Heterogeneous domain adaptation of connectomes across atlases using optimal transport

2206 

Abstract Submission 

Qinghao Liang¹, Dustin Scheinost¹

¹Yale University, New Haven, CT

Qinghao Liang  
Yale University
New Haven, CT

Dustin Scheinost  
Yale University
New Haven, CT

Brain parcellations are crucial for analyzing fMRI and DWI datasets, reducing dimensionality, noise, and enhancing interpretability. Connectomes built from preprocessed parcellated data are the standard for data sharing in neuroimaging research. However, the lack of standardized brain atlases limits comparability and data aggregation across studies. In this paper, we propose a method using optimal transport to map between atlas domains and transform connectomes across domains.

We employ optimal transport to map between two domains. For fMRI data, we calculate transportation matrices for parcel-wise signals from two atlases, yielding the functional mapping through averaging across all time points and subjects [1]. For structural connectomes, we use a graph-matching method to minimize the Gromov-Wasserstein discrepancy [2] and learn a transportation matrix, which is then averaged across all subjects. With the established mapping, we transform connectomes from the source to the target domain in three steps: 1. Decompose the connectome into its node representation. 2. Map this representation to the target domain. 3. Estimate the connectome in the target domain as the product of the transformed node representation.

In our experiments, we employed five atlases (Shen [3], Craddock [4], Brainnetome [5], Dosenbach [6], and Schaefer [7]). Functional mapping was learned from the Yale dataset [8], while structural mapping was learned from the HCP datasets [9]. We assessed method accuracy by evaluating functional connectomes in the HCP dataset and structural connectomes in the HCP-D dataset, using Pearson's correlation between the estimated connectomes and the "ground-truth" structural connectomes as the metric. To validate our domain adaptation approach, we applied it to predictive modeling tasks for fluid intelligence and age in cross-validation. The training and testing sets included connectomes from different atlases, with data in the training set transformed into the testing set's domain for modeling. This process was repeated 100 times, and predictive performance was assessed using Pearson's correlation between true and predicted values.

Figure 1a illustrates the precision of connectome estimation. In most scenarios, we observe notable correlations between the transformed connectome and its corresponding target connectome for both functional and structural mapping. These correlations provide insights into the degree of similarity or overlap between atlases. Particularly, when transforming data from the Shen atlas to the Craddock atlas, we observe high correlations (r = 0.744 for FC and r = 0.692 for SC), as both atlases are derived from clustering time series using variations of the N-cut algorithm. In contrast, the correlation between data transformed from the Dosenbach atlas, constructed based on task activation analysis, and other atlases is significantly lower. Figure 1b displays the predictive performance, with diagonal values representing results using "ground-truth" data. Notably, predictive performance aligns closely with estimation accuracy, indicating that connectomes estimated with higher precision also yield better predictive performance. It's worth noting that functional mapping demonstrates superior performance on both functional and structural data.

We devised a heterogeneous domain adaptation technique that utilizes optimal transport and matrix factorization to facilitate connectome transformation across different atlases. Our approach yielded substantial correlations between the estimated data and ground truth data, particularly when the atlases exhibited relatively high similarity.

Furthermore, we found that predictive performance remained largely consistent when the estimation accuracy was high. Functional mapping, which leverages data from multiple timepoints, outperformed structural mapping. In cases where timeseries data was unavailable, structural mapping provided a viable alternative.

Classification and Predictive Modeling

Connectivity (eg. functional, effective, structural) ²

Brain Atlases ¹

FUNCTIONAL MRI

Machine Learning

Open Data

WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC