Related Experiment Video
Updated: Oct 17, 2025
06:48
Automated Segmentation of Cortical Grey Matter from T1-Weighted MRI Images
Published on: January 7, 2019
9.1K
Convolutional Neural Network Based Frameworks for Fast Automatic Segmentation of Thalamic Nuclei from Native and
Lavanya Umapathy1,2, Mahesh Bharath Keerthivasan2,3, Natalie M Zahr4
1Department of Electrical and Computer Engineering, University of Arizona, Tucson, AZ, USA.
Neuroinformatics
|October 9, 2021
Summary
This study introduces a 3D convolutional neural network (CNN) for segmenting thalamic nuclei using Magnetization Prepared Rapid Gradient Echo (MPRAGE) images. Synthesizing White-Matter-nulled MPRAGE (WMn-MPRAGE) contrast improved segmentation accuracy and identified alcohol use disorder-related thalamic atrophy.
Area of Science:
- Neuroimaging
- Medical Image Analysis
- Artificial Intelligence in Medicine
Background:
- Thalamic nuclei segmentation is crucial for understanding neurological diseases but challenging with standard MRI due to poor contrast.
- Existing multi-atlas methods are computationally intensive and time-consuming.
- Limited spatial resolution and image distortion affect diffusion and functional MRI-based techniques.
Purpose of the Study:
- To develop and evaluate a 3D convolutional neural network (CNN) framework for automated thalamic nuclei parcellation using T1-weighted MPRAGE images.
- To investigate the impact of transforming MPRAGE images to a White-Matter-nulled MPRAGE (WMn-MPRAGE) contrast on segmentation performance.
- To assess the clinical utility of the proposed method in differentiating thalamic atrophy in alcohol use disorder (AUD).
Main Methods:
- A 3D CNN framework was trained for thalamic nuclei segmentation on MPRAGE images.
- Two approaches were compared: native contrast segmentation (NCS) and synthesized contrast segmentation (SCS) using WMn-MPRAGE.
- Thalamic nuclei labels were generated using the THOMAS multi-atlas technique. Segmentation accuracy and clinical utility were evaluated on healthy and AUD cohorts.
Main Results:
- Both NCS and SCS CNNs achieved high segmentation accuracy (Dice > 0.84 for large nuclei, > 0.7 for small nuclei).
- SCS demonstrated significant improvements in Dice scores for specific nuclei (medial geniculate, centromedian) and volume difference (ventral anterior, ventral posterior lateral) compared to NCS.
- In the AUD cohort, SCS accurately identified significant thalamic atrophy in the ventral lateral posterior nucleus, consistent with prior research, while NCS showed spurious atrophy.
Conclusions:
- CNN-based segmentation offers a fast and automated method for thalamic nuclei prediction from MPRAGE images.
- Transforming MPRAGE to WMn-MPRAGE contrast enhances segmentation performance for specific thalamic nuclei.
- The SCS approach shows promise for detecting subtle neuroanatomical changes, such as thalamic atrophy in AUD.
