Semi-supervised feature selection method for brain image of obsessive-compulsive disorder patient based on pheromone driving

Abstract

The invention provides an obsessive-compulsive disorder patient brain image semi-supervised feature selection method based on pheromone driving. The method comprises the following steps: step 1, reading obsessive-compulsive disorder patient brain image data; 2, determining two adaptive neighborhood spaces, selecting intersection parts of the spaces to form convex ball information particles, and forming a convex ball-based fuzzy similarity relationship between samples; 3, defining a supervised evaluation function and an unsupervised evaluation function, and evaluating feature performance from four different dimensions, namely correlation between features and decisions, joint information between the features, redundancy and interaction; and step 4, integrating evaluation results of four dimensions through a multi-criterion decision MCDM algorithm, generating comprehensive evaluation information, inputting the comprehensive evaluation information into an ant colony algorithm, and predicting an important brain region of the obsessive-compulsive disorder patient. According to the method, prediction of obsessive-compulsive disorder patients in the semi-supervised field is facilitated, the classification efficiency and precision are improved by removing redundant features, and a more accurate tool is provided for doctors and researchers.

Description

Technical Field

[0001] This invention belongs to the field of semi-supervised feature selection of medical images, and particularly relates to a semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients based on pheromone-driven methods. Background Technology

[0002] Obsessive-compulsive disorder (OCD) is a common mental illness characterized by persistent, disturbing obsessive thoughts or images, accompanied by repetitive behavioral patterns or rituals in an attempt to alleviate these unsettling feelings. This condition not only causes emotional discomfort, such as anxiety and depression, but can also significantly impact a patient's physical and mental health, severely affecting daily life and work in severe cases. Studies indicate that approximately 2.3% of the global population experiences OCD, with a lifetime prevalence of about 2% to 3%.

[0003] Currently, in research on feature selection and classification for obsessive-compulsive disorder (OCD), Tan et al., in their paper "Classification of EEGSynchronization Values ​​of Obsessive Compulsive Disorders Patients Using Support Vector Machine Method," proposed using Support Vector Machines (SVM) to classify the GFS values ​​of OCD patients and healthy groups. In this study, GFS was used to extract features associated with OCD, particularly in the theta and delta frequency bands. The study found significant differences in GFS values ​​between OCD patients and healthy controls. Sen In their paper "Sub-graphentropy based network approaches for classifying adolescent obsessive-compulsive disorder from resting-state functional MRI," researchers introduced a novel method for classifying adolescent obsessive-compulsive disorder from resting-state functional MRI data. This method utilizes information theory concepts such as subgraph entropy and marginal entropy, combined with leave-one-out cross-validation and feature selection techniques, to achieve high-precision classification of obsessive-compulsive disorder patients. However, neither of these methods considers the large amounts of unlabeled data available in real-world applications, thus possessing certain limitations.

[0004] Currently, the clinical diagnosis of obsessive-compulsive disorder (OCD) relies primarily on a comprehensive assessment of the patient's symptoms, behavioral manifestations, and cognitive state. This diagnostic method requires physicians to possess extensive clinical experience and involves a degree of subjective judgment during the diagnostic process; therefore, accurate identification of OCD and assessment of the patient's condition are crucial. This diagnostic process has significant research value for developing effective treatment plans. Given the limitations of existing diagnostic methods, exploring a new approach based on big data analysis is particularly urgent. By effectively extracting information about brain regions associated with OCD patients from large datasets and combining it with feature selection techniques, this method can assist physicians in identifying high-risk patients for OCD, opening up new research directions in the field of medical image analysis. Furthermore, this method holds the promise of improving the diagnosis and treatment outcomes of OCD, providing physicians and researchers with more precise tools, and offering more effective medical services to OCD patients. Summary of the Invention

[0005] Purpose of the Invention: Most semi-supervised feature selection methods based on granular computing rely on supervised and unsupervised evaluation functions to evaluate the feature quality of labeled and unlabeled data respectively. These functions are then integrated into a unified criterion, starting from an empty set or a set containing only core attributes, and employing a forward or backward greedy search strategy to select features. However, as the dimensionality increases, it easily gets trapped in local optima, leading to a sharp decline in search efficiency. To overcome these shortcomings, this invention proposes a pheromone-driven semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients. First, an adaptive fuzzy convex sphere is used to construct a fuzzy approximation space. Second, several necessary uncertainty measures are defined from an information theory perspective, such as neighborhood entropy, fuzzy entropy, and fuzzy complementary entropy. Next, several evaluation criteria are constructed based on the uncertainty measures, such as maximum joint information and maximum interaction. These measures can effectively evaluate the uncertainty and information contribution of features. Subsequently, these evaluation criteria are integrated into a multi-criteria decision-making (MCDM) algorithm to optimize the feature selection process. Finally, the results of the integrated evaluation criteria are used as heuristic information and further applied to an ant colony algorithm to select brain regions, improving prediction performance. It has strong application value for intelligent auxiliary diagnosis of obsessive-compulsive disorder patients.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients based on pheromones, comprising the following steps:

