Intelligent pigmented disease risk assessment method, device, equipment and medium

Abstract

The invention relates to an intelligent pigmented disease risk assessment method and device, equipment and a medium. The method comprises the following steps: acquiring a model pool containing a plurality of heterogeneous deep learning models, and generating a corresponding feature knowledge base and a calibration error parameter for each model in the model pool based on the same training data set; inputting a to-be-evaluated pigmented lesion image into each model in the model pool for forward reasoning, and outputting a lesion prediction probability vector and a high-dimensional feature vector of each model; calculating a dynamic fusion weight of each model for the to-be-evaluated pigmented lesion image; fusing the lesion prediction probability vector and the dynamic fusion weight of each model according to a preset combination rule, and generating a comprehensive risk probability and a fusion uncertainty; and generating a pigmented disease risk assessment report based on the comprehensive risk probability and the fusion uncertainty. By adopting the method, the robustness and reliability of intelligent evaluation of the pigmented disease risk can be effectively improved, and misdiagnosis and missed diagnosis risks caused by model deviation or uncertainty are reduced.

Description

Technical Field

[0001] This invention belongs to the field of disease risk assessment technology, and in particular relates to intelligent pigmentary disease risk assessment methods, devices, equipment and media. Background Technology

[0002] In the field of intelligent pigmentary disease risk assessment, deep learning-based convolutional neural network models have become mainstream due to their powerful feature extraction capabilities. However, single convolutional neural network models suffer from insufficient generalization, low recognition rates for rare lesions, and overconfidence or lack of confidence in their own predictions. This leads to a decrease in clinicians' trust in automated systems. While ensemble learning can improve overall robustness by combining multiple models, traditional ensemble methods (such as simple voting or averaging) have significant drawbacks: they fuse model results in a static, fixed manner, failing to dynamically adjust the weight of different models based on the specific characteristics of the lesion being assessed. For example, for a lesion with typical melanoma characteristics, all mainstream models may reach a consensus, resulting in a highly reliable ensemble result. However, for a rare lesion with ambiguous boundaries and atypical features, different models may give vastly different or even contradictory predictions based on the biases in their training data. If a fixed weight is still used for ensemble analysis, the reliability of the final result is actually very low, but the system cannot inform the physician of this uncertainty.

[0003] Therefore, it is urgent to solve the problem of how to dynamically assess the credibility of one's own decisions, and based on this, intelligently integrate the prediction results of multiple heterogeneous models to ultimately provide a risk assessment method that is not only accurate but also comes with quantitative credibility indicators. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, equipment, and medium for intelligent pigmentary disease risk assessment to address the aforementioned technical problems.

[0005] Firstly, this application provides a method for assessing the risk of pigmentary disorders, including:

[0006] Obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset;

[0007] The system receives images of pigmented lesions to be evaluated and inputs them into the models in the model pool for forward inference. It outputs the lesion prediction probability vector and high-dimensional feature vector of each model for the pigmented lesion image to be evaluated.

[0008] Based on the lesion prediction probability vector and high-dimensional feature vector output by each model, the dynamic fusion weight of each model for the pigmented lesion image to be evaluated is calculated.

[0009] According to the preset combination rules, the lesion prediction probability vectors and dynamic fusion weights of each model are fused to generate the comprehensive risk probability and fusion uncertainty.

[0010] A risk assessment report for pigmentary diseases is generated based on the comprehensive risk probability and fusion uncertainty.

[0011] Further, the system receives images of pigmented lesions to be evaluated and inputs them into the models in the model pool for forward inference. It outputs a lesion prediction probability vector and a high-dimensional feature vector for each model for the pigmented lesion image to be evaluated, including:

[0012] Receive images of pigmented lesions to be evaluated, and perform color correction on the images of pigmented lesions to be evaluated based on a standard color chart to obtain standardized images of pigmented lesions;

[0013] Each model in the model pool is input into a standardized pigmented lesion image, and the lesion prediction probability of each model for the standardized pigmented lesion image is output. The high-dimensional feature vector of the fully connected layer output of each model is extracted. The lesion prediction probability includes benign probability and malignant probability.

[0014] The benign and malignant probabilities output by each model are combined to obtain the lesion prediction probability vector of each model.

[0015] Furthermore, based on the same training dataset, a corresponding feature knowledge base and calibration error parameters are generated for each model in the model pool, including:

[0016] Obtain a predetermined number of pigmented lesion images with gold standard annotations, and construct a training dataset and an independent test dataset based on the pigmented lesion images; the gold standard annotations include both benign and malignant lesions;

[0017] All images of pigmented lesions in the training dataset are input into each model in the model pool for training, resulting in multiple models with the same learning objective;

[0018] High-dimensional feature vectors are extracted from models that share the same learning objective, and clustering is performed on each high-dimensional feature vector to generate a feature cluster center group for each model; the feature cluster center group serves as the feature knowledge base for the corresponding model.

[0019] The independent validation dataset is input into each model with the same learning objective to generate the expected calibration error for each model.

[0020] Furthermore, based on the lesion prediction probability vector and high-dimensional feature vector of each model, the dynamic fusion weights of each model for the pigmented lesion image to be evaluated are calculated, including:

[0021] Based on the lesion prediction probability vector of each model, the prediction information entropy of each model is calculated.

[0022] Based on the prediction information entropy and calibration error parameters of each model, the internal confidence of each model is calculated using the following formula:

[0023]

[0024] in, For internal confidence level, Let α and β be the predicted information entropy of the model, and α and β be the preset weight coefficients. These are the calibration error parameters for the model;

[0025] Based on the high-dimensional feature vectors of each model and the minimum cosine distance of all cluster centers in the feature knowledge base, the external matching weights of each model are calculated using the following formula:

[0026]

[0027] in, γ represents the external matching weight, and γ is a preset scaling factor. The minimum cosine distance between all cluster centers in the model feature library;

[0028] Based on the internal confidence and external matching weights of each model, the dynamic fusion weights of each model for the image of pigmented lesions to be evaluated are calculated using the following formula:

[0029]

[0030] in, For dynamic weight fusion, For internal confidence level, represents the external matching weight, and n represents the number of models.

