Human hair loss type and stage recognition system based on cross-domain semi-supervised learning
By employing a cross-domain semi-supervised learning system for identifying human hair loss types and stages, and combining multi-source data and professional knowledge, this system solves the problem of traditional diagnostic methods relying on manual judgment, achieving low-cost and efficient hair loss diagnosis and improving the early detection rate of the disease.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 杭州丝跃科技有限公司
- Filing Date
- 2022-07-06
- Publication Date
- 2026-07-10
AI Technical Summary
Current methods for diagnosing hair loss rely on the judgment of professional physicians, which leads to patients having to go to medical institutions for complicated treatments, making it difficult to afford the costs, and the disease cannot be detected in its early stages, making the condition prone to worsening.
A human hair loss type and stage identification system based on cross-domain semi-supervised learning is adopted. The system combines multi-source data and professional knowledge through classification and analysis modules for diagnosis. The classification model is trained using semi-supervised learning, multi-source domain generalization and course learning methods, and a comprehensive diagnosis is made by combining graph convolutional neural network analysis of medical history and medical knowledge graph.
It enables accurate identification of hair loss stages using a small amount of labeled data and a large amount of unlabeled data, reducing diagnostic costs and improving the early detection rate of the disease.
Smart Images
Figure CN115239993B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a human hair loss type and stage identification system based on cross-domain semi-supervised learning. Background Technology
[0002] Hair loss is a widespread health problem facing urban populations today. Current diagnostic methods primarily rely on patient consultations for diagnosis. Traditional pathological diagnoses of hair loss heavily depend on the judgment of specialist physicians, requiring patients to visit specialized medical institutions and undergo a series of cumbersome procedures to obtain diagnosis and treatment advice. Furthermore, this traditional consultation-based approach makes it difficult for the average patient to afford professional medical care and access such services. Simultaneously, the cumbersome procedures prevent the early detection of various hair loss conditions, leading to deterioration of the condition due to delayed treatment. Summary of the Invention
[0003] This invention provides a human hair loss type and stage identification system based on cross-domain semi-supervised learning to solve the aforementioned technical problems, specifically adopting the following technical solution:
[0004] A human hair loss type and stage recognition system based on cross-domain semi-supervised learning, comprising:
[0005] The classification module receives the images to be analyzed and classifies the input images using a trained classification model.
[0006] The analysis module receives the classification results, medical history sequence, and hair loss medical knowledge graph from the classification module, and outputs the analysis results through a comprehensive analysis using a trained analysis model.
[0007] The method for training the classification model is as follows:
[0008] Collect training image data, which consists of images with different degrees of hair loss.
[0009] The classification model is trained using a semi-supervised learning method;
[0010] The classification model is trained using a multi-source domain generalization method;
[0011] The classification model is trained using a course-based learning approach.
[0012] Furthermore, the training image data contains P datasets. Each dataset contains... A labeled dataset of samples and have Unlabeled dataset of samples ,in Representing the Zhang Image Representing the The labels corresponding to each image Representing the Zhang image.
[0013] Furthermore, the labels are the hair loss condition and severity corresponding to the image.
[0014] Furthermore, the specific method for training the classification model using semi-supervised learning is as follows:
[0015] Supervised training is performed on labeled samples. Perform forward prediction to obtain the probability of each category. and its predicted value With real labels Compare the results and correct the prediction error through backpropagation.
[0016] For unlabeled samples, during the training phase, the samples undergo one strong random augmentation transformation and one weak random transformation.
[0017]
[0018]
[0019] Forward prediction is performed on the data after the two transformations to obtain the probability of each category. and If the prediction confidence of the weak transform is higher than the threshold If the prediction result is used as a pseudo-label, then the pseudo-label is used to correct the strong transform prediction result.
[0020] Furthermore, in the process of training the classification model using a semi-supervised learning method, cross-entropy is used as the training loss function.
[0021] Furthermore, the specific method for training the classification model using multi-source domain generalization is as follows:
[0022] Jointly train the classification model on different data sources;
[0023] Perform regularization on the classification model.
[0024] Furthermore, the specific method for training the classification model using a course-based learning approach is as follows:
[0025] The confidence level of a sample is used as the criterion for judging the difficulty of the sample. The higher the confidence level, the easier the sample is; the lower the confidence level, the more difficult the sample is; and the higher the confidence level, the greater the weight of the sample.
