An AI-assisted decision support system for skin disease diagnosis

Abstract

The present invention relates to the technical field of skin disease diagnosis, and discloses an AI-assisted decision support system for skin disease diagnosis, which integrates four modules: image acquisition, image analysis, comprehensive analysis, and diagnostic decision-making. High-definition images of the patient's skin are acquired through high-precision skin photography equipment, and a training data set is formed in combination with the patient's basic information. The skin disease recognition model and the condition assessment model are trained using a deep learning algorithm to achieve accurate identification of the type of skin disease and objective assessment of the severity. The comprehensive analysis module fuses the outputs of the two models, constructs a fusion decision model, and provides comprehensive diagnostic recommendations. The diagnostic decision module generates a disease type judgment and condition assessment report based on the trained model and combined with real-time image data, effectively assisting doctors in making decisions. The present invention improves the accuracy, efficiency, and popularity of skin disease diagnosis, and meets the complex and changeable diagnostic needs in clinical practice.

Description

Technical Field

[0001] The present invention relates to the technical field of skin disease diagnosis, and specifically to an AI-assisted decision support system for skin disease diagnosis. Background Art

[0002] Skin diseases are common medical problems that cover a wide range of symptoms from mild to severe. Their accurate diagnosis is crucial for timely treatment, symptom relief, and prevention of complications. However, the diagnostic process of skin diseases is often complex and challenging due to the following reasons:

[0003] Disease diversity: There are many types of skin diseases. Different diseases may have similar clinical manifestations, while the same disease may manifest very differently in different individuals, which greatly increases the risk of misdiagnosis.

[0004] Subjectivity of diagnosis: Traditional diagnosis of skin diseases mainly relies on the doctor's visual inspection and empirical judgment. This diagnostic method is easily affected by the doctor's personal experience, knowledge level and subjective feelings, resulting in inconsistent diagnostic results.

[0005] Uneven distribution of resources: Especially in remote areas or areas with limited medical resources, the limited number of professional dermatologists makes it difficult for a large number of patients to obtain timely and accurate diagnostic services.

[0006] Difficulty in disease assessment: Assessment of the severity and development stage of skin diseases usually requires doctors to have extensive clinical experience and careful observation skills. However, the standardization and objectivity of this process are low, which is not conducive to the formulation of personalized treatment plans.

[0007] The rapid development of artificial intelligence technology, especially breakthroughs in computer vision and deep learning technology, has provided new solutions for the diagnosis of skin diseases. These technologies can automatically analyze skin images and identify subtle lesion features, thereby achieving rapid and accurate classification and disease assessment of skin diseases. However, existing AI-assisted diagnosis systems mostly focus on single disease identification and lack the ability to make comprehensive decisions that take into account the patient's basic information, disease type, and disease severity, making it difficult to meet the complex and changing diagnostic needs in clinical practice. Summary of the invention

[0008] The purpose of the present invention is to provide an AI-assisted decision support system for skin disease diagnosis to solve the problems raised in the above-mentioned background technology.

[0009] To achieve the above object, the present invention provides the following technical solution: an AI-assisted decision support system for skin disease diagnosis, the system comprising:

[0010] It includes image acquisition module, image analysis module, comprehensive analysis module and diagnosis decision module;

[0011] The image acquisition module is used to obtain high-definition images of the patient's skin, which is specifically achieved through high-precision skin photography equipment; at the same time, basic information such as the patient's age, gender, and medical history is collected as diagnostic reference data; the skin image data and the diagnostic reference data are integrated and preprocessed to form a skin disease training data set;

[0012] The image analysis module is used to construct a skin disease recognition model and a condition assessment model, wherein the recognition model is used to determine the type of skin disease, and the condition assessment model is used to assess the severity and development stage of the disease; the recognition model and the condition assessment model are trained respectively using a deep learning algorithm based on a skin disease training data set;

[0013] The comprehensive analysis module is used to construct a fusion decision model using the outputs of the skin disease recognition model and the condition assessment model, and the fusion decision model is used to make a comprehensive judgment and give a final diagnosis suggestion;

[0014] The diagnosis decision module is used to utilize the trained model in combination with real-time skin image data to generate skin disease type judgment results and condition assessment reports to assist doctors in making decisions.

