Newborn lung ultrasonic intelligent analysis system based on big data and machine learning
Through the intelligent analysis system for the neonatal lung ultrasound based on big data and machine learning, the problems of low quality of the neonatal lung ultrasound image and relying on a single data source are solved, and high accuracy and high intelligence evaluation and monitoring are achieved, improving The efficiency and reliability of diagnosis.
Patent Information
- Application Number
- CN202411085208.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-22
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing newborn lung ultrasound images are of low quality, diagnosis depends on a single data source and lacks wide applicability and accuracy, subjectivity and error in manual diagnosis, lack of real-time monitoring and remote diagnosis support, slow diagnostic feedback and system updates, automation and intelligence Inadequate degree of transformation.
Design a newborn lung ultrasound intelligent analysis system based on big data and machine learning, including data acquisition module, machine learning module, result output module, feedback learning module and analysis module, extract key features through deep learning network, conduct model training and Evaluate, generate analytical reports, and enable self-optimization and rapid updates.
It improves the accuracy and intelligence of neonatal lung status assessment, enhances the diversity and applicability of data sources, realizes real-time monitoring and rapid feedback, and improves the reliability and efficiency of diagnosis.
Smart Images

Figure CN119007913A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical devices, and in particular to a neonatal lung ultrasound intelligent analysis system based on big data and machine learning. Background Art
[0002] The coherence of existing neonatal lung ultrasound images will form speckle noise, which will cover the neonatal lung ultrasound images and become the main image feature of the neonatal lung ultrasound images. Therefore, speckle noise will greatly reduce the quality of neonatal lung ultrasound images. At the same time, the quality of ultrasound images of novice ultrasound doctors has always been checked and evaluated by experienced ultrasound doctors, and the evaluation results are inaccurate and inefficient.
[0003] The prior art still has the following technical defects:
[0004] 1. Single data source and limited data volume. Traditional diagnosis of neonatal lung diseases mainly relies on data from a single device or a single hospital, which limits the diversity and magnitude of the data, resulting in the diagnostic results that may lack wide applicability and high accuracy.
[0005] 2. Subjectivity and errors in manual diagnosis. Currently, the diagnosis of neonatal lung diseases relies heavily on the physician’s experience and subjective judgment, which may lead to inconsistencies in diagnosis. Especially in areas where medical resources are relatively scarce, differences in physician experience and level may have a significant impact on the diagnostic results.
[0006] 3. Lack of real-time monitoring and remote diagnosis support. In many cases, real-time and continuous monitoring of newborn lung health cannot be achieved, especially in remote areas. In addition, insufficient remote diagnosis support limits the expansion capacity and timeliness of high-quality medical resources.
[0007] 4. Diagnostic feedback and system updates are slow. Traditional diagnostic tools and methods often lack effective feedback mechanisms and rapid update capabilities, making it difficult for diagnostic tools to be quickly updated and optimized based on the latest clinical data and research results.
[0008] 5. Insufficient automation and intelligence. Although some automated tools have been used to analyze medical images, these tools generally lack optimization for specific applications (such as diagnosis of neonatal lung diseases) and are not intelligent enough to fully utilize the latest machine learning and big data technologies.
[0009] The present invention is made to solve the common problems in this field, such as inaccurate intelligent analysis, low intelligence level, poor timeliness of auxiliary diagnosis and feedback, lack of active monitoring and intervention, few data sources and poor evaluation methods. Summary of the invention
[0010] The purpose of the present invention is to propose a newborn lung ultrasound intelligent analysis system based on big data and machine learning to address the current deficiencies.
[0011] In order to overcome the shortcomings of the prior art, the present invention adopts the following technical solutions:
[0012] A neonatal lung ultrasound intelligent analysis system based on big data and machine learning, the neonatal lung ultrasound intelligent analysis system comprising a server, the neonatal lung ultrasound intelligent analysis system also comprising a data acquisition module, a machine learning module, a result output module, a feedback learning module, and an analysis module, the server being connected to the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module respectively;
[0013] The data acquisition module acquires neonatal lung ultrasound image data collected from different sources, the machine learning module evaluates the neonatal lung ultrasound image data acquired by the data acquisition module to form an evaluation result, the result output module outputs the evaluation result of the machine learning module to the medical staff, the feedback learning module acquires the feedback data of the medical staff and triggers the machine learning module to perform self-optimization, and the analysis module analyzes the evaluation result of the machine learning module and generates an analysis report;
[0014] The machine learning module includes a feature extraction unit and a model training unit. The feature extraction unit extracts key features of rigid neonatal lung health from the data acquisition module, and the model training unit uses the extracted key features to perform model training through a deep learning network to evaluate and determine the differences in neonatal lung ultrasound.
[0015] Optionally, the data acquisition module includes a data acquisition unit and a data transmission unit, the data transmission unit is used for data transmission by devices from different sources, and the data acquisition unit collects data transmitted by the data transmission unit and stores it in a database of the server;
[0016] The data transmission unit includes a data transmission interface and a data storage device. The data transmission interface is connected to at least one external device and transmits the neonatal lung ultrasound image data transmitted by the external device. The data storage device stores the neonatal lung ultrasound image data transmitted by the data transmission interface.
[0017] Optionally, the feature extraction unit extracts key features according to the following steps:
[0018] S1, receiving original neonatal lung ultrasound image data from a data acquisition module, and preprocessing the original neonatal lung ultrasound image data, wherein the preprocessing includes denoising, contrast enhancement and image normalization;
[0019] S2. performing image segmentation on the preprocessed original neonatal lung ultrasound image data, wherein the image segmentation includes segmenting the lung region using a fixed or adaptive threshold value and preliminarily extracting a region of interest;
[0020] S3, extracting features of the initially extracted region of interest through the gray-level co-occurrence matrix;
[0021] S4. Feature selection and optimization
[0022] For each extracted feature X, calculate the correlation coefficient between it and the binary target variable, and based on the calculation results, obtain a list of correlation coefficients between feature X and the binary target variable, sort them according to the absolute value of the correlation coefficient, set a threshold, and select at least two features with the strongest correlation with the target variable as key features;
[0023] S5. Output key features
[0024] The values corresponding to the key features of each sample are arranged in order to form a feature vector R, and the feature vector is output in a suitable format for use by the model training unit.
