Parkinson's disease evolution distinguishing method based on brain structure texture features

Abstract

The invention discloses a Parkinson's disease evolution distinguishing method based on brain structure texture features, and belongs to the technical field of biomedical image mode recognition. The main pathological feature of Parkinson's disease is gradual loss of dopaminergic neurons in the compact part of nigra, but due to the fact that a grey matter signal of the compact part of nigra is too small, the feature extracted according to the size of the signal does not achieve a good classification effect, and most researches also have conflict results. According to the method, the defects caused by too small signals are avoided by investigating the texture features of the compact part of the nigra. After ten-fold cross validation is carried out on test data, the average accuracy rate of classification tasks based on texture features is 71.11%, the area under a subject feature curve is 0.70, and the specificity is 80.00%. The average accuracy rate of classification tasks with the grey matter volume as the characteristic is 58.89%, the area under a subject characteristic curve is 0.53, and the specificity is 68.33%. Compared with the grey matter volume, the performance indexes of the textural features are improved, the average accuracy rate is improved by 12.22%, the area under a subject characteristic curve is improved by 0.17%, the specificity is improved by 11.67%, and the sensitivity is improved by 18.89%.

Description

Technical Field

[0001] The present invention belongs to the technical field of biomedical image pattern recognition, and specifically relates to functional magnetic resonance brain image data processing, image texture feature extraction and differentiation of Parkinson's disease evolution degree. Background Art

[0002] Parkinson's disease is the second largest neurodegenerative disease in the world, affecting more than 10 million people worldwide and is irreversible. As the disease progresses, Parkinson's patients will gradually lose their motor function. Therefore, scientific staging of Parkinson's disease is helpful for efficient and reasonable clinical grade management and treatment of patients. At present, the staging method of Parkinson's disease mainly uses the Hoehn-Yahr scale, which relies on the subjective observation and experience of doctors, and lacks intelligent disease progression monitoring methods. With the development of medical imaging technology, brain imaging tests are expected to provide objective biological markers for Parkinson's disease staging. The main advantages of magnetic resonance imaging (MRI) over positron emission tomography and single photon emission computed tomography are lower cost, convenient examination, and better suitability for promotion. However, the current research on T1-weighted magnetic resonance imaging technology in distinguishing patients with intermediate and late Parkinson's disease is limited, and the effect is not ideal. Traditional T1-weighted MRI image analysis methods usually evaluate brain regions through indicators such as gray matter volume and signal intensity, but due to the weak gray matter signal in the substantia nigra compacta, a key brain area for Parkinson's disease, the ability to distinguish between mid-stage and late-stage patients is insufficient. Although some studies have attempted to use voxel-based morphological analysis and cortical thickness analysis methods, the results are relatively conflicting. For example, some studies failed to find significant differences between healthy controls and late-stage patients, while others pointed out that mid-stage patients may have reduced gray matter volume in the frontal lobe. Summary of the invention

[0003] In order to solve the shortcomings of the technical background, this method first advances the gray matter signal of the substantia nigra compacta brain area, and performs 3σ normalization on the brain area to avoid the partial volume effect; then the gray matter signal is linearly transformed to 32 gray levels; finally, the texture features of the matrix are extracted based on the gray level co-occurrence matrix of the substantia nigra compacta.

[0004] The technical solution of the present invention is: a method for distinguishing the evolution of Parkinson's disease based on brain structure texture features, the method comprising:

[0005] Step 1: Acquire magnetic resonance structural images;

[0006] Collect T1-weighted magnetic resonance imaging data of the evolution of Parkinson's disease in patients with Parkinson's disease and save the records;

[0007] Step 2: Preprocessing;

[0008] Step 3: Extract texture features;

[0009] Step 4: Differentiation of Parkinson's disease progression prediction;

[0010] Step C1: Randomly extract multiple subsets from the sample with replacement, and the size of each subset is the same as the original data set;

[0011] Step C2: Random feature selection. When building a decision tree, not all texture features are used in the node partitioning process. Instead, a subset is randomly selected from all texture features, and the number of features in the subset is greater than 11.

[0012] Step C3: Root node division, based on the selected features of the samples in the subset with replacement each time, the sample classes are divided according to these features;

[0013] Step C4: Recursive partitioning: for each newly generated child node, continue to recursively select features and perform partitioning until the minimum number of leaf node samples is one, then stop the recursive partitioning;

[0014] Step C5: Multiple decision trees are constructed using the method from step C1 to step C4. Each tree will give a mid-term or late-term classification result. The random forest decides the final classification label by voting; that is, the category with the most occurrences is selected as the final prediction result.

