Paper ID: 2406.16619

No More Sliding-Windows: Dynamic Functional Connectivity Based On Random Convolutions Without Learning

Yongjie Duan, Zhiying Long

Compared to static functional connectivity, dynamic functional connectivity provides more detailed temporal information. The traditional sliding window method constructs functional connectivity matrices by applying a moving time window across the entire time series to calculate correlations between brain regions. However, as a method of feature extraction, it exhibits significant limitations, such as the dependency of feature dimensions on the window length and the generation of features lacking information from other time points within the window. This paper presents RandCon, a novel method for calculating dynamic functional connectivity (DFC), which employs randomly generated multi-dimensional convolution kernels. This method performs convolution operations directly on the BOLD signal without the need for learning, extracting functional connectivity features. Compared to the sliding window method, RandCon shows notable improvements in performance on simulated data, particularly in terms of temporal accuracy and noise resistance. Results from real data indicate that this method maintains stability within short time windows and better identifies gender differences. Furthermore, we propose a more comprehensive theoretical framework, the multi-dimensional convolution method, where the sliding window method and its variants are specific cases of this method. The proposed method is straightforward and efficient, significantly broadening the scope of dynamic functional connectivity research and offering substantial theoretical and practical potential.

Submitted: Jun 24, 2024