[0025] Optionally, the model training unit obtains the feature vector R as an input feature of the deep learning model, and outputs a score Z corresponding to the target category according to the predicted target category:
[0026] Z = f(R; θ);
[0027] In the formula, R is the input feature vector, f() is the model function corresponding to the convolutional neural network CNN, and θ is the model parameter, whose value is obtained through the training process;
[0028] The model training unit converts the score Z of the output corresponding to the target category into a probability score for each category through the softmax function:
[0029]
[0030] In the formula, The probability score of the i-th category, z i is the i-th score output by the deep learning model, K is the total number of categories, and e zj It means raising the natural constant e to the jth score z output by the deep learning model. j The power of
[0031] The model training unit selects the category with the largest probability score as the evaluation result of the deep learning model.
[0032] Optionally, the result output module includes a display unit and a verification unit, the verification unit verifies the identity of the medical staff, and the display unit displays the evaluation result to the medical staff who has passed the verification;
[0033] The verification unit includes a verification verifier and a verification code generated by the verification verifier, wherein the verification verifier obtains the identity ID of the medical staff and generates a verification code according to the identity ID;
[0034] If the verification code entered locally by the medical staff is the same as the verification code stored in the cloud, the verification is successful, otherwise the verification fails.
[0035] Optionally, the feedback learning module includes a feedback unit, a scoring unit and a trigger unit. The scoring unit is used by the medical staff to score the evaluation results. The feedback unit transmits the score of the scoring unit to the trigger unit. The trigger unit obtains the average score of the feedback unit and compares it with a set optimization threshold. If the average score is lower than the set optimization threshold, the model optimization is triggered.
[0036] Optionally, the analysis module includes a data receiving unit, a data integration unit and a report generating unit, wherein the data receiving unit is used to receive the optimized evaluation result from the machine learning module, the data integration unit integrates the analysis result with the associated auxiliary data to form an analysis view structure, and the report generating unit is used to automatically generate a detailed analysis report;
[0037] The associated auxiliary data include images of abnormal locations in the neonatal lung ultrasound image data.
[0038] Optionally, the optimization step of the deep learning module includes:
[0039] S11. Add the newly added case data to the training data set to increase the training sample size of the deep learning model;
[0040] S12. Retrain the deep learning model using the expanded training data and update the deep learning model parameters;
[0041] S13, evaluating the performance of the new deep learning model on the validation set, and comparing the differences before and after optimization;
[0042] S14. If the new deep learning model performs better than the old deep learning model, deploy the new deep learning model to replace the old deep learning model.
[0043] Optionally, the target categories include healthy and diseased.
[0044] Optionally, the report generating unit includes a report generator and a natural language generator, wherein the natural language generator converts the analysis result into an easy-to-understand natural language description, and the report generator inserts the description of the natural language generator into a predefined template and format and automatically generates a report.
[0045] The beneficial effects achieved by the present invention are:
[0046] 1. Through the cooperation of the data acquisition module and the machine learning module, the lung status of the newborn can be evaluated, ensuring that the entire system has the advantages of high accuracy of the evaluation process, good auxiliary effect, high timeliness of assistance, high intelligence, and precise evaluation methods;
[0047] 2. Receive multiple data sources through the data transmission module, making the data sources more extensive, ensuring that the entire system has the advantages of wide data sources and high data applicability;
[0048] 3. Through the cooperation between the feedback learning module and the machine learning module, the evaluation means can be updated in time, improving the reliability and accuracy of the evaluation, ensuring that the entire system has the advantages of high analysis accuracy, high intelligence, strong auxiliary diagnosis ability, strong feedback ability, high active monitoring ability and accurate evaluation means;
[0049] 4. Through the cooperation between the analysis module and the machine learning module, the report can be automatically organized and generated, which improves the efficiency and accuracy of the entire report generation, and ensures that the entire system has the advantages of high analysis accuracy, good interactive comfort, strong auxiliary capabilities, high real-time feedback and accurate evaluation methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate the same parts.
[0051] Figure 1 It is an overall block diagram of the present invention.
[0052] Figure 2 It is a schematic diagram of the verification unit, display unit and feedback process of medical staff of the present invention.
[0053] Figure 3 A schematic diagram of a process flow of triggering optimization of a scoring unit, a feedback unit and a trigger unit of the present invention.
[0054] Figure 4It is a flow chart of the data receiving unit, the data integration unit and the report generating unit of the present invention.
[0055] Figure 5 Schematic diagram of examples of reports included with the secondary analysis of the present invention.
[0056] Figure 6 for Figure 5 An example diagram of visualizing the results in . DETAILED DESCRIPTION
[0057] The following is an explanation of the embodiments of the present invention through specific embodiments. Those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual sizes. It is stated in advance. The following embodiments will further explain the relevant technical contents of the present invention in detail, but the disclosed contents are not intended to limit the scope of protection of the present invention.