[0015] Furthermore, the preprocessing method in step 2 is:

[0016] Step A1: Convert the original DICOM file exported from the MRI machine into a NIFTI file suitable for neuroimaging analysis;

[0017] Step A2: manually adjust the image so that the anterior and posterior joints remain in the same line;

[0018] Step A3: Map the structural image to a standard brain space template;

[0019] Step A4: Use the paranoid field correction technology to correct the signal offset change of the structural image caused by the magnetic field change;

[0020] Step A5: Separate the gray matter signal and the white matter signal in the structural image using the prior probability distribution of each tissue;

[0021] Step A6: Use the Jacobian determinant generated during spatial normalization to correct the gray matter signal inaccuracy caused by volume change.

[0022] Furthermore, the method of step 3 is:

[0023] Step B1: extract the gray matter signal in the region of interest, and perform 3σ normalization on the signal. The formula for calculating the normalized gray matter signal is (x-μ) / σ, where x represents the gray matter signal, μ represents the average value of the gray matter signal, and σ represents the standard deviation of the gray matter signal;

[0024] Step B2: using grayscale linear transformation to adjust the gray matter signal to 32 gray levels;

[0025] Step B3: Calculate the gray level co-occurrence matrix within the region of interest;

[0026] Step B4: Extract texture features of the gray-level co-occurrence matrix.

[0027] Furthermore, the texture features in step B4 include:

[0028] autocorrelation, contrast, correlation1, correlation2, cluster dominance, cluster shadow, dissimilarity, energy, entropy, homogeneity1, homogeneity2, maximum probability, variance, sum mean, sum variance, sum entropy, difference variance, difference entropy;

[0029] Among them, the correlation

[0030]

[0031]

[0032] N g is the number of discrete intensity levels in the image, i and j are the grayscale levels of the image, ∈ is 2.2×10 -16 , p(i, j) is the normalized co-occurrence matrix,

[0033] Relevance

[0034]

[0035] Homogeneity

[0036] Homogeneity

[0037] Aiming at the problem of disease evolution in Parkinson's patients, the present invention designs a system framework for distinguishing the degree of Parkinson's disease evolution based on image texture features. The system extracts subtle texture changes in the patient's T1 image and can distinguish between moderate and severe patients more accurately than gray matter volume.

[0038] The invention discloses a method for distinguishing the evolution of Parkinson's disease based on brain structure texture features. The method belongs to the field of biomedical image pattern recognition technology, and specifically relates to pattern recognition of magnetic resonance image data. The present invention uses the texture features of brain structure to achieve the classification task of patients with mid-to-late stage Parkinson's disease. Compared with traditional gray matter volume features, it has higher classification accuracy, area under the subject characteristic curve and specificity. The present invention proposes a new intelligent method for monitoring the evolution of Parkinson's disease. It is the first time that brain structure texture features are applied to the evolution monitoring of Parkinson's disease, which can be used as an auxiliary and supplementary means for clinicians to monitor the evolution of Parkinson's disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Flowchart for texture classification of patients with advanced Parkinson's disease.

[0040] Figure 2 A comparison of the confusion matrices for classification based on texture features and gray matter volume.

[0041] Figure 3 Comparison of ROC curves for classification based on texture features and gray matter volume.

[0042] Figure 4 Schematic diagram of the location of the substantia nigra pars compacta brain region. DETAILED DESCRIPTION

[0043] First, clinically collect T1 MRI image data of Parkinson's patients; before collecting MRI images, professional physicians complete the differentiation of the disease evolution degree of Parkinson's patients according to the Hoehn-Yahr grading scale;

[0044] Then proceed as follows:

[0045] A. Data Preprocessing

[0046] Step A1: Use DPABI_V8.2_240501 software of MATLAB R2022b to preprocess and convert the source file format from dcm to nii file;

[0047] Step A2: Adjust the image so that the front and back joints remain in the same line.

[0048] Step A3: Use CAT12_V12.9 software to register the T1 image to the Montreal Neurological Institute (MNI) space.

[0049] Step A3: Use the prior tissue distribution probability function on each voxel to segment the gray matter, white matter and cerebrospinal fluid signals.

[0050] Step A4: Modulate the gray matter signal size after registration according to the Jacobian matrix generated during image registration.

[0051] B. Texture feature extraction

[0052] Step B1: The brain structure magnetic resonance imaging data preprocessed in step A is

[0053] Step B2: Use Automated anatomical labelling atlas 3 to obtain the Parkinson's disease degenerative brain area, substantia nigra pars compacta;

[0054] Step B3: 3σ standardization of the substantia nigra pars compacta;

[0055] Step B4: Use grayscale linear change to transform the grayscale value of the substantia nigra compacta to 32 levels.

[0056] Step B5: Calculate the gray-level co-occurrence matrix, the gray-level co-occurrence matrix (GLCM) of the ROI. Only one GLCM is generated after each scan, and the neighboring attributes of the 26-connected neighborhood of all voxels in the ROI are accumulated. To adjust for the difference in discretization length, the neighbors of √3 voxels away from the central voxel contribute a value of √3 to the GLCM, the neighbors of √2 voxels away from the central voxel contribute a value of √2, and the neighbors of 1 voxel away from the central voxel contribute a value of 1.

