Dual model bias training and confidence fusion thermal image classification method and system

Abstract

The application discloses a kind of dual model bias training and confidence fusion's thermal image classification method and system.The method is inputted after pre-processing thermal image into benign bias model and malignant bias model respectively to obtain output value, and three kinds of index fusion are obtained, such as temperature scaling information entropy, Dirichlet evidence and maximum probability, to obtain comprehensive confidence.Training phase is according to loss learning curve to adaptively adjust bias intensity, and sample level bias weight is calculated by combining confidence and label to train with weighted loss function.Inference phase is according to the output value of two models to calculate prediction consistency, and according to the consistency high and low respectively equal weight fusion, confidence weighted fusion or expert rule decision is used, and classification result is output.The application overcomes the deficiency of static bias strategy and simple fusion method in class imbalance scene, and improves the recognition ability of malignant class and the robustness of fusion decision.

Description

Technical Field

[0001] This invention belongs to the field of intelligent thermal image analysis technology, specifically relating to a thermal image classification method and system that combines dual-model biased training with confidence fusion. Background Technology

[0002] Thermal imaging is a technique that uses infrared thermal imaging to capture images of the body surface temperature distribution. It provides a non-invasive, radiation-free, low-cost, and repeatable way to reflect local tissue metabolic activity and blood flow changes. In breast cancer screening, malignant lesions are often accompanied by increased local metabolism and angiogenesis, manifesting as abnormally high-temperature areas and abnormal vascular-like structures on thermal images. Therefore, thermal imaging can serve as an effective supplement to traditional examination methods such as mammography, ultrasound, and MRI. However, in real clinical settings, thermal imaging data commonly suffers from noise interference, significant differences in acquisition conditions and equipment, and inconsistent brightness and contrast. Furthermore, the distribution of benign and malignant samples is severely unbalanced, with malignant samples typically accounting for only a small proportion of the total sample. These factors collectively limit the ability of intelligent analysis models to identify malignant and other critical minority classes and the reliability of their outputs.

[0003] Currently, the technical solutions for classifying benign and malignant thermal images mainly fall into three categories. The first category is based on handcrafted features and traditional machine learning methods. This typically extracts statistical, texture, or shape features from the region of interest and uses classifiers such as support vector machines (SVMs) and random forests, combined with fixed-cost sensitive strategies, to address class imbalance. While simple to implement, these methods have limited feature representation capabilities and lack robustness to complex lesion patterns and varying imaging conditions. The second category is end-to-end classification schemes based on a single deep neural network. These generally use convolutional neural networks as the backbone network, combined with class-balanced cross-entropy loss functions or focal loss functions to increase the training weights of malignant samples. However, these methods often simply use the probability values ​​output by the softmax function directly as a confidence measure, lacking the ability to model uncertainty based on information entropy and evidence theory. They also fail to utilize the loss curve information during training to dynamically adjust the bias intensity, making it difficult to achieve sample-level, process-aware, and refined bias training. The third category is simple fusion schemes involving multiple models or branches. These schemes integrate the outputs of multiple models through probability averaging, fixed weighting, or voting. While they can improve classification performance to some extent, the fusion strategies are mostly static rules that do not comprehensively consider the predictive consistency between models and the confidence level of each model. They lack an intelligent fusion mechanism that performs hierarchical processing based on predictive consistency and are not stable enough when there is a large divergence in model predictions or when the uncertainty of samples is high.

[0004] In summary, existing technologies have not yet formed an integrated technical framework that takes comprehensive confidence assessment as the core, combines dynamic biased training of dual models with intelligent fusion decision-making, and makes it difficult to simultaneously take into account the recognition performance of key categories such as malignancy, the reliability of model output, and the robustness of fusion decision-making in thermal image classification scenarios with extremely imbalanced categories and high prediction uncertainty. Summary of the Invention

[0005] This invention proposes a thermal image classification method and system based on dual-model biased training and confidence fusion. The technical problems to be solved are: in the scenario of classifying benign and malignant thermal images, the extreme imbalance of categories makes it difficult for single-model and static reweighting methods to balance overall performance with the identification ability of a few key categories such as malignancy; existing methods lack comprehensive uncertainty quantification methods based on temperature scaling, information entropy and Dirichlet evidence theory, and cannot achieve sample-level training bias adjustment; and multi-model fusion often adopts simple averaging or fixed weights, without combining prediction consistency and confidence for adaptive fusion decision-making, resulting in unstable performance on samples with large discrepancies in model predictions.

[0006] To address the aforementioned technical problems, this invention provides a thermal image classification method that combines dual-model biased training with confidence fusion, comprising the following steps: Step S1: Input the preprocessed thermal image into the benign bias model and the malignant bias model respectively to obtain the output values ​​of the benign bias model and the malignant bias model; Step S2: Evaluate the overall confidence of the outputs of the benign bias model and the malign bias model to obtain the overall confidence of each model. The overall confidence is obtained by weighted fusion of entropy confidence, evidence confidence and maximum probability confidence. Step S3: During the training phase, the bias intensity is adaptively adjusted according to the loss learning curve, and the sample-level bias weight is calculated by combining the comprehensive confidence and the true label of the sample. The loss function is weighted with the sample-level bias weight to train the benign bias model and the malignant bias model. Step S4: In the inference stage, the prediction consistency is calculated based on the output values ​​of the benign bias model and the malignant bias model. A fusion strategy is determined based on the prediction consistency, and the output values ​​of the two models and the comprehensive confidence are fused according to the determined fusion strategy to obtain the prediction result used to assist in the classification of benign and malignant.

[0007] Preferably, in step S2, the comprehensive confidence assessment includes the following sub-steps: Step S2.1: Divide the model output logic value logits by the pre-calibrated temperature parameter. The scaled probability distribution is obtained by normalization using the softmax function. Calculate the information entropy of the scaled probability distribution. And calculate the entropy confidence. : ; in, For the scaled probability distribution, the th The probability of each category Number of categories; Step S2.2: Model the logistic values ​​logits using Dirichlet evidence theory, and obtain the non-negative evidence vector through soft addition function transformation. ,in It is a soft-addition function. These are preset prior parameters; The total amount of evidence is obtained by summing the evidence values ​​corresponding to each category in the non-negative evidence vector. ; Calculate the confidence level of the evidence based on the total amount of evidence. : ; Step S2.3: Combine the entropy confidence, the evidence confidence, and the maximum probability confidence. Weighted fusion is used to obtain the overall confidence level. : ; in, The preset weighting coefficients are set such that the sum of the three is 1. This outputs the maximum class probability value in the probability distribution of the model.

