Prototype and confidence driven medical image passive adaptive classification system and method

Abstract

This invention discloses a prototype- and confidence-driven passive adaptive classification system and method for medical images, relating to the field of medical image processing. It includes a data preparation module, a feature extraction and source domain prototype generation module, a joint prototype-guided pseudo-label generation module, a perceptual confidence tracking module, and a contrastive learning module. This invention employs the aforementioned prototype- and confidence-driven passive adaptive classification system and method for medical images to address the problem of medical image distribution differences caused by different medical institutions or equipment. It achieves efficient unsupervised domain adaptive training and progressive improvement from pseudo-labels to feature optimization without accessing the original source domain data, utilizing source domain information and unlabeled target domain data. It achieves significantly better performance than existing methods on multiple cross-domain datasets.

Description

Technical Field

[0001] This invention relates to the field of medical image processing, and in particular to a prototype and confidence-driven passive adaptive classification system and method for medical images. Background Technology

[0002] Source-Free Unsupervised Domain Adaptation (SFUDA) is a technique that adapts a target domain to its original data without access to the source domain. It utilizes only a pre-trained source model and unlabeled target domain data. The goal is to reduce the distributional differences between the source and target domain data without accessing the original source data, thereby mitigating model performance degradation caused by data distribution shifts. This technique enables cross-domain transfer learning while ensuring data privacy and security.

[0003] In the field of medical image classification, self-adjusting methods are the main technical approach of passive unsupervised domain adaptation (SFUDA). Its core idea is to rely on target domain data and iteratively optimize the model through pseudo-label generation, selection, and self-training to gradually adapt the model to the target domain distribution. Weighted K-means clustering is often used to construct class centroids to improve the stability of pseudo-labels and training reliability. In existing methods, researchers have improved the quality and feature representation capabilities of pseudo-labels through various strategies. On the one hand, neighborhood relationship modeling enhances the consistency of local features, thereby reducing label noise; on the other hand, mining domain-invariant features can generate robust representations, reducing the uncertainty of target domain training. Simultaneously, combining consistency constraints and knowledge distillation can enhance the consistency of cross-domain feature distributions, while improvements to information maximization loss and multi-label generation strategies improve the model's robustness to low-confidence samples. Furthermore, the combination of feature enhancement and spectral clustering can optimize the prediction space partitioning, making pseudo-label allocation more reasonable. Through these improvements, self-adjusting methods have significantly improved clustering quality, pseudo-label reliability, and cross-domain consistency, providing a robust technical foundation for the application of SFUDA in medical image classification.

[0004] The aforementioned passive, unsupervised domain adaptation methods generate initial pseudo-labels using a pre-trained source domain model, and then further update the pseudo-labels by calculating the target domain category prototype through K-means clustering, focusing on utilizing the structural information inherent in the target domain data itself. This self-regulating mechanism can gradually optimize the feature representation of the target domain, improving the model's adaptability. However, these methods still face challenges from pseudo-label noise and instability in practical applications, potentially leading to misleading optimization directions. Furthermore, some methods neglect the rich semantic information and domain-invariant features inherent in the source domain model, failing to fully utilize source domain knowledge to assist target domain adaptation, thus limiting the model's performance in complex cross-domain scenarios.

[0005] In summary, while existing passive unsupervised domain adaptation methods have made some progress in target domain feature alignment and self-adjustment mechanisms, they still have significant shortcomings in the deep mining and effective utilization of source domain knowledge. This limits the model's ability to cope with complex cross-domain distribution differences, and traditional methods struggle to achieve stable and well-generalized adaptation results. Most methods focus on achieving alignment by matching target domain features with the weight space of the source domain classifier, but neglect the rich semantic information and structural features in the source model, resulting in insufficient response to diverse and complex distribution changes in the target domain. Furthermore, strategies relying on fixed parameters or static assumptions lack flexibility and struggle to adapt to constantly changing feature representations in the target domain, further limiting the model's generalization ability. Summary of the Invention

[0006] The purpose of this invention is to provide a prototype and confidence-driven passive adaptive classification system and method for medical images, thereby solving the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a prototype and confidence-driven passive adaptive classification system for medical images, comprising:

[0008] The data preparation module is used to collect, segment, and label medical images in the source and target domains, unify data features, and improve the performance of subsequent model training and passive adaptive classification.

