Dynamic prototype multiple model medical image classification method based on feature credibility evaluation

Abstract

The application discloses a dynamic prototype multi-model medical image classification method based on feature credibility evaluation, and relates to the technical field of medical image processing.The method effectively solves the problems in traditional medical image subtype classification, such as lack of high-quality labeled data, high uncertainty of pseudo-labels, and difficulty of static prototypes in adapting to dynamic changes of lesions, and the like.Through construction of a multi-structure feature extraction network and completion of hierarchical feature fusion, the method combines feature cross learning, local attention modeling and expanded convolution to supplement context information, and strengthens semantic consistency and structural continuity of lesion region features.Meanwhile, the method constructs a spatial similarity graph through cosine similarity, and models and fuses a spatial uncertainty graph with any and cognitive uncertainty based on a Dirichlet distribution, obtains a reliable evidence graph through exponential fusion, and generates a pseudo-label with sample-level confidence, so that effective supervision information is accurately screened from a feature level, cumulative deviation of false pseudo-labels is greatly reduced, and stability of a semi-supervised learning process is improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a dynamic prototype multi-model medical image classification method based on feature credibility assessment. Background Technology

[0002] Medical image subtype classification is a key task in medical diagnosis, aiming to identify different subtypes within the same disease that have subtle morphological and biological differences. However, in practical applications, traditional supervised learning methods struggle to obtain stable models with good generalization ability due to factors such as the high cost of acquiring high-quality labeled data, insufficient consistency in sample annotation, and significant differences in imaging conditions.

[0003] Semi-supervised learning provides a feasible technical approach for medical image classification by jointly utilizing a small amount of labeled data with a large amount of unlabeled data. However, existing semi-supervised methods still face challenges when applied to medical image subtype classification: on the one hand, subtypes often have only subtle morphological differences, and traditional methods are unable to fully characterize their discriminative features; on the other hand, pseudo-labels generated from unlabeled samples are usually accompanied by high uncertainty, and erroneous pseudo-labels are easily amplified during training, introducing cumulative bias and significantly reducing model performance.

[0004] Prototype-based methods are an effective solution for semi-supervised learning, improving classification stability by learning class-representative prototypes. However, these methods often rely on static or global prototypes, making it difficult to adapt to the dynamic changes in lesion morphology, scale, and distribution in medical images. Furthermore, they lack explicit modeling of feature quality differences in unlabeled samples, limiting the model's generalization ability in complex clinical scenarios. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dynamic prototype multi-model medical image classification method based on feature credibility assessment that can effectively utilize labeled and unlabeled data and improve classification accuracy.

[0006] The dynamic prototype multi-model medical image classification method based on feature credibility assessment proposed in this invention includes the following steps: Step 1: Obtain a medical image dataset containing labeled and unlabeled samples; Step 2: Construct at least two initial feature extraction networks with different structures to extract features from the images in the medical image dataset, respectively. Step 3: Perform hierarchical fusion of the feature maps from multiple levels extracted by each initial feature extraction network to obtain the fused feature map corresponding to each network; Step 4: Based on the fused feature map and category labels of the labeled samples, construct an initial category prototype for each category; Step 5: For unlabeled samples, calculate the spatial similarity between their fused feature map and the prototypes of all categories to obtain a similarity map; Step 6: Perform spatial uncertainty modeling on the fused feature map of the unlabeled samples to obtain an uncertainty map; Step 7: Fuse the similarity map and the uncertainty map to obtain a reliable evidence map, and determine the pseudo-labels of the unlabeled samples and their corresponding sample-level confidence levels based on the reliable evidence map.

[0007] Preferably, step 3 involves hierarchical fusion of feature maps from multiple levels extracted by each initial feature extraction network, including: Select shallow and deep feature maps from different layers in the network; For two adjacent feature maps, feature cross-learning is first performed to obtain cross-layer fused features; Attention learning is applied to the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features. The attention-enhanced features are input into an extended feedforward network to supplement the contextual structure information across windows, resulting in the final fused feature map.

[0008] Preferably, the feature cross-learning includes: Taking the feature maps of two adjacent layers as input, let the first layer be denoted as . The layer feature map is Fl, the first... The layer feature map is Fl+1. When the spatial resolution of the two layers is inconsistent, bilinear interpolation is first used to upsample the deep features to obtain an aligned feature map, so as to ensure its consistency with Fl in spatial dimension.

[0009] The cross-learning stage includes two parallel branches: the basic feature extraction branch and the weight generation branch; the basic feature extraction branch uses a combination of BN, PConv

[33] and DSC followed by ReLU activation to extract the basic representations of adjacent layers; the weight generation branch has a corresponding structure, but uses Sigmoid activation at the end to generate a weight map to characterize the importance of different spatial locations; specifically, it can be expressed as: in Represents ReLU. It represents Sigmoid; Subsequently, element-wise multiplication is used to achieve cross-modulation across layers: deep layer weights modulate shallow layer features to guide them to focus on semantically meaningful regions; shallow layer weights constrain deep layer features, supplementing their spatial structure and boundary information, ultimately yielding cross-layer fused features. : .

[0010] This cross-learning mechanism enables bidirectional information interaction between the two layers of features during the fusion process: the shallow layer obtains a more consistent semantic response under the guidance of the deep layer weights, while the deep layer retains more spatial structure and detailed expression under the constraint of the shallow layer weights, thereby alleviating the representation offset between different semantic levels.

[0011] Preferably, attention learning is performed on the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features, as follows: In obtaining cross-layer fusion features Subsequently, to further model the spatial semantic consistency within the lesion region and enhance local discriminative ability, this paper introduces attention learning. By modeling the similarity relationships between features within a local window, the model can focus on pathological structural regions with discriminative value, thereby effectively suppressing the interference of background regions on feature expression.

[0012] Let the cross-layer fusion feature be First, divide it into several sizes. Non-overlapping local windows, each containing A spatial location; through sharing Convolution performs a linear mapping on the features within the window, generating query terms accordingly. Key items AND value term To enhance spatial awareness within the window, a learnable relative position offset matrix is ​​introduced. Used to describe the geometric relationships between different spatial locations; in the first Within a local window, the self-attention weights are calculated using the standard scaled dot product method, and feature aggregation is completed. This process is represented as follows: in , This represents the feature channel dimension. The attention matrix explicitly models the similarity relationship between any two spatial locations within the window, thereby enhancing the semantic consistency of expression within local regions.

