Cross-data heterogeneity intelligent optimization method, system and device based on data quality dynamic decision

Abstract

The invention relates to a cross-data heterogeneity intelligent optimization method, system and device based on data quality dynamic decision. The method mainly comprises the steps of S1, data collection and preprocessing, S2, data analysis and imaging quality division are conducted on preprocessed data, a training data set is formed, S3, a model is constructed and trained, the model comprises a backbone network and a plurality of decoders, the backbone network is used for extracting universal features, the decoders are used for processing specific tasks of different quality data, and the decoder is used for processing the specific tasks of different quality data. The training of the backbone network is joint training by uniformly using a plurality of training data sets with different qualities, and the training of the decoder is one-to-one training for the training data set with single quality, and S4, inputting a to-be-optimized data set into a trained model, and analyzing pixel value distribution and characteristic value distribution of input data in real time to obtain a to-be-optimized data set. And automatically matching the most suitable decoder to carry out output optimization. The method is beneficial to stable and accurate segmentation or classification of the image by a subsequent model.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a method, system and device for intelligent optimization across data heterogeneity based on dynamic decision-making of data quality. Background Art

[0002] In recent years, the development of artificial intelligence (AI) technology has brought revolutionary changes to the field of image analysis, especially the widespread application of deep learning in the field of image classification and segmentation. However, with the complexity of actual application scenarios, especially in the joint modeling of multiple data sets and heterogeneous data processing, existing technologies face obvious challenges in adaptability, generalization ability, and refined processing. Modern image processing systems need to perform a variety of tasks on image data in different fields. These tasks include not only common classification and segmentation, but also the ability of personalized optimization and intelligent task allocation when facing heterogeneous data.

[0003] When traditional single-task image processing models are jointly modeled on multiple data sets, they often face problems such as large differences in cross-domain data features and insufficient model generalization capabilities, which leads to poor performance when processing heterogeneous data sets. In addition, with the gradual application of deep learning in the field of medical imaging, the heterogeneity and acquisition differences of medical data have put forward higher requirements for image analysis technology. Medical images often show great data heterogeneity due to differences in acquisition equipment, regional differences, doctors' operating habits, etc. Figure 3 and Figure 4 This makes it difficult for image segmentation and classification models based on existing technologies to process these heterogeneous data stably and accurately, which in turn affects the accuracy and efficiency of medical diagnosis. Summary of the invention

[0004] Based on this, a cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality is provided. This method is conducive to improving the performance of image classification and segmentation, and is conducive to the subsequent model to process images stably and accurately.

[0005] A cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality, comprising:

[0006] S1. Data collection and preprocessing,

[0007] S2. Perform data analysis and image quality classification on the pre-processed data to form a training data set.

[0008] S3, model construction and training,

[0009] The model consists of a backbone network and multiple decoders. The backbone network is used to extract common features, and the multiple decoders are used to process specific tasks with data of different qualities. The backbone network is trained by jointly training multiple training data sets of different qualities, while the decoder training is one-to-one training for training data sets of a single quality.

[0010] S4. Input the data set to be optimized into the trained model, and use dynamic feature selection technology to automatically match the most suitable decoder for output by real-time analysis of the pixel value distribution and feature value distribution of the input data.

[0011] In one of the embodiments, the step further includes S5 of evaluating the performance of the model and displaying the evaluation result.

[0012] In one of the embodiments, data of different qualities are processed using different post-processing models.

[0013] In one of the embodiments, step S4 also includes a user operation platform, which is used to process abnormal data. When the data in the data set to be optimized cannot match any decoder, the data is abnormal data. When abnormal data appears, the user operation platform generates an alarm signal and retains the abnormal data.

[0014] In one embodiment, in the model, each decoder automatically adjusts weights and task assignments based on its performance in different tasks.

[0015] In one embodiment, in step S1, data enhancement strategy is used to pre-process the data. The data includes a fetal ultrasound data set or a cardiac ultrasound data set.