[0026] Furthermore, the specific method for training the analysis model is as follows:
[0027] Collect textual records of patients' historical diagnoses and filter out the diagnostic results, treatment opinions, and consultation times from the patients' medical history;
[0028] Transform the medical history into a medical history sequence of length L. Each element is a triple. ;
[0029] Constructing a medical knowledge graph related to hair loss. Each node in the graph represents a type of hair loss disease, and each edge represents the probability of mutual transfer between two diseases.
[0030] The analysis model is constructed, and it contains two graph convolutional neural networks. and ,in Responsible for processing medical history sequences It outputs a historical information feature. , Responsible for handling hair loss medical knowledge graph and output disease knowledge features. ;
[0031] Historical information features Disease knowledge characteristics The classification results of the classification model Perform a vector merging operation to output a unified vector. The final analysis and diagnostic results are predicted through a linear layer.
[0032] The analysis model is trained using real diagnostic labels, and the parameters are updated using gradient descent to obtain the trained analysis model.
[0033] Furthermore, the specific method for filtering out the patient's diagnosis results, treatment opinions, and consultation times from their medical history is as follows:
[0034] Construct a medical keyword dictionary;
[0035] Using regular expressions, keywords that meet the keyword requirements are matched in the text records of historical diagnostics.
[0036] The advantage of this invention lies in the fact that the human hair loss type and stage identification system based on cross-domain semi-supervised learning can identify hair loss stages by using a small amount of labeled data and a large amount of multi-source unlabeled data, and combined with professional knowledge of medical dermatology diagnosis, can conduct accurate and comprehensive analysis of patients. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of a human hair loss type and stage identification system based on cross-domain semi-supervised learning according to the present invention. Detailed Implementation
[0038] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0039] like Figure 1 The image shows a human hair loss type and stage identification system based on cross-domain semi-supervised learning, as proposed in this application. It mainly includes a classification module and an analysis module.
[0040] The classification module receives the images to be analyzed and classifies them using a trained classification model. The analysis module receives the classification results from the classification module, the patient history sequence, and the hair loss medical knowledge graph, and outputs the analysis results through a comprehensive analysis using a trained analysis model.
[0041] In this application, a comprehensive method is used to train the classification model in order to improve the classification efficiency.
[0042] Specifically, the method for training the classification model is as follows:
[0043] Collect training image data, which consists of images with different degrees of hair loss.
[0044] A semi-supervised learning method is used to train the classification model.
[0045] The classification model is trained using a multi-source domain generalization method.
[0046] The classification model is trained using a course-based learning approach.
[0047] In one preferred implementation, the training image data comprises P datasets. Each dataset contains... A labeled dataset of samples and have Unlabeled dataset of samples ,in Representing the Zhang Image Representing the The labels corresponding to each image Representing the The image contains [number] images. The labels indicate the degree and extent of hair loss corresponding to each image. It is understood that in this application, the labeled data comes from past desensitized patient diagnostic photos from multiple social hair transplant institutions and patient treatment photos with diagnostic information from multiple hospitals, ensuring accuracy and authenticity. Meanwhile, the unlabeled data comes from hair and facial data collected via internet crawlers and from publicly available datasets.
[0048] One objective of this application is to train a model with good classification performance using labeled and unlabeled samples from multiple datasets. ,in These are model parameters. It is the feature extractor of the model. It is a classifier.
[0049] As a preferred implementation method, the specific method for training the classification model using semi-supervised learning is as follows:
[0050] Supervised training is performed on labeled samples. The model first... Perform forward prediction to obtain the probability of each category. and its predicted value With real labels By comparing the results, the prediction error is corrected through backpropagation.
[0051] During training, cross-entropy is used as the loss function.
[0052]
[0053] The gradient descent algorithm can be used to calculate the gradient of the loss function with respect to each parameter. Then, the parameters are modified using the gradient to minimize the loss function.
[0054]
[0055] In the formula This represents the learning rate used for parameter updates.
[0056] For unlabeled samples, during the training phase, the samples need to undergo two data transformations, including one strong random augmentation transformation and one weak random transformation, as shown in the following equations.
