Cerebral disease auxiliary diagnosis method and system based on medical image

Through the integrated learning method combined with the characteristics of structural and functional NMR images, the Pearson coefficient and the maximum mutual information coefficient feature vector are used to solve the problem of insufficient recognition ability under the traditional diagnosis of small sample data, and achieve a more accurate and objective brain disease diagnosis.

CN119067915APending Publication Date: 2024-12-03SHENZHEN YUNYING MEDICAL TECH CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410954863.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Traditional medical images have a lack of objectivity and accuracy in diagnosis of brain diseases, especially in small sample data, which makes it difficult to make comprehensive judgments based on multiple types of medical images.

Method used

Using an integrated learning method based on medical images, combining structural NMR images and timing functional NMR images, feature extraction and model training are performed through Pearson coefficient and maximum mutual information coefficient feature vector, and multiple basic learners are integrated in parallel using the AdaBoost method. Improve the recognition ability of diagnostic models.

Benefits of technology

It improves the accuracy and recognition ability of brain disease auxiliary diagnosis, can effectively classify diseases under small sample data, and enhances the objectivity and comprehensive analysis ability of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067915A_ABST
    Figure CN119067915A_ABST
Patent Text Reader

Abstract

The invention discloses a brain disease auxiliary diagnosis method and system based on medical images, and relates to the field of medical image processing. A brain disease auxiliary diagnosis system based on medical images comprises an image acquisition module, a data preprocessing module, a feature construction module, a disease classification module and an auxiliary diagnosis report module. According to the method, an integrated learning method is used, three base learners, a structural image recognition model based on a neural network and two coefficient recognition models based on SVM are integrated in parallel, and the structural image recognition model analyzes the brain structure direction through a structural nuclear magnetic resonance image; the coefficient identification model analyzes the brain function characteristic coefficient direction through the time sequence function nuclear magnetic resonance image, the obtained integrated disease auxiliary diagnosis model can comprehensively analyze various medical images, and the accuracy of brain disease auxiliary diagnosis is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a brain disease auxiliary diagnosis method and system based on medical images. Background Art

[0002] As the average human life expectancy increases, the incidence of brain diseases is gradually increasing. The diagnosis of brain diseases is assisted by various medical data, among which medical imaging plays an extremely important role in the diagnosis of brain diseases. Medical imaging is a vital tool in medical diagnosis and treatment. Through different types of imaging technologies, medical images of the brain in various directions are obtained, observed and analyzed, and combined with the structure and function of the patient's brain to help make accurate diagnosis and treatment plans.

[0003] Traditionally, the diagnosis of brain diseases in medical images is usually done manually through empirical judgment. This method lacks objectivity and cannot be combined with multiple types of medical images for comprehensive judgment, resulting in inaccurate diagnosis results. The introduction of artificial intelligence to assist in the diagnosis of brain diseases can combine multiple types of medical data for judgment, but due to the small number of samples in medical image data, how to improve the recognition ability of classification models on small sample data needs to be improved. Summary of the invention

[0004] A brain disease auxiliary diagnosis method based on medical imaging, comprising: S1. Acquire structural MRI images and sequential functional MRI images; S2. Preprocessing the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, segmenting the time series functional MRI images into time slices of fixed length using a sliding window method, and obtaining a 2D data matrix corresponding to each time slice; S3. Pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). Use the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice to calculate the Pearson coefficient weight W of each brain region pair (i, j) ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; S4. Send the structural MRI image, the final eigenvector of the Pearson coefficient, and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; S5. Visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output the visualization results.

[0005] As a preferred technical solution of the present invention, the integrated disease auxiliary diagnosis model is divided into a data processing layer, a base learner recognition layer, and an integrated output layer; The data processing layer is used to preprocess the structural nuclear magnetic resonance image to obtain the structural image to be identified, perform feature selection on the final feature vector of the Pearson coefficient and the final feature vector of the maximum mutual information coefficient, and obtain the feature vector of the Pearson coefficient to be identified and the feature vector of the maximum mutual information coefficient to be identified; The base learner recognition layer includes three base learners, namely, the structural image recognition model, the Pearson coefficient recognition model and the maximum mutual information coefficient recognition model; the structural image recognition model is used to perform disease recognition on the structural image to be recognized, and outputs the brain disease label and the corresponding category probability; the Pearson coefficient recognition model is used to perform disease recognition on the Pearson coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; the maximum mutual information coefficient recognition model is used to perform disease recognition on the maximum mutual information coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; The integrated output layer is used to identify the outputs of the three base learners and output disease auxiliary diagnosis labels and corresponding category probabilities.

[0006] As a preferred technical solution of the present invention, an integrated disease-assisted diagnosis model is constructed, and the specific steps are as follows: A1. Collect structural MRI images and time-series functional MRI images of several samples, label the samples with brain disease labels, extract the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient corresponding to the time-series functional MRI images of each sample, and perform feature selection on the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient using the KS test; The structural MRI images of the samples are preprocessed and combined with brain disease labels to form a structural image recognition data set, and the structural image recognition data set is divided into a structural image recognition training set and a structural image recognition test set; The final feature vector of the Pearson coefficient of the sample after feature selection and the brain disease label are used to form a Pearson coefficient feature recognition data set, and the Pearson coefficient feature recognition data set is divided into a Pearson coefficient feature recognition training set and a Pearson coefficient feature recognition test set by sampling with replacement; The final feature vector of the maximum mutual information coefficient of the sample after feature selection and the brain disease label are used to form a maximum mutual information coefficient feature recognition data set, and the maximum mutual information coefficient feature recognition data set is divided into a maximum mutual information coefficient feature recognition training set and a maximum mutual information coefficient feature recognition test set by sampling with replacement; A2. Send the structural image recognition training set to the initial structural image recognition model based on the convolutional neural network for training, with the brain disease label as the target, to obtain the training structural image recognition model; send the structural image recognition test set to the training structural image recognition model for accuracy evaluation, with the brain disease label as the target, adjust the hyperparameters of the training structural image recognition model to obtain the structural image recognition model; The Pearson coefficient recognition model was obtained by combining the AdaBoost method with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The maximum mutual information coefficient recognition model was obtained by combining the AdaBoost method with the maximum mutual information coefficient feature recognition training set and the maximum mutual information coefficient feature recognition test set. A3. The structural image recognition model, the Pearson coefficient recognition model, and the maximum mutual information coefficient recognition model are used as base learners of the integrated disease-assisted diagnosis model; the three base learners are integrated in parallel to obtain the brain disease labels and corresponding category probabilities output by the three base learners on their respective training sets and test sets, and the brain disease labels and corresponding category probabilities output by the three base learners and the sample-labeled brain disease labels are combined into new recognition samples, all the recognition samples are combined into an integrated learning data set, and the integrated learning data set is divided into an integrated learning training set and an integrated learning test set; The ensemble learning training set is sent to the initial ensemble learning model based on the multi-layer perceptron for training, and the sample-labeled brain disease labels are used as the target to train the initial ensemble disease-assisted diagnosis model; the ensemble learning test set is sent to the initial ensemble disease-assisted diagnosis model for accuracy evaluation, and the sample-labeled brain disease labels are used as the target to adjust the hyperparameters of the initial ensemble disease-assisted diagnosis model to obtain the ensemble disease-assisted diagnosis model.

[0007] As a preferred technical solution of the present invention, the Pearson coefficient recognition model is obtained by training the AdaBoost method in combination with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The specific steps are as follows: B1. Send the samples in the Pearson coefficient feature recognition training set as training samples to the weak learner a based on the SVM model for training, and train the weak learner a with the brain disease label as the target. Calculate the proportion of training samples that are misclassified by the weak learner a as the error rate a; Use the Pearson coefficient feature recognition test set to calculate the accuracy a of the weak learner a, and the accuracy is based on the brain disease label as the target; B2. Adjust the weights of the training samples according to the error rate of the weak learner a, and increase the weights of the misclassified samples by half the error rate a; B3. Send the Pearson coefficient feature recognition training set after adjusting the sample weight to the initial weak learner b based on the SVM model for training, and train the weak learner b with the brain disease label as the target. Use the Pearson coefficient feature recognition test set to calculate the accuracy of the trained weak learner b, and compare whether the improvement rate of the accuracy of the trained weak learner b to the accuracy a is greater than the preset value. If it is greater than the preset value, the trained weak learner b is used as the weak learner b. If it is less than the preset value, repeat steps B2-B3 to obtain the weak learner b; calculate the proportion of samples misclassified by the weak learner b as the error rate b, and use the Pearson coefficient feature recognition test set to calculate the accuracy b of the weak learner b; B4. Repeat step B3 to obtain weak learner c and the accuracy c of weak learner c; add weights to the outputs of the weak learners according to the accuracy of the three weak learners; integrate the three weak learners into a Pearson coefficient recognition model, and calculate the weighted brain disease labels and weighted corresponding category probabilities output by the three weak learners as the output of the Pearson coefficient recognition model; In addition, the maximum mutual information coefficient identification model is obtained in the same way as above.