[0013] The output features of all windows are rearranged and concatenated to restore the original spatial resolution, and then combined with layer normalization and residual connections to obtain attention-enhanced feature representations: Building upon this, to further supplement cross-window contextual structure information, features are fed into a Dilated Feed-Forward Network (D-FFN) for nonlinear mapping. This module expands the receptive field through multi-level dilated convolutions, enhancing local contextual modeling capabilities while maintaining a relatively constant parameter scale. Its computation process is as follows: in Indicates the expansion rate The convolution operation is performed. This paper adopts a two-stage dilated convolution structure, with dilation rates set to... , This is to enhance context modeling capabilities while avoiding a significant increase in the number of parameters; Finally, the structural compensation features and attention enhancement features are fused through a second residual connection to obtain the module output: Through the synergistic effect of attention modeling and structural compensation, this module effectively enhances the local consistency and structural continuity within the lesion region while maintaining the semantic stability of cross-layer fusion features, providing a more reliable feature representation basis for the joint evaluation of the similarity map and uncertainty map.

[0014] Preferably, in step 4, based on the fused feature map and its category label of the labeled samples, an initial category prototype is constructed for each category, as follows: In step 4, class labels are assigned to samples by calculating the distance between sample features and prototype features of each class. This study utilizes a dual model to generate two prototypes for each class, thereby more effectively capturing the various variations and complexities within each class, reducing the reliance on large amounts of labeled data, and combining distance-based classification and voting mechanisms to effectively handle small sample sizes and class imbalance issues in medical image classification.

[0015] First, network models (ResNet and ViT are used in this paper) are trained on labeled data to extract features, which are then bound to ground truth labels. These features are then aggregated into prototypes pResNet and pViT. These prototypes represent the core features of each category, providing a benchmark for classifying unlabeled data. Assuming there are m categories, the prototype extracted by ResNet is pResNet,m, and the prototype extracted by ViT is pViT,m. For the m-th category, the formula for calculating its prototype feature pm is: Where Nm is the number of samples in the m-th category, and xi,m is the feature vector of the i-th sample in the m-th category; For unlabeled data, the same model is used to extract features, and the Euclidean distance between the feature vector and the prototype p of each category is calculated to measure its similarity with the prototypes of different categories. Specifically, for the feature x of the unlabeled image extracted by the model and the prototype p, their Euclidean distance d(x,p) is calculated as a similarity measure, as shown in the formula: Subsequently, a distance-based voting prediction mechanism is employed. For a sample feature x, its distance d(x, pm) with all prototype features is first calculated, and the distance is converted into a similarity score, as shown in the formula: Here, σ is a temperature parameter used to control the distribution of similarity scores.

[0016] If there are prototype features with a distance less than the threshold τ, then voting is performed based on the categories corresponding to these prototype features that meet the conditions: Where ι is the set of prototype indices with a distance less than the threshold, li is the category corresponding to the i-th prototype, and Ⅰ( () is the indicator function; the final predicted category is the category with the most votes. If all distances are greater than the threshold, the category corresponding to the prototype feature with the smallest distance is selected as the prediction result; then the results of the two prototypes, pResNet and pViT, are weighted and fused. The weights are inversely proportional to the distance; the smaller the distance, the higher the weight, indicating that the image feature is closer to the prototype. The initial weights are calculated as follows: Subsequently, weight normalization is performed so that the sum of the weights for each model is 1, ensuring that the influence of each prototype is appropriately scaled; the final weights are the average of the normalized weights of all models; by combining information from different models, a unified decision is made; the parameters of the weight layer are optimized through the training process to maximize classification accuracy. Each category prototype corresponds to a weight value, and these weight values ​​are updated through the backpropagation algorithm to minimize classification error. The final labels are determined as follows: This label represents the category to which the image is most likely to belong after combining information from all models. By combining the weights of multiple models, a more representative prototype is allowed to dominate the classification decision, avoiding the risk of misclassification from a single model prototype.

[0017] Preferably, in step 5, for unlabeled samples, the spatial similarity between their fused feature map and the prototypes of all categories is calculated to obtain a similarity map, as follows: If input images are classified using traditional methods, the entire feature extraction and classification process is an end-to-end black box operation, making it difficult to explain the model's decision-making process. Secondly, spatial information in the feature map is lost during global pooling, making it impossible to accurately locate features in a specific part of the image, resulting in insufficient representation of complex image structures. This paper designs a feature filtering layer that fuses similarity maps and uncertainty maps. This module, as the front-end processing unit of the model, operates on the multi-scale fused feature map to filter feature representations with reliable discriminative evidence from unlabeled data, providing high-quality input for subsequent prototype matching and pseudo-label generation. By jointly modeling the feature-prototype matching relationship and its spatial uncertainty distribution, this feature filtering layer can effectively suppress noisy samples at the feature level, thereby improving the stability of the semi-supervised learning process.

[0018] Similarity maps are used to depict the spatial matching relationship between sample features and class prototypes, focusing on fine-grained discriminative information of local features.

[0019] Given unlabeled samples The feature map is obtained after multi-scale feature fusion. For each category of prototype This paper matches the feature vectors of all spatial locations in the feature map one by one, and constructs a category-related similarity map by calculating the similarity relationship between the feature vectors and the prototype.

[0020] Unlike Euclidean distance-based metrics, this paper uses cosine similarity to match feature vectors at various spatial locations in the feature map with the class prototype. Cosine similarity assesses similarity by measuring the angle between two vectors, effectively eliminating interference from feature amplitude variations and exhibiting good scale-invariance properties.

[0021] set up Represents the first feature in the feature graph The feature vector at each spatial location, then its relationship with the prototype The similarity can be represented as: in The L2 norm of a vector.

[0022] The similarity map reflects the matching strength between each spatial location in the feature map and the prototype. Its high-response regions correspond to local discriminative features that are semantically consistent with the current category. By modeling fine-grained relationships between features and prototypes at the spatial level, the similarity map provides a more accurate and robust local matching basis for subsequent pseudo-label generation, rather than relying solely on the global feature vector.

[0023] Preferably, in step 6, spatial uncertainty modeling is performed on the fused feature map of the unlabeled samples to obtain an uncertainty map, as follows: In unlabeled samples, relying solely on the local matching information provided by the similarity map can still be affected by noisy features or accidental high-response regions. Although some spatial locations show high consistency with the class prototype in feature direction, their cognitive stability in classification decisions remains weak, easily introducing erroneous supervision during subsequent pseudo-label generation. To address this, this paper further introduces spatial classification uncertainty modeling to evaluate the cognitive reliability of the model when making class distinctions at different spatial locations from a feature perspective.