[0015] Preferably, a convolutional neural network (CNN) algorithm is used to construct a skin disease recognition model, and lesion features in skin images are extracted through multi-layer convolution and pooling operations, specifically: defining a CNN model, determining the number of convolutional layers, the size of each convolution kernel, the pooling method, the type of activation function, the structure of the fully connected layer, and the classifier of the output layer; wherein the convolutional layer is used to extract local features in the skin image, the pooling layer is used to reduce the dimension of the feature map and retain important information, the fully connected layer is used to map high-dimensional features to the classification space, and the output layer uses a softmax classifier to give the probability distribution of skin disease types.

[0016] Preferably, the steps of training the skin disease recognition model include:

[0017] S1: Input the preprocessed skin disease training data set into the CNN model, and calculate the loss between the model output and the true label through the forward propagation algorithm. The loss function adopts the cross entropy loss function, and its formula is:

[0018]

[0019] Where N is the sample size, is the true label of the i-th sample, is the predicted probability of the ith sample;

[0020] S2: The gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using the gradient descent algorithm. The parameter update formula is:

[0021]

[0022] in, represents the model parameters, including the weights and biases of the convolutional layers, is the learning rate, which is used to control the step size of parameter update. is the loss function with respect to the parameters The gradient of

[0023] S3: Evaluate the recognition accuracy of the trained CNN model on an independent validation set, and adjust the number of layers, convolution kernel size, and learning rate of the CNN model based on the evaluation results.

[0024] Preferably, the changes of skin diseases over time are processed by sequence analysis, and a disease assessment model is constructed to assess the development trend of the disease. The specific steps include:

[0025] Arrange each patient's medical records in chronological order to form medical condition time series data;

[0026] Construct a recurrent neural network RNN ​​structure, determine the number of hidden layers, the number of hidden units in each layer, the type of activation function, the learning rate, and the optimizer; wherein the RNN model is composed of recurrent units, each of which can receive the input at the current moment and the hidden state at the previous moment, and output the hidden state at the current moment, thereby realizing the memory and transmission of time series data; the hidden state update formula of the RNN model is:

[0027]

[0028] in, represents the hidden state at time t, represents the input at time t, and Represent the weight matrices from hidden layer to hidden layer and from input layer to hidden layer, represents the bias term of the hidden layer, represents the activation function;

[0029] The serialized disease data is input into the RNN model, and the change pattern of the disease over time is learned through the memory capacity of the recurrent unit; the time back propagation algorithm BPTT is used to calculate the gradient of the loss function with respect to the model parameters, and the Adam optimizer is used to update the model parameters;

[0030] According to the error between the disease assessment results predicted by the RNN model and the actual disease condition, the parameters and structure of the model are adjusted, including increasing the number of hidden layers, adjusting the number of hidden units, changing the learning rate and introducing regularization terms.

[0031] Preferably, the disease condition data includes lesion size, color, shape, location and patient symptoms.

[0032] Preferably, the step of constructing the fusion decision model includes:

[0033] Get the output of the skin disease recognition model, recorded as the disease type probability distribution ,in Indicates different types of skin diseases, Indicates the probability that the image belongs to each disease type; at the same time, the output of the disease assessment model is obtained and recorded as the disease assessment score , indicating the severity of the skin disease;

[0034] Disease type probability distribution and disease assessment scores With the original skin image data Combined to form a decision feature set Specifically, and As an additional feature, compared with the original skin image The image features extracted from the image are spliced ​​to form a decision feature set ;

[0035] The decision feature set is trained by ensemble learning method. Conduct training and build a fusion decision model; the structure of the fusion decision model includes base learners and a fusion strategy; each base learner receives a decision feature set As input, and output preliminary diagnosis results ; Fusion strategy based on Preliminary diagnosis results of the base learner Make a comprehensive judgment and give the final diagnosis suggestion .

[0036] Preferably, the training process of the base learner includes:

[0037] Using decision feature sets Train each base learner to optimize the model parameters by minimizing the loss function; the loss function uses the mean square error algorithm;

[0038] For each base learner, the k-fold cross-validation method is used to verify the model, and the model with the best performance is selected as the final base learner.

[0039] Preferably, the implementation method of the fusion strategy includes:

[0040] Simple voting method: Count the number of times each disease type is predicted by the base learner, and select the disease type with the most predictions as the final diagnosis recommendation;

[0041] Weighted voting method: assign different weights to each base learner, perform weighted voting based on the weights and preliminary diagnosis results, and select the disease type with the largest weighted voting result as the final diagnosis recommendation;

[0042] Stacking method: Use another logistic regression model as a meta-learner, take the preliminary diagnosis results of the base learner as input features, and train the meta-learner to output the final diagnosis recommendation.

[0043] Preferably, the method of preprocessing the skin image data and the diagnostic reference data includes missing value processing and outlier processing.

[0044] Preferably, the missing value processing method includes:

[0045] For numerical features, the median of the non-missing values ​​of the feature is used to fill the missing values;

[0046] For categorical features, the missing values ​​are filled with the mode of the non-missing values ​​of the feature;

[0047] For missing values ​​in time series data, the average value of the adjacent values ​​is used to fill them;

[0048] For features with obvious business meaning and few missing values, they are manually filled according to business logic;

[0049] The method for handling outliers includes:

[0050] For numerical features, the 3σ principle is used to identify and process outliers; specifically, the mean and standard deviation of the feature are calculated, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] This system combines two deep learning algorithms, convolutional neural network (CNN) and recurrent neural network (RNN), to build a skin disease recognition model and a disease assessment model respectively. CNN can efficiently extract lesion features in skin images and accurately classify skin diseases; while RNN processes the changes of skin diseases over time through sequence analysis and evaluates the development trend of the disease. This dual-model combination greatly improves the accuracy of skin disease diagnosis.

[0053] This system uses an integrated learning method to build a fusion decision model, and through the comprehensive judgment of multiple base learners, it improves the robustness and stability of the system. Even in the face of complex and changeable skin diseases, the system can give reliable diagnostic suggestions.

[0054] This system can quickly process and analyze a large amount of skin image data, give immediate diagnostic suggestions, and greatly improve diagnostic efficiency. This is especially important for areas with limited medical resources and a large number of patients, which can relieve doctors' work pressure and improve the accessibility of medical services. Through the disease assessment model, this system can assess the severity and development stage of skin diseases, providing important reference for doctors to formulate personalized treatment plans. This helps to achieve precision medicine and improve treatment effects and patient satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a working principle diagram of an AI-assisted decision support system for skin disease diagnosis according to the present invention;

[0056] Figure 2 Design schematic diagram of the disease assessment model;

[0057] Figure 3 The figure shows the working principle of the fusion decision model. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] See also Figure 1-3 The present invention provides a technical solution: an AI-assisted decision support system for skin disease diagnosis, the system comprising:

[0060] Image acquisition module: The image acquisition module is mainly responsible for obtaining high-definition images of the patient's skin and collecting relevant basic information. The specific implementation is as follows:

[0061] High-definition image acquisition: Use high-precision skin photography equipment, such as a high-definition digital camera or a dermatoscope, to photograph the patient's skin lesions to ensure that the image clarity and detail richness meet the needs of subsequent analysis.

[0062] Collection of basic information: Through electronic questionnaires or doctor input, collect the patient's age, gender, medical history (including past history of skin diseases, family genetic history, etc.) and other basic information as reference data for diagnosis.

[0063] Data integration and preprocessing: Integrate the acquired skin image data with the patient’s basic information to form a complete skin disease training data set. Preprocessing steps include image denoising, cropping, normalization, etc. to improve the effect of subsequent model training.

[0064] Image analysis module: The image analysis module is responsible for building a skin disease recognition model and a disease assessment model. The specific implementation is as follows:

[0065] Model construction: Based on the deep learning framework, a skin disease recognition model is constructed to extract lesion features in skin images and determine the type of disease. At the same time, a disease assessment model is constructed to process the changes of skin diseases over time through sequence analysis to assess the severity and development stage of the disease.

