Full-head segmentation of MRI with abnormal brain anatomy: model and data release

Area of Science:

• Neuroimaging and Medical Image Analysis
• Deep Learning and Artificial Intelligence
• Computational Neuroscience

Background:

• Accurate segmentation of whole-head magnetic resonance imaging (MRI) is crucial for various neuroimaging applications.
• Existing methods often struggle with abnormal anatomy and require atlas-based registration, limiting their robustness.
• There is a need for advanced segmentation tools and comprehensive datasets to address these challenges.

Purpose of the Study:

• To develop a deep learning network for whole-head segmentation of clinical MRI, including cases with abnormal anatomy.
• To create the first public benchmark dataset for whole-head MRI segmentation, comprising 98 MRIs with detailed volumetric labels.
• To evaluate the network's performance against existing tools and assess its utility in downstream applications.

Main Methods:

• A novel
• MultiAxial
• deep learning network was developed, utilizing three 2D U-Nets operating on sagittal, axial, and coronal planes.
• The network was trained and validated on a custom dataset of 98 MRIs with manual corrections for structures like skin, skull, CSF, gray matter, and white matter.
• The approach avoids atlas coregistration, enhancing robustness, particularly in regions with abnormal anatomy.

Main Results:

• The MultiAxial network achieved a high test-set Dice score of 0.88 ± 0.04 for whole-head segmentation, outperforming established tools like Multipriors (0.86 ± 0.04) and SPM12 (0.79 ± 0.10).
• The network demonstrated robustness on images with abnormal anatomy and de-identified scans.
• Improved segmentation accuracy facilitated more robust current flow modeling within the ROAST toolbox for transcranial electric stimulation.

Conclusions:

• The study presents a state-of-the-art deep learning tool for whole-head MRI segmentation, particularly effective in cases with abnormal anatomy.
• The release of the largest volume of labeled clinical head MRIs, including non-brain structures, establishes a new benchmark for the field.
• The developed model and dataset are expected to advance research in neuroimaging analysis and computational modeling.