[0008] As a preferred technical solution of the present invention, the KS test is used to perform feature selection on the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient. The specific steps of the KS test are: calculating the cumulative distribution function CDF of each feature value and brain disease label, calculating the maximum distance between the feature CDF and the category CDF, obtaining the D value of the feature, using a normal distribution to determine the distribution of the D value of the feature, calculating the P value by finding the probability of the D value of the feature appearing at the corresponding position (or a larger value) in its distribution, sorting the features according to the size of the P value, and selecting features with a P value less than a set threshold for subsequent analysis.

[0009] As a preferred technical solution of the present invention, the time series functional magnetic resonance imaging is preprocessed, and the specific steps of the preprocessing are: C1. Delete the first three functional MRI images in the time series functional MRI images; C2. Perform time slice correction, head motion correction and spatial standardization on the remaining time series functional MRI images to obtain standard functional MRI images; C3. Obtain the global blood oxygen level-dependent time series signal for each standard functional MRI image; C4. Based on the automatic anatomical labeling template, the global blood oxygen level dependent time series signal is divided into the average resting time series signals of multiple brain regions. The average resting time series signal of each brain region is processed using a bandpass filter with a frequency band of 0.01-0.08 Hz to obtain the blood oxygen level dependent signal value of each brain region, and a two-dimensional data matrix is ​​obtained, which stores the blood oxygen level dependent signal values ​​of the brain regions.

[0010] As a preferred technical solution of the present invention, a sliding window method is used to segment the time series functional magnetic resonance imaging into time slices of fixed length, and the formula for segmenting the time slices is: ; In the formula, num window is the number of time slices, window size For the length of each window, window step is the interval between two consecutive windows, num volume is the time series functional MRI image and is the total number of images.

[0011] As a preferred technical solution of the present invention, all Pearson coefficient weights W are used. ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) respectively calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient, and the steps of the calculation method are: D1. Obtain the Pearson coefficient weight W for each brain region pair (i, j) at each time slice ij (PCC); D2. All Pearson coefficient weights W on each time slice ij (PCC) is concatenated into a size of (num ROI ,num ROI ) and is called the brain functional connection network correlation matrix corresponding to the time slice, num ROI is the number of brain regions; D3. Reduce the dimension of the brain functional connection network correlation matrix on each time slice to a length of (num ROI *num ROI ), recorded as the feature vector of the corresponding time slice; D4. Take the average of the eigenvectors on all time slices as the final eigenvector of the Pearson coefficient; Furthermore, the final eigenvector of the maximum mutual information coefficient is obtained by the same steps.

[0012] A brain disease auxiliary diagnosis system based on medical images, comprising: An image acquisition module, used to acquire structural MRI images and time-series functional MRI images; A data preprocessing module is used to preprocess the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, and to divide the time series functional MRI images into time slices of fixed length using a sliding window method to obtain a 2D data matrix corresponding to each time slice; The feature construction module is used to pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). The Pearson coefficient weight W of each brain region pair (i, j) is calculated using the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice. ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; The disease classification module is used to input the structural MRI image, the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; The auxiliary diagnosis report module is used to visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output visualization results.

[0013] The present invention has the following advantages: 1. The present invention integrates three base learners in parallel by using an integrated learning method, including one structural image recognition model based on a neural network and two coefficient recognition models based on SVM. The structural image recognition model analyzes the direction of the brain structure through structural magnetic resonance imaging, and the coefficient recognition model analyzes the direction of the coefficient of brain function characteristics through time-series functional magnetic resonance imaging. The obtained integrated disease auxiliary diagnosis model can use a variety of medical images for comprehensive analysis, thereby improving the accuracy of auxiliary diagnosis of brain diseases.

[0014] 2. The present invention obtains a large number of features in the same sample by acquiring the Pearson coefficient features and the maximum mutual information coefficient features of the time-series functional magnetic resonance images, and selects effective features through the KS test method. By combining the two types of features, the recognition ability of the integrated disease-assisted diagnosis model on small sample data is improved.

