by Douglas Page © 1995
An unlikely digital courtship is developing between the far-sighted and the near-sighted. Researchers are attempting to
unite astronomical image processing techniques with digital mammography in an effort designed to help physicians get a closer
look at digital mammograms.
None of the researches will say this out loud, but the revolutionary marriage, if the engagement survives experimentation,
could eventually have enormous significance to almost all women: evidence of breast cancer could be detected sooner, in a
younger population, and treatment could therefore be less invasive.
A preliminary study funded by the National Science Foundation is underway at Johns Hopkins University, Georgetown University's
Lombardi Cancer Research Center and the Space Telescope Science Institute to investigate whether modern astronomical image
processing software can be useful when applied to the detection of cancer signs in digitized mammograms. Initial results using
the astronomical viewing tools - some of which were developed to correct the previously aberrant optics on the Hubble Space
Telescope - are promising.
The research team is careful not to raise expectations prematurely; no wedding date has been set. "It's getting a lot of
attention," said computer scientist Steven Salzberg, Johns Hopkins University, the project's principal investigator, "but
we really have just extremely preliminary results which look promising and that's really all we have right now."
The prospects are nevertheless exciting.
The collaboration happened serendipitously. Dr. Matthew Freedman, associate professor of radiology at Georgetown University,
a pioneer in the relatively new field of digital mammography, happened to see a digital image processing presentation given
by the Hubble Space Telescope people. Hubble astronomers use digital imaging all the time. Traditional photographic methods
quickly become obsolete when the cameras are in orbit. Who's going to change the film, for instance?
"I said, 'Gee, those images before they cleaned them up looked like our images'", Freedman said, "So I suggested to them
there might be benefit to us and to them if they tried their software on our digital mammography images." The result was that
a research team composed of astronomers and computer programmers was formed, all of whom are now looking at ways to help radiologists
get a sharper focus on digital mammograms.
The effort got a boost from Ben Snavely, program director for advanced technologies and instrumentation in the NSF division
of astronomical sciences, which awarded the study a $50,000 start-up grant from the Small Grants for Exploration Research
The problems of detection encountered by astronomers searching digital sky images and physicians examining digital mammograms
are similar. Both must locate extremely faint objects against gray, grainy, often blurred backgrounds. Astronomers are looking
for traces of cosmic ray hits, which mask the target stars and galaxies. Doctors are looking for microcalcifications, the
indicators of breast cancer.
"The thing that got us interested in this," said astronomer Bob Hanisch, one of the collaborators at the Space Telescope
Science Institute in Baltimore, "is that detecting the microcalcifications is similar to a problem in astronomical imaging
with CCD detectors, which is detection of cosmic rays hits."
Cosmic rays are elementary particles that dart around space at nearly the speed of light and basically ruin all the images
the Hubble collects. Every time a ray hits a CCD detector it leaves a trace in the form of a speck on the image. Terrestrial
telescopes are less likely to be effected by cosmic rays since the atmosphere does a good job of absorbing most of them. Orbiting
telescopes are unprotected.
"You get literally thousands of them that may occur on a single image, which may also have thousands of stars in it," Salzberg
said. "Even if you could discriminate between them with your eyes - which isn't so easy - it's extremely tedious. So, we worked
on developing automatic ways of doing that [for the astronomers]."
Astronomic software has become quite efficient at this technique. The digital sky images are first fed through software
filters, then are passed along to be sharpened and enhanced. Other tools are then used to actually classify those objects
It all happens automatically.
The object of the new NSF study is to see if these same tools used by the astronomers to automate the isolation of cosmic
ray specks might be likewise useful in automatically isolating microcalcifications. It's not quite as simple as it sounds.
Astronomers want the cosmic ray specks removed from their images.
Radiologists want the opposite. Radiologists want everything but microcalcifications removed from theirs.
Freedman supplies the digital mammogram data from his research operation at Georgetown's Lombardi Cancer Research Center.
He has one of the few experimental digital mammography operations in the country. The digital radiology hardware at Georgetown
is a Fuji 9000 Computed Radiography machine, with software modified by Fuji to Georgetown's specifications. "The operation
of the machine has been mathematically optimized for this purpose by us," Freedman said. "It is a Fuji machine used the way
Georgetown uses it and has modified it. [It] can obtain digital radiology of chest, bone, abdomen, kidney, skulls, etc. We
use this system for over 80 percent of our radiographic work."
Though digital mammography has not yet received FDA approval for clinical usage it seems likely it will in the not-too-distant
future. Digital mammography is currently being performed routinely in Japan, England and Denmark. The approval process, however,
in the US is lengthy and cumbersome.
"You run into a problem that first you have to prove its effectiveness [in trials] and then you have to convince the FDA
of its effectiveness," Freedman said, "and those are two different things. I think the FDA is setting very high standards,
which I think they should, but its very hard to predict at this point what they will really require in this regard, how many
cases they will want, what level of proof they will want. At the moment I don't think we've proved that digital mammography
should be a clinical tool, but that's what we're working toward."
