Full-Head Segmentation of MRI with Abnormal Brain Anatomy: Model and Data Release

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neuroimaging and Computational Anatomy

Background:

• Accurate whole-head segmentation is vital for clinical analysis and neuroscientific research.
• Existing segmentation tools struggle with abnormal anatomy and require atlas co-registration.

Purpose of the Study:

• To develop a deep learning network for robust whole-head MRI segmentation, including abnormal anatomy.
• To create the first public benchmark dataset for whole-head MRI segmentation.

Main Methods:

• A novel "MultiAxial" deep network was developed, combining three 2D U-Nets operating on different planes.
• A dataset of 98 MRIs with manual volumetric segmentation labels was compiled, including normal and abnormal anatomy cases.
• The network was trained and validated on diverse clinical MRIs, including those with stroke and disorders of consciousness.

Main Results:

• The MultiAxial network achieved a median Dice score of 0.88±0.04 for whole-head segmentation, outperforming standard tools like SPM12.
• The network demonstrated robustness in segmenting regions with abnormal anatomy and on de-identified images.
• Improved segmentation accuracy enhances current flow modeling in transcranial electric stimulation applications.

Conclusions:

• A state-of-the-art deep learning tool for whole-head MRI segmentation, particularly in cases of abnormal anatomy, has been developed.
• The largest publicly available dataset of labeled clinical head MRIs, including non-brain structures, is released.
• This model and dataset serve as a benchmark for future advancements in neuroimaging segmentation.