From Electrode to Global Brain: Integrating Multi- and Cross-Scale Brain Connections and Interactions Under Cross-Subject and Within-Subject Scenarios

TL;DR

A novel multi-scale spatial domain adaptation network (MSSDAN) consists of both multi-scale spatial feature extractor (MSSFE) and deep domain adaptation method called multi-scale spatial domain adaptation (MSSDA) which is to integrate the principles of multi-scale brain topological structures in order to solve the multi-scale spatial data distribution difference problem.

Abstract

The individual variabilities of electroencephalogram signals pose great challenges to cross-subject motor imagery (MI) classification, especially for the data-scarce single-source to single-target (STS) scenario. The multi-scale spatial data distribution differences can not be fully eliminated in MI experiments for the topological structure and connection are the inherent properties of the human brain. Overall, no literature investigates the multi-scale spatial data distribution problem in STS cross-subject MI classification task, neither intra-subject nor inter-subject scenarios. In this paper, a novel multi-scale spatial domain adaptation network (MSSDAN) consists of both multi-scale spatial feature extractor (MSSFE) and deep domain adaptation method called multi-scale spatial domain adaptation (MSSDA) is proposed and verified, our goal is to integrate the principles of multi-scale brain topological structures in order to solve the multi-scale spatial data distribution difference problem.

Figures (2)

• Figure 1: Overview of the MSSDAN. a, Set-up and the flowchart of MSSDAN in STS cross-subject MI classification experiment, where multi-scale spatial data distribution differences are adaptively minimized and the unsupervised classification of the target set is accomplished. b, Topological permutation of the raw MI data. c, Architecture of the proposed domain adaptation network, MSSDAN, which consists of two key components: (1) MSSFE, a multi-scale feature extractor which extracts the deep spatial features of MI data; (2) MSSDA, multi-scale spatial domain adaptation including electrode, region, and hemisphere adaptations, transferring the spatiotemporal knowledge from the source set to the target set.
• Figure 2: Topological permutation and multi-scale spatial feature extractor. a, Topological permutation, where the electrode serials are permutated with different topological rules to facilitate the feature extraction of deep features from different brain regions. b, Multi-scale spatial feature extractor, where the deep features from different brain scales are extracted with customized convolution kernels and sizes. c, Brain region atlas obtained by electrode convolution, representing the topological properties of the permutated EEG data.