Nevertheless, the NSF astronomical study is moving forward, in anticipation of eventual FDA approval. "The Space Telescope
software will likely have no effect positive or negative on the approval process," Freedman said, "in that I expect that approval
will be sought prior to this work being completed. The Space Telescope work will improve the quality of the system, but probably
after approval has been gained without it."
Three Step Process
Freedman's digital data is manipulated in a three step process at the Space Telescope Science Institute.
First, the digital data is fed through a computer process called 'unsharp masking', the result of which is a small-feature
enhanced version of the image. This is accomplished by smoothing the original and then subtracting the smoothed image from
the original image, which removes large-scale structures, such as blood vessels and tissue fibers. What remains is a jumble
of small-scale structures, some of which are microcalcifications.
"We have to find a way to distinguish between the really fine-scale structures of the microcalcifications and the background,"
explained STScI's Hanisch. "So the next thing we do is 'variance normalization', where what we want to do is be able to detect
small features at the same level of statistical significance regardless of whether they occur in a bright or a dark region
of the mammogram. You want the same statistical confidence level of detection throughout the picture. Our 'variance normalization'
involves dividing by that smoothed image we created in the first step.
"Now you have an unbiased picture that is just measurements in terms of local standard deviations. But it's still full
of all sorts of structures that are not really relevant," Hanisch continued. "So, in the third step, we're [using] an 'adaptive
filter', which allows the user to set a detection threshold, again in terms of standard deviation. Any area of the image where
structures exceed that noise threshold you leave as is.
In regions where the signal does not exceed that level you smooth it, and you smooth it on various spatial scales: if there
is some indication of structure you do a little smoothing, if there's no indication of structure you do a lot of smoothing."
Adaptive filtering leaves those events above the noise threshold. At this point what's been produced is an image where
most of the confusing background structures have been suppressed and what remains are the indications of microcalcifications.
"Basically, we've prepared the image in a way that makes it easier to see the events that might be of diagnostic interest,"
After the image has been sharpened in the preceding steps it goes to Steven Salzberg's lab at Johns Hopkins, where they
try the tricky process of attempting to classify what remains in the image. The STScI astronomic software is already largely
developed and in standard usage. Salzberg's link is the piece missing in the sequence. A significant amount of work remains
to be done in this area.
"I come in at the end of the process, after the potential microcalcifications have been highlighted. The final step is
to decide whether or not any of those spots, little clusters of spots, are anything to worry about, whether it indicates cancer.
That's image classification and that's where I would step in and measure things about the objects," Salzberg explained.
"This is what human experts do already, only they do it without all the automated help. What I've seen of what the astronomers
can do is quite impressive. The human eye is not able to detect all the nuances in a digital image. The reason people are
going to digital mammography and the reason astronomers use digital images is that they have tremendous dynamic range, they
detect extremely faint things in the same image where they're detecting quite bright things. They are in fact more sensitive
than the human eye."
Salzberg's expertise is in an area of computer science called 'machine learning'. "The idea is to develop computer programs
that can learn from data or data bases, meaning you program them to do something but what you really program them to do is
to get better at what they're doing."
One of the techniques Salzberg uses is 'decision trees'. He explained it this way.
"The idea would be for the astronomic work that we've been doing, as you look at a lot of images you train the program
by giving it images where you know what the objects are. For astronomy, you know you're looking at stars or galaxies or cosmic
rays. And you give it hundreds or thousands of these and [using decision trees] the program eventually figures out its own
set of rules to distinguish between them."
Salzberg's goal is to modify the cosmic ray-classification process to allow automated detection and classification of microcalcifications
in the digital mammograms. He has employed his cosmic ray-classification process with success for the astronomers now for
Microcalcifications as small as 100 microns are expected to be identifiable with these new methods. Radiologist Freedman
said, "What happens now [without the benefit of astronomical image processing] is that we can see them but we can't distinguish
whether or not they are noise [or disease]. So it's not that we will see something new, it's that we'll see it with greater
certainty that it is disease rather than noise."
Freedman believes that mammograms done in the new way will eventually be shown to be more effective than current methods
in women under the age of 40. "I think this will result in new population studies that will show that screening mammography
should start at age 35, and in high risk patients at an even younger age.
There is much study that will need to be done to prove that, but that is what I think the eventual result will be. This
proof will take about five years. In the meantime, no woman who should have a mammogram under current guidelines should avoid
having a mammogram done in the standard conventional method just because there may be a better method in the future."
Such clarity in the future conceivably could enable physicians to spot the presence of smaller and smaller signs of cancer,
yielding critical early detection - which, for everyone involved, especially women, could end up being the most stellar performance
Article appeared in the December, 1995, issue
of Digital Imaging Magazine.