[0031] Furthermore, the predicted probability vectors and dynamic fusion weights of each model are fused according to preset combination rules to generate a comprehensive risk probability and fusion uncertainty, including:

[0032] Based on the dynamic fusion weights of each model and the malign and benign probabilities in the predicted probability vector, a basic probability allocation function is constructed for each model. This basic probability allocation function includes the quality allocation for the malign hypothesis, the quality allocation for the benign hypothesis, and the quality allocation for the uncertainty domain.

[0033]

[0034]

[0035]

[0036] in, To allocate quality to malign hypotheses, For dynamic weight fusion, This represents the probability of malignancy. For benign probability, The quality assigned to benign assumptions, The mass assigned to the uncertain region;

[0037] According to the preset combination rules, the basic probability allocation functions of all models are recursively combined in pairs to calculate the global probability allocation.

[0038] The mass assigned to the malign hypothesis is extracted from the global probability allocation as the comprehensive risk probability, and the mass assigned to the uncertainty domain is extracted as the fused uncertainty.

[0039] Furthermore, according to the preset combination rules, the basic probability assignment functions of all models are recursively combined pairwise to calculate the global probability assignment, including:

[0040] The basic probability assignment function of each model is used as model evidence, and the fusion probability assignment is generated by recursively combining the evidence of each model pairwise using the Dempster-Shafer combination rule:

[0041]

[0042]

[0043] in, For fusion probability allocation, , Let K be the conflict coefficient, B and C be the subsets of hypotheses supported by the two models, and A be the hypothesis that we want to verify after fusion.

[0044] The fusion probability allocation generated after all models have been fused is used as the global probability allocation.

[0045] Furthermore, a risk assessment report for pigmentary diseases is generated based on the comprehensive risk probability and fusion uncertainty, including:

[0046] The overall risk probability is compared with a preset risk threshold, and the risk level of pigmentary diseases is determined based on the comparison results;

[0047] The fusion uncertainty is compared with a preset confidence threshold, and a confidence indicator is determined based on the comparison result.

[0048] Integrate pigmentary disease risk levels and confidence indicators into a pigmentary disease risk assessment report.

[0049] Secondly, this application also provides an intelligent pigmented disease risk assessment device, including:

[0050] The model pool construction module is used to obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset.

[0051] The model inference module is used to receive the pigmented lesion image to be evaluated, input the pigmented lesion image to be evaluated into each model in the model pool for forward inference, and output the lesion prediction probability vector and high-dimensional feature vector of each model for the pigmented lesion image to be evaluated.

[0052] The model's dynamic fusion weight generation module is used to calculate the dynamic fusion weight of each model for the image of pigmented lesions to be evaluated based on the lesion prediction probability vector and high-dimensional feature vector output by each model.

[0053] The module for generating comprehensive risk probability and fusion uncertainty is used to fuse the lesion prediction probability vectors and dynamic fusion weights of each model according to preset combination rules to generate comprehensive risk probability and fusion uncertainty.

[0054] The assessment report generation module is used to generate risk assessment reports for pigmentary diseases based on comprehensive risk probability and fusion uncertainty.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement any of the intelligent pigmentary disease risk assessment methods described in the embodiments of this application.

[0056] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the intelligent pigmentary disease risk assessment method described in any of the embodiments of this application.

[0057] The aforementioned intelligent pigmented disease risk assessment method, device, equipment, and medium acquire a model pool containing multiple heterogeneous deep learning models and generate corresponding feature knowledge bases and calibration error parameters for each model in the pool based on the same training dataset. The image of the pigmented lesion to be assessed is input into each model for forward inference, outputting the lesion prediction probability vector and high-dimensional feature vector for each model for the image. Then, the dynamic fusion weights of each model for the image are calculated. The lesion prediction probability vectors and dynamic fusion weights of each model are fused according to preset combination rules to generate a comprehensive risk probability and fusion uncertainty. Finally, a pigmented disease risk assessment report is generated based on the comprehensive risk probability and fusion uncertainty. This effectively improves the robustness, reliability, and clinical applicability of intelligent risk assessment for pigmented diseases, reducing the risk of misdiagnosis and missed diagnosis due to model bias or uncertainty. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a flowchart illustrating a method for assessing the risk of intelligent pigmented diseases in one embodiment;

[0060] Figure 2 This is a flowchart illustrating the steps of receiving an image of a pigmented lesion to be evaluated, inputting the image of the pigmented lesion to be evaluated into each model in the model pool for forward inference, and outputting the lesion prediction probability vector and high-dimensional feature vector of each model for the image of the pigmented lesion to be evaluated in one embodiment.

[0061] Figure 3 This is a schematic diagram of the structure of an intelligent pigmented disease risk assessment device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] In one embodiment, a method for intelligent pigmentary disease risk assessment is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0064] Step S101: Obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset.

[0065] Among them, heterogeneous deep learning models refer to models with different architectures. Different architectures have different feature extraction capabilities and can cover multi-dimensional features such as local texture and global structure of pigmented lesion images, such as convolutional neural networks and visual Transformers.

[0066] For example, a model pool containing multiple heterogeneous deep learning models is obtained, and each model in the pool is processed using the same training dataset. High-dimensional feature vectors from the fully connected layers during the training of each model are extracted. These high-dimensional feature vectors are then grouped using a clustering algorithm to obtain feature cluster centers. These centers constitute the feature knowledge base of the corresponding model. Simultaneously, an independent test dataset, which has no overlap with the training dataset but also contains gold-standard annotations, is input into each model to generate the model's expected calibration error. The clustering algorithm, without pre-defined category labels, automatically divides the dataset into multiple "clusters" by analyzing the feature similarity of the data itself, ensuring that data points within the same cluster are as similar as possible, and data points between different clusters are as different as possible.

