CN111415324A - Classification and identification method of spatial distribution characteristics of brain lesions based on magnetic resonance imaging - Google Patents

Abstract

The invention belongs to the technical field of image processing and application, and in particular relates to a method for classifying and identifying the spatial distribution characteristics of brain lesion images based on magnetic resonance imaging. The method of the present invention mainly includes the steps of lesion segmentation, individual image registration, spatial standardization, individualization of standard spatial template, feature extraction of spatial distribution of lesions, feature screening and modeling. Feature analysis, build a set of analysis methods for the spatial distribution of brain lesion images, and use machine learning to perform feature screening and modeling on this basis. This method can be used to classify and identify brain lesions of different brain diseases or brain states caused by different antibodies, different genes and other reasons using brain magnetic resonance imaging, and provide effective guidance for clinical and scientific research.

Description

Classification and identification method of spatial distribution characteristics of brain lesions based on magnetic resonance imaging

technical field

The invention belongs to the field of image processing and application, in particular to a method for classifying and identifying the spatial distribution characteristics of brain lesion images based on magnetic resonance imaging.

Background technique

In the prior art, in the classification and identification methods of different brain diseases or brain state images, brain magnetic resonance imaging technology plays an important role due to its non-invasiveness, timeliness, and excellent display effect on brain lesions. In clinical practice, doctors often summarize and summarize the different characteristics of lesions of different diseases on images through long-term clinical experience, and carry out visual classification, identification and reporting. However, this kind of experience-based classification and identification has shortcomings such as low efficiency, difficulty in discovering new features of lesions, and difficulty in automatically combining multiple features. For experienced new disease subtypes or clinically rare diseases, the accuracy and efficiency of the brain lesion image classification and identification are greatly reduced.

The radiomics and deep learning methods developed in recent years provide an idea for the classification and identification of brain lesions from the perspective of data analysis. Generally, deep learning methods [1] generally require a large amount of data, and the results are often not interpretable, so they are not suitable for small sample sizes and exploratory research stages of classification problems such as similar diseases, rare diseases, and new subtypes; The omics method [2] extracts quantitative features based on the grayscale information of image lesions, including statistical features, texture features, filtering features, etc., and then uses machine learning methods to build models; since features are artificially defined and extracted, the results have Interpretability; however, traditional radiomics often only focus on the characteristics of the lesions themselves, rather than their spatial distribution in the brain. In fact, according to clinical and research experience, diseases caused by different genotypes or antibodies often have important differences in the spatial location and distribution characteristics of the lesions in the brain, so they are of great significance in classification and identification.

At present, there is no systematic method to classify and identify the spatial distribution characteristics of brain lesions based on magnetic resonance imaging. Some research reports have paid attention to some spatial distribution characteristics, but they are generally identified by doctors or researchers with the naked eye (such as the presence or absence of lesions next to the ventricles), and then calibrated and counted [3]. This method of relying on qualitative identification and counting of doctors, on the one hand, extracts very limited features and is not objective enough, and there are differences between the results of different doctors or researchers; on the other hand, it does not realize automation and is inefficient.

Based on the shortcomings of the existing classification and identification technologies in terms of the spatial distribution characteristics of lesions, the inventor of the present application intends to provide a classification and identification method for the spatial distribution characteristics of brain lesion images based on magnetic resonance imaging. Calculate and analyze the spatial distribution characteristics of various categories, and obtain models and discrimination methods suitable for classification and identification, especially in the classification of different diseases caused by different genes or antibodies, which provides effective guidance for clinical and scientific research.

The prior art related to the present invention includes:

[1] Najafabadi M M, Villanustre F, Khoshgoftaar T M, et al. Deep learning applications and challenges in big data analytics [J]. Journal of Big Data, 2015, 2(1): 1.

[2] Ma X, Zhang L, Huang D, et al.Quantitative radiomic biomarkers fordiscrimination between neuromyelitis optica spectrum disorder and multiplesclerosis[J].Journal of Magnetic Resonance Imaging, 2018.

