CN112435742B - Neighborhood rough set method for feature reduction of fMRI brain function connection data - Google Patents
Neighborhood rough set method for feature reduction of fMRI brain function connection data Download PDFInfo
- Publication number
- CN112435742B CN112435742B CN202011137128.4A CN202011137128A CN112435742B CN 112435742 B CN112435742 B CN 112435742B CN 202011137128 A CN202011137128 A CN 202011137128A CN 112435742 B CN112435742 B CN 112435742B
- Authority
- CN
- China
- Prior art keywords
- fish
- feature
- brain function
- function connection
- individual
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/0033—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/0033—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room
- A61B5/004—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room adapted for image acquisition of a particular organ or body part
- A61B5/0042—Features or image-related aspects of imaging apparatus classified in A61B5/00, e.g. for MRI, optical tomography or impedance tomography apparatus; arrangements of imaging apparatus in a room adapted for image acquisition of a particular organ or body part for the brain
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/05—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves
- A61B5/055—Detecting, measuring or recording for diagnosis by means of electric currents or magnetic fields; Measuring using microwaves or radio waves involving electronic [EMR] or nuclear [NMR] magnetic resonance, e.g. magnetic resonance imaging
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/40—Detecting, measuring or recording for evaluating the nervous system
- A61B5/4058—Detecting, measuring or recording for evaluating the nervous system for evaluating the central nervous system
- A61B5/4064—Evaluating the brain
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Public Health (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Pathology (AREA)
- Physics & Mathematics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Heart & Thoracic Surgery (AREA)
- Neurology (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Molecular Biology (AREA)
- Data Mining & Analysis (AREA)
- Biophysics (AREA)
- Veterinary Medicine (AREA)
- Radiology & Medical Imaging (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Databases & Information Systems (AREA)
- High Energy & Nuclear Physics (AREA)
- Psychology (AREA)
- Neurosurgery (AREA)
- Physiology (AREA)
- Farming Of Fish And Shellfish (AREA)
- Magnetic Resonance Imaging Apparatus (AREA)
Abstract
The invention discloses a neighborhood rough set feature reduction method for performing feature reduction on fMRI brain function connection data, and provides a neighborhood rough set feature reduction method combined with fMRI brain function connection data based on double-fish-swarm intelligent search and information view neighborhood rough set theory. The method specifically comprises the following steps: acquiring resting fMRI data; preprocessing fMRI data; establishing a brain function connection decision table; searching a feature subset with strong classification and discrimination capabilities by using a double fish swarm algorithm; under the obtained reduced feature set, support vector machine classification is realized, and the feature scale Jian Nengli of the method provided by the invention is measured according to the classification result.
Description
Technical Field
The invention relates to a method for reducing the feature of fMRI brain function connection data, in particular to a method for reducing the feature of brain function connection based on a neighborhood rough set.
Background
The brain function connection characterizes the dynamic association of the neuron activities among different brain areas, and provides a brand-new visual angle for people to understand the pathological mechanism of nerve and mental diseases. In recent years, functional magnetic resonance imaging (functional Magnetic Resonance Imaging, fMRI) based brain functional connection classification has been a great interest for researchers because important brain functional connection characteristics related to a certain brain disease can be found, and the classification has been a great interest for understanding the pathogenesis of a brain disease and further for early diagnosis and treatment of a brain disease. However, the high-dimensional small sample characteristics of fMRI brain functional connection data are prone to cause dimension disaster problems, and provide great challenges for classification methods. To address this challenge, feature reduction of fMRI brain function connection data is a key research topic in brain function connection classification research.
In recent years, researchers have proposed several fMRI brain function connection feature reduction methods. These methods can be divided into two main categories, feature extraction and feature selection. The feature extraction maps the original brain function connection data from a high-dimensional space to a low-dimensional space through a certain transformation, so as to realize the feature dimension reduction targets, such as a principal component analysis (Principal Component Analysis, PCA), a local linear embedding (Locally Linear Embedding, LLE), an Independent component analysis (Independent ComponentAnalysis, ICA) and the like. Such methods have a natural limitation in that the resulting features are not easily understood and do not lend themselves to neurological interpretation. The feature selection directly selects important features from the original brain function connection features according to a certain measurement index, such as T test, F-score, correlation coefficient and the like. Although the method can maintain the meaning of the original brain function connection characteristics and is helpful for people to understand, the method mostly neglects the correlation between the brain function connection characteristics, so that the classification learning performance is also influenced.