[0067] Step S102: Receive the image of the pigmented lesion to be evaluated, input the image of the pigmented lesion to be evaluated into each model in the model pool for forward inference, and output the lesion prediction probability vector and high-dimensional feature vector of each model for the image of the pigmented lesion to be evaluated.

[0068] Forward reasoning refers to starting from "known conditions" and gradually deriving a "conclusion," similar to the human thought process of analyzing problems step by step based on facts.

[0069] For example, the image of the pigmented lesion to be evaluated is received and input into each model in the model pool for forward inference. The lesion prediction probability of each model for the image is output, including benign probability and malignant probability. The two are combined to form the lesion prediction probability vector of the model. At the same time, the high-dimensional feature vector that can characterize the global features of the image is extracted from the output of the fully connected layer of each model.

[0070] Step S103: Based on the lesion prediction probability vector and high-dimensional feature vector output by each model, calculate the dynamic fusion weight of each model for the pigmented lesion image to be evaluated.

[0071] For example, based on the lesion prediction probability vector and high-dimensional feature vector output by each model, the internal confidence and external matching weights of each model are calculated by combining the calibration error that reflects the reliability of the model itself and the feature knowledge base of each model. The calculated internal confidence and external matching weights are multiplied and normalized to generate the dynamic fusion weights of each model.

[0072] Step S104: According to the preset combination rules, the lesion prediction probability vectors and dynamic fusion weights of each model are fused to generate the comprehensive risk probability and fusion uncertainty.

[0073] Among them, the pre-set combination rules refer to a set of standardized mathematical logic and operational procedures that are pre-defined in order to achieve multi-model evidence fusion and generate comprehensive evaluation results.

[0074] For example, based on the lesion prediction probability vector and dynamic fusion weight of each model, a basic probability allocation function is constructed for each model. The basic probability allocation functions of all models are recursively combined in pairs using a preset combination rule until all models are fused to obtain a global probability allocation. Finally, the comprehensive risk probability and fusion uncertainty are extracted from the global probability allocation.

[0075] Step S105: Generate a risk assessment report for pigmentary diseases based on the comprehensive risk probability and fusion uncertainty.

[0076] For example, the overall risk probability is compared with a preset risk threshold to determine the risk level of pigmentary diseases. The fusion uncertainty is then compared with a preset confidence threshold to determine the uncertainty of the fusion results of each model. The pigmentary disease risk level and the uncertainty are integrated to generate a structured pigmentary disease risk assessment report. The preset risk threshold is a key reference standard used to convert the overall risk probability into a risk level (e.g., low, medium, high risk) that can directly guide clinical judgment; it is a probability cutoff point set according to clinical diagnostic needs. The preset confidence threshold is a benchmark parameter pre-set based on clinical diagnostic needs, model performance characteristics, and historical assessment data, used to judge the reliability of the overall assessment results: when the fusion uncertainty is below this threshold, it indicates high consistency among the models and low overall assessment uncertainty; when the fusion uncertainty is above this threshold, it indicates significant conflict between models or poor adaptation of some models to the current lesion characteristics, resulting in insufficient reliability of the overall assessment.

[0077] In this embodiment, feature knowledge bases and calibration error parameters are generated for each model in the heterogeneous model pool using the same training dataset. The pigmented lesion image to be evaluated is input into each model to extract high-dimensional feature vectors and lesion prediction probabilities for the image. The dynamic fusion weights of the model are calculated by calculating the internal confidence and external matching weights. The comprehensive risk probability is generated according to the preset combination rules, and the fusion uncertainty is output. The final evaluation report can effectively improve the robustness, reliability and clinical applicability of intelligent evaluation, and reduce the risk of misdiagnosis and missed diagnosis caused by model bias or uncertainty.

[0078] In one embodiment, such as Figure 2 As shown, the system receives images of pigmented lesions to be evaluated and inputs them into the models in the model pool for forward inference. It outputs a lesion prediction probability vector and a high-dimensional feature vector for each model for the pigmented lesion image to be evaluated, including:

[0079] Step S201: Receive the image of the pigmented lesion to be evaluated, and perform color correction on the image of the pigmented lesion to be evaluated based on the standard color chart to obtain a standardized image of the pigmented lesion.

[0080] The standard color chart contains multiple sets of standard color patches with known RGB (Red Green Blue) values.

[0081] For example, an image of a pigmented lesion to be evaluated is received, and the pixel values ​​of the corresponding color blocks in the image are compared with those of a standard color chart to establish a color mapping relationship. This corrects for color shifts caused by differences in the acquisition device, resulting in a standardized image of the pigmented lesion. The color mapping relationship refers to adjusting parameters such as hue, saturation, and brightness of the image to be evaluated, using the color reference of the standard color chart as the target.

[0082] Step S202: Input the standardized pigmented lesion image into each model in the model pool, output the lesion prediction probability of each model for the standardized pigmented lesion image; and extract the high-dimensional feature vector of the fully connected layer output of each model; the lesion prediction probability includes benign probability and malignant probability.

[0083] For example, the obtained standardized pigmented lesion images are successively input into each model in a pre-built model pool for forward inference: Features are extracted from the standardized images through convolutional layers or attention layers, gradually capturing low-dimensional features such as local texture of the lesions (e.g., whether the boundaries are regular) and pigment distribution (e.g., whether a pigment network exists), and feature dimensionality reduction and key feature preservation are performed through pooling layers; the extracted features are then fed into fully connected layers, which integrate the low-dimensional features into high-dimensional feature vectors, simultaneously extracting the high-dimensional feature vectors output by the fully connected layers of each model. Simultaneously, each model maps the high-dimensional feature vectors to probability values ​​through its output layer, outputting the benign and malignant probabilities for the current standardized image, respectively.

[0084] Step S203: Combine the benign and malignant probabilities output by each model to obtain the lesion prediction probability vector of each model.

