CN113080929A

CN113080929A - anti-NMDAR encephalitis image feature classification method based on machine learning

Info

Publication number: CN113080929A
Application number: CN202110399295.4A
Authority: CN
Inventors: 肖卓凌; 潘瑞祥; 魏然; 高宇; 陈劲涛; 杜周阳
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2021-04-14
Filing date: 2021-04-14
Publication date: 2021-07-09

Abstract

The invention discloses a feature classification method for anti-NMDAR encephalitis images based on machine learning. The original MRI image is preprocessed, then sent to a feature screening model based on the Lasso method, and finally the screened features are sent to a C-SVM combined with 3D‑CNN to obtain feature classification results. The method of the invention only uses the features of the cerebral cortex, and is highly correlated with autoimmune encephalitis. Compared with the existing manual classification method, the method greatly reduces the actual feature classification and reduces the dependence on domain expert knowledge; The method transforms the visual information of medical images into deep data features for quantitative research, providing physicians with objective, consistent and reproducible reference information.

Description

anti-NMDAR encephalitis image feature classification method based on machine learning

Technical Field

The invention belongs to the technical field of medical image processing, and particularly relates to an NMDAR encephalitis resisting image feature classification method based on machine learning.

Background

anti-NMDAR encephalitis is an inflammatory encephalopathy caused by anti-neuronal autoantibodies with complex pathogenesis, which not only seriously impairs physical and mental health of patients, but also imposes a heavy burden on families and society. The main etiological features of the disease are concentrated in the cerebral cortex, and Magnetic Resonance Imaging (MRI) is a common method for clinical examination of the brain, and can detect changes in various regions of the brain.

The Dutch scholars put forward the concept of radiology in 2012, and the concept emphasizes that a large amount of information is quickly extracted from images, so that feature extraction and model construction are realized; the deep mining of the mass data information can help the doctor to make judgment. Radiology can intuitively and immediately convert visual image information into quantitative research of deep features, and a machine learning method can be used for associating, reflecting and predicting the behavior of a patient by detecting changes of brain structural features, so that doctors are helped to perform follow-up operations.

For feature input of traditional machine learning methods, most studies still use magnetoencephalograms to acquire processed non-imaging features, such as cortical parameters, morphological feature measurements or GM/WM intensities and "doughnout" values calculated from cortical parameters. However, the conventional machine learning module has limited computational power and is difficult to process all parameters, and thus a method for extracting a large number of imaging features needs to be studied.

Disclosure of Invention

Aiming at the defects in the prior art, the anti-NMDAR encephalitis image feature classification method based on machine learning solves the problems that the existing brain structure image is difficult to identify the volume features only by naked eyes, and the existing machine learning method is limited in computing capacity and difficult to process all image parameters.

In order to achieve the purpose of the invention, the invention adopts the technical scheme that: a method for classifying anti-NMDAR encephalitis image features based on machine learning comprises the following steps:

s1, acquiring a corresponding MRI image based on the sample which is diagnosed as the NMDAR encephalitis resisting sample;

s2, processing the image data of the obtained MRI image, reconstructing the cerebral cortex and calculating the corresponding cortical features;

s3, performing data positioning and feature screening processing on all the calculated cortex features to form a data set;

s4, constructing an anti-NMDAR encephalitis image feature classification model, and training the model through a data set;

and S5, inputting the cortex characteristics to be processed into the trained NMDAR encephalitis resisting image characteristic classification model to obtain a characteristic classification result.

Further, in step S2, the method for processing the image data of the MRI image corresponding to the sample specifically includes:

a1, performing skull stripping on the MRI image, and performing B1 deviation field correction and grey white substance segmentation;

a2, reconstructing a cortex surface model based on the grey white substance segmentation result;

a3, marking the surface area of the cortex and the subcortical brain structure on the reconstructed cortical surface model;

a4, based on the marked surface region of the cortex and the subcortical brain structure, carrying out nonlinear registration on the surface of the cortex in the MRI image of each sample by using a stereotaxic map, reconstructing the cerebral cortex and calculating the cortical characteristic parameters of each cerebral region;

a5, selecting a plurality of cortical features with highest relevance to the autoimmune encephalitis in the cerebral cortex based on the cortical feature parameters.

Further, the step a5 specifically includes:

determining a standard brain as a surface template, mapping an annotation corresponding to an aparc.a2009 brain map into each cortical area of the surface template, and coloring the cortical areas interested in resisting NMDAR encephalitis in the cerebral cortex according to the cortical characteristic parameters to obtain the cortical characteristics of the cerebral cortex.

Further, in step S3, the method for performing data positioning on the cortical feature specifically includes:

b1, performing missing value check on the cortex characteristics corresponding to each sample, and deleting the cortex characteristics of the sample corresponding to the cortex characteristics with the missing values;

b2, detecting the outliers of all the cortical features of each current sample by a quartile spacing method, and deleting the outliers

B3, performing standardization processing on all cortical features of the current sample;

and B4, taking all the processed cortical features as the feature information of the sample.