[0008] Preferably, in step S3, the method of adaptively adjusting the bias intensity according to the loss learning curve includes: Maintain a fixed-length sliding window to record the training loss values ​​from the most recent few rounds. Fit the trend of the training loss values ​​with the number of training rounds using linear regression to obtain the trend slope. ; Calculate the standard deviation of the training loss values ​​within the sliding window as a volatility indicator. ; bias intensity It is obtained by summing the three parts: fundamental bias strength, trend adjustment, and volatility adjustment. ; in, Based on the strength bias, For trend adjustment volume, This is the volatility adjustment amount; the calculation method for the basic bias strength is as follows: , The initial bias intensity is preset. The preset attenuation coefficient and , This is the current training round; The trend adjustment amount Based on the trend slope The comparison result with the preset trend threshold determines the following: when the trend slope is negative and its absolute value is greater than the preset trend threshold, the trend adjustment amount takes a preset negative value; when the trend slope is positive and greater than the preset trend threshold, the trend adjustment amount takes a preset positive value; when the absolute value of the trend slope is not greater than the preset trend threshold, the trend adjustment amount is zero. The volatility adjustment amount According to the volatility indicator The comparison result with the preset volatility threshold determines that: when the volatility index is greater than the preset volatility threshold, the volatility adjustment amount takes a preset negative value; otherwise, the volatility adjustment amount is zero. The bias intensity The constraint is between a preset upper bound and a preset lower bound.

[0009] Preferably, in step S3, the method for calculating the sample-level bias weight by combining the comprehensive confidence level and the true label of the sample includes: The overall confidence level is divided into three levels: high confidence, medium confidence, and low confidence, based on a preset high confidence threshold and a preset low confidence threshold. Determine whether the sample belongs to the target category of the current model, wherein the target category of the benign bias model is benign, and the target category of the malignant bias model is malignant; Based on the combination of three confidence levels and target class attribution, a biased weight is assigned to each sample: for target class samples, the weight of high-confidence samples is... The weight of the medium confidence sample is The weight of low-confidence samples is For samples not belonging to the target class, the weight of high-confidence samples is... The weight of the medium confidence sample is The weight of low-confidence samples is ; in, These are the preset weight coefficients for the target category at high confidence, medium confidence, and low confidence levels, respectively, and they satisfy the following conditions: , These are the preset weight coefficients for non-target categories at high, medium, and low confidence levels, respectively, and they satisfy the following conditions: .

[0010] Preferably, before training in step S3, image segmentation and spatial coupling feature extraction steps are performed on the thermal image; the extracted spatial coupling features are weighted and combined to form a malignant pattern score. The malignant pattern score is used in the training process of step S3 to determine the routing direction of the sample gradient when the predictions of the benign bias model and the malignant bias model are inconsistent. The image segmentation and spatial coupling feature extraction method includes: Using the blood vessel segmentation model, the high-temperature area segmentation model, and the effective breast region segmentation model, blood vessel masks, high-temperature area masks, and effective breast region masks were extracted from the thermal images, respectively. Using the region defined by the effective breast region mask as the region of interest, calculate the spatial coupling features between the blood vessel mask and the high-temperature region mask; The spatial coupling features include: the proportion of the intersection of blood vessel pixels and high-temperature region pixels within the region of interest to the total number of blood vessel pixels, the proportion of blood vessel pixels in the high-temperature region pixels, the normalized distance from the centroid of the blood vessel mask to the centroid of the high-temperature region mask, and the normalized temperature difference between the average temperature of the high-temperature region and the average temperature of the region of interest.

[0011] Preferably, the malignant pattern scoring We obtain the following by weighted linear combination of spatial coupling features: ; in, The percentage of the total number of blood vessel pixels relative to the intersection of blood vessel pixels and high-temperature region pixels within the region of interest. This represents the proportion of pixels with accompanying blood vessels in the high-temperature region. The normalized distance from the centroid of the blood vessel mask to the centroid of the high-temperature region mask is... , Let be the Euclidean distance between the centroid of the blood vessel mask and the centroid of the high-temperature region mask. The difference between the average temperature of the high-temperature region and the average temperature of the region of interest is normalized to the hyperbolic tangent function. Temperature difference value after the interval, The malicious mode score is calculated with preset weighting coefficients and the sum of the four factors being 1. The value range is [0,1].

[0012] Preferably, step S3 further includes a Bayesian confidence consistency regularization and dynamic routing sub-step: In each training batch, the output values ​​and combined confidence scores of the benign bias model and the mafia bias model are acquired synchronously; the prediction consistency during the training period is calculated. ,in and These are the output values ​​of the benign bias model and the malignant bias model, respectively. Constructing fusion coefficients ,in and These are the combined confidence levels of the benign bias model and the malign bias model, respectively. According to the fusion coefficient Calculate the fusion target distribution Using the fusion target distribution as the regularization objective, KL divergence regularization is applied to both the benign bias model and the malignant bias model. This is to bring the predicted probability distributions of the two models closer to the fusion target distribution, wherein The predicted probability distribution of the currently regularized model; the regularization weights of the KL divergence regularization. The calculation method is as follows ,in This is the preset fusion weight scaling factor; Meanwhile, based on the training period prediction consistency and the malignant pattern score... Dynamic routing is performed as follows: When the training-period prediction consistency is greater than the first training-period consistency threshold, the sample gradient flows to both the benign biased model and the malignant biased model; when the training-period prediction consistency is not greater than the first training-period consistency threshold but greater than the second training-period consistency threshold, the sample gradient flows to both the benign biased model and the malignant biased model with reduced weights; when the training-period prediction consistency is not greater than the second training-period consistency threshold, the malignant pattern score is applied. Determine the routing direction: If the malicious mode score is greater than the preset malicious routing threshold, the sample gradient is routed to the malicious bias model; if the malicious mode score is less than the preset benign routing threshold, the sample gradient is routed to the benign bias model; if the malicious mode score is not less than the preset benign routing threshold and not greater than the preset malicious routing threshold, the sample gradient is routed to the model with higher overall confidence.

[0013] Preferably, in step S4, the prediction consistency The calculation method is as follows ,in and These are the output values ​​of the benign bias model and the malignant bias model, respectively. Determining a fusion strategy based on the prediction consistency includes: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The inference period is compared with a preset first consistency threshold and a second consistency threshold, wherein the first consistency threshold is greater than the second consistency threshold. When the prediction consistency When the value exceeds the first consistency threshold during the inference period, the output values ​​of the two models are taken as an equal-weighted average to obtain the fused output value. ; When the prediction consistency When the output value is not greater than the first consistency threshold during the inference period but greater than the second consistency threshold during the inference period, the output values ​​of the two models are weighted and fused using the combined confidence scores of the two models as weights to obtain the fused output value. : ,in and These are the combined confidence levels of the benign bias model and the malign bias model, respectively. When the prediction consistency If the sample value is not greater than the second consistency threshold of the inference period, the sample will be submitted to the expert rule for decision-making.

