Method for suppressing noise in brain function image based on spatiotemporal features

Abstract

The present application belongs to the technical field of medical image processing, and proposes a brain function image noise suppression method based on space-time features. The present application provides a brain function image noise suppression method and system based on space-time features. Compared with the traditional noise suppression method, the present application aims to independently implement targeted noise elimination and signal enhancement in the time domain and the spatial domain. Compared with the existing method which only focuses on global changes and overall trends, the present application adds a sliding window to process the segmented time-varying features, and uses methods such as independent covariate regression to effectively identify and remove the noise of the time and space features, thereby achieving better noise suppression effect.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for suppressing noise in brain functional images based on spatiotemporal features. Background Technology

[0002] Functional connectivity (FC) is one of the most important indicators of neural activity or brain function. For resting-state functional magnetic resonance imaging (rs-fMRI) data, FC is typically defined as the temporal coupling of blood oxygen level-dependent (BOLD) signals between different brain regions. Therefore, using resting-state functional MRI (rs-fMRI) to map functional networks and analyze functional connectivity has become a primary technical approach. Analyzing brain functional connectivity from a network theory perspective has been demonstrated by studies such as "Smith, SM, Vidaurre, D., Beckmann, CF, Glasser, MF, Jenkinson, M., Miller, KL, Nichols, TE, Robinson, EC, Salimi-Khorshidi, G., Woolrich, MW, Barch, DM, Uurbil, K., Essen, DCV, 2013. Functionalconnectomics from resting-state fMRI. Trends Cogn. Sci. 17, 666–682." This research has shown great potential to identify brain tissue characteristics and their connections to cognitive mechanisms and brain diseases. However, artifacts and noise sources (head movements, non-neuronal physiological signals, etc.) can disrupt the normal temporal fluctuations of neural activity and cause abrupt changes in signals, such as sudden spikes or anomalous trends. Therefore, this noise contamination can obscure the true functional connectivity (FC) when using seed point- or region-based analyses to calculate correlations, adversely affecting the analysis and interpretation of subsequent results.

[0003] To minimize the interference and impact of noise, basic denoising measures can be used to perform data cleansing. Previous studies, “Ciric R, Wolf DH, Power JD, et al. Benchmarking of Participant-Level Confound Regression Strategies for the Control of MotionArtifact in Studies of Functional Connectivity[J]. NeuroImage, 2017, 154:174–187.” and “Parkes L, Fulcher B, Yücel M, et al. An Evaluation of the Efficacy, Reliability, and Sensitivity of Motion Correction Strategies for Resting-State Functional MRI[J / OL]. NeuroImage, 2018, 171: 415–436.”, have compared the denoising performance of many typical denoising strategies. Various strategies generally achieve denoising effects, but the results vary considerably. Among them, the irrelevant covariate regression model is the main tool for suppressing head movement noise pollution by regressing different combinations of head motion parameters (HMP), white matter and cerebrospinal fluid signals (WCS), and global signals (GS).Some studies, such as "Hallquist MN, Hwang K, Luna B. The Nuisance of NuisanceRegression: Spectral Misspecification in a Common Approach to Resting-State FMRI Preprocessing Reintroduces Noise and Obscures Functional Connectivity[J / OL]. NeuroImage, 2013, 82: 208–225." and "Lindquist MA, Geuter S, Wager TD, et al. Modular Preprocessing Pipelines Can Reintroduce Artifacts into FMRI Data[J]. Human Brain Mapping, 2019, 40(8): 2358–2376," have raised concerns that some noise suppression strategies may even reintroduce artificial artifacts during computation. There are also some effective data-driven methods based on Independent Component Analysis (ICA) or Principal Component Analysis (PCA) that separate and discard noise components from the BOLD signal. ICA-AROMA (Pruim et al., 2015) / ICA-FIX (Salimi-Khorshidi et al., 2014) decompose rs-fMRI data and identify motion noise through cross-validation or pre-trained feature classification. However, the accuracy of noise identification and the loss of temporal degrees of freedom for many noise components remain significant problems in applications (Pruim et al., 2015; Salimi-Khorshidi et al., 2014). Other techniques, such as CompCorr (Behzadi et al., 2007), can be used to remove signals from tissues with high variance, while censoring (Power et al., 2012) can directly reject noise. However, they have been found to perform poorly when dealing with highly contaminated samples (Parkes et al., 2018b). In conclusion, the study “Caballero-Gaudes C, Reynolds R C. Methods for Cleaning the BOLD FMRI Signal[J / OL]. NeuroImage,2017, 154: 128–149.” demonstrates that even with these issues, data preprocessing and precise noise reduction methods are necessary steps to ensure reliable results from rs-fMRI data.

