Deep MRI brain extraction: A 3D convolutional neural network for skull stripping.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Brain extraction is vital for neuroimaging analysis.
• Existing methods are limited to specific MRI contrasts (e.g., T1-weighted) and struggle with pathological tissues.
• There is a need for robust brain extraction across diverse MRI modalities and conditions.

Purpose of the Study:

• To develop and evaluate a 3D convolutional deep learning architecture for accurate brain extraction.
• To overcome limitations of current methods, particularly with enhanced MRI contrasts and altered brain tissue.
• To demonstrate the method's performance on both clinical and public datasets.

Main Methods:

• A 3D convolutional neural network (CNN) architecture was designed for brain extraction.
• The model was trained to handle multiple MRI contrasts, including non-enhanced and contrast-enhanced T1w, T2w, and FLAIR.
• Performance was evaluated on a clinical dataset with brain tumors (N=53) and three public datasets (IBSR, LPBA40, OASIS; N=135).

Main Results:

• The CNN significantly outperformed six common tools on a challenging clinical dataset, achieving a mean Dice score of 95.19.
• On public datasets, the method achieved state-of-the-art or near state-of-the-art performance (e.g., IBSR: 96.32, LPBA40: 96.96, OASIS: 95.02).
• The model demonstrated high average specificity across all datasets, with sensitivity adjustable via thresholding.

Conclusions:

• The proposed 3D deep learning approach offers robust and accurate brain extraction across various MRI modalities and pathologies.
• This method surpasses current tools, showing significant potential for clinical trials and large-scale neuroimaging studies.
• Fast GPU implementation enables rapid predictions (<1 minute), facilitating clinical workflow integration.