Lung infection auxiliary differential diagnosis system based on self-supervised pre-training model
Through multi-dimensional data fusion and self-supervised learning of the self-supervised pre-training model, the problems of data dependence and insufficient feature capture in traditional machine learning in the diagnosis of lung infections are solved, and efficient and accurate auxiliary diagnosis of lung infections is achieved, especially flexible diagnosis in diverse and rare cases.
Patent Information
- Application Number
- CN202510806179.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-17
- Publication Date
- 2025-09-26
AI Technical Summary
Traditional machine learning algorithms rely on expensive professionally labeled data and manually designed features, making it difficult to accurately capture the complex features in medical images. This results in insufficient accuracy and flexibility in diagnosing lung infections, especially in diverse and rare cases.
A self-supervised pre-training model is adopted to realize unsupervised feature learning and transfer learning through multi-dimensional data acquisition, image preprocessing, multi-dimensional data analysis, model application and optimization modules, combined with multimodal data fusion, rare case identification and self-supervised learning, thereby improving feature extraction capabilities and diagnostic flexibility.
It has achieved low-dependence on data labeling, efficient and accurate auxiliary diagnosis of lung infections, can adapt to diverse lesions and identify rare cases, improve diagnostic efficiency and reliability, and provide intelligent hierarchical management and the ability to identify rare cases early.
Smart Images
Figure CN120708873A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the interdisciplinary field of medical artificial intelligence and medical image analysis, and specifically to an auxiliary differential diagnosis system for lung infection based on a self-supervised pre-training model. Background Art
[0002] In the field of medical image-assisted diagnosis, traditional machine learning technology faces many bottlenecks. Traditional machine learning algorithms such as random forests and support vector machines are highly dependent on large amounts of manually annotated data. The annotation of medical data, especially medical images, requires professional doctors with rich clinical experience. Not only is the annotation cost high, but it is also inefficient, which seriously restricts the acquisition of model training data. At the same time, traditional methods of manually designed features and shallow models are difficult to accurately capture complex features such as lesion morphology, texture, and spatial relationships in medical images, which greatly affects the accuracy and reliability of diagnosis. In addition, the regularized threshold judgment method based on expert experience lacks flexibility and cannot effectively cope with the diverse types of lung infections and complex lesion manifestations. It has significant difficulties in distinguishing similar lesions such as bacterial, viral, and fungal infections, and is even more difficult to handle the diagnosis needs of rare cases or infections caused by new pathogens. Therefore, there is an urgent need for a new technical solution to break through the limitations of traditional methods and achieve efficient and accurate auxiliary differential diagnosis of lung infections. Summary of the Invention
[0003] (1) Technical problems solved
[0004] In response to the shortcomings of the existing technology, the present invention provides an auxiliary differential diagnosis system for lung infection based on a self-supervised pre-training model, which has the advantages of low dependence on data labeling, strong feature extraction capability and high diagnostic flexibility. It solves the problems of high data labeling cost, insufficient capture of complex features and difficulty in dealing with diverse lesions and rare cases in traditional technologies.
[0005] (2) Technical solution
[0006] To achieve the above objectives, the present invention provides the following technical solutions: a system for auxiliary differential diagnosis of lung infection based on a self-supervised pre-training model, comprising a multidimensional data acquisition module, an image pre-processing module, a multidimensional data analysis module, a self-supervised pre-training model, a model application module, a model evaluation module, and a model optimization module;
[0007] The multi-dimensional data acquisition module collects dynamic monitoring data, complex feature data of medical images, annotation data and multimodal impact data, and transmits them to subsequent modules through a standardized interface;
[0008] The image preprocessing module processes the collected original medical images and performs denoising, normalization and lesion area enhancement operations;
[0009] The multidimensional data analysis module performs feature extraction, multimodal data fusion and rare case identification on the data collected by the multidimensional data acquisition module and the image information preprocessed by the image preprocessing module;
[0010] The self-supervised pre-training model uses the unlabeled data obtained by the multidimensional data acquisition module and the multidimensional data analysis module to perform unsupervised feature learning and transfer learning;
[0011] The model application module applies the self-supervised pre-trained model to actual scenarios to perform real-time diagnosis, lesion visualization, and early warning generation;
[0012] The model evaluation module performs analysis and evaluation based on the scenario output results and clinical real labels;
[0013] The model optimization module performs optimization measures based on evaluation feedback and active learning strategies.