[0004] The fragility of BOLD time series makes them too susceptible to noise, which then severely impacts functional connectivity. Studies have emerged that are heterogeneous in both spatial location and temporal period. “Power JD, Lynch CJ, Silver BM, et al. Distinctions among Real and Apparent Respiratory Motions in Human FMRI Data[J / OL]. NeuroImage, 2019, 201(July): 116041.” and “Ciric R, Rosen AFG, Erus G, et al. Mitigating Head Motion Artifact in Functional Connectivity MRI[J / OL]. Nature Protocols, 2018, 13(12): 2801–2826.” introduce and summarize three main effects of noise on functional connectivity: local diffusion, whole-brain diffusion, and whole-brain disruption. This suggests that the mixed effects of spatial types can be complex, and some noise reduction strategies may only target specific types of noise. Furthermore, noise and artifacts are temporally unstable due to dramatic intensity variations and mixed-type effects. Therefore, spatial domain-focused decomposition methods are less efficient at accurately separating mixed noise (Pruim RHR, Mennes M, Van Rooij D, et al. ICA-AROMA: A Robust ICA-Based Strategy for Removing Motion Artifacts from FMRI Data[J / OL]. NeuroImage, 2015, 112: 267–277.), while temporal domain-focused methods are better at decoding stable global signals only (Glasser MF, Coalson TS, Bijsterbosch JD, et al. Using Temporal ICA to Selectively Remove Global Noise While Preserving Global Signal in Functional MRI Data[J / OL]. NeuroImage, 2018, 181(3): 692–717.). Thus, the complex mixing of temporal and spatial noise makes traditional denoising effective for specific types of noise, but unable to handle multi-type mixed noise, temporally unstable noise, and drastic intensity changes. Summary of the Invention

[0005] To address current problems, this invention provides a robust and effective automatic feature space denoising method for brain functional imaging based on spatiotemporal features. First, a sliding window is used to accurately locate segmented noise while preserving the signal. Then, singular value decomposition (SVD) is used to transfer the data to a feature space containing temporal and spatial components. In the temporal domain, basic noise sources (HMP and WCS) and low-frequency drift noise are removed, while in the spatial domain, signal components are enhanced and modulated. This achieves accurate noise identification and efficient noise removal, improving the accuracy and reliability of functional networks (FC) and functional networks.

[0006] The technical solution of this invention:

[0007] A brain functional imaging noise suppression method based on spatiotemporal features, the specific steps of which are as follows:

[0008] (1) Sliding window and matrixing;

[0009] For the input raw 4D brain function image data, the data is first segmented into different time segments using the sliding window method; any time segment will be processed through a binary mask containing only voxel information inside the brain to extract effective spatial information, transforming the 4D data into a two-dimensional time x space matrix.