[0015] 3. The present invention trains the coefficient recognition model by using the AdaBoost method and uses multiple base learners for serial learning, so that the coefficient recognition model can learn the weights of the base learners in the misclassified samples, thereby improving the recognition accuracy of the coefficient recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the structure of a medical imaging-based brain disease auxiliary diagnosis system used in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0018] Embodiment 1, a method for assisting diagnosis of brain diseases based on medical imaging, comprising: S1. Acquire structural MRI images and sequential functional MRI images; Structural MRI (sMRI) usually acquires one image, while sequential functional MRI (fMRI) acquires a set of sequential images through continuous scanning. sMRI has high spatial resolution and can reflect the structural characteristics of brain tissue. fMRI has lower spatial resolution but has temporal resolution that sMRI does not have. Combining sMRI and fMRI analysis can make a more detailed diagnosis based on the brain structure and function. S2. Preprocessing the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, segmenting the time series functional MRI images into time slices of fixed length using a sliding window method, and obtaining a 2D data matrix corresponding to each time slice; When the brain is working, the relevant neurons will consume oxygen and glucose, and the blood flow in the corresponding functional brain area will increase, making the oxygenated hemoglobin in the area higher than the deoxygenated hemoglobin, thereby increasing the intensity of the local cortical signal. This phenomenon is called blood oxygen level dependence. Over a continuous period of time, the change pattern and intensity of blood oxygen level dependence between different brain regions are different, which is a unique feature. Patients with brain diseases have impaired function of some or most brain regions, which affects their blood oxygen level dependence. Therefore, the characteristics reflected by blood oxygen level dependence are used to classify and diagnose brain diseases; The brain is considered as a graph, and several brain regions (ROIs) in the brain correspond to the nodes in the graph, and the presence or absence of edges between nodes and the weights on the edges correspond to the correlation between brain regions; the number of brain regions (ROIs) is determined by the template used when performing brain region segmentation on fMRI data during data preprocessing; the graph obtained in this way is called a brain functional connection network; In step S2, the time series functional magnetic resonance imaging is segmented into time slices of fixed length using a sliding window method. The formula for segmenting the time slices is: ; In the formula, num window is the number of time slices, window size For the length of each window, window step is the interval between two consecutive windows, num volume is the time-series functional MRI image is the total number of images; The graph constructed by analyzing all time-series functional MRI images is a static brain functional connection network, which is a relatively rough processing and may dilute the obvious fluctuations in a certain time period; dividing the time period and constructing a static brain functional connection network on each time slice will be more conducive to highlighting the fluctuation pattern of the data on each time slice; using the time-series functional MRI images on each time slice to construct a static brain functional connection network, and analyzing multiple static brain functional connection networks, the multiple graphs obtained in this way are called dynamic brain functional connection networks; In step S2, the time series functional magnetic resonance imaging is preprocessed, and the specific steps of the preprocessing are: C1. Delete the first three functional MRI images in the time series functional MRI images in order to eliminate the interference caused by the instability of the samples in the early stage of scanning; C2. Perform time slice correction, head motion correction and spatial standardization on the remaining time series functional MRI images to obtain standard functional MRI images; Time layer correction includes slice time correction and scanning sequence correction to eliminate errors caused by time delay or time drift during scanning; Head motion correction includes voxel resampling and spatial interpolation to correct the effects caused by the subject's head movement during scanning; Spatial normalization maps the fMRI of the subject to a standard spatial template to facilitate comparison and merging between different individuals and studies; C3. Obtain the global blood oxygen level-dependent time series signal for each standard functional MRI image; C4. Based on the automatic anatomical labeling template, the global blood oxygen level dependent time series signal is divided into average resting time series signals of multiple brain regions, and the average resting time series signal of each brain region is processed using a bandpass filter with a frequency band of 0.01-0.08 Hz to obtain the blood oxygen level dependent signal value of each brain region, and a 2D data matrix is ​​obtained, in which the blood oxygen level dependent signal values ​​of the brain regions are stored; The size of the resulting 2D data matrix is ​​(number of images, number of brain regions); S3. Pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). Use the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice to calculate the Pearson coefficient weight W of each brain region pair (i, j) ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; The size of the 2D data matrix corresponding to each time slice is (number of images corresponding to the slice, number of brain regions); The brain functional connection network constructed based on the Pearson coefficient (PCC) has reliable performance in the diagnosis of various brain diseases. However, since the Pearson coefficient is only applicable to data with linear relationships, it is used in combination with the maximum mutual information coefficient (MIC). The maximum mutual information coefficient has excellent performance on both linear and nonlinear relationship data, and it can not only construct an effective brain functional connection network, but also has been successfully applied to the diagnosis of brain mental diseases. In step S3, all Pearson coefficient weights W are used. ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) respectively calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient, and the steps of the calculation method are: D1. Obtain the Pearson coefficient weight W for each brain region pair (i, j) at each time slice ij (PCC); D2. All Pearson coefficient weights W on each time slice ij (PCC) is concatenated into a size of (num ROI ,num ROI ) and is called the brain functional connection network correlation matrix corresponding to the time slice, num ROI is the number of brain regions; D3. Reduce the dimension of the brain functional connection network correlation matrix on each time slice to a length of (num ROI *num ROI ), recorded as the feature vector of the corresponding time slice; D4. Take the average of the eigenvectors on all time slices as the final eigenvector of the Pearson coefficient; Furthermore, the final eigenvector with the maximum mutual information coefficient was obtained by the same steps; S4. Send the structural MRI image, the final eigenvector of the Pearson coefficient, and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; The integrated disease-assisted diagnosis model is divided into a data processing layer, a base learner recognition layer, and an integrated output layer; The data processing layer is used to preprocess the structural nuclear magnetic resonance image to obtain the structural image to be identified, perform feature selection on the final feature vector of the Pearson coefficient and the final feature vector of the maximum mutual information coefficient, and obtain the feature vector of the Pearson coefficient to be identified and the feature vector of the maximum mutual information coefficient to be identified; The base learner recognition layer includes three base learners, namely, the structural image recognition model, the Pearson coefficient recognition model and the maximum mutual information coefficient recognition model; the structural image recognition model is used to perform disease recognition on the structural image to be recognized, and outputs the brain disease label and the corresponding category probability; the Pearson coefficient recognition model is used to perform disease recognition on the Pearson coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; the maximum mutual information coefficient recognition model is used to perform disease recognition on the maximum mutual information coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; The integrated output layer is used to identify the outputs of the three base learners and output disease auxiliary diagnosis labels and corresponding category probabilities; The structural image recognition model is built on the basis of convolutional neural network. The convolutional neural network can automatically extract the image features of the structural image and build a deep network structure. For the structural image, the spatial structure information of the image can be maintained, and the recognition rate is relatively high. The Pearson coefficient identification model and the maximum mutual information coefficient identification model are based on the SVM model, which is an unstable algorithm. Since the amount of medical imaging data is small, the use of an unstable algorithm to perform a serial integration model can solve the overfitting problem of the model. The integrated disease-assisted diagnosis model is built based on a multi-layer perceptron, which has a simple structure and is easy to construct to correspond to the outputs of different base learners in multiple hidden layers; The probability of the corresponding category is obtained in the neural network by using the softmax activation function in the last layer of the network, and in the SVM model by using the LibSVM model; Construct an integrated disease-assisted diagnosis model. The specific steps are: A1. Collect structural MRI images and time-series functional MRI images of several samples, label the samples with brain disease labels, extract the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient corresponding to the time-series functional MRI images of each sample, and perform feature selection on the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient using the KS test; The structural MRI images of the samples are preprocessed and combined with brain disease labels to form a structural image recognition data set, and the structural image recognition data set is divided into a structural image recognition training set and a structural image recognition test set; The final feature vector of the Pearson coefficient of the sample after feature selection and the brain disease label are used to form a Pearson coefficient feature recognition data set, and the Pearson coefficient feature recognition data set is divided into a Pearson coefficient feature recognition training set and a Pearson coefficient feature recognition test set by sampling with replacement; The final feature vector of the maximum mutual information coefficient of the sample after feature selection and the brain disease label are used to form a maximum mutual information coefficient feature recognition data set, and the maximum mutual information coefficient feature recognition data set is divided into a maximum mutual information coefficient feature recognition training set and a maximum mutual information coefficient feature recognition test set by sampling with replacement; The final feature vector of the constructed Pearson coefficient and the final feature vector of the maximum mutual information coefficient contain a large number of features, which are not necessarily effective. Too many invalid features reduce the recognition rate of the model. Feature selection before training can improve the recognition rate and training efficiency of the model. The purpose of sampling with replacement is to solve the problem of overfitting that the model is prone to on small sample data; A2. Send the structural image recognition training set to the initial structural image recognition model based on the convolutional neural network for training, with the brain disease label as the target, to obtain the training structural image recognition model; send the structural image recognition test set to the training structural image recognition model for accuracy evaluation, with the brain disease label as the target, adjust the hyperparameters of the training structural image recognition model to obtain the structural image recognition model; The Pearson coefficient recognition model was obtained by combining the AdaBoost method with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The maximum mutual information coefficient recognition model was obtained by combining the AdaBoost method with the maximum mutual information coefficient feature recognition training set and the maximum mutual information coefficient feature recognition test set. AdaBoost is an ensemble learning method that can be used to combine the advantages of multiple trained weak learners to integrate a high recognition rate ensemble model that learns the weights of error samples. This method can achieve a higher recognition rate on small sample data and improve the utilization rate of small sample data. A3. The structural image recognition model, the Pearson coefficient recognition model, and the maximum mutual information coefficient recognition model are used as base learners of the integrated disease-assisted diagnosis model; the three base learners are integrated in parallel to obtain the brain disease labels and corresponding category probabilities output by the three base learners on their respective training sets and test sets, and the brain disease labels and corresponding category probabilities output by the three base learners and the sample-labeled brain disease labels are combined into new recognition samples, all the recognition samples are combined into an integrated learning data set, and the integrated learning data set is divided into an integrated learning training set and an integrated learning test set; The ensemble learning training set is sent to the initial ensemble learning model based on the multi-layer perceptron for training, and the sample-labeled brain disease labels are used as the target to train the initial ensemble disease-assisted diagnosis model; the ensemble learning test set is sent to the initial ensemble disease-assisted diagnosis model for accuracy evaluation, and the sample-labeled brain disease labels are used as the target to adjust the hyperparameters of the initial ensemble disease-assisted diagnosis model to obtain the ensemble disease-assisted diagnosis model.

