Random grouping method for large-scale fMRI data coupling tensor decomposition

Abstract

The invention discloses a random grouping method for large-scale fMRI data coupling tensor decomposition, and belongs to the field of biomedical signal processing. The method comprises the following steps: firstly, randomly scrambling the sequence of all subjects, then carrying out equal-size sub-tensor division, and carrying out first-round decomposition by adopting a space coupling Tucker decomposition method; next, circularly moving left for several subjects, performing equal-size sub-tensor division again to form a new subject group, and performing a second round of space coupling Tucker decomposition; and analogy is carried out in sequence until displacement, sub-tensor division and space coupling Tucker decomposition of a predetermined round are completed, and decomposition results of group common and individual specific characteristics are obtained. According to the method, the adverse effect of the initial arrangement sequence of the large-scale subjects on feature extraction is weakened, the overall performance of the coupling tensor decomposition method is remarkably improved, especially the noise of networks such as central vision is reduced, and better feature support can be provided for brain function research and objective marker extraction of neuropsychiatric brain diseases.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing and relates to a random grouping method for coupled tensor decomposition of large-scale functional magnetic resonance imaging (fMRI) data. Background Technology

[0002] fMRI is a function of magnetic resonance imaging equipment, enabling advanced functional brain imaging with significant advantages such as safety, non-invasiveness, and high spatial resolution. Currently, multi-subject fMRI data is widely used in research on brain function and neuropsychiatric brain disorders (such as schizophrenia and depression).

[0003] Blind source separation is a method that extracts brain functional activation features from fMRI data without prior information, providing detailed supporting evidence for research on brain function and brain diseases. Specifically, the tensor decomposition method in blind source separation can utilize three-dimensional (spatial, temporal, and subject-specific) structural data from multi-subject fMRI without compression, and the Tucker decomposition method of tensors can simultaneously extract both group-shared and individual-specific spatial and temporal features, effectively eliminating feature confusion between the group and individuals, thus providing better feature support for research on brain function and brain diseases.

[0004] As research progresses and the volume of fMRI data expands, the number of participants has increased from dozens to hundreds, thousands, or even tens of thousands. This presents significant challenges to the Tucker decomposition method in terms of computational requirements and model matching. One existing solution is the coupled tensor decomposition method cCTKD (Han Yue. Research on Tucker Decomposition Method for Feature Extraction of Multi-Subject fMRI Brain Functional Networks, Doctoral Dissertation, Dalian University of Technology, 2024). Its approach is as follows: First, the large-scale fMRI data is divided into multiple small subtensors along the subject dimension, each of equal size (e.g., 10 participants per group). Then, the Tucker decomposition method is used to decompose each subtensor sequentially, and all subtensors are coupled in the spatial dimension. This involves using the shared spatial components of the current subtensor as the initialization matrix for the same components of the next subtensor. After several rounds of updates to all subtensors, the common features of the large-scale participants and the individual-specific features are finally converged.

[0005] In the existing Coupled Tensor Decomposition (ccTKD) method, sub-tensor partitioning is based on the initial arrangement of a large number of subjects, and the subjects in the sub-tensor groups remain unchanged during the update process. This sub-tensor grouping method is correlated with the initial arrangement of a large number of subjects, which can easily have a special impact on the feature extraction quality, thus limiting the overall performance of the Coupled Tensor Decomposition method. Summary of the Invention

[0006] This invention provides a random grouping method for coupled tensor decomposition of large-scale fMRI data, which weakens the influence of subject order in sub-tensor grouping and improves the performance of coupled tensor decomposition.

[0007] This invention randomly shuffles the order of all participants, then performs equal-sized subtensor partitioning, and uses the spatially coupled Tucker decomposition method for the first round of decomposition. Next, iteratively shifts several participants to the left, performs equal-sized subtensor partitioning again to form new participant groups, and performs the second round of spatially coupled Tucker decomposition. This process is repeated until the predetermined number of shifts, subtensor partitioning, and spatially coupled Tucker decompositions are completed, outputting the decomposition results of group and individual characteristics.

[0008] The technical solution adopted in this invention is as follows:

[0009] A random grouping method for large-scale fMRI data coupled with tensor decomposition includes the following steps:

[0010] Step 1: Randomize the order of fMRI participants

[0011] Assuming there is a total There were [number] participants. All participants' fMRI data underwent preprocessing including head movement correction, spatial normalization, and spatial smoothing. The total number of prime numbers and time points in the brain for each participant were [number]. and ; The initial order of all subjects {#1, …, # The subjects are uniformly shuffled to obtain a randomly arranged order {r(1), …, r(}. )};

[0012] Step 2: Round 1 Coupling Tucker Decomposition

[0013] Based on the above randomly shuffled subject order {r(1), …, r( Constructing a three-dimensional tensor for large fMRI data Perform a subtensor partition of size K to form I equal-sized subtensors. The first round of decomposition is performed using the spatially coupled Tucker decomposition method;

[0014] Step 3: Second round of coupling Tucker decomposition

[0015] The order of the subjects {r(1), …, r( )} Shift the subjects left by d numbers to form a new subject order {r(1+d), r(2+d), …, r( ), r(1), …, r(d)}, construct a three-dimensional tensor for large fMRI data The subjects are divided into I subtensors of size K to form new subject groups. Then, perform the second round of spatial coupling Tucker decomposition;

[0016] Step 4: Repeat step 3 until round R.

[0017] Repeat step three, that is, continue with the new round of cyclic left shifting of d subjects, subtensor partitioning, and spatial coupling Tucker decomposition based on the previous round, until the Rth round, R=K / d, and the subject order is: {r(1+(R-1)d), r(2+(R-1)d), …, r , r(1), …, r((R-1)d)};

[0018] Step 5: Output Results

[0019] Output the common and individual characteristics of the spatially coupled Tucker decomposition in the Rth round;

[0020] Step 6: Performance Evaluation

[0021] The absolute value of the correlation coefficient between the Smith template and the spatial reference component (denoted as ) is calculated and the Pearson correlation coefficient is determined. ), and the total number of voxels falling within the spatial reference component (denoted as ). The performance of spatial features of spatially coupled Tucker decomposition was evaluated and compared with the results of initial permutation order using a large number of subjects.

[0022] The Smith template mentioned refers to the template in the paper Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro R, Laird AR, Beckmann CF. Correspondence of the brain's functional architecture during activation and rest, Proceedings of the National Academy of Sciences, vol.106, no.31, pp.13040-13045, 2009.

[0023] This invention achieves random grouping of subjects within a subtensor by randomly scrambling the initial arrangement of large-scale fMRI data and combining this with cyclic left shifting. This mitigates the adverse effects of the initial arrangement order on feature extraction and significantly improves the overall performance of the coupled tensor decomposition method. Figure 3 Taking the extraction of shared spatial components from the three groups shown as an example, including the default mode network (DMN), auditory (AUD) network, and medial visual (MV) network, the method of this invention outperforms cctKD in terms of results. It can increase by 5.4%, in It can increase by 8.9% (see) Figure 4 ), especially MV network noise is significantly reduced (see Figure 3 As can be seen, the method of this invention can improve the performance of the coupled tensor decomposition method, and provide better feature support for brain function research and the extraction of objective biomarkers for neuropsychiatric brain diseases. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the present invention;

[0025] Figure 2 A schematic diagram of subtensor partitioning and coupled Tucker decomposition;

[0026] Figure 3 The three common components extracted from the fMRI data of 100 subjects by cCTKD and this invention are DMN, AUD, and MV from left to right. run1 to run3 are the results of three randomizations in this invention.

[0027] Figure 4 The average correlation coefficients of the common components of the three populations extracted by cCTKD and this invention are... Comparison (3 runs). Detailed Implementation

[0028] An embodiment of the present invention will now be described in detail with reference to the technical solution.

[0029] Example 1: Given fMRI data of 100 subjects after head movement correction, spatial standardization, and spatial smoothing preprocessing, Total number of primes in the brain =62336, number of time points .