[0010] (2) Construct feature vectors representing time and space;

[0011] For each time segment, the singular value decomposition method is used to decompose the 2D time-space matrix into time coefficient features and spatial distribution features: ;in It is a spatial feature vector, where P is the number of voxels and K is the window length. These are the corresponding time coefficients, and S is the sorted eigenvalue diagonal matrix;

[0012] (3) Denoising of temporal and spatial features;

[0013] For any given time segment, the data is considered to be a linear combination of spatiotemporal feature vectors: Noise removal in the feature space includes processing temporal features separately. and spatial features Implement time-domain noise suppression and spatial-domain signal modulation;

[0014] Spatial signal modulation consists of three steps:

[0015] First, the signal is purified by excluding clusters with excessively small effective volumes: the spatial feature vectors are first transformed into a three-dimensional spatial distribution, and the spatial distribution is adjusted to a Gaussian distribution by applying Z-score. Then, the spatial map is thresholded based on statistical properties to generate a binary map. All effective clusters under the connectivity criteria are calculated based on the Label-connected components method. Clusters with too few effective voxels are removed.

[0016] Secondly, redundant feature vectors are discarded based on the graph theory connection matrix: For each input data, all time segments have a certain proportion of overlap; therefore, the spatial distribution of different segments has strong commonalities; so by calculating the correlation of spatial feature vectors of different time segments, a graph theory connection matrix can be formed to represent the spatial feature similarity of all time segments; then, based on the classification method, the threshold is automatically estimated to filter out non-common spatial distributions with low correlation in the connection graph, so as to discard components that cannot exist stably in different time segments;

[0017] Finally, soft smoothing is applied to compensate for the effective information content of the spatial distribution: most of the original data feature vectors are in a low signal-to-noise ratio state due to thermal noise. A Gaussian smoothing kernel is applied to the cleaned signal clusters to enhance the local intensity of the signal-related voxels.

[0018] Temporal noise suppression consists of two steps:

[0019] First, irrelevant noise covariates are estimated. Realignment is used to perform preliminary motion correction and estimate head motion parameters, namely translation and rotation in the X, Y, and Z directions. Based on an autoregressive model, 24 covariates of head motion are obtained: 6 basic motion parameters, 6 time-domain derivatives, quadratic terms of the 6 basic motion parameters, and time-domain derivatives of the 6 quadratic terms. Then, the DARTEL algorithm is used to segment probability maps of gray matter, white matter, and cerebrospinal fluid. Thresholds are set, and a binary mask for white matter and cerebrospinal fluid is created. The n strongest tissue signals are calculated using aCompCorr as covariates. Drift signals with spectral energy less than 0.01 Hz are added, ultimately estimating a total of 25+n irrelevant noise covariates.

[0020] Then, noise energy in each time coefficient is eliminated based on linear regression variables; for each time segment, the eigenvector of any time-domain coefficient with a Pearson correlation coefficient exceeding a certain value is regressed to remove the associated noise covariate signal, in order to deeply remove noise;

[0021] (4) Data reconstruction and weighted average;

[0022] After denoising, the processed spatiotemporal feature vectors are reconstructed into the original 4D data based on linear combination. Due to the introduction of the sliding window, some time points will be reused multiple times in the calculation, and the frequency of reuse will be used as a weight to calculate the weighted average of the data at specific time points. Finally, the denoised 4D functional MRI image data is output for subsequent analysis.

[0023] The beneficial effects of this invention are as follows: This invention provides a method and system for suppressing noise in brain functional imaging based on spatiotemporal features. Compared with traditional noise suppression methods, this invention aims to independently and specifically eliminate noise and enhance signals for features in the time and spatial domains. Moreover, compared with existing methods that only focus on global changes and overall trends, this invention incorporates a sliding window to process segmented time-varying features and uses methods such as irrelevant covariate regression to effectively identify and remove noise from time and spatial features, thereby achieving better noise suppression results. Attached Figure Description

[0024] Figure 1 This is a flowchart of a brain functional imaging noise suppression method and system based on spatiotemporal features, according to an embodiment of the present invention.

[0025] Figure 2 The following are the typical noise indication sequences and their effects in three examples of embodiments of the present invention: (a) head movement noise; (b) breathing noise; and (c) cerebrospinal fluid signal.

[0026] Figure 3 In this embodiment of the invention, the noise recognition and removal effect diagrams of temporal and spatial features are shown. a) is the proportion of head-on motion noise and tissue signal recognition in the temporal features (right side of the dashed line); b) is the distribution change of spatial features after modulation enhancement; c) is the distribution change of temporal feature quantization index before and after noise suppression.

