Our paper examined how to segment the spinal cord. We took a machine learning approach, specifically we used auto-context, to do a voxel-wise labeling (i.e. label the voxel as spinal cord or background) and used this labeling as our segmentation.
Our main contribution is that in each stage of auto-context, we considered the connected components with the same label as a single candidate shape. Then, given this shape, we can compute features from the shape. These features we then use as input to the classifier in the next stage of auto-context.
We compared this approach to a more analytical approach we tried earlier and found that our new machine learning based approach gave superior results.
Our paper entitled, Augmenting auto-context with global geometric features for spinal cord segmentation, can be read here.
I did a video talk on our approaches to do spinal cord segmentation which I linked to in this post here.
If for some reason you would like the hi-res PDF version of the poster, feel free to email me or leave a comment below and I’ll post it 🙂