[0066] Model training: Use the preprocessed skin disease training data set to train the recognition model and condition assessment model. Use the back propagation algorithm and gradient descent algorithm to optimize the model parameters, and improve the accuracy and generalization ability of the model through multiple iterative training.

[0067] Comprehensive analysis module: The comprehensive analysis module is responsible for building a fusion decision model using the outputs of the recognition model and the disease assessment model. The specific implementation is as follows:

[0068] Feature fusion: The disease type probability distribution output by the recognition model and the condition assessment score output by the condition assessment model are fused with the image features extracted from the original skin image data to form a decision feature set.

[0069] Fusion decision model construction: The ensemble learning method is used to train the decision feature set and build a fusion decision model. The fusion decision model can integrate the preliminary diagnosis results of multiple base learners and give the final diagnosis recommendation.

[0070] Diagnosis decision module: The diagnosis decision module is responsible for using the trained model and real-time skin image data to generate diagnosis results and reports. The specific implementation is as follows:

[0071] Real-time image acquisition and processing: The patient's real-time skin image is acquired through the image acquisition module, and necessary preprocessing is performed.

[0072] Model prediction: The preprocessed real-time skin image data is input into the recognition model, condition assessment model and fusion decision model to obtain the disease type judgment, condition assessment score and final diagnosis recommendation respectively.

[0073] Report generation: Based on the model prediction results, a diagnostic report is generated that includes the skin disease type judgment results, condition assessment report and recommended treatment plan, providing auxiliary decision support for doctors.

[0074] The present invention will be further described below in conjunction with Examples 1 to 3:

[0075] Embodiment 1: The following will describe in detail the implementation method of building a skin disease recognition model using a convolutional neural network (CNN) algorithm, including model definition, training steps and parameter adjustment, in conjunction with a specific example.

[0076] ① CNN model definition:

[0077] Build a CNN model for skin disease recognition, which contains multiple convolutional layers, pooling layers, fully connected layers, and an output layer. The specific structure is as follows:

[0078] Input layer: receives the preprocessed skin image, and the image size is set to 224x224x3 (height x width x number of channels, assuming it is an RGB image).

[0079] Convolutional Layer:

[0080] The first convolution layer (Conv1): uses 32 3x3 convolution kernels, a stride of 1, a padding of 'same', and a ReLU activation function.

[0081] The second convolution layer (Conv2): uses 64 3x3 convolution kernels, a step size of 1, a padding method of 'same', and an activation function of ReLU.

[0082] Subsequent convolutional layers can be added as needed, and the number and size of convolution kernels in each layer can be adjusted according to the experimental results.

[0083] Pooling layer: Each convolution layer is followed by a maximum pooling layer (MaxPooling) with a pooling window size of 2x2 and a stride of 2 to reduce the dimension of the feature map and retain important information.

[0084] Fully connected layer: After the last pooling layer, the feature map is flattened into a one-dimensional vector and input into the fully connected layer (FC). Two fully connected layers are set, the first layer contains 128 neurons, the second layer (the layer before the output layer) contains 64 neurons, and the activation function is ReLU.

[0085] Output layer: Use the softmax classifier as the output layer to output the probability distribution of skin disease types. Assuming there are C types of skin diseases, the output layer contains C neurons.

[0086] ② Model training steps:

[0087] The skin disease image dataset was preprocessed, including image scaling, normalization and other operations, to ensure that the image data input into the CNN model had a uniform format and range.

[0088] S1. Forward propagation and loss calculation:

[0089] The preprocessed skin disease training dataset is input into the CNN model, and the loss between the model output and the true label is calculated through the forward propagation algorithm.

[0090] The loss function uses the cross entropy loss function, and the formula is as follows:

[0091]

[0092] Where N is the sample size, is the true label of the i-th sample, is the predicted probability of the ith sample.

[0093] S2, back propagation and parameter update:

[0094] The gradient of the loss function with respect to the model parameters is calculated through the back-propagation algorithm.

[0095] Use gradient descent algorithm (such as stochastic gradient descent SGD, Adam, etc.) to update model parameters. The parameter update formula is as follows:

[0096]

[0097] in, represents the model parameters (including the weights and biases of the convolutional layers), is the learning rate, is the loss function with respect to the parameters gradient.

