Group-representative functional network estimation from multi-subject fMRI data via MRF-based image segmentation
Bingjing Tang,Aditi Iyer,Vinayak Rao,Nan Kong
DOI: https://doi.org/10.1016/j.cmpb.2019.07.004
Abstract:Background and objective: There has been growing interest in using functional connectivity patterns, determined from fMRI data to characterize groups of individuals exhibiting common traits. However, the present challenge lies in efficient and accurate identification of distinct patterns observed consistently across multiple subjects. Existing approaches either impose strong assumptions, require aligning images before processing, or require data-intensive machine learning algorithms with manually labeled training datasets. In this paper, we propose a more principled and flexible approach to address this. Methods: Our approach redefines the problem of estimating the group-representative functional network as an image segmentation problem. After employing an improved clustering-based ICA scheme to pre-process the dataset of individual functional network images, we use a maximum a posteriori-Markov random field (MAP-MRF) framework to solve the image segmentation problem. In this framework, we propose a probabilistic model of the individual pixels of the fMRI data, with the model involving a latent group-representative functional network image. Given an observed dataset, we apply a novel and efficient variational Bayes algorithm to recover the associated latent group image. Our methodology seeks to overcome limitations in more traditional schemes by exploiting spatial relationships underlying the connectivity maps and accounting for uncertainty in the estimation process. Results: We validate our approach using synthetic, simulated and real data. First, we generate datasets from the proposed forward model with subject-specific binary masking and measurement noise, as well as from a variant of the model without measurement noise. We use both datasets to evaluate our model, along with two algorithms: coordinate-ascent algorithm and variational Bayes algorithm. We conclude that our proposed model with variational Bayes outperforms other competitors, even under model-misspecification. Using variational Bayes offers a significant improvement in performance, with almost no additional computational overhead. We next test our approach on simulated fMRI data. We show our approach is robust to initialization and can recover a solution close to the ground truth. Finally, we apply our proposed methodology along with baselines to a real dataset of fMRI recordings of individuals from two groups, a control group and a group suffering from depression, with recordings made while individuals were subjected to musical stimuli. Our methodology is able to identify group differences that are less clear under competing methods. Conclusions: Our model-based approach demonstrates the advantage of probabilistic models and modern algorithms that account for uncertainty in accurate identification of group-representative connectivity maps. The variational Bayes methodology yields highly accurate results without increasing the computational load compared to traditional methods. In addition, it is robust to model misspecification, and increases the ability to avoid local optima in the solution.