Limits...
Validation of Shared and Specific Independent Component Analysis (SSICA) for Between-Group Comparisons in fMRI

View Article: PubMed Central - PubMed

ABSTRACT

Independent component analysis (ICA) has been widely used to study functional magnetic resonance imaging (fMRI) connectivity. However, the application of ICA in multi-group designs is not straightforward. We have recently developed a new method named “shared and specific independent component analysis” (SSICA) to perform between-group comparisons in the ICA framework. SSICA is sensitive to extract those components which represent a significant difference in functional connectivity between groups or conditions, i.e., components that could be considered “specific” for a group or condition. Here, we investigated the performance of SSICA on realistic simulations, and task fMRI data and compared the results with one of the state-of-the-art group ICA approaches to infer between-group differences. We examined SSICA robustness with respect to the number of allowable extracted specific components and between-group orthogonality assumptions. Furthermore, we proposed a modified formulation of the back-reconstruction method to generate group-level t-statistics maps based on SSICA results. We also evaluated the consistency and specificity of the extracted specific components by SSICA. The results on realistic simulated and real fMRI data showed that SSICA outperforms the regular group ICA approach in terms of reconstruction and classification performance. We demonstrated that SSICA is a powerful data-driven approach to detect patterns of differences in functional connectivity across groups/conditions, particularly in model-free designs such as resting-state fMRI. Our findings in task fMRI show that SSICA confirms results of the general linear model (GLM) analysis and when combined with clustering analysis, it complements GLM findings by providing additional information regarding the reliability and specificity of networks.

No MeSH data available.


ROC curves representing the classification performance of shared and specific patches and three resting-state networks at different SNR values. For the SSICA, the ROC curve at each SNR level is generated using classification of 3200 extracted components (5 permutations × 4 anatomical noise level × 20 threshold values × 8 patches). For the regular approach, as there is no internal threshold, the ROC curve at each SNR level is generated using classification of 3200 extracted components (100 permutations × 4 anatomical noise level × 8 patches). RG01 and RG001 stand for the regular gICA approach with α = 0.01 and α = 0.001, respectively.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC5037228&req=5

Figure 4: ROC curves representing the classification performance of shared and specific patches and three resting-state networks at different SNR values. For the SSICA, the ROC curve at each SNR level is generated using classification of 3200 extracted components (5 permutations × 4 anatomical noise level × 20 threshold values × 8 patches). For the regular approach, as there is no internal threshold, the ROC curve at each SNR level is generated using classification of 3200 extracted components (100 permutations × 4 anatomical noise level × 8 patches). RG01 and RG001 stand for the regular gICA approach with α = 0.01 and α = 0.001, respectively.

Mentions: To investigate classification performance of shared and specific patches classification, ROC analysis was used as described in Section Patch Extraction Performance. The same above-mentioned parameters for SNR and anatomical noise levels, used in generating Figure 3, were used for the SSICA and the regular gICA approach. The ROC curves are based on the classification of the five simulated patches and three highly-reported resting-state networks (i.e., default mode, auditory, and visual networks) as described in Methods. Since in the SSICA the classification threshold needs to be specified in advance, we ran the algorithm multiple times (5 permutations) with 20 different threshold values ranging from 0 to 1, to estimate the ROC curve. Figure 4 shows the ROC curves at different SNR values for the SSICA in red, the regular gICA approach with α = 0.01 in blue, and the regular gICA approach with α = 0.001 in green (α is the two-sample t-test significance threshold). The ROC curves at each SNR level is generated using classification of 3200 (5 × 4 × 20 × 8) extracted components. Although the total number of active voxels in the difference map varies with the choice of α, as shown in Figure 4 the classification performance is quite stable using different α values, and is constantly superior in the SSICA compared to the regular approach. Particularly, the benefits of SSICA in the classification of specific patches (lower false positive rate) are greater at low SNR values.


Validation of Shared and Specific Independent Component Analysis (SSICA) for Between-Group Comparisons in fMRI
ROC curves representing the classification performance of shared and specific patches and three resting-state networks at different SNR values. For the SSICA, the ROC curve at each SNR level is generated using classification of 3200 extracted components (5 permutations × 4 anatomical noise level × 20 threshold values × 8 patches). For the regular approach, as there is no internal threshold, the ROC curve at each SNR level is generated using classification of 3200 extracted components (100 permutations × 4 anatomical noise level × 8 patches). RG01 and RG001 stand for the regular gICA approach with α = 0.01 and α = 0.001, respectively.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC5037228&req=5

Figure 4: ROC curves representing the classification performance of shared and specific patches and three resting-state networks at different SNR values. For the SSICA, the ROC curve at each SNR level is generated using classification of 3200 extracted components (5 permutations × 4 anatomical noise level × 20 threshold values × 8 patches). For the regular approach, as there is no internal threshold, the ROC curve at each SNR level is generated using classification of 3200 extracted components (100 permutations × 4 anatomical noise level × 8 patches). RG01 and RG001 stand for the regular gICA approach with α = 0.01 and α = 0.001, respectively.
Mentions: To investigate classification performance of shared and specific patches classification, ROC analysis was used as described in Section Patch Extraction Performance. The same above-mentioned parameters for SNR and anatomical noise levels, used in generating Figure 3, were used for the SSICA and the regular gICA approach. The ROC curves are based on the classification of the five simulated patches and three highly-reported resting-state networks (i.e., default mode, auditory, and visual networks) as described in Methods. Since in the SSICA the classification threshold needs to be specified in advance, we ran the algorithm multiple times (5 permutations) with 20 different threshold values ranging from 0 to 1, to estimate the ROC curve. Figure 4 shows the ROC curves at different SNR values for the SSICA in red, the regular gICA approach with α = 0.01 in blue, and the regular gICA approach with α = 0.001 in green (α is the two-sample t-test significance threshold). The ROC curves at each SNR level is generated using classification of 3200 (5 × 4 × 20 × 8) extracted components. Although the total number of active voxels in the difference map varies with the choice of α, as shown in Figure 4 the classification performance is quite stable using different α values, and is constantly superior in the SSICA compared to the regular approach. Particularly, the benefits of SSICA in the classification of specific patches (lower false positive rate) are greater at low SNR values.

View Article: PubMed Central - PubMed

ABSTRACT

Independent component analysis (ICA) has been widely used to study functional magnetic resonance imaging (fMRI) connectivity. However, the application of ICA in multi-group designs is not straightforward. We have recently developed a new method named “shared and specific independent component analysis” (SSICA) to perform between-group comparisons in the ICA framework. SSICA is sensitive to extract those components which represent a significant difference in functional connectivity between groups or conditions, i.e., components that could be considered “specific” for a group or condition. Here, we investigated the performance of SSICA on realistic simulations, and task fMRI data and compared the results with one of the state-of-the-art group ICA approaches to infer between-group differences. We examined SSICA robustness with respect to the number of allowable extracted specific components and between-group orthogonality assumptions. Furthermore, we proposed a modified formulation of the back-reconstruction method to generate group-level t-statistics maps based on SSICA results. We also evaluated the consistency and specificity of the extracted specific components by SSICA. The results on realistic simulated and real fMRI data showed that SSICA outperforms the regular group ICA approach in terms of reconstruction and classification performance. We demonstrated that SSICA is a powerful data-driven approach to detect patterns of differences in functional connectivity across groups/conditions, particularly in model-free designs such as resting-state fMRI. Our findings in task fMRI show that SSICA confirms results of the general linear model (GLM) analysis and when combined with clustering analysis, it complements GLM findings by providing additional information regarding the reliability and specificity of networks.

No MeSH data available.