Cognitive state classification method, device, equipment and storage medium

Abstract

The present invention provides a cognitive state classification method, apparatus, device, and storage medium, comprising: obtaining first quantitative magnetic susceptibility images of the brains of a first group of people with cognitive abnormalities within a target population, and second quantitative magnetic susceptibility images of the brains of a second group of people with normal cognition; extracting first radiomic features corresponding to the first quantitative magnetic susceptibility images, and second radiomic features corresponding to the second quantitative magnetic susceptibility images; determining multiple target features from the first radiomic features based on the first and second radiomic features, wherein the multiple target features differ from the second radiomic features and the correlations between the multiple target features satisfy a set threshold range; and training a cognitive state classification model based on the multiple target features, thereby performing cognitive state classification using the trained cognitive state classification model. This solution utilizes machine learning methods for cognitive state classification, which can effectively improve the accuracy of cognitive state classification results.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a cognitive state classification method, apparatus, device and storage medium. Background Art

[0002] End-stage renal disease (ESRD) refers to the terminal stage of various chronic kidney diseases. Patients with ESRD may experience cognitive abnormalities. Methods such as the Mini-Mental State Examination and the Montreal Cognitive Assessment are commonly used to classify the cognitive status of ESRD patients. However, these methods are often inaccurate because they are influenced by factors such as the educational level of the individuals being assessed. Summary of the Invention

[0003] Embodiments of the present invention provide a cognitive state classification method, apparatus, device, and storage medium to improve the accuracy of cognitive state classification results.

[0004] In a first aspect, an embodiment of the present invention provides a method for classifying cognitive states, the method comprising:

[0005] Acquiring a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population, wherein the target population is a group of people suffering from end-stage renal disease;

[0006] extracting a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image;

[0007] Determining, based on the first radiomics feature and the second radiomics feature, a plurality of target features from the first radiomics feature, wherein the plurality of target features are different from the second radiomics feature, and correlations between the plurality of target features satisfy a set threshold range;

[0008] A cognitive state classification model is trained based on the multiple target features, so as to perform cognitive state classification through the trained cognitive state classification model.

[0009] In a second aspect, an embodiment of the present invention provides a cognitive state classification device, the device comprising:

[0010] an acquisition module, configured to acquire a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population, wherein the target population is a group of people suffering from end-stage renal disease;

[0011] an extraction module, configured to extract a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image;

[0012] A processing module is used to determine multiple target features from the first imaging omics feature based on the first imaging omics feature and the second imaging omics feature, where the multiple target features are different from the second imaging omics feature and the correlation between the multiple target features meets a set threshold range; and train a cognitive state classification model based on the multiple target features to perform cognitive state classification using the trained cognitive state classification model.

[0013] In a third aspect, an embodiment of the present invention provides an electronic device comprising: a memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the cognitive state classification method as described in the first aspect.

[0014] In a fourth aspect, an embodiment of the present invention provides a non-transitory machine-readable storage medium having executable code stored thereon. When the executable code is executed by a processor of an electronic device, the processor can at least implement the cognitive state classification method described in the first aspect.

[0015] In an embodiment of the present invention, first, a quantitative magnetic susceptibility image of the brain of a target population with end-stage renal disease is obtained, wherein the quantitative magnetic susceptibility image of the brain of a first population with cognitive abnormalities in the target population is a first quantitative magnetic susceptibility image, and the quantitative magnetic susceptibility image of the brain of a second population with normal cognition in the target population is a second quantitative magnetic susceptibility image. Thereafter, a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image are extracted. Then, based on the first radiomics feature and the second radiomics feature, multiple target features are determined from the first radiomics feature, wherein the multiple target features differ from the second radiomics feature, and the correlation between the multiple target features satisfies a set threshold range. That is, the multiple target features are distinguishing features between the first radiomics feature and the second radiomics feature, and can be used to reflect the difference between the quantitative magnetic susceptibility images of the brains of a first population with cognitive abnormalities and a second population with normal cognition. Furthermore, the correlations between the multiple target features meet a set threshold range, for example, the correlations between the multiple target features are less than the set threshold. In other words, the features in the first radiomics can be screened based on the relationship between the correlations between the features and the set threshold range, so that only a certain number of target features are ultimately obtained whose correlations between the features meet the set threshold range. Finally, the multiple target features are input into a cognitive state classification model to train the cognitive state classification model based on the multiple target features. It is understood that the fewer the multiple target features, the higher the training efficiency when subsequently training the cognitive state classification model based on the multiple target features.

[0016] In summary, this scheme uses the imaging features reflected by the quantitative magnetic susceptibility images of the brain corresponding to people with cognitive abnormalities and normal cognition to determine multiple target features that affect the cognitive state. Based on the multiple target features, the cognitive state classification model for cognitive state classification is trained using machine learning methods. The cognitive state classification model can more objectively classify the cognitive state of patients with end-stage renal disease, which is not affected by factors such as the patient's education level and can effectively improve the accuracy of the cognitive state classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A flowchart of a cognitive state classification method provided by an embodiment of the present invention;

[0019] Figure 2 A flowchart of a target feature determination process provided by an embodiment of the present invention;

[0020] Figure 3 A flowchart of another target feature determination process provided by an embodiment of the present invention;

[0021] Figure 4 A flowchart of another cognitive state classification method provided by an embodiment of the present invention;

[0022] Figure 5 A schematic diagram of the structure of a cognitive state classification device provided by an embodiment of the present invention;

[0023] Figure 6 This is a schematic structural diagram of an electronic device provided in this embodiment. DETAILED DESCRIPTION

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] The terms used in the embodiments of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The singular forms "a," "the," and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates otherwise. "A plurality" generally includes at least two, but does not exclude the inclusion of at least one.