[0058] Embodiment 1: According to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 and Figure 6 As shown, this embodiment provides a neonatal lung ultrasound intelligent analysis system based on big data and machine learning, the neonatal lung ultrasound intelligent analysis system includes a server, the neonatal lung ultrasound intelligent analysis system also includes a data acquisition module, a machine learning module, a result output module, a feedback learning module, and an analysis module, the server is connected to the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module, respectively, and the intermediate data and process data of the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module are stored in the database of the server for query or call;
[0059] The data acquisition module acquires neonatal lung ultrasound image data collected from different sources, the machine learning module evaluates the neonatal lung ultrasound image data acquired by the data acquisition module to form an evaluation result, the result output module outputs the evaluation result of the machine learning module to the medical staff, the feedback learning module acquires the feedback data of the medical staff and triggers the machine learning module to perform self-optimization, and the analysis module analyzes the evaluation result of the machine learning module and generates an analysis report;
[0060] The neonatal lung ultrasound intelligent analysis system also includes a central processor, which is respectively connected to the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module, and centrally controls the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module based on the central processor, and stores the control data in the database of the server, thereby improving the analysis accuracy and reliability of the entire system;
[0061] The machine learning module includes a feature extraction unit and a model training unit, wherein the feature extraction unit extracts key features of rigid neonatal lung health from the data acquisition module, and the model training unit uses the extracted key features to perform model training through a deep learning network to evaluate and determine differences in neonatal lung ultrasound;
[0062] Optionally, the data acquisition module includes a data acquisition unit and a data transmission unit, the data transmission unit is used for data transmission by devices from different sources, and the data acquisition unit collects data transmitted by the data transmission unit and stores it in a database of the server;
[0063] The data transmission unit includes a data transmission interface and a data storage device, wherein the data transmission interface is connected to at least one external device and transmits the neonatal lung ultrasound image data transmitted by the external device, and the data storage device stores the neonatal lung ultrasound image data transmitted by the data transmission interface;
[0064] The data collection unit collects the data transmitted by the data transmission unit and aggregates the data in a cache database for query or call by the machine learning module;
[0065] Meanwhile, in this embodiment, the data transmission interface is suitable for commercially available or commonly used interfaces, which can cooperate with other medical device data transmission and communication to transmit the collected data through the data transmission interface, thereby realizing the collection of the entire data;
[0066] The data transmission module receives data from multiple sources, making the data sources more extensive, ensuring that the entire system has the advantages of wide data sources and high data applicability;
[0067] Optionally, the feature extraction unit extracts key features according to the following steps:
[0068] S1, receiving original neonatal lung ultrasound image data from a data acquisition module, and preprocessing the original neonatal lung ultrasound image data, wherein the preprocessing includes denoising, contrast enhancement and image normalization;
[0069] Among them, preprocessing the ultrasonic image data is a technical means well known to those skilled in the art. Therefore, in this embodiment, those skilled in the art can refer to relevant technical manuals to learn about the technology, and thus will not be described in detail in this embodiment.
[0070] S2. performing image segmentation on the preprocessed original neonatal lung ultrasound image data, wherein the image segmentation includes segmenting the lung region using a fixed or adaptive threshold value and preliminarily extracting a region of interest;
[0071] In this embodiment, a fixed threshold can be set based on experience, and the portion with a pixel value higher than the threshold is regarded as the foreground, and the portion with a pixel value lower than the threshold is regarded as the background;
[0072] S3, extracting features of the initially extracted region of interest through the gray-level co-occurrence matrix;
[0073] The step of extracting features of the initially extracted region of interest through the gray level co-occurrence matrix includes:
[0074] S31. Constructing gray-level co-occurrence matrix
[0075] Select grayscale level: quantize the grayscale value of the image into L grayscale levels; usually select L = 8, 16, 32, etc.;
[0076] Define direction and distance: define the relative position of pixel pairs, that is, direction (such as 0 degrees, 45 degrees, 90 degrees, 135 degrees) and distance (usually 1);
[0077] Construct a matrix: Construct an L×L gray-level co-occurrence matrix P(d,θ), where d is the distance and θ is the direction; the element P(i,j|d,θ) of the matrix represents the number of times a pixel pair with gray-level values i and j appears at a given distance and direction;
[0078] Each element C(i,j|d,θ) in the matrix represents the frequency of occurrence of pixel pairs with grayscale values i and j in the image at a given distance and direction.
[0079] S32, Normalized Gray Level Co-occurrence Matrix
[0080]
[0081] Where C(i,j|d,θ) is the number of times the pixel pair with grayscale values i and j appears at distance d and direction θ, and P(i,j|d,θ) is the element of the normalized gray-level co-occurrence matrix, which represents the probability of the pixel pair with grayscale values i and j appearing at distance d and direction θ.
[0082] C(i,j|d,θ) is the element of the normalized gray-level co-occurrence matrix, which represents the probability of a pixel pair with gray-level values i and j appearing at distance d and direction θ;
[0083] is the sum of all elements in the original co-occurrence matrix C(i,j|d,θ), that is, the total count of all pixel pairs in the image;
[0084] In this embodiment, in the gray level co-occurrence matrix (GLCM) analysis, the default value of the distance d is usually a small positive integer, which reflects the relative distance between pixel pairs in the image, and a smaller value can better reflect the local texture characteristics.
[0085] Meanwhile, in this embodiment, the default values of the distance d are: 1, 2, 3;
[0086] Among them, d=1: represents the relationship between adjacent pixel pairs. This is the most commonly used distance and can capture the most basic local texture features.
[0087] d=2: indicates that there is one pixel between each pair of pixels, which can capture texture information at a slightly longer distance.
[0088] d=3: indicates that there is a two-pixel gap between pixel pairs, further capturing a wider range of texture features.
[0089] These values are a common choice in the field of image processing and texture analysis because they effectively reflect local structure and texture patterns in the image.
[0090] In this embodiment, an example of a value of the distance d is also provided, specifically:
[0091] 1) Medical Image Analysis
[0092] In medical image analysis, especially in texture analysis of ultrasound images, CT images, and MRI images, smaller distance values such as d = 1 and d = 2 are usually selected to capture the subtle texture features of tissues. These distance values can help identify different types of tissues and lesions.