Further, the method for screening the feature information in step S3 specifically includes:

c1, forming a feature set by the feature information of all samples;

c2, taking the trained lasso model as a characteristic screening model;

c2, processing the feature information in the feature set through the trained feature screening model, distributing weight to each feature information in the processing process, and deleting the feature information smaller than a set weight threshold value from the feature set;

and C4, carrying out validity verification on the feature information retained in the feature set by an ROC regression method, and forming a data set by the verified feature information.

Further, the structure of the anti-NMDAR encephalitis image feature classification model in step S4 is:

on the basis of a 3D-CNN network, a full connection layer is modified into a support vector machine, and a feature classification result of the NMDAR encephalitis resisting image is output by the support vector machine.

Further, in the step S5, the performance evaluation parameters of the trained NMDAR encephalitis resisting image feature classification model include an accuracy value and a classification performance value;

wherein, the accuracy ACC is:

the categorical performance values F1 were:

where precision is precision, recall is recall,

TP is true positive, FP is false positive, TN is true negative and FN false negative.

The invention has the beneficial effects that:

(1) the method only uses the characteristics of the cerebral cortex, has high correlation with autoimmune encephalitis, and greatly reduces the actual characteristic classification and the dependence on domain expert knowledge compared with the existing artificial classification method;

(2) the method of the invention processes the medical image, extracts corresponding cortical parameters aiming at different brain images so as to extract the characteristics of the local brain lesion, greatly saves the cost of a detection instrument and improves the accuracy of characteristic classification;

(3) the method converts the visual information of the medical image into deep data characteristics for quantitative research, and provides objective, consistent and reproducible reference information for doctors.

Drawings

Fig. 1 is a flowchart of a method for classifying features of an anti-NMDAR encephalitis image based on machine learning according to the present invention.

Fig. 2 is a schematic view of visualization of an MRI image during image data processing according to the present invention.

Fig. 3 is a schematic diagram of feature visualization retained after feature screening of cortical features according to the present invention.

Detailed Description

The following description of the embodiments of the present invention is provided to facilitate the understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and it will be apparent to those skilled in the art that various changes may be made without departing from the spirit and scope of the invention as defined and defined in the appended claims, and all matters produced by the invention using the inventive concept are protected.

As shown in fig. 1, a method for classifying NMDAR encephalitis-resistant image features based on machine learning includes the following steps:

In step S1, a plurality of patient data are collected based on historical clinical work, a plurality of health samples are randomly selected from the database for comparison, after T1 weighted images are evaluated, Autoimmune Encephalitis (AE) and health data in the patient are analyzed, and corresponding serum and/or CSF samples are sent to patients in the cohort of autoimmune neurological disease markers in a remote special detection center, thereby identifying individuals diagnosed with NMDAR encephalitis.

In step S1, when the MRI image is acquired, structural MRI images of all subjects are acquired by acquiring signals using a Trio or Skyra 3.0 scanner and a standard 8-channel head coil. The T1 weighted structure brain image was acquired using a three-dimensional fourier transform fast spoiled gradient recall with a steady state because it provided excellent contrast between gray and white matter. In Trio, the repetition time (repetition time) is 450-500ms, the echo time (echo time) is 20-25ms, the slice thickness (slice thickness) is 5.0mm, the matrix (matrix) is 256256, and the number of axial slices (number of axial slices) is 20. On Skyra, repetition time is 1600-1660 ms, echo time is 8-8.6 ms, slice thickness is 5.0mm, matrix is 256256, and axial slice number is 20.

In step S2, the method for processing the image data of the MRI image corresponding to one sample specifically includes:

in this step, the reconstructed cerebral cortex is visually checked for the reconstructed result of each viewer, and any incorrect place due to the artifact image is corrected manually;

a5, selecting a plurality of cortical features with highest relevance to the autoimmune encephalitis in the cerebral cortex based on the cortical feature parameters;

wherein, the step A5 specifically comprises the following steps:

The visualization result obtained based on the above image data processing is shown in fig. 2.

In step S3, the method for performing data positioning on the cortical feature specifically includes:

wherein for the normalization process to subtract the sample mean, divided by the population standard deviation of each individual of the same characteristic, the standard scalar is applied to all datasets with fitted transformation ranges;

The method for screening the feature information in step S3 specifically includes:

c1, forming a feature set by the feature information of all samples;

c2, taking the trained lasso model as a characteristic screening model;

Specifically, in the step C2, when the lasso model is trained, the feature set obtained by positioning the previous data is randomly divided into k mutually exclusive subsets with mutually exclusive sizes, where a sum of the k subsets is used as a training set of the lasso model, and the remaining subsets are used as test sets, and the content of each training round is different, so that the generalization capability of the lasso model is improved.

The lasso model is an LR-based model that constructs penalty coefficients to obtain a more accurate model, compressing some of the regression coefficients, limits the sum of the absolute values of the forcing coefficients to less than some fixed value, and sets some of the regression coefficients to zero, thus, it retains the advantage of subset shrinkage, being a biased estimate of the data with multiple linearity.

Fig. 3 is a schematic diagram showing the visualization of the features retained after the feature screening.