[0098] S3. Model evaluation and parameter adjustment:

[0099] Evaluate the recognition accuracy of the trained CNN model on an independent validation set. Based on the evaluation results, adjust the CNN model's hyperparameters such as the number of layers, convolution kernel size, and learning rate to optimize the model performance. The optimal hyperparameter combination can be found through grid search, random search, or Bayesian optimization.

[0100] Example 2: This example is used to describe in detail the implementation of a skin disease condition assessment model based on a recurrent neural network (RNN).

[0101] Consider a common skin disease, psoriasis, whose progression is usually manifested by periodic changes in symptoms such as skin erythema and scaling. In order to assess the severity and development trend of psoriasis patients, a disease assessment model based on RNN can be constructed.

[0102] S1. Data preparation:

[0103] Data collection: Medical records of psoriasis patients were obtained from hospitals or research institutions, including descriptions of lesion size, color, morphology, and location at each visit, as well as records of patients' symptoms (such as itching, pain, etc.).

[0104] Data preprocessing: Arrange the collected data in chronological order to form a time series of each patient's condition. For non-numerical data (such as color and morphological descriptions), use coding to convert them into numerical data. At the same time, process missing values, such as using interpolation to fill in or delete records containing missing values.

[0105] Feature extraction: Based on the characteristics of the disease data, extract features that are valuable for disease assessment, such as the rate of change of lesion area, frequency of symptom occurrence, etc.

[0106] S2. Model construction:

[0107] RNN structure definition: Build an RNN model with two hidden layers. Each hidden layer contains 128 hidden units, and ReLU is used as the activation function. The learning rate is initially set to 0.001, and the Adam optimizer is used for parameter update.

[0108] Input layer and output layer design: The input layer receives the preprocessed disease time series data and inputs a feature vector at each time step. The output layer uses a fully connected layer to output a score indicating the severity of the disease. The higher the score, the more serious the disease.

[0109] Introducing the attention mechanism: In order to enhance the model's ability to capture key disease information, the attention mechanism is introduced. At each time step, the correlation between the current hidden state and all historical hidden states is calculated, and the historical information is weighted and summed according to the correlation weight to obtain the context vector containing the key information, which is combined with the current hidden state for subsequent calculations. The hidden state update formula of the RNN model is:

[0110]

[0111] in, represents the hidden state at time t, represents the input at time t, and Represent the weight matrices from hidden layer to hidden layer and from input layer to hidden layer, represents the bias term of the hidden layer, Represents the activation function.

[0112] S3, model training and optimization:

[0113] Loss function selection: The mean square error (MSE) is used as the loss function to measure the difference between the disease score predicted by the model and the actual score.

[0114] Model training: The preprocessed disease time series data is input into the RNN model, the BPTT algorithm is used to calculate the gradient of the loss function with respect to the model parameters, and the model parameters are updated through the Adam optimizer. During the training process, the early stopping method is used to prevent the model from overfitting, that is, the training is stopped when the loss on the validation set no longer decreases.

[0115] Model optimization: According to the evaluation results on the validation set, adjust the parameters and structure of the RNN model. For example, try to increase the number of hidden layers, adjust the number of hidden units, change the learning rate, etc. At the same time, optimize the correlation calculation method in the attention mechanism to improve the model's ability to capture key information.

[0116] S4. Model evaluation and application:

[0117] Model evaluation: Evaluate the performance of the trained RNN model on the test set, including the accuracy and stability of disease assessment. Indicators such as mean square error and accuracy can be used for evaluation.

[0118] Model application: The trained RNN model is applied in actual clinical practice to assist doctors in evaluating and predicting the condition of psoriasis patients. Doctors can formulate more personalized treatment plans based on the evaluation results of the model and the specific conditions of the patients.

[0119] Embodiment 3: The following describes in detail the implementation of the fusion decision model with reference to specific examples:

[0120] Data preparation:

[0121] Skin disease recognition model output: The skin disease recognition model is built based on a convolutional neural network (CNN) and can recognize five common types of skin diseases (d1, d2, d3, d4, d5). For each input skin image, the model outputs a 5-dimensional probability distribution vector , where each element represents the probability that the image belongs to the corresponding disease type.