[0007] Step 1: Read brain image data from obsessive-compulsive disorder patients, perform preprocessing, and analyze the data;

[0008] Step 2, establish an information granulation method based on the feature level: determine two adaptive neighborhood spaces by using the Pythagorean theorem, and select the intersection of the two adaptive neighborhood spaces to form a convex sphere information granule, thereby replacing the fixed neighborhood value in the traditional method and forming a fuzzy similarity relationship between samples based on the convex sphere.

[0009] Step 3: Based on labeled and unlabeled data, supervised evaluation functions and unsupervised evaluation functions are defined respectively. The performance of features is evaluated from four different dimensions: the correlation between features and decisions, the joint information between features, redundancy and interaction, and the multiple correlations between features are comprehensively considered.

[0010] Step 4: Integrate the evaluation results of the four dimensions in Step 3 using the Multi-Criteria Decision-Making (MCDM) algorithm to generate comprehensive evaluation information. This information is then used as heuristic information to input into the ant colony algorithm to select important features, thereby predicting important brain regions in obsessive-compulsive disorder patients.

[0011] Step 1 includes the following steps:

[0012] Step 1-1: Read the brain image data of obsessive-compulsive disorder patients and convert the data into a decision information system (SDS).<U,C∪D> Where U represents brain imaging data of obsessive-compulsive disorder patients, U = U L ∪U U U L Indicates that there is labeled data, U U U represents unlabeled data. L ≤U U ;

[0013] C = {c1, c2, ..., c} m} represents the conditional feature set in the brain image data of obsessive-compulsive disorder (OCD) patients, m represents the number of features in the image data of OCD patients, and c represents the conditional feature set in the brain image data of obsessive-compulsive disorder (OCD) patients. m This represents the brain image data characteristics of the m-th obsessive-compulsive disorder patient;

[0014] D = {D1,D2,...,D} l} represents a non-empty finite set of category labels for obsessive-compulsive disorder (OCD) patient data, l represents the number of decision categories in the OCD patient data, and D l This represents the l-th decision category.

[0015] Steps 1-2 involve normalizing the unlabeled data in the brain image data of obsessive-compulsive disorder patients using the following formula:

[0016]

[0017] Where c1 refers to the value of a feature in the original sample, c'1 represents the value of a feature in the original sample after normalization, max(c1) represents the maximum value of a feature in all samples of the brain image data of obsessive-compulsive disorder patients, and min(c1) represents the minimum value of a feature in all samples of the brain image data of obsessive-compulsive disorder patients.

[0018] Step 2 includes the following steps:

[0019] Step 2-1, for the brain image data U=U of obsessive-compulsive disorder patients L ∪U U Based on the conditional feature set C = {c1, c2, ..., c m}, calculate the sample center points of all sample data at the feature level. Indicates feature c m The center sample point below;

[0020] Step 2-2: Calculate brain image data of obsessive-compulsive disorder patients. Euclidean distance between dis(x) i ,x j ), x i ,x j It is any sample belonging to U;

[0021] Steps 2-3: Calculate brain image data of obsessive-compulsive disorder patients. and sample center point Euclidean distance between each pair

[0022] Steps 2-4: Construct a model with the current sample x i Centered on the circle, to the center point of the sample A spherical neighborhood with a radius between them Defined as:

[0023]

[0024] Steps 2-5: Construct a system with the sample center point as the reference point. The distance obtained using the Pythagorean theorem with the center as the center. spherical neighborhood with radius Defined as:

[0025]

[0026] Steps 2-6: Select the intersection of two spherical neighborhoods to obtain the current sample x at the feature level. i convex spherical neighborhood δ c (x i ), defined as:

[0027]

[0028] Steps 2-7: Based on the convex sphere neighborhood, obtain brain image data x of obsessive-compulsive disorder patients. i With x j Regarding the formation of new fuzzy similarity relations by conditional feature set C

[0029] In steps 2-4, Defined as:

[0030]

[0031] Where e represents the natural constant.