[0014] Preferably, the expert rule decisions are made and executed sequentially in descending order of priority: Rule 1: If the overall confidence of either the benign bias model or the malign bias model is greater than the preset rule 1 confidence threshold, then the output value of the model with the higher overall confidence is used as the fusion result. Rule 2: If Rule 1 is not satisfied, and the absolute value of the difference between the overall confidence scores of the two models is greater than the preset confidence score difference threshold, then the output value of the model with the higher overall confidence score shall be used as the fusion result. Rule 3: If the output value of any model is greater than the preset upper threshold for extreme prediction or less than the preset lower threshold for extreme prediction, the output value of the model whose output value deviates more from 0.5 by an absolute value shall be used as the fusion result. Rule 4: When none of the above rules 1 to 3 are satisfied, the arithmetic mean of the output values ​​of the two models shall be taken as the fusion result.

[0015] This invention also provides a thermal image classification system with dual-model biased training and confidence fusion, comprising: The dual-model feature extraction and classification module is configured to input the preprocessed thermal image into a benign bias model and a malignant bias model respectively, and obtain the output values ​​of the benign bias model and the malignant bias model. The confidence assessment module is configured to perform a comprehensive confidence assessment on the outputs of the benign bias model and the malign bias model respectively, and obtain the comprehensive confidence of each model. The comprehensive confidence is obtained by weighted fusion of entropy confidence, evidence confidence and maximum probability confidence. The dynamic bias training module is configured to adaptively adjust the bias intensity according to the loss learning curve during the training phase, and calculate the sample-level bias weight by combining the comprehensive confidence and the true label of the sample. The loss function is weighted by the sample-level bias weight to train the benign bias model and the malignant bias model. The intelligent fusion decision module is configured to calculate the prediction consistency based on the output values ​​of the benign bias model and the malign bias model during the inference phase, determine the fusion strategy based on the prediction consistency, and fuse the output values ​​of the two models and the comprehensive confidence level according to the determined fusion strategy to obtain a prediction result for assisting in the classification of benign and malignant.

[0016] The beneficial effects of the present invention include at least the following: First, this invention constructs a dual-model architecture of benign bias model and malignant bias model, which performs specialized learning on benign and malignant samples respectively. By combining the learning curve-driven adaptive scheduling of bias intensity and the sample-level weighting mechanism of comprehensive confidence level, dynamic bias training of minority class, high-risk sample and difficult sample is realized, overcoming the problem of fixed bias strategy and difficulty in adaptive adjustment with the training process and sample difficulty in the existing technology.

[0017] Second, this invention employs a weighted fusion of three complementary indicators: temperature-scaled prediction entropy, Dirichlet evidence confidence, and maximum class probability. This achieves sample-level uncertainty quantification and provides a unified and interpretable confidence measure for biased weight allocation during training and fusion decision-making during the inference stage. Compared with the single measurement method that relies solely on softmax probability in existing technologies, this invention is more comprehensive and stable.

[0018] Third, in the inference stage, this invention divides the samples into different cases based on the consistency of prediction and adopts differentiated fusion strategies. When the consistency is high, the samples are fused with equal weights; when the consistency is medium, the samples are fused with weights based on confidence; and when the consistency is low, the expert rule decision is initiated. Compared with the simple averaging or fixed weighting methods in the prior art, this invention is more robust on samples with large discrepancies in model prediction and high uncertainty. Moreover, the fusion process is traceable and interpretable.

[0019] Fourth, this invention introduces the spatial coupling feature of blood vessel-high temperature zone, encodes the medical prior knowledge in the thermal image into a computable structured score, and helps determine the routing direction of sample gradient when the two models are inconsistent during the training phase, thereby improving the accuracy and interpretability of low consistency sample processing. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a performance comparison diagram of an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] like Figure 1 As shown, this invention provides a thermal image classification method based on dual-model biased training and confidence fusion, the overall process of which includes four core steps. Step S1 involves inputting the preprocessed thermal image into a benign bias model and a malignant bias model respectively to obtain their respective output values; Step S2 performs a comprehensive confidence assessment on the outputs of the two models; Step S3, during the training phase, adaptively adjusts the bias intensity according to the loss learning curve, and calculates the sample-level bias weights by combining the comprehensive confidence and the true labels of the samples, and performs weighted training on the loss function; Step S4, during the inference phase, calculates the prediction consistency based on the output values ​​of the two models, determines the fusion strategy based on the prediction consistency, completes the fusion, and outputs the final benign / malignant classification prediction result. The training phase may also optionally include image segmentation and spatial coupling feature extraction steps, as well as Bayesian confidence consistency regularization and dynamic routing sub-steps, to further improve the model's collaborative ability and the stability of processing low-consistency samples. Each step is described in detail below.

[0023] Step S1: Input the preprocessed thermal images into the benign bias model and the malignant bias model respectively to obtain the output values ​​of the benign bias model and the malignant bias model.

[0024] This step receives a pre-processed thermal image of the breast as input. The pre-processing includes normalizing the image to a preset size (480×640 pixels in this embodiment) and number of channels, and standardizing brightness and contrast to eliminate differences caused by different acquisition devices and environmental conditions.

[0025] The preprocessed thermal images are fed into two classification models: a benign bias model and a malignant bias model. Both models employ the same convolutional neural network backbone architecture (ResNet-50 is used as the feature extraction backbone in this embodiment, but it can also be replaced with other backbone networks such as DenseNet-121 or EfficientNet-B0). Both models include convolutional layers, non-linear activation functions, pooling layers, and fully connected layers. The output uses a binary classification head to output the logits and corresponding probabilities for benign and malignant classes. The two models can share a single encoder for extracting general visual features, or they can each perform the entire feature extraction process independently, depending on computational resources and data scale. The network structures of the two models can be the same or different, but the output category dimension must be consistent, and the output values ​​will serve as key inputs for subsequent confidence assessment and fusion decisions.

[0026] The difference between benign and malignant biased models lies not in their network structure, but in the dynamic bias weight mechanism in step S3 during the training phase, which applies differentiated training weights to samples of different categories. Benign biased models assign higher bias weights to benign samples during training, making their feature representation of benign samples more comprehensive. Malignant biased models, on the other hand, assign higher bias weights to malignant samples, giving them stronger specialization capabilities in terms of sensitivity and recall for malignant lesions. This differentiated bias training allows the two models to produce different outputs on the same sample, and this predictive difference provides valuable information for subsequent confidence assessment and fusion decisions.

