CAU-Net: A Deep Learning Method for Deep Gray Matter Nuclei Segmentation.

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Neurological Disorders

Background:

• Abnormal iron deposition in deep gray matter nuclei is linked to neurological diseases.
• Quantitative susceptibility mapping (QSM) allows in vivo measurement of brain iron content.
• Accurate segmentation of deep gray matter nuclei is essential for QSM analysis.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for segmenting deep gray matter nuclei.
• To assess the performance of the proposed model on diverse MRI datasets.
• To investigate the impact of training and inference strategies on segmentation accuracy.

Main Methods:

• Proposed a contrast attention U-Net architecture for nuclei segmentation.
• Evaluated the model on two datasets from different MRI devices and sequences.
• Investigated the effects of test time augmentation, data augmentation, deep supervision, and nonuniform patch sampling.

Main Results:

• The proposed contrast attention U-Net outperformed commonly adopted network structures on both datasets.
• Test time augmentation significantly improved segmentation performance during inference.
• Sufficient data augmentation, deep supervision, and nonuniform patch sampling were critical for enhancing segmentation accuracy during training.

Conclusions:

• The developed contrast attention U-Net provides accurate segmentation of deep gray matter nuclei for QSM.
• Optimal training and inference strategies are as vital as advanced network design for improving segmentation.
• This work contributes to more reliable in vivo brain iron quantification for neurological disease research.