[0026] It should be understood that the term "and / or" as used herein is merely a description of the relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the associated objects are in an "or" relationship.

[0027] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or system. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or system comprising the element.

[0028] In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0029] The cognitive state classification method provided in the embodiments of the present invention can be performed by an electronic device, which can be a terminal device with data processing capabilities such as a PC, laptop, or smartphone, or a server. The server can be a physical server including an independent host, a virtual server, a cloud server, or a server cluster.

[0030] Figure 1 A flowchart of a cognitive state classification method provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the method includes the following steps:

[0031] 101. Obtain a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population. The target population is people with end-stage renal disease.

[0032] 102. Extract a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image.

[0033] 103. Based on the first radiomics feature and the second radiomics feature, determine multiple target features from the first radiomics feature, the multiple target features are different from the second radiomics feature, and correlations between the multiple target features meet a set threshold range.

[0034] 104. Training a cognitive state classification model based on multiple target features to perform cognitive state classification using the trained cognitive state classification model.

[0035] Cognitive state is usually related to the human brain's cognitive abilities such as perception, memory, and thinking. In practical applications, cognitive states can be divided into multiple types based on different application requirements. For example, cognitive state can be divided into normal cognition, mild cognitive impairment, and cognitive impairment, etc. according to the decline of cognitive abilities such as memory and thinking. For another example, cognitive state can be divided into normal cognition / mild cognitive impairment / cognitive impairment for children, normal cognition / mild cognitive impairment / cognitive impairment for young people, and normal cognition / mild cognitive impairment / cognitive impairment for the elderly according to age groups. Of course, it can also be subdivided into more types based on a single evaluation dimension or multiple evaluation dimensions such as gender and physical condition (such as whether or not suffering from underlying diseases). In this embodiment, the classification method of cognitive state is not restricted.

[0036] Quantitative susceptibility imaging (Quantitative Susceptibility Mapping, QSM) is a kind of technology for the distribution of magnetic susceptibility in quantitative measurement tissue, wherein, magnetic susceptibility is the intrinsic property of material, mainly derived from white matter myelin and iron deposition in brain tissue.In the growth and aging process of brain, the neurodegenerative disease that abnormal brain iron deposition and neuronal structure and function lose gradually is closely connected, and neurodegenerative disease can cause cognitive impairment etc., and that is, brain iron deposition situation can have an impact on people's cognitive state.Therefore, in the present embodiment, when carrying out cognitive state classification, the quantitative susceptibility imaging result (i.e. quantitative susceptibility image) of brain is selected as the basis for analyzing cognitive state category.

[0037] In summary, this approach first uses quantitative magnetic susceptibility images of the brains of individuals with cognitive abnormalities and normal cognition to identify multiple target features that influence cognitive status. Then, based on these target features, a cognitive status classification model is trained using machine learning methods. This allows for more objective classification of the cognitive status of patients with end-stage renal disease, unaffected by factors such as the patient's educational level, effectively improving the accuracy of cognitive status classification results.

[0038] Given that the cognitive status classification method in this embodiment is applied to classify the cognitive status of patients with end-stage renal disease, the quantitative brain susceptibility images obtained are quantitative brain susceptibility images of a target population of patients with end-stage renal disease. The target population includes a first population with cognitive abnormalities and a second population with normal cognition. For ease of distinction, the quantitative brain susceptibility images of the first population are referred to as first quantitative brain susceptibility images, and the quantitative brain susceptibility images of the second population are referred to as second quantitative brain susceptibility images.

[0039] Optionally, in different cognitive status classification scenarios, different groups of people may be selected as target groups, for example, people of a certain age group may be selected as target groups, or people of a certain gender may be selected as target groups.

[0040] After acquiring the first and second quantitative magnetic susceptibility images, radiomic features are extracted from the first and second quantitative magnetic susceptibility images, respectively. For ease of distinction, the radiomic features extracted from the first quantitative magnetic susceptibility image are referred to as first radiomic features, and the radiomic features extracted from the second quantitative magnetic susceptibility image are referred to as second radiomic features.

[0041] Optionally, to improve the accuracy of radiomic feature extraction and reduce the impact of individual anatomical differences on feature extraction results, the first and second quantitative magnetic susceptibility images can be standardized before performing radiomic feature extraction. This involves aligning the first and second quantitative magnetic susceptibility images with standard brain images. Radiomic feature extraction is then performed on the aligned first and second quantitative magnetic susceptibility images. The standard brain image is a healthy brain image with a clear anatomical structure and standard anatomical landmark positioning. The standard brain image includes, but is not limited to, a quantitative magnetic susceptibility image.

[0042] As an optional standardization method, anatomical landmarks can be identified in the first quantitative magnetic susceptibility image, the second quantitative magnetic susceptibility image, and the standard brain image. Anatomical landmarks are anatomical structures that are highly recognizable and consistent across individuals, such as sulci, gyri, and ventricles. Next, a first spatial transformation matrix is ​​determined for aligning the first quantitative magnetic susceptibility image with the standard brain image based on the positional coordinates of the anatomical landmarks identified in the first quantitative magnetic susceptibility image and the positional coordinates of the anatomical landmarks identified in the standard brain image. A second spatial transformation matrix is ​​determined for aligning the second quantitative magnetic susceptibility image with the standard brain image based on the positional coordinates of the anatomical landmarks identified in the second quantitative magnetic susceptibility image and the positional coordinates of the anatomical landmarks identified in the standard brain image. Finally, the first quantitative magnetic susceptibility image is aligned with the standard brain image using the first spatial transformation matrix, and the second quantitative magnetic susceptibility image is aligned with the standard brain image using the second spatial transformation matrix.