[0009] The feature extraction and source domain prototype generation module is used to extract feature representations of medical images in the source and target domains, and store source domain model weights and category prototypes as the basis for fine-tuning the target domain.

[0010] The joint prototype-guided pseudo-label generation module is used to perform weighted fusion of the source domain category prototype and the target domain category prototype to generate a joint category prototype and redistribute pseudo-labels, thereby improving the feature alignment and classification performance of subsequent iterations.

[0011] The perceptual confidence tracking module is used to calculate the current pseudo-label confidence and generate the final confidence. High-confidence samples are selected for subsequent feature alignment and model optimization.

[0012] The contrastive learning module is used to construct positive and negative sample pairs, and optimize the feature representation of the target domain through contrastive learning to improve the robustness and discriminative ability of the model.

[0013] Preferably, the data preparation module includes:

[0014] The image collection module is used to collect medical images from multiple medical institutions or multiple imaging devices, and divide the medical images into source domain datasets and target domain datasets. The source domain datasets have classification labels, while the target domain datasets do not.

[0015] The image annotation module is used to annotate the lesion types in source domain medical images;

[0016] The data preprocessing module is used to preprocess medical images in the source and target domains.

[0017] Preferably, the feature extraction and source domain prototype generation module includes:

[0018] Convolutional layers are used to encode features of medical images in the source and target domains, extract high-level semantic features, and store the parameters of the source domain convolutional layers.

[0019] The bottleneck layer is used to perform dimensionality reduction and normalization on the features extracted by the convolutional layer, and to store the bottleneck layer parameters.

[0020] Preferably, the feature extraction uses ResNet50 pre-trained on ImageNet as the backbone network for feature extraction.

[0021] Preferably, the joint prototype-guided pseudo-tag generation module includes:

[0022] The target domain prototype calculation module is used to dynamically calculate the category prototype of the target domain samples by averaging the features of the target domain samples based on the target domain pseudo-labels generated in the current iteration.

[0023] The adaptive prototype fusion module is used to adaptively determine the weighted fusion ratio based on the cosine similarity between the source domain and target domain category prototypes to generate joint category prototypes.

[0024] The pseudo-label generation module is used to calculate the cosine distance between the target domain sample features and the joint prototype based on the fused joint category prototype, and assign the sample to the nearest prototype.

[0025] Preferably, the perception confidence tracking module includes:

[0026] The pseudo-label confidence calculation module is used to calculate the current pseudo-label confidence for each sample based on the pseudo-labels of the K nearest neighbor samples of the target domain sample through weighted voting.

[0027] The confidence fusion module is used to dynamically weight and fuse the current pseudo-label confidence with the historical confidence to generate a stable final confidence.

[0028] The high-confidence sample screening module is used to screen high-confidence target domain samples by category according to the final confidence level after fusion and a preset screening ratio.

[0029] Preferably, the confidence fusion module uses the exponential moving average (EMA) method for weighted fusion, and uses the EMA fusion result as the historical confidence level for the next iteration;

[0030] The fusion factor of EMA is adaptively adjusted according to the number of iterations of model training.

[0031] Preferably, the comparison learning module includes:

[0032] The positive and negative sample pair construction module is used to construct positive and negative sample pairs based on the high-confidence target domain samples selected by the perceptual confidence tracking module.

[0033] The feature comparison optimization module is used to optimize the feature representation of the target domain by maximizing the feature similarity of positive sample pairs and minimizing the feature similarity of negative sample pairs.

[0034] Preferably, positive sample pairs consist of samples with the same pseudo-label category, and negative sample pairs consist of samples with different pseudo-label categories.

[0035] A method for a prototype and confidence-driven passive adaptive classification system for medical images includes the following steps:

[0036] S1. Set up the hardware and software environment required for model operation, and construct a passive, unsupervised, adaptive joint prototype and confidence tracking system for medical image classification;

[0037] S2. Acquire medical image data to form source and target domains, and perform bias field correction, grayscale normalization, voxel resampling and random cropping operations on the source and target domain images.