[0024] Given an input CT image After passing through the backbone network and multi-scale fusion module, the feature map is obtained: On this feature map, this paper constructs a convolutional evidence head. It is used to model local evidence while preserving spatial structure information, and its output is expressed as: Among them, the channel dimension is expanded to This means that at each spatial location, four sets of evidence parameters are predicted for each category.

[0025] Within the framework of evidence-based deep learning and subjective logic modeling, for spatial location and categories The parameters output by the evidence header are used to construct a pixel-by-pixel Dirichlet distribution. To ensure numerical stability, this paper adopts the following mapping method: in It is a very small constant used to avoid numerical instability.

[0026] Subsequently, based on the principle of subjective logic modeling, the pixel-by-pixel Dirichlet concentration parameter is defined as: After obtaining the pixel-by-pixel Dirichlet parameters, this paper further aggregates the uncertainties of all categories to construct a spatial uncertainty map. Specifically, spatial location... Uncertainty at a given point is defined as: The first term corresponds to arbitrary uncertainty, which mainly comes from imaging noise and local structural blurring; the second term corresponds to epistemic uncertainty, which reflects the instability of the model's discrimination at that spatial location due to insufficient knowledge.

[0027] This yields the spatial uncertainty diagram: To facilitate subsequent fusion and interpretable analysis, this paper normalizes the uncertainty graph to the [0,1] interval, enabling it to be jointly modeled with the similarity graph on the same spatial scale.

[0028] This spatial uncertainty map characterizes the cognitive credibility of the model when participating in global classification decisions at different spatial locations. Higher uncertainty values ​​indicate that the features in that region are less likely to provide consistent and reliable evidence for category judgment; regions with lower uncertainty typically correspond to stable and discriminative key semantic regions. By explicitly modeling spatial uncertainty, the model can effectively distinguish between pseudo-correlation responses that are semantically consistent but discriminatively unstable, and stable and reliable discriminative evidence, laying the foundation for subsequent uncertainty-guided feature selection and pseudo-label generation.

[0029] Preferably, in step 7, the similarity map and the uncertainty map are fused to obtain a reliable evidence map, and the pseudo-labels of the unlabeled samples and their corresponding sample-level confidence levels are determined based on the reliable evidence map, as follows: In the feature selection process for unlabeled samples, relying solely on the local matching information provided by the feature prototype similarity map is insufficient to guarantee the quality of pseudo-labels. While some spatial locations may show high consistency with a certain category's prototype in the feature direction, their corresponding prediction uncertainty remains high. Such responses often originate from noisy regions, blurred local structures, or feature patterns that the model has not yet fully learned. Directly using these as the criterion for pseudo-label generation can easily introduce unstable or even erroneous supervisory signals.

[0030] To address this, this paper introduces spatial prediction uncertainty as a reliability constraint during the feature selection stage and employs an exponential fusion strategy to modulate similarity evidence, thereby enabling the automatic discovery of stable discriminative regions in unlabeled samples. This spatial uncertainty includes both arbitrary and cognitive uncertainties, comprehensively characterizing the reliability of local features as a supervisory basis from both data noise and model cognition perspectives.

[0031] Specifically, for unlabeled samples Let the feature-prototype similarity graph related to its category be... Spatial prediction uncertainty diagram is as follows Exponential fusion modulates the similarity map point-by-point by mapping uncertainty to exponentially decaying weights. Its fusion form is defined as follows: in This indicates an element-wise multiplication operation. This is the uncertainty attenuation coefficient, used to control the suppression strength of uncertainty on the similarity response.

[0032] This exponential fusion mechanism has a clear discriminative meaning: when the uncertainty at a spatial location is high, the corresponding similarity response will be exponentially weakened; while in low-uncertainty regions, feature responses consistent with the semantic direction of the category prototype will be effectively preserved. In this way, the model can explicitly suppress noisy responses with high similarity but high uncertainty, and highlight stable discriminative regions with high similarity and low uncertainty.

[0033] In obtaining reliable evidence maps related to categories Subsequently, this paper further aggregates spatial-level evidence into sample-level confidence scores. Considering that lesion areas in medical images typically exhibit a locally sparse distribution, averaging the entire evidence image would easily dilute it with background areas. Therefore, this paper employs a top-k spatial aggregation strategy, selecting only the local regions with the strongest responses for sample-level discrimination. Specifically, for each category c, from... Select the one with the largest response value Each spatial location is used as the sample-level confidence score for that category, and their mean is taken as the score. This allows us to obtain the sample-level confidence sets for each category. This paper uses the category with the highest confidence level as the pseudo-label for the unlabeled sample. The highest confidence level is then used as the final credibility score for that sample. This sample-level confidence score comprehensively reflects the consistency of the most discriminative local regions in unlabeled samples in terms of both semantic relevance and prediction stability, providing a reliable basis for subsequent confidence-based sample utilization strategies.

[0034] The dynamic prototype multi-model medical image classification method based on feature credibility assessment proposed in this invention has the following beneficial technical effects: This method effectively addresses the problems of scarce high-quality labeled data, high uncertainty of pseudo-labels, and difficulty in adapting static prototypes to dynamic changes in lesions in traditional medical image subtype classification. By constructing a multi-structure feature extraction network and completing hierarchical feature fusion, and combining feature cross-learning, local attention modeling, and dilated convolution to supplement contextual information, the semantic consistency and structural continuity of lesion region features are enhanced. At the same time, a spatial similarity map is constructed by cosine similarity, and a spatial uncertainty map based on Dirichlet distribution modeling is fused with arbitrary and cognitive uncertainty. Reliable evidence maps are obtained through exponential fusion, and pseudo-labels with sample-level confidence are generated. Effective supervision information is accurately screened at the feature level, which greatly reduces the cumulative bias of erroneous pseudo-labels and improves the stability of the semi-supervised learning process. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0037] like Figure 1 The dynamic prototype multi-model medical image classification method based on feature credibility assessment, as shown, includes the following steps: Step 1: Obtain a medical image dataset containing labeled and unlabeled samples; Step 2: Construct at least two initial feature extraction networks with different structures to extract features from the images in the medical image dataset, respectively. Step 3: Perform hierarchical fusion of the feature maps from multiple levels extracted by each initial feature extraction network to obtain the fused feature map corresponding to each network; Step 4: Based on the fused feature map and category labels of the labeled samples, construct an initial category prototype for each category; Step 5: For unlabeled samples, calculate the spatial similarity between their fused feature map and the prototypes of all categories to obtain a similarity map; Step 6: Perform spatial uncertainty modeling on the fused feature map of the unlabeled samples to obtain an uncertainty map; Step 7: Fuse the similarity map and the uncertainty map to obtain a reliable evidence map, and determine the pseudo-labels of the unlabeled samples and their corresponding sample-level confidence levels based on the reliable evidence map.