[0032] Step 3 includes the following steps:

[0033] Step 3-1, for labeled sample x L ,Right now Based on the convex sphere neighborhood, labeled sample features are defined. Mutual information between non-empty finite sets D and their neighborhoods

[0034]

[0035] in, Represents the sample and x in the neighborhood of the convex sphere L The set, Let D(x) represent the b-th sample within the neighborhood of the convex sphere. L () represents a specific class consisting of samples that have the same category;

[0036] Step 3-2, Feature Subset Calculate feature subset C based on neighborhood mutual information. s Maximum correlation N rel (C s ), defined as:

[0037]

[0038] Step 3-3, for unlabeled sample x U ,Right now Based on the fuzzy similarity relationships formed by all unlabeled samples under the conditional feature set C The fuzzy complementary entropy Fce(C) of the conditional feature set C is:

[0039]

[0040] Where, for x i ,x j ∈x U , It is relative to fuzzy similarity. Fuzzy similarity classes, i.e. Based on The resulting fuzzy convex spherical information particles;

[0041] Steps 3-4: Select the feature with the maximum fuzzy complementary entropy according to formula (8). Let the currently selected feature set Use C u This represents the feature set that is not currently selected;

[0042] Steps 3-5, in set C u Select the feature that has the maximum joint information with the currently selected feature set Red. c is C u For a given feature, the maximum joint information JInf is defined as:

[0043]

[0044] in, It is a fuzzy convex spherical information particle formed based on the currently selected feature set Red;

[0045] Steps 3-6, for It is C u One of the features, and Not equal to c, based on complementary mutual information, a minimum redundancy criterion REDU is further introduced, defined as:

[0046]

[0047] Where Red represents the selected (r-1) feature sets, and r represents the total number of unselected features. Based on features The resulting fuzzy convex spherical information particles;

[0048] Steps 3-7, for The maximum interaction INT based on complementary conditional mutual information is defined as:

[0049]

[0050] Where c'∈C u -{c}, c' represents C u Remove a certain feature from feature c. It is a fuzzy convex spherical information particle formed based on feature c'.

[0051] Step 4 includes the following steps:

[0052] Step 4-1: Sort the features based on the maximum relevance, maximum joint information, minimum redundancy criterion, and maximum interaction respectively, and construct the decision matrix DM;

[0053] Step 4-2: Assign a score to each feature based on its ranking. The highest-ranking feature, represented by a value of 1 in matrix DM, will be assigned the maximum score, which is the total number of features m. The second-highest ranked feature (with a value of 2) will receive m-1 points, and the scoring process will continue in this manner until all features have received new scores. Finally, a new decision matrix DM* is obtained.

[0054] Step 4-3: Each row of the decision matrix DM* (representing a feature) corresponds to an alternative. The weight of each feature evaluation method is 1 / 4. By using the multi-criteria decision MCDM algorithm, the features are sorted using the decision matrix DM*, and the quality Q of each feature is evaluated.

[0055] Step 4-4: Input the quality Q as a heuristic vector into the ant colony algorithm (ACO). In the probabilistic method, ant k uses a probabilistic action selection rule based on the following equation.

[0056]

[0057] in, τ represents the probability that ant k reaches feature i at time t; t (c i () represents the pheromone value of feature i at time t. η represents the neighborhood value of ant k based on feature i. t (Q i ) represents the heuristic information of feature i at time t, η t (Q u Let ) represent the heuristic information of feature u at time t, q,q0∈[0,1] are random variables, and the parameters α and β represent the exponential weights used to control pheromones and heuristic information, respectively; here it is assumed that pheromones and heuristic information are equally important, and the β parameter is set to 1.

[0058] Steps 4-5: At the end of each iteration, update the pheromone value τ;

[0059] Steps 4-6: After the Ant Colony Algorithm (ACO) is completed, the features are sorted in descending order according to their final pheromone values, and the brain regions associated with obsessive-compulsive disorder patients are selected from them.

[0060] In step 4-1, the maximum relevance, maximum joint information, and maximum interaction are all sorted in descending order of features, while the minimum redundancy criterion sorts features in ascending order of features.