[0038] S3. Extract stable source domain category prototypes using the source domain pre-trained model, and dynamically calculate the target domain category prototypes based on the current pseudo-labels of the target domain.

[0039] S4. Based on the similarity between the source domain and target domain category prototypes, dynamically adjust the fusion ratio of the source domain prototype and the target domain prototype to generate a joint prototype that integrates prior knowledge and dynamic features.

[0040] S5. Based on the joint prototype, the target domain pseudo-labels are reassigned to the target domain samples, and the current pseudo-label confidence and historical pseudo-label confidence are fused through the perceptual confidence tracking module to obtain a more stable final confidence of the target domain samples and screen high-confidence samples.

[0041] S6. Construct positive and negative sample pairs using high-confidence samples, and optimize the target domain feature representation through the contrastive learning module to improve the model's discriminative ability and robustness;

[0042] S7. Based on the updated pseudo-labels and joint prototypes, iteratively execute steps S5 to S6 to continuously optimize the model parameters until the model converges or reaches the preset performance index.

[0043] S8. Evaluate the performance of the trained model on the target domain samples, calculate the area under the curve (AUC) and classification accuracy (ACC) of the target domain samples, and compare and verify with existing methods.

[0044] Therefore, the present invention employs the above-mentioned prototype and confidence-driven passive adaptive classification system and method for medical images, which has the following beneficial effects:

[0045] (1) An innovative, unsupervised, passive-free medical image classification framework that does not require access to the original source domain data is proposed. Without accessing the source data, the framework combines source domain prior information and confidence tracking mechanism, and improves the model's cross-domain adaptation ability through a contrastive learning strategy. It is suitable for clinical scenarios where domain bias is common.

[0046] (2) This invention makes fuller use of source domain structure information, combines the dynamic category prototype of the current target domain with the stable source domain category prototype, realizes the combination of source domain prior knowledge and target domain dynamic features, significantly improves the accuracy of pseudo-labels and the robustness of sample selection, and improves the classification accuracy and stability of the model in the target domain.

[0047] (3) By using the perceptual confidence tracking module to introduce historical confidence information, the current confidence is weighted and smoothed to suppress random fluctuations in a single prediction, maintain the continuity and stability of confidence changes, select high-confidence target domain samples, construct positive and negative sample pairs, and maximize the similarity of positive samples and minimize the similarity of negative samples through comparative learning, thereby strengthening the discriminativeness and domain invariance of features and continuously improving the generalization performance of the model in multiple rounds of iterative optimization.

[0048] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0049] Figure 1 This is a two-stage optimization structure diagram of a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention.

[0050] Figure 2 This is a data preparation module diagram for a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention.

[0051] Figure 3 This is an overall framework diagram of a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention.

[0052] Figure 4 This is a diagram of the feature extraction and source domain prototype generation module in an embodiment of a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention.

[0053] Figure 5 This is a diagram of a joint prototype-guided pseudo-label generation module in an embodiment of a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention.

[0054] Figure 6 This is a diagram of the perceptual confidence tracking module in an embodiment of a prototype and confidence-driven passive adaptive classification system and method for medical images according to the present invention. Detailed Implementation

[0055] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] Example

[0058] Please see Figures 1-6 This invention provides a prototype and confidence-driven passive adaptive classification system for medical images, comprising:

[0059] The data preparation module is used to collect, segment, and label medical images in the source and target domains, unify data features, and improve the performance of subsequent model training and passive adaptive classification.

[0060] The feature extraction and source domain prototype generation module is used to extract feature representations of medical images in the source and target domains, and store source domain model weights and category prototypes as the basis for fine-tuning the target domain.

[0061] The joint prototype-guided pseudo-label generation module is used to perform weighted fusion of the source domain category prototype and the target domain category prototype to generate a joint category prototype and redistribute pseudo-labels, thereby improving the feature alignment and classification performance of subsequent iterations.

[0062] The perceptual confidence tracking module is used to calculate the current pseudo-label confidence and generate the final confidence. High-confidence samples are selected for subsequent feature alignment and model optimization.

[0063] The contrastive learning module is used to construct positive and negative sample pairs, and optimize the feature representation of the target domain through contrastive learning to improve the robustness and discriminative ability of the model.

[0064] This invention also provides a method for a prototype and confidence-driven passive adaptive classification system for medical images, the specific steps of which are as follows.

[0065] The implementation process of this invention is divided into two stages, as shown in the appendix. Figure 1 The diagram illustrates the process from feature extraction to joint prototype-guided pseudo-label generation, then to perceptual-driven confidence tracking, and finally to the output result. (See attached image.) Figure 3 This is an overall framework diagram of the method of the present invention, which further illustrates the relationship between the modules, the data flow, and the internal details of data processing.

[0066] S1, Image Collection.

[0067] like Figure 2The image collection module shown constructs two multicenter datasets: one for adrenal tumor diagnosis and the other for intrahepatic cholangiocarcinoma diagnosis, sourced from hospitals A, B, C, and D, respectively. The data are as follows: 1938 CT slices from 101 patients with adrenal adenomas and 92 patients with adrenal metastases obtained by GE Discovery HD750 64-slice CT scan at Hospital A, and 1171 slices from 116 patients with intrahepatic cholangiocarcinoma metastases and 63 patients with non-metastatic intrahepatic cholangiocarcinoma obtained by Philips Achieva 1.5T MRI scan at Hospital B; 1052 CT slices from 31 patients with adrenal adenomas and 40 patients with adrenal metastases obtained by Philips 256-slice Brilliance iCT scan at Hospital B; 1052 CT slices from 31 patients with adrenal adenomas and 40 patients with adrenal metastases obtained by Siemens Prisma 3.0T MRI scan at Hospital C; and 1171 slices from 8 patients with intrahepatic cholangiocarcinoma metastases and 7 patients with non-metastatic intrahepatic cholangiocarcinoma obtained by Siemens Amira 1.5T MRI scan at Hospital D. In the adrenal tumor diagnosis task, multi-center datasets from different hospitals served as both the source and target domains to validate the method's adaptability under various domain shift conditions. In the intrahepatic cholangiocarcinoma diagnosis task, Hospital A, with its abundant data, served as the source domain, while Hospitals C and D, with their smaller sample sizes, served as the target domains, simulating a real-world clinical scenario where the model is being extended from high-level medical centers to primary care hospitals.

[0068] S2, Image annotation.

[0069] like Figure 2 The image annotation module shown employs three experienced radiologists to delineate lesion regions in an image, obtaining pixel-level labels. The annotation results from the three physicians are integrated using a majority voting method, and a consistency analysis is performed on the annotated regions to verify the accuracy and reliability of the annotations, providing high-quality supervision for the training of the domain adaptive model.

[0070] S3, Data Preprocessing.

[0071] like Figure 2The data preprocessing module shown performs bias field correction, normalization, resampling, and random cropping preprocessing operations on the source and target domain images. Bias field correction makes the gray values ​​of the same tissue more uniform in the image, reducing artifacts and brightness variations; normalization scales the pixel values ​​to the [0, 1] range and crops them into 512×512 images; resampling adjusts the spatial resolution of the image to ensure that images obtained from different devices or scanning protocols have a uniform voxel size, guaranteeing consistency between different data sources; random cropping involves cropping the image at random locations during training, increasing sample diversity and preventing model overfitting. Data preprocessing unifies data features and provides reliable input for subsequent model training and passive adaptive classification.

[0072] S4. Feature extraction and source domain prototype generation.

[0073] like Figure 4 The feature extraction and source domain prototype generation module shown employs a ResNet50 pre-trained on ImageNet as the backbone network for feature extraction, used to extract deep semantic features from medical images. The convolutional layers are responsible for feature encoding of the source and target domain images, extracting high-level features rich in semantic information, and storing the source domain convolutional layer parameters, providing a solid foundation for fine-tuning the target domain. The bottleneck layer performs dimensionality reduction and normalization on the features extracted by the convolutional layers, further optimizing feature representation and storing the corresponding bottleneck layer parameters to support effective fine-tuning and adaptation of the target domain features.

[0074] S5A, joint prototype guides pseudo-tag generation.