[0038] In an optional embodiment, step 3 involves hierarchically fusing the feature maps from multiple levels extracted by each initial feature extraction network, including: Select shallow and deep feature maps from different layers in the network; For two adjacent feature maps, feature cross-learning is first performed to obtain cross-layer fused features; Attention learning is applied to the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features. The attention-enhanced features are input into an extended feedforward network to supplement the contextual structure information across windows, resulting in the final fused feature map.

[0039] In an optional embodiment, the feature cross-learning includes: Taking the feature maps of two adjacent layers as input, let the first layer be denoted as . The layer feature map is Fl, the first... The layer feature map is Fl+1. When the spatial resolution of the two layers is inconsistent, bilinear interpolation is first used to upsample the deep features to obtain an aligned feature map, so as to ensure its consistency with Fl in spatial dimension.

[0040] The cross-learning stage includes two parallel branches: the basic feature extraction branch and the weight generation branch; the basic feature extraction branch uses a combination of BN, PConv

[33] and DSC followed by ReLU activation to extract the basic representations of adjacent layers; the weight generation branch has a corresponding structure, but uses Sigmoid activation at the end to generate a weight map to characterize the importance of different spatial locations; specifically, it can be expressed as: in Represents ReLU. It represents Sigmoid; Subsequently, element-wise multiplication is used to achieve cross-modulation across layers: deep layer weights modulate shallow layer features to guide them to focus on semantically meaningful regions; shallow layer weights constrain deep layer features, supplementing their spatial structure and boundary information, ultimately yielding cross-layer fused features. : .

[0041] This cross-learning mechanism enables bidirectional information interaction between the two layers of features during the fusion process: the shallow layer obtains a more consistent semantic response under the guidance of the deep layer weights, while the deep layer retains more spatial structure and detailed expression under the constraint of the shallow layer weights, thereby alleviating the representation offset between different semantic levels.

[0042] In an optional embodiment, attention learning is performed on the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features, as follows: In obtaining cross-layer fusion features Subsequently, to further model the spatial semantic consistency within the lesion region and enhance local discriminative ability, this paper introduces attention learning. By modeling the similarity relationships between features within a local window, the model can focus on pathological structural regions with discriminative value, thereby effectively suppressing the interference of background regions on feature expression.

[0043] Let the cross-layer fusion feature be First, divide it into several sizes. Non-overlapping local windows, each containing A spatial location; through sharing Convolution performs a linear mapping on the features within the window, generating query terms accordingly. Key items AND value term To enhance spatial awareness within the window, a learnable relative position offset matrix is ​​introduced. Used to describe the geometric relationships between different spatial locations; in the first Within a local window, the self-attention weights are calculated using the standard scaled dot product method, and feature aggregation is completed. This process is represented as follows: in , This represents the feature channel dimension. The attention matrix explicitly models the similarity relationship between any two spatial locations within the window, thereby enhancing the semantic consistency of expression within local regions.

[0044] The output features of all windows are rearranged and concatenated to restore the original spatial resolution, and then combined with layer normalization and residual connections to obtain attention-enhanced feature representations: Building upon this, to further supplement cross-window contextual structure information, features are fed into a Dilated Feed-Forward Network (D-FFN) for nonlinear mapping. This module expands the receptive field through multi-level dilated convolutions, enhancing local contextual modeling capabilities while maintaining a relatively constant parameter scale. Its computation process is as follows: in Indicates the expansion rate The convolution operation is performed. This paper adopts a two-stage dilated convolution structure, with dilation rates set to... , This is to enhance context modeling capabilities while avoiding a significant increase in the number of parameters; Finally, the structural compensation features and attention enhancement features are fused through a second residual connection to obtain the module output: Through the synergistic effect of attention modeling and structural compensation, this module effectively enhances the local consistency and structural continuity within the lesion region while maintaining the semantic stability of cross-layer fusion features, providing a more reliable feature representation basis for the joint evaluation of the similarity map and uncertainty map.

[0045] In an optional embodiment, in step 4, based on the fused feature map and its category label of the labeled samples, an initial category prototype is constructed for each category as follows: In step 4, class labels are assigned to samples by calculating the distance between sample features and prototype features of each class. This study utilizes a dual model to generate two prototypes for each class, thereby more effectively capturing the various variations and complexities within each class, reducing the reliance on large amounts of labeled data, and combining distance-based classification and voting mechanisms to effectively handle small sample sizes and class imbalance issues in medical image classification.

[0046] First, network models (ResNet and ViT are used in this paper) are trained on labeled data to extract features, which are then bound to ground truth labels. These features are then aggregated into prototypes pResNet and pViT. These prototypes represent the core features of each category, providing a benchmark for classifying unlabeled data. Assuming there are m categories, the prototype extracted by ResNet is pResNet,m, and the prototype extracted by ViT is pViT,m. For the m-th category, the formula for calculating its prototype feature pm is: Where Nm is the number of samples in the m-th category, and xi,m is the feature vector of the i-th sample in the m-th category; For unlabeled data, the same model is used to extract features, and the Euclidean distance between the feature vector and the prototype p of each category is calculated to measure its similarity with the prototypes of different categories. Specifically, for the feature x of the unlabeled image extracted by the model and the prototype p, their Euclidean distance d(x,p) is calculated as a similarity measure, as shown in the formula: Subsequently, a distance-based voting prediction mechanism is employed. For a sample feature x, its distance d(x, pm) with all prototype features is first calculated, and the distance is converted into a similarity score, as shown in the formula: Here, σ is a temperature parameter used to control the distribution of similarity scores.