[3] Jurynczyk M, Geraldes R, Probert F, et al. Distinct brain imaging characteristics of autoantibody-mediated CNS conditions and multiplesclerosis[J]. Brain, 2017, 140(3): 617-627.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for classifying and identifying the spatial distribution characteristics of brain lesions based on magnetic resonance imaging based on the deficiencies of the existing classification and identification technologies in terms of the spatial distribution characteristics of lesions. The method mainly includes lesion segmentation, individual image matching Standardization, spatial standardization, individualization of standard spatial template, feature extraction of spatial distribution of lesions, feature screening and modeling (as shown in Figure 1). The method of the present invention obtains a model and a discrimination method suitable for classification and identification by calculating and analyzing the spatial distribution characteristics of various categories of brain lesion images in magnetic resonance imaging, especially in the classification of different diseases caused by different genes or antibodies. and scientific research to provide effective guidance.

Specifically, the magnetic resonance imaging-based method for classifying and identifying the spatial distribution characteristics of brain lesion images of the present invention includes the following steps:

1), lesion image segmentation:

1)-1, preparation and selection of brain imaging data: at least two groups of G1 and G2 test images; for each individual subject, two images are prepared, the first image is a "lesion display image", which is obtained from the patient's clinical brain magnetic resonance imaging. Among the images (including but not limited to T1-weighted images, T2-weighted images, FLAIR, DWI, ADC, SWI, etc.), one modality is selected as the image for subsequent lesion segmentation; the second image is "brain structure image", and T1-weighted image is generally selected image; both images can be the same;

1)-2, Lesion segmentation: For the "lesion display image", use image segmentation software, toolkits and segmentation algorithms, including but not limited to workstations and segmentation software provided by MRI manufacturers, MRIcro, MRIcron, MIPAV, ITK-SNAP, MITK , regional growth algorithm, etc., to segment the whole brain lesion, and save it as a binary image or data to generate a "lesion image";

2) Individual image registration: Rigid registration of individual "brain structure image" and "lesion display image", available software and methods include but are not limited to SPM, FSL, etc.;

3) Space standardization: register the individual "brain structure image" to a standard space (such as MNI standard space, Talairach standard space, etc.), using software and methods including but not limited to SPM, FSL, AFNI, ANTs, based on deformation field registration method, etc.; the transformation parameters generated in this process are applied to the "lesion image", and the binarized "standard spatial lesion image" is obtained by selecting a threshold (0.5 by default, 0-1 optional); using the same The software and method of the individual "brain structure image" are registered with the standard symmetric brain template (the template with completely symmetrical left and right brain space), and the binarized "symmetric spatial lesion image" is obtained;

4) Individualization of standard space template: This step is applicable to the situation that in the subsequent steps 5)-4-6, the partition feature needs to be calculated in the individual space, if it is carried out in the standard space, it can be omitted; for the partition template of the standard space (including But not limited to gray matter template, white matter template, cerebrospinal fluid template, cerebral lobe segmentation template, subtentorial structure segmentation template, AAL segmentation template, Brodmann segmentation template, HarvardOxford segmentation template, white matter segmentation such as corpus callosum, ventricle template, and midbrain aqueduct, etc. Partition, etc.), extract each partition in the template as an independent binary template, use the inverse transformation process of the individual to the standard space in step 3, register all the partition templates to the individual space, and select the threshold (default 0.5, optional) 0 to 1) to obtain the binarized "individual spatial partition image" corresponding to each partition;

5) Feature extraction of spatial distribution of lesions: Among them, the following categories of spatial distribution feature sets of lesions and their extraction methods are established:

5)-1, distribution size features: using the individual space "lesion image", according to the principle of inter-voxel connectivity, extract a single 2D or 3D lesion, and calculate the size features such as the volume and longest diameter of a single lesion; on this basis, obtain The maximum, mean, sum, median, number of lesions, and the number of large lesions (such as more than 200mm ³ ) of all the above characteristics of each subject for each individual lesion were taken as the distribution size of the subjects. feature set;

5)-2, Symmetrical features: Using the "symmetric spatial lesion image" and the corresponding partition template, in the whole brain and within each group of symmetrical brain partitions, calculate the symmetrical features of the lesions in the left and right hemispheres, including unilateral/bilateral Laterality (marked as -1/0/1 on the left/double/right side respectively), the difference or ratio of the number of voxels in the left and right hemispheres of the lesion, the Dice coefficient to measure the similarity of distribution, etc.; The Dice coefficient of i is defined as the ratio of the ratio of the number of voxels in the intersection of the two lesions to the number of voxels in the union after the left-side (VL) or right-side (VR) images of symmetrical lesions are flipped left and right. times, that is