[0014] Preferably, the multidimensional data acquisition module includes a dynamic monitoring data unit, a medical image complex feature data unit, a medical image annotation data unit and a multimodal influence data unit. The multidimensional data analysis module includes a multidimensional data fusion unit, a self-supervised comparative learning unit and a rare case adaptive detection unit.
[0015] Preferably, the dynamic monitoring data unit collects the patient's vital signs and condition change data in real time through wearable devices or medical Internet of Things.
[0016] Preferably, the medical image complex feature data unit automatically extracts the lesion morphology, texture and spatial relationship features in the medical image through a deep learning algorithm.
[0017] Preferably, the medical image annotation data unit acquires the annotated image data with the assistance of a semi-automatic tool.
[0018] Preferably, the multimodal impact data unit collects laboratory test data and electronic medical record text non-image information through the medical information system interface and natural language processing and structured conversion, and transmits it to the image preprocessing module through the network for further image preprocessing.
[0019] Preferably, the multidimensional data fusion unit calculates the multidimensional data fusion index Ez based on the collected data and preprocessing information, and the calculation formula is:
[0020]
[0021] In the formula, Ez represents the multidimensional data fusion index, Y d Represents the dynamic monitoring data vector, Y b Represents the lesion feature data vector, Y grepresents the labeled data vector, Y l represents the epidemiological data vector, a1, a2, a3 and a4 represent the weight coefficients of the dynamic monitoring data vector, lesion feature data vector, annotation data vector and epidemiological data vector respectively, and k represents the multimodal influence coefficient.
[0022] Preferably, the self-supervised contrastive learning unit calculates the infection type score Qv based on the collected data and pre-processed information j , and its calculation formula is:
[0023]
[0024] In the formula, Qv j Indicates the infection type score, s i represents the score of the i-th feature for the j-th type of infection, w i represents the weight of the i-th feature, and m represents the number of features.
[0025] Preferably, the rare case adaptive detection unit calculates the rare case adaptive detection index Ux based on the collected data and preprocessing information, and the calculation formula is:
[0026]
[0027] In the formula, Ux represents the adaptive detection index of rare cases, f represents the embedding vector of the sample to be detected in the feature space, μ represents the mean of the feature vector of normal cases, α represents the standard deviation of the feature vector of normal cases, n represents the number of cluster centers, and T i The eigenvector representing the i-th cluster center.
[0028] Preferably, the self-supervised pre-training model uses the multidimensional data fusion index Ez in the multidimensional data analysis module to perform unsupervised feature learning and representation space construction, and uses the infection type score Qv j , perform supervisory signal-guided feature weight optimization, and use the rare case adaptive detection index Ux to mine abnormal samples and enhance model robustness.
[0029] Compared with the existing technology, the present invention provides a lung infection auxiliary differential diagnosis system based on a self-supervised pre-training model, which has the following beneficial effects:
[0030] 1. The present invention calculates the multidimensional data fusion index Ez and substitutes it into the loss function of the self-supervised pre-training model to replace the traditional single modality data input, thereby guiding the model to learn the collaborative feature representation of multimodal data. When the multidimensional data fusion index Ez is compared with the average fusion index of similar historical cases, when the multidimensional data fusion index Ez deviates from the mean by more than ±2 times the standard deviation, it indicates that the modal fusion characteristics of the current case data are significantly different from those of common cases. The system adjusts the parameter update direction of the self-supervised pre-training accordingly, ultimately achieving the effect of enhancing the model's adaptability to complex case data.
[0031] 2. The present invention calculates the infection type score Qv j , as the quantitative decision basis for the model to judge the infection type, when the infection type score Qv j When the infection type score Qv is ≥0.8, the system directly outputs a high-confidence diagnosis result of the infection type and automatically generates a standardized treatment plan; when 0.5≤ infection type score Qv j When the infection type score Qv is less than 0.8, the system will mark the case as suspected infection, triggering the secondary cross-validation of multimodal data, combining the dynamic monitoring data change trend with the similar case database for comprehensive analysis, and assisting doctors in making more accurate diagnostic decisions; when ... j When the score is less than 0.5, the system will place the case in the low-risk observation queue and continue to track dynamic monitoring data. Once the score shows an upward trend or key vital signs are abnormal, the upgraded diagnosis process will be immediately initiated. Through the above differentiated processing of different scoring intervals, the system can achieve intelligent hierarchical management from preliminary screening to accurate diagnosis, thereby improving the efficiency and reliability of differential diagnosis of lung infection.