[0057]
[0058]
[0059] Forward prediction is performed on the data after the two transformations to obtain the probability of each category. and If the prediction confidence of the weak transform is higher than the threshold Then, the prediction result is used as a pseudo-label, and this pseudo-label is used to correct the strong transform prediction result, as follows:
[0060]
[0061] in, It is a threshold function, that is, when Output 1 if the prediction is true, otherwise output 0. This process mainly uses the predicted values with high confidence as pseudo-labels to train the classification model, while ensuring the consistency of predictions after different transformations. Then, we again use the aforementioned loss function for gradient descent, and then modify the parameters through the gradient to minimize the loss function.
[0062]
[0063] After each iteration, the classification model is updated using a weighted average. Given a weighted average parameter... We then perform a weighted sum using the parameters obtained after gradient descent and the parameters from the previous step, as shown below.
[0064]
[0065] The above processing improves the generalization ability of the classification model.
[0066] As a preferred implementation method, the specific method for training the classification model using multi-source domain generalization is as follows:
[0067] The classification model is jointly trained on different data sources. This process is equivalent to the supervised training process described above, except that it requires joint training on multiple different datasets.
[0068] Specifically, given K different data sources, labeled data is obtained. The model is trained in a supervised manner using a joint dataset. In each training epoch, K samples are sampled from the joint dataset and input into the model to obtain the predicted probabilities. One sample is taken from each dataset. During training, the sum of the cross-entropy of multiple samples is used again as the training loss function to update the model.
[0069]
[0070] The gradient descent algorithm can be used to calculate the gradient of the loss function with respect to each parameter. Then, the parameters are modified using the gradient to minimize the loss function.
[0071]
[0072] In the formula This represents the learning rate used for parameter updates.
[0073] Regularization is applied to the classification model. The specific method for regularizing the classification model is as follows:
[0074] For data of the same category from different sources, we aim to ensure that their feature spaces belong to the same distribution. Given sample data, there are certain differences between data from different data sources. We hope to learn a unified feature space to reduce the data discrepancies caused by different data sources.
[0075]
[0076]
[0077] For two similar data from different sources and This reduces the L2 distance in its feature space.
[0078] As a preferred implementation method, the specific method for training the classification model using a course-based learning approach is as follows:
[0079] The confidence level of a sample is used to determine its difficulty; higher confidence indicates an easier sample, lower confidence indicates a more difficult sample, and higher confidence results in a greater training weight for that sample. This process employs the same supervised learning approach, using predicted probabilities. of The loss function is as follows: Each sample is weighted by a power factor.
[0080]
[0081] This loss function design ensures that simpler samples are learned first. By scoring and learning samples from easy to difficult, low-quality data is filtered out, preventing the model from overly focusing on noisy or low-quality samples.
[0082] Based on the given image prediction results, we aim to make a comprehensive diagnosis of the patient by combining their historical medical data and professional dermatological medical knowledge. In this process, we need to train an analytical model. It predicts the probability of images. Medical history sequence Hair loss medical knowledge graph As input, output a current comprehensive analysis and diagnostic result. These are the parameters of the model. This transforms the hair loss diagnosis problem into a conditional probability inference problem.
[0083] The analytical model was obtained through the following process:
[0084] First, we collect patients' historical diagnostic records. Using keyword detection technology, we first filter out the diagnostic results, treatment opinions, and consultation times from the patients' medical history. It is understandable that different types of hair loss are strongly correlated. Hair loss exhibits certain stages, ranging from mild to severe. Different types of hair loss can also transform into each other; for example, it may progress from M-shaped hair loss to comprehensive hair loss.
[0085] In this process, we first communicate with doctors to build a medical keyword dictionary, and then use regular expression technology to find keywords that meet the keyword requirements in medical texts. The diagnosis result refers to a string that provides a definitive assessment of the current condition, such as severe / moderate / mild / no hair loss, etc. The treatment opinion refers to a string of suggested treatment methods for the current condition, such as taking a certain medication / undergoing a certain surgery / increasing exercise, etc. The consultation time refers to the date on which the diagnosis was made, such as a specific year, month, and day.
[0086] Then, the medical history is transformed into a triple sequence of length L. This refers to the sequence of medical history, where each element is a triple. This sequence summarizes the patient's medical history.