[0122] Condition assessment model output: The condition assessment model is built based on a recurrent neural network (RNN). It can assess the severity of the current condition based on the patient's historical condition data and output a condition assessment score S ranging from 0 to 100. The higher the score, the more serious the condition.

[0123] Original skin image data: For each skin image, in addition to the outputs of the above two models, the image’s color, texture, shape and other features are extracted to form an image feature vector .

[0124] Construction of decision feature set:

[0125] Disease type probability distribution , disease assessment score S and image feature vector Combined to form a decision feature set F. Specifically, .

[0126] Fusion decision model construction:

[0127] Base learner selection: Three different types of machine learning models are selected as base learners, namely support vector machine (SVM), random forest (RF) and gradient boosted decision tree (GBDT). These models perform well in classification tasks and each has different characteristics, which helps to improve the diversity and accuracy of the fusion decision model.

[0128] Base learner training:

[0129] Each base learner is trained using the decision feature set F. For SVM, a linear kernel function is selected and the penalty parameter C is optimized by grid search; for RF, the number of trees is set to 100 and the optimal number of features is selected by cross-validation; for GBDT, the number of trees is set to 200, the learning rate is 0.1, and the optimal tree depth is selected by cross-validation.

[0130] K-fold cross validation (k=5) was used to evaluate the performance of each base learner, and the model with the best performance was selected as the final base learner.

[0131] Fusion strategy selection: Use weighted voting as the fusion strategy. Different weights are assigned to each base learner according to its performance in cross-validation (such as accuracy, F1 score, etc.). The principle of weight assignment is that the base learner with better performance is assigned a larger weight.

[0132] Fusion decision model training:

[0133] Input the decision feature set F into each base learner to obtain N preliminary diagnosis results. .

[0134] According to the weighted voting method, the weighted voting results of each disease type are calculated, and the disease type with the largest weighted voting result is selected as the final diagnosis recommendation. .

[0135] Assume that there is a new skin image that needs to be diagnosed. First, the skin disease recognition model is used to obtain the disease type probability distribution P(d) = [0.1, 0.3, 0.2, 0.1, 0.3], and the disease assessment model is used to obtain the disease assessment score S = 70. At the same time, features are extracted from the image to obtain These features are combined into a decision feature set F and input into the fusion decision model. After preliminary diagnosis by the three base learners, three preliminary diagnosis results are obtained: = d2, = d5, = d2. Based on the weighted voting method, the final diagnosis recommendation is = d2.

[0136] Through the above steps, the construction and application of the fusion decision model are realized. The model can comprehensively consider the type probability, severity of the disease and image characteristics of skin diseases, and improve the accuracy and robustness of skin disease diagnosis.

[0137] Before building a skin disease recognition model, it is necessary to preprocess the skin image data and diagnostic reference data. Preprocessing mainly includes missing value processing and outlier processing to ensure the quality and consistency of the data, thereby improving the accuracy and reliability of the model. The following describes in detail how to implement these preprocessing methods.

[0138] Missing value handling:

[0139] The following strategies are used to handle missing values ​​in skin image data and diagnostic reference data:

[0140] ① Processing of missing values ​​of numerical features:

[0141] Traverse the numerical features in the data set, and for features with missing values, calculate the median of the non-missing values ​​of the feature.

[0142] Use the calculated median to fill missing values ​​in the feature. For example, if a patient's age feature is missing and the median of the non-missing values ​​of the feature is 35 years old, fill the missing value of the patient's age with 35 years old.

[0143] ② Processing of missing values ​​of categorical features:

[0144] For categorical features, count the mode of non-missing values ​​of the feature, that is, the value with the highest frequency.

[0145] Use the mode to fill missing values ​​in the feature. For example, if a patient's skin type feature is missing and the mode of non-missing values ​​in the feature is "oily skin", fill the missing values ​​of the patient's skin type with "oily skin".