[0093] 2) Remote sensing image processing
[0094] In remote sensing image processing and classification, smaller distance values (such as d = 1 and d = 2) are usually used to identify local texture features of objects, while larger distance values (such as d = 3) can capture larger-scale texture information and help distinguish different types of surface cover.
[0095] 3) Texture classification and segmentation
[0096] In texture classification and segmentation tasks in computer vision, d = 1 and d = 2 are usually used to capture the local texture patterns of the image. These distance values can provide enough information to distinguish different texture types.
[0097] In short, the value of the distance d needs to be set or taken according to the actual application scenario to meet the actual needs. This is a technical means well known to those skilled in the art. Those skilled in the art can refer to relevant technical manuals to learn about the technology, so it will not be described in detail in this embodiment.
[0098] For each pair of pixels in the image, their grayscale values are calculated; if the grayscale value of a pixel is i, and the grayscale value of its adjacent pixel in the specified direction and distance is j, then C(i,j|d,θ) is increased by 1;
[0099]
[0100] Where, I(x,y): grayscale value of the image at position (x,y), M is the height of the image, indicating the number of pixels in the vertical direction; N is the width of the image, indicating the number of pixels in the horizontal direction, i,j are grayscale values, d is the distance, θ is the direction, Δx, Δy: pixel offset determined by θ and d;
[0101] S33, extracting texture features
[0102] The following texture features are calculated from the normalized gray-level co-occurrence matrix:
[0103] 1) Energy (indicates the uniformity of texture):
[0104]
[0105] 2) Entropy (indicates the complexity of the texture):
[0106]
[0107] 3) Contrast (indicates the local change of gray value distribution):
[0108]
[0109] Where n: represents the grayscale difference, reflecting the difference in grayscale values between pixel pairs, n 2 : Square the grayscale difference, emphasizing that larger grayscale differences contribute more to contrast, P(i,j|d,θ)for|ij|=n: In the grayscale co-occurrence matrix, find the probability of all pixel pairs with a grayscale difference of n, and accumulate these probability values;
[0110] 4) Homogeneity (indicates the smoothness of the texture):
[0111]
[0112] S4. Feature selection and optimization
[0113] For each texture feature X extracted, the correlation coefficient between it and the binary target variable is calculated, and based on the calculation results, a list of correlation coefficients between feature X and the binary target variable is obtained, and the list is sorted according to the absolute value of the correlation coefficient, a threshold is set, and at least two texture features with the strongest correlation with the target variable are selected as key features;
[0114] Among them, the binary target variables include lesions (lesions are 1) and health (health is 0);
[0115] In this embodiment, the texture feature X includes: Energy, Entropy, Contrast, and Homogeneity;
[0116] The correlation coefficient r between the texture feature X and the binary target variable pb :
[0117]
[0118] In the formula, is the mean of feature X when target variable Y is 1 (lesion), X 0 is the mean of feature X when target variable Y is 0 (healthy), S is the standard deviation of feature X, n 1 is the number of samples where the target variable Y is 1, n 0 is the number of samples with the target variable Y being 0, and n is the total number of samples (n = n 1 +n 2 );
[0119] According to the above formula, the size of each correlation coefficient of each texture feature of the neonatal lung ultrasound image data is sorted to form a correlation coefficient list;
[0120] At the same time, the system or operator sets a threshold based on experience, and if at least two texture features exceeding this threshold are considered as key features;
[0121] The threshold is set by the system or the operator according to the actual situation. This is a technical means well known to those skilled in the art. Those skilled in the art can refer to relevant technical manuals to learn about the technology, so it will not be described in detail in this embodiment.
[0122] S5. Output key features
[0123] Arrange the values corresponding to each key feature of each sample in order to form a feature vector R, and output the feature vector in an appropriate format (such as a CSV file, a database record, etc.) for use by the model training unit;
[0124] Optionally, the model training unit obtains the feature vector R as an input feature of the deep learning model, and outputs a score Z corresponding to the target category according to the predicted target category:
[0125] Z = f(R; θ);
[0126] In the formula, R is the input feature vector, f() is the model function corresponding to the deep learning model, and θ is the model parameter, whose value is obtained through the training process;
[0127] In this embodiment, the specific steps for determining the model parameter θ are:
[0128] S41. Data preparation
[0129] Data Collection
[0130] Collect sufficient quantity and diversity of sample data; for neonatal lung ultrasound image analysis, the data includes image data and corresponding health status labels (such as healthy, diseased);
[0131] Data preprocessing
[0132] Image preprocessing: grayscale, denoising, normalization, etc.;
[0133] Feature extraction: extract important texture features;
[0134] S42. Model selection
[0135] Choose the appropriate model:
[0136] Choose a model type that is appropriate for the task, such as convolutional neural network (CNN), random forest, support vector machine (SVM), etc.
[0137] S43. Loss function definition
[0138] Define the loss function
[0139] The loss function is used to measure the difference between the model prediction result and the true label; commonly used loss functions include cross entropy loss (classification task) and mean square error (regression task);
[0140] S44. Model training
[0141] 1. Initialize the model parameters θ. Parameter initialization can use random initialization, He initialization or Xavier initialization;
[0142] 2. Forward Propagation
[0143] Pass the input data through the model to calculate the predicted value:
[0144]
[0145] In the formula, is the predicted value, X is the input feature, and θ is the model parameter;
[0146] 3. Calculate the loss
[0147] The loss function is used to calculate the difference between the predicted value and the true label:
[0148]
[0149] Where Y is the true label;
[0150] 4. Back Propagation
[0151] Calculate the gradient of the loss function with respect to the model parameters;
[0152] The gradient calculation method used in the back propagation process is a technical means well known to those skilled in the art, and thus will not be described in detail in this embodiment;
[0153] 5. Parameter Update
[0154] Update the model parameters θ using an optimization algorithm (such as gradient descent, Adam, etc.):
[0155]
[0156] Where η is the learning rate;
[0157] In this embodiment, the learning rate (η) is a hyperparameter in the deep learning model training, which is usually set by the operator. The learning rate determines the magnitude of the weight update at each iteration of the model. In simple terms, it is a parameter for adjusting the learning speed of the model. The learning rate (η) generally ranges from 0.0001 to 0.1.