[0019] In step A2, the Pearson coefficient recognition model is obtained by using the AdaBoost method combined with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The specific steps are as follows: B1. Send the samples in the Pearson coefficient feature recognition training set as training samples to the weak learner a based on the SVM model for training, and train the weak learner a with the brain disease label as the target. Calculate the proportion of training samples that are misclassified by the weak learner a as the error rate a; Use the Pearson coefficient feature recognition test set to calculate the accuracy a of the weak learner a, and the accuracy is based on the brain disease label as the target; B2. Adjust the weights of the training samples according to the error rate of the weak learner a, and increase the weights of the misclassified samples by half the error rate a; B3. Send the Pearson coefficient feature recognition training set after adjusting the sample weight to the initial weak learner b based on the SVM model for training, and train the weak learner b with the brain disease label as the target. Use the Pearson coefficient feature recognition test set to calculate the accuracy of the trained weak learner b, and compare whether the improvement rate of the accuracy of the trained weak learner b to the accuracy a is greater than the preset value. If it is greater than the preset value, the trained weak learner b is used as the weak learner b. If it is less than the preset value, repeat steps B2-B3 to obtain the weak learner b; calculate the proportion of samples misclassified by the weak learner b as the error rate b, and use the Pearson coefficient feature recognition test set to calculate the accuracy b of the weak learner b; B4. Repeat step B3 to obtain weak learner c and the accuracy c of weak learner c; add weights to the outputs of the weak learners according to the accuracy of the three weak learners; integrate the three weak learners into a Pearson coefficient recognition model, and calculate the weighted brain disease labels and weighted corresponding category probabilities output by the three weak learners as the output of the Pearson coefficient recognition model; In addition, the maximum mutual information coefficient identification model is obtained in the same way as above.

