Nuclear medicine image prediction model construction system

Abstract

The invention provides a nuclear medicine image prediction model construction system, and relates to five modules, namely a data acquisition and preprocessing module, a feature extraction and selection module, a model training and optimization module, a prediction and evaluation module and a feedback and iteration module. According to the system, image features are extracted through deep learning technologies such as a convolutional neural network, feature selection is performed in combination with methods such as principal component analysis (PCA), and a prediction model is trained and optimized by using various machine learning algorithms. The system supports multi-modal image data fusion, realizes comprehensive analysis of multi-source data, has dynamic feedback and adaptive learning functions, can automatically adjust model parameters according to clinical feedback, and improves prediction precision. A user can flexibly configure a model structure, and personalized medical requirements are met. The system remarkably improves the efficiency and accuracy of nuclear medicine image analysis, is widely applicable to various clinical scenes such as tumor diagnosis and curative effect evaluation, and can provide more reliable auxiliary diagnosis support for doctors.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear medicine image analysis and artificial intelligence application, and specifically to a system for building a nuclear medicine image prediction model. Background Art

[0002] With the rapid development of medical imaging technology, nuclear medicine imaging plays a vital role in clinical diagnosis, disease assessment, and treatment monitoring. Nuclear medicine imaging technologies, such as positron emission tomography (PET / CT) and single photon emission computed tomography (SPECT), have been widely used in oncology, cardiovascular disease, neurological diseases and other fields. These imaging technologies can provide rich information at the functional and metabolic levels, providing support for early diagnosis and precise treatment. However, the current analysis and application of nuclear medicine imaging still faces the following technical bottlenecks:

[0003] Traditional nuclear medicine image analysis mainly relies on the experience of doctors, who judge image results through visual observation and quantitative analysis. However, due to the complexity and heterogeneity of image information, different doctors may have large differences in their interpretation of the same image, resulting in a certain degree of subjectivity in the diagnosis results. In addition, when doctors are faced with a large amount of image data, they are prone to fatigue, which in turn affects the accuracy of diagnosis.

[0004] With the advancement of imaging technology, the resolution and quantity of nuclear medicine images have been continuously improved, generating massive amounts of imaging data. However, the existing data processing methods are relatively traditional and it is difficult to process these large-scale data quickly and effectively, resulting in low analysis efficiency. At the same time, the integration and analysis of multimodal imaging data also face technical challenges, such as the registration of different modal images, feature extraction and data fusion problems, which further increase the complexity of diagnosis.

[0005] With the rise of artificial intelligence, especially deep learning technology, some studies have attempted to apply these technologies to nuclear medicine image analysis and train predictive models to assist diagnosis. However, most of these attempts focus on the research of a single model or a specific application scenario, lacking a systematic integration solution, resulting in limited applicability and promotion of the model in actual clinical applications. In addition, these models often lack dynamic feedback and continuous optimization mechanisms, making it difficult to meet changes in clinical needs.

[0006] In practical applications, doctors often need to combine the results of multiple imaging technologies, such as PET, SPECT, CT, and MRI, to comprehensively assess the patient's condition. Existing multimodal data fusion methods are relatively simple, usually processed by linear weighting or direct splicing, ignoring the heterogeneity and interrelationships of each modality data, resulting in low prediction accuracy.

[0007] With the in-depth promotion of the concepts of precision medicine and personalized medicine, how to provide personalized diagnosis and treatment recommendations based on the patient's specific condition, physical signs and imaging data has become an important direction of clinical application. However, most of the existing systems and models are general-purpose, lacking the ability of personalized configuration and flexible adjustment, and it is difficult to provide effective diagnostic support for special diseases or rare diseases.

[0008] Based on the above problems, there is an urgent need for an integrated nuclear medicine imaging prediction model construction system that can automatically process image data, extract features, train and optimize models, and support multimodal data fusion and personalized configuration to meet the application needs of different clinical scenarios. At the same time, by introducing a dynamic feedback mechanism, the prediction ability of the model can be continuously improved, thereby greatly improving the efficiency and accuracy of nuclear medicine imaging analysis and assisting doctors in making more accurate diagnostic decisions. Summary of the invention

[0009] The purpose of the present invention is to provide a nuclear medicine image prediction model construction system to address the deficiencies of the prior art and to solve the problems raised in the background technology.