[0043] Optionally, anatomical landmarks can be customized based on alignment requirements. For example, when there is a high requirement for spatial consistency between the anatomical structures in the first and second quantitative magnetic susceptibility images and the anatomical structures in the standard brain image, such as when the aligned anatomical structures reach a preset ratio (e.g., 95%), a larger number of anatomical landmarks can be selected for identification, thereby improving the alignment effect by increasing the number of anatomical landmarks. When there is a lower requirement for spatial consistency between the anatomical structures in the first and second quantitative magnetic susceptibility images and the anatomical structures in the standard brain image, and a higher requirement for alignment processing efficiency, a smaller number of anatomical landmarks can be selected for identification, thereby completing the identification and alignment of the anatomical landmarks in a shorter time.

[0044] Optionally, the imaging omics features of the first quantitative magnetic susceptibility image and the second quantitative magnetic susceptibility image can be extracted using open source tools such as Pyrad iomics.

[0045] The extracted imaging omics features include but are not limited to shape features, texture features, and grayscale value statistical features. Through these features, the image features of the first quantitative magnetic susceptibility image and the second quantitative magnetic susceptibility image can be more comprehensively represented from different dimensions.

[0046] The brain is divided into different regions, namely brain areas, based on their structure, function, or neural connection characteristics. Shape features are used to describe the geometric shapes and sizes of different brain areas in the first quantitative magnetic susceptibility image and the second quantitative magnetic susceptibility image, including but not limited to the volume, surface area, maximum diameter, aspect ratio, etc. of each area.

[0047] Texture features are used to describe the magnetic susceptibility distribution of different brain regions in the first and second quantitative magnetic susceptibility images, including but not limited to the peak and mean magnetic susceptibility values ​​for each brain region. The magnetic susceptibility distribution reveals the distribution of iron in the brain. Iron is a key neuromodulator, involved in various physiological processes such as energy metabolism and neurotransmitter synthesis. Higher magnetic susceptibility in a brain region indicates greater iron deposition and a greater likelihood of cognitive impairment.

[0048] Grayscale features are used to describe the tissue density and heterogeneity of different brain regions in the first and second quantitative magnetic susceptibility images, including but not limited to statistical information such as the average grayscale value and grayscale standard deviation of the corresponding regions of each brain region. In the quantitative magnetic susceptibility images, tissue density is used to reflect the cell density or myelination level of each brain region. High tissue density usually indicates healthy neural tissue, while low tissue density usually indicates tissue atrophy or cell loss. Cognitive impairment may be associated with reduced tissue density. Heterogeneity is used to describe the unevenness and diversity of the internal tissue structure of the brain. Heterogeneity can be represented by the iron deposition corresponding to the magnetic susceptibility images. The smaller the difference in iron deposition between different brain regions, the lower the heterogeneity and the greater the possibility of cognitive impairment.

[0049] After extracting a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image, multiple target features for cognitive status classification are further screened from the extracted first radiomics feature based on the first radiomics feature and the second radiomics feature. The multiple target features differ from the second radiomics feature, and correlations between the multiple target features satisfy a set threshold range.

[0050] It should be noted that the feature types corresponding to the first and second radiomic features are the same. The term "selecting multiple target features for cognitive state classification from the first radiomic feature" refers to selecting multiple target feature types for cognitive state classification from the various feature types corresponding to the first radiomic feature. The various feature types corresponding to the second radiomic feature also include multiple target feature types corresponding to multiple target features.

[0051] It is understandable that the first imaging omics features and the second imaging omics features contain a large number of features corresponding to different descriptive dimensions, some of which are related to cognitive state, and some are not related to cognitive state. If all of the first imaging omics features and the second imaging omics features are used for cognitive state classification model training, on the one hand, the large number of features will lead to slow training of the cognitive state classification model. On the other hand, due to the presence of some interference features that are not related to cognitive state classification, the cognitive state classification model will be easily affected by the interference features, and the final classification result will be inaccurate. Therefore, in this embodiment, based on the first imaging omics features and the second imaging omics features, the purpose of screening out multiple target features for cognitive state classification from the extracted first imaging omics features is to: screen out target features related to cognitive state, and reduce the number of target features used for cognitive state classification model training.

[0052] Next, the process of determining multiple target features is described in detail.

[0053] Figure 2 A flowchart of a target feature determination process provided by an embodiment of the present invention is shown in FIG. Figure 2 As shown, it at least includes the following steps:

[0054] 201. Determine, from the first radiomics feature, a plurality of first target features that are different from the second radiomics feature.

[0055] 202. Determine a plurality of second target features from the plurality of first target features based on correlations between the plurality of first target features, wherein the correlations between the plurality of second target features satisfy a set threshold range.

[0056] 203. Use the multiple second target features as multiple target features.

[0057] It is easy to understand that since the first imaging omics feature represents the image features of the quantitative magnetic susceptibility image of the brain of the first population with cognitive abnormalities and the second imaging omics feature represents the image features of the quantitative magnetic susceptibility image of the brain of the second population with normal cognition, when the first imaging omics feature is compared with the second imaging omics feature, the features that differ between the two (i.e., multiple first target features) are more likely to be features related to cognitive status.

[0058] Optionally, statistical methods such as T-test and variance analysis can be used to determine the features in the first radiomics feature that are significantly different from the second radiomics feature, so that the features with significant differences are used as multiple first target features.

