also includes monitoring the frequency with which newly extracted morphological feature vectors assign each disease progression state node to the path; if a disease progression state node is not assigned within a preset long period, its weight in the path is reduced; when the weight falls below a failure threshold, the node is removed from the path; the reduction in weight follows the formula below:

[0073] ;

[0074] in: This indicates the updated weights. Represents the original weights. This represents the attenuation coefficient, and its value ranges from zero to one. This represents the time interval during which a node remains unassigned within a preset long period. In some embodiments, the monitoring frequency is achieved by maintaining an assignment counter for each disease progression node; the preset long period is set according to the disease type, for example, ninety days for chronic diseases; the expiration threshold is set to 0.1 to ensure that only completely outdated nodes are removed. Optionally, the initial weight value is set to 1.0; the decay coefficient is adjusted based on clinical experience, with a smaller value resulting in a faster decrease in weight; the path connection relationship is updated synchronously after a node is removed.

[0075] Example 4: In specific implementation, the system performs abnormal disease course pattern detection and calculates the feature difference between the newly extracted morphological feature vector and the feature difference degree of its corresponding disease course state node in real time. The feature difference degree is calculated using the Euclidean distance method, comparing the deviation of the newly extracted morphological feature vector and the feature vector corresponding to the corresponding disease course state node in each dimension. If the feature difference degree continuously exceeds the abnormal threshold, the disease course data of that time period is marked as an abnormal disease course segment. The abnormal threshold is set based on the statistical distribution of historical normal variations, for example, taking three times the standard deviation of the normal distribution as the benchmark value. In some embodiments, the calculation period of the feature difference degree is synchronized with the image acquisition frequency to ensure real-time performance. The marking of abnormal disease course segments includes a start timestamp, an end timestamp, and the corresponding feature difference degree sequence. When extracting the feature pattern of an abnormal disease course segment, the mean vector of all morphological feature vectors in the segment is calculated as the representative pattern. It can be understood that by monitoring the feature difference degree in real time, the system can promptly capture abnormal events that deviate from the normal disease course.

[0076] In practice, after extracting the feature patterns of abnormal disease course segments, the matching degree is calculated by comparing them with the feature patterns of historical abnormal disease course segments. The matching degree is calculated using the following formula:

[0077] ;

[0078] in: This represents the matching degree value. This represents the feature pattern vector of the newly extracted abnormal disease course fragment. A feature pattern vector representing a specific historical abnormal disease course segment, point operations. This represents the vector dot product. The matching degree value ranges from zero to one, with higher values ​​indicating greater pattern similarity. When the matching degree between the feature pattern of a newly emerging abnormal disease segment and the feature pattern of any historical abnormal disease segment is lower than a preset matching degree, the feature pattern of the newly emerging abnormal disease segment is stored as a new abnormal disease pattern. The preset matching degree is usually set to 0.8 to ensure that only significantly different patterns are recorded as new patterns. Optionally, storing the new pattern includes assigning a unique identifier, recording the feature pattern vector, and initializing the occurrence frequency to one. In some embodiments, the feature patterns of historical abnormal disease segments are stored in a database, and the matching degree calculation is completed by traversing all existing patterns; a system notification is triggered after the new pattern is created for clinicians to review.

[0079] In practice, when the feature pattern of a newly emerging abnormal disease course segment matches the feature pattern of a historical abnormal disease course segment better than a preset match, the feature pattern of that historical abnormal disease course segment is updated, and the frequency of the abnormal pattern is recorded. The update operation uses a weighted average method, merging the new feature pattern and the historical feature pattern according to frequency, and increasing the frequency value by one. See Table 1 for the recording of abnormal disease course patterns.

[0080] Table 1: Recording Table of Abnormal Disease Progression Patterns

[0081] Optionally, feature pattern vectors are stored as arrays to support efficient matching calculations; frequency of occurrence is used to assess the commonness of abnormal patterns and assist clinical decision-making. In some embodiments, the update process includes checking pattern similarity thresholds to avoid over-merging of different patterns; frequency records help identify recurring abnormality types.

[0082] In practice, the abnormal disease course pattern detection process is closely integrated with the state determination module. The newly extracted morphological feature vectors immediately enter the difference calculation stage after the state assignment is completed. The adjustment of the abnormal threshold is based on the feature fluctuation of recent normal disease course segments, and the statistics are recalculated using a sliding window method.

