Normalization of Multimodal Brain Networks for Integral and Predictive Mapping of Neurological Disorders

Project Information

NormNets

Grant agreement ID: 101003403

DOI

10.3030/101003403

Project closed

EC signature date 20 March 2020

Start date 1 April 2020

End date 31 March 2022

Funded under

• SPREADING EXCELLENCE AND WIDENING PARTICIPATION

Total cost

€ 145 355,52

EU contribution

€ 145 355,52

Coordinated by

ISTANBUL TEKNIK UNIVERSITESI
Türkiye

Periodic Reporting for period 1 - NormNets (Normalization of Multimodal Brain Networks for Integral and Predictive Mapping of Neurological Disorders)

Reporting period: 2020-04-01 to 2022-03-31

Summary of the context and overall objectives of the project

• What is the problem/issue being addressed?

Modern network science has introduced exciting new opportunities for understanding the brain as a complex system of interacting units in both health and disease and across the human lifespan. Despite the rapidly growing interdisciplinary science of complex networks, the brain remains one of greatest challenges for network science. In the face of the oncoming ‘tsunami’ of disordered neuroimaging datasets on brain networks, namely the ongoing 14 connectomic brain data collection studies for Connectome Related to Human Disease (CRHD) initiative funded by the National Institutes of Health, there is an urgent need for handling connectomic datasets with unprecedented scale and heterogeneity via developing for the first time advanced learning-based models of multimodal brain networks that will enable accurate mapping of disordered neural pathways, and a rational approach to designing effective interventions and treatments in the context of public health.

Creating a connectional brain template (CBT) for a population of subjects can be a powerful tool to integrate complementary information from different neuroimaging modalities in order to understand different connectional patterns of both the healthy and the disordered brain. CBTs can be the basis for automated brain analysis in clinical practices when used as references in comparative studies and classification. However, it is not always obvious how to integrate multimodal connectomic data together, nor easy to do so in practice, in order to first understand how the brain's structural, morphological and functional levels interlink to form this integrated complex system, and then identify typical and atypical connectional trends fingerprinting the human brain. This is substantially due to the large variability in brain connectivity across individuals, which limits our ability to distinguish between `healthy' brain connectional variability and `pathological' variability. To disentangle such connectional variability, we need to define a `normalization' or `standardization' process of brain networks. Here, we set out to design geometric deep learning-based methods that can meet such challenges by normalizing a population of multimodal brain networks and thereby using the learned template in downstream brain state classification and prediction tasks.


• Why is it important for society?

With yearly costs of about 800 billion euros and an estimated 179 million people afflicted in 2010 in Europe, neurological disorders are an unquestionable emergency and a grand challenge for neuroscientists. The NormNets Project has contributed to better understanding the underpinning of neurological disorders by providing population-driven connectional brain templates (CBTs) that allow for an easy identification of neurological biomarkers. We also demonstrated that the learned CBTs have prognostic predictive potential that can translate into the development of brain connectivity-targeting clinical trials for treating neurological disorders in addition to integral mapping of the brain across its lifespan.

In addition to that, using these compact and easy to process and load brain templates, one can train and design artificial intelligence (AI) methods that are affordable —i.e. do not require high-computation resources for training. We have shown this “affordable” aspect of CBT-based geometric deep learning model training in our recent published works. This has a strong societal contribution in terms of democratizing AI with application to network neuroscience in low-middle income countries and low-regime environments.


• What are the overall objectives?

The NormNets Project has:
(i) produced the most centered and representative connectional brain template that consistently captures the unique and distinctive traits of a population of multimodal brain networks,
(ii) reliably revealed the integral signature of a particular brain disorder by comparing the estimated disordered CBT with a healthy CBT, and
(iii) helped better predict integral changes in brain connectivity for an individual patient from first brain scan (i.e. baseline acquisition) for early diagnosis.

• Conclusions
We conclude this action with a set of geometric deep learning methods for CBT learning published in high-impact journals and conferences/workshops. More importantly, all source codes and estimated CBTs be publicly shared via GitHub to facilitate the reproducibility of our findings and diffusing the created knowledge by NormNets globally and openly.

GitHub channel: https://github.com/basiralab
YouTube presentation channel: https://bit.ly/3lt7Su7
YouTube code demo channel: https://bit.ly/3wL2Q1s

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

NormNets addressed big gaps in network neuroscience and neurological disorders literature including:

(1) Creating a centered and well-representative connectional brain template for the healthy/disordered human brain using multimodal connectomic data. This is the first project to produce graph neural network methods that normalize multimodal brain networks, thereby creating connectional templates for the human brain.
Our “Deep Graph Normalizer” (DGN) and “Multi-view Graph Normalizer Network” (MGN-Net) codes are openly available on the BASIRA Lab GitHub website.
The first DGN paper got an early accept at MICCAI 2020 and was selected for oral presentation. We then extended it into a more comprehensive MGN-Net paper published in the journal of Medical Image Analysis 2021 (IF:8.54).
** DGN GitHub: https://github.com/basiralab/DGN
** MGN-Net GitHub: https://github.com/basiralab/MGN-Net

(2) Estimating disorder-specific connectional brain templates that capture the integral signature of a neurological disorder from large-scale multimodal connectomes using the nascent field of geometric deep learning.
We have successfully met all these goals which were disseminated through 9 peer-reviewed published journal and MICCAI/workshop conference papers including the first DGN 2020 and MGN-Net 2021 papers. We have also written a survey paper which covers our pioneering works on CBT learning, entitled “Comparative Survey of Multigraph Integration Methods for Holistic Brain Connectivity Mapping” (https://arxiv.org/abs/2204.05110).

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

In addition to meeting the aims of the NormNets project, we have not only advanced the state-of-the-art in the field of brain connectivity mapping but more importantly introduced the new research frontier on “geometric deep learning & integration connectomics”. Our recent 2022 survey paper presents a detailed overview of our pioneering works at this frontier, which present the results of this project (GitHub: https://github.com/basiralab/survey-multigraph-integration-methods).

We have also tapped into the nascent field of federated learning where we set out to learn a CBT in a fully data-preserving manner while leveraging large-scale datasets collected from different hospitals (https://github.com/basiralab/Fed-CBT). Besides, our CBT presents a compact and single representative sample that can be used to train one-shot deep neural networks in low-regime settings, thereby promoting AI affordability in network neuroscience. For more, check our recent works at https://github.com/basiralab/one-representative-shot-learning.

graphical-abstract.png