[0059] To elaborate, first, the first imaging omics feature and the second imaging omics feature are taken as two different feature groups, and the features contained in each feature group are quantified respectively, and the quantitative indicators such as skewness, kurtosis and variance corresponding to each feature are extracted. Then, for any feature group, the features contained in the feature group are clustered unsupervisedly. Specifically, the quantitative indicators corresponding to each feature in the feature group are clustered by clustering algorithms such as K-means, and the optimal number of clusters corresponding to the feature group is determined by, for example, the elbow rule, so as to divide the features contained in the feature group into multiple cluster groups based on the optimal number of clusters. Afterwards, the Pearson correlation coefficients between the features contained in each cluster group are calculated respectively to evaluate the linear relationship between the features in the same cluster group; based on the calculated Pearson correlation coefficients between the features in the cluster group, the features in the cluster group are screened to screen and retain features whose corresponding Pearson correlation coefficients are less than or equal to the set threshold (such as 0.8), thereby reducing the redundancy between features. Finally, according to the screening processing results corresponding to the multiple cluster groups contained in the feature grouping, the features retained in the feature grouping after the screening processing are determined; based on the retained features corresponding to the two feature groups, statistical methods such as T-test or variance analysis are used to determine the features that are significantly different from the feature grouping corresponding to the first imaging omics feature and the feature grouping corresponding to the second imaging omics feature, as the multiple first target features corresponding to the first imaging omics feature and the second imaging omics feature.

[0060] In the above process, by identifying multiple first target features that differ from the second radiomics features from the first radiomics features, features that do not differ significantly between people with cognitive abnormalities and normal cognition, that is, features that are not significantly correlated with cognitive status, can be effectively filtered out. This effectively avoids overfitting during the subsequent training of the cognitive status classification model. Furthermore, in the process of determining the first target feature, unsupervised clustering is used to generate cluster groups and calculate feature correlations within the cluster groups. This allows for the rapid and effective elimination of features with a Pearson correlation coefficient greater than a set threshold, reducing the number of features subsequently subjected to T-tests or variance analysis, and improving the efficiency of determining the first target feature.

[0061] In practical applications, among the multiple first target features, there may be some features that are correlated with each other, for example, the first target feature A is linearly positively correlated with the first target feature B. When training a cognitive state classification model, if the first target feature A and the first target feature B are used simultaneously, there is a problem of feature redundancy because the first target feature A and the first target feature B have the same impact on the cognitive state classification. Therefore, in this embodiment, after determining the multiple first target features, further, based on the correlation between the multiple first target features, multiple second target features are determined from the multiple first target features to filter out features with a high linear correlation among the multiple first target features, thereby reducing redundancy between features.

[0062] Optionally, a threshold range may be set to determine multiple second target features whose correlations satisfy the set threshold range from multiple first target features, that is, the first target feature whose correlations satisfy the set threshold range is determined as the second target feature.

[0063] Specifically, the features of the plurality of first target features can be arbitrarily combined in pairs, and the correlation between the two first target features contained in each feature combination can be calculated, for example, the Pearson correlation coefficient between the two first target features is calculated, and the correlation between the two first target features is represented by the Pearson coefficient. Afterwards, for any feature combination, if the absolute value of the corresponding Pearson coefficient is less than or equal to a set threshold (for example, 0.8), both first target features in the feature combination are used as second target features; if the absolute value of the corresponding Pearson coefficient is greater than the set threshold (for example, 0.8), one first target feature is removed from both first target features contained in the feature combination, and the remaining first target feature is used as the second target feature. For example, for the two first target features contained in a certain feature combination (respectively denoted as first target feature x and first target feature y), if the Pearson correlation coefficient between the first target feature x and the first target feature y is 0.9 (greater than the set threshold 0.8), the first target feature y is deleted and the first target feature x is retained, that is, the first target feature x is used as the second target feature.

[0064] It is understandable that when the Pearson correlation coefficient between two first target features is greater than a set threshold, it indicates that the two first features are highly collinear. If these two first target features are used simultaneously for model training, multicollinearity may occur during the training of the cognitive state classification model, affecting the stability and interpretability of the model. In this solution, by utilizing the correlation between the first target features and performing feature screening on multiple first target features, the highly collinear first target features are screened out. The second target features determined after screening are then used to train the cognitive state classification model, effectively improving the model's generalization ability and predictive performance.

[0065] The above process is described by first determining multiple first target features from the first imaging omics feature, and then screening multiple second target features from the multiple first target features as an example. Optionally, based on the correlation between each pair of features in the first imaging omics feature, multiple second target features' whose correlations meet a set threshold range can be screened from the first imaging omics feature; then, based on the difference between the multiple second target features' corresponding to the first imaging omics feature and the second imaging omics feature, multiple first target features' that are significantly different from the multiple second target features' of the second imaging omics feature can be screened from the multiple second target features' of the first imaging omics feature, so as to use the multiple first target features' as multiple target features.

[0066] Finally, the multiple second target features determined from the multiple first target features are used as multiple target features for training the cognitive state classification model.

[0067] In this solution, on the one hand, the multiple first target features are distinguishing features between the first imaging omics feature and the second imaging omics feature, which can be used to reflect the difference between the quantitative magnetic susceptibility images of the brains of the first population with cognitive abnormalities and the second population with normal cognition. On the other hand, the multiple second target features are obtained by screening the multiple first target features using the relationship between the correlation between the multiple first target features and the set threshold range. The multiple second target features are the first target features within the set threshold range, that is, the number of target features that can be used to train the cognitive state classification model is further reduced through correlation screening. Therefore, the multiple target features (i.e., the multiple second target features) determined based on this solution are both related to the cognitive state and small in number, thereby improving the training speed and training efficiency of the cognitive state classification model.