[0083] See Figure 5 This graph, with time points on the horizontal axis and feature difference degree on the vertical axis, clearly presents the characteristic fluctuations during the course of the disease: the blue line represents normal difference degree, the red dashed line represents the abnormal threshold, the orange line represents the actual detected feature difference degree, and the light yellow filled area represents the marked abnormal interval. This graph is an intuitive representation of the abnormality detection logic. By quantifying the relationship between feature difference degree and threshold, it can quickly locate abnormal fluctuation periods in the course of the disease, providing a visual basis for subsequent extraction of abnormal feature patterns and matching of historical abnormal cases, helping clinicians to identify abnormalities in the course of the disease and take intervention measures in a timely manner.

[0084] Example 5: In specific implementation, the construction steps of the state recognition model include collecting sequential image data of historical dermatological cases. Each case contains images of the dermatological area at multiple time points and their corresponding disease status labels. The data collection process ensures the diversity and representativeness of cases. The sequential image data sources cover different medical institutions and patient groups. The disease status labels are marked by professional dermatologists according to clinical guidelines. The labeling system includes standard states such as acute phase, chronic phase, and remission phase. In some embodiments, data collection also involves ethical review and patient informed consent to ensure compliance. The image data of each case is stored together with metadata such as collection time and device type to form a structured dataset.

[0085] In practice, preprocessing is performed on the skin lesion area images at each time point, including image normalization, noise removal, and lesion segmentation, and morphological feature vector extraction. Image normalization adjusts the image size to a uniform pixel resolution and standardizes brightness and contrast. Noise removal uses a Gaussian filtering algorithm to reduce random noise interference. Lesion segmentation uses a deep learning-based semantic segmentation model to accurately identify the boundaries of the skin lesion area. Optionally, the preprocessing process is automated, processing batches of images through a pipeline. Morphological feature vector extraction includes calculating the color statistics, texture features, and geometric parameters of the lesion area to generate high-dimensional feature vectors. In some embodiments, feature extraction is implemented using open-source computer vision libraries to ensure consistency and repeatability of computation.

[0086] In practice, the morphological feature vectors of each case are arranged chronologically to construct a training sequence of skin disease progression features. The chronological order is strictly determined by the timestamps of image acquisition, forming an independent time series for each case. The sequence length is variable to accommodate different observation periods. Optionally, for cases with inconsistent sequence lengths, padding or truncation methods are used to align the time dimension. Time interval information is retained during sequence construction as auxiliary input to enhance temporal modeling. It can be understood that the construction of the disease progression feature sequence captures the dynamic evolution of disease development.

[0087] In practice, a recurrent neural network (RNN) model is trained using the skin disease course feature sequences and corresponding course state labels. This model outputs the course state sequence by learning temporal dependencies. The RNN employs a gated recurrent unit (GRN) architecture with 256 hidden layer nodes. The training objective is to minimize the prediction error. The loss function used during training is calculated as follows:

[0088] ;

[0089] in: Indicates batch loss value. Indicates batch size, Indicates the number of disease course status categories. Indicates the first The sample at the th The one-hot encoded value of the true label in each category. The model predicts the first The sample belongs to the first The probability values ​​for each category. In some embodiments, the model is initialized with random weights, and the training data is divided into training, validation, and test sets with proportions of 70%, 15%, and 15%, respectively; the backpropagation algorithm is used to calculate gradients during training. It can be understood that by optimizing the cross-entropy loss, the model learns an accurate mapping from the feature sequence to the state sequence.

[0090] In practice, during training, the model parameters are optimized using the cross-entropy loss function and gradient descent algorithm until the model's accuracy on the validation set reaches a preset threshold. The gradient descent algorithm uses an adaptive momentum estimation optimizer, with an initial learning rate of 0.001 and a maximum of 1000 training epochs. Optionally, the preset threshold can be set to 0.90, representing 90% classification accuracy, depending on the application scenario. The validation set performance is monitored during training, and training is terminated early when the accuracy no longer improves after 10 consecutive epochs. In some embodiments, model parameters are saved as a file for easy loading during deployment; training logs record loss values ​​and accuracy curves for debugging and analysis.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A skin disease progression management system based on artificial intelligence, characterized in that, The system includes: The image receiving module is used to receive sequential images of skin disease areas from the image acquisition device; The feature extraction module is used to preprocess the sequence images of the skin disease area and extract the morphological feature vectors of the skin lesions; The sequence construction module is used to arrange the morphological feature vectors in chronological order to construct a feature sequence of the skin disease course. The state recognition module is used to analyze the characteristic sequence of the skin disease course using a state recognition model and output the state sequence of the skin disease course. The path generation module is used to generate a disease evolution path based on the transition probability of adjacent state nodes in the skin disease disease state sequence. The state determination module is used to determine the state attribution of subsequently received morphological feature vectors based on the disease progression path.

