Related Experiment Video
Updated: Sep 24, 2025
10:06
High-resolution Functional Magnetic Resonance Imaging Methods for Human Midbrain
Published on: May 10, 2012
13.0K
Multi-Scale Self-Supervised Learning for Multi-Site Pediatric Brain MR Image Segmentation with Motion/Gibbs Artifacts
Yue Sun1,2, Kun Gao2, Weili Lin2
1Department of Shandong Provincial Key Laboratory of Network based Intelligent Computing, University of Jinan, Jinan 250022, China.
Machine Learning in Medical Imaging. MLMI (Workshop)
|May 9, 2022
Summary
This study introduces a multi-scale self-supervised learning (M-SSL) framework for accurate pediatric brain MRI segmentation. The method effectively handles artifacts and multi-site data variations, improving early brain development characterization.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate segmentation of pediatric brain MRI is crucial for understanding early brain development.
- Challenges include imaging artifacts (motion, Gibbs) and multi-site data variability (domain shift).
- Existing methods struggle with these complexities in large-scale datasets.
Purpose of the Study:
- To develop a robust framework for accurate brain tissue segmentation in multi-site pediatric MRI data.
- To address challenges posed by imaging artifacts and domain shift.
- To improve the characterization of early brain development through enhanced segmentation.
Main Methods:
- A multi-scale self-supervised learning (M-SSL) framework was proposed.
- Coarse segmentation on downsampled images provided global anatomic guidance.
- Fine segmentation on original images was refined using this guidance.
- An iterative self-supervised strategy trained site-specific models to mitigate multi-site issues.
Main Results:
- The M-SSL method demonstrated superior performance compared to state-of-the-art approaches.
- Experiments were conducted on the iSeg2019 challenge dataset, featuring real artifacts and multi-site pediatric brain MR images.
- The framework successfully segmented brain tissues with high accuracy despite data complexities.
Conclusions:
- The proposed M-SSL framework offers an effective solution for segmenting pediatric brain MR images with artifacts and multi-site variations.
- This advancement facilitates more accurate characterization of early brain development.
- The method shows significant potential for clinical and research applications in pediatric neuroimaging.
