An information enhancement-based misdiagnosed lesion screening method

Abstract

This invention relates to an information-enhanced method for screening misdiagnosed lesions, which can alleviate the problem of high uncertainty in misdiagnosis of lesions with similar appearances or ambiguous boundaries. The method includes the following steps: First, a counterfactual gradient perturbation method is proposed, which improves the learning ability for complex lesion features by inversely optimizing the loss adversarial strategy; second, attention maps are generated through weakly supervised learning to locate lesion-distinguishing regions, and information entropy is used to obtain local information containing these regions; then, the attention map is inverted to extract inverse local information, and counterfactual gradient perturbation is applied to obtain additional task-related information; finally, by combining local and inverse local information, contrastive loss is used to jointly train and optimize the model. This method can improve the screening ability for misdiagnosed lesions under conditions of scarce medical data.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a method for screening misdiagnosed lesions based on information enhancement. Background Technology

[0002] In the field of medical image analysis, accurate lesion identification is crucial for the early diagnosis and treatment of diseases. However, in actual clinical applications, lesions exhibit complex and varied morphologies, especially those with similar appearances or blurred boundaries, making accurate differentiation challenging and prone to misdiagnosis with high uncertainty. Therefore, improving the identification accuracy of complex lesions has become a pressing problem in the field of medical image analysis. With the rapid development of technologies such as computer vision, deep learning-based automatic lesion identification methods have shown great potential. Due to the inherent heterogeneity of medical images, traditional deep learning methods such as ResNet are easily affected by noise and subtle changes in medical image analysis, leading to reduced diagnostic accuracy.

[0003] To overcome the aforementioned challenges, the goal of fine-grained visual recognition is to classify genus-level categories into base-level categories. A key step lies in extracting local features that effectively distinguish multiple base-level categories, which has been well-applied to medical lesion identification. However, obtaining fine-grained labels often requires significant manual labor and expertise from medical professionals, making it costly and time-consuming. How to perform fine-grained recognition using only coarse labels remains a critical challenge. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of traditional deep networks and fine-grained recognition techniques in screening for misdiagnosed lesions, and to propose a method for screening misdiagnosed lesions based on information augmentation. The technical solution is as follows: A method for screening misdiagnosed lesions based on information augmentation: (1) Counterfactual global adversarial information enhancement: By introducing gradient representation of basic features, the direction of the gradient is determined by the sign function sign; in each iteration, a gradient ascent step size is applied in the feature space to generate high information content features with counterfactual properties; in , (·)and These represent the adaptive step size, the positive gradient direction, and the enhanced feature set, respectively, used to control the magnitude of the perturbation. Then, a contrastive learning framework was introduced to minimize the difference representation learning; by forming sample pairs with the base features and counterfactual features of the same sample, the Euclidean distance d was used to measure the output similarity between the sample pairs, and the model was optimized by minimizing d during training. in , ,and Let represent the contrast difference loss, the basic feature representation of the i-th sample, and its counterfactual enhancement feature, respectively. (2) Adaptive weighted representation of local connectivity regions based on information entropy: This method uses weak supervision to locate regions of interest and binarizes the attention map to extract salient regions. Specifically, for each sample, an initial threshold is set, and regions in the attention map that are higher than the threshold are marked as salient regions, and a binarized attention map is generated. Subsequently, the binarization result is analyzed using a connected component labeling algorithm to obtain the number of labeled regions. If no effective region is detected and the current threshold is higher than the preset minimum threshold, the threshold is gradually reduced and the above process is repeated until a salient region is detected or the minimum threshold is reached. In addition, a constraint strategy based on cosine similarity is introduced, which treats global features as prototype templates and uses them as reference standards for global semantics to match local feature regions one by one; by calculating the cosine similarity between local features and global prototypes, the semantic relevance of each local region is evaluated. Subsequently, information entropy is introduced to evaluate the representational power and potential learning value of these regions. During training, the loss values ​​of these local regions are calculated; larger loss values ​​generally indicate that the model has not fully understood the feature information of that region. The information content of each local region is quantified and subjected to softmax normalization. The information content score of each region is then converted into a weighting factor to measure the relative contribution of different regions to the overall information domain of the sample. (3) Local external information enhancement representation based on gradient ascent. During training, the gradient information of the model is back-guided, and appropriate perturbation signals are applied to regions that are irrelevant to global features and are judged as unimportant by the model to obtain local external representation enhancement information; The representation of a single lesion is trained through the coordinated interaction of global representation, local representation, and external representation.

[0005] The beneficial effects of this invention are: (1) An information-enhanced method for screening misdiagnosed lesions is proposed to improve the model’s ability to analyze lesion boundary features, alleviate the problem of high uncertainty in some lesions with similar appearance or fuzzy boundaries, improve clinical decision-making ability, and reduce non-invasive biopsy procedures. (2) Under the condition of scarce medical samples, traditional data augmentation will introduce uncontrollable noise interference and create information bottlenecks. The method provided by this invention can provide high-information feature information and provide effective support for training data. Attached Figure Description

[0006] Figure 1 This is a schematic diagram illustrating information enhancement according to the present invention; Figure 2 This is a roadmap for the counterfactual global adversarial information enhancement technology of the present invention; Figure 3 This is a roadmap for the adaptive weighted representation technology of local connectivity regions based on information entropy in this invention; Detailed Implementation