2. The artificial intelligence-based skin disease management system according to claim 1, characterized in that, The constructed characteristic sequence of skin disease progression includes: The difference between the morphological feature vector at each time point and the morphological feature vector at the initial time point is calculated to obtain the relative feature change. Arrange the relative feature changes in chronological order to form a relative feature change sequence; The relative feature change sequence is spliced ​​with the skin disease course feature sequence to generate an enhanced skin disease course feature sequence.

3. The artificial intelligence-based skin disease management system according to claim 2, characterized in that, The analysis of the skin disease course feature sequence using a state recognition model includes: The enhanced skin disease course feature sequence is input into a pre-trained state recognition model; The state recognition model identifies time periods with stable features by analyzing the similarity of feature vectors at adjacent time points in the enhanced skin disease course feature sequence. Each stable time period is marked as a disease course state node, thus outputting a skin disease course state sequence composed of multiple disease course state nodes.

4. The artificial intelligence-based skin disease management system according to claim 3, characterized in that, The generated pathogenesis evolution path includes: Calculate the feature distance between any two disease state nodes in the skin disease disease course state sequence; Based on the characteristic distance and the time interval between the occurrence of the state, the transition probability between any two disease state nodes is calculated; Connect all state transition relationships whose transition probabilities exceed a preset threshold to form the disease progression path.

5. The artificial intelligence-based skin disease management system according to claim 4, characterized in that, The process of determining the state attribution of subsequently received morphological feature vectors includes: The newly extracted morphological feature vectors are mapped onto the disease progression path; Calculate the feature similarity between the newly extracted morphological feature vector and the feature similarity between each disease state node on the disease evolution path; The disease state node with the highest feature similarity is determined as the belonging state of the newly extracted morphological feature vector.

6. The artificial intelligence-based skin disease management system according to claim 1, characterized in that, The system also includes dynamically adjusting the disease progression path: When the ownership status of newly extracted morphological feature vectors is determined to point to the same new disease state node for a preset number of consecutive times, the new disease state node is created in the disease evolution path. Recalculate the transition probabilities between all disease state nodes in the disease evolution path, and update the disease evolution path.

7. The artificial intelligence-based skin disease management system according to claim 6, characterized in that, The dynamic adjustment of the disease progression path also includes: Monitor the frequency at which newly extracted morphological feature vectors are assigned to each disease state node on the disease evolution path; If a disease state node is not assigned within a preset long period, the weight of that disease state node in the disease evolution path will be reduced. When the weight is lower than the failure threshold, the disease state node is removed from the disease evolution path.

8. The artificial intelligence-based skin disease management system according to claim 1, characterized in that, The system also includes abnormal disease progression pattern detection: Real-time calculation of the feature difference between the newly extracted morphological feature vector and the feature difference between the node to which it belongs in the disease course state; If the feature difference degree continuously exceeds the abnormal threshold, then the disease data corresponding to the time interval in which the feature difference degree continuously exceeds the abnormal threshold is marked as an abnormal disease segment. The feature patterns of the abnormal disease course segments are extracted, and the matching degree is calculated with the feature patterns of historical abnormal disease course segments.

9. The artificial intelligence-based skin disease management system according to claim 8, characterized in that, The abnormal disease course pattern detection also includes: When the feature pattern of a newly emerging abnormal disease segment matches the feature pattern of any historical abnormal disease segment less than the preset match, the feature pattern of the newly emerging abnormal disease segment is stored as a new abnormal disease pattern. When the feature pattern of a newly emerging abnormal disease segment matches the feature pattern of a certain historical abnormal disease segment better than the preset match, the feature pattern of the historical abnormal disease segment is updated, and the frequency of occurrence of the abnormal pattern corresponding to the historical abnormal disease segment is recorded.

10. The artificial intelligence-based skin disease management system as described in claim 1, characterized in that, The steps for constructing the state recognition model include: Collect sequential image data of historical skin disease cases, with each case containing images of the skin disease area at multiple time points and their corresponding disease status labels; The images of the skin disease area at each time point were preprocessed, including image normalization, noise removal and lesion segmentation, and morphological feature vectors were extracted. Arrange the morphological feature vectors of each case in chronological order to construct a sequence of skin disease course features for training. Using the skin disease course feature sequence and corresponding course state labels used for training, a recurrent neural network model is trained. This model outputs a course state sequence by learning temporal dependencies. During training, the model parameters are optimized using the cross-entropy loss function and gradient descent algorithm until the model's accuracy on the validation set reaches a preset threshold.

Images