[0027] Step S2: Evaluate the overall confidence of the outputs of the benign bias model and the malign bias model to obtain the overall confidence of each model.

[0028] This step performs a comprehensive confidence assessment of the outputs of both the benign and malignant biased models, providing a unified measure of uncertainty for bias weight allocation in the subsequent training phase and fusion decision-making in the inference phase. The comprehensive confidence score is obtained by weighted fusion of three complementary confidence indices: entropy confidence based on temperature-scaled information entropy, evidence confidence based on Dirichlet evidence theory, and maximum probability confidence. These three indices measure the model's confidence in predicting the current sample from three different perspectives: the concentration of the probability distribution, the amount of evidence accumulated by the model, and the maximum class prediction strength.

[0029] S2.1: Entropy confidence based on temperature scaling and information entropy.

[0030] First, divide the logic value output by the model by the preset temperature parameter. The scaled probability distribution is obtained by normalization using the softmax function. .when When the value is greater than 1, the scaled probability distribution is flatter, causing the model to no longer give overly extreme probability values ​​for uncertain samples; when When the value is less than 1, the distribution becomes more acute. In this embodiment, the temperature parameter... The model parameters are frozen and then optimized by minimizing the negative log-likelihood loss on the validation set. The typical value range is 0.5 to 2.0.

[0031] Calculate the information entropy for the scaled probability distribution: ; in, For the scaled probability distribution, the th The probability of each class is given by log, where log is the natural logarithm. Information entropy measures the degree of uncertainty in the probability distribution: the entropy is at its maximum when the model's predicted probabilities for all classes are close to a uniform distribution, indicating that the model is most uncertain about that sample; when the model concentrates most of the probability on a particular class, the entropy is close to zero, indicating that the model is very certain.

[0032] Based on this, the information entropy is normalized to the [0,1] interval, and the entropy confidence level is defined: ; in, Number of categories (in binary classification) =2), This represents the maximum possible entropy value. The entropy confidence score ranges from [0,1], with higher values ​​indicating a more concentrated distribution of model predictions and lower uncertainty.

[0033] S2.2: Confidence of evidence based on Dirichlet's evidence theory.

[0034] Dirichlet evidence theory interprets the output of a neural network as the accumulated amount of evidence for each class, and parameterizes the distribution of class probabilities using a Dirichlet distribution, thereby simultaneously obtaining prediction results and uncertainty estimates in a single forward propagation. This step uses this theory to model the logistic values ​​of the model output, obtaining a non-negative evidence vector through a soft-addition function transformation: ; Where softplus is the soft addition function, defined as softplus(x)=log(1+exp(x)), which guarantees that the output is non-negative; logits is the logical value of the model output; The preset prior parameter is set to 1.0 in this embodiment, corresponding to a uniform prior. Each element of the evidence vector corresponds to the evidence amount for a category. When the model has accumulated more evidence for a certain category, the evidence value for that category is higher.

[0035] The total amount of evidence is obtained by summing the evidence values ​​corresponding to each category in the non-negative evidence vector. Based on this, the confidence level of the evidence is calculated: ; When each category of evidence is relatively small, the total amount of evidence is close to... (Due to the existence of prior parameters), the confidence level of the evidence is close to zero, indicating high uncertainty; as the model accumulates more evidence from the samples, the total amount of evidence increases, and the confidence level of the evidence tends to 1.

[0036] S2.3: Comprehensive confidence level fusion.

[0037] The three indicators—entropy confidence, evidence confidence, and maximum probability confidence—are weighted and fused together to obtain the comprehensive confidence score: ; in, Assuming the preset weighting coefficients and the sum of the three is 1, in this embodiment, we take... =0.4、 =0.3、 =0.3. The maximum class probability value in the model's output probability distribution is directly taken as the maximum value of the scaled probability distribution. These three metrics provide complementary confidence information from three dimensions: probability distribution shape, amount of accumulated evidence, and maximum predicted probability.

[0038] Step S3: During the training phase, the bias intensity is adaptively adjusted according to the loss learning curve, and the sample-level bias weight is calculated by combining the comprehensive confidence and the true label of the sample. The loss function is weighted with the sample-level bias weight to train benign bias models and malignant bias models.

[0039] This step is performed during the training phase and includes three stages: bias intensity scheduling, sample-level bias weight calculation, and weighted loss training. Optional steps may also include image segmentation and spatial coupling feature extraction, Bayesian confidence consistency regularization, and dynamic routing sub-steps.

[0040] S3.1: Adaptive scheduling based on bias intensity according to learning curve.

[0041] The bias intensity controls the overall strength of the bias weights applied to the target class samples in the current training epoch. The core design of this invention lies in correlating the bias intensity with the trend of the loss curve during training, enabling it to adaptively adjust according to the model's current learning state. Specifically, the system maintains a fixed-length sliding window (10 epochs in this embodiment) and records the training loss values ​​from the most recent epochs. The trend slope (slope) is obtained by fitting the trend of the training loss values ​​within the sliding window with the training epochs using linear regression; simultaneously, the standard deviation of the training loss values ​​within the sliding window is calculated as a volatility indicator. Bias intensity It is obtained by summing the three parts: fundamental bias strength, trend adjustment, and volatility adjustment. ; The fundamental bias strength employs an exponential decay strategy: ,in The initial bias strength is preset (1.5 in this embodiment). The decay coefficient is a preset value (0.95 in this example), and epoch is the current training round. This design results in a higher bias in the early stages of training, helping the model quickly learn the discriminative features of the target category, which gradually decreases to a stable level as training progresses.

[0042] Trend Adjustment Volume Based on the comparison between the trend slope and the preset trend threshold (0.01 in this embodiment), the following is determined: when the trend slope is negative and its absolute value is greater than the threshold (i.e., the loss decreases rapidly), the preset negative value is taken. 0.1 is used to reduce the bias intensity and avoid overfitting due to excessive bias; when the trend slope is positive and greater than this threshold (i.e., the loss increases or convergence is difficult), a preset positive value +0.2 is used to enhance the bias and strengthen the model's learning of the target class; when the absolute value of the trend slope is not greater than this threshold (i.e., the loss is stable), a zero value is used without adjustment.

[0043] Volatility Adjustment According to volatility indicators The comparison result with the preset volatility threshold (0.1 in this embodiment) determines the value: when the volatility index is greater than the threshold, a preset negative value is taken. Set the bias strength to 0.1 for stable training, otherwise set it to zero. Ultimately, constrain the bias strength between a preset lower bound of 0.5 and a preset upper bound of 2.0 to ensure the bias strength remains within a reasonable range.