[0016] In one embodiment, in step S2, data analysis and imaging quality classification includes: performing in-depth analysis on the pre-processed data, and classifying the data quality according to image clarity, contrast, and integrity of diagnostic information.

[0017] In one embodiment, in step S1, the data is medical imaging data.

[0018] In one of the embodiments, step S4 further includes: if the user needs to select a decoder independently, the user independently selects a decoder of appropriate quality.

[0019] In one of the embodiments, the user is further provided with the ability to independently select a decoder of appropriate quality.

[0020] A computer storage medium stores at least one executable instruction, wherein the executable instruction enables a processor to execute operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality.

[0021] A computer device includes: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality.

[0022] The beneficial effects of this application are:

[0023] This application significantly improves the performance of image classification and segmentation through adaptive optimization and multi-task collaboration. With a multi-decoder architecture and dynamic weight allocation mechanism, the model has higher generalization ability, can efficiently adapt to data sets in different fields, and shows excellent stability in handling diverse tasks; through a cross-level feature calculation strategy, the consistency of the feature transfer process is guaranteed, the convergence and robustness of the model are improved, and the risk of overfitting and underfitting is reduced; the dynamic weight allocation mechanism enables the model to adaptively select the best parameters, realize personalized optimization for specific tasks, and significantly improve the task processing accuracy; multi-decoder joint evaluation and dynamic optimization realize collaborative learning between decoders, effectively improving the overall performance and continuous optimization capabilities of the model in classification and segmentation tasks.

[0024] In the method of the present application, the user platform can accurately display which port the input image is output from, and an alarm will be issued if there is no suitable port. In the field of medical imaging, when processing highly heterogeneous data, the present application reduces data errors, improves the standardization of analysis, shortens diagnosis time, and supports remote diagnosis, significantly improving the efficiency and accuracy of diagnosis. The present application has demonstrated significant advantages and excellent performance over existing technologies in terms of adaptability, accuracy, and performance in multi-task environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making of this application.

[0026] Figure 2 A schematic diagram of a model used in the cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making of the present application, wherein the model has a backbone network and multiple decoders.

[0027] Figure 3The following are fetal ultrasound datasets, from top to bottom: European fetal ultrasound dataset, African fetal ultrasound dataset, and Chinese Macao fetal ultrasound dataset. Figure 3 It can be seen that due to the differences in ultrasound instruments and the level of ultrasound physicians, there are differences in the quality of ultrasound data.

[0028] Figure 4 There are two four-chamber heart ultrasound public datasets, Figure 4 It can be seen that due to the differences in ultrasound instruments and the level of ultrasound physicians, there are differences in the quality of ultrasound data. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] like Figure 1 As shown, the first embodiment of the present application provides a cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality, the method comprising:

[0031] S1. Data collection and preprocessing. Data enhancement strategy is used to preprocess data.

[0032] S2. Perform data analysis and imaging quality classification on the preprocessed data to form a training data set, wherein the data analysis and imaging quality classification include: performing in-depth analysis on the preprocessed data, and classifying the data quality according to the clarity, contrast, and integrity of the diagnostic information of the image.

[0033] S3, model construction and training,

[0034] The model consists of a backbone network and multiple decoders. The backbone network is used to extract common features, and the multiple decoders are used to process specific tasks with data of different qualities. The backbone network is trained by jointly training multiple training data sets of different qualities, while the decoder training is one-to-one training for training data sets of a single quality.

[0035] S4. Input the data set to be optimized into the trained model, and use dynamic feature selection technology to automatically match the most suitable decoder for output optimization by analyzing the pixel value distribution and feature value distribution of the input data in real time. Each decoder automatically adjusts the weight and task allocation according to its performance in different tasks. Furthermore, if the user needs to select a decoder independently, the user can independently select a decoder of appropriate quality to use. Furthermore, it also includes a user operation platform, which is used to process abnormal data. When the data in the data set to be optimized cannot be matched with any decoder, the data is abnormal data. When abnormal data appears, the user operation platform generates an alarm signal and retains the abnormal data.

[0036] S5, evaluates the performance of the model and displays the evaluation results.

[0037] It should be noted that the data in step S1 of the present application may be medical imaging data. In step S4, the data output by different decoders are processed using different subsequent processing models. It also includes the user independently selecting a decoder of appropriate quality. For example, there are a total of 3 decoders, and the user only needs to use 2 of them, then the user can select 2 of them for use.

[0038] The following is a detailed introduction to the above-mentioned cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making of the present application.

[0039] 1. Data collection and preprocessing:

[0040] Acquire multiple datasets, such as fetal ultrasound datasets from Europe, Africa, and Macau, China, for downstream classification tasks. Four-chamber heart ultrasound datasets constructed by two research institutions in the United States and France are used for downstream segmentation tasks.

[0041] For image datasets with large heterogeneity from different centers, we first perform preprocessing operations, including image cropping and label unification. In addition, we apply a variety of data enhancement strategies such as random flipping, rotation, and brightness adjustment to increase the diversity and complexity of the data and improve the generalization ability of the model.

[0042] 2. Data analysis and imaging quality classification:

[0043] The preprocessed data is deeply analyzed and combined with image quality assessment, the data is divided into data set quality categories according to image clarity, contrast, and completeness of diagnostic information.

[0044] 3. Model construction and training:

[0045] The model is constructed using a multi-head network framework, including a backbone network and multiple decoders. The backbone network is used to extract common features, and multiple decoders are used to process specific tasks with different data qualities. The choice of the backbone network is relatively broad. In the fetal ultrasound classification task, the vit model is selected as the backbone network; in the cardiac ultrasound segmentation task, the UNet model is selected as the backbone network. For example Figure 2 As shown in the figure, a USHydraNet framework based on multiple decoders is designed to detect the quality of the input image and select the appropriate decoder for corresponding processing according to the quality of the image. Figure 2 NM in D stands for common and universal network model. i,i∈{1，2，3} Represents decoders that process different data qualities, for example, D1 represents low-quality images, D2 represents medium-quality images, and D3 represents high-quality images. Specifically, USHydraNet calculates the data distribution of the image pixel value level before entering the neural network model and the data distribution of the feature level obtained after the neural network, combines the pixel-level and feature-level distributions, and dynamically selects the most suitable decoder to perform the corresponding operation. For example, there are 1500 test images in a data set, the first port (first decoder) outputs 200, the second port (second decoder) outputs 300, and the third port (third decoder) outputs 1000. The first port represents low-quality images, the second port represents medium-quality images, and the third port represents high-quality images. From the above output results, it can be seen that the proportion of high-quality images in this data set is very large. This data set can be defined as a high-quality data set, and the data set can be processed by a subsequent processing model that matches the data set. Similarly, if a data set is a low-quality data set, a dedicated subsequent processing model can be matched to the data set to process the data set. Furthermore, as an improvement, in addition to outputting the specific number of images, different ports can also output specific codes of the images, and the image codes correspond to the images one-to-one. In this way, the images corresponding to the image codes output by different ports can be formed into different data sets, and then matched with different trained subsequent processing models for targeted processing.

[0046] 4. Dynamic feature selection:

[0047] In the inference process of the model, dynamic feature selection technology is used to automatically match the most suitable decoder for output optimization by analyzing the pixel value distribution and feature value distribution of the input data in real time. In addition, in order to avoid the problem that the input data does not belong to the range of any decoder, a user operation platform is added to handle abnormal data.

[0048] 5. Evaluation and results presentation:

[0049] The overall performance of the model is comprehensively evaluated through multi-dimensional performance indicators, including accuracy, precision, recall, F1 score, etc. Visualization tools such as confusion matrix, PR curve, ROC curve, etc. are used to display the performance of the model in each task, combined with the results of dynamic evaluation, to help identify the advantages and bottlenecks of the model. In the result presentation stage, the visualization of feature maps and attention mechanisms is used to conduct an in-depth analysis of the model prediction process to further support the improvement and optimization of the model, especially in high-demand application scenarios such as medical imaging, to ensure the interpretability and diagnostic value of the prediction results.

