Youtube Kanal CG
Sie sind hier
Spine segmentation and analysis
One of the routine examination procedures for patients with back pain is magnetic resonance imaging (MRI). MRI uses no ionizing radiation, unlike X-ray and computed tomography (CT), so patients can be scanned with no time constraints regarding exposure.
In spite constant improvements of scanning devices, routinely acquired images are not of very high quality. This especially concerns spatial resolution in Z axis. Reason for this is that doctors still heavily rely on diagnosis from 2D slices - hence, the inter-slice distance is not of such a great concern to them.
There are many methods to segment the spine, but some do not result in explicit surface geometry (which makes subsequent symbolic analysis hard), and most are focused on CT images.
There are several challenges in front of anyone who wants to do segmentation of the spine.
- Spine is anatomically complex - 33 vertebrae, 23 intervertebral disks, spinal cord, branching nerve roots, connecting ribs, blood vessels etc. Anatomically correct model, besides being complex and hard to create, would be computationally very demanding - possibly even unfeasible.
- MR images have good visibility of soft tissues (tissues with higher water content), but bones are sometimes captured with intensities similar to air.
- As all real-life data, MR images contain noise. Algorithms must have tolerance for noise, because noise can only be reduced, not removed.
One benefit of spine's complexity is that its structure is subject to many constraints. Not all of these constraints can be employed by segmentation algorithms, but using even a few constraints could have a significant impact on the algorithm's performance and robustness.
In order to keep number of shape parameters low, the surface of vertebral bodies does not have to be freely deformable (to best fit the data), but can be a smooth subdivision surface. Only control points of this surface will have to be considered during optimization of model fitting (fitting prior shape to given image data).
The ultimate long-term goal is to have an abstract spine model, which can then be analyzed to detect certain types of abnormalities, such as spondylolisthesis (misalignment of vertebrae), scoliosis (abnormal curvature of the spine) and vertebral collapse (crushed vertebral body).
We publicly release our source code, datasets and their segmentations in support of paper published in Computer Graphics Forum. Please cite this paper if you use any of these in your work. In support of my thesis I publicly release newer version of the source code.