D(i)=2×|VL∩VR|/|VL∪VR|

5)-3, Probability distribution map features: For each group of subjects, use the binarized "standard spatial lesion image" of all individuals in the group, calculate the number of non-zero values in each voxel in the group, and divide by the individuals in the group The total number of lesions is obtained to obtain the spatial probability distribution map of the group of lesions (the voxel value ranges from 0 to 1, the closer to 1, the higher the probability of the voxel lesion exists), and the threshold value (default 0.2, optional 0 to 1) is obtained. The high probability distribution map of the group; defining the characteristic value of the probability distribution map of the gth group of the lesions of a subject i in the M groups, the value of the voxels in the intersection of the lesion area of i and the high probability distribution lesion area of the gth group The difference between the number and the sum of the number of intersections of this lesion and other groups with high probability distribution lesions:

Following steps 5)-4-6, the analysis can be performed in the individual space using the "lesion image" and the "individual space partition image", or in the standard space using the "standard space lesion image" and the partition template of the standard space, or both;

5)-4, distribution characteristics of a single brain region: for each individual brain region, calculate the distribution characteristics of the subject's brain lesions in the region: whether the lesions exist in the region (1 for existence, 0 for non-existence), The volume of the lesion in the partition, the proportion of the volume of the lesion in the partition to the volume of the entire partition, etc.;

5)-5, distribution characteristics of brain regions: For the landmark brain regions of interest (such as ventricle, midbrain aqueduct, corpus callosum, white matter area, etc.), small regional expansion operations (generally no more than 5mm) are performed to generate The para-brain area, and calculate the para-brain distribution characteristics of the tested lesions for the corresponding brain area (such as paraventricular, midbrain periaqueduct, near the corpus callosum, near the white matter, etc.), including the existence of para-brain lesions (the existence of 1, if it does not exist, it is 0), the volume of the lesions next to the brain region, etc.;

5)-6, cross-brain distribution characteristics: According to the results of the above steps 5)-4, calculate the cross-brain distribution characteristics of the tested brain lesions, including the number of covered brain tissue types (gray matter/white matter/cerebrospinal fluid) and pairwise ratio, the number of brain regions covered, etc.;

6), feature screening and modeling:

6)-1, feature screening: on the basis of the feature set of spatial distribution of lesions extracted in step 5), various feature screening methods in machine learning are used to screen features; available feature screening methods include variance selection, chi-square test, U test, mutual information, recursive elimination method, feature selection method of logistic regression model based on L1 penalty term/L2 penalty term/L1 combined with L2 penalty term (such as LASSO method), exhaustive method, etc.; some of these methods (such as LASSO method) ), can carry out feature screening and generate classification discrimination model at the same time, thus can combine 6-(2) steps;

6)-2, establish a classification and identification model: take the characteristics of the subject lesions after screening as input, and use the actual subject grouping category as a label, train the classifier, and generate a linear or nonlinear classification and identification model; available classification The tools include logistic regression, random forest, support vector machine, artificial neural network, etc. After the model is established, the classification and discrimination effect of the model is evaluated by the AUC value, accuracy, sensitivity, specificity and other values of the ROC curve.

The invention provides a method for classifying and discriminating the spatial distribution characteristics of brain lesion images based on magnetic resonance imaging. By calculating and analyzing the spatial distribution characteristics of various categories of brain lesion images in magnetic resonance imaging, a model and discriminant suitable for classification and discrimination are obtained. The method, especially in the classification of different diseases caused by different genes or antibodies, can provide effective guidance for clinical and scientific research.

Description of drawings

Fig. 1 is a flowchart of a classification and identification method based on the spatial distribution characteristics of MRI brain lesion images.

FIG. 2 shows the results of segmentation and extraction of lesions in the MOG and AQP4 groups in Example 1. FIG.

Figure 3 shows the feature selection results of Example 1.

FIG. 4 shows the ROC curve of the model classification discrimination effect of Example 1.