[0010] To achieve the above object, the present invention provides the following technical solution: a nuclear medicine image prediction model construction system, comprising:

[0011] Data acquisition and preprocessing module: used to obtain raw image data from nuclear medicine imaging equipment and perform standardized processing;

[0012] Feature extraction and selection module: used to extract image features through deep learning networks, and perform dimension reduction and screening on the features;

[0013] Model training and optimization module: used to train models based on collected image data and extracted features, and optimize model performance through cross-validation and hyperparameter tuning;

[0014] Prediction and evaluation module: used to predict the newly input image data and measure the model accuracy through evaluation indicators;

[0015] Feedback and iteration module: used to dynamically adjust model parameters and structure according to clinical feedback to achieve continuous optimization of the model.

[0016] As a preferred technical solution of the present invention, the data acquisition and preprocessing module includes:

[0017] The data denoising unit is used to remove the noise in the original image data. It uses a Gaussian filter and the filtering formula is:

[0018]

[0019] in:

[0020] I filtered (x,y): pixel value of the image at position (x,y) after Gaussian filtering;

[0021] I original (x,y): pixel value of the original image at position (x,y);

[0022] σ: standard deviation of the Gaussian filter, used to control the filtering strength;

[0023] x, y: coordinates of the pixel in the image;

[0024] The data registration unit is used to spatially align multimodal image data. It uses an affine transformation algorithm based on gradient descent. The registration formula is:

[0025] T(x)=Ax+b

[0026] in:

[0027] b: Translation vector of affine transformation.

[0028] T(x): coordinates of the pixel in the image after affine transformation;

[0029] A: Linear transformation matrix of affine transformation.

[0030] As a preferred technical solution of the present invention, the feature extraction and selection module includes: a convolutional neural network for extracting deep features from image data, and the convolution operation formula is:

[0031]

[0032] in:

[0033] h i,j : The value of the convolution output at position (i, j);

[0034] w m,n : weight parameters of the convolution kernel;

[0035] x i,j : Pixel value of the input image at position (i, j);

[0036] b: bias term in convolution operation; k: radius of convolution kernel;

[0037] The feature selection algorithm is used to reduce the dimension and screen the extracted features. The principal component analysis (PCA) is used. The feature dimension reduction formula is:

[0038] Z=XW

[0039] in:

[0040] Z: feature matrix after dimensionality reduction;

[0041] X: original feature matrix;

[0042] W: A matrix composed of feature vectors, used to map high-dimensional features to low-dimensional space.

[0043] As a preferred technical solution of the present invention, the model training and optimization module includes: a multi-model training unit for training using a variety of machine learning models such as random forest, support vector machine, deep neural network, etc.;

[0044] The hyperparameter tuning unit is used to optimize the model's hyperparameters through grid search and cross-validation. The loss function of cross-validation is:

[0045]

[0046] in:

[0047] L: average loss value of the model;

[0048] y i : The true label value of the i-th sample;

[0049] The predicted label value of the i-th sample;

[0050] n: the total number of samples.

[0051] As a preferred technical solution of the present invention, the prediction and evaluation module includes a prediction unit: used to predict the disease risk of the input image data to be analyzed;

[0052] The evaluation unit is used to evaluate the prediction results based on indicators such as ROC curve, confusion matrix, accuracy, recall rate, etc. The area under the ROC curve (AUC) calculation formula is:

[0053] AUC = ∫0 1 TPR(FPR)d(FPR)

[0054] in:

[0055] AUC: Area under the ROC curve;

[0056] TPR: True Rate, i.e. Where TP is the number of true positive cases, FN is the number of false negative cases; FPR: false positive rate, that is Among them, FP is the number of false positives and TN is the number of true negatives.

