Graphics Reference
In-Depth Information
10
Globally Optimal Energy Minimization
Techniques
10.1 Introduction and Timeline
In 1992, we began a research project to automatically segment cell images
from Pap smear slides for the detection of cancer of the cervix. We investigated
simple low-level techniques based on edge detection, grayscale thresholding,
and grayscale morphology (e.g., watersheds), but could only achieve accurate
segmentation on about 60% of cell images (Figure 10.1). In 1997, we started
looking at dual active contour segmentation techniques as proposed by Gunn
and Nixon [72], but this method suffered from poor robustness on our images.
However, Gunn [70] also suggested a fast globally optimal method based on
converting the problem of finding the best circular contour into a linear trellis
and then applying the Viterbi algorithm to determine the minimum energy
path. This approach worked remarkably well, as reported by Bamford and
Lovell [13] in 1998, and yielded 99.5% correct segmentation on a cell database
of nearly 20,000 cell images.
(a)
(b)
(c)
Fig. 10.1.
Traditional bottom-up approach to cell image segmentation. (a) Original
graylevel image, (b) thresholded image showing voids and artifacts, and (c) Canny
edge map showing a partially complete border and other spurious edges (from [11]).
Search WWH ::
Custom Search