Classical Rough Set (Rough Set) theory is a mathematical theory that has been proposed by the poland math z. Pawlak in the 80 s of the 20 th century to deal with imprecise, inconsistent, incomplete knowledge. The method can obtain the core knowledge of the data without providing any priori information, realizes feature reduction, and is widely applied to the fields of machine learning, data mining, pattern recognition and the like at present. However, this theory is only applicable to discrete data, and when continuous data is faced, discretization of the data is required, and the discretization causes serious information loss. For this reason Hu Qinghua et al propose a neighborhood rough set theory based on classical rough set theory and neighborhood concepts, which can well process continuous data.
Disclosure of Invention
Aiming at the defects of the method for reducing the brain function connection characteristics, and considering that fMRI brain function connection shows the degree of functional correlation among different brain regions, the invention provides a neighborhood rough set method for reducing the characteristics of fMRI function connection data. The method fully utilizes the advantages of the neighborhood rough set theory, adopts the double-fish-swarm algorithm to comprehensively explore the feature subset space, and overcomes the defect that the traditional rough set theory is easy to fall into local optimum by using a local search method. Compared with other methods, the method can more effectively reduce the dimension of the fMRI brain function connection data, and the obtained reduction features can ensure the performance of fMRI brain function connection classification.
The main idea for realizing the invention is as follows: acquiring resting fMRI data; preprocessing fMRI data; establishing a brain function connection decision table; searching a feature subset with strong classification and discrimination capabilities by using a double fish swarm algorithm; under the obtained reduced features, support vector machine classification is realized, and the feature reduction Jian Nengli of the invention is measured according to classification performance.
A neighborhood rough set method for brain function connection feature reduction from fMRI data comprising the steps of:
(step 1) fMRI data acquisition; the present experiment used the following two published brain disease resting state fMRI datasets: attention deficit hyperactivity disorder alliance (Attention Deficit Hyperactivity Disorder, ADHD) dataset and autism brain imaging data exchange alliance (Autism Brain Imaging Data Exchange, ABIDE) dataset. ADHD included both Attention Deficit Hyperactivity Disorder (ADHD) and Normal (NC) groups tested; ABIDE contained Autism (Autism) and normal two groups of subjects.
(step 2) fMRI data preprocessing; to remove the interference and obtain brain function connection data, the raw fMRI data needs to be preprocessed. Firstly, performing interlayer time correction, head motion correction, spatial standardization, gaussian smoothing, drift removal, filtering, image registration and other image preprocessing operations on fMRI data by using a statistical parameter map software package SPM 8; then 90 brain regions of the cortex were selected according to AAL (AnatomicalAutomatic Labeling) and the average time series of the 90 brain regions was extracted; finally, the pearson correlation coefficient of the average time sequence of every two brain regions is calculated, and a 90 multiplied by 90 brain function connection matrix is obtained. It is apparent that the brain function connection matrix is a symmetric matrix, so the original brain function connection number is 89 x 90/2=4005.
(step 3) establishing a brain function connection decision table; each tested brain functional attachment feature and label together form a sample. The brain function connection decision table is then denoted (U, a, d), where U is the set of all sample components, a is the set of all brain function connection features, and d denotes the decision feature of whether or not a disease is present.
(step 4) searching a feature subset with strong classification and discrimination capabilities by using a double fish swarm algorithm; the method comprises the following steps: initializing parameters; generating an initial fish school; calculating an adaptation value of each fish individual, wherein the size of the adaptation value reflects the classification discrimination capability of the corresponding feature subset, and the calculation method of the adaptation value is related to neighborhood mutual information and the number of features in the feature subset; before each iteration, the fish shoal is divided into elite fish shoal and common fish shoal in average; the elite fish subgroup executes a foraging mechanism to randomly search for a brain function connection feature subset with strong classification discrimination capability in a visual field range; the common fish subgroup selects and executes a mechanism which can greatly improve the adaptation value from an aggregation mechanism and a rear-end collision mechanism, and searches for a brain function connection feature subset with strong classification and discrimination capability under the guidance of excellent fish individuals; thereafter, both the elite and normal fish herds perform a swallowing mechanism to select excellent individuals in the normal fish herd; at the end of each iteration, two fish shoals mutually merge and exchange information and judge whether to continue the next generation search, and when the iteration search is finished, the optimal fish individual, namely the optimal brain function connection feature subset is output, namely the reduction features obtained by the invention.
