Medical image X-ray pneumonia auxiliary evaluation system based on AI
By using conditional generative adversarial networks with multi-expert consensus and knowledge distillation techniques, a high-quality medical image dataset is generated, which solves the problems of high cost and poor consistency of manual annotation, realizes a high-precision pneumonia auxiliary assessment model, and improves the robustness of the model and human-machine collaborative diagnostic capabilities.
Patent Information
- Application Number
- CN202511430938.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-09
- Publication Date
- 2026-02-03
AI Technical Summary
Existing technologies rely on manually labeled medical image datasets, which are costly to obtain, have poor label consistency, and are difficult to meet the needs of deep learning models. Furthermore, the quality of the generated synthetic images is not high, resulting in low accuracy of pneumonia auxiliary assessment models.
By using a conditional generative adversarial network based on multi-expert consensus, high-quality medical image training data is generated. Knowledge distillation technology is used to integrate expert annotations, outputting high-fidelity soft labels, and training a high-performance pneumonia auxiliary assessment model.
It significantly improved the accuracy and robustness of the pneumonia auxiliary assessment model, reduced the cost and time of manual annotation, realized human-machine collaborative diagnosis, and enhanced the model's ability to identify rare cases.
Smart Images
Figure CN121460147A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence image processing, and in particular to an AI-based medical imaging X-ray pneumonia auxiliary assessment system. Background Technology
[0002] Artificial intelligence (AI)-based medical image analysis techniques, particularly the assisted diagnosis of pneumonia using X-ray images, offer new avenues for improving the efficiency and accuracy of clinical diagnosis. The performance of these techniques heavily relies on large-scale, high-quality labeled datasets for model training. However, obtaining ideal datasets that effectively support model training remains a fundamental challenge in this field.
[0003] Current technologies typically rely on manual data annotation by radiologists. This process has inherent bottlenecks: expert annotation is extremely time-consuming and labor-intensive, resulting in high costs and hindering large-scale implementation. This core bottleneck further leads to three interconnected superficial problems: First, the total amount of annotated data is scarce, making it difficult to meet the needs of data-driven deep learning models; second, due to differences in clinical experience and judgment criteria among different experts, the annotation results for the same image often differ, leading to poor annotation consistency, and these noisy labels limit the upper limit of model performance; third, it is difficult to obtain a sufficient number of rare case data, resulting in class imbalance in the training data, making the trained model insufficient in identifying clinically valuable minority cases.
[0004] To alleviate data scarcity, conventional techniques often employ basic data augmentation (such as image rotation and flipping) or generative adversarial networks (GANs) to generate synthetic images. However, the former offers limited diversity, while the latter, although capable of generating images, lacks reliable annotations and fails to provide the labeling information needed for model training, thus not addressing the fundamental bottleneck of annotation. The training data generated by these two methods is of low quality, resulting in low accuracy of the trained pneumonia-aided assessment models. Summary of the Invention
[0005] In view of the aforementioned deficiencies of the prior art, the technical problem to be solved by the present invention is to provide an AI-based medical imaging X-ray pneumonia auxiliary assessment system, which aims to generate high-quality medical image training data to train the model and improve the model's auxiliary assessment performance.
[0006] To achieve the above objectives, this invention discloses an AI-based medical imaging X-ray pneumonia auxiliary assessment system, the system comprising:
[0007] The training data acquisition module is used to obtain a raw medical image dataset containing raw medical images with hard labels independently annotated by different experts.
[0008] The expert annotation model training module is used to input the original medical images from the same expert into the same first training model for training, so as to obtain expert annotation models corresponding to different experts.
[0009] The training data multi-labeling module is used to input the original medical image with the original hard labels removed into each of the expert labeling models to obtain different hard labels for the original medical image by the expert labeling models.
[0010] The consensus soft label generation module is used to fuse the various hard labels of the original medical image through knowledge distillation to obtain soft labels that characterize the degree of consensus and disagreement of the various expert annotation models.
