A review on brain structures segmentation in magnetic resonance imaging

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Automatic brain structure segmentation in magnetic resonance imaging (MRI) is crucial for diagnosing and monitoring neurological diseases.
• Recent years have seen significant advancements in automated segmentation techniques, moving from structure-specific methods to whole-brain parcellation.

Purpose of the Study:

• To review the state-of-the-art automatic methods for brain structure segmentation.
• To classify and compare different segmentation strategies based on their performance and target structures.

Main Methods:

• Methods are categorized by target structures and segmentation strategies (atlas-based, learning-based, deformable, region-based, hybrid).
• A qualitative and quantitative comparison of each category's strengths, weaknesses, and performance across various brain structures is provided.

Main Results:

• The hippocampus and caudate nucleus are the most frequently segmented structures.
• The accumbens is the most challenging structure to segment (mean DSC 0.69), while the brainstem (0.88), thalamus (0.87), and putamen (0.86) yield the best results.
• Atlas-based methods excel in segmenting the hippocampus, thalamus, and lateral ventricles, whereas deformable methods are effective for the caudate nucleus and putamen.

Conclusions:

• No single automatic segmentation method is currently suitable as a clinical standard.
• Future directions involve integrating multi-atlas techniques with learning-based or deformable approaches for enhanced segmentation accuracy.