Related Experiment Video
Updated: Oct 7, 2025
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
Generalizing deep learning brain segmentation for skull removal and intracranial measurements
Yue Liu1, Yuankai Huo2, Blake Dewey3
1College of Information Science and Engineering, Northeastern University, Shenyang 110819, China; Electrical Engineering and Computer Science, Vanderbilt University, TN, USA.
This study introduces a deep learning method for automatic brain segmentation and volume estimation, including total intracranial volume (TICV) and posterior fossa volume (PFV), on MRI scans with or without skulls. The approach effectively addresses data limitations using transfer learning for enhanced neuroimaging analysis.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Image Analysis
Background:
- Accurate brain volumetric measurements, such as total intracranial volume (TICV) and posterior fossa volume (PFV), are crucial for structural magnetic resonance imaging (MRI) analyses.
- Neuroimaging data is increasingly distributed in a skull-stripped format for privacy, posing challenges for existing segmentation methods.
- Limited availability of manually traced atlases with comprehensive labels hinders the development of robust deep learning models for brain segmentation and volumetric measurements.
Purpose of the Study:
- To develop a generalized deep learning approach for simultaneous whole brain segmentation and intracranial volume estimation (TICV, PFV) on both skull-intact and skull-stripped T1-weighted MRI.
- To overcome data scarcity by employing a transfer learning strategy for automatic label estimation.
- To provide an open-source tool for reproducible neuroimaging research.
Main Methods:
- Utilized U-Net tiles for simultaneous automatic TICV estimation and whole brain segmentation.
- Implemented a transfer learning approach: pre-training U-Net tiles on large-scale atlases without TICV/PFV labels, followed by refinement using limited atlases with these labels.
- Extended the method to effectively process skull-stripped brain MRI data.
Main Results:
- Achieved promising results in whole brain segmentation and volume estimation for both skull-intact and skull-stripped brains.
- Demonstrated high performance based on mean Dice similarity coefficients, mean surface distance, and absolute volume similarity metrics.
- The developed method shows significant potential for accurate and efficient neuroimaging analysis.
Conclusions:
- The proposed U-Net tile-based transfer learning method enables accurate brain segmentation and volumetric measurements (TICV, PFV) on diverse MRI data.
- The approach successfully handles both skull-intact and skull-stripped images, addressing a key limitation in current neuroimaging practices.
- This open-source tool facilitates advanced brain volumetric analysis in research settings.