[0032] 3. The present invention calculates the rare case adaptive detection index Ux and substitutes the rare case adaptive detection index Ux into the abnormal sample screening mechanism of the model to trigger the special labeling process and data enhancement strategy in the system. The rare case adaptive detection index Ux is compared with the preset rare case threshold. When the threshold is exceeded, the system automatically marks the case as a rare case, pushes it to the expert consultation platform as a priority, and extracts more detailed feature information from the multimodal data for secondary analysis; at the same time, the case data is included in the active learning sample pool for subsequent model optimization training. Finally, the system has the ability to identify rare cases early and accumulate knowledge, thereby improving the comprehensiveness of the overall diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is a flow chart of the system of the present invention. DETAILED DESCRIPTION
[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0035] See also Figure 1 , an auxiliary differential diagnosis system for lung infection based on a self-supervised pre-training model, including a multi-dimensional data acquisition module, an image pre-processing module, a multi-dimensional data analysis module, a self-supervised pre-training model, a model application module, a model evaluation module and a model optimization module;
[0036] The multi-dimensional data acquisition module collects dynamic monitoring data, complex feature data of medical images, annotation data, and multimodal impact data, and transmits them to subsequent modules through standardized interfaces;
[0037] The image preprocessing module processes the collected original medical images (such as X-rays and CT scans) and performs denoising, normalization, and lesion area enhancement operations;
[0038] The multidimensional data analysis module combines the data collected by the multidimensional data acquisition module with the image information preprocessed by the image preprocessing module to perform feature extraction, multimodal data fusion and rare case identification;
[0039] The self-supervised pre-training model uses the unlabeled data obtained by the multi-dimensional data acquisition module and the multi-dimensional data analysis module to perform unsupervised feature learning and transfer learning;
[0040] The model application module applies the self-supervised pre-trained model to actual scenarios for real-time diagnosis, lesion visualization, and warning generation. The warning generation logic is as follows:
[0041] Warning generation logic: When the infection type score Qv calculated by the self-supervised contrastive learning unit j When the set high-risk threshold is exceeded (e.g., score ≥ 0.8), and the rare case adaptive detection index Ux does not trigger a rare case alarm, the system determines it as a common high-risk infection and generates a red warning; when the rare case adaptive detection index Ux is higher than the rare case threshold (e.g., index ≥ 3), it is determined to be a rare case and generates an orange warning; when the infection type score Qv j When the patient is in the medium-to-low risk range (e.g., 0.3 < score < 0.8), but key vital signs in the dynamic monitoring data (e.g., blood oxygen saturation < 90%) show abnormal fluctuations, a yellow warning is generated to remind medical staff to pay attention to changes in the patient's condition;
[0042] The model evaluation module analyzes and evaluates the scenario output results and clinical real labels;
[0043] The model optimization module takes optimization measures based on evaluation feedback and active learning strategies (when the model evaluation module detects that the diagnostic accuracy rate has continuously decreased by more than 5%, the missed diagnosis rate of rare cases is higher than 10%, or the distribution difference between the newly emerged case data and the existing training data exceeds the set threshold, the model optimization module's active learning needs to be optimized by actively selecting unlabeled samples with rich information for labeling, incorporating them into model training, and adjusting model parameters and algorithm structure to improve model performance).
[0044] The multidimensional data acquisition module includes a dynamic monitoring data unit, a medical image complex feature data unit, a medical image annotation data unit, and a multimodal influence data unit. The multidimensional data analysis module includes a multidimensional data fusion unit, a self-supervised comparative learning unit, and a rare case adaptive detection unit.
[0045] The dynamic monitoring data unit collects patients' vital signs (such as body temperature, blood oxygen, respiratory rate, etc.) and disease condition change data in real time through wearable devices or medical Internet of Things.
[0046] The complex feature data unit of medical images automatically extracts the lesion morphology (shape, edge), texture (grayscale distribution, gradient) and spatial relationship (lesion location, adjacent tissue) features in medical images through deep learning algorithms (such as convolutional neural networks and semantic segmentation models).
[0047] The medical image annotation data unit obtains annotated image data (such as lesion classification labels) with the assistance of semi-automatic tools.
[0048] The multimodal impact data unit collects laboratory test data (such as blood routine, C-reactive protein), electronic medical record text (such as symptom description) and non-imaging information through the medical information system (HIS / LIS) interface and natural language processing (NLP) and structured conversion, and transmits it to the image preprocessing module through the network for further image preprocessing.