[0061] In steps 4-5, the pheromone value τ is updated using the following formula:

[0062] τ i (t+1)=(1-ρ)τ i (t)+Δτ i (13),

[0063] Where, τ i (t) represents the pheromone value of feature i at time t, τ i (t+1) represents the pheromone value of feature i at time t+1, Δτ i ρ represents the increment of pheromone after each iteration; ρ represents the pheromone decay rate parameter.

[0064] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0065] The present invention also provides a storage medium storing a computer program or instructions that, when the computer program or instructions are run on a computer, execute the steps of the method described.

[0066] Compared with the prior art, the beneficial effects of the present invention include: (1) The pheromone-driven semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients provided by the present invention is more suitable for the common situation of limited data and scarce labeled data in many practical applications. It reduces the cost of data collection and labeling and improves the generalization ability of the model.

[0067] (2) A novel granulation method is proposed based on the feature level. Two adaptive neighborhood spaces are determined according to the Pythagorean theorem, and a convex spherical neighborhood is further constructed. Weakly correlated information particles are removed by utilizing average distance and geometric properties, thereby forming a corresponding fuzzy similarity relationship for each feature.

[0068] (3) Several evaluation criteria, such as maximum joint information and maximum interaction, were constructed based on uncertainty measures. These metrics can effectively evaluate the uncertainty and information contribution of features. These evaluation criteria were integrated into the multi-criteria decision-making (MCDM) algorithm, and the integrated results were used as heuristic information to further apply to the ant colony algorithm for feature selection. The feature subset selected in this way has higher application value. This method significantly improves feature selection for partially labeled obsessive-compulsive disorder patient data, thereby improving the accuracy and efficiency of intelligent assisted diagnosis. Attached Figure Description

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0070] Figure 1 This is an overall flowchart of the method of the present invention.

[0071] Figure 2 This is a diagram illustrating the overall data processing framework of the method of the present invention. Detailed Implementation

[0072] See Figure 1 and Figure 2 ( Figure 2 In this example, there are n samples in total, x1 represents the first sample, x2 represents the second sample, and there are m features in total, represented by sample x1 and the center sample point x1. c (To illustrate the process of granulation of convex sphere information), this embodiment provides a semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients based on pheromones, including the following steps:

[0073] Step 1: Read brain image data from obsessive-compulsive disorder patients, perform preprocessing, and analyze the data;

[0074] Step 2, establish an information granulation method based on the feature level: determine two adaptive neighborhood spaces by using the Pythagorean theorem, and select the intersection of the two adaptive neighborhood spaces to form a convex sphere information granule, thereby replacing the fixed neighborhood value in the traditional method and forming a fuzzy similarity relationship between samples based on the convex sphere.

[0075] Step 3: Based on labeled and unlabeled data, supervised evaluation functions and unsupervised evaluation functions are defined respectively. The performance of features is evaluated from four different dimensions: the correlation between features and decisions, the joint information between features, redundancy and interaction, and the multiple correlations between features are comprehensively considered.

[0076] Step 4: Integrate the evaluation results of the four dimensions in Step 3 using the Multi-Criteria Decision-Making (MCDM) algorithm to generate comprehensive evaluation information. This information is then used as heuristic information to input into the ant colony algorithm to select important features, thereby predicting important brain regions in obsessive-compulsive disorder patients.

[0077] Step 1 includes the following steps:

[0078] Step 1-1: Read brain image sample data from obsessive-compulsive disorder patients and convert the data into a decision information system (SDS).<U,C∪D> , where U=U L ∪U U A sample set representing brain images of obsessive-compulsive disorder patients, U L This indicates that there is labeled data and U. U To represent unlabeled data, typically U L ≤U UC = {c1, c2, ..., c} m} represents the conditional features in the brain image sample data of obsessive-compulsive disorder patients, m represents the number of features in the image data of obsessive-compulsive disorder patients, and D = {D1, D2, ..., D} l} represents a non-empty finite set of category labels for obsessive-compulsive disorder (OCD) patient data, and l represents the number of decision categories in the OCD patient data, usually divided according to whether there is a category label. L and U U ,and

[0079] Steps 1-2 involve normalizing the unlabeled data in the brain image data of obsessive-compulsive disorder patients using the following formula:

[0080]

[0081] Where c1 refers to the value of a certain feature in the original sample, c1 represents the value of a certain feature in the original sample after normalization, max(c1) represents the maximum value of a certain feature among all samples, and min(c1) represents the minimum value of a certain feature among all samples.

[0082] The original dataset was transformed into a decision information system and normalized. The results are shown in Table 1 below.