[0075] like Figure 5 The joint prototype-guided pseudo-label generation module shown uses the same feature extractor to extract features from both the target domain samples and the source domain samples. First, it calculates the similarity between the target sample and the source domain category prototype to obtain initial pseudo-labels. Then, it dynamically calculates the target domain category prototype based on the pseudo-labels. Subsequently, it fuses the source and target domain category prototypes to reflect the changes and distribution of target domain features in real time, improving the reliability of the fusion.

[0076] (1) The target domain prototype calculation module calculates the feature representation of the target domain sample and performs a weighted average according to its pseudo-label category to obtain the feature center of each category in the target domain. It can provide the latest and most representative category feature information of the target domain and provide dynamic target domain feature information for prototype fusion.

[0077] (2) The adaptive prototype fusion module dynamically calculates the fusion factor α based on the cosine similarity between the source domain and target domain category prototypes. This factor is used to adjust the weight ratio of the source domain category prototypes and the target domain category prototypes in the joint prototype. Specifically, the cosine similarity of each category between the source domain and the target domain is first calculated. Then, the similarity is converted into a fusion weight α through a linear mapping. This ensures that when the similarity is higher, the weight of the joint prototype depends more on the target domain prototype; when the similarity is lower, it depends more on the source domain prototype, thus achieving adaptive fusion of stable priors and dynamic features. The fusion factor α maps the cosine similarity (within the range of [-1, 1]) to the interval [0, 1] through a linear mapping function. When the similarity is 0.8, α is 0.5; when the similarity is 1.0, α is 1.0, thus achieving adaptive and effective fusion of the source domain and target domain prototypes and improving the adaptive performance.

[0078] (3) The pseudo-label generation module calculates the cosine similarity between the features of each target domain sample and the prototypes of each category based on the fused joint category prototypes, and assigns the sample to the category with the highest similarity as a new pseudo-label. By continuously using the updated joint prototypes to redistribute pseudo-labels, the module effectively improves the alignment accuracy and classification performance of the target domain features, and promotes the iterative optimization and generalization ability of the model in the target domain.

[0079] S5B, Perception Confidence Tracking Module.

[0080] like Figure 6 The perceptual confidence tracking module shown uses a dynamic evaluation and exponential averaging mechanism to fuse the current and historical confidence of pseudo-labels of target domain samples, and selects samples with high confidence to improve the generalization performance and stability of the model in the target domain.

[0081] (1) Calculation of pseudo-label confidence: By counting the number of samples with the same pseudo-label as the sample among the K nearest neighbor samples, the current confidence of the sample is obtained, which reflects its classification certainty. In the pseudo-label confidence calculation module, K=4 is selected, that is, for each target sample, the consistency of its pseudo-labels among its four nearest neighbor samples in the feature space is analyzed, and the reliability and confidence of the pseudo-label of the sample are evaluated by weighted voting, so as to dynamically reflect the uncertainty of sample classification in the target domain.

[0082] (2) The pseudo-label confidence score obtained in the current iteration is weighted and fused with the confidence scores in the historical iterations. The exponential moving average (EMA) mechanism is used to reduce the random fluctuations of a single prediction and ensure the continuous stability of the confidence score. The smoothing coefficient is set to 0.9. This means that when calculating the final confidence score, the historical confidence score and the current confidence score each account for 50% of the weight, thereby effectively smoothing the fluctuations of a single prediction, reducing the impact of fluctuations in the confidence score of a single prediction, and improving the stability and continuity of the confidence score.

[0083] (3) Sort all target domain samples in descending order according to the final confidence score after fusion, and select the samples with the highest confidence scores in the top 60% by category. These samples are considered to have relatively reliable pseudo-labels, which are used for subsequent feature alignment and model optimization processes to enhance the performance and robustness of the model in the target domain and promote the accuracy of model iterative training.

[0084] S6. Optimize the feature representation of the target domain.

[0085] The contrastive learning module constructs positive and negative sample pairs based on high-confidence target domain samples selected by the perceptual confidence tracking module. Positive samples consist of samples with the same pseudo-label category, while negative samples consist of samples with different pseudo-label categories. By maximizing the feature similarity of positive sample pairs and minimizing the feature similarity of negative sample pairs, the robustness and discriminative ability of the target domain feature representation are optimized.

[0086] S7. Iterative optimization.