[0057] Step B6: Calculate 22 features of GLCM, autocorrelation, contrast, correlation 1, correlation 2, cluster dominance, cluster shadow, dissimilarity, energy, entropy, homogeneity 1, homogeneity 2, maximum probability, variance, sum mean, sum variance, sum entropy, difference variance, difference entropy.

[0058] C. Methods for distinguishing the degree of Parkinson's disease evolution

[0059] Step C1. A ten-fold cross validation was performed on the test data, with an average accuracy of 71.11%, an area under the receiver operating characteristic curve of 0.70, and a specificity of 80.00%.

[0060] Step C2. To illustrate the superiority of this feature over traditional features, this test extracts the traditional feature gray matter volume from the image processed by method A and compares it with the texture feature. The comparison results with gray matter volume as the feature are as follows: the average accuracy is 58.89%, the area under the receiver characteristic curve is 0.53, and the specificity is 68.33%.

[0061] Figure 2It shows that using texture features, 48 ​​patients in the middle stage were correctly classified, and 16 patients in the late stage were correctly classified; using gray matter volume, 41 patients in the middle stage were correctly classified, and 12 patients in the late stage were correctly classified. The texture features used in the present invention are superior to the traditional features of gray matter volume. Figure 3 This shows that in the classification task, texture features are superior to gray matter volume in terms of two evaluation indicators: area under the curve and accuracy.

Claims

1. A method for distinguishing the evolution of Parkinson's disease based on brain structure texture features, the method comprising: Step 1: Acquire magnetic resonance structural images; Collect T1-weighted magnetic resonance imaging data of the evolution of Parkinson's disease in patients with Parkinson's disease and save the records; Step 2: Preprocessing; Step 3: Extract texture features; Step 4: Differentiation of Parkinson's disease progression prediction; Step C1: Randomly extract multiple subsets from the sample with replacement, and the size of each subset is the same as the original data set; Step C2: Random feature selection. When building a decision tree, not all texture features are used in the node partitioning process. Instead, a subset is randomly selected from all texture features, and the number of features in the subset is greater than 11. Step C3: Root node division, based on the selected features of the samples in the subset with replacement each time, the sample classes are divided according to these features; Step C4: Recursive partitioning: for each newly generated child node, continue to recursively select features and partition until the minimum number of leaf node samples is one, then stop the recursive partitioning; Step C5: Multiple decision trees are constructed using the method from step C1 to step C4. Each tree will give a mid-term or late-term classification result. The random forest decides the final classification label by voting; that is, the category with the most occurrences is selected as the final prediction result.

2. A method for distinguishing Parkinson's disease evolution based on brain structure texture features as claimed in claim 1, characterized in that: The preprocessing method in step 2 is: Step A1: Convert the original DICOM files exported from the MRI machine into NIFTI files suitable for neuroimaging analysis; Step A2: manually adjust the image so that the anterior and posterior joints remain in the same line; Step A3: Map the structural image to a standard brain space template; Step A4: Use the paranoid field correction technology to correct the signal offset change of the structural image caused by the magnetic field change; Step A5: Separate the gray matter signal and the white matter signal in the structural image using the prior probability distribution of each tissue; Step A6: Use the Jacobian determinant generated during spatial normalization to correct the gray matter signal inaccuracy caused by volume change.

3. The method for distinguishing the evolution of Parkinson's disease based on brain structure texture features as claimed in claim 1, characterized in that: The method for step 3 is: Step B1: extract the gray matter signal in the region of interest, and perform 3σ normalization on the signal. The formula for calculating the normalized gray matter signal is (x-μ) / σ, where x represents the gray matter signal, μ represents the average value of the gray matter signal, and σ represents the standard deviation of the gray matter signal; Step B2: using grayscale linear transformation to adjust the gray matter signal to 32 gray levels; Step B3: Calculate the gray level co-occurrence matrix within the region of interest; Step B4: Extract texture features of the gray-level co-occurrence matrix.

4. A method for distinguishing the evolution of Parkinson's disease based on brain structure texture features as claimed in claim 3, characterized in that: The texture features in step B4 include: autocorrelation, contrast, correlation1, correlation2, cluster dominance, cluster shadow, dissimilarity, energy, entropy, homogeneity1, homogeneity2, maximum probability, variance, sum mean, sum variance, sum entropy, difference variance, difference entropy; Among them, the correlation N g is the number of discrete intensity levels in the image, i and j are the gray levels of the image, ∈ is 2.2×10 -16 , p(i, j) is the normalized co-occurrence matrix, Relevance Homogeneity Homogeneity

Images