[0083] Table 1

[0084] U <![CDATA[a1]]> <![CDATA[a2]]> <![CDATA[a3]]> <![CDATA[a4]]> <![CDATA[a5]]> <![CDATA[a6]]> <![CDATA[a7]]> <![CDATA[a8]]> <![CDATA[a9]]> D <![CDATA[x1]]> 0.2000 0.0371 0.1539 0.1601 0.3109 0.4797 0.5783 0.2611 0.4827 1 <![CDATA[x2]]> 0.0453 0.0523 0.2156 0.3481 0.3337 0.7818 0.5212 0.3098 0.3019 1 <![CDATA[x3]]> 0.0262 0.0582 0.2431 0.3771 0.5598 0.1619 0.7974 0.1567 0.6401 1 <![CDATA[x4]]> 0.0100 0.0171 0.1098 0.1276 0.0598 0.3793 0.5556 0.2629 0.6830 2 <![CDATA[x5]]> 0.0762 0.0666 0.1209 0.2467 0.3564 0.4636 0.4148 0.3176 0.5663 2

[0085] Step 2 includes the following steps:

[0086] Step 2-1, for the brain image data U=U of obsessive-compulsive disorder patients L ∪U U Based on the feature set C = {c1, c2, ..., c m}, calculate the sample center points of all sample data at the feature level.

[0087] In this embodiment, the number of samples is 5 and the number of features is 9. The sample center points corresponding to each feature are shown in Table 2.

[0088] Table 2

[0089]

[0090] Step 2-2: Calculate brain image data of obsessive-compulsive disorder patients. Euclidean distance between dis(x) i ,x j ), x i ,x j It is any sample belonging to U;

[0091] In this embodiment, there are a total of five samples, and the Euclidean distances between the samples are as follows:

[0092]

[0093] Steps 2-3: Calculate brain image data of obsessive-compulsive disorder patients. and sample center point Euclidean distance between each pair

[0094]

[0095] Steps 2-4: Construct a model with the current sample x i Centered on the circle, to the center point of the sample A spherical neighborhood with a radius between two points is defined as follows:

[0096]

[0097] Where x j Dis(x) represents another sample. i ,x j ) is x i and x j The Euclidean distance between them;

[0098] Steps 2-5: Construct a sample centered on the center The distance obtained using the Pythagorean theorem with the center as the center. A spherical neighborhood with radius is defined as follows:

[0099]

[0100] Steps 2-6: Select the intersection of two spherical neighborhoods to obtain the sample x at the feature level. i The convex spherical neighborhood is defined as follows:

[0101]

[0102] For ease of understanding, sample x1 and sample center point in the embodiment are described here. To elaborate on the narrative, neighborhood set

[0103] Steps 2-7: Based on the above convex spherical neighborhood, obtain the obsessive-compulsive disorder patient data x. i With x j Regarding the formation of new fuzzy similarity relationships in feature set C The definition is as follows:

[0104]

[0105] This example calculates the distance to each sample based on the first feature level, determines whether they satisfy the neighborhood relationship, and thus obtains the fuzzy similarity relationship between samples:

[0106]

[0107] Step 3 includes the following steps:

[0108] Step 3-1: First, consider that there are labeled samples in the brain images of obsessive-compulsive disorder patients, i.e. Based on the aforementioned convex spherical neighborhood, labeled sample features are defined. Mutual information between neighbors of category D:

[0109]

[0110] in, D(x) represents the sample and itself within the neighborhood of the convex sphere. L () represents a specific class consisting of samples that have the same category.

[0111] Step 3-2, Feature Subset Calculate feature subset C based on neighborhood mutual information s Maximum correlation N rel (C s ), defined as:

[0112]

[0113] In this embodiment, it is assumed that x1 and x4 are labeled samples, and x2, x3, and x5 are unlabeled samples. That is, only samples x1 and x4 are considered when calculating neighborhood mutual information. In the example, C... s =5, taking the first feature as an example, |D(x L Since |) = 2, the neighborhood mutual information (NMI) is calculated to be 0.0739. Similarly, we obtain N... rel (C s = [0.0739, 0.1696, 0.1070, 0.1138, 0.1152].

[0114] Step 3-3, for unlabeled sample x U ,Right now Based on the fuzzy similarity relationships formed by all unlabeled samples under the conditional feature set C The fuzzy complementary entropy Fce(C) of the conditional feature set C is:

[0115]

[0116] in, It is relative to the fuzzy similarity relation r cFuzzy similarity class.