[0044] The three-layered adjustment mechanism described above corresponds to the long-term decay strategy, the medium-term trend response, and the short-term stability protection, respectively, forming a multi-layered, multi-time-scale adaptive bias control framework. The training losses of both models are incorporated into the loss history to fully reflect the overall learning state of the system.

[0045] S3.2: Sample-level biased weight calculation based on confidence level.

[0046] To obtain the bias strength in the current round After calculating the overall confidence level for each sample, this step assigns differentiated training weights to each sample based on the combination of the overall confidence level and the sample's true label. First, the overall confidence level is divided into three levels: high confidence, medium confidence, and low confidence, according to a preset high confidence threshold (0.8 in this example) and a preset low confidence threshold (0.5 in this example). Simultaneously, it is determined whether the sample belongs to the target category of the current model: the target category for a benign-biased model is benign, and the target category for a malignant-biased model is malignant.

[0047] Based on the combination of three confidence levels and the target class classification, a biased weight is assigned to each sample. For target class samples: the weight of high-confidence samples is... (In this embodiment) =1.5), the weight of the medium confidence sample is ( =1.2), the weight of low confidence samples is ( =0.8). For non-target class samples: the weight of high-confidence samples is ( =0.5), the weight of the medium confidence sample is ( =0.7), the weight of low confidence samples is ( =1.0). These preset weighting coefficients satisfy... and The constraint relationship is designed with the following logic: apply the maximum weight to high-confidence samples of the target class to strengthen the existing advantages, and apply the minimum weight to high-confidence samples of non-target classes to avoid over-focusing on easily classified non-target samples, thereby achieving refined dynamic bias training in the two-dimensional space of class × confidence.

[0048] S3.3: Weighted Loss Construction and Model Update

[0049] After obtaining the sample-level bias weights, a sample-by-sample cross-entropy loss function is applied to each model. The loss value of each sample is multiplied element-wise by its corresponding bias weight, and the batch average is calculated to form a weighted loss function. This weighted loss function retains the convergence property of the standard cross-entropy loss while achieving adaptive weighting of key categories and high-value samples through a confidence- and learning curve-driven weight design. The weighted losses of the two models are calculated separately and backpropagated independently to update parameters. Simultaneously, their respective loss values ​​are recorded in the loss history of the bias intensity scheduling module, providing a basis for the next round of bias intensity calculation.

[0050] S3.4: Image segmentation and spatial coupling feature extraction.

[0051] After obtaining the model's output value in step S1 and before training in step S3, image segmentation and spatial coupling feature extraction steps can optionally be performed on the preprocessed thermal image. This step extracts three types of masks from the input thermal image using three independent segmentation models: a blood vessel segmentation model is used to detect vascular-like structures such as thin strips and branching networks, outputting a blood vessel mask; a high-temperature region segmentation model is used to identify abnormally high-temperature regions, outputting a high-temperature region mask; and a breast effective region segmentation model is used to identify the effective breast region and remove background and artifacts, outputting a breast effective region mask. All three segmentation models adopt the U-Net architecture and output a binary mask of shape [B, H, W], where 1 represents the corresponding region and 0 represents the background.

[0052] The region defined by the effective breast region mask is taken as the region of interest. Within this region, four spatial coupling features between the blood vessel mask and the high-temperature region mask are calculated: the first feature is the proportion of the intersection of blood vessel pixels and high-temperature region pixels within the region of interest to the total number of blood vessel pixels. The higher this ratio, the closer the spatial overlap between the blood vessels and the high-temperature zone; the second item is the proportion of pixels in the high-temperature zone that are accompanied by blood vessels. The first term reflects the degree of co-occurrence between the high-temperature zone and blood vessels; the third term is the normalized distance from the centroid of the blood vessel mask to the centroid of the high-temperature zone mask. ,in The first term represents the Euclidean distance between the two centroids. After normalization, a smaller distance indicates a larger value, reflecting the spatial proximity between the two. The fourth term is the difference between the average temperature of the high-temperature region and the average temperature of the region of interest, normalized to ( ) using the hyperbolic tangent function. Temperature difference value after interval 1,1) The greater the temperature difference, the more prominent the high-temperature zone.

[0053] The above four spatial coupling features are used to calculate the malignancy score through a weighted linear combination: ; in The preset weighting coefficients are set such that the sum of the four is 1 (in this embodiment, the weighting coefficients are set to 1). =0.3、 =0.3、 =0.3、 =0.1), Malignant Pattern Score The value range is [0,1], with higher scores indicating a greater tendency towards malignancy. The weights of the four factors are set based on clinical experience: the proportion of blood vessels passing through the high-temperature zone, the proportion of blood vessels accompanying the high-temperature zone, and the spatial distance between the two factors all contribute equally to malignancy detection, while the temperature difference factor has a lower weight as an auxiliary reference factor.

[0054] S3.5: Bayesian confidence consistency regularization and dynamic routing.

[0055] This sub-step builds upon S3.4 by introducing a dual-model co-optimization mechanism during the training phase, enabling the fusion strategy to take effect during training. In each training batch, the output values ​​and combined confidence scores of both models are acquired simultaneously, and the training-phase prediction consistency is calculated. ; in and These are the output values ​​of the benign bias model and the malign bias model, respectively. The training-period prediction consistency value ranges from [0,1], with higher values ​​indicating closer similarities between the two models.

[0056] Constructing fusion coefficients ,in and These represent the combined confidence scores of the two models. The distribution of the fusion target is calculated based on the fusion coefficient. This distribution integrates the prediction and confidence information from both models, serving as the regularization objective. KL divergence regularization is applied to both models. This is to bring the predicted probability distributions of the two models closer to the fusion target distribution, where This represents the predicted probability distribution of the currently regularized model. Regularization weights. ,in The preset fusion weight scaling factor is 0.5 in this embodiment. The regularization weight increases with the increase of prediction consistency and overall confidence during training: the regularization constraint is strongest when the two models predict highly consistent and are both confident; when the prediction discrepancy is large or the confidence is low, the regularization weight decreases to avoid excessive constraint on noisy samples.