[0011] The image data generation model training module is used to train a first conditional generative adversarial network using the soft labels as conditional information to obtain the image data generation model. The training process includes: having the generator output a synthesized medical image based on the soft labels and a noise vector; having a discriminator receive the original medical image and its corresponding soft labels, and the synthesized medical image and its corresponding soft labels, and perform a real / fake judgment on the input medical image; and iteratively training until the first conditional generative adversarial network converges to obtain the image data generation model.
[0012] The pneumonia auxiliary assessment model training module is used to input the synthetic training medical images generated by the image data generation model and their corresponding soft labels into the second training model for training, so as to obtain the pneumonia auxiliary assessment model.
[0013] The pneumonia auxiliary assessment module is used to input the lung medical image to be labeled into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model obtains a first soft label for the lung medical image and labels the lung medical image based on the value of the first soft label.
[0014] Optionally, the expert annotation model training module is specifically used for:
[0015] The original medical images from the same expert annotation are input into the same first training model, so that the first training model clusters the original medical images according to the annotation style features of the original medical images to obtain the original medical images of multiple annotation style categories; the first training model is trained according to the original medical images of different annotation style categories to obtain expert annotation models; wherein, the same expert annotation model includes multiple expert annotation sub-models corresponding to different annotation style categories, that is, the same expert annotation model outputs a corresponding number of hard labels according to the number of its annotation style categories.
[0016] Optionally, the hard label is a clear annotation of the lesion area in the original medical image, and the soft label is a pixel-level probability map corresponding to the lesion area. The pneumonia auxiliary assessment module is specifically used for:
[0017] The lung medical image to be labeled is input into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model obtains the pixel-level probability map of the lung medical image, and the lesion area of the lung medical image is labeled based on the probability value corresponding to each pixel in the pixel-level probability map; wherein, when the probability value is greater than a preset threshold, the corresponding pixel is considered to belong to the lesion area.
[0018] Optionally, the consensus soft tag generation module is specifically used for:
[0019] The distribution of each hard label is calculated, and the corresponding label weight of each hard label is obtained; wherein, the label weight is determined based on the labeling accuracy of the corresponding expert.
[0020] Based on the distribution results of the hard labels and the label weights, the hard labels are fused to obtain soft labels that characterize the degree of consensus and disagreement of the various expert labeling models.
[0021] Optionally, in the image data generation model training module, the input of the discriminator is the spliced data of image-soft label pairs, and its task is to determine whether the image and the soft label match and are real.
[0022] Optionally, the image data generation model can receive manually specified target soft tags as conditions and generate synthetic medical images corresponding to the target soft tags for targeted augmentation of specific types of pneumonia data.
[0023] Optionally, the system further includes a data preprocessing module.
[0024] The data preprocessing module is used to perform intensity normalization and contrast enhancement operations on the original medical image dataset and / or the synthetic medical images, and to send the processed data to subsequent modules.
[0025] Optionally, the pneumonia auxiliary assessment model training module adopts a course learning strategy, first using the synthetic training medical images for preliminary training, and then gradually introducing the original medical images for refined training.
[0026] The beneficial effects of this invention are as follows: 1. This invention, through a conditional generative adversarial network based on multi-expert consensus, can automatically generate a large amount of high-fidelity synthetic pneumonia X-ray image data, and each synthetic data comes with high-quality soft labels. This greatly reduces the cost and time pressure of collecting and labeling real clinical data, providing a data foundation for training high-performance models. 2. This invention, by training corresponding expert labeling models to intelligently label each original medical image with only a single hard label, can not only effectively reduce the time and effort of manual data labeling, but also avoid labeling errors caused by visual observation. 3. This invention, through knowledge distillation technology, merges the independent labels of multiple experts into a unified consensus soft label, effectively eliminating the subjective bias and random noise of single expert labels. The conditional generative adversarial network trained with these soft labels and the final evaluation model learn the "wise consensus" of the expert group, thereby significantly improving the accuracy, robustness, and generalization ability of the final pneumonia auxiliary evaluation model. 4. The expert annotation model training module of this invention can cluster original medical images attributed to the same expert based on their annotation style characteristics, obtaining original medical images with multiple annotation style categories. This allows the trained expert annotation model to include multiple expert annotation sub-models corresponding to different annotation style categories, enabling the same expert annotation model to output multiple hard labels. This effectively solves the problem of annotation consistency discrepancies caused by shared accounts under the same author (generally due to borrowing or external factors such as shared departments), making the hard labels output by the expert annotation model more consistent with the corresponding expert's clinical experience and style, thus making the subsequent training of the evaluation model more accurate. 5. Because the training data itself contains soft labels with probabilistic information, the final trained pneumonia auxiliary evaluation model can output pixel-level probability maps, rather than arbitrary binary segmentation results. This allows doctors to intuitively understand the model's confidence level in their judgments, focusing on areas that the model also considers "uncertain," thereby achieving true human-machine collaboration and improving the reliability of diagnostic decisions. 6. This invention allows users to manually specify the attributes of the target soft labels (such as disease type and severity) to generate targeted pneumonia image data of scarce categories. This on-demand generation capability can perfectly balance the training dataset, enabling the model to have excellent recognition capabilities for various types of pneumonia, especially rare types, and avoiding model bias towards the majority class.