[0068] In practical applications, based on Figure 2 While the multiple second target features determined by the illustrated target feature determination process all have an impact on cognitive state classification, the magnitude of their influence varies. Specifically, different second target features contribute differently to cognitive state classification. For example, some second target features are crucial for the classification results, while others have a negligible or negligible impact. To improve the training efficiency and classification efficiency of the cognitive state classification model, the multiple second target features can be further screened to retain only those with the greatest impact. This is explained below.

[0069] Figure 3 A flowchart of another target feature determination process provided by an embodiment of the present invention is shown in FIG. Figure 3 As shown, it at least includes the following steps:

[0070] 301. Determine, from the first radiomics feature, a plurality of first target features that are different from the second radiomics feature.

[0071] 302. Determine a plurality of second target features from the plurality of first target features based on correlations between the plurality of first target features, wherein the correlations between the plurality of second target features satisfy a set threshold range.

[0072] 303. Determine, based on a pre-established feature contribution evaluation model, the feature contributions of the plurality of second target features to determining the cognitive state classification result.

[0073] 304. Determine, from the plurality of second target features, a third target feature whose feature contribution is greater than a set contribution threshold, based on the feature contribution degrees corresponding to the plurality of second target features.

[0074] 305. Use the multiple third target features as multiple target features.

[0075] The specific implementation process of step 301 and step 302 can refer to the above embodiment and will not be described again here.

[0076] As a way to optionally screen out features that have a greater impact on the cognitive state classification results from multiple second target features (referred to as third target features in this embodiment), a feature contribution evaluation model can be pre-established. The feature contribution evaluation model can be used to calculate the contribution of each second target feature involved in the cognitive state classification to the cognitive state classification result. The feature contribution evaluation model includes but is not limited to a random forest model. Then, according to the pre-established feature contribution evaluation model, the feature contributions of the multiple second target features to the determination of the cognitive state classification result are determined. Finally, according to the feature contributions corresponding to the multiple second target features, a third target feature whose feature contribution is greater than the set contribution threshold is determined from the multiple second target features.

[0077] The method of determining the feature contribution of each of the plurality of second target features to the cognitive state classification result according to a pre-established feature contribution evaluation model includes:

[0078] dividing the plurality of second target features into a plurality of second target feature groups, wherein the second target features included in different second target feature groups are not completely the same;

[0079] For any second target feature group, the cognitive state classification model is trained using the second target features contained in the second target feature group, and the feature contribution corresponding to each second target feature in the second target feature group is calculated using the feature contribution evaluation model.

[0080] During the specific implementation process, when grouping multiple second target features, you can first customize the number of second target features contained in each second target feature group, for example: each second target feature group contains 5 second target features or 10 second target features, etc.; then, determine to take out multiple combinations consisting of the customized number of second target features from the multiple second target features as multiple second target feature groups.

[0081] After determining multiple second target feature groups, taking the feature contribution evaluation model as a random forest model as an example, the process of calculating the feature contribution corresponding to each second target feature in a second target feature group is explained.

[0082] In summary, for each second target feature in a second target feature group, the Gini index of each second target feature for each decision tree in the random forest model is calculated to obtain the Gini index of each second target feature for each decision tree in the random forest model.

[0083] Specifically, in this embodiment, the feature contribution of the second target feature is recorded as VIM, and the Gini index value is recorded as GI. Assuming that a second target feature group X contains m second target features X1, X2, ...Xm, the feature contribution corresponding to the Gini index of each second target feature Xi is VIMi, that is, the average change in node split impurity of the i-th second target feature in all decision trees of the random forest model. According to the Gini index calculation formula, in the i-th decision tree, the Gini index of node k is:

[0084] Among them, k represents that there are k categories at feature node i, p mk Indicates the proportion of category k in node m.

[0085] The feature contribution of the second target feature Xi at node m, that is, the change in the Gini index before and after the branch of node m, is:

[0086]

[0087] Among them, GI l and GI r Represent the Gini index of the two new nodes after branching.

[0088] If the node where the feature contribution Xi appears in decision tree j is in the set M, then the feature contribution Xi in the jth tree is:

[0089] Assuming that there are n trees in the random forest, the feature contribution of the second target feature Xi is:

[0090] By analogy, the feature contribution corresponding to each second target feature in each second target feature group can be obtained.

[0091] In an optional embodiment, determining, from the plurality of second target features, a third target feature having a feature contribution greater than a set contribution threshold according to the feature contribution degrees respectively corresponding to the plurality of second target features, includes:

[0092] Calculating a statistical value of the feature contribution of the same second target feature in multiple second target feature groups according to the feature contribution of each second target feature corresponding to each second target feature group;

[0093] Sorting the statistical values ​​of the feature contributions of the plurality of second target features in descending order, and determining the second target feature whose statistical value of the feature contribution is greater than the set feature contribution threshold as the third target feature; or

[0094] The statistical values ​​of the feature contributions of the multiple second target features are sorted in descending order, and the second target features corresponding to the statistical values ​​of the first K feature contributions are determined as the third target features, where K is an integer greater than 1 and can be customized.