[0057] Simultaneously, dynamic routing is performed based on training-period prediction consistency and malignant mode scoring to determine the gradient flow direction for each sample. When the training-period prediction consistency is greater than the first consistency threshold (0.7 in this embodiment), the sample gradient flows to both models simultaneously, and both models learn fully from this sample. When the training-period prediction consistency is not greater than the first consistency threshold but greater than the second consistency threshold (0.3 in this embodiment), the sample gradient still flows to both models, but the weights are reduced to minimize the perturbation of model parameters by divergent samples. When the training-period prediction consistency is not greater than the second consistency threshold, it indicates that the two models have significant disagreements on this sample. In this case, a malignant mode scoring is introduced to assist in decision-making: if... If the malignant pattern score is greater than the preset malignant routing threshold (0.6 in this embodiment), the sample exhibits a malignant pattern of "high blood vessel density and high temperature area closely attached to abnormal blood vessels," and the sample gradient is routed to the malignant bias model. If the malignant pattern score is less than the preset benign routing threshold (0.4 in this embodiment), the sample exhibits a benign pattern of "scattered small high temperature areas and regular blood vessel structure," and the sample gradient is routed to the benign bias model. If the malignant pattern score is not less than the benign routing threshold and not greater than the malignant routing threshold, the image features are insufficient to make a judgment, and the sample gradient is routed to the model with higher overall confidence.

[0058] Step S4: In the inference stage, the prediction consistency is calculated based on the output values ​​of the benign bias model and the malign bias model. The fusion strategy is determined based on the prediction consistency, and the output values ​​and comprehensive confidence of the two models are fused according to the determined fusion strategy to obtain the benign and malignant classification prediction results.

[0059] This step is performed during the inference phase. After completing steps S1 and S2 to obtain the output values ​​and combined confidence of the two models, the prediction consistency is first calculated as the basis for selecting the fusion strategy.

[0060] S4.1: Prediction Consistency Calculation.

[0061] Prediction Consistency The calculation method is as follows: ; in and These are the output values ​​of the benign bias model and the malign bias model, respectively. The prediction consistency value ranges from [0,1]. The consistency is 1 when the output values ​​of the two models are exactly the same, and the consistency is 0 when the two models give extreme predictions of 0 and 1, respectively.

[0062] S4.2: Hierarchical fusion strategy.

[0063] Will predict consistency The results are compared with preset first and second consistency thresholds during the inference period (in this embodiment, the first consistency threshold is set to 0.7 and the second consistency threshold is set to 0.3), and three different fusion strategies are determined based on the comparison results: When predictions are consistent When the value exceeds the first consistency threshold during the inference period, the opinions of the two models are considered to be highly consistent. An equally weighted average of the output values ​​of the two models is then taken to obtain the fused output value. This approach is simple and efficient, and, provided that the two models agree, it is sufficient to obtain robust fusion results.

[0064] When predictions are consistent When the consensus value is not greater than the first consistency threshold during the inference period but greater than the second consistency threshold during the inference period, the two models have some disagreement but it has not yet reached a serious level. In this case, the combined confidence scores of the two models are used as weights for weighted fusion to obtain the fused output value: ; in and These represent the combined confidence scores of the benign biased model and the malignant biased model, respectively. This weighting method ensures that the model with the higher combined confidence score contributes more to the fusion result.

[0065] When the predictive consistency is not greater than the second consistency threshold during the inference period, there is a serious discrepancy between the two models, and simple weighted fusion may not provide reliable results. In this case, the samples are handed over to expert rule decision-making. The expert rules are judged and executed in descending order of priority as follows: Rule 1: If the overall confidence of either model is greater than the preset confidence threshold of Rule 1 (0.9 in this embodiment), the output value of the model with the higher overall confidence is used as the fusion result, because the model shows a very high degree of confidence in the current sample and its prediction reliability is higher; Rule 2: If the absolute value of the difference between the overall confidence of the two models is greater than the preset confidence difference threshold (0.3 in this embodiment), the output value of the model with the higher overall confidence is used as the fusion result; Rule 3: If the output value of either model is greater than the preset upper threshold for extreme prediction (0.8 in this embodiment) or less than the preset lower threshold for extreme prediction (0.2 in this embodiment), the output value of the model with the larger absolute value of the output value deviating from 0.5 is used as the fusion result, because extreme prediction usually means that the model has detected a strong class-specific signal; Rule 4: When none of the above three rules are satisfied, the arithmetic mean of the output values ​​of the two models is taken as the conservative fusion result.

[0066] After obtaining the fused output value, the system converts it into a benign or malignant classification label based on a preset classification threshold (0.5 in this embodiment) as the final benign / malignant classification prediction result. Simultaneously, the system retains the fused output value, the output values ​​of each of the two models, the overall confidence level, prediction consistency, the type of fusion strategy used, and the malignancy pattern score, among other intermediate indicators, for clinicians' reference. Samples with low confidence levels or significant model disagreement can be marked in the interface as requiring focused review or as high-uncertainty samples, prompting doctors to conduct manual review.

[0067] To verify the effectiveness of the method of this invention, experiments were conducted on a dataset containing 1200 thermal breast images. This dataset contained 960 benign samples and 240 malignant samples, a benign-to-malignant ratio of 4:1. The dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The training run consisted of 100 epochs with a batch size of 32. The Adam optimizer was used, with an initial learning rate of 0.001.

[0068] The method of this invention was compared with the following baseline methods: (1) a single ResNet-50 model with standard cross-entropy loss; (2) a single ResNet-50 model with focus loss (γ=2); (3) a single ResNet-50 model with fixed class weights (the weight of the malignant class is 4); and (4) a simple average fusion of two ResNet-50 models. Evaluation metrics included overall accuracy, malignant class recall, malignant class F1 score, and AUC. Experimental results are as follows: Figure 2 As shown, the method of this invention achieves an overall accuracy of 93.2%, a recall rate of 91.7% for the malicious class, an F1 score of 89.4% for the malicious class, and an AUC of 0.962. It outperforms the baseline method mentioned above in all metrics, especially in the recall rate of the malicious class, which is 5.6 percentage points higher than the best baseline method (focus loss scheme, malicious recall rate of 86.1%). This verifies the effectiveness of the dynamic biased training mechanism and intelligent fusion strategy in class imbalance scenarios.

[0069] All threshold parameters involved in this invention are determined through system calibration on a validation set. The temperature parameter T is optimized on the validation set by minimizing the negative log-likelihood loss. The trend threshold of 0.01 and the fluctuation threshold of 0.1, which are related to the bias intensity, are determined based on statistical analysis of the loss curve during training: in multiple pre-experiments, the inter-cycle variation of the training loss was observed to be typically between 0.005 and 0.05, and therefore 0.01 was set as the critical value to distinguish between "significant trend" and "stable trend"; the fluctuation threshold of 0.1 corresponds to the typical dividing point of the loss standard deviation, and exceeding this value indicates that the training has experienced large oscillations.