Detailed ways

Example 1 Classification and identification of the spatial distribution characteristics of brain lesions in patients with MOG antibody positive and AQP4 antibody positive NMOSD patients

Classification identification:

1) The clinical MRI images of the two groups of patients with MOG antibody positive and AQP4 antibody positive NMOSD were selected, 28 cases and 57 cases, respectively, each containing FLAIR images as "lesion display images", and T1-weighted images as "brain structure images" ;

2) For the FLAIR image, use MRIcron to segment the whole brain lesion, and save it as a binary image as a "lesion image";

3) Rigid registration of individual T1-weighted images with FLAIR images using SPM;

4) Use SPM to register individual T1-weighted images to MNI standard space; apply transformation parameters to FLAIR images, and select a threshold of 0.5 to obtain a binarized "standard space lesion image";

5) Using the individual space "lesion image", extract a single 3D lesion (as shown in Figure 2), and calculate the volume of a single lesion; obtain the maximum value, average value, sum, and number of lesions of all individual 3D lesions for each subject. number, the number of large lesions (over 200mm ³ );

6) For each individual brain partition, calculate whether the subject's brain lesion exists in the partition (1 for presence, 0 for non-existence), and the volume of the lesion in the partition;

7) Calculate the proportion of gray matter and white matter covered by the tested brain lesions and the number of brain regions covered;

8) Feature screening and modeling: The features extracted in step 5 are used as input, and the actual subject grouping category is used as a label (MOG group is 1, AQP4 group is 0), using the LASSO method, which uses 5-fold cross-validation, Perform feature screening and modeling; the final model contains 9 parameters, 1 of which is constant, and the remaining 8 are spatially distributed features (as shown in Figure 3).

The results show that the ROC curve of the established model has AUC=0.959, accuracy=0.959, sensitivity=1, specificity=0.86, which has a good classification and discrimination effect (as shown in Figure 4).

Claims (7)

1. a method for classifying and discriminating based on the spatial distribution feature of a brain lesion image of magnetic resonance imaging, it is characterized in that, it comprises the steps: 1), lesion image segmentation: 2), individual image registration: using SPM, FSL software and methods, rigid registration of individual "brain structure image" and "lesion display image"; 3) Spatial standardization: Using software and methods SPM, FSL, AFNI, ANTs, and deformation field-based registration methods, the individual "brain structure images" are registered to the standard space. image", and by selecting the threshold value, the binarized "standard spatial lesion image" was obtained; using the same software and method, the individual "brain structure image" was registered with the standard symmetric brain template, and the binarized "symmetrical image" was obtained. Spatial Lesion Image"; 4), standard space template individualization: for the partition template of the standard space, extract each partition in the template as an independent binary template, and use the inverse transformation process of the individual to the standard space in step 3 to register all the partition templates to the individual respectively. space, and by selecting the threshold (default 0.5, optional 0 to 1), the binarized "individual spatial partition image" corresponding to each partition is obtained; 5) Feature extraction of spatial distribution of lesions: Establish several categories of spatial distribution feature sets of lesions and their extraction methods: 6), feature screening and modeling. 2. by the described method of claim 1, it is characterised in that in described step 3), The individual "brain structure image" is registered to the MNI standard space or the Talairach standard space; the threshold is 0.5 by default, ranging from 0 to 1; the standard symmetrical brain template is a template that is completely symmetrical in the left and right brain spaces. 3. by the described method of claim 1, it is characterized in that, described step 4) is applicable to the situation that needs to calculate partition feature in individual space in follow-up 5)-4-6 steps; The partition templates of the standard space include but are not limited to gray matter template, white matter template, cerebrospinal fluid template, cerebral lobe partition template, subtentorial structure partition template, AAL partition template, Brodmann partition template, HarvardOxford partition template, corpus callosum partition template, ventricle template , and the self-made partition of the midbrain aqueduct. 4. by the described method of claim 1, it is characterized in that, described step 1) comprises sub-step: 1)-1, preparation and selection of brain imaging data: at least two groups of G1 and G2 test images; for each individual subject, two images are prepared. Magnetic resonance images include but are not limited to T1-weighted images, T2-weighted images, FLAIR, DWI, ADC, and SWI, and one modality is selected as the image for subsequent lesion segmentation; the second image is "brain structure image", which is selected from T1-weighted images ; Both images can be the same; 1)-2, Lesion segmentation: For the "lesion display image", use image segmentation software, toolkits and segmentation algorithms, including but not limited to workstations and segmentation software provided by MRI manufacturers, MRIcro, MRIcron, MIPAV, ITK-SNAP, MITK Or the regional growth algorithm, segment the whole brain lesion image, and save it as a binary image or data to generate a "lesion image". 5. by the described method of claim 1, it is characterized in that, described step 5) comprises the extraction method of following feature set: 5)-1, distribution size features: using individual space "lesion images", according to the principle of inter-voxel connectivity, extract a single 2D or 3D lesion, and calculate the volume and longest diameter size features of a single lesion image; obtain all The maximum value, average value, sum, median, number of lesions, and statistical features of the number of large lesions exceeding 200 mm ³ of the above features of a single lesion image are used as the feature set of the distribution size of the lesion image for the subject; 5)-2, Symmetrical features: Using the "symmetrical spatial lesion image" and the corresponding partition template, in the whole brain and within each group of symmetrical brain partitions, respectively calculate the symmetrical features of the lesion image in the left and right hemispheres, including unilateral/ Bilateral, left/bilateral/right are marked as -1/0/1 respectively, the difference or ratio of the number of voxels of the lesion image in the left and right hemispheres, and the Dice coefficient to measure the similarity of distribution; The Dice coefficient of i is defined as the ratio of the ratio of the number of voxels in the intersection of the two lesions to the number of voxels in the union after the left-side (VL) or right-side (VR) images of symmetrical lesions are flipped left and right. times, that is D(i)=2×|VL∩VR|/|VL∪VR| 5)-3, Probability distribution map features: For each group of subjects, use the binarized "standard spatial lesion image" of all individuals in the group, calculate the number of non-zero values in each voxel in the group, and divide by the individuals in the group The total number of lesions is obtained to obtain the spatial probability distribution map of the group of lesions, in which the voxel value ranges from 0 to 1, and the closer to 1, the higher the probability of the voxel lesion exists. The characteristic value of the probability distribution map of the group g of the lesion of a subject i in the M groups is defined as the number of voxels in the intersection of the lesion area of i and the high probability distribution lesion area of the group g. The difference between the number and the sum of the number of intersections of this lesion and other groups with high probability distribution lesions:

Following steps 5)-4-6, you can use "lesion image" and "individual space partition image" to analyze in individual space, or use "standard space lesion image" and partition template of standard space to analyze in standard space, or two both; 5)-4, distribution characteristics of a single brain region: for each individual brain region, calculate the distribution features of the subject's brain lesion image in the region: whether the lesion exists in the region, the presence is 1, the absence is 0, The volume of the lesion in the partition, the proportion of the volume of the lesion in the partition to the volume of the entire partition; 5)-5, distribution characteristics of parabrain areas: for the landmark brain areas of interest, such as ventricle, midbrain aqueduct, corpus callosum, and white matter area, perform small regional expansion operations, no more than 5mm, to generate parabrain areas , and calculate the para-brain distribution characteristics of the tested lesions for the corresponding brain regions, including the paraventricular, midbrain aqueduct, near the corpus callosum, and near the white matter area, including the presence of para-brain lesions, with the presence of 1 and the absence of 0 , and the volume of adjacent lesions in the brain region; 5)-6, cross-brain distribution characteristics: According to the results of steps 5)-4, calculate the cross-brain distribution characteristics of the brain lesion images of the subjects, including the types of brain tissue covered: the number of gray matter/white matter/cerebrospinal fluid and Pairwise ratio, covering the number of brain regions. 6. by the described method of claim 1, it is characterized in that, described step 6) comprises following substep: 6)-1, feature screening: on the basis of the feature set of the spatial distribution of the lesion image extracted in step 5), various feature screening methods in machine learning are used to screen features; the feature screening methods used include variance selection, chi-square test, U test, mutual information, recursive elimination method, feature selection method or exhaustive method of logistic regression model based on L1 penalty term/L2 penalty term/L1 combined with L2 penalty term; the LASSO method can simultaneously perform feature screening and generate classification identification model so that steps 6-(2) can be incorporated; 6)-2, establish a classification and identification model: take the screened image features of the tested lesions as input, and use the actual subject grouping categories as labels to train the classifier to generate a linear or non-linear classification and identification model; Classifiers include logistic regression, random forests, support vector machines or artificial neural networks; After the model is established, the classification and discrimination effect of the model is evaluated by the AUC value, accuracy, sensitivity, and specificity value of the ROC curve. 7. The method according to claim 6, wherein the feature selection method of the logistic regression model in the step 6)-1 is the LASSO method.

Images