[0020] The KS test was used to select features for the final feature vector of the Pearson coefficient and the final feature vector of the maximum mutual information coefficient. The specific steps of the KS test were as follows: the cumulative distribution function CDF was calculated for each feature value and brain disease label, the maximum distance between the feature CDF and the category CDF was calculated, the D value of the feature was obtained, the distribution of the D value of the feature was determined using the normal distribution, the P value was calculated by finding the probability of the D value of the feature appearing at the corresponding position (or a larger value) in its distribution, the features were sorted according to the P value, and the features with a P value less than the set threshold were selected for subsequent analysis. Features with a smaller P value had a higher analytical value. S5. Visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output the visualization results.

[0021] Example 2, a brain disease auxiliary diagnosis system based on medical imaging, see Figure 1 As shown, including: An image acquisition module, used to acquire structural MRI images and time-series functional MRI images; A data preprocessing module is used to preprocess the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, and to divide the time series functional MRI images into time slices of fixed length using a sliding window method to obtain a 2D data matrix corresponding to each time slice; The feature construction module is used to pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). The Pearson coefficient weight W of each brain region pair (i, j) is calculated using the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice. ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; The disease classification module is used to input the structural MRI image, the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; The auxiliary diagnosis report module is used to visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output visualization results.

[0022] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A brain disease auxiliary diagnosis method based on medical imaging, characterized in that: include: S1. Acquire structural MRI images and sequential functional MRI images; S2. Preprocessing the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, segmenting the time series functional MRI images into time slices of fixed length using a sliding window method, and obtaining a 2D data matrix corresponding to each time slice; S3. Pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). Use the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice to calculate the Pearson coefficient weight W of each brain region pair (i, j) ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; S4. Send the structural MRI image, the final eigenvector of the Pearson coefficient, and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; S5. Visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output the visualization results.

2. The method for assisting diagnosis of brain diseases based on medical imaging according to claim 1, characterized in that: The integrated disease-assisted diagnosis model is divided into a data processing layer, a base learner recognition layer, and an integrated output layer; The data processing layer is used to preprocess the structural nuclear magnetic resonance image to obtain the structural image to be identified, perform feature selection on the final feature vector of the Pearson coefficient and the final feature vector of the maximum mutual information coefficient, and obtain the feature vector of the Pearson coefficient to be identified and the feature vector of the maximum mutual information coefficient to be identified; The base learner recognition layer includes three base learners, namely, the structural image recognition model, the Pearson coefficient recognition model and the maximum mutual information coefficient recognition model; the structural image recognition model is used to perform disease recognition on the structural image to be recognized, and outputs the brain disease label and the corresponding category probability; the Pearson coefficient recognition model is used to perform disease recognition on the Pearson coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; the maximum mutual information coefficient recognition model is used to perform disease recognition on the maximum mutual information coefficient feature vector to be recognized, and outputs the brain disease label and the corresponding category probability; The integrated output layer is used to identify the outputs of the three base learners and output disease auxiliary diagnosis labels and corresponding category probabilities.