[0117] Steps 3-4: Select the feature with the maximum fuzzy complementary entropy according to formula (8). Let the currently selected feature set ; Use C u This represents the feature set that is not currently selected;

[0118] In this embodiment, the sample labels are consistent with the above assumptions. Feature C takes the above nine conditional features as an example, where the fuzzy complementary entropy value of each conditional feature is as follows: Fce(C) = [0.7024, 0.7070, 0.7005, 0.7134, 0.6989, 0.6839, 0.5769, 0.6731, 0.6940]. It can be seen that the maximum value is 0.7134, which is the fourth feature. At this time, Red = {c4}

[0119] Steps 3-5, in set C u Select the feature that has the maximum joint information with the currently selected feature set Red. c is C u For a given feature, the maximum joint information JInf is defined as:

[0120]

[0121] In this embodiment, Cu = {c1, c2, c3, c5, c6, c7, c8, c9}. The joint information of each feature and Red is calculated to obtain JInf = [0.9154, 0.9043, 0.9174, 0.9024, 0.9089, 0.9090, 0.9166, 0.9004].

[0122] Steps 3-6, for It is C u One of the features, and Not equal to c, based on complementary mutual information, a minimum redundancy criterion REDU is further introduced, defined as:

[0123]

[0124] Where Red represents the simplified set of the selected (r-1) features, and r represents the total number of unselected features.

[0125] In this embodiment, the redundancy among the nine features in Cu is calculated, and REDU = [0.5409, 0.5500, 0.5315, 0.5462, 0.5367, 0.5361, 0.5280, 0.5442] is obtained.

[0126] Steps 3-7, for The maximum interaction INT based on complementary conditional mutual information is defined as:

[0127]

[0128] In this embodiment, the interaction information between the remaining features and the selected feature c4 is calculated, resulting in INT = [0.1230, 0.1141, 0.1288, 0.1148, 0.1215, 0.1197, 0.1242, 0.1158].

[0129] Step S includes the following steps:

[0130] Step 4-1: Sort the features based on the maximum relevance, maximum joint information, minimum redundancy criterion, and maximum interaction respectively, and construct the decision matrix DM;

[0131] Step 4-2: Assign a score to each feature based on its ranking. The highest-ranking feature, represented by a value of 1 in matrix DM, will be assigned the maximum score, which is the total number of features m. The second-highest ranked feature (with a value of 2) will receive m-1 points, and the scoring process will continue in this manner until all features have received new scores. Finally, a new decision matrix DM* is obtained.

[0132] Four feature evaluation methods are used to evaluate eight features. The rows of the decision matrix DM represent the features, and the columns represent the evaluation methods. Each column indicates the importance ranking of each evaluation method for the feature. The decision matrix DM is shown below:

[0133]

[0134] Given the ranking of each feature, we obtain a new decision matrix DM*, as shown below:

[0135]

[0136] Step 4-3: Each row of the decision matrix DM* (representing a feature) corresponds to an alternative. Since there are four evaluation methods, each feature evaluation method has a weight of 1 / 4. By employing the MCDM technique, the features are ranked using the matrix DM*, thereby evaluating the quality Q of each feature.

[0137] In this embodiment, Q = [0.1336, 0.1219, 0.1123, 0.0997, 0.0832, 0.0658, 0.1103, 0.1229].

[0138] Step 4-4: Input Q as a heuristic vector into the ACO algorithm. In the probabilistic method, ant k uses a probabilistic action selection rule based on the following equation, as shown below.

[0139]

[0140] in, τ represents the probability that ant k reaches feature i at time t; t (c i () represents the pheromone value of feature i at time t. η represents the neighborhood value of ant k based on feature i. t (Q i ) represents the heuristic information of feature i at time t, η t (Q u Let ) represent the heuristic information of feature u at time t, q,q0∈[0,1] are random variables, and the parameters α and β represent the exponential weights used to control pheromones and heuristic information, respectively; here it is assumed that pheromones and heuristic information are equally important, and the β parameter is set to 1.

[0141] Steps 4-5: At the end of each iteration, update the pheromone value τ, as shown in the formula below:

[0142] τ i (t+1)=(1-ρ)τ i (t)+Δτ i (13),

[0143] Where, τ i (t) represents the pheromone value of feature i at time t, τ i (t+1) represents the pheromone value of feature i at time t+1, Δτ i This represents the increment of pheromone after each iteration; ρ represents the pheromone decay rate parameter, which is generally set to 0.1.

[0144] Steps 4-6: After the Ant Colony Algorithm (ACO) is completed, the features are sorted in descending order according to their final pheromone values, and the brain regions associated with obsessive-compulsive disorder patients are selected from them.

[0145] Q is input as prior knowledge into the ant colony algorithm, and k is set to 5. The final output selected by each ant based on probability is Sort = [4 6 9 5 1 2 7 3 8]. According to the obtained feature sequence, the importance of features gradually decreases from front to back. Therefore, by selecting only some of the top features, instead of selecting all features for prediction, efficiency can be effectively improved. This method has a significant improvement effect on feature selection in partially labeled obsessive-compulsive disorder patient data, thereby further improving the accuracy and efficiency of intelligent assisted diagnosis.

[0146] This invention provides a pheromone-driven semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A pheromone-driven semi-supervised feature selection method for brain images of obsessive-compulsive disorder patients, characterized in that, Includes the following steps: Step 1: Read brain image data from obsessive-compulsive disorder patients, perform preprocessing, and analyze the data; Step 2, establish an information granulation method based on the feature level: determine two adaptive neighborhood spaces by using the Pythagorean theorem, select the intersection of the two adaptive neighborhood spaces to form a convex sphere information granule, and form a fuzzy similarity relationship between samples based on the convex sphere; Step 3: Based on labeled and unlabeled data, define supervised and unsupervised evaluation functions respectively, and evaluate feature performance from four different dimensions: correlation between features and decisions, joint information between features, redundancy and interaction. Step 4: Integrate the evaluation results of the four dimensions in Step 3 through the Multi-Criterion Decision Making (MCDM) algorithm to generate comprehensive evaluation information. This comprehensive evaluation information is then used as heuristic information to input into the ant colony algorithm to select important features, thereby predicting important brain regions in obsessive-compulsive disorder patients.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1-1: Read the brain image data of obsessive-compulsive disorder patients and convert the data into a decision information system (SDS).<U,C∪D> Where U represents brain imaging data of obsessive-compulsive disorder patients, U = U L ∪U U U L Indicates labeled data, U U U represents unlabeled data. L ≤U U ; C = {c1, c2, ..., c} m } represents the conditional feature set in the brain image data of obsessive-compulsive disorder (OCD) patients, m represents the number of features in the image data of OCD patients, and c represents the conditional feature set in the brain image data of obsessive-compulsive disorder (OCD) patients. m This represents the brain image data characteristics of the m-th obsessive-compulsive disorder patient; D ={D1,D2,...,D l } represents a non-empty finite set of category labels for obsessive-compulsive disorder (OCD) patient data, l represents the number of decision categories in the OCD patient data, and D l This represents the l-th decision category. Steps 1-2 involve normalizing the unlabeled data in the brain image data of obsessive-compulsive disorder patients using the following formula: Where c1 refers to the value of a feature in the original sample, c'1 represents the value of a feature in the original sample after normalization, max(c1) represents the maximum value of a feature in all samples of the brain image data of obsessive-compulsive disorder patients, and min(c1) represents the minimum value of a feature in all samples of the brain image data of obsessive-compulsive disorder patients.

3. The method according to claim 2, characterized in that, Step 2 includes the following steps: Step 2-1, for the brain image data U=U of obsessive-compulsive disorder patients L ∪U U Based on the conditional feature set C = {c1, c2, ..., c m }, calculate the sample center points of all sample data at the feature level. Indicates feature c m The center sample point below; Step 2-2: Calculate brain image data of obsessive-compulsive disorder patients. Euclidean distance between dis(x) i ,x j ), x i ,x j It is any sample belonging to U; Steps 2-3: Calculate brain image data of obsessive-compulsive disorder patients. and sample center point Euclidean distance between each pair Steps 2-4: Construct a model with the current sample x i Centered on the circle, to the center point of the sample A spherical neighborhood with a radius between them Defined as: Steps 2-5: Construct a system with the sample center point as the reference point. The distance obtained using the Pythagorean theorem with the center as the center. spherical neighborhood with radius Defined as: Steps 2-6: Select the intersection of two spherical neighborhoods to obtain the current sample x at the feature level. i convex spherical neighborhood δ c (x i ), defined as: Steps 2-7: Based on the convex sphere neighborhood, obtain brain image data x of obsessive-compulsive disorder patients. i With x j Regarding the formation of new fuzzy similarity relations by conditional feature set C 4. The method according to claim 3, characterized in that, In steps 2-4, Defined as: Where e represents the natural constant.