[0085] For example, for each model, the output benign probability and malignant probability are structurally combined to generate the lesion prediction probability vector corresponding to that model. The probability vector is constructed using a fixed dimension and order, for example, uniformly set as a two-dimensional vector form of [benign probability, malignant probability].

[0086] In this embodiment, the pigmented lesion image to be evaluated is color-corrected; the color-corrected image is input into each model in the model pool for forward inference to obtain the probability of each model determining whether the image lesion is benign or malignant and a high-dimensional feature vector that can characterize the global features of the lesion. The probability of benign or malignant is constructed into a standardized lesion prediction probability vector, which can effectively improve the accuracy and reliability of the data.

[0087] In one embodiment, a corresponding feature knowledge base and calibration error parameters are generated for each model in the model pool based on the same training dataset, including:

[0088] Step S301: Obtain a preset number of pigmented lesion images with gold standard annotations, and construct a training dataset and an independent test dataset based on the pigmented lesion images; the gold standard annotations include benign and malignant.

[0089] For example, a predetermined number of images of pigmented lesions are acquired, and all images must carry gold standard annotations. The gold standard annotations refer to the lesion nature verified by clinical pathological examination (such as tissue biopsy), including only benign and malignant types. For the acquired pigmented lesion images, invalid images with blurriness, inconsistent resolution, or missing annotation information are removed to obtain the valid images. These valid images are then divided into a training dataset with no data overlap and an independent test dataset. The division of valid images can be based on image acquisition time, patient group stratification, etc., to avoid overfitting caused by data from a single source.

[0090] Step S302: Input all the pigmented lesion images in the training dataset into each model in the model pool for training, and obtain multiple models with the same learning objective.

[0091] For example, all images of pigmented lesions in the training dataset are input into each model in the model pool. All models are given the same learning objective: to perform a binary classification task of pigmented lesions and output a prediction of benign or malignant. A uniform loss function, optimizer, and training hyperparameters are used to ensure that all models learn under a consistent training benchmark. During training, the training loss and training set accuracy of each model are monitored in real time. Training is stopped when the loss stabilizes and the accuracy no longer significantly improves, resulting in multiple models with the same learning objective but different feature extraction capabilities due to architectural differences. Among them, the loss function can be the cross-entropy loss function, which is used to measure the difference between the model's predicted probability and the gold standard label; the optimizer is the core algorithmic tool used to minimize the loss function during model training. By adjusting model parameters such as weights and biases, it allows the model's prediction results to gradually approach the true label, achieving the goal of the model "learning" the data patterns, such as the Adam (Adaptive Moment Estimation) optimizer; training hyperparameters are parameters that are manually set before the machine learning model is trained, rather than being learned autonomously during the model training process. They are used to regulate the model training process and affect the final performance and convergence effect of the model, such as the number of iterations and the learning rate decay strategy.

[0092] Step S303: Extract the high-dimensional feature vectors of models that have the same learning objective, and perform clustering processing on each high-dimensional feature vector to generate a feature clustering center group for each model; the feature clustering center group is the feature knowledge base of the corresponding model.

[0093] For example, after the model training is completed, high-dimensional feature vectors of each model for all images in the training dataset are extracted. Clustering is performed separately on the high-dimensional feature vectors of each model. The similarity between feature vectors (such as Euclidean distance) is calculated using a clustering algorithm. Vectors with similar features are grouped into the same category, generating several feature cluster centers. All cluster centers constitute the feature knowledge base of the model. This knowledge base is essentially a "set of lesion features familiar to the model".

[0094] Step S304: Input the independent validation dataset into each model with the same learning objective to generate the expected calibration error for each model.

[0095] Among them, the expected calibration error is the deviation between the model's predicted probability and the actual accuracy. If the expected calibration error is small, it indicates that the model's predicted probability can truly reflect its prediction accuracy; if the expected calibration error is large, it indicates that the model has the problem of "overconfidence" or "overconservatism".

[0096] For example, independent test datasets are input into models with the same learning objective, and forward inference is performed to output the lesion prediction probability and corresponding prediction category for each image. The lesion prediction probability of each model is divided into several intervals (such as 0-0.1, 0.1-0.2, etc.). The number of samples in each interval and the number of correctly predicted samples in that interval (i.e., samples whose predicted category is consistent with the gold standard label) are counted. The absolute difference between the actual accuracy (number of correct samples / number of samples in the interval) and the average prediction probability of the interval is calculated. Then, the absolute differences of all intervals are weighted and summed using the proportion of the number of samples in each interval to the total number of test samples to obtain the expected calibration error of each model.

[0097] In this embodiment, each model is trained using a gold standard annotation to unify the learning objective and training benchmark. A feature knowledge base is generated by extracting high-dimensional feature vectors from all images in the training dataset for each model and performing clustering. Each image from the independent test dataset is input into each model to generate the lesion prediction probability for each image. The expected calibration error of each model is calculated by combining the gold standard annotation from the independent test dataset. This effectively avoids evaluation errors caused by poor data quality, inconsistent model benchmarks, lack of feature references, or lack of calibration basis, thus improving the accuracy and robustness of the overall risk assessment.

[0098] In one embodiment, based on the lesion prediction probability vector and high-dimensional feature vector of each model, the dynamic fusion weights of each model for the pigmented lesion image to be evaluated are calculated, including:

[0099] Step S401: Calculate the prediction information entropy of each model based on the lesion prediction probability vector of each model.

[0100] For example, based on the lesion prediction probability vector of each model, the information entropy of the prediction probability of each model is calculated:

[0101]

[0102] in, To predict information entropy, The probability of being classified as good. This represents the probability of being classified as malignant.

[0103] Step S402: Based on the prediction information entropy and calibration error parameters of each model, calculate the internal confidence of each model using the following formula:

[0104]

[0105] in, For internal confidence level, Let α and β be the predicted information entropy of the model, and α and β be the preset weight coefficients. These are the calibration error parameters for the model.