[0049] The multidimensional data fusion unit calculates the multidimensional data fusion index Ez based on the collected data and preprocessing information. The calculation formula is:
[0050]
[0051] In the formula, Ez represents the multidimensional data fusion index, Y d Represents dynamic monitoring data vector (body temperature, blood oxygen), Y b Represents the lesion feature data vector (morphology, texture and spatial relationship), Y g represents the labeled data vector (lesion classification label annotated by semi-automatic tools), Y lrepresents the epidemiological data vector, a1, a2, a3, and a4 represent the weight coefficients of the dynamic monitoring data vector, lesion feature data vector, labeled data vector, and epidemiological data vector, respectively. This formula is used to achieve the organic integration of multi-source data and reduce the dependence on a single labeled data. k represents the multimodal influence coefficient.
[0052] The advantages are: by calculating the multidimensional data fusion index Ez and substituting it into the loss function of the self-supervised pre-training model, it replaces the traditional single modal data input, thereby guiding the model to learn the collaborative feature representation of multimodal data. When the multidimensional data fusion index Ez is compared with the average fusion index of similar historical cases, when the multidimensional data fusion index Ez deviates from the mean by more than ±2 times the standard deviation, it means that the modal fusion characteristics of the current case data are significantly different from those of common cases. The system adjusts the parameter update direction of the self-supervised pre-training accordingly, and ultimately achieves the effect of enhancing the model's adaptability to complex case data.
[0053] The self-supervised contrastive learning unit calculates the infection type score Qv based on the collected data and preprocessed information j , and its calculation formula is:
[0054]
[0055] In the formula, Qv j Indicates the infection type score, s i represents the score of the i-th feature for the j-th type of infection (such as the score of morphological features supporting bacterial pneumonia), w i represents the weight of the i-th feature (optimized through active learning or grid search), and m represents the number of features (such as texture, spatial relationship, and the total amount of clinical indicators and other indicators). The weighted summation is used to achieve infection type segmentation to cope with diverse lesions (such as ground-glass opacities and cavities) and rare cases (such as fungal infections).
[0056] Advantages: By calculating the infection type score Qv j , as the quantitative decision basis for the model to judge the infection type, when the infection type score Qv j When the infection type score Qv is ≥0.8, the system directly outputs a high-confidence diagnosis result of the infection type and automatically generates a standardized treatment plan; when 0.5≤ infection type score Qv j When the infection type score Qv is less than 0.8, the system will mark the case as suspected infection, triggering the secondary cross-validation of multimodal data, combining the dynamic monitoring data change trend with the similar case database for comprehensive analysis, and assisting doctors in making more accurate diagnostic decisions; when ... jWhen the score is less than 0.5, the system will place the case in the low-risk observation queue and continue to track dynamic monitoring data. Once the score shows an upward trend or key vital signs are abnormal, the upgraded diagnosis process will be immediately initiated. Through the above differentiated processing of different scoring intervals, the system can achieve intelligent hierarchical management from preliminary screening to accurate diagnosis, thereby improving the efficiency and reliability of differential diagnosis of lung infection.
[0057] The rare case adaptive detection unit calculates the rare case adaptive detection index Ux based on the collected data and preprocessing information. The calculation formula is:
[0058]
[0059] In the formula, Ux represents the adaptive detection index of rare cases, f represents the embedding vector of the sample to be detected in the feature space (extracted by self-supervised contrastive learning), μ represents the mean of the normal case feature vector, which is used to measure the average difference between the sample and the normal case group, reflecting the degree to which the sample deviates from the normal data distribution center, α represents the standard deviation of the normal case feature vector, which reflects the degree of dispersion of the normal case data and is used to normalize the difference value to make the feature differences of different dimensions comparable and ensure the stability of the formula calculation, n represents the number of cluster centers, which is obtained by clustering analysis of the normal case feature vector and is used to determine the category structure of the data distribution, and T i Represents the eigenvector of the i-th cluster center. This formula achieves adaptive detection of rare cases by comprehensively considering the distance between the sample and the normal case group and the similarity with the known category structure.
[0060] The advantages are: by calculating the adaptive detection index Ux of rare cases and substituting the adaptive detection index Ux of rare cases into the abnormal sample screening mechanism of the model, the special labeling process and data enhancement strategy in the system are triggered, and the adaptive detection index Ux of rare cases is compared with the preset rare case threshold (adaptive detection index Ux of rare cases ≥ 3). When the threshold is exceeded, the system automatically marks the case as a rare case, pushes it to the expert consultation platform as a priority, and extracts more detailed feature information from the multimodal data for secondary analysis; at the same time, the case data is included in the active learning sample pool for subsequent model optimization training, and finally enables the system to have the ability to identify rare cases early and accumulate knowledge, thereby improving the comprehensiveness of the overall diagnosis.
