A recent thread on this newsgroup has mentioned a common problem---no
automated edge detection methods work well in medical images. At Colorado
State, we are developing a semi-automated method in which a person traces a
small number of pixels along the boundary of a region of interest. These
pixels are used to train a neural network to duplicate the human's decisions,
then the neural network carries the trace forward until its confidence falls
below a threshold. The human can then do more tracing by hand to train the
network further. The human can also correct errors in the automated part of
the trace by hand-tracing more pixels.
We are just beginning to test this technique on images from the Visible Human
Project and on MRI head scans. We are very interested in testing our
technique on other images. If you have sets of images that contain some
hand-traced regions of interests, would you be interested in sharing the
images with us? We would keep you posted on the results. The ideal image set
is one containing multiple slices, each with the cross section of a volume of
interest hand-traced.
Preliminary results can be found at
http://www.cs.colostate.edu/~anderson#CG
----------------------------------------------------------------------
Chuck Anderson
Department of Computer Science anderson at cs.colostate.edu
Colorado State University http://www.cs.colostate.edu/~anderson
Fort Collins, CO 80523-1873 office: 970-491-7491, FAX: 970-491-2466