(step 5) support vector machine classification; under the obtained reduced features, a support vector machine classification model is learned, testing is carried out on a test set to obtain four performance index values of the classification model, and the performance of the method is evaluated according to the four performance index values, so that the method can find a feature subset with stronger classification and discrimination capabilities.
Compared with the prior art, the invention has the following obvious advantages and beneficial effects;
(1) The method applies the neighborhood rough set to carry out brain function connection feature reduction, considers the association between different features, and has stronger classification discrimination capability because of no feature transformation, and is easy to understand and has application in giving neurological explanation.
(2) According to the invention, the neighborhood rough set is utilized to reduce the brain function connection characteristics, discretization of brain function connection data is not needed like classical rough set theory, information loss caused by discretization is avoided, and a brain function connection characteristic subset with stronger classification and discrimination capability can be obtained.
(3) According to the invention, the candidate brain function connection feature subset is searched by using the double-fish swarm algorithm, so that the defect that the classical rough set theory is easy to fall into local optimum by adopting a local searching method is effectively overcome, and the performance of feature subset searching is ensured.
Drawings
Figure 1 is a flow chart of a method according to the invention.
FIG. 2 is a flow chart of the construction of a brain function connection data decision table.
Fig. 3 is a flow chart of a double fish school feature subset based search.
Detailed Description
Specific embodiments and detailed steps of the invention are set forth below by the disclosed real attention deficit hyperactivity disorder dataset ADHD and autism dataset ABIDE (flow chart shown in FIG. 1):
and (3) obtaining fMRI data in the step (1).
This experiment used two published sets of brain disease resting state fMRI data: attention deficit hyperactivity disorder alliance (Attention Deficit Hyperactivity Disorder, ADHD) dataset and autism brain imaging data exchange alliance (Autism Brain Imaging Data Exchange, ABIDE) dataset. ADHD comprises two groups of subjects with Attention Deficit Hyperactivity Disorder (ADHD) and Normal (NC) and acquires a website of http:// fcon\1000. Subjects. Nitrc. Org\indi/ADHD 200; ABIDE contains Autism (Autism) and normal two groups of subjects, and the acquired website is: http:// fcon \1000. Projects. Nitrc. Org/indi/abide. The statistics associated with the two data sets are shown in table 1.
Table 1 data set statistics
And (3) preprocessing fMRI data.
In order to remove the interference and obtain brain function connection data, the ABIDE data needs to be preprocessed. Firstly, image preprocessing operations such as ABIDE data line interlayer time correction, head motion correction, spatial standardization, gaussian smoothing, drift removal, filtering, image registration and the like are facilitated by a statistical parameter map software package SPM 8; then 90 brain regions of the cortex were selected and the average time series of 90 brain regions were extracted according to AAL (AnatomicalAutomatic Labeling); finally, the pearson correlation coefficient of the average time sequence of every two brain regions is calculated, and a 90 multiplied by 90 brain function connection matrix is obtained. It is apparent that the brain function connection matrix is a symmetric matrix, so the original brain function connection number is 89 x 90/2=4005.
And (3) establishing a brain function connection decision table.
Each tested brain functional attachment feature and label together form a sample. The brain function connection decision table is then denoted (U, a, d), where u= { U 1 ,u 2 ,...,u N The sample set (i.e., data set) consisting of N samples, a= { a 1 ,a 2 ,...,a n And d represents whether the sample has autism decision feature or not. When the feature set A generates a group of neighborhood relations, the neighborhood decision system is called (U, A, d). The overall process of creating a decision table is shown in fig. 2.
And (4) searching a feature subset (shown in fig. 3) with strong classification and discrimination capabilities by using the double fish swarm algorithm.
Initializing parameters; the number P of fish individuals represents the size of the fish shoal, and the value of the fish individuals is required to ensure the convergence of the shoal and does not cause excessive redundancy; maximum number of iterations T max The method comprises the steps of carrying out a first treatment on the surface of the A neighborhood radius delta; the maximum number Try_number of fish individuals attempting to move in the foraging mechanism determines the exploration degree around a candidate solution; visual field range V of fish individuals in aggregation mechanism max And a view weight ω; eliminating coefficient eta in the swallowing mechanism; two weight factors beta and gamma in the adaptation function.