[0050] The present invention uses a multi-decoder architecture and a dynamic weight allocation mechanism to enable the model to have high adaptability and generalization capabilities, and can effectively handle cross-domain data sets and diverse tasks. At the same time, cross-level feature calculation and dynamic feature selection strategies are adopted to ensure consistency in the feature transfer process, so that the model has stronger robustness and convergence. Through the joint evaluation and dynamic optimization of multiple decoders, the present invention realizes knowledge sharing and adaptive task allocation between decoders, improves the accuracy and efficiency of task collaborative processing, especially in a multi-task environment. The ability of continuous optimization is demonstrated, and it has good functions in the human-computer interaction interface, which can display which port is output. If there is no output, an alarm will be directly triggered to feedback which images cannot be classified into high, medium and low. For example, if the quality is too low or too high, it cannot be output, and an alarm will be automatically triggered at this time.

[0051] In addition, the present invention has shown significant advantages in the field of medical imaging. It reduces the errors caused by data collection differences through a dynamic decision-making mechanism, improves the level of analysis standardization and diagnostic efficiency, supports remote cross-border diagnosis, and helps to efficiently utilize medical resources. The present invention combines cutting-edge technologies such as multi-task learning, dynamic feature selection, and flexible weight optimization, breaking through the bottleneck of existing technologies, and realizing intelligent optimization and refined processing under complex data conditions. It has broad application potential and significant performance improvement.

[0052] Table 1 shows the evaluation results of image data from different regions after applying the method of this application (Hydra-vit, that is, the improvement on the Vit model) and the existing Vit model and the existing Mixup-CIFAR10 model. Among them, European represents European image data, African represents African image data, and Macao China represents Macau image data. It can be seen from the evaluation results that the method of this application can bring significant improvements. Compared with the original Vit model, the method of this application has been improved in multiple indicators of multiple data sets and is better than the existing Mixup-CIFAR10 method.

[0053] Table 1

[0054]

[0055] 1. This application goes beyond the traditional multi-problem parallel processing method and innovatively constructs a parallel multi-decoder architecture based on a combination of cardiac ultrasound and prenatal ultrasound related neural network models. This architecture can not only efficiently perform classification and segmentation tasks on multi-domain image datasets, but also optimizes the adaptation problems of images of different qualities. By introducing a multi-decoder structure, the model can dynamically adjust the processing strategy according to the quality of the input image, thereby effectively improving the performance on low-quality and high-quality images, and significantly improving the adaptability and generalization capabilities of cross-domain tasks.

[0056] 2. Compared with the existing technology, this application divides images of different qualities into low-quality images, medium-quality images and high-quality images by comparing the quality classification of different images, such as clarity, contrast and other related image feature indicators. By using a flexible dynamic port allocation mechanism, the model can automatically select the optimal port to achieve personalized processing of data sets of different qualities after division. The multi-decoder structure further enhances the refined processing capabilities of specific tasks, and can also achieve self-learning and knowledge sharing between decoders, effectively improving the task targeting and performance stability of the model. Effectively improve the task targeting and performance stability of the model.

[0057] 3. In the feature extraction stage, this application adopts a cross-level and cross-dimensional statistical feature strategy, through the step-by-step mean and variance calculation from the input image to the model feature layer, and considers relevant indicators such as contrast and clarity. At the same time, the data set model of the same quality is evaluated to ensure the consistency of data distribution in the feature mapping process. On this basis, this application further introduces a quality scoring mechanism, combined with the distribution distance of image features, to comprehensively evaluate and select the most suitable decoder. This globally consistent feature transfer mechanism enables the model to automatically adapt to different decoders according to image quality and distribution characteristics, thereby improving the feature adaptation capability and ensuring the efficient execution of cross-domain tasks.

