Biomedical Engineering Reference
In-Depth Information
Uncertainty in the delineation of the target volume
The delineation of the tumor is subject to uncertainty and can be quite
non-reproducible as was demonstrated in Figure 3.19 above. This
non-reproducibility troubles many people - and one cannot say that it
is desirable. However, in this connection, I want to make two points:
First, much of the variability seen in Figure 3.19 is, I believe, less a
matter of inconsistency than a consequence of genuine uncertainties
as to precisely where gross and subclinical disease is present. The
fault, one might say, is in asking clinicians to draw a single sharp line.
We would do much better to express our uncertainties explicitly.
Figure 3.20. Three approaches to delineating a target volume:
(a) traditional; (b) as a fuzzy boundary; and (c) as a most likely surface,
with a confidence band around it (in 3D).
Figure 3.20 suggests a couple of ways to make the uncertainties
in delineation both explicit and quantitative, as was also proposed by
by Waschek
et al.
(1997). Panel (a) is the traditional single line
delineation. In panel (b), a spray-can like tool has been used to ex-
press the delineator's guess at the probability that the surface of the
volume passes through at a certain point (the probability density
function). In panel (c), a triplet of contours has been drawn defining:
the most likely outline; the outline outside of which the therapist is
confident at a defined confidence level that there is no disease; and an
outline within which he or she is confident, at the same level of
confidence, that there is disease. Often the margin will not be the
same all around the tumor, and may well be different for the upper
and lower confidence contours. If we were to delineate out target
volumes in such a manner, I would be willing to wager that there
would be much more overlap between observers and for repeat
delineations by a single observer than is suggested by Figure 3.19.
The second point relates to the fact that many people respond to
the demonstrated uncertainties in target volume delineation by