[0087] Next, with the assistance of expert doctors, a medical knowledge graph related to hair loss was constructed. Each node in the graph represents a type of hair loss condition, and each edge represents the probability of transition between two conditions. This graph was constructed by a professional physician.
[0088] The analysis model contains two graph convolutional neural networks. and .in Responsible for processing medical history sequences It outputs a historical information feature. . Responsible for handling hair loss medical knowledge graph and output disease knowledge features.
[0089]
[0090] We will use historical information features Disease knowledge characteristics The classification result of the classification model, i.e., the output probability. Perform a vector merging operation to output a unified vector. The final analytical and diagnostic results are predicted through a linear layer.
[0091]
[0092] Finally, the analysis model is trained using real diagnostic labels, and the parameters are updated using gradient descent to obtain the trained analysis model.
[0093]
[0094]
[0095] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.
Claims
1. A system for identifying human hair loss types and stages based on cross-domain semi-supervised learning, characterized in that, Include: The classification module receives the images to be analyzed and classifies the input images using a trained classification model. The analysis module is used to receive the classification results, medical history sequence and hair loss medical knowledge graph from the classification module, and output the analysis results through comprehensive analysis using the trained analysis model; The method for training the classification model is as follows: Collect training image data, which consists of images with different degrees of hair loss; The classification model is trained using a semi-supervised learning method; The classification model is trained using a multi-source domain generalization method; The classification model is trained using a course-based learning approach; The specific method for training the classification model using semi-supervised learning is as follows: Supervised training is performed on labeled samples. Perform forward prediction to obtain the probability of each category. and its predicted value With real labels Compare the results and correct the prediction error through backpropagation. For unlabeled samples, during the training phase, the samples undergo one strong random augmentation transformation and one weak random transformation. Forward prediction is performed on the data after the two transformations to obtain the probability of each category. and If the prediction confidence of the weak transform is higher than the threshold If the prediction result is used as a pseudo-label, then the pseudo-label is used to correct the strong transform prediction result. The specific method for training the classification model using multi-source domain generalization is as follows: The classification model is jointly trained on different data sources; Perform regularization on the classification model; The specific method for training the classification model using the course learning approach is as follows: The confidence level of a sample is used as the criterion for judging the difficulty of the sample. The higher the confidence level, the easier the sample is; the lower the confidence level, the more difficult the sample is; and the higher the confidence level, the greater the weight of the sample. The specific method for training the analysis model is as follows: Collect textual records of patients' historical diagnoses and filter out the diagnostic results, treatment opinions, and consultation times from the patients' medical history; Transform the medical history into a medical history sequence of length L. Each element is a triple. ; Building a knowledge graph for hair loss treatment Each node in the graph represents a type of hair loss disease, and each edge represents the probability of mutual transfer between two diseases. The analysis model is constructed, and the analysis model contains two graph convolutional neural networks. and ,in Responsible for processing medical history sequences It outputs a historical information feature. , Responsible for handling hair loss medical knowledge graph and output disease knowledge features. ; Historical information features Disease knowledge characteristics The classification results of the classification model Perform a vector merging operation to output a unified vector. The final analysis and diagnostic results are predicted through a linear layer. The analysis model is trained using real diagnostic labels, and the parameters are updated using gradient descent to obtain the trained analysis model.
2. The human hair loss type and stage identification system based on cross-domain semi-supervised learning according to claim 1, characterized in that, The training image data contains P datasets. Each dataset contains... A labeled dataset of samples and have Unlabeled dataset of samples ,in Representing the Zhang Image Representing the The labels corresponding to each image Representing the Zhang image.
3. The human hair loss type and stage identification system based on cross-domain semi-supervised learning according to claim 2, characterized in that, The labels corresponding to the images indicate the type and severity of hair loss.
4. The human hair loss type and stage identification system based on cross-domain semi-supervised learning according to claim 1, characterized in that, In the process of training the classification model using a semi-supervised learning method, cross-entropy is used as the training loss function.
5. The human hair loss type and stage identification system based on cross-domain semi-supervised learning according to claim 1, characterized in that, The specific method for filtering out the diagnostic results, treatment opinions, and consultation times from a patient's medical history is as follows: Construct a medical keyword dictionary; Using regular expressions, keywords that meet the keyword requirements are matched in the text records of historical diagnostics.