Step (4.2) generating an initial fish school; since brain function connections weigh the functional connection patterns between different nodes (nodes may be brain regions, voxels, regions of interest, etc.), the number of which is equal to the square of the number of nodes, brain function connection data has a high dimension. However, brain function connections associated with a certain brain disease may be only a small fraction of all brain function connections, so brain function connection data contains a large number of redundant features. In order to preliminarily exclude some redundant brain function connection features, the search space of the fish shoal is reduced, the feature reduction efficiency is improved, the neighborhood mutual information of each brain function connection feature and the class feature is calculated firstly, and the importance of each brain function connection feature on classification is measured according to the neighborhood mutual information; then, selecting front top features as a primary brain function connection feature set, marking as C, and marking the feature number contained in C as n; and finally, searching the brain function connection feature subset with strong classification and discrimination capability from the n brain function connection feature sets by using a fish swarm algorithm. In the search, each artificial fish is represented as an n-dimensional binary vector X. If the j-th component value of the vector is 1, the j-th feature in the C is selected into the candidate feature subset; in order to quickly complete the initialization of the fish school at the beginning of the search, each individual fish is randomly initialized to have only one n-dimensional binary vector X with a component value of 1.
Step (4.3) calculating individual adaptation values of the fishes; the searching of the double fish shoal is completed under the guidance of an adaptation function, and people hope to obtain the characteristic subsets with the classification discrimination capability as strong as possible and the characteristic number as few as possible, so that the following adaptation function which simultaneously considers the characteristic subset classification discrimination capability and the characteristic number is defined to evaluate fish individuals:
wherein f (X) i ) Is fish individual X i Is adapted to the value of (a); b is fish individual X i The corresponding brain function connection feature subset; beta and gamma are two weight factors used for controlling influence degree of neighborhood mutual information and feature quantity of feature subsets on an adaptive function; NMI (B, d) is neighborhood mutual information of the intentional decision feature d of the feature subset B, and is used for measuring the classification discrimination capability of the feature subset B, and is defined as follows:
the meaning of each character in the formula is as described above.
Step (4.4) dividing fish populations; before each iteration, the fish shoals are sorted in descending order of the adaptive value, the first half of the fish individuals are divided into elite fish subgroups, and the second half of the fish individuals are divided into normal fish subgroups.
Step (4.5), executing a foraging mechanism by the elite roe group; the foraging mechanism causes each individual elite to swim around itself randomly to find a high fitness location. The specific implementation process is as follows: for any elite fish individual i, firstly calculating a corresponding candidate solution X i Regarding the neighborhood mutual information of the decision feature, it is compared with the neighborhood mutual information of the initial selection feature set C regarding the decision feature. If the two are equal, then the current candidate solution X i A bit is randomly selected from the component with the median value of 1, and the value is changed to 0, namely, a feature is randomly deleted in the current candidate solution to try to further reject redundant features. Otherwise, in the current candidate solution X i A bit is randomly selected from the components with the median value of 0, and the value of the bit is changed to 1, namely, a feature which is not in the feature subset is randomly added to the current brain function connection feature subset. And obtaining a new candidate solution after the movement. And then the elite fish individual i tries the Try_number for the movement to obtain Try_number candidate solutions. The current solution X from Try_number candidate solutions and elite individual i i Selecting the solution with the highest adaptation value as a new solution X after i forages of elite fish individuals inew 。
Step (4.5) the ordinary shoal tries to execute an aggregation mechanism; in the implementation process of the aggregation mechanism, the field-of-view range center and the global optimal solution X of the common fish subgroup best Two kinds of heuristic information guide the individual search of the common fish to accelerate the pace of searching for excellent candidate solutions. The specific process of the common fish individual i executing the aggregation mechanism is as follows: firstly, calculating the distance between the fish and other common fish individuals, and finding out all the common fish individuals in the visual field range; then determining the central position of the visual field range and the global optimal solution, and updating the own weatherAnd (5) selecting a solution. The method for measuring the distance between fish individuals, the definition of the visual field range, the definition of the central position and the realization mode of moving to the central position are key to the realization of the aggregation mechanism. These key points are described below, respectively.
