Related Experiment Video
Updated: Oct 2, 2025
04:25
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
3.0K
Automated Segmentation of Midbrain Structures in High-Resolution Susceptibility Maps Based on Convolutional Neural
Weiwei Zhao1, Yida Wang1, Fangfang Zhou1
1Shanghai Key Laboratory of Magnetic Resonance, School of Physics and Electronic Science, East China Normal University, Shanghai, China.
Frontiers in Neuroscience
|February 28, 2022
Summary
This study introduces a convolutional neural network (CNN) method for precise segmentation of midbrain nuclei, including the red nucleus (RN), substantia nigra (SN), and subthalamic nucleus (STN), in susceptibility maps. The CNN approach demonstrates high accuracy, comparable to manual delineations, aiding neuroimaging research.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Neuroscience
Background:
- Accurate delineation of midbrain nuclei (red nucleus, substantia nigra, subthalamic nucleus) is crucial for neuroimaging studies.
- Neurodegenerative diseases and other conditions affect these critical midbrain structures.
- Current segmentation methods may lack precision for high-resolution susceptibility maps.
Purpose of the Study:
- To develop and validate a convolutional neural network (CNN) based method for segmenting midbrain nuclei.
- To assess the accuracy of CNN segmentation against manual delineations using Dice coefficients.
- To compare extracted volume and magnetic susceptibility values from automated and manual segmentations.
Main Methods:
- Acquired high-resolution susceptibility maps (75 subjects, 0.83 × 0.83 × 0.80 mm³ voxel size) on a 3T MRI.
- Utilized a pre-trained deeply supervised attention U-net for initial weight provision.
- Employed five-fold cross-validation and evaluated models on a separate test cohort, using Dice coefficients for accuracy assessment.
Main Results:
- Achieved high mean Dice scores: 0.903 (RN), 0.864 (SN), and 0.777 (STN).
- CNN segmentation accuracy was not significantly different from inter-rater reliability (p > 0.05).
- Automated CNN extraction showed significant correlation with manual tracing for volume and magnetic susceptibility values (p < 0.01).
Conclusions:
- A CNN-based method enables precise segmentation of midbrain structures in high-resolution susceptibility maps.
- The automated approach offers accuracy comparable to manual delineation, improving efficiency in neuroimaging.
- This method facilitates more reliable quantitative analysis of midbrain nuclei in clinical and research settings.
