A deep learning framework for brain extraction in humans and animals with traumatic brain injury

Automatic brain extraction or skull stripping from magnetic resonance images (MRI) is an important pre-processing step in many image processing pipelines. Most skull stripping methods are optimized for normal brains and applicable to single T
₁
-w MR images. However, other contrasts, such as T
₂
, can provide complementary information about the boundary. This is especially true in the presence of traumatic brain injury (TBI) and other diseases, where lesions can confound boundary definitions. In this paper, we propose a deep learning based framework to extract intracranial tissues from multi-contrast MR images in the presence of TBI. Our approach is based on state-of-the-art convolutional neural network architecture to learn a transformation from multi-contrast atlas MR images to their stripping masks without using any deformable registration. An advantage of our framework is that it can be applied to different species. We applied our approach to 19 human patients with mild to severe TBI, as well as 16 normal mice images, and another 10 mice brains with TBI. We compared the approach with 3 separate state-of-the-art human and rodent brain extraction methods. Using only a few manually delineated atlases, we showed significant improvement in brain extraction accuracy in both healthy and pathological human and rodent images.