Jaccard distance between fish individuals: in view of sparsity of brain function connections, jaccard distance is more suitable for measuring the distance between fish individuals, and the formula is:
wherein E and H each represent an individual fish X i And X j The corresponding brain function connection feature subset.
Adaptive field of view: to balance the global and local search capabilities of the algorithm, the field of view is automatically adjusted herein using an adaptive linear function, expressed as follows:
where t is the current iteration number. In the early stages of fish shoal searches, the field of view is relatively large and individuals can explore excellent solutions over a large range. As the search progresses, the field of view gradually decreases and fish individuals carefully seek for superior solutions in smaller areas.
Center of field of view: n in the visual field f Individual fish isCenter of fish school X c =(x c,1 ,x c,2 ,...,x c,n ) Wherein each component is calculated according to the following two equations:
wherein x 'is' c,k Is X' c Is the kth component of (c).
Moving toward the center of the field of view: when the center of the visual field is determined, the fish individual i performing the aggregation mechanism obtains a new solution according to the following formula
In the method, in the process of the invention,is->Is the kth component of (c).
Step (4.6) the ordinary shoal tries to execute a rear-end collision mechanism; in the rear-end collision mechanism, the optimal individual X in the visual field range max And a global optimal solution X best Guiding the search of the common fish individuals and accelerating the convergence of the method. The specific implementation mode is that for common fish individual X i Firstly, all artificial fish in the visual field range are determined, and the artificial fish X with the best adaptation value is found max The method comprises the steps of carrying out a first treatment on the surface of the Then a new solution is obtained according to the following two formulas
Step (4.7), executing a swallowing mechanism by the two shoals of fish; to ensure that the shoal search jumps out of local optimum, both shoals have to perform a swallowing mechanism after each iteration. The specific implementation method comprises the following steps: selecting and reinitializing a certain number of fish individuals with low adaptability, wherein the number of reinitialized individuals is as follows:
in the formula, iter represents the iteration times of the optimal individual adaptation value, and the meaning of other symbols is the same as that of the above description. The number of the eliminated fish individuals can be adaptively adjusted according to the searching process, and the more the optimal individual adaptation value is unchanged, the more the fish individuals are subjected to initialization again, so that more opportunities are given to the fish shoals to jump out of local optimal.
Step (4.8) updating two shoal population: at the end of each iteration, two shoals of fish can fuse and exchange information with each other and judge whether search is completed. If the searching is not completed, the fused fish shoal is divided into elite fish subgroups and common fish subgroups according to the adaptation value, and the searching is continued for new candidate brain function connection feature subsets. And if the search is completed, outputting an optimal fish individual, namely an optimal brain function connection feature subset.
(step 5) support vector machine classification; the data set is partitioned in a 10-fold cross-validation manner. 9 parts in the data set are used as training sets, and under the obtained reduced characteristics, the support vector machine classification model is learned. Testing the classification model on 1 part in the data set to obtain four performance index values of Accuracy (Accumey), precision (Precision), recall rate (Recall) and F-measure (F-measure) of the classification model, evaluating the performance of the invention based on the four performance index values, and finding out the feature subset with stronger classification discrimination capability by the method.
TABLE 2 Performance index values of different reduction algorithms on ADHD and ABIDE datasets
NMIDFSA in table 2 represents the method provided by the present invention. Table 2 shows four evaluation index values of NMIDFSA and five comparison methods, and also lists the Number of reduced features (numbers) obtained for each method. From this table, it can be seen that NMIDFSA gives the best results on both data sets for the four evaluation criteria, and that the number of features in the resulting reduced feature subset is also relatively small. These results indicate that the method of the invention can obtain key features with stronger classification and discrimination capability and smaller number.