[0007] To make the technical solution of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. The present invention is implemented in specific steps: (1) Counterfactual global adversarial information enhancement; Define the features of the original samples as basic features, and use gradient information to increase the loss by adjusting the basic features themselves instead of adjusting the model parameters, that is, changing the feature values ​​in the direction of maximum loss through positive gradient. By introducing the gradient representation of the basic features, the direction of the gradient is determined by the sign function sign; in , (·)and These represent the adaptive step size, the positive gradient direction, and the enhanced feature set, respectively, used to control the magnitude of the perturbation. Subsequently, the feature space is optimized along the ascending direction of the gradient to maximize the information gain of the objective function. In each iteration, a gradient ascending step size is applied to the feature space to generate highly informative features with counterfactual properties. Then, a contrastive learning framework is introduced to minimize the difference representation learning. By forming sample pairs with the base features and counterfactual features of the same sample, the Euclidean distance d is used to measure the output similarity between sample pairs, and the model is optimized by minimizing d during training; in , ,and Let represent the contrast difference loss, the basic feature representation of the i-th sample, and its counterfactual enhancement feature, respectively. (2) Adaptive weighted representation of local connectivity regions based on information entropy; This invention utilizes weak supervision to locate regions of interest and binarizes the attention map to extract salient regions; Specifically, for each sample, an initial threshold is set, and regions in the attention map that are above the threshold are marked as salient regions, and a binarized attention map is generated. Then, the binarization result is analyzed using a connected component labeling algorithm to obtain the number of labeled regions. If no effective region is detected and the current threshold is higher than the preset minimum threshold, the threshold is gradually lowered and the above process is repeated until a salient region is detected or the minimum threshold is reached. In addition, a constraint strategy based on cosine similarity is introduced, which treats global features as prototype templates and uses them as reference standards for global semantics to match local feature regions one by one. The semantic relevance of each local region is evaluated by calculating the cosine similarity between local features and the global prototype; regions with low similarity to the prototype features are considered potential noise regions and are removed. Then, information entropy is introduced to evaluate the representational ability and potential learning value of these regions. During training, the loss value of these local regions is calculated. A larger loss value usually indicates that the model has not fully understood the feature information of the region. The information content of each local region is quantified and subjected to softmax normalization. The information content score of each region is transformed into a weighting factor to measure the relative contribution of different regions in the overall information domain of the sample. in, It is a region The prediction confidence of class c is then normalized using softmax to form a weight distribution, which is used for weighting local features. (3) Local external information enhancement representation based on gradient ascent. During training, the gradient information of the model is back-guided, and appropriate perturbation signals are applied to regions that are irrelevant to global features and are judged as unimportant by the model to obtain local external representation enhancement information; in, For spatial attention matrix, This represents the binarized region after inverting the spatial attention matrix; (4) The feature information of a single lesion is trained through the synergy of global representation, local representation and inverse local representation: Among them, B, (·) and CE represent the batch size, confidence function, and cross-entropy function, respectively.

[0008] Through the above implementation methods, the present invention can achieve fine-grained information enhancement screening of misdiagnosed lesions without the need for detailed manual annotation, saving annotation costs while improving screening accuracy. It is applicable to upstream tasks of medical lesion classification and provides support for various downstream intelligent algorithms.

Claims

1. A method for screening misdiagnosed lesions based on information enhancement, characterized in that, Includes the following steps: (1) Counterfactual global adversarial information enhancement: Define the feature map extracted by the ResNet34 convolutional layer as the base feature, introduce the gradient representation of the base feature, and use the sign function sign to determine the direction of the gradient; in each iteration, apply the gradient ascent step size in the feature space to generate high information content features with counterfactual properties. ； ； in , (·)and These represent the adaptive step size, the positive gradient direction, and the enhanced feature set, respectively, used to control the magnitude of the perturbation. Then, a contrastive learning module is designed to learn the representation of high-information features and base features by minimizing the difference. By forming sample pairs with base features and counterfactual features of the same sample, the Euclidean distance d is used to measure the output similarity between sample pairs, and the model is optimized by minimizing d during training. ； in , ,and Let represent the contrast difference loss, the basic feature representation of the i-th sample, and its counterfactual enhancement feature, respectively. (2) Adaptive weighted representation of local connectivity regions based on information entropy: This method uses weak supervision to locate regions of interest and binarizes the attention map to extract salient regions. Specifically, for each sample, an initial threshold is set, and regions in the attention map that are higher than the threshold are marked as salient regions, and a binarized attention map is generated. Subsequently, the binarization result is analyzed using a connected component labeling algorithm to obtain the number of labeled regions. If no valid region is detected and the current threshold is higher than the preset minimum threshold, the threshold is gradually lowered and the above process is repeated until a significant region is detected or the minimum threshold is reached. In addition, this method introduces a constraint strategy based on cosine similarity, which treats global features as prototype templates and uses them as reference standards for global semantics to match local feature regions one by one; by calculating the cosine similarity between local features and global prototypes, the semantic relevance of each local region is evaluated. ； Subsequently, information entropy is introduced to evaluate the representational power and potential learning value of these regions. During training, the loss values ​​of these local regions are calculated. Larger loss values ​​usually indicate that the model has not fully understood the feature information of that region. The information content of each local region is quantified and subjected to softmax normalization. The information content score of each region is converted into a weighting factor to measure the relative contribution of different regions to the overall information domain of the sample. ； in, It is a region The prediction confidence of class c is then normalized using softmax to form a weight distribution, which is used for weighting local features. ； (3) Local external information enhancement representation based on gradient ascent: During the training process, the gradient information of the model is guided in reverse, the spatial attention matrix is ​​inverted, and regions that are irrelevant to global features and are judged as unimportant by the model are obtained. Appropriate perturbation signals are applied to obtain local external representation enhancement information. ； in, For spatial attention matrix, This represents the binarized region after inverting the spatial attention matrix; (4) The feature information of a single lesion is trained through global representation, local representation and inverse local representation: ； ； ； Among them, B, (·) and CE represent the batch size, confidence function, and cross-entropy function, respectively.

Images