[0070] The confidence level thresholds of 0.8 and 0.5 were determined by analyzing the distribution of the overall confidence level on the validation set: using the 75th and 40th percentiles of the overall confidence level as initial reference values, a grid search was used to find the threshold combination that maximizes the maladaptive F1 score within the range of [0.7, 0.9] × [0.4, 0.6]. The inference consistency thresholds of 0.7 and 0.3 were determined by observing the relationship between prediction consistency and classification error rate on the validation set: samples with consistency higher than 0.7 had a classification error rate of less than 3%, samples with consistency between 0.3 and 0.7 had a classification error rate of approximately 12%, and samples with consistency lower than 0.3 had a classification error rate exceeding 25%. Therefore, 0.7 and 0.3 were set as the cutoff points for tiered fusion.

[0071] This invention also provides a thermal image classification system with dual-model biased training and confidence fusion, comprising the following functional modules: The dual-model feature extraction and classification module is configured to input the preprocessed thermal image into the benign bias model and the malignant bias model respectively, and obtain the output values ​​of each model. The specific implementation of this module is consistent with the description in step S1 above.

[0072] The confidence assessment module is configured to perform a comprehensive confidence assessment on the outputs of the two models to obtain the overall confidence of each model. The specific implementation of this module is consistent with the description in step S2 above.

[0073] The dynamic biased training module is configured to adaptively adjust the bias strength based on the loss learning curve during the training phase, and calculate sample-level bias weights by combining the comprehensive confidence score and the true labels of the samples. These sample-level bias weights are then used to weight the loss function for training the two models. The specific implementation of this module is consistent with the description in step S3 above.

[0074] The intelligent fusion decision module is configured to calculate the prediction consistency based on the output values ​​of the two models during the inference phase, determine the fusion strategy based on the prediction consistency, and complete the fusion to output the final benign / malignant classification prediction result. The specific implementation of this module is consistent with the description in step S4 above.

[0075] The modules described above can be deployed on servers or embedded computing platforms equipped with GPU computing cards, implemented as software modules. Alternatively, the inference process can be deployed on edge computing devices to meet real-time requirements. The system provides standardized input / output interfaces, receiving preprocessed thermal images and outputting classification prediction results along with relevant confidence scores and intermediate analysis information.

[0076] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0077] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A thermal image classification method based on dual-model biased training and confidence fusion, characterized in that, Includes the following steps: Step S1: Input the preprocessed thermal image into the benign bias model and the malignant bias model respectively to obtain the output values ​​of the benign bias model and the malignant bias model; Step S2: Evaluate the overall confidence of the outputs of the benign bias model and the malign bias model to obtain the overall confidence of each model. The overall confidence is obtained by weighted fusion of entropy confidence, evidence confidence and maximum probability confidence. Step S3: During the training phase, the bias intensity is adaptively adjusted according to the loss learning curve, and the sample-level bias weight is calculated by combining the comprehensive confidence and the true label of the sample. The loss function is weighted with the sample-level bias weight to train the benign bias model and the malignant bias model. Step S4: In the inference stage, the prediction consistency is calculated based on the output values ​​of the benign bias model and the malignant bias model. A fusion strategy is determined based on the prediction consistency, and the output values ​​of the two models and the comprehensive confidence are fused according to the determined fusion strategy to obtain the prediction result used to assist in the classification of benign and malignant. In step S3, the methods for adaptively adjusting the bias strength based on the loss learning curve include: We maintain a fixed-length sliding window to record the training loss values ​​from the most recent few rounds, and use linear regression to fit the trend of the training loss values ​​as a function of training rounds to obtain the trend slope. ; Calculate the standard deviation of the training loss values ​​within the sliding window as a volatility indicator. ; bias intensity It is obtained by summing the three parts: fundamental bias strength, trend adjustment, and volatility adjustment. ； in, Based on the strength bias, For trend adjustment volume, This is the volatility adjustment amount; the calculation method for the basic bias strength is as follows: , The initial bias intensity is preset. The preset attenuation coefficient and , This is the current training round; The trend adjustment amount Based on the trend slope The comparison result with the preset trend threshold determines the following: when the trend slope is negative and its absolute value is greater than the preset trend threshold, the trend adjustment amount takes a preset negative value; when the trend slope is positive and greater than the preset trend threshold, the trend adjustment amount takes a preset positive value; when the absolute value of the trend slope is not greater than the preset trend threshold, the trend adjustment amount is zero. The volatility adjustment amount According to the volatility indicator The comparison result with the preset volatility threshold determines that: when the volatility index is greater than the preset volatility threshold, the volatility adjustment amount takes a preset negative value; otherwise, the volatility adjustment amount is zero. The bias intensity The constraint is between a preset upper bound and a preset lower bound.

2. The method according to claim 1, characterized in that, In step S2, the comprehensive confidence assessment includes the following sub-steps: Step S2.1: Divide the model output logic value logits by the pre-calibrated temperature parameter. The scaled probability distribution is obtained by normalization using the softmax function. Calculate the information entropy of the scaled probability distribution. And calculate the entropy confidence. : ； in, For the scaled probability distribution, the th The probability of each category Number of categories; Step S2.2: Model the logistic values ​​logits using Dirichlet evidence theory, and obtain the non-negative evidence vector through soft addition function transformation. ,in It is a soft-addition function. These are preset prior parameters; The total amount of evidence is obtained by summing the evidence values ​​corresponding to each category in the non-negative evidence vector. ; Calculate the confidence level of the evidence based on the total amount of evidence. : ； Step S2.3: Combine the entropy confidence, the evidence confidence, and the maximum probability confidence. Weighted fusion is used to obtain the overall confidence level. : ； in, The preset weighting coefficients are set such that the sum of the three is 1. This outputs the maximum class probability value in the probability distribution of the model.

3. The method according to claim 2, characterized in that, In step S3, the method for calculating the sample-level bias weight by combining the comprehensive confidence score and the true label of the sample includes: The overall confidence level is divided into three levels: high confidence, medium confidence, and low confidence, based on a preset high confidence threshold and a preset low confidence threshold. Determine whether the sample belongs to the target category of the current model, wherein the target category of the benign bias model is benign, and the target category of the malignant bias model is malignant; Based on the combination of three confidence levels and target class attribution, a biased weight is assigned to each sample: for target class samples, the weight of high-confidence samples is... The weight of the medium confidence sample is The weight of low-confidence samples is For samples not belonging to the target class, the weight of high-confidence samples is... The weight of the medium confidence sample is The weight of low-confidence samples is ; in, These are the preset weight coefficients for the target category at high confidence, medium confidence, and low confidence levels, respectively, and they satisfy the following conditions: , These are the preset weight coefficients for non-target categories at high, medium, and low confidence levels, respectively, and they satisfy the following conditions: .

