Related Experiment Video
Updated: Jul 2, 2025
09:33
Diffusion Tensor Magnetic Resonance Imaging in the Analysis of Neurodegenerative Diseases
Published on: July 28, 2013
28.4K
Structural MRI Harmonization via Disentangled Latent Energy-Based Style Translation
Mengqi Wu1,2, Lintao Zhang1, Pew-Thian Yap1
1Department of Radiology and Biomedical Research Imaging Center, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599, USA.
Machine Learning in Medical Imaging. MLMI (Workshop)
|February 23, 2024
Summary
This study introduces a new method, disentangled latent energy-based style translation (DLEST), to harmonize multi-site brain MRI data. DLEST effectively removes site-specific variations, improving image quality and generalizability for research.
Area of Science:
- Medical Imaging
- Neuroimaging
- Machine Learning
Background:
- Multi-site brain magnetic resonance imaging (MRI) data is susceptible to non-biological variations from different scanning sites.
- Existing harmonization methods, often generative adversarial networks (GANs), are computationally intensive and lack generalizability.
- These variations can confound clinical and research findings.
Purpose of the Study:
- To propose a novel image-level harmonization framework for structural MRI data.
- To develop a method that is computationally efficient and generalizes well to independent data.
- To address the challenge of site effects in multi-site neuroimaging studies.
Main Methods:
- Introduced a disentangled latent energy-based style translation (DLEST) framework.
- Utilized a latent autoencoder for encoding images and a generative model.
- Incorporated an energy-based model for implicit style translation within the latent space.
- Trained the model on 4,092 T1-weighted MRIs across three tasks.
Main Results:
- DLEST effectively disentangles site-invariant image generation from site-specific style translation.
- The method demonstrated superior performance in histogram comparison, site classification, and brain tissue segmentation tasks.
- Qualitative and quantitative evaluations showed DLEST outperforms existing state-of-the-art harmonization techniques.
- Achieved highly generalizable image generation and efficient style translation through the latent space.
Conclusions:
- DLEST offers an effective and efficient solution for harmonizing multi-site brain MRI data.
- The proposed framework improves the reliability and generalizability of neuroimaging data for research.
- DLEST presents a promising advancement in addressing site effects in large-scale MRI studies.