[0106] The preset weighting coefficients α and β are used to balance the influence of prediction information entropy and calibration error parameters on the internal confidence level based on the actual application scenario. For example, if more emphasis is needed on the certainty of model prediction, the value of α can be increased; if more emphasis is needed on the historical calibration performance of the model, the value of β can be increased. The value of the internal confidence level directly reflects the reliability of the model's own prediction results: the closer the value is to 1, the higher the certainty of the model's current prediction and the smaller the historical calibration error, and the stronger its own reliability; the closer the value is to 0, the higher the uncertainty of the model's current prediction or the poor historical calibration performance, and the weaker its own reliability.

[0107] For example, the internal confidence of each model is calculated based on the prediction information entropy and calibration error parameters of each model.

[0108] Step S403: Based on the high-dimensional feature vectors of each model and the minimum cosine distance of all cluster centers in the feature knowledge base, calculate the external matching weight of each model using the following formula:

[0109]

[0110] in, γ represents the external matching weight, and γ is a preset scaling factor. It is the minimum cosine distance between all cluster centers in the model feature library.

[0111] The external matching weight is used to measure the model's suitability for the image features of the pigmented lesions to be evaluated.

[0112] For example, the similarity between the high-dimensional feature vectors of each model and all cluster centers in the feature knowledge base can be calculated using the cosine distance formula. The smaller the value, the closer the directions of the two vectors are, meaning that the features of the image to be evaluated are more similar to the "typical lesion features" (features represented by cluster centers) learned by the model during training. The minimum cosine distance is selected from all cosine distances and substituted into the exponential formula containing a preset scaling factor γ. The exponential function converts the distance value into a weight value in the interval [0,1]. The smaller the minimum cosine distance, the greater the weight of the external matching degree, which means that the model is more adaptable to the current image features, and its prediction results should have a higher say in the fusion.

[0113] Step S404: Based on the internal confidence and external matching weights of each model, the dynamic fusion weights of each model for the pigmented lesion image to be evaluated are calculated using the following formula:

[0114]

[0115] in, For dynamic weight fusion, For internal confidence level, represents the external matching weight, and n represents the number of models.

[0116] For example, based on the internal confidence and external matching weights of each model, the product of "internal confidence × external matching weight" for each individual model is first calculated. This product combines the reliability of the model's own predictions with the model's adaptability to the current image. Then, the sum of the above products of all models is calculated. The product of each individual model is divided by this sum to complete the normalization process, making the sum of the dynamic fusion weights of all models equal to 1, ensuring that the weights meet the basic requirements of probability allocation. Unlike traditional fixed weights, these dynamic fusion weights are adjusted according to the changes in the features of the image to be evaluated. For example, a model with high adaptability to image A features has a high weight, while one with low adaptability to image B features has a low weight, thus avoiding the defect of fixed weights being unable to adapt to different image features.

[0117] In this embodiment, the prediction information entropy of each model is calculated using the lesion prediction probability vector output by each model; internal confidence is generated by combining historical calibration errors; external matching weights are calculated using feature similarity; and the internal confidence and external weight coefficients are integrated to generate dynamically adjusted fusion weights, which can ensure that the weights of each model can accurately match the features and model performance of the current image to be evaluated.

[0118] In one embodiment, the predicted probability vectors and dynamic fusion weights of each model are fused according to a preset combination rule to generate a comprehensive risk probability and fusion uncertainty, including:

[0119] Step S501: Based on the dynamic fusion weights of each model and the malignant and benign probabilities in the predicted probability vector, construct the basic probability allocation function for each model; the basic probability allocation function includes the quality allocation of the malignant hypothesis, the quality allocation of the benign hypothesis, and the quality allocation of the uncertainty region:

[0120]

[0121]

[0122]

[0123] in, To allocate quality to malign hypotheses, For dynamic weight fusion, This represents the probability of malignancy. For benign probability, The quality assigned to benign assumptions, The mass assigned to the uncertain region.

[0124] For example, based on the dynamic fusion weights of each model and the malignant and benign probabilities in the lesion prediction probability vector output by the model, a corresponding Basic Probability Assignment (BPA) function is constructed. The constructed BPA function must satisfy the constraint that "the sum of all trust qualities is 1". Here, BPA is a core element of the Dempster-Shafer (DS) evidence theory, used to transform the model's predictive information into a trust assignment of the two types of hypotheses, "malignant" and "benign", and the "uncertainty domain".

[0125] Step S502: According to the preset combination rules, the basic probability allocation functions of all models are recursively combined in pairs to calculate the global probability allocation.

[0126] For example, according to the preset combination rules, the BPA functions of any two models are first selected as the initial fusion objects for fusion to generate a fusion quality assignment. This fusion quality assignment is then used as the new BPA function, and the above fusion process is repeated with the BPA function of the next unfused model until the BPA functions of all models have been fused to obtain a global probability assignment containing global trust assignment information.

[0127] Step S503: Extract the mass assigned to the malign hypothesis from the global probability allocation as the comprehensive risk probability, and extract the mass assigned to the uncertainty domain as the fused uncertainty.

[0128] Among them, the overall risk probability integrates the predictions and dynamic weights of all models, comprehensively reflecting the likelihood that the pigmented lesion under evaluation is malignant. The larger the value, the higher the probability that the lesion has a malignant risk. The fusion uncertainty integrates the uncertain parts of all models, quantifying the credibility of the overall evaluation results. The smaller the value, the higher the consistency of opinions among the models, the lower the uncertainty of the overall evaluation, and the stronger the credibility of the results. The larger the value, the more conflicts there are between the models or the overall reliability is insufficient, and further verification with other clinical information is required.

[0129] For example, the trust quality assigned to the "malicious hypothesis" in the global probability allocation is directly extracted and defined as the comprehensive risk probability. Simultaneously, the trust quality assigned to the "uncertainty domain" in the global probability allocation is extracted. The quality of trust is defined as the fusion uncertainty.

