Individualized ROI Optimization via Maximization of Group-wise Consistency of Structural and Functional Profiles

Kaiming Li, Lei Guo, Carlos Faraco, Dajiang Zhu, Fan Deng, Tuo Zhang, Xi Jiang, Degang Zhang, Hanbo Chen, Xintao Hu, L. Stephen Miller, Tianming LiuNeurIPS 2010

Introduction

In neuroimaging, defining a Region of Interest (ROI)—the specific "patch" of the brain responsible for a task—is notoriously difficult. While we have general brain atlases, every individual's brain is folded differently. This paper introduces a novel method to precisely locate functional ROIs in individual brains by maximizing group-wise consistency. By jointly modeling anatomical locations, structural fiber connections, and functional activity patterns, the researchers use simulated annealing to "nudge" ROIs into positions that are more biologically consistent across a group than standard fMRI activation maps.

The Problem with Current ROI Definitions

Current methods for defining ROIs suffer from several critical issues:

  1. Inter-subject variability: One person’s "working memory center" might be millimeters away from another's.
  2. Preprocessing artifacts: Common techniques like spatial smoothing can shift the apparent peak of brain activity by over 12mm.
  3. Inconsistency: A region defined by functional activity (fMRI) might not align well with the physical "wiring" (DTI) of the brain.

If ROIs are misaligned, any subsequent analysis of how brain regions communicate (connectivity modeling) becomes unreliable. The authors seek a way to "fine-tune" these locations so that the ROIs across a group of subjects are truly comparable in both form and function.

Methods: An Energy Minimization Framework

The authors propose an Energy Minimization framework. They define a total energy function ($E$) that represents the "cost" of an ROI's location. The goal is to minimize this energy, which is composed of three distinct constraints:

Optimization framework integrating anatomical, structural, and functional constraints.
  • Anatomical Constraint ($E_a$): Uses a Gaussian model in atlas space to ensure the ROI doesn't drift too far from its expected biological landmark, punishing extreme outliers.
  • Structural Connectivity Constraint ($E_c$): Analyzes DTI (Diffusion Tensor Imaging) data to create a "fingerprint" of where white-matter fibers from the ROI lead. It minimizes the variance of these fingerprints across the group.
  • Functional Connectivity Constraint ($E_f$): Analyzes fMRI data to see how the ROI's activity correlates with other ROIs, seeking to make the functional network pattern as consistent as possible across subjects.

Because the "search space" of the brain's surface is rugged and full of local minima, the authors use Simulated Annealing. This algorithm allows the system to occasionally accept "worse" positions early on to avoid getting stuck, eventually settling into a global optimum where the ROIs are most consistent across the whole group.

Datasets

The study utilized data from 25 university students scanned with a 3T GE Signa scanner, providing both fMRI and DTI data. Subjects performed a modified OSPAN (Operation Span) task, a standard test for working memory. The researchers focused on 16 specific ROIs associated with the working memory system (e.g., insula, precentral gyrus, dorsolateral prefrontal cortex).

Key Insights

  • The Structure-Function Link: The most significant insight is the validation of the "close relationship" between structural and functional profiles. Optimizing for structural consistency (fibers) naturally led to a drop in functional energy, and vice versa. This suggests that biological "truth" lies at the intersection of these two modalities.
  • Beyond Activation Peaks: Standard fMRI analysis looks for the "brightest" spot of activity. This paper argues that the brightest spot isn't always the most representative spot. By looking at the context—how that spot is wired to the rest of the brain—we find more stable landmarks.
  • Individualization via Group Wisdom: Paradoxically, to find the best location for one person, you need to look at everyone. The group-wise consistency acts as a North Star for individual optimization.

Key Results

Visual comparisons of ROIs before and after optimization, showing structural fiber tracts and functional activation maps. Source, p. 2

Visual comparisons of ROIs before and after optimization, showing structural fiber tracts and functional activation maps.
  • Improved Structural Consistency: Visualizations of fiber bundles showed that after optimization, subjects who previously looked like "outliers" possessed fiber patterns that matched the group consensus.
  • Reduced Functional Variance: The standard deviation of functional connectivity matrices was significantly reduced across the 16 ROIs after optimization, indicating a more stable network definition.
  • Convergence: The simulated annealing algorithm proved robust, with multiple runs from random starting points converging to nearly the same energy level after about 24 temperature steps.
  • Anatomical Realism: Optimized ROIs moved toward clear anatomical landmarks, such as the tip of the upper bank of the parieto-occipital sulcus, even if the initial fMRI activation was slightly off-target.

Limitations and Future Work

The current model treats ROIs as spheres of a fixed size (5mm radius). In reality, functional regions have complex, irregular shapes and varying sizes. Future work should include optimizing ROI size and shape. Additionally, simulated annealing is computationally expensive, especially as the number of subjects or ROIs increases. While this study focused specifically on the working memory network, the authors plan to apply this framework to other systems—such as visual, auditory, and emotion networks—to see if the consistency gains hold across the entire brain.

Summary updated

BibTeX
@inproceedings{li2010individualized,
  author = {Kaiming Li and Lei Guo and Carlos Faraco and Dajiang Zhu and Fan Deng and Tuo Zhang and Xi Jiang and Degang Zhang and Hanbo Chen and Xintao Hu and L. Stephen Miller and Tianming Liu},
  title = {{Individualized ROI Optimization via Maximization of Group-wise Consistency of Structural and Functional Profiles}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2010},
}
Figure