I have developed a new algorithm for fitting templates to images by "elastic" deformation. By precomputing orthogonal curves to the template, the algorithm guarantees that the fit is globally optimal over the entire region of deformation. Most deformable contour algorithms give only a local optimum.
For details, see reference [1].
An early version of the algorithm was used to reconstruct wrist bones in 3-D from CT images (collaboration with Prof. Gil Hillman of UTMB). Reference [3] has details.
A typical 3-D reconstruction looks like this:
In this image, the thumb is located at the top left.
A template and its precomputed orthogonal
curves:
The initial placement of the template on the image, and the final fit of the template are shown below. The image is a CT of a human wrist.
Request references by email.
Accepted for publication by I.E.E.E. Trans. on Medical Imaging.
An unprocessed image is just an array of intensities or gray levels. To process the image and get quantitative information from it, we usually have to outline the objects visible in the image.
Manual outlining can become very tedious in large studies involving hundreds of images. It is more convenient to develop techniques for semi-automatic outlining.
One possibility is to design templates which the user places over a region of interest and which the computer automatically deforms to fit the object. This is called deformable template matching .
If the deformation algorithm is not well designed it can be sensitive to noise and other artifacts in the image. On the other hand, if the algorithm is complicated, it can run very slowly. The key is to develop an algorithm that is reasonably fast and guarantees optimal behavior. This research led to one such algorithm.
Hemant's Home Page.
Hemant's Research Page.