[0095] For ease of understanding, for example, assuming there are three second target features, namely feature 1, feature 2, and feature 3, and grouping them according to the fact that each second target feature group contains two second target features, we get second target feature group 1 {feature 1, feature 2}, second target feature group 2 {feature 2, feature 3}, and second target feature group 3 {feature 1, feature 3}. Calculate the feature contribution 1-1 corresponding to feature 1 and the feature contribution 1-2 corresponding to feature 2 in second target feature group 1 {feature 1, feature 2}; the feature contribution 2-2 corresponding to feature 2 and the feature contribution 2-3 corresponding to feature 3 in second target feature group 2 {feature 2, feature 3}; and the feature contribution 3-1 corresponding to feature 1 and the feature contribution 3-3 corresponding to feature 3 in second target feature group 3 {feature 1, feature 3}.

[0096] Based on the calculation of the above feature contribution, feature 1, feature 2, and feature 3 each correspond to two different feature contributions. For feature 1, the sum of feature contribution 1-1 and feature contribution 3-1 is used as the statistical value of the feature contribution of feature 1; for feature 2, the sum of feature contribution 1-2 and feature contribution 2-2 is used as the statistical value of the feature contribution of feature 2; for feature 3, the sum of feature contribution 2-3 and feature contribution 3-3 is used as the statistical value of the feature contribution of feature 3. If the statistical value of the feature contribution of feature 1 > the statistical value of the feature contribution of feature 2 > the statistical value of the feature contribution of feature 3, and the statistical value of the feature contribution of feature 1 and the statistical value of the feature contribution of feature 2 are greater than the set feature contribution threshold, and the statistical value of the feature contribution of feature 3 is less than or equal to the set feature contribution threshold, then feature 1 and feature 2 are determined to be the third target feature. Alternatively, when the value of K is 1, feature 1 is determined to be the third target feature, and when the value of K is 2, feature 1 and feature 2 are determined to be the third target feature.

[0097] In another optional embodiment, according to the feature contributions respectively corresponding to the plurality of second target features, determining a third target feature having a feature contribution greater than a set contribution threshold from the plurality of second target features, further comprising:

[0098] Determine, according to the feature contribution of each second target feature corresponding to each second target feature group and the set feature contribution threshold, a second target feature whose feature contribution is greater than the set feature contribution threshold in each second target feature group;

[0099] Count the number of times the feature contribution corresponding to the same second target feature is greater than the set feature contribution threshold;

[0100] Sort the target times corresponding to the multiple second target features in descending order, and determine the second target features corresponding to the first K target times as the third target feature, where K is an integer greater than 1 and can be customized.

[0101] For ease of understanding, for example, we still assume that there are three second target features, namely feature 1, feature 2, and feature 3. Grouping is performed based on the fact that each second target feature group contains two second target features, thus obtaining second target feature group 1 {feature 1, feature 2}, second target feature group 2 {feature 2, feature 3}, and second target feature group 3 {feature 1, feature 3}. Calculate the feature contribution 1-1 corresponding to feature 1 and the feature contribution 1-2 corresponding to feature 2 in second target feature group 1 {feature 1, feature 2}; the feature contribution 2-2 corresponding to feature 2 and the feature contribution 2-3 corresponding to feature 3 in second target feature group 2 {feature 2, feature 3}; and the feature contribution 3-1 corresponding to feature 1 and the feature contribution 3-3 corresponding to feature 3 in second target feature group 3 {feature 1, feature 3}.

[0102] Assume that in the second target feature group 1 {feature 1, feature 2}, the feature contribution 1-1 corresponding to feature 1 is greater than the set feature contribution threshold, in the second target feature group 2 {feature 2, feature 3}, the feature contribution 2-2 corresponding to feature 2 is greater than the set feature contribution threshold, and in the second target feature group 3 {feature 1, feature 3}, the feature contribution 3-1 corresponding to feature 1 is greater than the set feature contribution threshold. Then, if both the feature contribution 1-1 and the feature contribution 3-1 corresponding to feature 1 are greater than the set feature contribution threshold, the corresponding target number is 2. Only the feature contribution 2-2 of feature 2 is greater than the set feature contribution threshold, the corresponding target number is 1. The feature contribution 2-3 and the feature contribution 3-3 corresponding to feature 3 are both less than or equal to the set feature contribution threshold, and the corresponding target number is 0. The corresponding feature order in descending order of target number is: feature 1, feature 2, and feature 3. When K is 1, feature 1 is determined to be the third target feature. When K is 2, features 1 and 2 are determined to be the third target features.

[0103] Finally, the plurality of third target features determined from the plurality of second target features are used as the plurality of target features for training the cognitive state classification model.

[0104] Optionally, multiple target features corresponding to the first and second radiomic features can be used as target sample features to train the cognitive state classification model. Alternatively, additional quantitative brain susceptibility images can be obtained as training samples, and multiple target features can be extracted from these training samples for use in training the cognitive state classification model. In this embodiment, the cognitive state classification model can be a random forest model, a logistic regression model, a naive Bayes model, or the like.

[0105] Optionally, a backpropagation algorithm can be used to train the cognitive state classification model, and the Adam optimizer can be used to adjust the model parameters to improve the model's convergence speed and efficiency. Furthermore, to quantify the difference between the cognitive state classification model's predictions and the actual cognitive state category labels, a cross-entropy loss function can be used to measure the difference between the model output and the true label, enabling the model to more accurately predict cognitive state categories.

[0106] Figure 4 A flowchart of another cognitive state classification method provided by an embodiment of the present invention is shown in FIG. Figure 4 As shown, it at least includes the following steps:

[0107] 401. Acquire a target quantitative magnetic susceptibility image of the target subject's brain.

[0108] The target object is the object that needs to be classified into cognitive states.

[0109] 402. Extract multiple target features corresponding to the target quantitative magnetic susceptibility image.