Claims (6)
1. A neighborhood rough set method for feature reduction of fMRI brain function connection data, characterized by: the method specifically comprises the following steps of,
step 1: fMRI data acquisition; the following two classes of published brain disease resting state fMRI datasets were used: attention deficit hyperactivity disorder coalition data set ADHD and autism brain imaging data exchange coalition data set ABIDE;
step 2: preprocessing fMRI data; the image preprocessing operation is carried out by the SPM8 which is a statistical parameter map software package; then selecting 90 brain regions of the cerebral cortex according to AAL and extracting an average time sequence of the 90 brain regions; finally, calculating the pearson correlation coefficient of the average time sequence of every two brain areas;
step 3: establishing a brain function connection decision table; the brain function connecting features and the labels of each tested are combined to form a sample; the brain function connection decision table is expressed as (U, A, d), wherein U is a set composed of all samples, A is a set composed of all brain function connection features, and d represents decision features of whether a disease exists or not;
step 4: the double fish swarm algorithm searches for a feature subset with strong classification and discrimination capabilities; the method comprises the following steps: initializing parameters; generating an initial fish school; calculating an adaptation value of each fish individual, wherein the size of the adaptation value reflects the classification discrimination capability of the corresponding feature subset, and the calculation method of the adaptation value is related to neighborhood mutual information and the number of features in the feature subset; before each iteration, the fish shoal is divided into elite fish shoal and common fish shoal in average; the elite fish subgroup executes a foraging mechanism to randomly search for a brain function connection feature subset with strong classification discrimination capability in a visual field range; both elite and normal fish herds perform a swallowing mechanism to select excellent individuals in the normal fish herd; at the end of each iteration, two fish shoals mutually merge and exchange information and judge whether to continue the next generation search, and when the iteration search is finished, an optimal fish individual, namely an optimal brain function connection feature subset, namely a reduction feature is output;
step 5: classifying a support vector machine; under the reduced features, a support vector machine classification model is learned, testing is conducted on a test set, four performance index values of the classification model are obtained, performance evaluation is conducted on the four performance index values serving as the basis, and a feature subset with stronger classification discrimination capability is found.
2. A neighborhood rough set method for feature reduction of fMRI brain function connection data according to claim 1, wherein: in fMRI data acquisition of step 1, ADHD included two groups of subjects, attention deficit hyperactivity disorder ADHD and normal NC; ABIDE contains Autism Autism and normal two groups of subjects.
3. A neighborhood rough set method for feature reduction of fMRI brain function connection data according to claim 1, wherein: in the fMRI data preprocessing of the step 2, in order to remove interference and obtain brain function connection data, ABIDE data is preprocessed; image preprocessing operations include ABIDE data line inter-layer time correction, head motion correction, spatial normalization, gaussian smoothing, drift removal, filtering, and image registration.
4. A neighborhood rough set method for feature reduction of fMRI brain function connection data according to claim 1, wherein: step 3, establishing a brain function connection decision table, wherein each tested brain function connection feature and the label are combined to form one sample; the brain function connection decision table is then denoted (U, a, d), where u= { U 1 ,u 2 ,...,u N The data set is a sample set consisting of N samples, a= { a 1 ,a 2 ,...,a n The } is descriptive of the sampleA feature set consisting of n features, d representing whether an autism decision feature is present; when the feature set A generates a group of neighborhood relations, the neighborhood decision system is called (U, A, d).
5. A neighborhood rough set method for feature reduction of fMRI brain function connection data according to claim 1, wherein: the double fish swarm algorithm in the step 4 searches the characteristic subset with strong classifying and distinguishing capability,
initializing parameters; the number P of fish individuals represents the size of the fish shoal, and the value of the fish individuals is required to ensure the convergence of the shoal and does not cause excessive redundancy; maximum number of iterations T max The method comprises the steps of carrying out a first treatment on the surface of the A neighborhood radius delta; the maximum number Try_number of fish individuals attempting to move in the foraging mechanism determines the exploration degree around a candidate solution; visual field range V of fish individuals in aggregation mechanism max And a view weight ω; eliminating coefficient eta in the swallowing mechanism; two weight factors beta and gamma in the adaptation function;