[0146] ③ Processing of missing values ​​in time series data:

[0147] For missing values ​​in time series data, such as missing results of a patient's examination, the average of the adjacent values ​​is used to fill in the missing values. For example, if the patient's second examination result is missing, and the results of the first and third examinations are 8 and 12 respectively, the missing value of the second examination is filled in with (8+12) / 2=10.

[0148] ④ Feature processing with obvious business meaning and few missing values:

[0149] For features with clear business logic and few missing values, they can be filled manually according to business rules. For example, if the dataset contains the feature "whether a certain drug has been used" and there are few missing values, the missing values ​​of the feature can be manually determined and filled based on the patient's medical history or other relevant information.

[0150] Outlier handling:

[0151] For outliers in skin image data and diagnostic reference data, the 3σ principle is mainly used for identification and processing:

[0152] Numerical feature outlier processing:

[0153] For each numerical feature, calculate its mean (μ) and standard deviation (σ).

[0154] According to the 3σ principle, values ​​beyond the mean ±3 times the standard deviation are considered outliers. That is, for any value x, if x <μ - 3σ or x > μ + 3σ, x is considered an outlier.

[0155] For the identified outliers, you can adopt a replacement or deletion strategy. When replacing, you can replace the outliers with the mean, median, or a specific threshold (such as within the limit of the maximum and minimum values). When deleting, directly remove the samples containing outliers.

[0156] For example, if a patient's test result is 200, and the mean of the feature is 50 and the standard deviation is 10, then 200 is considered an outlier. According to business logic, the outlier can be replaced with the mean 50 plus 3 times the standard deviation 30, that is, 80, or the sample can be deleted directly.

[0157] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0158] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-assisted decision support system for skin disease diagnosis, characterized by: It includes image acquisition module, image analysis module, comprehensive analysis module and diagnosis decision module; The image acquisition module is used to obtain high-definition images of the patient's skin, which is specifically achieved through high-precision skin photography equipment; at the same time, basic information of the patient, including age, gender and medical history, is collected as diagnostic reference data; the skin image data and the diagnostic reference data are integrated and preprocessed to form a skin disease training data set; The image analysis module is used to construct a skin disease recognition model and a condition assessment model, wherein the recognition model is used to determine the type of skin disease, and the condition assessment model is used to assess the severity and development stage of the disease; the recognition model and the condition assessment model are trained respectively using a deep learning algorithm based on a skin disease training data set; The comprehensive analysis module is used to construct a fusion decision model using the outputs of the skin disease recognition model and the condition assessment model, and the fusion decision model is used to make a comprehensive judgment and give a final diagnosis suggestion; The diagnosis decision module is used to generate skin disease type judgment results and condition assessment reports by using the trained model in combination with real-time skin image data to assist doctors in making decisions; The steps to build a fusion decision model include: Get the output of the skin disease recognition model, recorded as the disease type probability distribution ,in Indicates different types of skin diseases, Indicates the probability that the image belongs to each disease type; at the same time, the output of the disease assessment model is obtained and recorded as the disease assessment score , indicating the severity of the skin disease; Disease type probability distribution and disease assessment scores With the original skin image data Combined to form a decision feature set Specifically, and As an additional feature, compared with the original skin image The image features extracted from the image are spliced ​​to form a decision feature set ; The decision feature set is trained by ensemble learning method. Conduct training and build a fusion decision model; the structure of the fusion decision model includes base learners and a fusion strategy; each base learner receives a decision feature set As input, and output preliminary diagnosis results ; Fusion strategy based on Preliminary diagnosis results of the base learner Make a comprehensive judgment and give the final diagnosis suggestion .

2. The AI-assisted decision support system for skin disease diagnosis according to claim 1, characterized in that: A convolutional neural network (CNN) algorithm was used to construct a skin disease recognition model. The lesion features in skin images were extracted through multi-layer convolution and pooling operations. Specifically, the CNN model was defined to determine the number of convolutional layers, the size of the convolution kernel in each layer, the pooling method, the type of activation function, the structure of the fully connected layer, and the classifier of the output layer. The convolutional layer was used to extract local features in the skin image, the pooling layer was used to reduce the dimension of the feature map and retain important information, the fully connected layer was used to map high-dimensional features to the classification space, and the output layer used a softmax classifier to give the probability distribution of skin disease types.