[0030] Step 1: Randomize the order of fMRI participants

[0031] The initial arrangement of all subjects is {#1, …, #}. The subjects were uniformly randomized to obtain a randomly arranged order of subjects {r(1), …, r(100)};

[0032] Step 2: Round 1 Coupling Tucker Decomposition

[0033] Based on the randomly scrambled subject order {r(1), …, r(100)} mentioned above, a three-dimensional tensor of large fMRI data is constructed. Perform subtensor partitioning of size K=10, forming I=10 equal-sized subtensors. The first round of decomposition was performed using the spatially coupled Tucker decomposition method, see [link to documentation]. Figure 2 ;

[0034] Step 3: Second round of coupling Tucker decomposition

[0035] The subject order {r(1), …, r(100)} is cyclically shifted left by d=2 subjects to form a new subject order {r(3), r(4), …, r(100), r(1), r(2)}, and a three-dimensional tensor of the large fMRI data is constructed. The subjects were divided into 10 subtensors of size 10 to form new subject groups, resulting in 10 subtensors. Perform the second round of spatially coupled Tucker decomposition, see... Figure 2 ;

[0036] Step 4: Repeat step 3 until the 5th round.

[0037] Repeat step 3, that is, continue to perform a new round of 2-subject cyclic left shift, subtensor partitioning and spatial coupling Tucker decomposition based on the previous round, until the Rth round, R=K / d=10 / 2=5, the subject order is {r(9), r(10), …, r(100), r(1), …, r(8)};

[0038] Step 5: Output Results

[0039] Output the shared spatial features DMN, AUD, and MV of the three populations in the fifth round of spatially coupled Tucker decomposition;

[0040] Step 6: Repeat steps 1 through 5 twice to obtain the results of run1 through run3 in total;

[0041] Step 7: Performance Evaluation

[0042] Calculate the absolute value of the Pearson correlation coefficient with the spatial reference component (Smith template). and the total number of voxels falling within the spatial reference component. The results were compared with those of cCTKD, which used an initial ranking of a large number of participants. Figure 3 and Figure 4 As shown. The performance extracted by this invention is superior to cctKD, especially in terms of significantly reduced noise in the MV network, see... Figure 3 .

Claims

1. A random grouping method for large scale fMRI data coupled tensor decomposition, characterized in that The following steps: Step 1: fMRI subject order randomization Assume there are subjects in total, all of which have been pre-processed fMRI data, and the total number of voxels and time points in each subject's brain are and ; the initial arrangement order of all subjects {#1, …, # } is uniformly randomly shuffled to obtain the randomly arranged subject order {r(1), …, r( )}. Step 2: 1st round coupled Tucker decomposition Based on the random permutation of the subject order {r(1), …, r( )}, construct the three-dimensional tensor of large fMRI data , perform sub-tensor division of size K , form I sub-tensors of equal size , and perform first-round decomposition using the spatial coupling Tucker decomposition method; Step 3: 2nd round coupled Tucker decomposition The order of the subjects is {r(1), …, r( )} Shift the subjects left by d numbers to form a new subject order {r(1+d), r(2+d), …, r( ), r(1), …, r(d)}, construct a three-dimensional tensor for large fMRI data , for a size of K The subtensor partitioning forms new subject groups, constituting I Size tensor Then, perform the second round of spatial coupling Tucker decomposition; Step 4: Repeat Step 3 until Rth round The third step is repeated, i.e. a new round of d-trial cyclic left shift, sub-tensor partitioning and spatially coupled Tucker decomposition is continued on the basis of the last round, until the Rth round, R = K / d, the trial order is: {r(1+(R-1)d), r(2+(R-1)d),…, r , r(1), …, r((R-1)d)}; Step 5: Output results Output group common and individual-specific features from Rth round spatially coupled Tucker decomposition Step 6: Performance evaluation Computing the absolute value of the Pearson correlation coefficient of the spatially referenced component Smith's template and the total number of voxels falling within the spatially referenced component The performance of the spatially coupled Tucker decomposition is evaluated on the spatial features and compared to the results of the initial permutation order using a large scale set of subjects.

2. The random grouping method for large-scale fMRI data coupled tensor decomposition according to claim 1, wherein, In Step 6, the Smith template refers to the template in Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro R, Laird AR, Beckmann CF. Correspondence of the brain’s functional architecture during activation and rest, Proceedings of the National Academy of Sciences, vol. 106, no. 31, pp. 13040-13045, 2009.

Images