Non-local means filtering for cortical parcellation of resting fMRI

2194 

Abstract Submission 

Monday, June 27 & Tuesday, June 28 

Chitresh Bhushan¹, Minqi Chong¹, Soyoung Choi¹, Anand Joshi¹, Justin Haldar¹, Hanna Damasio¹, Richard Leahy¹

¹University of Southern California, Los Angeles, CA, United States

Intensity variations over time in resting fMRI exhibit spatial correlation patterns consistent with a set of large scale cortical networks [1]. However, visualizations of this data on the brain surface, even after extensive preprocessing, are dominated by local intensity fluctuations that obscure larger scale behavior. Our novel adaptation of non-local means (NLM) filtering [2], which we refer to as temporal NLM (tNLM), reduces these local fluctuations without the spatial blurring that occurs when using standard Gaussian or Laplace-Beltrami (LB) filtering. tNLM filtering respects functional boundaries and shows superior performance in functional parcellation of the cortex in a population of 40 subjects available from Human Connectome Project dataset [3].

tNLM is an edge-preserving denoising method that uses weighted average of data in a large neighborhood where weights are chosen adaptively depending on similarities between the fMRI time series at the vertices. In comparison, weights in LB filtering are chosen based on the spatial proximity between the vertices. tNLM directly exploits the temporal information in the data by using a weight based on the correlation between the time series. Specifically, tNLM uses a weight w(s,r) = exp(corr(s,r)/h), where h is the filtering parameter and corr(s,r) is the correlation between time series at the vertices s and r. This avoids mixing across functional boundaries, since the time series in different functional areas will be less strongly correlated than within each distinct functional area. This effect is shown in Fig.1 where we compare weights of tNLM and LB filtering for several filtering parameters for a vertex lying in visual cortex. While LB weights are isotropic around the vertex, the tNLM weights are largest within visual cortex.

We show a comparison of original, tNLM and LB filtered fMRI data in Fig.2. We see that LB results have several isolated patches which seems to be uncorrelated to any large network. We use the filtered and unfiltered fMRI data with N-cuts classification [4] with fully connected graphs (no spatial contiguity constrain) to parcellate the cortex into several networks. Fig.3 shows an example of parcellation obtained with 6 classes. LB results contain several regions lying along the boundaries of larger parcels that are identified as separate parcels, indicating mixing across functional regions. tNLM results on the other hand show large contiguous regions classified as separate networks. Next, we evaluate the quality of parcellations by quantifying the fractional agreement of the parcels with the regions identified independently using task experiments and with probabilistic Brodmann areas (BA). We obtained the distribution of performances across several filtering parameters (h=0.6, 0.72, 1.73 for tNLM and t=2, 4, 6 for LB) and difference numbers of cuts (N=2 to 400). Then we compared the mean of the peak performance of both the methods across a population of 40 subjects in Fig.4, which shows that tNLM has higher agreement for several task and BA labels. We also performed a Wilcoxon signed-rank test on peak performances, which revealed that tNLM had significantly higher agreement (p<0.01) for 6 out of 9 task labels and 15 out of 26 BAs.

tNLM filtering is able to denoise resting fMRI data while also retaining much of the spatial structure that reflects ongoing dynamic brain activity. tNLM achieves significant improvement over traditional Laplace-Beltrami filtering in quantitative comparisons with task and BA labels. Correlated variations in activity are also directly visible in the image and movie rendering of the filtered data, which may facilitate exploratory data analysis leading to new insights into the dynamics of spontaneous brain activity.

[1] Smith et at. (2009) PNAS, vol 106(31); [2] Buades et al. (2005) CVPR, vol 2; [3] Van Essen et al. (2013) NeuroImage vol 80; [4] Shi & Malik (2000) PAMI, vol 22(8)

fMRI Connectivity and Network Modeling

Segmentation and Parcellation ¹

Task-Independent and Resting-State Analysis ²

FUNCTIONAL MRI