Magnetic Resonance Imaging is an important tool for characterizing volumetric changes of the piglet brain during development. These analyses have been aided by the development of piglet brain atlases which are based on averages drawn from multiple piglets. Because these atlases typically contain only brain tissue, their use is limited to “brain extracted” images from which the surrounding tissues have been removed. Brain extractions, or segmentations, are typically performed manually. This approach is time-intensive and can lead to variation between segmentations when multiple raters are used. Automated segmentations processes are important for reducing the time required for analyses and improving the uniformity of the segmentations. Here we demonstrate the use of region-based recurrent convolutional neural networks (RCNNs) on a dataset consisting of 32 piglet brains. The RCNNs are trained from manual segmentations of sets of 27 piglets and then applied to sets of the remaining 5 piglets. The volumes of the machine-generated brain masks are highly correlated with those of the manually generated masks, and visual inspection of the segmentations show acceptable accuracy. These results demonstrate that neural networks provide a viable tool for the segmentation of piglet brains.
Multiclass Brain Tissue Segmentation in 4D CT Using Convolutional Neural Networks