Related Experiment Video
Updated: Jun 9, 2025
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
15.6K
mSPD-NN: A Geometrically Aware Neural Framework for Biomarker Discovery from Functional Connectomics Manifolds
Niharika S D'Souza1, Archana Venkataraman2
1IBM Research, Almaden, San Jose, USA.
Information Processing in Medical Imaging : Proceedings of the ... Conference
|October 29, 2024
Summary
This study introduces a novel neural network, the mSPD-NN, for analyzing brain connectomes by respecting their geometric properties. It accurately estimates the mean of connectivity data, revealing biomarkers in ADHD-ASD comorbidities.
Area of Science:
- Neuroimaging
- Machine Learning
- Statistics
Background:
- Connectomics utilizes neuroimaging for connectivity data analysis.
- Current methods often overlook the matrix manifold geometry of connectomes.
- Estimating the mean of connectivity matrices is challenging due to complex geometry.
Purpose of the Study:
- To propose a geometrically aware neural framework, mSPD-NN, for estimating the geodesic mean of symmetric positive definite (SPD) matrices.
- To address the limitations of existing analytical frameworks in handling connectome data geometry.
- To develop a robust method for analyzing complex neuroimaging data.
Main Methods:
- Developed the mSPD-NN, a neural network utilizing bilinear fully connected layers with tied weights.
- Implemented a novel loss function to optimize the matrix-normal equation for Fréchet mean estimation.
- Tested on synthetic data for performance evaluation against alternative SPD mean estimation methods.
Main Results:
- The mSPD-NN demonstrated competitive performance in scalability and robustness to noise on synthetic data.
- Experiments on resting-state fMRI (rs-fMRI) data showcased the framework's real-world flexibility.
- Identified stable biomarkers linked to network differences in ADHD-ASD comorbidities.
Conclusions:
- The mSPD-NN provides an effective, geometrically informed approach for analyzing connectome data.
- This method advances statistical and machine learning applications in neuroimaging.
- The framework has potential for biomarker discovery in neurological and psychiatric disorders.