[0047] If there are prototype features with a distance less than the threshold τ, then voting is performed based on the categories corresponding to these prototype features that meet the conditions: Where ι is the set of prototype indices with a distance less than the threshold, li is the category corresponding to the i-th prototype, and Ⅰ( () is the indicator function; the final predicted category is the category with the most votes. If all distances are greater than the threshold, the category corresponding to the prototype feature with the smallest distance is selected as the prediction result; then the results of the two prototypes, pResNet and pViT, are weighted and fused. The weights are inversely proportional to the distance; the smaller the distance, the higher the weight, indicating that the image feature is closer to the prototype. The initial weights are calculated as follows: Subsequently, weight normalization is performed so that the sum of the weights for each model is 1, ensuring that the influence of each prototype is appropriately scaled; the final weights are the average of the normalized weights of all models; by combining information from different models, a unified decision is made; the parameters of the weight layer are optimized through the training process to maximize classification accuracy. Each category prototype corresponds to a weight value, and these weight values ​​are updated through the backpropagation algorithm to minimize classification error. The final labels are determined as follows: This label represents the category to which the image is most likely to belong after combining information from all models. By combining the weights of multiple models, a more representative prototype is allowed to dominate the classification decision, avoiding the risk of misclassification from a single model prototype.

[0048] In an optional embodiment, in step 5, for unlabeled samples, the spatial similarity between their fused feature map and all category prototypes is calculated to obtain a similarity map, as follows: If input images are classified using traditional methods, the entire feature extraction and classification process is an end-to-end black box operation, making it difficult to explain the model's decision-making process. Secondly, spatial information in the feature map is lost during global pooling, making it impossible to accurately locate features in a specific part of the image, resulting in insufficient representation of complex image structures. This paper designs a feature filtering layer that fuses similarity maps and uncertainty maps. This module, as the front-end processing unit of the model, operates on the multi-scale fused feature map to filter feature representations with reliable discriminative evidence from unlabeled data, providing high-quality input for subsequent prototype matching and pseudo-label generation. By jointly modeling the feature-prototype matching relationship and its spatial uncertainty distribution, this feature filtering layer can effectively suppress noisy samples at the feature level, thereby improving the stability of the semi-supervised learning process.

[0049] Similarity maps are used to depict the spatial matching relationship between sample features and class prototypes, focusing on fine-grained discriminative information of local features.

[0050] Given unlabeled samples The feature map is obtained after multi-scale feature fusion. For each category of prototype This paper matches the feature vectors of all spatial locations in the feature map one by one, and constructs a category-related similarity map by calculating the similarity relationship between the feature vectors and the prototype.

[0051] Unlike Euclidean distance-based metrics, this paper uses cosine similarity to match feature vectors at various spatial locations in the feature map with the class prototype. Cosine similarity assesses similarity by measuring the angle between two vectors, effectively eliminating interference from feature amplitude variations and exhibiting good scale-invariance properties.

[0052] set up Represents the first feature in the feature graph The feature vector at each spatial location, then its relationship with the prototype The similarity can be represented as: in The L2 norm of a vector.

[0053] The similarity map reflects the matching strength between each spatial location in the feature map and the prototype. Its high-response regions correspond to local discriminative features that are semantically consistent with the current category. By modeling fine-grained relationships between features and prototypes at the spatial level, the similarity map provides a more accurate and robust local matching basis for subsequent pseudo-label generation, rather than relying solely on the global feature vector.

[0054] In an optional embodiment, step 6 involves performing spatial uncertainty modeling on the fused feature map of the unlabeled samples to obtain an uncertainty map, as follows: In unlabeled samples, relying solely on the local matching information provided by the similarity map can still be affected by noisy features or accidental high-response regions. Although some spatial locations show high consistency with the class prototype in feature direction, their cognitive stability in classification decisions remains weak, easily introducing erroneous supervision during subsequent pseudo-label generation. To address this, this paper further introduces spatial classification uncertainty modeling to evaluate the cognitive reliability of the model when making class distinctions at different spatial locations from a feature perspective.

[0055] Given an input CT image After passing through the backbone network and multi-scale fusion module, the feature map is obtained: On this feature map, this paper constructs a convolutional evidence head. It is used to model local evidence while preserving spatial structure information, and its output is expressed as: Among them, the channel dimension is expanded to This means that at each spatial location, four sets of evidence parameters are predicted for each category.

[0056] Within the framework of evidence-based deep learning and subjective logic modeling, for spatial location and categories The parameters output by the evidence header are used to construct a pixel-by-pixel Dirichlet distribution. To ensure numerical stability, this paper adopts the following mapping method: in It is a very small constant used to avoid numerical instability.

[0057] Subsequently, based on the principle of subjective logic modeling, the pixel-by-pixel Dirichlet concentration parameter is defined as: After obtaining the pixel-by-pixel Dirichlet parameters, this paper further aggregates the uncertainties of all categories to construct a spatial uncertainty map. Specifically, spatial location... Uncertainty at a given point is defined as: The first term corresponds to arbitrary uncertainty, which mainly comes from imaging noise and local structural blurring; the second term corresponds to epistemic uncertainty, which reflects the instability of the model's discrimination at that spatial location due to insufficient knowledge.

[0058] This yields the spatial uncertainty diagram: To facilitate subsequent fusion and interpretable analysis, this paper normalizes the uncertainty graph to the [0,1] interval, enabling it to be jointly modeled with the similarity graph on the same spatial scale.

[0059] This spatial uncertainty map characterizes the cognitive credibility of the model when participating in global classification decisions at different spatial locations. Higher uncertainty values ​​indicate that the features in that region are less likely to provide consistent and reliable evidence for category judgment; regions with lower uncertainty typically correspond to stable and discriminative key semantic regions. By explicitly modeling spatial uncertainty, the model can effectively distinguish between pseudo-correlation responses that are semantically consistent but discriminatively unstable, and stable and reliable discriminative evidence, laying the foundation for subsequent uncertainty-guided feature selection and pseudo-label generation.

[0060] In an optional embodiment, step 7 involves fusing the similarity map and the uncertainty map to obtain a reliable evidence map, and determining the pseudo-labels of the unlabeled samples and their corresponding sample-level confidence levels based on the reliable evidence map, as follows: In the feature selection process for unlabeled samples, relying solely on the local matching information provided by the feature prototype similarity map is insufficient to guarantee the quality of pseudo-labels. While some spatial locations may show high consistency with a certain category's prototype in the feature direction, their corresponding prediction uncertainty remains high. Such responses often originate from noisy regions, blurred local structures, or feature patterns that the model has not yet fully learned. Directly using these as the criterion for pseudo-label generation can easily introduce unstable or even erroneous supervisory signals.

[0061] To address this, this paper introduces spatial prediction uncertainty as a reliability constraint during the feature selection stage and employs an exponential fusion strategy to modulate similarity evidence, thereby enabling the automatic discovery of stable discriminative regions in unlabeled samples. This spatial uncertainty includes both arbitrary and cognitive uncertainties, comprehensively characterizing the reliability of local features as a supervisory basis from both data noise and model cognition perspectives.

[0062] Specifically, for unlabeled samples Let the feature-prototype similarity graph related to its category be... Spatial prediction uncertainty diagram is as follows Exponential fusion modulates the similarity map point-by-point by mapping uncertainty to exponentially decaying weights. Its fusion form is defined as follows: in This indicates an element-wise multiplication operation. This is the uncertainty attenuation coefficient, used to control the suppression strength of uncertainty on the similarity response.

[0063] This exponential fusion mechanism has a clear discriminative meaning: when the uncertainty at a spatial location is high, the corresponding similarity response will be exponentially weakened; while in low-uncertainty regions, feature responses consistent with the semantic direction of the category prototype will be effectively preserved. In this way, the model can explicitly suppress noisy responses with high similarity but high uncertainty, and highlight stable discriminative regions with high similarity and low uncertainty.

[0064] In obtaining reliable evidence maps related to categories Subsequently, this paper further aggregates spatial-level evidence into sample-level confidence scores. Considering that lesion areas in medical images typically exhibit a locally sparse distribution, averaging the entire evidence image would easily dilute it with background areas. Therefore, this paper employs a top-k spatial aggregation strategy, selecting only the local regions with the strongest responses for sample-level discrimination. Specifically, for each category c, from... Select the one with the largest response value Each spatial location is used as the sample-level confidence score for that category, and their mean is taken as the score. This allows us to obtain the sample-level confidence sets for each category. This paper uses the category with the highest confidence level as the pseudo-label for the unlabeled sample. The highest confidence level is then used as the final credibility score for that sample. This sample-level confidence score comprehensively reflects the consistency of the most discriminative local regions in unlabeled samples in terms of both semantic relevance and prediction stability, providing a reliable basis for subsequent confidence-based sample utilization strategies.

[0065] It also includes a closed-loop learning process consisting of pseudo-label generation, filtering, and expansion to gradually uncover effective supervisory information in unlabeled data, thereby improving the model's generalization ability in liver cancer subtyping tasks. Specifically, the model first uses a feature filtering layer to model discriminative evidence for unlabeled samples, and by fusing feature-prototype similarity with spatial prediction uncertainty, obtains a reliable evidence map related to the category, and calculates sample-level confidence based on this map.

[0066] In each training round, for all unlabeled samples, this paper uses the sample-level confidence level corresponding to the final reliable evidence map. Calculate the mean of the built-in confidence distribution of the current epoch. with standard deviation Based on this, a dynamic confidence threshold is constructed: When the sample satisfies When a sample is identified as a high-quality pseudo-label, its corresponding pseudo-label is added to the training set to participate in the next round of model updates; the remaining samples are not used as strong supervision signals for the time being.

[0067] This dynamic thresholding mechanism is based on the confidence statistical distribution of all unlabeled samples, enabling the sample selection criteria to adaptively adjust as the overall distribution predicted by the model changes. By simultaneously considering the central tendency and dispersion of the confidence distribution, this method avoids the problem of being too strict or too lenient that a fixed threshold may cause at different training stages, ensuring that the pseudo-label selection process remains consistent with the current discrimination state of the model.

[0068] By employing the dynamic pseudo-label selection strategy based on reliable evidence graphs, the model can gradually introduce credible unlabeled samples into training while ensuring the quality of pseudo-labels, thereby forming a stable semi-supervised closed-loop learning process and effectively improving the model's generalization ability to complex liver cancer subtype samples.

[0069] Unlabeled data is categorized into high-quality and low-quality data through a feature filtering layer. Discarding low-quality data during training would reduce the diversity of training data, potentially hindering the model's ability to handle boundary or complex situations. Therefore, this study employs a co-training module, adding high-quality pseudo-labeled samples to the training set to update model parameters, while simultaneously utilizing feature feedback from low-quality pseudo-labeled samples to update prototype parameters, thereby enhancing the model's generalization ability.

[0070] Each model (such as ResNet and ViT) independently maintains its own class prototype and fine-tunes it using low-confidence pseudo-label samples it generates. Through feature feedback from these low-quality pseudo-labels, it dynamically adjusts the prototype parameters, balancing model robustness and update efficiency. Specifically, for each class m, the prototype update formula is: Where poldm is the old prototype, f(x) is the feature of the low-confidence pseudo-label sample x corresponding to category m, Dlowq,m is the set of low-confidence pseudo-label samples of category m, and Nlowq is the number of samples in Dlowq,m.

[0071] The core of the collaborative training module lies in the complementarity between independent learning by multiple models and dynamic prototype updates. Different models extract features from different perspectives and maintain prototypes. They adjust the prototypes based on feature feedback from low-quality pseudo-labels, leveraging the implicit differences between models to form complementarity and adaptively improve the reliability of pseudo-label selection.

[0072] To simultaneously ensure the model's classification and uncertainty modeling effectiveness, this paper designs a joint optimization composite loss function that collaboratively constrains the classification branch and the uncertainty estimation branch. The overall training objective is defined as: in, For classification cross-entropy loss, For the uncertainty calibration loss, This is the balance coefficient between the two parts of the loss. In the experiment presented in this paper, we take... The model only participates in loss calculation on labeled samples and high-quality pseudo-labeled samples obtained through screening; unlabeled samples that are not selected do not participate in the backpropagation process.

[0073] The classification loss is used to supervise the model's learning of stable class discrimination ability. For input samples, the model outputs sample-level prediction logits through the classification head, and optimizes it using the cross-entropy loss function, which has the following form: in For the number of categories, Indicates the true class label of the sample. This is the predicted probability obtained after applying softmax normalization to the output of the classification head.

[0074] This loss serves as the primary supervisory signal during model training, guiding the network to form clear and stable class discrimination boundaries in the feature space, thus providing a reliable foundation for subsequent prototype comparison and uncertainty modeling.

[0075] To ensure that the uncertainty map output by the model accurately reflects the reliability of the classification prediction, this paper further introduces an uncertainty calibration loss to constrain the evidence branch. The model generates a spatial uncertainty map U on the multi-scale fused feature map using the evidence head, with its value normalized to [0,1]. A larger value indicates greater uncertainty in the model's classification decision at that spatial location. Since explicit uncertainty labeling does not exist in real-world tasks, this paper employs a weakly supervised approach based on global prediction correctness to construct the uncertainty learning signal. Specifically, the model's global classification result is used to determine whether a sample is correctly predicted. The image-level accuracy scalar is then extended to the spatial dimension, and a target uncertainty map is constructed accordingly. When the model makes a wrong prediction of a sample, its overall discrimination result is unreliable and should respond with a high level of uncertainty; when the prediction is correct, a lower level of uncertainty should be output.

[0076] Based on this, Mean Squared Error (MSE) is used to constrain the difference between the predicted uncertainty map and the target uncertainty map, which is defined as: This loss encourages the model to output a higher uncertainty response on samples that are mispredicted and to maintain a lower uncertainty on samples that are correctly predicted, thereby achieving effective calibration of the uncertainty estimation results.

[0077] It should be noted that this uncertainty supervision is not pixel-level manual annotation, but a weak supervision signal derived from the image-level classification results. Its goal is not to directly improve classification accuracy, but to improve the credibility and stability of the uncertainty map in subsequent sample screening and pseudo-label quality assessment.

[0078] This method effectively addresses the problems of scarce high-quality labeled data, high uncertainty of pseudo-labels, and difficulty in adapting static prototypes to dynamic changes in lesions in traditional medical image subtype classification. By constructing a multi-structure feature extraction network and completing hierarchical feature fusion, and combining feature cross-learning, local attention modeling, and dilated convolution to supplement contextual information, the semantic consistency and structural continuity of lesion region features are enhanced. At the same time, a spatial similarity map is constructed by cosine similarity, and a spatial uncertainty map based on Dirichlet distribution modeling is fused with arbitrary and cognitive uncertainty. Reliable evidence maps are obtained through exponential fusion, and pseudo-labels with sample-level confidence are generated. Effective supervision information is accurately screened at the feature level, which greatly reduces the cumulative bias of erroneous pseudo-labels and improves the stability of the semi-supervised learning process.

[0079] Furthermore, this method dynamically divides unlabeled samples into high-quality and low-quality samples based on confidence levels. It employs a collaborative training mechanism to use high-quality samples to update network parameters and low-quality samples to dynamically fine-tune category prototypes. This ensures the reliability of the supervision signal while preserving the diversity of training data, preventing the model from experiencing a decline in generalization ability due to the discarding of low-quality samples. Combined with the joint optimization of the loss function using classification cross-entropy and uncertainty calibration, it achieves a synergistic improvement in classification discrimination ability and uncertainty modeling. Ultimately, it significantly improves the accuracy of medical image subtype classification, enhances the model's generalization ability in complex clinical scenarios with dynamic changes in lesion morphology, scale, and distribution, and effectively handles small sample size and class imbalance problems, greatly reducing the dependence on high-quality labeled data.

[0080] For clarification, "acquisition" in this application refers to obtaining the required content or data using existing technical means.

[0081] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0082] In the embodiments provided by this invention, it should be understood that the disclosed system or method can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0083] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0085] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0086] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic prototype multi-model medical image classification method based on feature credibility assessment, characterized in that, Includes the following steps: Step 1: Obtain a medical image dataset containing labeled and unlabeled samples; Step 2: Construct at least two initial feature extraction networks with different structures to extract features from the images in the medical image dataset, respectively. Step 3: Perform hierarchical fusion of the feature maps from multiple levels extracted by each initial feature extraction network to obtain the fused feature map corresponding to each network; Step 4: Based on the fused feature map and category labels of the labeled samples, construct an initial category prototype for each category; Step 5: For unlabeled samples, calculate the spatial similarity between their fused feature map and the prototypes of all categories to obtain a similarity map; Step 6: Perform spatial uncertainty modeling on the fused feature map of the unlabeled samples to obtain an uncertainty map; Step 7: Fuse the similarity map and the uncertainty map to obtain a reliable evidence map, and determine the pseudo-labels of the unlabeled samples and their corresponding sample-level confidence levels based on the reliable evidence map.

2. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 1, characterized in that, Step 3 involves hierarchical fusion of feature maps from multiple levels extracted by each initial feature extraction network, including: Select shallow and deep feature maps from different layers in the network; For two adjacent feature maps, feature cross-learning is first performed to obtain cross-layer fused features; Attention learning is applied to the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features. The attention-enhanced features are input into an extended feedforward network to supplement the contextual structure information across windows, resulting in the final fused feature map.

3. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 2, characterized in that, The feature cross learning includes: Taking the feature maps of two adjacent layers as input, let the first layer be denoted as . The layer feature map is Fl, the first... The feature map of the layer is Fl+1. When the spatial resolution of the two layers of features is inconsistent, bilinear interpolation is first used to upsample the deep features to obtain the aligned feature map, so as to ensure that it is consistent with Fl in spatial dimension. The cross-learning stage includes two parallel branches: the basic feature extraction branch and the weight generation branch; the basic feature extraction branch uses a combination of BN, PConv[33] and DSC followed by ReLU activation to extract the basic representations of adjacent layers; the weight generation branch has a corresponding structure, but uses Sigmoid activation at the end to generate a weight map to characterize the importance of different spatial locations; specifically, it can be expressed as: in Represents ReLU. It represents Sigmoid; Subsequently, element-wise multiplication is used to achieve cross-modulation across layers: deep layer weights modulate shallow layer features to guide them to focus on semantically meaningful regions; shallow layer weights constrain deep layer features, supplementing their spatial structure and boundary information, ultimately yielding cross-layer fused features. : 。 4. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 2, characterized in that, Attention learning is applied to the cross-layer fusion features to enhance the semantic consistency of local discriminative regions, resulting in attention-enhanced features, as follows: Let the cross-layer fusion feature be First, divide it into several sizes. Non-overlapping local windows, each containing A spatial location; through sharing Convolution performs a linear mapping on the features within the window, generating query terms accordingly. Key items AND value term To enhance spatial awareness within the window, a learnable relative position offset matrix is ​​introduced. Used to describe the geometric relationships between different spatial locations; in the first Within a local window, the self-attention weights are calculated using the standard scaled dot product method, and feature aggregation is completed. This process is represented as follows: in , Indicates the feature channel dimension; This attention matrix explicitly models the similarity relationship between any two spatial locations within the window, thereby enhancing the semantic consistency of expression within local regions; The output features of all windows are rearranged and concatenated to restore the original spatial resolution, and then combined with layer normalization and residual connections to obtain attention-enhanced feature representations: Building upon this, to further supplement cross-window contextual structure information, features are fed into a dilated feedforward network for nonlinear mapping. This module expands the receptive field through multi-level dilated convolutions, enhancing local contextual modeling capabilities while maintaining a relatively constant parameter scale. The computation process is as follows: in Indicates the expansion rate The convolution operation; this paper adopts a two-stage dilated convolution structure, with dilation rates set to 1000 and 1000 respectively. , This is to enhance context modeling capabilities while avoiding a significant increase in the number of parameters; Finally, the structural compensation features and attention enhancement features are fused through a second residual connection to obtain the module output: 。 5. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 1, characterized in that, In step 4, based on the fused feature map and category labels of the labeled samples, an initial category prototype is constructed for each category, as follows: First, the network model is used to train the labeled data and extract features, which are then bound to the real labels and subsequently aggregated into prototypes pResNet and pViT. These prototypes represent the core features of each category, providing a benchmark for classifying unlabeled data. Suppose there are m categories, the prototype extracted by ResNet is pResNet,m, and the prototype extracted by ViT is pViT,m. For the m-th category, the formula for calculating its prototype feature pm is: Where Nm is the number of samples in the m-th category, and xi,m is the feature vector of the i-th sample in the m-th category; For unlabeled samples, the same model is used to extract features, and the Euclidean distance between the feature vector and the prototype p of each category is calculated to measure its similarity with the prototypes of different categories. Specifically, for the feature x of the unlabeled image extracted by the model and the prototype p, their Euclidean distance d(x,p) is calculated as a similarity measure, as shown in the formula: Subsequently, a distance-based voting prediction mechanism is employed. For a sample feature x, its distance d(x, pm) with all prototype features is first calculated, and the distance is converted into a similarity score, as shown in the formula: Where σ is a temperature parameter used to control the distribution of similarity scores; If there are prototype features with a distance less than the threshold τ, then voting is performed based on the categories corresponding to these prototype features that meet the conditions: Where ι is the set of prototype indices with a distance less than the threshold, li is the category corresponding to the i-th prototype, and Ⅰ( () is the indicator function; the final predicted category is the category with the most votes. If all distances are greater than the threshold, the category corresponding to the prototype feature with the smallest distance is selected as the prediction result; then the results of the two prototypes, pResNet and pViT, are weighted and fused. The weights are inversely proportional to the distance; the smaller the distance, the higher the weight, indicating that the image feature is closer to the prototype. The initial weights are calculated as follows: Subsequently, weight normalization is performed so that the sum of the weights of each model is 1, ensuring that the influence of each prototype is appropriately scaled; the final weight is the average of the normalized weights of all models; a unified decision is made by combining information from different models; the parameters of the weight layer are optimized through the training process to maximize classification accuracy; each category prototype corresponds to a weight value, and these weight values ​​are updated through the backpropagation algorithm to minimize classification error; the final label is determined as follows: 。 6. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 1, characterized in that, In step 5, for unlabeled samples, the spatial similarity between their fused feature map and the prototypes of all categories is calculated to obtain a similarity map, as follows: Given unlabeled samples The feature map is obtained after multi-scale feature fusion. For each category prototype The feature vectors are matched one by one with the feature vectors of all spatial locations in the feature map. By calculating the similarity relationship between the feature vectors and the prototype, a category-related similarity map is constructed. Unlike Euclidean distance-based metrics, cosine similarity is used to match feature vectors at each spatial location in the feature map with the class prototype. set up Represents the first feature in the feature graph The feature vector at each spatial location, then its relationship with the prototype The similarity can be represented as: in The second norm of a vector; The similarity map reflects the matching strength between each spatial location in the feature map and the prototype. Its high-response regions correspond to local discriminative features that are consistent with the semantics of the current category.

7. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 1, characterized in that, In step 6, spatial uncertainty modeling is performed on the fused feature map of the unlabeled samples to obtain an uncertainty map, as follows: Given an input CT image After passing through the backbone network and multi-scale fusion module, the feature map is obtained: On this feature map, this paper constructs a convolutional evidence head. It is used to model local evidence while preserving spatial structure information, and its output is expressed as: Among them, the channel dimension is expanded to This means that at each spatial location, four sets of evidence parameters are predicted for each category; Within the framework of evidence-based deep learning and subjective logic modeling, for spatial location and categories The parameters output by the evidence header are used to construct a pixel-by-pixel Dirichlet distribution; to ensure numerical stability, this paper adopts the following mapping method: in It is a very small constant used to avoid numerical instability; Subsequently, based on the principle of subjective logic modeling, the pixel-by-pixel Dirichlet concentration parameter is defined as: After obtaining the pixel-by-pixel Dirichlet parameters, this paper further aggregates the uncertainties of all categories to construct a spatial uncertainty mapping; specifically, spatial location... Uncertainty at a given point is defined as: The first term corresponds to arbitrary uncertainty, which mainly comes from imaging noise and local structural blur; the second term corresponds to cognitive uncertainty, which reflects the instability of the model's discrimination at this spatial location due to insufficient knowledge. This yields the spatial uncertainty diagram: The uncertainty graph is normalized to the [0,1] interval.

8. The dynamic prototype multi-model medical image classification method based on feature credibility assessment according to claim 1, characterized in that, In step 6, spatial uncertainty modeling is performed on the fused feature map of the unlabeled samples to obtain an uncertainty map, as follows: Given an input CT image After passing through the backbone network and multi-scale fusion module, the feature map is obtained: On this feature map, this paper constructs a convolutional evidence head. It is used to model local evidence while preserving spatial structure information, and its output is expressed as: Among them, the channel dimension is expanded to This means that at each spatial location, four sets of evidence parameters are predicted for each category; Within the framework of evidence-based deep learning and subjective logic modeling, for spatial location and categories The parameters output by the evidence header are used to construct a pixel-by-pixel Dirichlet distribution; to ensure numerical stability, this paper adopts the following mapping method: in It is a very small constant used to avoid numerical instability; Subsequently, based on the principle of subjective logic modeling, the pixel-by-pixel Dirichlet concentration parameter is defined as: After obtaining the pixel-by-pixel Dirichlet parameters, this paper further aggregates the uncertainties of all categories to construct a spatial uncertainty mapping; specifically, spatial location... Uncertainty at a given point is defined as: The first term corresponds to arbitrary uncertainty, which mainly comes from imaging noise and local structural blur; the second term corresponds to cognitive uncertainty, which reflects the instability of the model's discrimination at this spatial location due to insufficient knowledge. This yields the spatial uncertainty diagram: The uncertainty graph is normalized to the [0,1] interval.

Images