Biomedical Engineering Reference
In-Depth Information
which scans an object rotationally. There were 13-17 frames per cardiac cycle,
and the angular slice spacing was 3 degrees, resulting in 60 images slices in each
frame. Therefore, there were about a thousand images in each 3D sequence. The
resolution of images was 240 × 256.
Obviously, it is too tedious to segment all the images manually using the
traditional method. Our method deals with all the images without too much manual
intervention. Neither it is sensitive to the initial zero level set nor the prior region,
as long as it is just between the valve and the cardiac wall. Given several prior
regions, our method may segment the whole image sequence automatically without
too much parameter choosing or adjusting. The images from the same scanning
time can share the same prior region, as do images from neighboring times. Without
too many manual processes, segmentation can be as efficient as possible and the
segmenting result precise. Some results obtained with the region prior-based
geodesic snake are shown below in Figure 22. The images in there are at the
same scanning position as a different scanning time from a 3D valve sequence. To
facilitate the display, we cut out the valve region.
Most of the above segmentation results may satisfy the needs of 3D recon-
struction and diagnosis. Because of contamination of noise and movement blur,
there will be inevitable errors in some contours, such as the 8th, 12th, and 13th
contours in Figure 23. We can segment this kind of noise-disturbed image guided
by a prior shape. The prior shape can either come from an output of a neighboring
slice or from a manual outline provided by an expert physician. In the manual
outlining process, it is unnecessary to draw out the whole shape: a part of the
stained edge is just enough. We depict some results using the shape prior-based
geodesic snake in Figure 23. We can alleviate almost all segmentation errors under
the guidance of a shape prior.
6.5. Conclusions
The inherent noise, blur, and large quantity of data in echocardiographic se-
quences make it difficult to segment valve structures, which hinders clinical ap-
plication. We present a new algorithm to incorporate prior knowledge into the
geodesic snake. The priors are expressed as a speed field that directly draws the
zero level set to the ideal contour. The region prior limits the zero level set evolving
within a certain region and the shape prior draws the curve to the ideal contour. An
actual application on 3D echocardiographic sequences demonstrates that the algo-
rithm segments the valve structure accurately and reduces the need for a manual
procedure, resulting in a greatly accelerated process.
Prior-based image segmentation is the subject of active interest at present. We
can express more prior knowledge as a speed field and embed them in image the
segmentation process in future work. Guided by prior information, we can make
image segmentation more accurate and efficient.
Search WWH ::




Custom Search