[0027] In summary, this invention, through an innovative and systematic process, collaboratively solves several interrelated core technical challenges such as data scarcity, annotation noise, and uncertainty quantification, ultimately producing a technical effect greater than the sum of its parts, and significantly advancing the development of medical imaging AI technology. Attached Figure Description
[0028] Figure 1This is a schematic diagram of the structure of an AI-based medical imaging X-ray pneumonia auxiliary assessment system provided in a specific embodiment of the present invention;
[0029] Figure 2 This is a flowchart illustrating an AI-based medical imaging X-ray-assisted assessment method for pneumonia, provided in a specific embodiment of the present invention. Detailed Implementation
[0030] This invention discloses an AI-based medical imaging X-ray pneumonia auxiliary assessment system. Those skilled in the art can refer to the content of this document and appropriately improve the technical details for implementation. It should be particularly noted that all similar substitutions and modifications are obvious to those skilled in the art and are considered to be included in this invention. The apparatus and application of this invention have been described through preferred embodiments. Those skilled in the art can obviously make modifications or appropriate alterations and combinations to the apparatus and application described herein without departing from the content, spirit, and scope of this invention to implement and apply the technology of this invention.
[0031] The applicant's research revealed that, to alleviate data scarcity, conventional techniques often employ basic data augmentation (such as image rotation and flipping) or Generative Adversarial Networks (GANs) to generate synthetic images. However, the former offers limited diversity; while the latter can generate images, its output lacks reliable annotations, failing to provide the label information needed for model training, thus not addressing the fundamental bottleneck of annotation. More importantly, existing methods cannot systematically resolve the core contradiction caused by the high cost of manual annotation—the inability to simultaneously achieve data scale, annotation consistency, and class balance.
[0032] Therefore, embodiments of the present invention provide an AI-based medical imaging X-ray pneumonia auxiliary assessment system, such as... Figure 1 As shown, the system includes:
[0033] The training data acquisition module 101 is used to obtain the original medical image dataset containing original medical images with hard labels independently annotated by different experts.
[0034] The expert annotation model training module 102 is used to input the original medical images from the same expert annotation into the same first training model for training, so as to obtain expert annotation models corresponding to different experts.
[0035] The training data multi-labeling module 103 is used to input the original medical images with the original hard labels removed into various expert labeling models to obtain different hard labels for the original medical images by different expert labeling models.
[0036] The consensus soft label generation module 104 is used to fuse the various hard labels of the original medical image through knowledge distillation to obtain soft labels that characterize the degree of consensus and disagreement of various expert annotation models.
[0037] The image data generation model training module 105 is used to train a first conditional generative adversarial network using soft labels as conditional information to obtain an image data generation model. The training process includes: enabling the generator to output a synthetic medical image based on soft labels and noise vectors; enabling the discriminator to receive the original medical image and its corresponding soft label and the synthetic medical image and its corresponding soft label, and to judge the authenticity of the input medical image; iteratively training until the first conditional generative adversarial network converges to obtain the image data generation model.
[0038] The pneumonia auxiliary assessment model training module 106 is used to input the synthetic training medical images generated by the image data generation model and their corresponding soft labels into the second training model for training, so as to obtain the pneumonia auxiliary assessment model.
[0039] The pneumonia auxiliary assessment module 107 is used to input the lung medical images to be labeled into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model can obtain the first soft label of the lung medical images and label the lung medical images based on the value of the first soft label.
[0040] It should be noted that:
[0041] 1. Training Data Acquisition Module 101
[0042] Function: Collect the most original input materials.
[0043] Input: Raw dataset of pneumonia X-ray images.
[0044] Processing: Each image was independently annotated by K (K≥2) radiologists (e.g., delineating pneumonia lesions on the image).
[0045] Output: A collection of images and their corresponding (potentially inconsistent) expert annotations.
[0046] Objective: To obtain raw data that includes human expert knowledge and its inherent uncertainties.
[0047] 2. Expert Annotation Model Training Module 102
[0048] Function: Learn each expert's personal "annotation style".
[0049] Input: Data from training data acquisition module 101, but grouped by expert (i.e., all labeled data of expert A, all labeled data of expert B, etc.).
[0050] Process: Train a separate machine learning model for each expert (called the "first training model"). The learning objective of this model is: given an X-ray image, predict how that particular expert would annotate it.
[0051] Output: K expert labeling models, each model mimicking the labeling behavior of a specific expert.
[0052] Objective: To digitize and model the annotation capabilities of human experts in preparation for subsequent steps.
[0053] 3. Training data multi-labeling module 103
[0054] Function: Use digital "experts" to re-annotate data and generate multi-perspective labels.
[0055] Input: The original X-ray image (with original labels removed) and K pre-trained expert models.
[0056] Processing: Each original image is sequentially input into K expert models to obtain K new annotation results.
[0057] Output: K different digital annotation results for the same image.
[0058] Objective: To eliminate random errors in the original annotations and generate more regular multi-labels that can be used for efficient consensus calculation.
[0059] 4. Consensus Soft Tag Generation Module 104
[0060] Function: Integrates opinions from multiple experts to generate probabilistic labels.
[0061] Input: K digitized annotations for the same image generated by the training data multi-labeling module 103.
[0062] Processing: Apply knowledge distillation techniques. Taking pixel labeling as an example, for each pixel in the image, calculate the average of the K labels. For instance, if a pixel is considered a lesion by 3 models and normal by 2, then the soft label value for that pixel is 0.6.
[0063] Output: Soft label - a probability map with the same size as the original image, where each pixel value is between 0 and 1, and the value represents the probability that the point is a lesion.
[0064] Objective: To integrate multiple potentially contradictory hard labels into a high-quality probability graph that reflects expert consensus and uncertainty.
[0065] 5. Image Data Generation Model Training Module 105
[0066] Function: Train a conditional generative adversarial network (cGAN).
[0067] Input: Soft tags generated by consensus soft tag generation module 104.
[0068] deal with:
[0069] The first step is to input soft labels as conditional information into cGAN.
[0070] The second step, the generator, learns to receive random noise and a soft label, and generates an X-ray image that looks realistic and whose content closely matches the soft label.
[0071] The third step is the discriminator, which learns to determine whether the input is a "real image-soft label pair" or a "generated image-soft label pair".
[0072] The fourth step involves adversarial training between the two generators until the generator can produce incredibly realistic images.
[0073] Output: A trained image data generation model (i.e., the generator of cGAN).
[0074] Objective: To create a factory that can generate an unlimited number of "synthetic image-precise soft label pairs" to completely solve the problem of data scarcity.
[0075] 6. Training Module for Pneumonia Auxiliary Assessment Model 106
[0076] Function: To train the final pneumonia diagnostic model.
[0077] Input: A large number of "synthetic image-soft label pairs" generated by the image data generation model.
[0078] Processing: Train a deep learning model using synthetic data (“second training model”).
[0079] Output: The trained pneumonia auxiliary assessment model.
[0080] Objective: To train a high-performance, robust final application model using high-quality synthetic data.
[0081] 7. Pneumonia Auxiliary Assessment Module 107
[0082] Function: Practical application.
[0083] Input: A new, undiagnosed chest X-ray image.
[0084] Processing: Input the image into the trained model of the pneumonia auxiliary assessment model training module 106.
[0085] Output: The first soft label output by the model is a pixel-level probability map that identifies the lesion area and its probability.
[0086] Objective: To provide doctors with visualized diagnostic results that include information about uncertainty.
[0087] In this specific embodiment, the expert annotation model training module 102 is specifically used for:
[0088] Original medical images from the same expert annotation are input into the same first training model, so that the first training model can cluster according to the annotation style features of the original medical images to obtain original medical images with multiple annotation style categories; the first training model is trained according to original medical images with different annotation style categories to obtain expert annotation models; wherein, the same expert annotation model includes multiple expert annotation sub-models corresponding to different annotation style categories, that is, the same expert annotation model outputs a corresponding number of hard labels according to the number of its annotation style categories.
[0089] It should be noted that in actual annotation processes, account sharing often occurs due to multiple experts working in the same department, account borrowing, or other factors. This means that the original medical image annotation styles corresponding to the same author may differ. If the first model is trained to output a single hard label, the inconsistent training data styles will cause the hard labels to deviate from the styles of these experts. Therefore, this embodiment of the invention first performs clustering before training, so that the annotation model for the same expert outputs a corresponding number of hard labels based on the number of its annotation style categories. These hard labels correspond to the annotation styles of various experts under the same authorship, effectively solving the aforementioned problem.
[0090] In this specific embodiment, hard labels are explicit markings of lesion areas in the original medical images, and soft labels are pixel-level probability maps corresponding to the lesion areas. The pneumonia auxiliary assessment module 107 is specifically used for:
[0091] The lung medical images to be labeled are input into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model can obtain a pixel-level probability map of the lung medical images. Based on the probability value corresponding to each pixel in the pixel-level probability map, the lesion areas of the lung medical images are labeled. When the probability value is greater than a preset threshold, the corresponding pixel is considered to belong to the lesion area.
[0092] It should be noted that assisted lesion area identification is one of the functions of the corresponding evaluation model in this embodiment of the invention. In this specific embodiment, the invention uses pixel-level probability maps, rather than arbitrary binary segmentation results. This allows doctors to intuitively understand the model's confidence level in its judgments, focusing on areas that the model also considers "uncertain," thereby achieving true human-machine collaboration and improving the reliability of diagnostic decisions.
[0093] In this specific embodiment, the consensus soft tag generation module 104 is specifically used for:
[0094] The distribution of each hard label is calculated, and the corresponding label weight of each hard label is obtained; the label weight is determined based on the labeling accuracy of the corresponding expert.
[0095] Based on the distribution results and label weights of the hard labels, the hard labels are fused to obtain soft labels that characterize the degree of consensus and disagreement among the various expert labeling models.
[0096] It should be noted that different experts have different clinical experience, so the accuracy of labeling varies. Based on this, weighting can further improve the credibility of soft labels.
[0097] In this specific embodiment, in the image data generation model training module 105, the input of the discriminator is the spliced data of image-soft label pairs, and its task is to determine whether the image and the soft label match and are real.
[0098] It should be noted that the conditional generative adversarial network (cGAN) in this invention is conditional, and the discriminator performs a pairwise discrimination task. This is a key difference from ordinary generative adversarial networks (GANs).
[0099] In this specific embodiment, the image data generation model can receive manually specified target soft tags as conditions and generate synthetic medical images corresponding to the target soft tags for targeted augmentation of specific types of pneumonia data.
[0100] It should be noted that the data in this embodiment is generated on demand, which solves the problem of data category imbalance.
[0101] In this specific embodiment, the system further includes a data preprocessing module.
[0102] The data preprocessing module performs intensity normalization and contrast enhancement operations on the raw medical image dataset and / or synthetic medical images, and then sends the processed data to subsequent modules.
[0103] It should be noted that preprocessing data can be a specific method to improve data quality and enhance model training performance.
[0104] In this specific embodiment, the pneumonia auxiliary assessment model training module 106 adopts a course learning strategy, first using synthetic training medical images for preliminary training, and then gradually introducing original medical images for refined training.
[0105] It should be noted that this strategy can improve training stability and the final model performance.
[0106] The steps of the AI-based medical imaging X-ray-assisted assessment method for pneumonia corresponding to the system of this embodiment are as follows: Figure 2 As shown, it includes:
[0107] S1. Obtain the original medical image dataset. The hard labels corresponding to the original medical images in the original medical image dataset come from independent annotations by different experts.
[0108] S2. Input the original medical image data from the same expert into the same first training model for training, so as to obtain expert label annotation models corresponding to different experts.
[0109] S3. Input the original medical images with the original hard labels removed into each expert annotation model to obtain the hard labels of the original medical images for different expert annotation models.
[0110] S4. Through knowledge distillation, the various hard labels of the original medical image data are fused to obtain the soft labels of the original medical image data.
[0111] S5. Using soft labels as conditional information, train the first conditional generative adversarial network to obtain the image data generation model.
[0112] The training process includes enabling the generator to output synthetic medical images based on soft labels and noise vectors; enabling the discriminator to receive the original medical image data and its corresponding soft labels, as well as the synthetic medical images and their corresponding soft labels, and to judge the authenticity of the input medical images; and iteratively training until the first conditional generative adversarial network converges to obtain the image data generation model.
[0113] S6. Input the synthetic training medical images generated by the image data generation model and their corresponding soft labels into the second training model for training to obtain a pneumonia auxiliary assessment model.
[0114] S7. Input the lung medical images to be labeled into the pneumonia auxiliary assessment model, so that the pneumonia auxiliary assessment model can obtain the first soft label of the lung medical images, and label the lung medical images based on the value of the first soft label.
[0115] This invention utilizes a conditional generative adversarial network based on multi-expert consensus to automatically generate a large amount of high-fidelity synthetic pneumonia X-ray image data, with each synthetic image containing high-quality soft labels. This significantly reduces the cost and time burden of collecting and labeling real clinical data, providing a data foundation for training high-performance models.
[0116] This invention provides an intelligent annotation method for original medical images with only a single hard label by training a corresponding expert annotation model. This not only effectively reduces the time and effort required for manual data annotation, but also avoids annotation errors caused by visual observation.
[0117] This invention utilizes knowledge distillation technology to fuse independent annotations from multiple experts into a unified consensus soft label, effectively eliminating the subjective bias and random noise inherent in single expert annotations. The conditional generative adversarial network trained with this soft label, and the final evaluation model, learn the "wisdom consensus" of the expert group, thereby significantly improving the accuracy, robustness, and generalization ability of the final pneumonia-assisted evaluation model.
[0118] The expert annotation model training module 102 of this invention can cluster original medical images attributed to the same expert based on their annotation style characteristics, obtaining original medical images with multiple annotation style categories. This results in the trained expert annotation model comprising multiple expert annotation sub-models corresponding to different annotation style categories, allowing the same expert annotation model to output multiple hard labels. This effectively solves the problem of inconsistent annotations caused by shared accounts under the same author (generally due to borrowing or external factors such as shared departments), making the hard labels output by the expert annotation model more consistent with the corresponding expert's clinical experience and style, thereby making the subsequent trained evaluation model more accurate.
[0119] Because the training data itself contains soft labels with probabilistic information, the pneumonia auxiliary assessment model trained in this invention can output pixel-level probability maps, rather than arbitrary binary segmentation results. This allows doctors to intuitively understand the model's confidence level in their judgments, focusing on areas that the model also considers "uncertain," thereby achieving true human-machine collaboration and improving the reliability of diagnostic decisions.
[0120] This invention allows users to specify the attributes of target soft labels (such as disease type and severity) to generate targeted pneumonia image data for rare categories. This on-demand generation capability perfectly balances the training dataset, enabling the model to have excellent recognition capabilities for various types of pneumonia, especially rare types, and avoiding model bias towards the majority class.
[0121] In summary, the embodiments of the present invention, through an innovative and systematic process, collaboratively solve multiple interrelated core technical challenges such as data scarcity, annotation noise, and uncertainty quantification, ultimately producing a technical effect greater than the sum of its parts, and significantly advancing the development of medical imaging AI technology.
[0122] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.
[0123] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0124] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. An AI-based medical imaging X-ray pneumonia auxiliary assessment system, characterized in that, The system includes: The training data acquisition module is used to obtain a raw medical image dataset containing raw medical images with hard labels independently annotated by different experts. The expert annotation model training module is used to input the original medical images from the same expert into the same first training model for training, so as to obtain expert annotation models corresponding to different experts. The training data multi-labeling module is used to input the original medical image with the original hard labels removed into each of the expert labeling models to obtain different hard labels for the original medical image by the expert labeling models. The consensus soft label generation module is used to fuse the various hard labels of the original medical image through knowledge distillation to obtain soft labels that characterize the degree of consensus and disagreement of the various expert annotation models. The image data generation model training module is used to train a first conditional generative adversarial network using the soft labels as conditional information to obtain the image data generation model. The training process includes: having the generator output a synthesized medical image based on the soft labels and a noise vector; having a discriminator receive the original medical image and its corresponding soft labels, and the synthesized medical image and its corresponding soft labels, and perform a real / fake judgment on the input medical image; and iteratively training until the first conditional generative adversarial network converges to obtain the image data generation model. The pneumonia auxiliary assessment model training module is used to input the synthetic training medical images generated by the image data generation model and their corresponding soft labels into the second training model for training, so as to obtain the pneumonia auxiliary assessment model. The pneumonia auxiliary assessment module is used to input the lung medical image to be labeled into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model obtains a first soft label for the lung medical image and labels the lung medical image based on the value of the first soft label.
2. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The expert annotation model training module is specifically used for: The original medical images from the same expert annotation are input into the same first training model, so that the first training model clusters the original medical images according to the annotation style features of the original medical images to obtain the original medical images of multiple annotation style categories; the first training model is trained according to the original medical images of different annotation style categories to obtain expert annotation models; wherein, the same expert annotation model includes multiple expert annotation sub-models corresponding to different annotation style categories, that is, the same expert annotation model outputs a corresponding number of hard labels according to the number of its annotation style categories.
3. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The hard label is a clear annotation of the lesion area in the original medical image, and the soft label is a pixel-level probability map corresponding to the lesion area. The pneumonia auxiliary assessment module is specifically used for: The lung medical image to be labeled is input into the pneumonia auxiliary assessment model so that the pneumonia auxiliary assessment model obtains the pixel-level probability map of the lung medical image, and the lesion area of the lung medical image is labeled based on the probability value corresponding to each pixel in the pixel-level probability map; wherein, when the probability value is greater than a preset threshold, the corresponding pixel is considered to belong to the lesion area.
4. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The consensus soft tag generation module is specifically used for: The distribution of each hard label is calculated, and the corresponding label weight of each hard label is obtained; wherein, the label weight is determined based on the labeling accuracy of the corresponding expert. Based on the distribution results of the hard labels and the label weights, the hard labels are fused to obtain soft labels that characterize the degree of consensus and disagreement of the various expert labeling models.
5. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, In the image data generation model training module, the input of the discriminator is the spliced data of image-soft label pairs, and its task is to determine whether the image and the soft label match and are real.
6. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The image data generation model can receive manually specified target soft tags as conditions and generate synthetic medical images corresponding to the target soft tags, which can be used to target and augment specific types of pneumonia data.
7. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The system also includes a data preprocessing module. The data preprocessing module is used to perform intensity normalization and contrast enhancement operations on the original medical image dataset and / or the synthetic medical images, and to send the processed data to subsequent modules.
8. The AI-based medical imaging X-ray pneumonia auxiliary assessment system according to claim 1, characterized in that, The training module of the pneumonia auxiliary assessment model adopts a course learning strategy. First, it uses the synthetic training medical images for preliminary training, and then gradually introduces the original medical images for refined training.