The structure of the anti-NMDAR encephalitis image feature classification model in the step S4 is as follows:

The basic model of Support Vector Machine (SVM) is a linear classifier with maximum separation defined in feature space, and a C-support vector machine classifier (C-SVM) is used in the present invention, which defines the classification of nonlinear samples. The learning strategy of the support vector machine is interval maximization, which can be formalized as a convex quadratic programming problem and is also equivalent to a regularized hinge loss function minimization problem, the optimal classification hyperplane of the SVM depends on some support vector machines, and in order to prevent model transition fitting, we fit data through a limited data set (only 400 samples but 304 features).

The invention adopts SVM-based ensemble learning (SVM + AdaBoost): because of limitations of medical image data and difficulties in feature selection, only SVM cannot obtain satisfactory classification results and high accuracy, and thus, an attempt has been made to improve classification results based on integrated learning of SVM, AdaBoost is a weak classification enhancement method with an integrated boosting algorithm that combines a group of weak classifiers into a weighted sum to create a stronger enhanced classifier, the principle of which is to update the weak classifiers by adjusting sample weights according to a training set, the weight of each sample being estimated from the output of the classifier in the previous step, so as to improve the next classifier to cope with more challenging examples.

In step S5, the performance evaluation parameters of the trained NMDAR encephalitis resisting image feature classification model include an accuracy value and a classification performance value;

wherein, the accuracy ACC is:

the categorical performance values F1 were:

where precision is precision, recall is recall,

TP is true positive, FP is false positive, TN is true negative and FN false negative. Wherein, the accuracy value is used for characterizing the accuracy of the classification task of the model, and the classification performance value F1 is used for characterizing the classification recognition capability of the model to the features. The classification performance value achieves a perfect balance between accuracy and recall, thereby providing a correct assessment of the model's performance in classifying images.

Claims

1. an anti-NMDAR encephalitis image feature classification method based on machine learning, is characterized in that, comprises the following steps:

S1. Obtain corresponding MRI images based on samples diagnosed as anti-NMDAR encephalitis;

S2. Perform image data processing on the acquired MRI image to reconstruct the cerebral cortex and calculate the corresponding cortical features;

S3. Perform data positioning and feature screening on all the calculated cortical features to form a data set;

S4. Build an anti-NMDAR encephalitis image feature classification model, and train it through the data set;

S5. Input the cortical features to be processed into the trained anti-NMDAR encephalitis image feature classification model to obtain a feature classification result.

2. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 1, is characterized in that, in described step S2, the method that the MRI image corresponding to a sample is carried out image data processing is specifically:

A1. Perform skull stripping on MRI images, and perform B1 bias field correction and gray-white material segmentation;

A2. Reconstruct the cortical surface model based on the gray-white material segmentation results;

A3. Mark the cortical surface area and subcortical brain structures on the reconstructed cortical surface model;

A4. Based on the marked cortical surface area and subcortical brain structure, use stereotaxic atlas to non-linearly register the cortical surface in the MRI image of each sample, reconstruct the cerebral cortex and calculate the cortical characteristic parameters of each brain area;

A5. Based on cortical feature parameters, select several cortical features in the cerebral cortex that are most correlated with autoimmune encephalitis.

3. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 2, is characterized in that, in described step A5, be specifically:

Determine the standard brain as the surface template, map annotations corresponding to the aparc.a2009 brain atlas into each cortical region of the surface template, color the cortical regions of interest for anti-NMDAR encephalitis in the cerebral cortex according to the cortical feature parameters, and obtain the brain Cortical features of the cortex.

4. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 1, is characterized in that, in described step S3, the method that cortical feature is carried out data location is specifically:

B1. Check the cortical features corresponding to each sample for missing values, and delete the cortical features of the samples corresponding to the cortical features with missing values;

B2. Detect the outliers of all cortical features of each current sample by the interquartile range method, and delete them

B3. Standardize all cortical features of the current sample;

B4. Use all processed cortical features as the feature information of the sample.

5. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 4, is characterized in that, in described step S3, the method that feature information is screened is specifically:

C1. Combine the feature information of all samples into a feature set;

C2. Use the trained lasso model as a feature screening model;

C2. Process the feature information in the feature set through the trained feature screening model. During the processing, assign a weight to each feature information, and delete the feature information less than the set weight threshold from the feature set;

C4. The validity of the feature information retained in the feature set is verified by the ROC regression method, and the verified feature information is formed into a data set.

6. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 1, is characterized in that, the structure of the anti-NMDAR encephalitis image feature classification model in described step S4 is:

Based on the 3D-CNN network, its fully connected layer is modified into a support vector machine, and the support vector machine outputs the feature classification results of anti-NMDAR encephalitis images.

7. the anti-NMDAR encephalitis image feature classification method based on machine learning according to claim 6, is characterized in that, in described step S5, for the performance evaluation parameter of the trained anti-NMDAR encephalitis image feature classification model including accurate rate value and classification performance value;

Among them, the accuracy rate ACC is:

The classification performance value F1 is:

In the formula, precision is the precision, recall is the recall rate,

TP is true example, FP is false positive example, TN is true negative example and FN false negative example.