[0057] As a preferred technical solution of the present invention, the feedback and iteration module includes: a clinical feedback collection unit for receiving feedback information from clinicians;

[0058] The dynamic optimization unit is used to adjust the model parameters and structure according to the feedback information. The loss function of the model adjustment is:

[0059]

[0060] in:

[0061] J(θ): model loss function;

[0062] θ: parameter vector of the model;

[0063] h θ (x (i) ): The predicted value of the i-th sample under parameter θ;

[0064] y (i) : The true value of the i-th sample; m: The number of samples.

[0065] As a preferred technical solution of the present invention, a multimodal data fusion module is used to fuse and analyze different image data to improve the accuracy of the prediction model. The fusion strategy adopts a weighted average method, and the formula is:

[0066] S=α1S PET +α2S CT

[0067] in:

[0068] S: fused image data;

[0069] S PET :PET image data from PET / CT equipment;

[0070] S CT : CT image data from PET / CT equipment;

[0071] α1, α2: weighting coefficients, satisfying α1+α2=1.

[0072] As a preferred technical solution of the present invention, the system is deployed on a cloud platform and supports distributed processing and parallel computing of large-scale image data.

[0073] As a preferred technical solution of the present invention, the system has an adaptive learning function and can dynamically update the model according to newly added image data and feedback.

[0074] As a preferred technical solution of the present invention, the system supports users to customize model structure and parameter settings to adapt to different clinical needs and data types.

[0075] Beneficial effects:

[0076] The nuclear medicine image prediction model construction system of the present invention realizes the automation of the whole process from image data acquisition to prediction result output by integrating five modules: data acquisition and preprocessing, feature extraction and selection, model training and optimization, prediction and evaluation, feedback and iteration. This systematic solution significantly improves the efficiency and accuracy of image analysis and reduces the dependence on manual experience. Through deep learning networks and multimodal data fusion technology, the system can extract efficient features from large-scale nuclear medicine image data and build accurate prediction models, solving the problems of strong subjectivity and low data processing efficiency in traditional image analysis.

[0077] In addition, the system's dynamic feedback mechanism and adaptive learning function ensure the continuous optimization of the model in clinical applications and can respond to the needs of different clinical scenarios in a timely manner. Especially in personalized medicine, the system supports users to flexibly configure the model architecture and parameters according to specific conditions or data types to meet the diagnosis needs of special diseases and rare diseases. Through the fusion analysis of multimodal data and the configuration of personalized models, the system has significant advantages in improving diagnostic accuracy and adaptability, providing important support for precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is a system block diagram of a nuclear medicine image prediction model construction system proposed by the present invention. DETAILED DESCRIPTION

[0079] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] Example 1: Construction of a full-process prediction model based on nuclear medicine imaging

[0081] Data acquisition and preprocessing module

[0082] The system of this embodiment first obtains the patient's image data through nuclear medicine imaging equipment (such as PET / CT). The image data undergoes the following processing steps:

[0083] Data denoising: Gaussian filter is used to denoise the image data. The filtering formula is as follows:

[0084]

[0085] in:

[0086] I filtered(x,y): pixel value of the image at position (x,y) after Gaussian filtering;

[0087] I original (x,y): pixel value of the original image at position (x,y);

[0088] σ: standard deviation of the Gaussian filter, used to control the filtering strength;

[0089] x, y: coordinates of the pixel in the image;

[0090] Data registration: Use affine transformation to spatially register multimodal image data to ensure that image data from different sources are aligned. The formula for affine transformation is:

[0091] T(x)=Ax+b

[0092] in:

[0093] b: Translation vector of affine transformation.

[0094] T(x): coordinates of the pixel in the image after affine transformation;

[0095] A: Linear transformation matrix of affine transformation.

