Automated sub-cortical brain structure segmentation combining spatial and deep convolutional features.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Image Analysis

Background:

• Accurate segmentation of sub-cortical brain structures in Magnetic Resonance Images (MRI) is crucial for understanding neurodegenerative disorders.
• Manual segmentation is time-consuming and inconsistent, necessitating automated methods.
• Existing automated methods often struggle with accuracy and variability.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) for accurate automated segmentation of sub-cortical brain structures in MRI.
• To improve segmentation accuracy by integrating convolutional features with spatial prior information.
• To enhance network training through a restricted sample selection strategy focusing on challenging regions.

Main Methods:

• A novel CNN architecture combining convolutional and spatial prior features was developed.
• A restricted sample selection strategy was employed during training to focus on difficult segmentation areas.
• The method was evaluated on the MICCAI 2012 challenge and IBSR 18 datasets.

Main Results:

• The proposed method achieved performance comparable to the top strategy on the MICCAI 2012 dataset.
• It significantly outperformed traditional methods like FreeSurfer and FIRST on both datasets.
• On the IBSR 18 dataset, it showed comparable or superior results to recent deep learning approaches.

Conclusions:

• The novel CNN approach effectively segments sub-cortical brain structures with high accuracy.
• Both spatial priors and restricted sampling significantly contribute to the method's performance.
• The publicly available tool promotes reproducibility and adoption in neuroimaging research.