[0061] The self-supervised pre-training model uses the multidimensional data fusion index Ez in the multidimensional data analysis module as a metric for the joint representation of multimodal data, driving the model to mine potential correlations between data in the unsupervised learning stage, perform unsupervised feature learning and representation space construction, and use the infection type score Qv j, as a quantitative indicator of supervisory signals, by minimizing the difference between predicted scores and true scores, dynamically adjusting the weight distribution of each feature, and optimizing the feature weights guided by supervisory signals. The rare case adaptive detection index Ux is used as the basis for screening abnormal samples. By identifying samples that deviate from the normal distribution, the diversity of model training is expanded, and abnormal sample mining and model robustness enhancement are carried out.
[0062] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A lung infection auxiliary differential diagnosis system based on a self-supervised pre-training model, characterized by: It includes multi-dimensional data acquisition module, image pre-processing module, multi-dimensional data analysis module, self-supervised pre-training model, model application module, model evaluation module and model optimization module; The multi-dimensional data acquisition module collects dynamic monitoring data, complex feature data of medical images, annotation data and multimodal impact data, and transmits them to subsequent modules through a standardized interface; The image preprocessing module processes the collected original medical images and performs denoising, normalization and lesion area enhancement operations; The multidimensional data analysis module performs feature extraction, multimodal data fusion and rare case identification on the data collected by the multidimensional data acquisition module and the image information preprocessed by the image preprocessing module; The self-supervised pre-training model uses the unlabeled data obtained by the multidimensional data acquisition module and the multidimensional data analysis module to perform unsupervised feature learning and transfer learning; The model application module applies the self-supervised pre-trained model to actual scenarios to perform real-time diagnosis, lesion visualization, and early warning generation; The model evaluation module performs analysis and evaluation based on the scenario output results and clinical real labels; The model optimization module performs optimization measures based on evaluation feedback and active learning strategies.
2. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 1 is characterized by: The multidimensional data acquisition module includes a dynamic monitoring data unit, a medical image complex feature data unit, a medical image annotation data unit and a multimodal influence data unit. The multidimensional data analysis module includes a multidimensional data fusion unit, a self-supervised comparative learning unit and a rare case adaptive detection unit.
3. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The dynamic monitoring data unit collects the patient's vital signs and condition change data in real time through wearable devices or medical Internet of Things.
4. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The medical image complex feature data unit automatically extracts the lesion morphology, texture and spatial relationship features in the medical image through a deep learning algorithm.
5. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The medical image annotation data unit acquires the annotated image data with the assistance of a semi-automatic tool.
6. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The multimodal impact data unit collects laboratory test data, electronic medical record text and non-imaging information through a medical information system interface and natural language processing and structured conversion, and transmits it to an image preprocessing module through a network for further image preprocessing.
7. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The multidimensional data fusion unit calculates the multidimensional data fusion index Ez based on the collected data and preprocessing information, and its calculation formula is: In the formula, Ez represents the multidimensional data fusion index, Y d Represents the dynamic monitoring data vector, Y b Represents the lesion feature data vector, Y g represents the labeled data vector, Y l represents the epidemiological data vector, a1, a2, a3 and a4 represent the weight coefficients of the dynamic monitoring data vector, lesion feature data vector, annotation data vector and epidemiological data vector respectively, and k represents the multimodal influence coefficient.
8. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2 is characterized by: The self-supervised contrastive learning unit calculates the infection type score Qv based on the collected data and pre-processed information j , and its calculation formula is: In the formula, Qv j Indicates the infection type score, s i represents the score of the i-th feature for the j-th type of infection, w i represents the weight of the i-th feature, and m represents the number of features.
9. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 2, characterized in that: The rare case adaptive detection unit calculates the rare case adaptive detection index Ux based on the collected data and preprocessing information, and the calculation formula is: In the formula, Ux represents the adaptive detection index of rare cases, f represents the embedding vector of the sample to be detected in the feature space, μ represents the mean of the feature vector of normal cases, α represents the standard deviation of the feature vector of normal cases, n represents the number of cluster centers, and T i The eigenvector representing the i-th cluster center.
10. The lung infection auxiliary differential diagnosis system based on the self-supervised pre-training model according to claim 1, characterized in that: The self-supervised pre-training model uses the multidimensional data fusion index Ez in the multidimensional data analysis module to perform unsupervised feature learning and representation space construction, and uses the infection type score Qv j , perform supervisory signal-guided feature weight optimization, and use the rare case adaptive detection index Ux to mine abnormal samples and enhance model robustness.