3. The method for auxiliary diagnosis of brain diseases based on medical imaging according to claim 2, characterized in that: Construct an integrated disease-assisted diagnosis model. The specific steps are: A1. Collect structural MRI images and time-series functional MRI images of several samples, label the samples with brain disease labels, extract the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient corresponding to the time-series functional MRI images of each sample, and perform feature selection on the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient using the KS test; The structural MRI images of the samples are preprocessed and combined with brain disease labels to form a structural image recognition data set, and the structural image recognition data set is divided into a structural image recognition training set and a structural image recognition test set; The final feature vector of the Pearson coefficient of the sample after feature selection and the brain disease label are used to form a Pearson coefficient feature recognition data set, and the Pearson coefficient feature recognition data set is divided into a Pearson coefficient feature recognition training set and a Pearson coefficient feature recognition test set by sampling with replacement; The final feature vector of the maximum mutual information coefficient of the sample after feature selection and the brain disease label are used to form a maximum mutual information coefficient feature recognition data set, and the maximum mutual information coefficient feature recognition data set is divided into a maximum mutual information coefficient feature recognition training set and a maximum mutual information coefficient feature recognition test set by sampling with replacement; A2. Send the structural image recognition training set to the initial structural image recognition model based on the convolutional neural network for training, with the brain disease label as the target, to obtain the training structural image recognition model; send the structural image recognition test set to the training structural image recognition model for accuracy evaluation, with the brain disease label as the target, adjust the hyperparameters of the training structural image recognition model to obtain the structural image recognition model; The Pearson coefficient recognition model was obtained by combining the AdaBoost method with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The maximum mutual information coefficient recognition model was obtained by combining the AdaBoost method with the maximum mutual information coefficient feature recognition training set and the maximum mutual information coefficient feature recognition test set. A3. The structural image recognition model, the Pearson coefficient recognition model, and the maximum mutual information coefficient recognition model are used as base learners of the integrated disease-assisted diagnosis model; the three base learners are integrated in parallel to obtain the brain disease labels and corresponding category probabilities output by the three base learners on their respective training sets and test sets, and the brain disease labels and corresponding category probabilities output by the three base learners and the sample-labeled brain disease labels are combined into new recognition samples, all the recognition samples are combined into an integrated learning data set, and the integrated learning data set is divided into an integrated learning training set and an integrated learning test set; The ensemble learning training set is sent to the initial ensemble learning model based on the multi-layer perceptron for training, and the sample-labeled brain disease labels are used as the target to train the initial ensemble disease-assisted diagnosis model; the ensemble learning test set is sent to the initial ensemble disease-assisted diagnosis model for accuracy evaluation, and the sample-labeled brain disease labels are used as the target to adjust the hyperparameters of the initial ensemble disease-assisted diagnosis model to obtain the ensemble disease-assisted diagnosis model.

4. The method for auxiliary diagnosis of brain diseases based on medical imaging according to claim 3, characterized in that: The Pearson coefficient recognition model is obtained by using the AdaBoost method combined with the Pearson coefficient feature recognition training set and the Pearson coefficient feature recognition test set. The specific steps are as follows: B1. Send the samples in the Pearson coefficient feature recognition training set as training samples to the weak learner a based on the SVM model for training, and train the weak learner a with the brain disease label as the target. Calculate the proportion of training samples that are misclassified by the weak learner a as the error rate a; Use the Pearson coefficient feature recognition test set to calculate the accuracy a of the weak learner a, and the accuracy is based on the brain disease label as the target; B2. Adjust the weights of the training samples according to the error rate of the weak learner a, and increase the weights of the misclassified samples by half the error rate a; B3. Send the Pearson coefficient feature recognition training set after adjusting the sample weight to the initial weak learner b based on the SVM model for training, and train the weak learner b with the brain disease label as the target. Use the Pearson coefficient feature recognition test set to calculate the accuracy of the trained weak learner b, and compare whether the improvement rate of the accuracy of the trained weak learner b to the accuracy a is greater than the preset value. If it is greater than the preset value, the trained weak learner b is used as the weak learner b. If it is less than the preset value, repeat steps B2-B3 to obtain the weak learner b; Calculate the proportion of training samples that are misclassified by weak learner b as the error rate b, and use the Pearson coefficient feature recognition test set to calculate the accuracy rate b of weak learner b; B4. Repeat step B3 to obtain weak learner c and the accuracy c of weak learner c; add weights to the outputs of the weak learners according to the accuracy of the three weak learners; integrate the three weak learners into a Pearson coefficient recognition model, and calculate the weighted brain disease labels and weighted corresponding category probabilities output by the three weak learners as the output of the Pearson coefficient recognition model; In addition, the maximum mutual information coefficient identification model is obtained in the same way as above.

