We are attempting to segment medical images of various types, i.e. to divide the images into anatomically significant regions.
The images are characterised by:
The methods we are deriving are demonstrated by segmenting retinal images - useful in the diagnosis and treatment of glaucoma, and hand X-rays - useful in the diagnosis and treatment of short stature in children.
Glaucoma is the major cause of loss of sight in the 1 % of the population who are registered blind. This project is concerned with producing software tools to aid in the diagnosis and assessment of the disease. These are tasks which are presently done by an expert viewing images of the optic nerve head, such as these healthy and diseased eyes. (The images are negatives.)
Click an image to enlarge
The images show the region of the retina where nerve fibres and blood vessels pass through the eye, i.e. the blind spot. In a normal, healthy eye, the nerve head region is approximately circular. In an eye effected by glaucoma, nerve damage results in a change to a more irregular shape.
Identifying the structures is difficult as they are indistinct, of lower contrast than and partially obscured by blood vessels.
Our initial work has resulted in a semi automatic method of identifying the nerve head boundaries. It works by asking the user to indicate the approximate location of the boundary in a median filtered version of the image. The software moves the user's points to a consistent location on the boundary. It also inserts additional points by interpolation. The boundary between the points is tracked. The work was presented at the American Academy of Optometrists Biennial Meeting in Europe in 1994 and at the 2nd Internet World Congress on Biomedical Science.
A major disadvantage of this approach is its sensitivity to the user's initial selection of points. The intra observer variability in the latest version is in the order of 4 % (in a normal eye, the variability is typically 8 % and in a diseased eye 4 %). This may be reduced to zero by automating the boundary idetification.
We have described in a recent paper how a dynamic contour (snake) may be shrunk onto the boundaries in the optic nerve head. The dynamic contour requires that the image is enhanced to increase the contrast between the nerve head and the retinal regions. The snake can then shrink reliably onto the boundary.
We are currently working on multiresolution methods incorporating the wavelet transform to enhance the images. More examples will follow in due course.
This work has been partially funded by the International Glaucoma Association.
Last modified
March 2008
Click to enlarge