Project Proposal: Attention-based fusion of multiple graph heat networks for structural to

functional brain mapping

Team Name/No – Alienware/06

Background

The paper tries to address a cognitive science research problem of mapping the structural connectivity (SC) that
refers to the anatomical connections between different regions of the brain, to steady-state functional connectivity
(FC) of brain that represents the temporal synchronization or coordination of neural activity between different regions
of the brain, using graph-based modelling. Characterizing the SC-FC mapping is an open and challenging research
problem in cognitive neuroscience. Such models can be used to identify the biomarkers that underlie any deviation
from the expected FC based on the SC in various diseases such as Autism Spectrum Disorder (ASD), Dementia, and
many more. The paper specifically implements a variant of Graph Convolutional network - Graph Heat Network (GHN)
for SC-FC mapping.

Paper Summary

As key contribution, the paper proposes an end-to-end learnable A-GHN architecture for learning the SC-FC mapping
on brain graphs. The method is grounded in the theory of the reaction-diffusion process in the cognitive domain while
retaining the key properties of generalizability, scalability, and tractability in the deep learning framework. The paper
also provides comprehensive empirical analysis, including perturbation experiments and a detailed ablation study, to
demonstrate the proposed model’s robustness and validity on a large publicly available dataset.

Project Description

Most of the existing methods in SC-FC mapping space suffer from the computational overhead related to
scalability or exhibit sub-optimal performance over brain graphs. The authors of the paper used attention-
based fusion of multiple GraphHeat networks (A-GHN) that efficiently employs multi-scale diffusion to promote
computational tractability and scalability.
The primary objective of the project is to verify the claims of the authors that justify the usage of multiple graph
heat networks aggregated with attention layer. We planned to run the code on current model along with
baseline and related models mentioned in the paper to verify if claims and also to see if metrics can be
improved tweaking parameters of exiting models mentioned in paper.
The objective also includes attempts to enhance the performance on SC-FC mapping by utilizing the current
SOTA models in Graph space such as GraphMAE, DiffusionGraph, GraphSAGE, graph attention networks
(GAT), attention-based graph neural network (AGNN), variational graph autoencoder (VGAE), Graph2Gauss
(G2G) and, deep graph infomax (DGI).

What task/objective will you address

• Investigating the need of using multiple sub graphs (GHN) used in the paper on the task of mapping SC-FC.
Is there an alternative to using multiple graphs? And Is there an optimal value to the number of sub graphs?
• Attempt to enhance the accuracy of the existing models described in baseline and related work (by altering
model architecture and tweaking hyper parameters).
• Attempt to enhance the accuracy on the SC-FC mapping task, by replacing Attention - Multiple GraphHeat
networks (A-GHN) with some of the new SOTA models for graphs, such as GraphMAE, DiffusionNet,
GraphSAGE, graph attention networks (GAT), attention-based graph neural network (AGNN), variational
graph autoencoder (VGAE), Graph2Gauss (G2G) and, deep graph infomax (DGI). We planned to try and
implement only some of the above models, which ever are compatible to our dataset and problem.
What data will you use? Already existing code?

The authors of the paper considered a popular and widely used dataset from the human connectome project (HCP).
The structural connectivity - functional connectivity (SC-FC) pairs of a total of 1058 subjects were utilized from the HCP
repository. The structural connectivity (SC) matrix, derived from diffusion tensor imaging (DTI), reveals the white-matter
fiber connections between regions of interest (ROIs). The elements of the SC matrix correspond to the normalized
count of streamlines connecting pairs of regions. On the other hand, the FC matrix is characterized by Pearson’s
correlation of time series from resting state fMRI for different brain regions. The code for the paper is available. We
plan to run it to validate the results, enhance the base line and existing methods by tweaking any possible parameters,
and by using current SOTA in graph net space.

What baseline(s) will you use? How will you evaluate your results?

We will use the two baseline models multiple kernel learning (MKL) and GCN-based Encoder-Decoder architecture
used in this paper. If baseline working models are available the we will download and run to reproduce the results,
otherwise we will use the previously published scores. For evaluation will use the Pearson correlation between the
ground-truth FC and predicted FC. The mean Pearson correlation on test subjects of HCP dataset, using 2-layered A-
GHN is 0.799, which is current SOTA. Using newer models in graph net space such as GraphMAE or VGAE, we will
try to enhance the correlation score. Also we plan to provide better insights on 86 ROI regions correlation analysis.

Timeframe + Work Distribution (For two months)

Task Work Distribution Start and End Dates

Phase One Reproduce the results of the paper Reproduce the results: Vijay, 01/03/2024 to
and explore the limitation of the Janak 20/03/2024
paper Find limitations of paper: Pradeep,
Shruti

Phase Two Improve the limitations of the 4 new models will be distributed 21/03/2024 to
paper (try to implement new between 4 members. 10/04/2024
methods on given dataset)

Phase Three Presentation and final report All will document their respective 10/04/2024 to
preparation work 20/04/2024

References

• Oota, S.R., Yadav, A., Dash, A. et al. Attention-based fusion of multiple graphheat networks for structural to
functional brain mapping. Sci Rep 14, 1184 (2024).
• Surampudi, S. G. et al. Multiple kernel learning model for relating structural and functional connectivity in the
brain. Sci. Rep. 8, 1–14 (2018).
• Li, Y., Shafpour, R., Mateos, G. & Zhang, Z. Mapping brain structural connectivities to functional networks
via graph encoderdecoder with interpretable latent embeddings. In 2019 IEEE Global Conference on Signal
and Information Processing (GlobalSIP), 1–5 (IEEE, 2019).