Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An Adaptively Weighted Averaging Method for Regional Time Series Extraction of fMRI-based Brain Decoding

Jianfei Zhu, Baichun Wei, Jiaru Tian, Feng Jiang, Chunzhi Yi

TL;DR

This work addresses the loss of spatial information when extracting regional fMRI time series by ROI-wise averaging. It introduces Adaptively Weighted Average Time Series (AWATS), which builds three axis-wise spatial representations per ROI, resamples them to a fixed length $q$ (set to 10), and learns adaptive voxel weights via a fully connected neural network jointly trained with a downstream cognitive-state decoder (STDfMRI). AWATS is evaluated on the HCP task fMRI data across Schaefer atlas scales and compared to ATS and voxel-wise methods, showing up to a 5 percentage-point gain in decoding accuracy, improved separability in manifold embeddings, and interpretable ROI contributions via Shapley values. The approach demonstrates robustness across window sizes, resampling sizes, and training data and offers a principled, task-driven improvement to the fMRI processing pipeline with potential extensions to other brain-imaging tasks and datasets.

Abstract

Brain decoding that classifies cognitive states using the functional fluctuations of the brain can provide insightful information for understanding the brain mechanisms of cognitive functions. Among the common procedures of decoding the brain cognitive states with functional magnetic resonance imaging (fMRI), extracting the time series of each brain region after brain parcellation traditionally averages across the voxels within a brain region. This neglects the spatial information among the voxels and the requirement of extracting information for the downstream tasks. In this study, we propose to use a fully connected neural network that is jointly trained with the brain decoder to perform an adaptively weighted average across the voxels within each brain region. We perform extensive evaluations by cognitive state decoding, manifold learning, and interpretability analysis on the Human Connectome Project (HCP) dataset. The performance comparison of the cognitive state decoding presents an accuracy increase of up to 5\% and stable accuracy improvement under different time window sizes, resampling sizes, and training data sizes. The results of manifold learning show that our method presents a considerable separability among cognitive states and basically excludes subject-specific information. The interpretability analysis shows that our method can identify reasonable brain regions corresponding to each cognitive state. Our study would aid the improvement of the basic pipeline of fMRI processing.

An Adaptively Weighted Averaging Method for Regional Time Series Extraction of fMRI-based Brain Decoding

TL;DR

This work addresses the loss of spatial information when extracting regional fMRI time series by ROI-wise averaging. It introduces Adaptively Weighted Average Time Series (AWATS), which builds three axis-wise spatial representations per ROI, resamples them to a fixed length (set to 10), and learns adaptive voxel weights via a fully connected neural network jointly trained with a downstream cognitive-state decoder (STDfMRI). AWATS is evaluated on the HCP task fMRI data across Schaefer atlas scales and compared to ATS and voxel-wise methods, showing up to a 5 percentage-point gain in decoding accuracy, improved separability in manifold embeddings, and interpretable ROI contributions via Shapley values. The approach demonstrates robustness across window sizes, resampling sizes, and training data and offers a principled, task-driven improvement to the fMRI processing pipeline with potential extensions to other brain-imaging tasks and datasets.

Abstract

Brain decoding that classifies cognitive states using the functional fluctuations of the brain can provide insightful information for understanding the brain mechanisms of cognitive functions. Among the common procedures of decoding the brain cognitive states with functional magnetic resonance imaging (fMRI), extracting the time series of each brain region after brain parcellation traditionally averages across the voxels within a brain region. This neglects the spatial information among the voxels and the requirement of extracting information for the downstream tasks. In this study, we propose to use a fully connected neural network that is jointly trained with the brain decoder to perform an adaptively weighted average across the voxels within each brain region. We perform extensive evaluations by cognitive state decoding, manifold learning, and interpretability analysis on the Human Connectome Project (HCP) dataset. The performance comparison of the cognitive state decoding presents an accuracy increase of up to 5\% and stable accuracy improvement under different time window sizes, resampling sizes, and training data sizes. The results of manifold learning show that our method presents a considerable separability among cognitive states and basically excludes subject-specific information. The interpretability analysis shows that our method can identify reasonable brain regions corresponding to each cognitive state. Our study would aid the improvement of the basic pipeline of fMRI processing.
Paper Structure (26 sections, 10 equations, 4 figures, 6 tables)

This paper contains 26 sections, 10 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. fMRI data acquisition and preprocessing
  4. AWATS extraction
  5. Spatial representation for each ROI
  6. Resampling the representation vectors
  7. Adaptively weighted average using a fully connected NN
  8. Cognitive state decoding using STDfMRI
  9. Time windowing
  10. The STDfMRI model architecture
  11. Experiment setting
  12. Manifold learning for fMRI regional time series data
  13. Interpretability analysis using Shapley value method
  14. Results
  15. Cognitive state decoding performance
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Overview of the workflow of this paper. (a) The AWATS extraction method measures the data of a single ROI and of a TR from x, y and z axes, generating three spatial representation vectors. Then, the representation vectors are resampled to the same length and adaptively averaged by a fully connected network that is jointly trained with the downstream tasks (i.e. cognitive state decoding). (b) The structure of the STDfMRI model is designed for comparing AWATS and ATS in cognitive state decoding. (c) AWATS and ATS are spatially embedded into a two-dimensional space for temporal trajectory visualization. (d) Interpretability analysis is performed for the STDfMRI model using the Shapley value method and the ROI contributions are presented as brain maps.
  • Figure 2: BOLD signals of 5 voxels from the fist ROI of Schaefer 100.
  • Figure 3: Visualization of the regional time series embedding obtained by UMAP for the WM task. (a) The color of the scatters represents different subjects. (b) The color of the scatters represents different task conditions.
  • Figure 4: Contribution maps calculated using the Shapley value method. (a) Contribution maps for AWATS. (b) Contribution maps for ATS. The contributions have been normalized by dividing them by the maximum positive contribution to depict the relative importance of different brain areas in identifying specific task conditions.