5. The method according to claim 4, characterized in that, Step 3 includes the following steps: Step 3-1, for labeled sample x L ,Right now Based on the convex sphere neighborhood, labeled sample features are defined. Mutual information between non-empty finite sets D and their neighborhoods in, Represents the sample and x in the neighborhood of the convex sphere L The set, Let D(x) represent the b-th sample within the neighborhood of the convex sphere. L () represents a specific class consisting of samples that have the same category; Step 3-2, Feature Subset Calculate feature subset C based on neighborhood mutual information. s Maximum correlation N rel (C s ), defined as: Step 3-3, for unlabeled sample x U ,Right now Based on the fuzzy similarity relationships formed by all unlabeled samples under the conditional feature set C The fuzzy complementary entropy Fce(C) of the conditional feature set C is: Where, for x i ,x j ∈x U , It is relative to fuzzy similarity. Fuzzy similarity classes, i.e. Based on The resulting fuzzy convex spherical information particles; Steps 3-4: Select the feature with the maximum fuzzy complementary entropy according to formula (8). Let the currently selected feature set Use C u This represents the feature set that is not currently selected; Steps 3-5, in set C u Select the feature that has the maximum joint information with the currently selected feature set Red. c is C u One of the features, the maximum joint information JInf, is defined as: in, It is a fuzzy convex spherical information particle formed based on the currently selected feature set Red; Steps 3-6, for It is C u One of the features, and Not equal to c, based on complementary mutual information, a minimum redundancy criterion REDU is further introduced, defined as: Where Red represents the selected (r-1) feature sets, and r represents the total number of unselected features. Based on features The resulting fuzzy convex spherical information particles; Steps 3-7, for The maximum interaction INT based on complementary conditional mutual information is defined as: Where c'∈C u -{c}, c' represents C u Remove a certain feature from feature c. It is a fuzzy convex spherical information particle formed based on feature c'.

6. The method according to claim 5, characterized in that, Step 4 includes the following steps: Step 4-1: Sort the features based on the maximum relevance, maximum joint information, minimum redundancy criterion, and maximum interaction respectively, and construct the decision matrix DM; Step 4-2: Assign a score to each feature based on its ranking. The highest-ranking feature, represented by a value of 1 in matrix DM, will be assigned the maximum score, which is the total number of features m. The second-highest ranked feature will receive m-1 points, and the scoring process will continue in this manner until all features have received new scores. Finally, a new decision matrix DM* is obtained. Step 4-3: Each row of the decision matrix DM* corresponds to an alternative, and the weight of each feature evaluation method is 1 / 4. By using the multi-criteria decision MCDM algorithm, the features are sorted using the decision matrix DM*, and the quality Q of each feature is evaluated. Step 4-4: Input the quality Q as a heuristic vector into the ant colony algorithm (ACO). In the probabilistic method, ant k uses a probabilistic action selection rule based on the following equation. in, τ represents the probability that ant k reaches feature i at time t; t (c i () represents the pheromone value of feature i at time t. η represents the neighborhood value of ant k based on feature i. t (Q i ) represents the heuristic information of feature i at time t, η t (Q u ) represents the heuristic information of feature u at time t, q,q0∈[0,1] are random variables, and the parameters α and β represent the exponential weights used to control the pheromone and the exponential weights used to control the heuristic information, respectively. Steps 4-5: At the end of each iteration, update the pheromone value τ; Steps 4-6: After the Ant Colony Algorithm (ACO) is completed, the features are sorted in descending order according to their final pheromone values, and the brain regions associated with obsessive-compulsive disorder patients are selected from them.

7. The method according to claim 6, characterized in that, In step 4-1, the maximum relevance, maximum joint information, and maximum interaction are all sorted by feature order from largest to smallest, while the minimum redundancy criterion sorts features by feature order from smallest to largest.

8. The method according to claim 7, characterized in that, In steps 4-5, the pheromone value τ is updated using the following formula: t i (t+1)=(1-ρ)τ i (t)+Δτ i (13), Where, τ i (t) represents the pheromone value of feature i at time t, τ i (t+1) represents the pheromone value of feature i at time t+1, Δτ i ρ represents the increment of pheromone after each iteration; ρ represents the pheromone decay rate parameter.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program or instructions that, when run on a computer, perform the steps of the method as described in any one of claims 1 to 8.

Images