Extended Technical and Clinical Validation of Deep Learning-Based Brainstem Segmentation for Application in...

Area of Science:

• Neuroimaging
• Medical Artificial Intelligence
• Neuroscience

Background:

• Central nervous system disorders, including neurodegenerative diseases, often impact the brainstem, leading to focal atrophy.
• Accurate brainstem segmentation is crucial for evaluating disease progression and developing biomarkers.

Purpose of the Study:

• Optimize deep learning-based brainstem segmentation for diverse pathologies and MRI parameters.
• Validate segmentation methods technically and clinically.
• Enhance segmentation in the presence of brainstem lesions.
• Provide an open-source brainstem segmentation tool.

Main Methods:

• Trained deep learning models (MD-GRU, nnU-Net) on a heterogeneous dataset (n=257).
• Validated performance against ground truth and FreeSurfer.
• Assessed scan-rescan repeatability (n=46) and inter-scanner reproducibility (n=20).
• Evaluated clinical utility in multiple system atrophy and multiple sclerosis cohorts.

Main Results:

• Both MD-GRU and nnU-Net achieved high segmentation performance (Dice ≥ 0.95), outperforming previous models.
• Achieved excellent scan-rescan repeatability and inter-scanner reproducibility.
• Demonstrated comparable performance in detecting brainstem atrophy across methods.
• Lesion filling improved segmentation accuracy in multiple sclerosis patients.

Conclusions:

• Enhanced and validated two automated deep learning brainstem segmentation methods.
• Public release of these tools facilitates broader research into brainstem volume as a neurodegeneration biomarker.