Open-source pipeline for multi-class segmentation of the spinal cord with deep learning

Area of Science:

• Medical Imaging
• Neuroscience
• Machine Learning

Background:

• Accurate segmentation of anatomical structures in Magnetic Resonance Imaging (MRI) is crucial for quantitative analysis and diagnosis.
• Manual segmentation is time-consuming and requires specialized expertise, limiting its scalability.
• Developing automated or semi-automated methods is essential for efficient and reproducible neuroimaging research.

Purpose of the Study:

• To present an open-source computational pipeline for training neural networks to segment structures from MRI data.
• To demonstrate the pipeline's effectiveness on homogeneous datasets with limited manual segmentation requirements.
• To showcase its application in segmenting spinal cord white and grey matter in both non-human primates and humans.

Main Methods:

• Development of an open-source pipeline utilizing neural networks for image segmentation.
• Training the pipeline on datasets requiring a relatively low number of manual segmentations (dozens or fewer).
• Validation through two use-case scenarios: marmoset spinal cord segmentation (with lesions) and human grey matter segmentation.

Main Results:

• The pipeline successfully segments structures of interest from MRI data.
• It demonstrates efficacy on homogeneous datasets, requiring minimal manual annotations.
• Successful application in segmenting spinal cord white and grey matter in complex scenarios, including lesioned marmoset brains and human datasets.

Conclusions:

• The presented open-source pipeline offers an efficient and accessible tool for MRI-based structure segmentation.
• It significantly reduces the need for extensive manual segmentation, making advanced neuroimaging analysis more feasible.
• The pipeline's versatility is confirmed by its successful application across different species and anatomical regions.