[0158] Learning rate is too high: the model is prone to skipping the optimal solution of the loss function during training, resulting in instability and divergence, causing the loss function to suddenly increase or oscillate during training;
[0159] Learning rate is too low: Model training converges very slowly, and it may take a very large number of iterations to get a satisfactory result, resulting in a very slow decrease in the loss function.
[0160] In this embodiment, it can be set based on experience. Common empirical values of η are 0.001, 0.01 or 0.1. Try several common values and observe the training results.
[0161] Of course, the gradual attenuation method is also used for adjustment:
[0162] That is, adjust the learning rate η according to the following formula:
[0163]
[0164] Where η t is the learning rate of the tth iteration, η 0 is the initial learning rate, t is the number of iterations, and Times is the total number of training rounds;
[0165] 6. Iterative training
[0166] The above process is iterated multiple times (multiple epochs) on the entire training data set until the loss function converges or reaches the preset stopping condition, thereby determining the model parameter θ;
[0167] The model training unit converts the score Z of the output corresponding to the target category into a probability score for each category through the softmax function:
[0168]
[0169] In the formula, The probability score of the i-th category, z i is the i-th score output by the deep learning model, K is the total number of categories, and e zj It means raising the natural constant e to the jth score z output by the deep learning model. j The power of, that is, in the Softmax function, we perform an exponential operation on each score zj output by the model, that is, calculate e zj ;
[0170] The model training unit selects the category with the largest probability score as the evaluation result of the deep learning model;
[0171] Optionally, the target categories include healthy and diseased;
[0172] Through the cooperation between the data acquisition module and the machine learning module, the lung status of the newborn can be evaluated, ensuring that the entire system has the advantages of high accuracy of the evaluation process, good auxiliary effect, high timeliness of assistance, high intelligence, and precise evaluation means;
[0173] Optionally, the result output module includes a display unit and a verification unit, the verification unit verifies the identity of the medical staff, and the display unit displays the evaluation result to the medical staff who has passed the verification;
[0174] The verification unit includes a verification verifier and a verification code generated by the verification verifier, wherein the verification verifier obtains the identity ID of the medical staff and generates a verification code according to the identity ID;
[0175] If the verification code entered locally by the medical staff is the same as the verification code stored in the cloud, the verification is successful, otherwise the verification fails;
[0176] Specifically: if the verification code entered locally by the medical staff is different from the verification code stored in the cloud, the verification fails;
[0177] The verification code generated by the verification device is sent to the server and the service terminal of the medical staff, and is input locally through the service terminal;
[0178] The service terminal includes but is not limited to the following: mobile phones or mobile devices, hospital service counters, etc.;
[0179] During the local input process, input is performed through a web page or a terminal service interface, which is a technical means well known to those skilled in the art, and thus will not be described in detail in this embodiment;
[0180] In this embodiment, the verification verifier generates the verification code according to the following formula (assuming that the verification code has k bits):
[0181] Code(i)=((D m ·i+N)+(D p ·i+N)mod10000)mod10;
[0182] In the formula, Code(i) is the value corresponding to the i-th digit of the verification code, D m The last four digits extracted from the medical staff’s ID; D p is the last four digits extracted from the patient's ID, i is the bit index of the verification code (from 1 to k), i∈{1,2,…,k}, and N is the number of historical verifications of the patient;
[0183] In addition, the newly generated verification code must be inconsistent with the historically generated verification code to be valid, ensuring the uniqueness of the verification code;
[0184] The verification unit verifies the identity of medical staff to ensure that the patient's privacy data is more secure, improve the comfort, reliability, uniqueness and security of the entire data interaction, prevent identity forgery and data tampering, and ensure that the entire system has the advantages of high data security, good interaction effect and high intelligence;
[0185] Optionally, the feedback learning module includes a feedback unit, a scoring unit and a trigger unit, wherein the scoring unit is used by the medical staff to score the evaluation results, the feedback unit transmits the score of the scoring unit to the trigger unit, and the trigger unit obtains the average score of the feedback unit and compares it with a set optimization threshold. If the average score is lower than the set optimization threshold, the model optimization is triggered;
[0186] The scoring unit includes a scoring collector and a scoring evaluation subunit, wherein the scoring collector is used by the medical staff to score the evaluation results, and the scoring evaluation subunit obtains the scoring results collected by the scoring collector for evaluation;
[0187] The score collector may be built into the mobile terminal of the medical staff, so that the medical staff can easily check the evaluation results and give scores in time;
[0188] The scoring evaluation subunit obtains the scoring data collected by the scoring collector and calculates the average score S according to the following formula: avg :
[0189]
[0190] In the formula, S i is the score of the i-th evaluation result, and N is the number of scores, that is, the number of feedbacks from medical staff;
[0191] The trigger unit obtains the average score of the feedback unit and compares it with the set optimization threshold. If the average score is lower than the set optimization threshold, the model optimization is triggered;
[0192] The set optimization threshold is set by the system or the operator according to the actual situation. This is a technical means well known to those skilled in the art. Those skilled in the art can refer to relevant technical manuals to learn about the technology, so it will not be described in detail in this embodiment.
