ICV

ICV segments intracranial volumes (ICV) and ventricles from single- and multi-contrast MRI using convolutional neural networks (CNNs) to enable quantitative analysis of structural changes in populations with cerebrovascular lesions, white matter hyperintensity, and atrophy.

Key Features:

• iCVMapper and VentMapper: Separate CNN-based models tailored for total intracranial vault segmentation (iCVMapper) and ventricular segmentation (VentMapper), optimized for data from older adults with cerebrovascular lesions and atrophy.
• Single- and Multi-contrast MRI Support: Models process both single-contrast and multi-contrast MRI datasets.
• Extensive Training Dataset: Trained on multi-site datasets comprising 528 subjects for ICV segmentation and 501 subjects for ventricular segmentation.
• High Accuracy and Comparative Evaluation: Multi-contrast models achieved average Dice coefficients of 0.98 for ICV and 0.94 for ventricles and were compared against MONSTR, BET, Deep Extraction, and FreeSurfer.
• Time Efficiency: Processing times are reported to be orders of magnitude faster than many existing methods.
• Post-processing Enhancement: A conditional random field (CRF) is applied as a post-processing step to improve segmentation accuracy.
• Robustness to Image Quality Variations: Validated robustness to lower resolution and reduced signal-to-noise ratio, outperforming other tested techniques under such conditions.

Scientific Applications:

• Neurodegenerative research: Precise ICV and ventricular segmentations to study structural brain changes associated with neurodegeneration.
• Normal aging studies: Quantification of intracranial and ventricular changes across aging cohorts.
• Cerebrovascular-related cognitive impairment: Segmentation in cohorts with cerebrovascular lesions and high white matter hyperintensity burden to support studies of vascular contributions to cognitive impairment.
• Multi-site population analyses: Application in large multi-site studies to enable comparative analyses across diverse imaging protocols and populations.

Methodology:

Convolutional neural networks were trained on multi-site MRI datasets (528 subjects for ICV, 501 for ventricles), evaluated using Dice coefficients and comparisons to MONSTR, BET, Deep Extraction, and FreeSurfer, tested on independent cohorts including subjects with vascular cognitive impairment and high white matter hyperintensity burden, and refined with a conditional random field (CRF) post-processing step.