[0027] Figure 4 This is a schematic diagram of the algorithm framework based on spatiotemporal noise suppression in an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] A method and system for suppressing noise in brain functional images based on spatiotemporal features, characterized in that the method and system for suppressing noise in brain functional images includes two independent denoising functions in the temporal domain and the spatial domain.

[0030] The system first segments the input subject data and transforms it into a feature space. Then, it performs temporal noise suppression and spatial signal modulation on the temporal and spatial features respectively. Finally, it re-integrates the temporal and spatial features to reconstruct the denoised data. The specific steps are as follows:

[0031] (1) Sliding window and matrixing;

[0032] For the input raw 4D (3D spatial x temporal) brain functional imaging data, certain preprocessing is required to meet the needs of the streamlined analysis. First, a sliding window method is used to segment the data into different time segments. The minimum time point (window length) contained in any segment is 50, and the window overlap rate (the number of reused time points) is set to 0.4. This sliding window parameter can effectively preserve the stability of dynamic brain functional networks in functional time-varying connectivity analysis. Furthermore, it ensures a proper balance between artifact capture accuracy and system computational efficiency in the subsequent noise removal stage. Each time segment will be processed through a binary mask containing only voxel information from within the brain to extract effective spatial information, transforming the 4D data into a 2D temporal x spatial matrix.

[0033] (2) Construct feature vectors representing time and space;

[0034] For each time segment, singular value decomposition is used to decompose the 2-dimensional time-space matrix into temporal coefficient features and spatial distribution features. . It is a spatial feature vector, where P is the number of voxels and K is the window length. These are the corresponding time coefficients, and S is the sorted eigenvalue diagonal matrix;

[0035] (3) Denoising of temporal and spatial features;

[0036] For any time segment of data, it can be viewed as a linear combination of spatiotemporal feature vectors, such as... In short, noise removal in the feature space emphasizes the separate processing of temporal features. and spatial features Implement time-domain noise suppression and spatial-domain signal modulation;

[0037] Spatial signal modulation mainly consists of three steps:

[0038] First, the signal is purified by excluding clusters with excessively small effective volumes. The spatial feature vectors are first transformed into a three-dimensional spatial distribution, and the spatial distribution is adjusted to a Gaussian distribution using Z-score. Then, based on statistical properties, the spatial map is thresholded with a p-value < 0.05 to generate a binary map. The Label-connected components method is used to calculate all effective clusters with a connection criterion of 18. Clusters with fewer than 50 effective voxels are removed.

[0039] Secondly, redundant feature vectors are discarded based on graph theory connection matrices. For each input subject, all time periods have a certain degree of overlap. Therefore, the spatial distributions of different segments have strong commonalities (especially feature vectors related to the signal). Based on this assumption, by calculating the correlation of spatial feature vectors in different time periods, a graph theory connection matrix can be formed to represent the spatial feature similarity of all time segments. Then, based on classification methods such as K-nearest neighbors, a threshold is automatically estimated to filter out non-common spatial distributions with low correlation in the connection graph, so as to discard components that cannot exist stably in different segments.

[0040] Finally, soft smoothing is applied to compensate for the effective information content of the spatial distribution. Most of the original data feature vectors are in a low signal-to-noise ratio state due to thermal noise. Therefore, applying a 3-voxel Gaussian smoothing kernel to the surviving signal clusters can effectively enhance the local intensity of the signal-related voxels.

[0041] Temporal noise suppression mainly consists of two steps:

[0042] First, irrelevant noise covariates are estimated. Realignment is used to perform preliminary motion correction and estimate head motion parameters (translation and rotation in the X, Y, and Z directions). Based on an autoregressive model, 24 covariates for head motion are obtained (6 basic motion parameters, 6 time-domain derivatives, quadratic terms of the 6 basic motion parameters, and time-domain derivatives of the 6 quadratic terms). Then, the DARTEL algorithm is used to segment gray matter, white matter, and cerebrospinal fluid tissue probability maps. To maintain the accuracy of noise estimation, a binary mask for white matter and cerebrospinal fluid is created with a threshold of 0.99. The five strongest tissue signals are calculated using aCompCorr as covariates. Drift signals with spectral energy less than 0.01 Hz are added, resulting in a total of 30 irrelevant noise covariates estimated.

