Related Experiment Video
Updated: Aug 5, 2025
08:05
Measuring Statistical Learning Across Modalities and Domains in School-Aged Children Via an Online Platform and Neuroimaging Techniques
Published on: June 30, 2020
7.6K
Learning Strategies for Contrast-agnostic Segmentation via SynthSeg for Infant MRI data
Ziyao Shang1, Md Asadullah Turja1, Eric Feczko2
1University of North Carolina, Chapel Hill, USA.
Proceedings of Machine Learning Research
|March 27, 2023
Summary
Infant-SynthSeg is a new deep learning model for segmenting infant brain MRI scans across the first year of life. It overcomes age-related image changes, providing consistent results for neurodevelopmental disorder research.
Area of Science:
- Medical Imaging
- Neuroscience
- Artificial Intelligence
Background:
- Longitudinal infant brain MRI studies are crucial for identifying neurodevelopmental disorders.
- Current deep learning segmentation models are limited to narrow age ranges due to significant postnatal brain development changes.
- Using multiple age-specific models causes inconsistencies and biases in longitudinal data.
Purpose of the Study:
- To develop an age-agnostic deep learning segmentation model for infant brains within the first year of life.
- To address limitations of existing models in handling variations in MRI intensity and contrast during early development.
- To create a unified segmentation framework for consistent analysis of longitudinal infant brain data.
Main Methods:
- Extended the contrast-agnostic SynthSeg framework to create Infant-SynthSeg.
- Focused on advanced synthetic data generation and augmentation strategies tailored for infant brain features.
- Trained and evaluated models using diverse infant MRI data across the first year of life.
- Compared Infant-SynthSeg performance against traditional contrast-aware models like nnU-net.
Main Results:
- Infant-SynthSeg demonstrated consistently high segmentation performance across all ages within the first year of life.
- The model successfully segmented structures present at specific ages (e.g., cerebellar white matter at 1 month) across the entire age range.
- While robust across ages, Infant-SynthSeg's performance was slightly lower than highly specialized, age-specific models.
Conclusions:
- Infant-SynthSeg offers a reliable, age-agnostic solution for segmenting infant brain MRI scans, crucial for longitudinal neurodevelopmental research.
- The model's ability to handle diverse developmental stages reduces inconsistencies in analyzing infant brain development.
- Further research could explore hybrid approaches combining Infant-SynthSeg's generalizability with age-specific model precision.
Keywords:
Convolutional neural networksData augmentationDeep learningInfant brain segmentationNeurodevelopmental DisordersNeuroimaging