[0193] Optionally, the optimization step of the deep learning module includes:
[0194] S11. Add the newly added case data to the training data set to increase the training sample size of the deep learning model;
[0195] S12. Retrain the deep learning model using the expanded training data and update the deep learning model parameters;
[0196] S13, evaluating the performance of the new deep learning model on the validation set, and comparing the differences before and after optimization;
[0197] S14. If the new deep learning model performs better than the old deep learning model, deploy the new deep learning model to replace the old deep learning model;
[0198] Through the cooperation between the feedback learning module and the machine learning module, the evaluation means can be updated in time, the reliability and accuracy of the evaluation can be improved, and the whole system can be guaranteed to have the advantages of high analysis accuracy, high intelligence, strong auxiliary diagnosis ability, strong feedback ability, high active monitoring ability and accurate evaluation means;
[0199] Optionally, the analysis module includes a data receiving unit, a data integration unit and a report generating unit, wherein the data receiving unit is used to receive the optimized evaluation result from the machine learning module, the data integration unit integrates the analysis result with the associated auxiliary data to form an analysis view structure, and the report generating unit is used to automatically generate a detailed analysis report;
[0200] The associated auxiliary data includes an image of an abnormal position of the neonatal lung ultrasound image data;
[0201] Optionally, the report generating unit includes a report generator and a natural language generator, the natural language generator converts the analysis result into an easily understandable natural language description, the report generator inserts the description of the natural language generator into a predefined template and format, and automatically generates a report;
[0202] In this embodiment, the report generator is a software module responsible for inserting analysis results, images, natural language descriptions and other contents into a predefined report template to generate a final analysis report; at the same time, the report generator is used to receive analysis results and natural language descriptions, insert auxiliary data such as images and charts, apply predefined report templates and formats, and automatically generate and export report documents (such as PDF, HTML, etc.);
[0203] The natural language generator is a software module that uses natural language processing (NLP) technology to convert structured data and analysis results into easy-to-understand natural language descriptions. The natural language generator is used to receive structured analysis result data, generate descriptive text, explain the analysis results and findings, and provide detailed diagnosis and suggestion information;
[0204] In this embodiment, the working steps and processes of the report generator and the natural language generator are provided:
[0205] S51. Data reception and preprocessing
[0206] 1) Data receiving unit:
[0207] Receive optimized evaluation results from the machine learning module;
[0208] Receive associated auxiliary data (such as images of abnormal locations);
[0209] 2) Data preprocessing:
[0210] Clean and standardize the received data;
[0211] Ensure data consistency and integrity;
[0212] S52. Data Integration
[0213] Data Integration Unit:
[0214] Integrate the analysis results with the associated auxiliary data to form a complete analysis view structure;
[0215] Steps: 1) Matching the assessment results with the image data;
[0216] 2) Extract images of abnormal locations and annotate relevant information;
[0217] 3) Form a comprehensive analytical view, including images, charts and statistics;
[0218] S53. Natural Language Generation
[0219] Natural Language Generator (NLG): converts structured analysis results into natural language descriptions;
[0220] step:
[0221] 1) Data input: receiving analysis results and auxiliary data;
[0222] 2) Language generation: Use NLG algorithms to generate natural language descriptions, including diagnostic results, abnormal findings, recommendations, and conclusions;
[0223] 3) Text optimization: ensure that the generated language is fluent, easy to read, and conforms to professional medical terminology;
[0224] Example:
[0225] Diagnostic results: The model detected abnormal lesions in the right lower lobe of the lung, which may be neonatal pneumonia;
[0226] Recommendation: Further chest X-ray and blood tests are recommended to confirm the diagnosis;
[0227] S54. Report Generation
[0228] Report Generator: Insert the description, images and analysis results of the natural language generator into the predefined report template to generate the final report; Figure 5 Example of report shown;
[0229] Steps: 1) Template application: load predefined report template;
[0230] 2) Content insertion: inserting natural language descriptions, images, and analysis results into corresponding locations;
[0231] 3) Format adjustment: ensure that the report format meets the standards and all contents are displayed correctly;
[0232] 4) Report export: Generate final report document (PDF, HTML, etc.);
[0233] In this embodiment, an example process is provided, specifically:
[0234] S61, Load Template: Load a report template containing title, chapters, tables and chart placeholders;
[0235] S62, insert description: insert the description of the natural language generator into the diagnostic conclusion, suggestion, etc. part of the report;
[0236] S63, inserting an image: inserting an image of the abnormal position into a corresponding position in the report;
[0237] S64, insert data: insert the data table and chart of the analysis results into the report;
[0238] S65, Format adjustment: adjust the position and size of text and images to ensure a beautiful format;
[0239] S66, Export report: Generate a report in PDF format and save or send it to medical staff;
[0240] The natural language generator is responsible for converting structured data into easy-to-understand text descriptions, and the report generator integrates these descriptions with images and analysis results in predefined templates to ultimately generate professional analysis reports. Through the above steps and processes, detailed medical analysis reports can be generated efficiently and accurately to support medical staff's diagnosis and decision-making;
[0241] In this embodiment, through the mutual cooperation between the analysis module and the machine learning module, the report can be automatically organized and generated, which improves the efficiency and accuracy of the entire report generation and ensures that the entire system has the advantages of high analysis accuracy, good interactive comfort, strong auxiliary capabilities, high real-time feedback and precise evaluation methods.