3. The AI-assisted decision support system for skin disease diagnosis according to claim 2, characterized in that: The steps to train a skin disease recognition model include: S1: Input the preprocessed skin disease training data set into the CNN model, and calculate the loss between the model output and the true label through the forward propagation algorithm. The loss function adopts the cross entropy loss function, and its formula is: ； Where N is the sample size, is the true label of the i-th sample, is the predicted probability of the ith sample; S2: The gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using the gradient descent algorithm. The parameter update formula is: ； in, represents the model parameters, including the weights and biases of the convolutional layers, is the learning rate, which is used to control the step size of parameter update. is the loss function with respect to the parameters The gradient of S3: Evaluate the recognition accuracy of the trained CNN model on an independent validation set, and adjust the number of layers, convolution kernel size, and learning rate of the CNN model based on the evaluation results.

4. The AI-assisted decision support system for skin disease diagnosis according to claim 1, characterized in that: Through sequence analysis, the changes of skin diseases over time are processed, and a disease assessment model is constructed to evaluate the development trend of the disease. The specific steps include: Arrange each patient's medical records in chronological order to form medical condition time series data; Construct a recurrent neural network RNN ​​structure, determine the number of hidden layers, the number of hidden units in each layer, the type of activation function, the learning rate, and the optimizer; wherein the RNN model consists of recurrent units, each of which can receive the input at the current moment and the hidden state at the previous moment, and output the hidden state at the current moment, thereby realizing the memory and transmission of time series data; the hidden state update formula of the RNN model is: ； in, represents the hidden state at time t, represents the input at time t, and Represent the weight matrices from hidden layer to hidden layer and from input layer to hidden layer, represents the bias term of the hidden layer, represents the activation function; The serialized disease data is input into the RNN model, and the change pattern of the disease over time is learned through the memory capacity of the recurrent unit; the time back propagation algorithm BPTT is used to calculate the gradient of the loss function with respect to the model parameters, and the Adam optimizer is used to update the model parameters; According to the error between the disease assessment results predicted by the RNN model and the actual disease condition, the parameters and structure of the model are adjusted, including increasing the number of hidden layers, adjusting the number of hidden units, changing the learning rate and introducing regularization terms.

5. The AI-assisted decision support system for skin disease diagnosis according to claim 4, characterized in that: The disease condition data includes lesion size, color, shape, location and patient symptoms.

6. The AI-assisted decision support system for skin disease diagnosis according to claim 1, characterized in that: The training process of the base learner includes: Using decision feature sets Train each base learner to optimize the model parameters by minimizing the loss function; the loss function uses the mean square error algorithm; For each base learner, the k-fold cross-validation method is used to verify the model, and the model with the best performance is selected as the final base learner.

7. The AI-assisted decision support system for skin disease diagnosis according to claim 1, characterized in that: The implementation method of the fusion strategy includes: Simple voting method: Count the number of times each disease type is predicted by the base learner, and select the disease type with the most predictions as the final diagnosis recommendation; Weighted voting method: assign different weights to each base learner, perform weighted voting based on the weights and preliminary diagnosis results, and select the disease type with the largest weighted voting result as the final diagnosis recommendation; Stacking method: Use another logistic regression model as a meta-learner, take the preliminary diagnosis results of the base learner as input features, and train the meta-learner to output the final diagnosis recommendation.

8. The AI-assisted decision support system for skin disease diagnosis according to claim 1, characterized in that: The preprocessing methods for skin image data and diagnostic reference data include missing value processing and outlier processing.

9. The AI-assisted decision support system for skin disease diagnosis according to claim 8, characterized in that: The missing value processing method includes: For numerical features, the median of the non-missing values ​​of the feature is used to fill the missing values; For categorical features, the missing values ​​are filled with the mode of the non-missing values ​​of the feature; For missing values ​​in time series data, the average value of the adjacent values ​​is used to fill them; For features with obvious business meaning and few missing values, they are manually filled according to business logic; The method for handling outliers includes: For numerical features, the 3σ principle is used to identify and process outliers; specifically, the mean and standard deviation of the feature are calculated, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers.

Images