Intracranial volume segmentation for neurodegenerative populations using multicentre FLAIR MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Intracranial volume (ICV) segmentation is vital for neurodegenerative disease research using MRI.
• Existing automated methods often struggle with FLAIR MRI, varying acquisition parameters, and diverse pathologies.
• FLAIR MRI is increasingly important for cerebrovascular and neurodegenerative disease analysis.

Purpose of the Study:

• To develop and evaluate automated ICV segmentation algorithms specifically for FLAIR MRI.
• To assess algorithm performance across multicentre datasets with varying parameters and patient cohorts.
• To introduce a comprehensive framework for evaluating segmentation tools for clinical translation.

Main Methods:

• Developed and tested 10 algorithms (2 traditional, 8 deep learning) for ICV segmentation in FLAIR MRI.
• Trained and tested on 175 volumes (8317 images) from dementia and vascular disease cohorts across 47 imaging centres.
• Utilized a human phantom dataset for reliability analysis and developed a novel multicentre evaluation framework.

Main Results:

• The MultiResUNet deep learning model achieved a mean Dice similarity coefficient >98%.
• The top-performing model demonstrated robustness across different pathologies and spatial locations.
• The evaluation framework successfully assessed accuracy, generalization, robustness, and reliability.

Conclusions:

• FLAIR-based ICV segmentation pipelines are effective for large-scale neurodegenerative disease research.
• The MultiResUNet model shows significant promise for automated brain extraction in diverse clinical FLAIR MRI datasets.
• The presented evaluation framework can guide the assessment of automated MRI analysis tools for clinical use.