[0242] Embodiment 2: This embodiment should be understood to include all the features of any of the above embodiments, and further improve upon them. Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5and Figure 6 As shown, the neonatal lung distension ultrasonic intelligent analysis system also includes an alarm module, which obtains the correlation coefficient list of the machine learning module and the correlation coefficient corresponding to the correlation coefficient list, and judges the lung state of the neonate according to the correlation coefficient of the texture feature pair to form a judgment result, and generates an alarm prompt of a corresponding level to prompt the medical staff of the current patient's state;
[0243] The alarm module includes a result collection unit, an early warning evaluation unit and an early warning prompt unit. The result collection unit collects the correlation coefficient list of the machine learning module and the correlation coefficient corresponding to the correlation coefficient list, and transmits them to the early warning evaluation module for evaluation. The early warning evaluation module evaluates the lung state of the newborn according to the correlation coefficient corresponding to the correlation coefficient list to form an evaluation result, and generates an early warning prompt of a corresponding level. The early warning prompt unit prompts the medical staff with the early warning prompt;
[0244] The result collection unit includes an API accessor and a data buffer, wherein the API accessor is used to access the correlation coefficient list of the machine learning module and store the correlation coefficient list in the data buffer;
[0245] The specific access steps of the API accessor include:
[0246] S51. Set API endpoint and request parameters
[0247] Define the API endpoints and request parameters for the machine learning module
[0248] S52: Send HTTP request
[0249] Use the requests library in Python to send an HTTP GET request to access the API endpoint to obtain a list of correlation coefficients;
[0250] S53, Data Cache
[0251] The obtained correlation coefficient list is stored in a data cache; the data cache can be a simple memory cache or a cache database such as Redis;
[0252] By defining the API interface, implementing the API accessor and the data cache, it is possible to obtain the correlation coefficient list from the machine learning module. By accessing and calling the correlation coefficient list, timely acquisition and efficient access to data are ensured, which is helpful for further evaluation and alarm prompt functions.
[0253] The early warning evaluation unit obtains the correlation coefficient list of the machine learning module and calculates the neonatal lung state index State according to the following formula:
[0254]
[0255] In the formula, μ r is the mean of the correlation coefficient, σ r is the standard deviation of the correlation coefficient, r max is the maximum value of the correlation coefficient, r min is the minimum value of the correlation coefficient, ε is the regularization term, and its value is: ε=e -5 ;
[0256] For the mean value μ of the correlation coefficient r Calculate according to the following formula:
[0257]
[0258] In the formula, r pb_i is the i-th correlation coefficient r pb The value of , n is the total number of features in the correlation coefficient list;
[0259] For the standard deviation of the correlation coefficient σ r Calculate according to the following formula:
[0260]
[0261] In the formula, μ r is the mean of the correlation coefficient, r pb_i is the i-th correlation coefficient r pb The value of , n is the total number of features in the correlation coefficient list;
[0262] The maximum value of the correlation coefficient r max The minimum value of the correlation coefficient r min Determined according to the following formula:
[0263] r max =max(r pb_1 ,r pb_2 ,r pb_3 ,...,r pb_n )
[0264] r min =min(r pb_1 ,r pb_2 ,r pb_3 ,...,r pb_n );
[0265] In the formula, r pb_1 is the first correlation coefficient r pbThe value of r pb_2 is the second correlation coefficient r pb The value of r pb_3 is the third correlation coefficient r pb The value of r pb_n is the nth correlation coefficient r pb The value of
[0266] When the State is less than the preset normal threshold, Warn normal When , it means that the newborn's lungs are in normal condition;
[0267] When the State is between the preset normal threshold Warnnormal and the preset warning threshold Warn warning When it is between, it means that a warning is needed and there is a potential risk;
[0268] When State is greater than or equal to the preset warning threshold, warning When it is on, it indicates that there is a serious abnormality and needs to be dealt with immediately;
[0269] Among them, the preset normal threshold Warn normal , Warning threshold warning All of them are set by the system or the administrator / operator according to the actual situation. This is a technical means well known to those skilled in the art. Those skilled in the art can refer to the relevant technical manuals to learn about the technology, so it will not be described in detail in this embodiment.
[0270] Normal Threshold Warn normal : Defines the upper limit of normal lung status of newborns; when the status index is lower than this value, it means that the lung status of the newborn is within the normal range;
[0271] Warning Threshold warning : Defines the lower limit for warning; when the status index is higher than the normal threshold but lower than the warning threshold, it indicates that the newborn may be at potential risk and needs warning;
[0272] Severe abnormality: When the status index is higher than or equal to the warning threshold, it means that there is a serious abnormality in the neonatal lung status and immediate treatment is required;
[0273] The warning prompt unit prompts the medical staff with the warning prompt, and the prompting method includes but is not limited to the following: screen prompt, SMS prompt, email prompt, and in-app notification
[0274] The screen prompt may be implemented by using a pop-up window or a notification; it may also be combined with a sound prompt to draw attention;
[0275] The SMS reminder is implemented by transmitting the early warning reminder information through the pre-bound mobile phone number of the medical staff, so as to achieve accurate delivery and push of the early warning reminder information;
[0276] The email prompt is implemented by sending the warning information to the medical staff's mailbox (the mailbox has been pre-bound) via email, thereby achieving accurate delivery of the warning prompt information;
[0277] The implementation of the in-application notification includes: sending warning information through the notification function of the hospital management system or medical application, that is, using the built-in notification function of the application to send real-time notifications to users;
[0278] The early warning prompt unit conveys the early warning information to medical staff through various means (screen prompts, text messages, emails, in-app notifications, etc.) to ensure that they can take timely measures; these various prompting methods can be selected and combined according to actual needs to provide flexible and efficient early warning prompt services;
[0279] Through the mutual cooperation between the alarm module and the machine learning module, abnormal early warning prompts can be given quickly, and the interaction between medical staff and the system is comfortable, ensuring that the entire system has the advantages of good interaction effect, high interaction comfort, high intelligence, strong alarm capability, good feedback timeliness, strong active feedback and monitoring capabilities, and accurate evaluation methods.
[0280] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.