[0110] Alternatively, open source tools such as Pyrad iomics can be used to first extract radiomics features from the target quantitative magnetic susceptibility image, and then multiple target features can be screened from the radiomics features; or, multiple target features can be directly extracted from the target quantitative magnetic susceptibility image using a feature extraction algorithm corresponding to multiple target features.

[0111] 403. Input multiple target features corresponding to the target quantitative magnetic susceptibility image into the trained cognitive state classification model to determine the cognitive state classification result of the target object.

[0112] In summary, this protocol first obtains quantitative magnetic susceptibility images of the brains of a target population with end-stage renal disease (ESRD). The quantitative magnetic susceptibility images of the brains of a first group of patients with cognitive abnormalities within the target population serve as first quantitative magnetic susceptibility images, and the quantitative magnetic susceptibility images of the brains of a second group of patients with normal cognition serve as second quantitative magnetic susceptibility images. Next, first radiomic features corresponding to the first quantitative magnetic susceptibility image and second radiomic features corresponding to the second quantitative magnetic susceptibility image are extracted. Then, based on the differences between the first and second radiomic features, the correlations between the features, and the contribution of the features to cognitive status classification, multiple target features are selected from the first influencing omics features, whose feature correlations meet a set threshold, and whose contributions to cognitive status classification are significant. A cognitive status classification model for cognitive status classification is trained using machine learning methods using these multiple target features. This cognitive status classification model can more objectively classify the cognitive status of ESRD patients, unaffected by factors such as the patient's education level, and effectively improves the accuracy of cognitive status classification results.

[0113] The following describes in detail one or more embodiments of the cognitive state classification device of the present invention. Those skilled in the art will appreciate that these devices can be constructed using commercially available hardware components and configured according to the steps taught in this solution.

[0114] Figure 5 A schematic diagram of the structure of a cognitive state classification device provided by an embodiment of the present invention is shown in FIG. Figure 5 As shown, the device includes: an acquisition module 11, an extraction module 12, and a processing module 13.

[0115] The acquisition module 11 is used to acquire a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population, wherein the target population is a group of people with end-stage renal disease.

[0116] The extraction module 12 is configured to extract a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image.

[0117] The processing module 13 is used to determine multiple target features from the first imaging omics feature based on the first imaging omics feature and the second imaging omics feature, where the multiple target features are different from the second imaging omics feature and the correlation between the multiple target features meets a set threshold range; and train a cognitive state classification model based on the multiple target features to perform cognitive state classification using the trained cognitive state classification model.

[0118] Optionally, the acquisition module 11 is further configured to acquire a target quantitative magnetic susceptibility image of the target subject's brain.

[0119] Correspondingly, the extraction module 12 is further configured to extract the multiple target features corresponding to the target quantitative magnetic susceptibility image.

[0120] The processing module 13 is further configured to input the multiple target features corresponding to the target quantitative magnetic susceptibility image into the trained cognitive state classification model to determine a cognitive state classification result of the target object.

[0121] Optionally, the processing module 13 is further specifically used to determine multiple first target features that are different from the second imaging omics features from the first imaging omics features; determine multiple second target features from the multiple first target features based on the correlation between the multiple first target features, and the correlation between the multiple second target features satisfies a set threshold range; and use the multiple second target features as the multiple target features.

[0122] Optionally, the processing module 13 is further specifically used to determine the feature contributions of the multiple second target features to determining the cognitive state classification result based on a pre-established feature contribution evaluation model; determine a third target feature whose feature contribution is greater than a set contribution threshold from the multiple second target features based on the feature contributions corresponding to the multiple second target features; and use the multiple third target features as the multiple target features.

[0123] Optionally, the processing module 13 is further specifically used to treat the first imaging genomics feature and the second imaging genomics feature as two different feature groups; extract quantitative indicators corresponding to the features contained in each feature group, and the quantitative indicators include skewness, kurtosis and variance; perform unsupervised clustering on each feature group according to the quantitative indicators to determine multiple cluster groups corresponding to each feature group; screen the features in each cluster group according to the Pearson correlation coefficient between the features contained in each cluster group in the multiple cluster groups; determine the features retained in each feature group after the screening process according to the screening processing results corresponding to the multiple cluster groups contained in each feature group; and determine the features that are different from the feature group corresponding to the first imaging genomics feature according to the features retained in the two different feature groups, as the multiple first target features.

[0124] Optionally, the processing module 13 is further specifically used to arbitrarily combine the features of the multiple first target features in pairs to obtain multiple feature combinations; calculate the correlation between the two first target features contained in each feature combination; and determine that the first target feature whose correlation meets the set threshold range is the second target feature.

[0125] Optionally, the processing module 13 is further specifically used to divide the multiple second target features into multiple second target feature groups, and the second target features contained in different second target feature groups are not exactly the same; for any second target feature group among the multiple second target feature groups, the cognitive state classification model is trained using the second target features contained in any second target feature group, and the feature contribution corresponding to each second target feature in any second target feature group is calculated through a feature contribution evaluation model.

[0126] Figure 5 The device shown can execute the steps in the aforementioned embodiments. For detailed execution process and technical effects, please refer to the description in the aforementioned embodiments and will not be repeated here.

[0127] In one possible design, the above Figure 5 The structure of the cognitive state classification device shown can be implemented as an electronic device. Figure 6 As shown, the electronic device may include: a processor 21, a memory 22, and a communication interface 23. The memory 22 stores executable code, which, when executed by the processor 21, enables the processor 21 to at least implement the cognitive state classification method provided in the aforementioned embodiment.

[0128] In addition, an embodiment of the present invention provides a non-transitory machine-readable storage medium, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor can at least implement the cognitive state classification method provided in the aforementioned embodiment.