[0096] Feature extraction and selection module: Feature extraction: The system uses convolutional neural network (CNN) to extract deep features from image data. The convolution operation formula is:

[0097]

[0098] in:

[0099] h i,j : The value of the convolution output at position (i, j);

[0100] w m,n : weight parameters of the convolution kernel;

[0101] x i,j : Pixel value of the input image at position (i, j);

[0102] b: bias term in convolution operation; k: radius of convolution kernel;

[0103] Feature selection: The extracted features are reduced in dimension through principal component analysis (PCA), and the formula is as follows:

[0104] Z=XW

[0105] in:

[0106] Z: feature matrix after dimensionality reduction;

[0107] X: original feature matrix;

[0108] W: A matrix composed of feature vectors, used to map high-dimensional features to low-dimensional space.

[0109] Model training and optimization module: Model training: The system uses a variety of machine learning models (such as random forest, support vector machine, deep neural network) for training. The loss function of the model is:

[0110]

[0111] in:

[0112] L: average loss value of the model;

[0113] y i : The true label value of the i-th sample;

[0114] The predicted label value of the i-th sample;

[0115] n: the total number of samples.

[0116] Hyperparameter tuning: The system performs hyperparameter tuning through grid search and cross-validation to ensure the stability and efficiency of the model on different data sets.

[0117] Prediction and evaluation module: Prediction: After preprocessing and feature extraction, the newly input image data is input into the trained model for disease risk prediction.

[0118] Evaluation: The model performance is evaluated by the ROC curve. The area under the ROC curve (AUC) is calculated as:

[0119] AUC = ∫0 1 TPR(FPR)d(FPR)

[0120] in:

[0121] AUC: Area under the ROC curve;

[0122] TPR: True Rate, i.e. Where TP is the number of true positive cases, FN is the number of false negative cases; FPR: false positive rate, that is Among them, FP is the number of false positives and TN is the number of true negatives.

[0123] Feedback and Iteration Module

[0124] Feedback collection: The system receives feedback from clinicians and uses it as new training data to continuously optimize the model.

[0125] Dynamic Optimization: The system uses the following loss function for iterative optimization:

[0126]

[0127] in:

[0128] J(θ): model loss function;

[0129] θ: parameter vector of the model;

[0130] h θ (x (i) ): The predicted value of the i-th sample under parameter θ;

[0131] y (i) : The true value of the i-th sample; m: The number of samples.

[0132] Example 2: Cardiovascular disease diagnosis based on multimodal data fusion

[0133] Data acquisition and preprocessing: The system acquires image data from PET and SPECT devices, representing myocardial perfusion and anatomical structure, respectively. The image data is spatially registered by affine transformation and noise is removed using Gaussian filters.

[0134] The multimodal data fusion module is used to fuse and analyze different image data to improve the accuracy of the prediction model. The fusion strategy adopts the weighted average method, and the formula is:

[0135] S=α1S PET +α2S CT

[0136] in:

[0137] S: fused image data;

[0138] S PET :PET image data from PET / CT equipment;

[0139] S CT : CT image data from PET / CT equipment;

[0140] α1, α2: weighting coefficients, satisfying α1+α2=1.

[0141] Model training and prediction: The system uses support vector machines to train based on the fused data to predict whether a patient is at risk of cardiovascular disease. Hyperparameters are adjusted through cross-validation to ensure the efficiency of the model.

[0142] Evaluation and feedback: The prediction results are evaluated through confusion matrix, accuracy and AUC, and the model is continuously optimized based on doctor feedback to improve the prediction effect.

[0143] Example 3: Rare disease detection based on personalized model configuration

[0144] Custom model configuration: The system allows users to customize the model structure, including the choice of convolutional neural network, recurrent neural network or hybrid model. Users can adjust parameters such as the number of model layers, learning rate and activation function according to clinical needs.

[0145] Data input and feature extraction: The user selects the appropriate image data type, such as PET / CT. The system extracts features through the corresponding network structure and performs feature dimensionality reduction based on PCA.

[0146] Personalized risk prediction: The system uses a customized model to predict the input rare disease data and outputs personalized diagnosis results, which are suitable for testing specific populations or special diseases.

[0147] Feedback and iteration: In combination with clinical feedback, the system dynamically adjusts the model to gradually improve the predictive accuracy in rare disease detection.

