Brain gray matter nuclei segmentation on quantitative susceptibility mapping using dual-branch convolutional neural...

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Abnormal brain iron accumulation is linked to neurodegenerative diseases.
• Quantitative susceptibility mapping (QSM) measures magnetic susceptibility for iron detection.
• Accurate segmentation of brain nuclei is crucial but challenging for clinicians.

Purpose of the Study:

• To develop an automated method for segmenting brain gray matter nuclei.
• To improve the accuracy and efficiency of quantitative susceptibility mapping (QSM) analysis.
• To address limitations of existing 3D CNN methods in handling memory constraints and spatial context.

Main Methods:

• Proposed a dual-branch residual-structured U-Net (DB-ResUNet) using 3D convolutional neural networks (CNNs).
• Incorporated image patches at different resolutions to balance accuracy and memory efficiency.
• Utilized both QSM and 3D T1-weighted imaging (T1WI) as input modalities.

Main Results:

• The DB-ResUNet achieved superior segmentation accuracy compared to single-branch networks, atlas-based methods, and classical 3D CNNs.
• The model demonstrated high correlation between measured susceptibility values/volumes and manually annotated regions of interest.
• The approach effectively addressed the spatial contextual information loss issue in patch-based 3D CNNs.

Conclusions:

• The proposed DB-ResUNet offers an accurate and efficient automated solution for brain gray matter nuclei segmentation.
• This method facilitates reliable quantitative susceptibility mapping (QSM) for studying neurodegenerative diseases.
• The dual-branch architecture with multi-resolution patches represents a significant advancement in medical image segmentation.