5. The method for auxiliary diagnosis of brain diseases based on medical imaging according to claim 3, characterized in that: The KS test was used to select features for the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient. The specific steps of the KS test were as follows: the cumulative distribution function CDF was calculated for each feature value and brain disease label, the maximum distance between the feature CDF and the category CDF was calculated, the D value of the feature was obtained, the distribution of the D value of the feature was determined using a normal distribution, the P value was calculated by finding the probability of the D value of the feature appearing at the corresponding position (or a larger value) in its distribution, the features were sorted according to the P value, and the features with a P value less than the set threshold were selected for subsequent analysis.

6. The method for auxiliary diagnosis of brain diseases based on medical imaging according to claim 1, characterized in that: Preprocess the time series functional magnetic resonance imaging. The specific steps of preprocessing are as follows: C1. Delete the first three functional MRI images in the time series functional MRI images; C2. Perform time slice correction, head motion correction and spatial standardization on the remaining time series functional MRI images to obtain standard functional MRI images; C3. Obtain the global blood oxygen level-dependent time series signal for each standard functional MRI image; C4. Based on the automatic anatomical labeling template, the global blood oxygen level dependent time series signal is divided into the average resting time series signals of multiple brain regions. The average resting time series signal of each brain region is processed using a bandpass filter with a frequency band of 0.01-0.08 Hz to obtain the blood oxygen level dependent signal value of each brain region, and a two-dimensional data matrix is ​​obtained, which stores the blood oxygen level dependent signal values ​​of the brain regions.

7. The method for assisting diagnosis of brain diseases based on medical imaging according to claim 1, characterized in that: The time series functional magnetic resonance imaging is segmented into time slices of fixed length using a sliding window method. The formula for segmenting time slices is: ; In the formula, num window is the number of time slices, window size For the length of each window, window step is the interval between two consecutive windows, num volume is the time series functional MRI image and is the total number of images.

8. The method for assisting diagnosis of brain diseases based on medical imaging according to claim 1, characterized in that: Using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) respectively calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient, and the steps of the calculation method are: D1. Obtain the Pearson coefficient weight W for each brain region pair (i, j) at each time slice ij (PCC); D2. All Pearson coefficient weights W on each time slice ij (PCC) is concatenated into a size of (num ROI ,num ROI ) and is called the brain functional connection network correlation matrix corresponding to the time slice, num ROI is the number of brain regions; D3. Reduce the dimension of the brain functional connection network correlation matrix on each time slice to a length of (num ROI *num ROI ), recorded as the feature vector of the corresponding time slice; D4. Take the average of the eigenvectors on all time slices as the final eigenvector of the Pearson coefficient; Furthermore, the final eigenvector of the maximum mutual information coefficient is obtained by the same steps.

9. A brain disease auxiliary diagnosis system based on medical images, characterized in that: The system uses a brain disease auxiliary diagnosis method based on medical images as described in any one of claims 1 to 8, including: An image acquisition module, used to acquire structural MRI images and time-series functional MRI images; A data preprocessing module is used to preprocess the time series functional MRI images to obtain a 2D data matrix storing the blood oxygen level-dependent signal values ​​of the brain regions, and to divide the time series functional MRI images into time slices of fixed length using a sliding window method to obtain a 2D data matrix corresponding to each time slice; The feature construction module is used to pair different brain regions in pairs to obtain multiple groups of brain region pairs (i, j). The Pearson coefficient weight W of each brain region pair (i, j) is calculated using the brain region blood oxygen level dependence signal value in the 2D data matrix corresponding to each time slice. ij (PCC) and the maximum mutual information coefficient weight W ij (MIC); using all Pearson coefficient weights W ij (PCC) and all maximum mutual information coefficient weights W ij (PCC) calculates the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient respectively; The disease classification module is used to input the structural MRI image, the final eigenvector of the Pearson coefficient and the final eigenvector of the maximum mutual information coefficient into the integrated disease-assisted diagnosis model for disease classification, and output the disease-assisted diagnosis label and the corresponding category probability; The auxiliary diagnosis report module is used to visualize the disease auxiliary diagnosis labels and corresponding category probabilities and output visualization results.

Citation Information

Cited By

  • Multi-modal fusion-based explainable rare disease auxiliary diagnosis system and method

    CN120496811A