[0129] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Those skilled in the art can understand and implement the present invention without inventive effort.

[0130] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by a combination of hardware and software. Based on this understanding, the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a computer product. The present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A cognitive state classification method, characterized in that: include: Acquiring a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population, wherein the target population is a group of people suffering from end-stage renal disease; extracting a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image; Determining, based on the first radiomics feature and the second radiomics feature, a plurality of target features from the first radiomics feature, wherein the plurality of target features are different from the second radiomics feature, and correlations between the plurality of target features satisfy a set threshold range; Training a cognitive state classification model based on the multiple target features, so as to perform cognitive state classification using the trained cognitive state classification model; Wherein, the determining of multiple target features from the first radiomics feature according to the first radiomics feature and the second radiomics feature includes: grouping the first radiomics feature and the second radiomics feature as two different features; Extracting quantitative indicators corresponding to the features contained in each feature group, wherein the quantitative indicators include skewness, kurtosis and variance; According to the quantitative indicators, unsupervised clustering is performed on each feature group to determine multiple cluster groups corresponding to each feature group; Performing screening processing on the features within each cluster group according to the Pearson correlation coefficient between each of the features included in the plurality of cluster groups; Determining the features retained in each feature group after the screening process based on the screening process results corresponding to the multiple cluster groups contained in each feature group; Determining, from the feature group corresponding to the first radiomics feature, features that are different from the feature group corresponding to the second radiomics feature based on the features retained in the two different feature groups, as the multiple first target features; Determining a plurality of second target features from the plurality of first target features according to correlations between the plurality of first target features, wherein the correlations between the plurality of second target features satisfy a set threshold range; The plurality of second target features are used as the plurality of target features.

2. The method according to claim 1, characterized in that The method further comprises: acquiring a target quantitative magnetic susceptibility image of the target subject's brain; extracting the plurality of target features corresponding to the target quantitative magnetic susceptibility image; The multiple target features corresponding to the target quantitative magnetic susceptibility image are input into the trained cognitive state classification model to determine the cognitive state classification result of the target object.

3. The method according to claim 1, characterized in that The method further comprises: determining, according to a pre-established feature contribution evaluation model, feature contributions of the plurality of second target features to determining a cognitive state classification result; Determining, from the plurality of second target features, according to the feature contribution degrees respectively corresponding to the plurality of second target features, a third target feature whose feature contribution degree is greater than a set contribution degree threshold; The plurality of third target features are used as the plurality of target features.

4. The method according to claim 2, characterized in that The determining of a plurality of second target features from the plurality of first target features according to correlations between the plurality of first target features comprises: Arbitrarily combining any two of the features in the plurality of first target features to obtain a plurality of feature combinations; Calculating the correlation between the two first target features included in each feature combination; The first target feature whose correlation satisfies a set threshold range is determined as the second target feature.

5. The method according to claim 3, characterized in that Determining the feature contribution of each of the plurality of second target features to determining the cognitive state classification result according to a pre-established feature contribution evaluation model includes: dividing the plurality of second target features into a plurality of second target feature groups, wherein the second target features included in different second target feature groups are not completely the same; For any second target feature group among the multiple second target feature groupings, the cognitive state classification model is trained using the second target features contained in any second target feature group, and the feature contribution corresponding to each second target feature in any second target feature group is calculated using a feature contribution evaluation model.

6. A cognitive state classification device, characterized in that: include: an acquisition module, configured to acquire a first quantitative magnetic susceptibility image of the brain of a first group of people with cognitive abnormalities and a second quantitative magnetic susceptibility image of the brain of a second group of people with normal cognition in a target population, wherein the target population is a group of people suffering from end-stage renal disease; an extraction module, configured to extract a first radiomics feature corresponding to the first quantitative magnetic susceptibility image and a second radiomics feature corresponding to the second quantitative magnetic susceptibility image; a processing module, configured to determine, based on the first radiomics feature and the second radiomics feature, a plurality of target features from the first radiomics feature, wherein the plurality of target features differ from the second radiomics feature and correlations between the plurality of target features satisfy a set threshold range; and train a cognitive state classification model based on the plurality of target features to perform cognitive state classification using the trained cognitive state classification model; The processing module is further configured to group the first imaging omics feature and the second imaging omics feature as two different feature groups; extract quantitative indicators corresponding to the features contained in each feature group, wherein the quantitative indicators include skewness, kurtosis, and variance; perform unsupervised clustering on each feature group based on the quantitative indicators to determine multiple cluster groups corresponding to each feature group; perform screening processing on the features within each cluster group based on the Pearson correlation coefficient between the features contained in each cluster group in the multiple cluster groups; determine the features retained in each feature group after the screening processing based on the screening processing results corresponding to the multiple cluster groups contained in each feature group; and determine, from the feature group corresponding to the first imaging omics feature and the feature group corresponding to the second imaging omics feature, features that are different from each other based on the features retained in the two different feature groups, as the multiple first target features; According to the correlation between the multiple first target features, multiple second target features are determined from the multiple first target features, and the correlation between the multiple second target features meets the set threshold range; the multiple second target features are used as the multiple target features.

7. An electronic device, characterized in that: include: A memory, a processor, and a communication interface; wherein the memory stores executable code, and when the executable code is executed by the processor, the processor executes the cognitive state classification method according to any one of claims 1 to 5.

8. A non-transitory machine-readable storage medium, characterized in that The non-transitory machine-readable storage medium stores executable code, and when the executable code is executed by a processor of an electronic device, the processor is caused to perform the cognitive state classification method according to any one of claims 1 to 5.

Images