[0087] In each iteration, the model updates the pseudo-labels based on the joint prototype and high-confidence samples, and further optimizes the model parameters through contrastive learning. This process is repeated until the model's performance in the target domain converges or the preset number of iterations is reached.

[0088] S8. Model Evaluation.

[0089] The model's performance was evaluated on clinical data in the target domain. The fine-tuned model was applied to the diagnosis of adrenal tumors and intrahepatic cholangiocarcinoma. The classification performance of the model was comprehensively evaluated by calculating key indicators such as classification accuracy (ACC), precision (PRE), sensitivity (SEN), specificity (SPE), and area under the curve (AUC). A comparative analysis was also conducted with traditional unsupervised domain adaptation methods that require access to source data and passive unsupervised domain adaptation methods.

[0090] Table 1 Comparison of classification and prediction performance between adrenal adenoma and metastatic tumor

[0091]

[0092] Table 1 compares the performance of this invention in classifying and predicting adrenal adenomas and metastatic tumors. Experimental results show that this invention demonstrates superior performance in both data migration directions without accessing the source data. Compared to other methods that also do not access the source data, this invention achieves significant superiority in key indicators such as accuracy, sensitivity, specificity, and AUC, realizing high-precision cross-center tumor diagnosis and possessing outstanding advanced technology and application potential.

[0093] Table 2 Comparison of classification and prediction performance between metastatic and non-metastatic intrahepatic cholangiocarcinoma.

[0094]

[0095] Table 2 compares the performance of this invention in classifying and predicting metastatic and non-metastatic intrahepatic cholangiocarcinoma, verifying its universality and effectiveness across different clinical diagnostic tasks. Experimental results show that this invention also demonstrates superior performance in the diagnosis of intrahepatic cholangiocarcinoma. In a real-world clinical scenario simulating the transfer of technology from high-level medical centers to primary hospitals, this invention, without accessing the source data, significantly outperformed other similar methods in core metrics such as accuracy, sensitivity, specificity, and AUC in both the "Hospital A → Hospital C" and "Hospital A → Hospital D" migration directions. This, coupled with the results from the adrenal tumor diagnosis task, further demonstrates that this invention not only performs exceptionally well on specific tasks but also possesses strong generalization capabilities and application potential across diseases, centers, and modalities.

[0096] Therefore, this invention employs the aforementioned prototype- and confidence-driven passive adaptive classification system and method for medical images. Addressing the issues of high noise and instability in pseudo-labels during passive domain adaptation, this invention innovatively introduces a joint prototype-guided pseudo-label generation and perceptual confidence tracking module. This mechanism significantly improves the reliability of pseudo-labels by adaptively fusing source and target domain prototypes and dynamically selecting high-confidence samples, thereby effectively enhancing the model's learning and generalization capabilities.

[0097] This invention not only demonstrates outstanding performance in the diagnosis of adrenal tumors but has also been validated in the diagnosis of intrahepatic cholangiocarcinoma. This proves that the invention is not limited to specific diseases or data types and has the ability to be generalized across different devices, centers, and even different disease types, laying the foundation for future multi-tasking and interdisciplinary applications.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A prototype and confidence-driven passive adaptive classification system for medical images, characterized in that, include: The data preparation module is used to collect, segment, and label medical images in the source and target domains, unify data features, and improve the performance of subsequent model training and passive adaptive classification. The feature extraction and source domain prototype generation module is used to extract feature representations of medical images in the source and target domains, and store source domain model weights and category prototypes as the basis for fine-tuning the target domain. The joint prototype-guided pseudo-label generation module is used to perform weighted fusion of the source domain category prototype and the target domain category prototype to generate a joint category prototype and redistribute pseudo-labels, thereby improving the feature alignment and classification performance of subsequent iterations. The perceptual confidence tracking module is used to calculate the current pseudo-label confidence and generate the final confidence. High-confidence samples are selected for subsequent feature alignment and model optimization. The contrastive learning module is used to construct positive and negative sample pairs, and optimize the feature representation of the target domain through contrastive learning to improve the robustness and discriminative ability of the model.

2. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 1, characterized in that, The data preparation module includes: The image collection module is used to collect medical images from multiple medical institutions or multiple imaging devices, and divide the medical images into source domain datasets and target domain datasets. The source domain datasets have classification labels, while the target domain datasets do not. The image annotation module is used to annotate the lesion types in source domain medical images; The data preprocessing module is used to preprocess medical images in the source and target domains.

3. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 1, characterized in that, The feature extraction and source domain prototype generation module includes: Convolutional layers are used to encode features of medical images in the source and target domains, extract high-level semantic features, and store the parameters of the source domain convolutional layers. The bottleneck layer is used to perform dimensionality reduction and normalization on the features extracted by the convolutional layer, and to store the bottleneck layer parameters.

4. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 3, characterized in that: Feature extraction uses ResNet50 pre-trained on ImageNet as the backbone network.

5. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 1, characterized in that, The joint prototype-guided pseudo-tag generation module includes: The target domain prototype calculation module is used to dynamically calculate the category prototype of the target domain samples by averaging the features of the target domain samples based on the target domain pseudo-labels generated in the current iteration. The adaptive prototype fusion module is used to adaptively determine the weighted fusion ratio based on the cosine similarity between the source domain and target domain category prototypes to generate joint category prototypes. The pseudo-label generation module is used to calculate the cosine distance between the target domain sample features and the joint prototype based on the fused joint category prototype, and assign the sample to the nearest prototype.

6. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 1, characterized in that, The perception confidence tracking module includes: The pseudo-label confidence calculation module is used to calculate the current pseudo-label confidence for each sample based on the pseudo-labels of the K nearest neighbor samples of the target domain sample through weighted voting. The confidence fusion module is used to dynamically weight and fuse the current pseudo-label confidence with the historical confidence to generate a stable final confidence. The high-confidence sample screening module is used to screen high-confidence target domain samples by category according to the final confidence level after fusion and a preset screening ratio.

7. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 6, characterized in that: The confidence fusion module uses the exponential moving average (EMA) method for weighted fusion, and uses the EMA fusion result as the historical confidence level for the next iteration. The fusion factor of EMA is adaptively adjusted according to the number of iterations of model training.

8. The prototype and confidence-driven passive adaptive classification system for medical images according to claim 1, characterized in that, The contrastive learning module includes: The positive and negative sample pair construction module is used to construct positive and negative sample pairs based on the high-confidence target domain samples selected by the perceptual confidence tracking module. The feature comparison optimization module is used to optimize the feature representation of the target domain by maximizing the feature similarity of positive sample pairs and minimizing the feature similarity of negative sample pairs.

9. A prototype and confidence-driven passive adaptive classification system for medical images according to claim 8, characterized in that: Positive sample pairs consist of samples with the same pseudo-label category, while negative sample pairs consist of samples with different pseudo-label categories.

10. A method for a prototype and confidence-driven passive adaptive classification system for medical images according to any one of claims 1-9, characterized in that, Includes the following steps: S1. Set up the hardware and software environment required for model operation, and construct a passive, unsupervised, adaptive joint prototype and confidence tracking system for medical image classification; S2. Acquire medical image data to form source and target domains, and perform bias field correction, grayscale normalization, voxel resampling and random cropping operations on the source and target domain images. S3. Extract stable source domain category prototypes using the source domain pre-trained model, and dynamically calculate the target domain category prototypes based on the current pseudo-labels of the target domain. S4. Based on the similarity between the source domain and target domain category prototypes, dynamically adjust the fusion ratio of the source domain prototype and the target domain prototype to generate a joint prototype that integrates prior knowledge and dynamic features. S5. Based on the joint prototype, the target domain pseudo-labels are reassigned to the target domain samples, and the current pseudo-label confidence and historical pseudo-label confidence are fused through the perceptual confidence tracking module to obtain a more stable final confidence of the target domain samples and screen high-confidence samples. S6. Construct positive and negative sample pairs using high-confidence samples, and optimize the target domain feature representation through the contrastive learning module to improve the model's discriminative ability and robustness; S7. Based on the updated pseudo-labels and joint prototypes, iteratively execute steps S5 to S6 to continuously optimize the model parameters until the model converges or reaches the preset performance index. S8. Evaluate the performance of the trained model on the target domain samples, calculate the area under the curve (AUC) and classification accuracy (ACC) of the target domain samples, and compare and verify with existing methods.

Images