[0130] In this embodiment, by constructing a BPA function for each model, using the DS combination rule to recursively fuse the BPA functions of all models in pairs to generate a global probability allocation, and extracting the comprehensive risk probability and fusion uncertainty from the global probability allocation, the shortcomings of traditional assessments that only output risk results and lack credibility indicators can be effectively compensated for, thereby improving the accuracy and reliability of risk assessment for pigmentary diseases.

[0131] In one embodiment, according to a preset combination rule, the basic probability allocation functions of all models are recursively combined pairwise to calculate the global probability allocation, including:

[0132] Step S601: The basic probability assignment function of each model is used as model evidence, and the fusion probability assignment is generated by recursively combining the model evidence pairwise using the Dempster-Shafer combination rule.

[0133]

[0134]

[0135] in, For fusion probability allocation, , Let K be the two model evidences to be fused, K be the conflict coefficient, B and C be the subsets of hypotheses supported by the two models, and A be the hypothesis that we want to verify after fusion.

[0136] For example, the basic probability assignment function of each model is used as model evidence. First, any two pieces of model evidence to be fused are selected and denoted as follows: and By traversing All hypothesis subsets B and Given all the hypothetical subsets C, summing them up such that "the intersection of B and C is empty, meaning B and C represent mutually exclusive hypotheses") and If the product of K and C is K≠1, meaning there exist non-conflicting opinions, then we can sum over all subsets that satisfy the condition "the intersection of B and C is A". and The product of the two models is then divided by the normalization factor "1-K" to complete the fusion of the evidence from the two models and obtain the fusion probability allocation. The K value directly reflects the intensity of the conflict between the two models. The larger the K value, the more significant the difference in the two models' judgments on the nature of the lesion. The role of "1-K" is to eliminate the interference of the conflicting parts on the fusion result and ensure that the sum of the trust quality of all hypotheses after fusion is 1.

[0137] Step S602: The fusion probability allocation generated after all models have been fused is used as the global probability allocation.

[0138] For example, the generated intermediate fusion probability is assigned. For the new fusion object, repeat the DS combination process with the next model evidence that has not participated in the fusion: first calculate the conflict coefficient K between the new fusion object and the model evidence to be added, determine whether there is a conflict and perform normalization processing to obtain the updated fusion probability allocation. The recursive fusion process continues until the evidence from all models in the model pool has been fused. That is, each round of fusion combines the integrated multi-model information with the information from a single new model. The final fusion probability assignment is the global probability assignment, which includes the global confidence in the "malicious hypothesis", "benign hypothesis" and "uncertainty domain (Θ)".

[0139] In this embodiment, the Dempster-Shafer combination rule is used to achieve pairwise fusion of model evidence. Through recursive fusion, all model evidence is gradually integrated to generate a global probability assignment. This approach can combine the advantages of all models, avoid the problem of insufficient generalization of a single model, and effectively improve the comprehensiveness and reliability of risk assessment for pigmentary diseases.

[0140] In one embodiment, a risk assessment report for pigmentary diseases is generated based on the comprehensive risk probability and fusion uncertainty, including:

[0141] Step S701: Compare the overall risk probability with the preset risk threshold, and determine the risk level of pigmentary diseases based on the comparison results.

[0142] For example, the extracted comprehensive risk probability is compared with the preset risk threshold one by one: if the comprehensive risk probability is lower than the low risk threshold, the lesion is judged as "low risk", indicating that the lesion is less likely to be malignant, and clinical practice usually recommends regular follow-up observation; if the comprehensive risk probability is between the low risk threshold and the high risk threshold, it is judged as "medium risk", indicating that the nature of the lesion is uncertain, and further judgment is required by combining detailed observation with dermoscopy or imaging examination; if the comprehensive risk probability is higher than the high risk threshold, it is judged as "high risk", which means that the probability of the lesion is malignant is higher, and clinical practice should prioritize pathological biopsy to clarify the diagnosis.

[0143] Step S702: Compare the fusion uncertainty with a preset confidence threshold, and determine the confidence indicator based on the comparison result.

[0144] For example, the extracted fusion uncertainty is compared with a preset confidence threshold: if the fusion uncertainty is lower than the confidence threshold, a "high confidence" indicator is generated, which means that the comprehensive risk probability can truly reflect the actual risk of the lesion and can be used as an important reference for clinical decision-making; if the fusion uncertainty is higher than the confidence threshold, a "low confidence" indicator is generated, along with a prompt message to remind doctors that the current assessment results have limitations and that they should avoid relying solely on the results to make a diagnosis. For example, there may be significant conflicts between the models, and it is recommended to combine the results with other clinical examinations for verification.

[0145] Step S703 integrates the risk level and confidence index of pigmentary diseases into a pigmentary disease risk assessment report.

[0146] The structure of the risk assessment report for pigmentary diseases follows clinical document reading habits. It typically includes the acquisition time, imaging equipment model, basic information about the images to be assessed, a risk assessment result clearly marked with the risk level and corresponding comprehensive risk probability, a statement of the reliability of the result with a confidence index and fusion uncertainty value, and targeted clinical recommendations based on the risk level and confidence index.

[0147] For example, the risk level of a pigmentary disease is structurally integrated with the generated credibility indicator, while supplementing key assessment criteria, such as: the name of the model with the highest contribution, the benign / malignant probability of the model, the range of conflict coefficients in the fusion process, etc., to generate a standardized risk assessment report for pigmentary diseases.

[0148] In this embodiment, the abstract comprehensive risk probability is transformed into a concrete risk level by setting a clinical threshold, and the reliability of the assessment results is quantified by the credibility indicator. The risk level, credibility, and key evidence are integrated to generate a pigmentary disease risk assessment report, which can effectively solve the shortcomings of traditional intelligent assessment that "only gives results, but not evidence and credibility", and effectively improve the clinical practicality, reliability and operability of intelligent pigmentary disease risk assessment.

[0149] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0150] Based on the same inventive concept, this application also provides an intelligent pigmentary disease risk assessment device for implementing the intelligent pigmentary disease risk assessment method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the intelligent pigmentary disease risk assessment device provided below can be found in the limitations of the intelligent pigmentary disease risk assessment method described above, and will not be repeated here.