4. The method according to claim 1, characterized in that, Before training in step S3, image segmentation and spatial coupling feature extraction steps are performed on the thermal image; the extracted spatial coupling features are weighted and combined to form a malignant pattern score. The malignant pattern score is used in the training process of step S3 to determine the routing direction of the sample gradient when the predictions of the benign bias model and the malignant bias model are inconsistent. The image segmentation and spatial coupling feature extraction method includes: Using the blood vessel segmentation model, the high-temperature area segmentation model, and the effective breast region segmentation model, blood vessel masks, high-temperature area masks, and effective breast region masks were extracted from the thermal images, respectively. Using the region defined by the effective breast region mask as the region of interest, calculate the spatial coupling features between the blood vessel mask and the high-temperature region mask; The spatial coupling features include: the proportion of the intersection of blood vessel pixels and high-temperature region pixels within the region of interest to the total number of blood vessel pixels, the proportion of blood vessel pixels in the high-temperature region pixels, the normalized distance from the centroid of the blood vessel mask to the centroid of the high-temperature region mask, and the normalized temperature difference between the average temperature of the high-temperature region and the average temperature of the region of interest.

5. The method according to claim 4, characterized in that, The malignancy pattern score We obtain the following by weighted linear combination of spatial coupling features: ； in, The percentage of the total number of blood vessel pixels relative to the intersection of blood vessel pixels and high-temperature region pixels within the region of interest. This represents the proportion of pixels with accompanying blood vessels in the high-temperature region. The normalized distance from the centroid of the blood vessel mask to the centroid of the high-temperature region mask is... , Let be the Euclidean distance between the centroid of the blood vessel mask and the centroid of the high-temperature region mask. The difference between the average temperature of the high-temperature region and the average temperature of the region of interest is normalized to the hyperbolic tangent function. Temperature difference value after the interval, The malicious mode score is calculated with preset weighting coefficients and the sum of the four factors being 1. The value range is [0,1].

6. The method according to claim 5, characterized in that, Step S3 also includes Bayesian confidence consistency regularization and dynamic routing sub-steps: In each training batch, the output values ​​and combined confidence scores of the benign bias model and the mafia bias model are acquired synchronously; the prediction consistency during the training period is calculated. ,in and These are the output values ​​of the benign bias model and the malignant bias model, respectively. Constructing fusion coefficients ,in and These are the combined confidence levels of the benign bias model and the malign bias model, respectively. According to the fusion coefficient Calculate the distribution of fused targets Using the fusion target distribution as the regularization objective, KL divergence regularization is applied to both the benign bias model and the malignant bias model. This is to bring the predicted probability distributions of the two models closer to the fusion target distribution, wherein The predicted probability distribution of the currently regularized model; the regularization weights of the KL divergence regularization. The calculation method is as follows ,in This is the preset fusion weight scaling factor; Meanwhile, based on the training period prediction consistency and the malignant pattern score... Dynamic routing is performed as follows: When the training-period prediction consistency is greater than the first training-period consistency threshold, the sample gradient flows to both the benign biased model and the malignant biased model; when the training-period prediction consistency is not greater than the first training-period consistency threshold but greater than the second training-period consistency threshold, the sample gradient flows to both the benign biased model and the malignant biased model with reduced weights; when the training-period prediction consistency is not greater than the second training-period consistency threshold, the malignant pattern score is applied. Determine the routing direction: If the malicious mode score is greater than the preset malicious routing threshold, the sample gradient is routed to the malicious bias model; if the malicious mode score is less than the preset benign routing threshold, the sample gradient is routed to the benign bias model; if the malicious mode score is not less than the preset benign routing threshold and not greater than the preset malicious routing threshold, the sample gradient is routed to the model with higher overall confidence.

7. The method according to claim 1, characterized in that, In step S4, the prediction consistency The calculation method is as follows ,in and These are the output values ​​of the benign bias model and the malignant bias model, respectively. Determining a fusion strategy based on the prediction consistency includes: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The inference period is compared with a preset first consistency threshold and a second consistency threshold, wherein the first consistency threshold is greater than the second consistency threshold. When the prediction consistency When the value exceeds the first consistency threshold during the inference period, the output values ​​of the two models are taken as an equal-weighted average to obtain the fused output value. ; When the prediction consistency When the output value of the two models is not greater than the first consistency threshold during the inference period but is greater than the second consistency threshold during the inference period, the output values ​​of the two models are weighted and fused using the combined confidence scores of the two models as weights to obtain the fused output value. : ,in and These are the combined confidence levels of the benign bias model and the malign bias model, respectively. When the prediction consistency If the sample value is not greater than the second consistency threshold of the inference period, the sample will be submitted to the expert rule for decision-making.

8. The method according to claim 7, characterized in that, The expert rule decisions are made and executed in descending order of priority as follows: Rule 1: If the overall confidence of either the benign bias model or the malign bias model is greater than the preset rule 1 confidence threshold, then the output value of the model with the higher overall confidence is used as the fusion result. Rule 2: If Rule 1 is not satisfied, and the absolute value of the difference between the overall confidence scores of the two models is greater than the preset confidence score difference threshold, then the output value of the model with the higher overall confidence score shall be used as the fusion result. Rule 3: If the output value of any model is greater than the preset upper threshold for extreme prediction or less than the preset lower threshold for extreme prediction, the output value of the model whose output value deviates more from 0.5 by an absolute value shall be used as the fusion result. Rule 4: When none of the above rules 1 to 3 are satisfied, the arithmetic mean of the output values ​​of the two models shall be taken as the fusion result.

9. A thermal image classification system based on dual-model biased training and confidence fusion, applicable to the method described in claim 1, characterized in that, include: The dual-model feature extraction and classification module is configured to input the preprocessed thermal image into a benign bias model and a malignant bias model respectively, and obtain the output values ​​of the benign bias model and the malignant bias model. The confidence assessment module is configured to perform a comprehensive confidence assessment on the outputs of the benign bias model and the malign bias model respectively, and obtain the comprehensive confidence of each model. The comprehensive confidence is obtained by weighted fusion of entropy confidence, evidence confidence and maximum probability confidence. The dynamic bias training module is configured to adaptively adjust the bias intensity according to the loss learning curve during the training phase, and calculate the sample-level bias weight by combining the comprehensive confidence and the true label of the sample. The loss function is weighted by the sample-level bias weight to train the benign bias model and the malignant bias model. The intelligent fusion decision module is configured to calculate the prediction consistency based on the output values ​​of the benign bias model and the malign bias model during the inference phase, determine the fusion strategy based on the prediction consistency, and fuse the output values ​​of the two models and the comprehensive confidence level according to the determined fusion strategy to obtain a prediction result for assisting in the classification of benign and malignant.

Images