DBSegment: Fast and robust segmentation of deep brain structures considering domain generalization.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Deep brain structure segmentation is crucial for clinical applications like diagnosis and surgical planning.
• Current segmentation methods often rely on time-consuming registration-based pipelines, limiting their clinical utility.

Purpose of the Study:

• To develop a one-step, efficient, and robust deep learning solution for segmenting deep brain structures directly in native MRI space.
• To significantly reduce the computational time for deep brain segmentation compared to existing methods.

Main Methods:

• A convolutional neural network within the nnU-Net framework was employed for segmentation.
• The method involved a preprocessing step for MRI orientation conformity, followed by deep learning segmentation.
• Training and validation utilized seven datasets, while seven independent datasets were used for testing, including leave-one-dataset-out cross-validation.

Main Results:

• The deep learning model achieved an average Dice score similarity of 0.89 ± 0.04 against a registration-based gold standard.
• Computational time was drastically reduced from 43 minutes for registration-based methods to 1.3 minutes with the proposed deep learning approach.
• The method demonstrated robustness and generalizability across multiple datasets and domains.

Conclusions:

• The proposed deep learning method offers a fast, reliable, and accurate solution for deep brain segmentation.
• This approach has the potential to enhance clinical workflows and can be extended to segment other brain structures.
• The tool is publicly available, facilitating broader adoption in research and clinical settings.