Discriminating schizophrenia and bipolar disorder by fusing fMRI and DTI in a multimodal CCA+ joint ICA model

Abstract

Diverse structural and functional brain alterations have been identified in both schizophrenia and bipolar disorder, but with variable replicability, significant overlap and often in limited number of subjects. In this paper, we aimed to clarify differences between bipolar disorder and schizophrenia by combining fMRI (collected during an auditory oddball task) and diffusion tensor imaging (DTI) data. We proposed a fusion method, "multimodal CCA+ joint ICA", which increases flexibility in statistical assumptions beyond existing approaches and can achieve higher estimation accuracy. The data collected from 164 participants (62 healthy controls, 54 schizophrenia and 48 bipolar) were extracted into "features" (contrast maps for fMRI and fractional anisotropy (FA) for DTI) and analyzed in multiple facets to investigate the group differences for each pair-wised groups and each modality. Specifically, both patient groups shared significant dysfunction in dorsolateral prefrontal cortex and thalamus, as well as reduced white matter (WM) integrity in anterior thalamic radiation and uncinate fasciculus. Schizophrenia and bipolar subjects were separated by functional differences in medial frontal and visual cortex, as well as WM tracts associated with occipital and frontal lobes. Both patients and controls showed similar spatial distributions in motor and parietal regions, but exhibited significant variations in temporal lobe. Furthermore, there were different group trends for age effects on loading parameters in motor cortex and multiple WM regions, suggesting that brain dysfunction and WM disruptions occurred in identified regions for both disorders. Most importantly, we can visualize an underlying function-structure network by evaluating the joint components with strong links between DTI and fMRI. Our findings suggest that although the two patient groups showed several distinct brain patterns from each other and healthy controls, they also shared common abnormalities in prefrontal thalamic WM integrity and in frontal brain mechanisms.

Figure 1

presents a summary of various blind data-driven methods for data fusion. Note that both mCCA and sCCA belong to CCA, but maximize the correlation in different parts of the data. In mCCA+jICA, mCCA automatically links two datasets via correlation between mixing matrices; while jICA further decomposes the remaining mixtures in the associated maps.

Figure 2

Flow chart of the mCCA+jICA analysis process for the clinical fMRI and DTI data fusion

Figure 3

illustrates the whole simulation results for 3 factors and in 2 ways. The 3 factors are source estimation (Ŝ), mixing matrix estimation (Â) and modal linkage shown by correlation between mixing coefficients of each modality(corr(A₁(:,i),A₂(:,i))), which are displayed in three rows. For each factor, we compared 3 algorithms in different noise conditions (left column) and for each source index (right column). Under each noise condition (PSNR), we illustrate the average of all sources’ estimation. For each specific source, we plotted its mean estimation and standard derivation across all noisy conditions. It is indicated that mCCA+jICA is robust for noise and source type, and its source decomposition performance is the best.

Figure 4

simulation results of comparing separation performance of 3 methods 8 sources for each simulated modality: S₁ (left) and S₂ (right) are designed on purpose with greatly different final vector length (65536 vs. 2000, implying multimodal data); 100 mixtures are generated for each PSNR, here we display PSNR=6dB. The correlations of mixing coefficients between corresponding sources of each modality are listed in the middle, so do their estimations. See jICA separate sources accurately except the 3^rd S₂ signal, while mCCA estimates the modal connection accurately except that it can not decompose images quite well. mCCA+jICA combine both advantages and improve the performance remarkably.

Figure 5

Summary of components with significant group differences in four pair-wise group combinations. Two ICs with the same frame color in two modalities represent joint ICs. Those p values displayed with yellow text pass the Bonferroni correction for multiple comparison (p<0.005), others pass the uncorrected significance level (p<0.05).

Figure 6

Variation of functional spatial maps (AOD_IC9) among groups. Left: The AOD_IC9 derived from mCCA+jICA. Right: the back-reconstructed group sources S_BP, S_SZ and S_{g common} (the common activation of S_BP, S_SZ and S_HC) are overlaid one by one from bottom to top. For display, the components were converted to Z-values and thresholded at |Z|>2.

Figure 7

demonstrates the scatter plots and linear trends between subjects’ age and loading parameters. Specifically, HC in red line, SZ in blue line, BP in green line and trend of all subjects in black line. For FA_IC1, All groups have very significant correlation, thus were plotted using one marker.

Figure 8

This figure attempts to verify that the “linked”(joint) components do correspond to FA changes in known tracts and functional changes in distant regions connected to that tract. Note that we did not perform actual fiber tractography but provided a type of summary statistic. Three joint components are selected with their A₁–A₂ correlation from moderate to high. On the right diagrams, functional region with a red solid line frame indicates that a major portion of it is activated. Region with dotted line frame indicates that only small part of it is activated. We plotted only FA fiber tracts with more than 1 cm³ volume(R+L) activated. Abbreviations are defined below, F: frontal lobe, P: parietal lobe, T: temporal lobe, O: occipital lobe, PH: parahippocampal gyrus, SLF: Superior longitudinal fasciculus, CGC: Cingulum, CST: Corticospinal tract, UF: Uncinate fasciculus IFO: Inferior fronto-occipital fasciculus, ILF: Inferior longitudinal fasciculus, FMIN: Forceps minor, FMAJ: Forceps major, ATR: Anterior thalamic radiation.