Automated Deep Learning-based Segmentation of the Dentate Nucleus Using Quantitative Susceptibility Mapping MRI

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate segmentation of the dentate nucleus (DN) is crucial for understanding neurological disorders.
• Quantitative susceptibility mapping (QSM) offers valuable insights into brain tissue properties.
• Deep learning (DL) has shown promise in automating complex medical image segmentation tasks.

Purpose of the Study:

• To develop and validate a deep learning-based tool for automated segmentation of the dentate nucleus (DN) on brain MRI quantitative susceptibility mapping (QSM) images.
• To assess the performance and generalizability of the developed DL model across diverse datasets.

Main Methods:

• A retrospective study collected brain QSM images from healthy controls and patients with cerebellar ataxia or multiple sclerosis across nine international datasets.
• A two-step deep learning approach, involving a localization model followed by a segmentation model (nnU-Net framework), was employed.
• Performance was evaluated using intraclass correlation coefficient (ICC), Dice scores, and Pearson correlation coefficients against manual delineations.

Main Results:

• The developed DL model achieved high Dice scores for left and right DN segmentation (0.90 ± 0.03 and 0.89 ± 0.04, respectively).
• The automated tool significantly outperformed the leading existing automated method in external testing (mean Dice scores 0.86 vs 0.57 for left DN, 0.84 vs 0.58 for right DN).
• The model demonstrated strong generalizability across unseen datasets, with automated segmentations highly correlated with manual annotations (r = 0.74 for left DN, r = 0.48 for right DN).

Conclusions:

• The proposed deep learning model provides accurate and efficient segmentation of the dentate nucleus from brain QSM images.
• This tool has the potential to advance research in neurological conditions affecting the cerebellum.
• The publicly available model (https://github.com/art2mri/DentateSeg) facilitates wider adoption and research.