step (4.2) generating an initial fish school; firstly, calculating neighborhood mutual information of each brain function connection feature and category feature, and measuring importance of each brain function connection feature to classification based on the neighborhood mutual information; then, selecting front top features as a primary brain function connection feature set, marking as C, and marking the feature number contained in C as n; finally, searching a brain function connection feature subset with strong classification and discrimination capability from the n brain function connection feature sets by using a fish swarm algorithm; in searching, each artificial fish is expressed as an n-dimensional binary vector X; if the j-th component value of the binary vector is 1, the j-th feature in C is selected into the candidate feature subset; if the j-th component value is 0, the corresponding feature is not selected into the candidate feature subset; at the beginning of the search, each individual fish is randomly initialized to have only one n-dimensional binary vector X with a component value of 1;
step (4.3) calculating individual adaptation values of the fishes; defining an adaptive function considering the classifying and distinguishing capacity of the feature subset and the number of the features at the same time to evaluate the individual fish;
step (4.4) dividing fish populations; before each iteration, sorting fish shoals in descending order according to the magnitude of the adaptation value, dividing the first half of fish individuals into elite fish subgroups, and dividing the second half of fish individuals into common fish subgroups;
step (4.5), executing a foraging mechanism by the elite roe group; the foraging mechanism enables each elite fish individual to randomly swim around the foraging mechanism to find a position with a high adaptation value; for any elite fish individual i, firstly calculating a corresponding candidate solution X i Comparing the neighborhood mutual information about the decision feature with the neighborhood mutual information about the decision feature of the initial feature set C; if the two are equal, then the current candidate solution X i Randomly selecting one bit from the components with the median value of 1, and changing the value of the bit into 0, namely randomly deleting one feature in the current candidate solution to try to further reject redundant features; otherwise, in the current candidate solution X i Randomly selecting a bit from the components with the median value of 0, and changing the value of the bit into 1, namely randomly adding a feature which is not in the feature subset into the current brain function connection feature subset; obtaining a new candidate solution after the movement; then, the elite fish individual i tries the Try_number for the movement to obtain Try_number candidate solutions; the current solution X from Try_number candidate solutions and elite individual i i Selecting the solution with the highest adaptation value as a new solution after i foraging of the elite fish individual
Step (4.5) the ordinary shoal tries to execute an aggregation mechanism; in the implementation process of the aggregation mechanism, the field-of-view range center and the global optimal solution X of the common fish subgroup best Two pieces of heuristic information guide the individual search of the common fish to accelerate the pace of searching for excellent candidate solutions; the specific process of the common fish individual i executing the aggregation mechanism is as follows: firstly, calculating the distance between the fish and other common fish individuals, and finding out all the common fish individuals in the visual field range; then determining the central position of the visual field range and the global optimal solution, and updating the candidate solution of the user; the method for measuring the distance between fish individuals, the definition of the visual field range, the definition of the central position and the realization mode of moving to the central position are key to the realization of an aggregation mechanism;
step (4.6) common shoal of fishAttempting to execute a rear-end collision mechanism; in the rear-end collision mechanism, the optimal individual X in the visual field range max And a global optimal solution X best Guiding the individual search of the common fish, and accelerating the convergence of the method; the specific implementation mode is as follows: for common fish individual X i First, X is determined i Finding artificial fish X with the best adaptation value for all artificial fish in the visual field range max ;
Step (4.7), executing a swallowing mechanism by the two shoals of fish; in order to ensure the ability of the shoal search to jump out of local optimum, two shoal are required to execute a swallowing mechanism after each round of iteration; the specific implementation method comprises the following steps: selecting and reinitializing individual fish, and reinitializing the number of individual fish;
step (4.8) updating two shoal population: at the end of each iteration, two shoals of fish can mutually fuse and exchange information and judge whether searching is completed or not; if the searching is not completed, the fused fish shoals are divided into elite fish subgroups and common fish subgroups according to the adaptation value, and new candidate brain function connection feature subsets are continuously searched; and if the search is completed, outputting an optimal fish individual, namely an optimal brain function connection feature subset.