[0043] Then, in order to eliminate the noise components mixed in the time coefficient features, the noise energy in each time coefficient is eliminated based on the linear regression variables; for each time period, any time-domain coefficient feature vector with a Pearson correlation coefficient greater than 0.05 is regressed to remove the correlated noise covariate signal, so as to deeply remove the noise.

[0044] (4) Data reconstruction and weighted average;

[0045] After denoising, the processed spatiotemporal feature vectors are reconstructed into the original 4D data (3D space x time) based on a linear combination. Due to the introduction of a sliding window, some time points are reused multiple times in the calculation, and the frequency of reuse is used as a weight to calculate a weighted average of the data at specific time points. Finally, the denoised 4D functional MRI image data can be output for subsequent analysis.

[0046] Figure 1 This is a flowchart of the brain functional imaging noise suppression method and system based on spatiotemporal features according to the present invention. Figure 1 As shown, embodiments of the present invention provide a method and system for suppressing noise in brain functional images based on spatiotemporal features. The main feature is in-depth noise removal based on the spatiotemporal features of rs-fMRI data.

[0047] In order to ensure that the data can be processed stably and efficiently in a streamlined process, the input data is first processed in segments by a sliding time window and decomposed into temporal and spatial features; step 101, noise identification and suppression are performed based on the temporal feature vector; step 102, distributed modulation and signal enhancement are performed based on the spatial feature vector; finally, the denoised output is obtained by reconstructing each segment of data and weighting the data based on the overlap coefficient.

[0048] In this embodiment of the invention, steps 101 and 102 are performed independently; in step 101, various noise sequences that need to be removed are first obtained, and the key noise types can be calculated from the original data using estimation algorithms.

[0049] Figure 2 Images (a)-(c) show the indicator sequences for head movement noise, respiratory noise, and cerebrospinal fluid signals, respectively, along with the distortions caused by noise to the BOLD signal sequence and the changes in the correlation strength of different brain regions in the FC. These three examples are calculated based on typical noise-contaminated data from the ABIDEI and HCP public databases. In the first row, the noise indicator sequences are divided into those using head movement frame displacement indices, respiratory band detection signals, and cerebrospinal fluid tissue to explain the maximum variance noise; the second row shows the BOLD signals for the corresponding time intervals of the noise indicator sequences, with the changes in signal affected by noise clearly visible at the corresponding time points; the third row shows the functional connectivity presented by the time series of BOLD signals contaminated by noise. By calculating the Pearson correlation coefficient between the time feature vector and different noise estimates, the corresponding coefficient value distribution plot can be obtained.

[0050] Figure 3Figure 'a' primarily displays the squared distribution of the noise mixing coefficients related to head movement noise and tissue signals. This mainly helps in analyzing and distinguishing the proportional relationship between noise and non-noise features. For relevant time features identified as having noise mixing effects, linear regression methods will be used to eliminate the noise influence contained within them, thus achieving data purification. Figure 3 Figure 'c' shows the improvement in four time-quantization performance indicators of the temporal characteristics before and after noise elimination, illustrating the effectiveness and accuracy of temporal noise suppression. Step 102 mainly modulates and enhances the spatial distribution pattern of spatial features. This improvement is achieved primarily through three aspects: eliminating invalid clusters with insufficient voxel counts based on binarized 3D spatial clusters, eliminating signal-irrelevant feature vectors based on graph theory connectivity, and enhancing the intensity of locally effective clusters based on Gaussian smoothing. Figure 3 The b in the paper demonstrates the improvement of three quantization characteristics of spatial features before and after modulation processing: kurtosis, entropy, and kurtosis. This effectively suppresses extreme noise-related characteristic indicators.

