Automatic Segmentation of the Scoliotic Spine from Mr Images

Segmenting vertebral bodies (VBs) and intervertebral discs (IVDs) in magnetic resonance imaging (MRI) data is an important step towards the creation of 3D spine models for image-guided surgical treatement of adolescent idiopathic scoliosis (AIS). Recent advances in deep learning have established state-of-the-art results in medical image segmentation. Thus, in this paper, we present a method based on convolutional neural networks to simultaneously segment VBs and IVDs in MRI data sets of AIS patients, a difficult problem which has not yet been adressed in the literature. Our architecure is inspired by the U-net architecture, combined with the recently proposed and promising squeeze-and-excitation (SE) block and an objective function incorporating Cohen's kappa to deal with the imbalamce class problem. Our model is first trained using a public dataset of non-AIS patients, then fine tuned using a few images of AIS patients. Results in 8 test AIS patient MRI volumes show that the fine tuning and SE strategies both improve segmentation considerably while remaining highly complementary to each other.