Related Experiment Video
Updated: Oct 5, 2025
05:33
Three-Dimensional Shape Modeling and Analysis of Brain Structures
Published on: November 14, 2019
7.2K
Automated 2D Slice-Based Skull Stripping Multi-View Ensemble Model on NFBS and IBSR Datasets
Anam Fatima1,2, Tahir Mustafa Madni3,4, Fozia Anwar5
1Medical Imaging and Diagnostic Lab, National Centre of Artificial Intelligence, Park Road, Tarlai Kalan, Islamabad, 45550, Pakistan.
Journal of Digital Imaging
|January 27, 2022
Summary
A novel multi-view U-Net (MVU-Net) efficiently strips skulls from brain MRI scans. This 2D approach matches 3D model performance with reduced computational needs, improving accuracy and efficiency.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate brain extraction from Magnetic Resonance Imaging (MRI) is crucial for neurological studies.
- Existing skull stripping methods often require significant computational resources or compromise accuracy.
- Three-dimensional (3D) models offer high performance but are computationally intensive.
Purpose of the Study:
- To introduce and evaluate a novel two-dimensional (2D) slice-based multi-view U-Net (MVU-Net) architecture for enhanced skull stripping.
- To compare the performance of MVU-Net against existing U-Net and skip connection U-Net (SCU-Net) architectures.
- To assess the efficiency and accuracy of the proposed MVU-Net in generating brain masks from multimodal MRI views.
Main Methods:
- Development of a 2D MVU-Net architecture that fuses axial, coronal, and sagittal TI-weighted brain MRI views.
- Linear fusion of predictions from all three views to generate a final brain mask.
- Training and testing on two public datasets: Internet Brain Segmentation Repository (IBSR) and Neurofeedback Skull-stripped (NFBS).
Main Results:
- MVU-Net achieved superior performance compared to U-Net and SCU-Net on the IBSR dataset, with mean Dice Score Coefficient (DSC) of 0.9184, sensitivity of 0.9397, and specificity of 0.9908.
- On the NFBS dataset, MVU-Net demonstrated improved results over U-Net and SC-UNet, achieving a mean DSC of 0.9681, sensitivity of 0.9763, and specificity of 0.9954.
- The 2D MVU-Net approach matched 3D model performance while utilizing fewer computational resources.
Conclusions:
- The proposed MVU-Net architecture offers an accurate and computationally efficient solution for brain MRI skull stripping.
- Fusion of multiple 2D MRI views in the MVU-Net framework significantly enhances segmentation accuracy.
- MVU-Net presents a viable alternative to traditional 3D methods, particularly in resource-constrained environments.