6. A neighborhood rough set method for feature reduction of fMRI brain function connection data according to claim 1, wherein: step 5, classifying the support vector machine; dividing the data set by adopting a 10-fold cross validation mode; 9 parts in the data set are used as training sets, and under the obtained reduced characteristics, a support vector machine classification model is learned; testing the classification model on 1 part in the data set to obtain four performance index values of accuracy, precision, recall rate and F measurement of the classification model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011137128.4A CN112435742B (en) | 2020-10-22 | 2020-10-22 | Neighborhood rough set method for feature reduction of fMRI brain function connection data |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011137128.4A CN112435742B (en) | 2020-10-22 | 2020-10-22 | Neighborhood rough set method for feature reduction of fMRI brain function connection data |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112435742A CN112435742A (en) | 2021-03-02 |
CN112435742B true CN112435742B (en) | 2023-10-20 |
Family
ID=74695881
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011137128.4A Active CN112435742B (en) | 2020-10-22 | 2020-10-22 | Neighborhood rough set method for feature reduction of fMRI brain function connection data |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112435742B (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1588363A (en) * | 2004-10-14 | 2005-03-02 | 上海交通大学 | Backward coarse collecting attribute reducing method using directed search |
CN103336790A (en) * | 2013-06-06 | 2013-10-02 | 湖州师范学院 | Hadoop-based fast neighborhood rough set attribute reduction method |
CN103942567A (en) * | 2014-04-22 | 2014-07-23 | 张擎 | Automatic discrimination analysis method of mild cognitive impairment based on support vector machine |
CN106651021A (en) * | 2016-12-16 | 2017-05-10 | 新疆疆天航空科技有限公司 | Plant protection unmanned aerial vehicle formation path planning method based on elite artificial shoal |
CN109190851A (en) * | 2018-10-25 | 2019-01-11 | 上海电机学院 | A kind of optimal configuration algorithm based on the independent wind-light storage microgrid for improving fish-swarm algorithm |
-
2020
- 2020-10-22 CN CN202011137128.4A patent/CN112435742B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1588363A (en) * | 2004-10-14 | 2005-03-02 | 上海交通大学 | Backward coarse collecting attribute reducing method using directed search |
CN103336790A (en) * | 2013-06-06 | 2013-10-02 | 湖州师范学院 | Hadoop-based fast neighborhood rough set attribute reduction method |
CN103942567A (en) * | 2014-04-22 | 2014-07-23 | 张擎 | Automatic discrimination analysis method of mild cognitive impairment based on support vector machine |
CN106651021A (en) * | 2016-12-16 | 2017-05-10 | 新疆疆天航空科技有限公司 | Plant protection unmanned aerial vehicle formation path planning method based on elite artificial shoal |
CN109190851A (en) * | 2018-10-25 | 2019-01-11 | 上海电机学院 | A kind of optimal configuration algorithm based on the independent wind-light storage microgrid for improving fish-swarm algorithm |
Also Published As
Publication number | Publication date |
---|---|
CN112435742A (en) | 2021-03-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Połap | An adaptive genetic algorithm as a supporting mechanism for microscopy image analysis in a cascade of convolution neural networks | |
CN111553127B (en) | Multi-label text data feature selection method and device | |
CN114730463A (en) | Multi-instance learner for tissue image classification | |
CN107133651B (en) | The functional magnetic resonance imaging data classification method of subgraph is differentiated based on super-network | |
CN113408605A (en) | Hyperspectral image semi-supervised classification method based on small sample learning | |
CN114038037B (en) | Expression label correction and identification method based on separable residual error attention network | |
CN111000553A (en) | Intelligent classification method for electrocardiogram data based on voting ensemble learning | |
CN108154924A (en) | Alzheimer's disease tagsort method and system based on support vector machines | |
CN116805533A (en) | Cerebral hemorrhage operation risk prediction system based on data collection and simulation | |
CN117238379A (en) | Storage medium storing gene selection method program | |
CN108256423A (en) | Alzheimer's disease feature extracting method and system based on population correlation coefficient | |
CN117195027A (en) | Cluster weighted clustering integration method based on member selection | |
CN117520914A (en) | Single cell classification method, system, equipment and computer readable storage medium | |
CN112435742B (en) | Neighborhood rough set method for feature reduction of fMRI brain function connection data | |
CN116226629B (en) | Multi-model feature selection method and system based on feature contribution | |
CN111832645A (en) | Classification data feature selection method based on discrete crow difference collaborative search algorithm | |
CN113723485B (en) | Hypergraph processing method for brain image of mild hepatic encephalopathy | |
Angayarkanni | Predictive analytics of chronic kidney disease using machine learning algorithm | |
CN112200224B (en) | Medical image feature processing method and device | |
CN110941994B (en) | Pedestrian re-identification integration method based on meta-class-based learner | |
CN115310491A (en) | Class-imbalance magnetic resonance whole brain data classification method based on deep learning | |
CN115098699A (en) | Link prediction method based on knowledge graph embedded model | |
CN114821157A (en) | Multi-modal image classification method based on hybrid model network | |
Cao et al. | Alzheimer’s Disease Stage Detection Method Based on Convolutional Neural Network | |
CN108304546B (en) | Medical image retrieval method based on content similarity and Softmax classifier |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |