Using Deep Clustering to Improve fMRI Dynamic Functional Connectivity Analysis

• Published in: bioRxiv
• Main Authors: Spencer, Arthur P C, Goodfellow, Marc
• Format: Paper
• Language: English
• Published: Cold Spring Harbor Cold Spring Harbor Laboratory Press 16.12.2021; Cold Spring Harbor Laboratory
• Edition: 1.1
• Subjects: Feature selection; Functional magnetic resonance imaging; Neuroscience; deep learning; autoencoders; dimensionality reduction; Dynamic functional connectivity; sliding window correlations
• DOI: 10.1101/2021.12.14.472680
• ISSN: 2692-8205

Dynamic functional connectivity (dFC) analysis of resting-state fMRI data is commonly per- formed by calculating sliding-window correlations (SWC), followed by k-means clustering in order to assign each window to a given state. Studies using synthetic data have shown that k-means per- formance is highly dependent on sliding window parameters and signal-to-noise ratio. Additionally, sources of heterogeneity between subjects may affect the accuracy of group-level clustering, thus affecting measurements of dFC state temporal properties such as dwell time and fractional occu- pancy. This may result in spurious conclusions regarding differences between groups (e.g. when comparing a clinical population to healthy controls). Therefore, is it important to quantify the ability of k-means to estimate dFC state temporal properties when applied to cohorts of multiple subjects, and to explore ways in which clustering performance can be maximised. Here, we explore the use of dimensionality reduction methods prior to clustering in order to map high-dimensional data to a lower dimensional space, providing salient features to the subse- quent clustering step. We assess the use of deep autoencoders for feature selection prior to applying k-means clustering to the encoded data. We compare this deep clustering method to feature selec- tion using principle component analysis (PCA), uniform manifold approximation and projection (UMAP), as well as applying k-means to the original feature space using either L1 or L2 distance. We provide extensive quantitative evaluation of clustering performance using synthetic datasets, representing data from multiple heterogeneous subjects. In synthetic data we find that deep clus- tering gives the best performance, while other approaches are often insufficient to capture temporal properties of dFC states. We then demonstrate the application of each method to real-world data from human subjects and show that the choice of feature selection method has a significant effect on group-level measurements of state temporal properties. We therefore advocate for the use of deep clustering as a precursor to clustering in dFC. Competing Interest Statement The authors have declared no competing interest.