[0151] In one exemplary embodiment, such as Figure 3 As shown, a smart pigmented lesion risk assessment device 300 is provided, including:

[0152] The model pool construction module 301 is used to obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset.

[0153] The model inference module 302 is used to receive the pigmented lesion image to be evaluated, input the pigmented lesion image to be evaluated into each model in the model pool for forward inference, and output the lesion prediction probability vector and high-dimensional feature vector of each model for the pigmented lesion image to be evaluated.

[0154] The model's dynamic fusion weight generation module 303 is used to calculate the dynamic fusion weight of each model for the pigmented lesion image to be evaluated based on the lesion prediction probability vector and high-dimensional feature vector output by each model.

[0155] The comprehensive risk probability and fusion uncertainty generation module 304 is used to fuse the lesion prediction probability vectors and dynamic fusion weights of each model according to the preset combination rules to generate comprehensive risk probability and fusion uncertainty.

[0156] The assessment report generation module 305 is used to generate a risk assessment report for pigmentary diseases based on the comprehensive risk probability and fusion uncertainty.

[0157] In one embodiment, the model inference module 302 is further configured to:

[0158] Receive images of pigmented lesions to be evaluated, and perform color correction on the images of pigmented lesions to be evaluated based on a standard color chart to obtain standardized images of pigmented lesions;

[0159] Each model in the model pool is input into a standardized pigmented lesion image, and the lesion prediction probability of each model for the standardized pigmented lesion image is output. The high-dimensional feature vector of the fully connected layer output of each model is extracted. The lesion prediction probability includes benign probability and malignant probability.

[0160] The benign and malignant probabilities output by each model are combined to obtain the lesion prediction probability vector of each model.

[0161] In one embodiment, the model pool construction module 301 is further configured to:

[0162] Obtain a predetermined number of pigmented lesion images with gold standard annotations, and construct a training dataset and an independent test dataset based on the pigmented lesion images; the gold standard annotations include both benign and malignant lesions;

[0163] All images of pigmented lesions in the training dataset are input into each model in the model pool for training, resulting in multiple models with the same learning objective;

[0164] High-dimensional feature vectors are extracted from models that share the same learning objective, and clustering is performed on each high-dimensional feature vector to generate a feature cluster center group for each model; the feature cluster center group serves as the feature knowledge base for the corresponding model.

[0165] The independent validation dataset is input into each model with the same learning objective to generate the expected calibration error for each model.

[0166] In one embodiment, the dynamic fusion weight generation module 303 of the model is further configured to:

[0167] Based on the lesion prediction probability vector of each model, the prediction information entropy of each model is calculated.

[0168] Based on the prediction information entropy and calibration error parameters of each model, the internal confidence of each model is calculated using the following formula:

[0169]

[0170] in, For internal confidence level, Let α and β be the predicted information entropy of the model, and α and β be the preset weight coefficients. These are the calibration error parameters for the model;

[0171] Based on the high-dimensional feature vectors of each model and the minimum cosine distance of all cluster centers in the feature knowledge base, the external matching weights of each model are calculated using the following formula:

[0172]

[0173] in, γ represents the external matching weight, and γ is a preset scaling factor. The minimum cosine distance between all cluster centers in the model feature library;

[0174] Based on the internal confidence and external matching weights of each model, the dynamic fusion weights of each model for the image of pigmented lesions to be evaluated are calculated using the following formula:

[0175]

[0176] in, For dynamic weight fusion, For internal confidence level, represents the external matching weight, and n represents the number of models.

[0177] In one embodiment, the combined risk probability and fusion uncertainty generation module 304 is further configured to:

[0178] Based on the dynamic fusion weights of each model and the malign and benign probabilities in the predicted probability vector, a basic probability allocation function is constructed for each model. This basic probability allocation function includes the quality allocation for the malign hypothesis, the quality allocation for the benign hypothesis, and the quality allocation for the uncertainty domain.

[0179]

[0180]

[0181]

[0182] in, To allocate quality to malign hypotheses, For dynamic weight fusion, This represents the probability of malignancy. For benign probability, The quality assigned to benign assumptions, The mass assigned to the uncertain region;

[0183] According to the preset combination rules, the basic probability allocation functions of all models are recursively combined in pairs to calculate the global probability allocation.

[0184] The mass assigned to the malign hypothesis is extracted from the global probability allocation as the comprehensive risk probability, and the mass assigned to the uncertainty domain is extracted as the fused uncertainty.

[0185] In one embodiment, the combined risk probability and fusion uncertainty generation module 304 is further configured to:

[0186] The basic probability assignment function of each model is used as model evidence, and the fusion probability assignment is generated by recursively combining the evidence of each model pairwise using the Dempster-Shafer combination rule:

[0187]

[0188]

[0189] in, For fusion probability allocation, , Let K be the conflict coefficient, B and C be the subsets of hypotheses supported by the two models, and A be the hypothesis that we want to verify after fusion.

[0190] The fusion probability allocation generated after all models have been fused is used as the global probability allocation.

[0191] In one embodiment, the evaluation report generation module 305 is further configured to:

[0192] The overall risk probability is compared with a preset risk threshold, and the risk level of pigmentary diseases is determined based on the comparison results;

[0193] The fusion uncertainty is compared with a preset confidence threshold, and a confidence indicator is determined based on the comparison result.

[0194] Integrate pigmentary disease risk levels and confidence indicators into a pigmentary disease risk assessment report.

[0195] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the intelligent pigmented disease risk assessment method as described above.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0197] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0198] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent risk assessment of pigmentary disorders, characterized in that, The method includes: Obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset; Receive the image of the pigmented lesion to be evaluated, and input the image of the pigmented lesion to be evaluated into each model in the model pool for forward inference, and output the lesion prediction probability vector and high-dimensional feature vector of each model for the image of the pigmented lesion to be evaluated; Based on the lesion prediction probability vector and the high-dimensional feature vector output by each model, calculate the dynamic fusion weight of each model for the pigmented lesion image to be evaluated; According to the preset combination rules, the lesion prediction probability vectors of each model and the dynamic fusion weights are fused to generate a comprehensive risk probability and fusion uncertainty. A risk assessment report for pigmentary diseases is generated based on the comprehensive risk probability and the fusion uncertainty.

2. The method according to claim 1, characterized in that, The process involves receiving an image of a pigmented lesion to be evaluated, inputting the image into each model in the model pool for forward inference, and outputting a lesion prediction probability vector and a high-dimensional feature vector for each model for the image of the pigmented lesion to be evaluated, including: Receive images of pigmented lesions to be evaluated, and perform color correction on the images of pigmented lesions to be evaluated based on a standard color chart to obtain standardized images of pigmented lesions; The standardized pigmented lesion image is input into each model in the model pool, and the lesion prediction probability of each model for the standardized pigmented lesion image is output; and the high-dimensional feature vector of the fully connected layer output of each model is extracted; the lesion prediction probability includes benign probability and malignant probability; The benign probability and malignant probability output by each model are combined to obtain the lesion prediction probability vector of each model.

3. The method according to claim 1, characterized in that, The process of generating a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset includes: A predetermined number of pigmented lesion images with gold standard annotations are obtained, and a training dataset and an independent test dataset are constructed based on the pigmented lesion images; the gold standard annotations include benign and malignant lesions. All pigmented lesion images in the training dataset are input into each model in the model pool for training, resulting in multiple models with the same learning objective; High-dimensional feature vectors of models with the same learning objective are extracted, and clustering is performed on each high-dimensional feature vector to generate a feature clustering center group for each model; the feature clustering center group is the feature knowledge base of the corresponding model. The independent validation dataset is input into each of the models that have the same learning objective to generate the expected calibration error for each model.

4. The method according to claim 1, characterized in that, The calculation of dynamic fusion weights for each model for the pigmented lesion image to be evaluated, based on the lesion prediction probability vector and the high-dimensional feature vector of each model, includes: Based on the lesion prediction probability vector of each model, the prediction information entropy of each model is calculated; Based on the predicted information entropy and calibration error parameters of each model, the internal confidence of each model is calculated using the following formula: in, For internal confidence level, Let α and β be the predicted information entropy of the model, and α and β be the preset weight coefficients. These are the calibration error parameters for the model; Based on the high-dimensional feature vectors of each model and the minimum cosine distance of all cluster centers in the feature knowledge base, the external matching weight of each model is calculated using the following formula: in, γ represents the external matching weight, and γ is a preset scaling factor. The minimum cosine distance between all cluster centers in the model feature library; Based on the internal confidence and external matching weights of each model, the dynamic fusion weights of each model for the image of the pigmented lesion to be evaluated are calculated using the following formula: in, For dynamic weight fusion, For internal confidence level, represents the external matching weight, and n represents the number of models.

5. The method according to claim 1, characterized in that, The step of fusing the predicted probability vectors and dynamic fusion weights of each model according to a preset combination rule to generate a comprehensive risk probability and fusion uncertainty includes: Based on the dynamic fusion weights of each model and the malign and benign probabilities in the predicted probability vector, a basic probability allocation function is constructed for each model; the basic probability allocation function includes the quality allocation of the malign hypothesis, the quality allocation of the benign hypothesis, and the quality allocation of the uncertainty domain. in, To allocate quality to malign hypotheses, For dynamic weight fusion, This represents the probability of malignancy. For benign probability, The quality assigned to benign assumptions, The mass assigned to the uncertain region; According to the preset combination rules, the basic probability allocation functions of all models are recursively combined in pairs to calculate the global probability allocation; The mass assigned to the malign hypothesis is extracted from the global probability allocation as the comprehensive risk probability, and the mass assigned to the uncertainty domain is extracted as the fused uncertainty.

6. The method according to claim 5, characterized in that, The step of recursively combining the basic probability allocation functions of all models pairwise according to a preset combination rule to calculate the global probability allocation includes: The basic probability assignment function of each model is used as model evidence, and the fusion probability assignment is generated by recursively combining each model evidence pairwise using the Dempster-Shafer combination rule: in, For fusion probability allocation, , Let K be the conflict coefficient, B and C be the subsets of hypotheses supported by the two models, and A be the hypothesis that we want to verify after fusion. The fusion probability allocation generated after all models have been fused is used as the global probability allocation.

7. The method according to claim 1, characterized in that, The generation of a pigmentary disease risk assessment report based on the comprehensive risk probability and the fusion uncertainty includes: The overall risk probability is compared with a preset risk threshold, and the risk level of pigmentary diseases is determined based on the comparison results; The fusion uncertainty is compared with a preset confidence threshold, and a confidence indicator is determined based on the comparison result; The risk level of the pigmentary disease and the confidence index are integrated into a pigmentary disease risk assessment report.

8. An intelligent risk assessment device for pigmentary disorders, characterized in that, The device includes: The model pool construction module is used to obtain a model pool containing multiple heterogeneous deep learning models, and generate a corresponding feature knowledge base and calibration error parameters for each model in the model pool based on the same training dataset. The model inference module is used to receive the pigmented lesion image to be evaluated, input the pigmented lesion image to be evaluated into each model in the model pool for forward inference, and output the lesion prediction probability vector and high-dimensional feature vector of each model for the pigmented lesion image to be evaluated. The model's dynamic fusion weight generation module is used to calculate the dynamic fusion weight of each model for the pigmented lesion image to be evaluated based on the lesion prediction probability vector and the high-dimensional feature vector output by each model. The module for generating comprehensive risk probability and fusion uncertainty is used to fuse the lesion prediction probability vectors of each model and the dynamic fusion weights according to preset combination rules to generate comprehensive risk probability and fusion uncertainty. The assessment report generation module is used to generate a risk assessment report for pigmentary diseases based on the comprehensive risk probability and the fusion uncertainty.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images