Brain Lesion Parcellation Using a Precomputed Human Brain Connectome Improves Symptom Localization

1773 

Abstract Submission 

William Drew¹, Alexander Cohen², Amy Brodtmann³, Maurizio Corbetta⁴, Natalia Egorova-Brumley⁵, Sophia Gozzi³, Jordan Grafman⁶, Andrew Naidech⁶, Joel Voss⁷, B. T. Thomas Yeo⁸, Michael Fox⁹, Shan Siddiqi⁹

¹Brigham and Women's Hospital, Harvard Medical School, Boston, MA, ²Boston Children's Hospital, Harvard Medical School, Boston, MA, ³Monash University, Melbourne, Victoria, Australia, ⁴Clinica Neurologica, Department of Neuroscience, University of Padova, Padova, Italy, ⁵University of Melbourne, Melbourne, Victoria, Australia, ⁶Northwestern University, Chicago, IL, ⁷University of Chicago, Chicago, IL, ⁸National University of Singapore, Singapore, ⁹Brigham and Women’s Hospital, Harvard Medical School, Boston, MA

William Drew  
Brigham and Women's Hospital, Harvard Medical School
Boston, MA

Alexander Cohen  
Boston Children's Hospital, Harvard Medical School
Boston, MA

Amy Brodtmann, MBBS FRACP PhD  
Monash University
Melbourne, Victoria, Australia

Maurizio Corbetta  
Clinica Neurologica, Department of Neuroscience, University of Padova
Padova, Italy

Natalia Egorova-Brumley  
University of Melbourne
Melbourne, Victoria, Australia

Sophia Gozzi  
Monash University
Melbourne, Victoria, Australia

Jordan Grafman  
Northwestern University
Chicago, IL

Andrew Naidech, MD MSPH  
Northwestern University
Chicago, IL

Joel Voss  
University of Chicago
Chicago, IL

B. T. Thomas Yeo  
National University of Singapore
Singapore

Michael Fox  
Brigham and Women’s Hospital, Harvard Medical School
Boston, MA

Shan Siddiqi  
Brigham and Women’s Hospital, Harvard Medical School
Boston, MA

Lesion network mapping (LNM) uses brain lesions to causally link symptoms to functional brain networks. However, there are two main limitations of this method. First, LNM is computationally inefficient, inhibiting the use of large connectome datasets like the 40,000+ subject UK Biobank. Second, LNM assumes that a brain lesion is connected to a single functional brain network. However, lesions may span multiple functionally distinct brain regions, potentially introducing noise. Here, we attempt to resolve these limitations by developing a method to parcellate brain lesions using resting-state functional connectivity into regions connected to functionally distinct networks and a method to compute lesion network maps rapidly and efficiently. We also apply these methods to improve lesion-symptom prediction of depression.

First, we generated a precomputed human brain connectome (PHBC) by computing the mean whole-brain functional connectivity of each voxel across 1,000 healthy individuals. Next, to address LNM's first limitation, we developed the "precomputed" LNM method. Using the PHBC, a weighted average of the functional connectivity maps associated with the ROI's voxels is computed, and a scaling factor is applied to account for differences between individual voxel BOLD signal strengths. Weights are computed as the standard deviation of a voxel's BOLD signal amplitude. To address LNM's second limitation, we used the PHBC to parcellate stroke lesions into distinct regions that share common patterns of functional connectivity. For each lesion, we extract a connectivity matrix that consists of connectivity measures between every pair of voxels in the lesion. Next, the connectivity matrix is thresholded to remove weak connections between voxels. The thresholded connectivity matrix is then clustered using Infomap, a modular community detection algorithm that is commonly used for resting-state network parcellation, to group voxels into clusters with similar connectivity profiles. For each lesion, we used the largest parcel as a seed to generate functional connectivity maps and compared these maps to depression outcomes in five lesion datasets (n=449), yielding a map of connectivity of lesion parcels associated with depression. The largest parcel was selected as a simple metric for the optimal component. We hypothesized that LNM using lesion parcels would explain more variance in post-lesion depression than whole lesions.

Functional connectivity maps of whole lesions generated using both the "precomputed" and conventional LNM methods were similar (spatial r=0.997) and were far more efficiently computed (~7X faster) when using a 1,000-subject normative functional connectome. In a leave-one-dataset-out cross-validation, lesion network maps derived from the largest parcel of each lesion from four datasets predicted depression outcomes in the fifth (r=0.155, p<0.001). This was significantly stronger (p=0.0011) than the predictive value of whole lesions.

The PHBC and the "precomputed" LNM method accelerate LNM, enabling functional connectivity analyses of even single voxel lesions in a computationally efficient manner. The PHBC also enables functional parcellation of lesions, potentially removing noise from lesion analyses. This parcellation method significantly improved lesion-symptom localization when using each lesion's largest parcel compared to whole lesions. If lesion-symptom localization can be improved even when using a simple size metric to pick the optimal parcel, we hypothesize that further improvements in LNM can be achieved by using a more specialized parcel selection method that could include selecting multiple relevant lesion parcels to consider their interacting network effects. The PHBC enables functional brain network analysis with a higher level of granularity than conventional LNM, making it possible to discriminate between the interacting network effects of a functionally heterogeneous brain lesion.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling

fMRI Connectivity and Network Modeling ¹

Methods Development ²

Segmentation and Parcellation

Computational Neuroscience

Computing

FUNCTIONAL MRI

Modeling

Open-Source Code

Open-Source Software

Other - Connectivity; Connectome