[0051] Figure 4 This is a framework diagram of a brain functional imaging noise suppression method and system based on spatiotemporal features. After inputting 4-dimensional data, the data is first segmented and feature decomposed. Then, noise suppression and elimination are performed based on temporal and spatial feature vectors respectively. Finally, the final denoising output is obtained by weighted averaging and reconstructing the data.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for suppressing noise in brain functional images based on spatiotemporal features, characterized in that, The specific steps are as follows: (1) Sliding window and matrixing; For the input raw 4D brain function image data, the data is first segmented into different time segments using the sliding window method; any time segment will be processed through a binary mask containing only voxel information inside the brain to extract effective spatial information, transforming the 4D data into a two-dimensional time x space matrix. (2) Construct feature vectors representing time and space; For each time segment, the singular value decomposition method is used to decompose the 2D time-space matrix into time coefficient features and spatial distribution features: ;in It is a spatial feature vector, where P is the number of voxels and K is the window length. These are the corresponding time coefficients, and S is the sorted eigenvalue diagonal matrix; (3) Denoising of temporal and spatial features; For any given time segment, the data is considered to be a linear combination of spatiotemporal feature vectors: Noise removal in the feature space includes processing temporal features separately. and spatial features Implement time-domain noise suppression and spatial-domain signal modulation; Spatial signal modulation consists of three steps: First, the signal is purified by excluding clusters with excessively small effective volumes: the spatial feature vectors are first transformed into a three-dimensional spatial distribution, and the spatial distribution is adjusted to a Gaussian distribution by applying Z-score. Then, the spatial map is thresholded based on statistical properties to generate a binary map. All effective clusters under the connectivity criteria are calculated based on the Label-connected components method. Clusters with too few effective voxels are removed. Secondly, redundant feature vectors are discarded based on the graph theory connection matrix: For each input data, all time segments have a certain proportion of overlap; therefore, the spatial distribution of different segments has strong commonalities; so by calculating the correlation of spatial feature vectors of different time segments, a graph theory connection matrix can be formed to represent the spatial feature similarity of all time segments; then, based on the classification method, the threshold is automatically estimated to filter out non-common spatial distributions with low correlation in the connection graph, so as to discard components that cannot exist stably in different time segments; Finally, soft smoothing is applied to compensate for the effective information content of the spatial distribution: most of the original data feature vectors are in a low signal-to-noise ratio state due to thermal noise. A Gaussian smoothing kernel is applied to the cleaned signal clusters to enhance the local intensity of the signal-related voxels. Temporal noise suppression consists of two steps: First, irrelevant noise covariates are estimated; alignment is used to perform preliminary motion correction and estimate head motion parameters, namely translation and rotation in the XYZ directions; based on an autoregressive model, 24 covariates of head motion are obtained, namely 6 basic motion parameters, 6 time-domain derivatives, quadratic terms of the 6 basic motion parameters, and time-domain derivatives of the 6 quadratic terms; then, the probability maps of gray matter, white matter, and cerebrospinal fluid tissue are segmented based on the DARTEL algorithm. A threshold was set and a binary mask for white matter and cerebrospinal fluid was created; the n strongest tissue signals were calculated as covariates based on aCompCorr; drift signals with spectral energy less than 0.01Hz were added, and a total of 25+n irrelevant noise covariates were finally estimated. Then, noise energy in each time coefficient is eliminated based on linear regression variables; for each time segment, the eigenvector of any time-domain coefficient with a Pearson correlation coefficient exceeding a certain value is regressed to remove the associated noise covariate signal, in order to deeply remove noise; (4) Data reconstruction and weighted average; After denoising, the processed spatiotemporal feature vectors are reconstructed into the original 4D data based on linear combination. Due to the introduction of the sliding window, some time points will be reused multiple times in the calculation, and the frequency of reuse will be used as a weight to calculate the weighted average of the data at specific time points. Finally, the denoised 4D functional MRI image data is output for subsequent analysis.

Images