[0058] 4. In order to achieve refined task processing across multiple domains and data sets, this application proposes a joint evaluation and progressive optimization strategy for multiple decoders. By real-time monitoring of the performance of each decoder, and optimizing the dynamic output of classification or segmentation ports of different qualities, the classification and segmentation ports include three types of ports: low, medium, and high. In order to further improve the accuracy of task processing, this application sets a port in the classification and segmentation tasks respectively, and each port is further subdivided into three sub-ports, corresponding to images of different qualities (low, medium, and high). The model can automatically adapt to decoders of different qualities, and users can also choose decoders of appropriate quality according to their needs. In terms of image quality assessment, the model scores by comprehensively considering indicators such as clarity, contrast, and signal-to-noise ratio of the heart and fetal organs, and divides the images into three grades of data. Based on these quality scores, this application trains decoders corresponding to data of different qualities to ensure that each decoder can achieve optimal performance within its adapted quality range. This strategy can gradually optimize the performance of each decoder and effectively improve the overall performance of the model in classification and segmentation tasks.

[0059] 5. In response to the heterogeneity and complexity of data collection in the field of medical imaging, this application provides a breakthrough solution for medical image analysis. By combining dynamic weight allocation with the refined processing capabilities of multiple decoders, this application can effectively reduce data errors caused by doctors' personal habits, equipment differences, regional factors, etc., and significantly improve the degree of data standardization. At the same time, this application supports remote cross-border diagnosis, and can greatly reduce the diagnostic burden of doctors based on intelligent task allocation and data adaptive processing.

[0060] The second embodiment of the present application also provides a computer storage medium, in which at least one executable instruction is stored, and the executable instruction enables a processor to execute operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making on data quality.

[0061] The third embodiment of the present application also provides a computer device, including: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making on data quality.

[0062] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A cross-data heterogeneity intelligent optimization method based on dynamic decision-making of data quality, characterized in that: include: S1. Data collection and preprocessing, S2. Perform data analysis and image quality classification on the pre-processed data to form a training data set. S3, model construction and training, The model consists of a backbone network and multiple decoders. The backbone network is used to extract common features, and the multiple decoders are used to process specific tasks with data of different qualities. The backbone network is trained by jointly training multiple training data sets of different qualities, while the decoder training is one-to-one training for training data sets of a single quality. S4. Input the data set to be optimized into the trained model, and use dynamic feature selection technology to automatically match the most suitable decoder for output by real-time analysis of the pixel value distribution and feature value distribution of the input data.

2. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: S5 is also included to evaluate the performance of the model and display the evaluation results.

3. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: In step S1, a data enhancement strategy is used to preprocess the data, where the data includes a fetal ultrasound data set or a cardiac ultrasound data set.

4. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: Step S4 also includes a user operation platform, which is used to process abnormal data. When the data in the data set to be optimized cannot match any decoder, the data is abnormal data. When abnormal data appears, the user operation platform generates an alarm signal and retains the abnormal data.

5. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: In the model, each decoder automatically adjusts its weight and task assignment according to its performance in different tasks.

6. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: In step S2, data analysis and imaging quality classification include: performing in-depth analysis on the pre-processed data, and classifying the data quality according to image clarity, contrast, and integrity of diagnostic information.

7. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: In step S4, data of different qualities are processed using different subsequent processing models.

8. The cross-data heterogeneity intelligent optimization method based on dynamic data quality decision-making according to claim 1 is characterized in that: Step S4 also includes: if the user needs to select a decoder independently, the user independently selects a decoder of appropriate quality.

9. A computer storage medium, characterized in that: The computer storage medium stores at least one executable instruction, which enables the processor to perform operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making on data quality as described in any one of claims 1 to 8.

10. A computer device, characterized in that: include: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus, and the memory is used to store at least one executable instruction, wherein the executable instruction enables the processor to perform operations corresponding to the cross-data heterogeneity intelligent optimization method based on dynamic decision-making on data quality as described in any one of claims 1 to 8.

Images