Connectomic profiles for individualized resting state networks and regions of interest

Abstract

Functional connectivity analysis of human brain resting state functional magnetic resonance imaging (rsfMRI) data and resultant functional networks, or RSNs, have drawn increasing interest in both research and clinical applications. A fundamental yet challenging problem is to identify distinct functional regions or regions of interest (ROIs) that have accurate functional correspondence across subjects. This article presents an algorithmic framework to identify ROIs of common RSNs at the individual level. It first employed a dual-sparsity dictionary learning algorithm to extract group connectomic profiles of ROIs and RSNs from noisy and high-dimensional fMRI data, with special attention to the well-known inter-subject variability in anatomy and then identified the ROIs of a given individual by employing both anatomic and group connectomic profile constraints using an energy minimization approach. Applications of this framework demonstrated that it can identify individualized ROIs of RSNs with superior performance over commonly used registration methods in terms of functional correspondence, and a test-retest study revealed that the framework is robust and consistent across both short-interval and long-interval repeated sessions of the same population. These results indicate that our framework can provide accurate substrates for individualized rsfMRI analysis.

Keywords: anatomical variability; connectomic profiles; cortical parcellation; dictionary learning; individualized ROIs/RSNs.

FIG. 1.

Illustration of the connectomic profile of posterior cingulate cortex or PCC. The left panel shows the whole-brain connectivity patterns of PCC for 193 subjects, and the right panel depicts the corresponding learned dictionary elements. For each of the five elements, the normalized summation of loadings [Eq. (4)] was displayed on the top. The first element was chosen as the connectomic profile of PCC. The orange arrows highlight the correspondence between connectivity pattern and the connectomic profile.

FIG. 2.

Illustration of the connectomic profile for default mode network (DMN). The left column is the dorsal view. The middle column shows the lateral (top) and medial (bottom) views of the left hemisphere. The right column provides the lateral (top) and medial (bottom) views of the right hemisphere. The color bar displays the intensity of DMN's profile, which indicates how much a region is involved in DMN.

FIG. 3.

Illustration of region of interest (ROI) section for the DMN. (A) The profile of the DMN in Figure 2 mapped onto an inflated cortical surface. (B) The selected ROIs for the DMN.

FIG. 4.

Cortical parcellation by the framework. (A–C) Are illustrations of the cortical parcellation. (A) Dorsal view. (B, C) Lateral views. The parcels are colored in a way that neighboring parcels have different colors, and parcels with the same color mean nothing, except that they are not neighbors. (D) Shows the box-and-whisker diagram of η values for both registration based parcellation and the present method.

FIG. 5.

Histograms for fractions of subjects sharing same dictionary elements. (A, B) Are the histograms of fractions of subjects that share dominant and trivial dictionary elements, respectively. For both figures, the horizontal axis refers to the fraction of subjects sharing the same dictionary elements, for example, 0.8 at the X-axis in (A) means 80% of subjects have the same dominant connectomic pattern, whereas the vertical axis indicates how many parcels out of 400 have the same fraction. For instance, the red star in (A) indicates that there are ∼4% parcels, where 60% subjects share the same dominant connectomic patterns, whereas the blue star in (B) indicates there are ∼2% parcels where 4% subjects have the same trivial connectomic patterns. (A) Small fraction value (averaged at 3.43%) in (B) indicates five dictionary elements in the present model is sufficient.

FIG. 6.

Common group resultant functional networks (RSNs) and their profiles. (A–C) Illustrate profiles of three group RSNs, that is, temporal–parietal junction, DMN, and ventral attention network, respectively. For each RSN profile, there are four subfigures that show different views, that is, dorsal, ventral, and bilateral. (D) Summarizes the common RSNs across three datasets D1, D2, and D3.

FIG. 7.

Comparison of individual RSNs with surface registration. The two figures compare the individual ROIs' functional correspondence between surface-based registration (REG) and our framework (Profile) for datasets D1 and D2. Horizontal axis denotes the 27 consistent RSNs, and vertical axis is the correlation coefficient of a ROI's connectivity pattern with the connectomic profile of corresponding RSN. All RSNs except those with red stars have significant improvements.

FIG. 8.

Similarity of individual RSNs with the group connectomic profiles. Horizontal axis is RSN ID, and vertical axis is the similarity of individual RSNs with the connectomic profiles. Each similarity was averaged across 23 subjects, and the error bar shows the standard deviation. For each RSN, similarities from three sessions of the group are depicted.

FIG. 9.

Test and retest of the individual ROIs/RSNs through our framework. Horizontal axis is RSN ID, and vertical axis is the corrected p-value using false discovery rate (FDR). P12 compares the connectivity patterns of individual ROIs from sessions 1 and 2 with the group connectomic profile of corresponding RSN. P13 compares sessions 1 and 3, and P23 is for sessions 2 and 3.