[0148] The contents not described in detail in this specification belong to the prior art known to professional and technical personnel in the field. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A nuclear medicine image prediction model construction system, characterized in that: include: Data acquisition and preprocessing module: used to obtain raw image data from nuclear medicine imaging equipment and perform standardized processing; Feature extraction and selection module: used to extract image features through deep learning networks, and perform dimension reduction and screening on the features; Model training and optimization module: used to train models based on collected image data and extracted features, and optimize model performance through cross-validation and hyperparameter tuning; Prediction and evaluation module: used to predict the newly input image data and measure the model accuracy through evaluation indicators; Feedback and iteration module: used to dynamically adjust model parameters and structure according to clinical feedback to achieve continuous optimization of the model.

2. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The data acquisition and preprocessing module includes: The data denoising unit is used to remove the noise in the original image data. It uses a Gaussian filter and the filtering formula is: in: I filtered (x,y): pixel value of the image at position (x,y) after Gaussian filtering; I original (x,y): pixel value of the original image at position (x,y); σ: standard deviation of the Gaussian filter, used to control the filtering strength; x, y: coordinates of the pixel in the image; The data registration unit is used to spatially align multimodal image data. It uses an affine transformation algorithm based on gradient descent. The registration formula is: T(x)=Ax+b in: b: Translation vector of affine transformation. T(x): coordinates of the pixel in the image after affine transformation; A: Linear transformation matrix of affine transformation.

3. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The feature extraction and selection module includes: a convolutional neural network, which is used to extract deep features from image data. The convolution operation formula is: in: h i,j : The value of the convolution output at position (i, j); w m,n : weight parameters of the convolution kernel; x i,j : Pixel value of the input image at position (i, j); b: bias term in convolution operation; k: radius of convolution kernel; The feature selection algorithm is used to reduce the dimension and screen the extracted features. The principal component analysis is used. The feature dimension reduction formula is: Z=XW in: Z: feature matrix after dimensionality reduction; X: original feature matrix; W: A matrix composed of feature vectors, used to map high-dimensional features to low-dimensional space.

4. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The model training and optimization module includes: a multi-model training unit for training using a variety of machine learning models such as random forest, support vector machine, deep neural network, etc.; The hyperparameter tuning unit is used to optimize the model's hyperparameters through grid search and cross-validation. The loss function of cross-validation is: in: L: average loss value of the model; y i : The true label value of the i-th sample; The predicted label value of the i-th sample; n: the total number of samples.

5. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The prediction and evaluation module includes a prediction unit: used to predict the disease risk of the input image data to be analyzed; The evaluation unit is used to evaluate the prediction results based on indicators such as ROC curve, confusion matrix, accuracy, recall rate, etc. The area calculation formula under the ROC curve is: in: AUC: Area under the ROC curve; TPR: True Rate, i.e. TP is the number of true positive cases, FN is the number of false negative cases; FPR is the false positive rate, that is, Among them, FP is the number of false positives and TN is the number of true negatives.

6. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The feedback and iteration module includes: a clinical feedback collection unit for receiving feedback information from clinicians; The dynamic optimization unit is used to adjust the model parameters and structure according to the feedback information. The loss function of the model adjustment is: in: J(θ): model loss function; θ: parameter vector of the model; hθ(x(i)): the predicted value of the i-th sample under parameter θ; y (i) : The true value of the i-th sample; m: The number of samples.

7. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: Also includes: The multimodal data fusion module is used to fuse and analyze different image data to improve the accuracy of the prediction model. The fusion strategy adopts the weighted average method, and the formula is: S=α1S PET +α2S CT in: S: fused image data; S PET : PET image data from PET / CT equipment; S CT : CT image data from PET / CT equipment; α1, α2: weighting coefficients, satisfying α1+α2=1.

8. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The system is deployed on a cloud platform and supports distributed processing and parallel computing of large-scale image data.

9. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The system has adaptive learning capabilities and can dynamically update the model based on new image data and feedback.

10. The nuclear medicine image prediction model construction system according to claim 1, characterized in that: The system supports users to customize model structure and parameter settings to adapt to different clinical needs and data types.

Images