Claims
1. A newborn lung ultrasound intelligent analysis system based on big data and machine learning, the newborn lung ultrasound intelligent analysis system comprising a server, characterized in that: The neonatal lung ultrasound intelligent analysis system also includes a data acquisition module, a machine learning module, a result output module, a feedback learning module, and an analysis module, and the server is connected to the data acquisition module, the machine learning module, the result output module, the feedback learning module, and the analysis module respectively; The data acquisition module acquires neonatal lung ultrasound image data collected from different sources, the machine learning module evaluates the neonatal lung ultrasound image data acquired by the data acquisition module to form an evaluation result, the result output module outputs the evaluation result of the machine learning module to the medical staff, the feedback learning module acquires the feedback data of the medical staff and triggers the machine learning module to perform self-optimization, and the analysis module analyzes the evaluation result of the machine learning module and generates an analysis report; The machine learning module includes a feature extraction unit and a model training unit. The feature extraction unit extracts key features of rigid neonatal lung health from the data acquisition module, and the model training unit uses the extracted key features to perform model training through a deep learning network to evaluate and determine the differences in neonatal lung ultrasound.
2. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 1 is characterized in that: The data acquisition module includes a data acquisition unit and a data transmission unit. The data transmission unit is used for data transmission by devices from different sources. The data acquisition unit collects the data transmitted by the data transmission unit and stores it in the database of the server. The data transmission unit includes a data transmission interface and a data storage device. The data transmission interface is connected to at least one external device and transmits the neonatal lung ultrasound image data transmitted by the external device. The data storage device stores the neonatal lung ultrasound image data transmitted by the data transmission interface.
3. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 2 is characterized in that: The feature extraction unit extracts key features according to the following steps: S1, receiving original neonatal lung ultrasound image data from a data acquisition module, and preprocessing the original neonatal lung ultrasound image data, wherein the preprocessing includes denoising, contrast enhancement and image normalization; S2, performing image segmentation on the preprocessed original neonatal lung ultrasound image data, wherein the image segmentation includes segmenting the lung area using a fixed or adaptive threshold value and preliminarily extracting a region of interest; S3, extracting features from the initially extracted region of interest using the gray-level co-occurrence matrix; S4. Feature selection and optimization For each extracted feature X, calculate the correlation coefficient between it and the binary target variable, and based on the calculation results, obtain a list of correlation coefficients between feature X and the binary target variable, sort them according to the absolute value of the correlation coefficient, set a threshold, and select at least two features with the strongest correlation with the target variable as key features; S5. Output key features The values corresponding to the key features of each sample are arranged in order to form a feature vector R, and the feature vector is output in a suitable format for use by the model training unit.
4. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 3 is characterized in that: The model training unit obtains the feature vector R as the input feature of the deep learning model, and outputs the score Z of the corresponding target category according to the predicted target category: Z = f(R; θ); In the formula, R is the input feature vector, f() is the model function corresponding to the deep learning model, and θ is the model parameter, whose value is obtained through the training process; The model training unit converts the score Z of the output corresponding to the target category into a probability score for each category through the softmax function: In the formula, The probability score of the i-th category, z i is the i-th score output by the deep learning model, K is the total number of categories, and e zj It means raising the natural constant e to the jth score z output by the deep learning model. j The power of The model training unit selects the category with the largest probability score as the evaluation result of the deep learning model.
5. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 4 is characterized in that: The result output module includes a display unit and a verification unit, wherein the verification unit verifies the identity of the medical staff, and the display unit displays the evaluation result to the medical staff who have passed the verification; The verification unit includes a verification verifier and a verification code generated by the verification verifier, wherein the verification verifier obtains the identity ID of the medical staff and generates a verification code according to the identity ID; If the verification code entered locally by the medical staff is the same as the verification code stored in the cloud, the verification is successful, otherwise the verification fails.
6. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 5 is characterized in that: The feedback learning module includes a feedback unit, a scoring unit and a trigger unit. The scoring unit is used by the medical staff to score the evaluation results. The feedback unit transmits the score of the scoring unit to the trigger unit. The trigger unit obtains the average score of the feedback unit and compares it with a set optimization threshold. If the average score is lower than the set optimization threshold, the model optimization is triggered.
7. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 6 is characterized in that: The analysis module includes a data receiving unit, a data integration unit and a report generating unit, wherein the data receiving unit is used to receive the optimized evaluation result from the machine learning module, the data integration unit integrates the analysis result with the associated auxiliary data to form an analysis view structure, and the report generating unit is used to automatically generate a detailed analysis report; The associated auxiliary data include images of abnormal locations in the neonatal lung ultrasound image data.
8. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 7 is characterized in that: The optimization steps of the deep learning module include: S11. Add the newly added case data to the training data set to increase the training sample size of the deep learning model; S12. Retrain the deep learning model using the expanded training data and update the deep learning model parameters; S13, evaluating the performance of the new deep learning model on the validation set, and comparing the differences before and after optimization; S14. If the new deep learning model performs better than the old deep learning model, deploy the new deep learning model to replace the old deep learning model.
9. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 8, characterized in that: The target categories include healthy and diseased.
10. The neonatal lung ultrasound intelligent analysis system based on big data and machine learning according to claim 9, characterized in that: The report generating unit comprises a report generator and a natural language generator, wherein the natural language generator converts the analysis result into an easily understandable natural language description, and the report generator inserts the description of the natural language generator into a predefined template and format and automatically generates a report.
Citation Information
Patent Citations
Tumor identification method and system based on artificial intelligence
CN116862840A
Thoracic surgery image analysis method based on big data
CN117994238A
Medical image quality management system based on artificial intelligence
CN118098518A
Method and Apparatus of Intelligent Analysis for Liver Tumor
US20240225588A1
Cited By
Deep learning-based children pneumonia lung image analysis and evaluation system and method
CN119991669A
