快猫短视频

鈥淲HENEVER there鈥檚 another report of smear tests going wrong, I always hold my
breath until I hear that my local hospital isn鈥檛 affected,鈥� said a woman last
month on BBC radio. She was reacting to Britain鈥檚 worst cervical cancer
screening scandal. Over five years, staff at the Kent and Canterbury Hospital in
southeast England missed abnormalities in smear tests and mistakenly gave many
women the all-clear. Of these women, eight have since died and 30 needed
hysterectomies to halt the progress of cervical cancer. In all, 91 000 smears
had to be rechecked.

An inquiry by the hospital and local health authority blamed the tragedy on
inadequate training for cytotechnologists鈥攖he people who assess the
smears鈥攁nd, ultimately, a breakdown in the programme鈥檚 management. It was
an extreme event, but it reinforces a more mundane lesson: even in the
best-managed laboratory with well-trained staff, smear testing will always be
fallible. Judging whether a smear is positive or negative is a subjective test
and some abnormalities will inevitably be missed.

Against this background, the National Health Service has been funding a
trial鈥攄ue to end next month鈥攖hat specialists hope will reduce the
number of women who fall through the screening net. The trial is based at five
British hospitals with a group at St Mary鈥檚 Hospital in London taking the lead.
The researchers are testing PAPNET, which uses neural networks to spot
potentially abnormal smears. The system was approved two years ago in the US for
secondary screening鈥攃hecking the assessments made by cytotechnologists.
But in the British study, PAPNET is being studied as a tool to help
cytotechnologists in their initial assessment. 鈥淭here is a shortage of trained
cytotechnologists in the UK,鈥� says Dulcie Coleman, professor of cytopathology
and cytogenetics at St Mary鈥檚. 鈥淚f PAPNET is reliable, we could make much better
use of our available skills.鈥�

Screening for cancer of the cervix, the neck of the womb, is possible because
cervical cells pass through a series of detectable changes before they become
cancerous. To screen for these changes, a doctor takes a sample of cells from
the cervix and smears them across a glass slide. The cytotechnologist examines
those cells through a microscope for abnormalities. Treating the condition at an
early stage is relatively straightforward.

Success story

Despite all the scare stories, smear testing is one of medicine鈥檚 success
stories. According to the US National Cancer Institute in Bethesda, Maryland,
between 1950 and 1970 deaths from cervical cancer fell by 70 per cent. Last year
in Britain, cervical cancer killed nearly 1400 women. Many others were saved by
cytotechnologists, who found 20 000 positive slides among the 4 million slides
they examined. Some of the women from whom these smears were taken would have
developed cancer if not for early intervention.

Screening may sound simple, but specialists say it is rather like searching
for a needle in a haystack. There can be 300 000 overlapping cells on one slide,
and fewer than fifty may be precancerous. These cells may be scattered across
the slide, and while they tend to smaller than usual, with larger, darkish
nuclei, the changes are subtle.

To compound the problem, cytotechnologists can become habituated to finding
negative results, because they know that more than 90 per cent of slides they
see will have no abnormalities. A screener might review 50 slides a day in
Britain, or up to 100 a day in the US. And for each slide, there are 1000 fields
of view, so a cytotechnologist in the US might manipulate the microscope 100 000
times per day. All these difficulties combine to thwart even the most vigilant
screener. And because standards and practices can vary widely from one hospital
to another, so too can the screening results. Estimates from medical studies
suggest that the false negative rate鈥攖he percentage of women with abnormal
smears who are given the all-clear鈥攔anges from 10 to 40 per cent.

PAPNET鈥檚 role is to reduce these figures. The man behind the machine is Mark
Rutenberg, the founder of Neuromedical Systems in Suffern, New York. The seeds
of his idea were sown back in the 1970s while he was working as a biomedical
engineer at Case Western Reserve University in Ohio. NASA had developed a
digital image processor to analyse pictures returned from Mars by the Mariner
spacecraft. Keen to justify its budget, the agency asked hospitals if the
processor would be of benefit to them. Rutenberg looked at the technology鈥檚
potential for automating smears. There was none.

Rutenberg had come up against what some call the 鈥渃ombinatorial explosion鈥�.
Put simply, this means that complex images, such as cervical cells on a glass
slide, cannot be described according to the rules (or algorithms) that computers
understand. Normal cervical cells are not uniform in shape or size. Abnormal
cells also vary in many ways, and may be isolated or in small clumps. Then there
might be an artefact, such as an air bubble, on the slide. An algorithm-based
computer, such as the PC on your desk, would need a set of rules to deal with
every eventuality. Rutenberg realised that he faced an impossible programming
task, so he shelved the idea.

There the matter rested until a series of reports in The Wall Street
Journal by Walt Bogdanich (for which he won a Pulitzer prize in 1988)
highlighted poor testing in American medical laboratories. Among the
mistakes鈥攎ade under financial pressures鈥攚ere faulty interpretations
of cervical smears. By this time, Rutenberg had moved into military
research.

A shower of dummies

His job, part of the Strategic Defense Initiative, was to identify live
warheads among a shower of dummy warheads. If the US came under nuclear attack,
enemy missiles would deploy the dummies to act as decoys. Distinguishing between
the two relied on spotting subtle differences in their trajectories, detected by
high-resolution radars or infrared sensors. Since speed was clearly essential,
Rutenberg鈥檚 group enlisted the help of a neural network.

The beauty of a neural net is that it does not need an algorithm to cope with
every eventuality. Rather, it 鈥渓earns鈥� just as a person does鈥攆rom
examples. Train a neural net to identify warheads by their trajectories, for
example, then present it with a trajectory it hasn鈥檛 seen before and it will
draw on its experience to decide which type of warhead it鈥檚 likely to be.

Rutenberg tells that he was in the shower when the idea came to him to target
neural nets, not on warheads, but cell cytology in general and cervical smears
in the immediate future. True to the 1980s techno-hero myth, he quit his job,
took out a patent and retired to his garage. He founded his company in 1987,
scrounged software and hardware, built a prototype in 1988 and touted it around
in search of backing鈥攚hich he received.

The result is PAPNET, which recognises potentially abnormal cells on a
microscope slide and shows a cytologist where to find them. The machine scans
the slide first at 50X magnification, optimising the focus, defining hundreds of
small fields of view that contain cells and excluding areas with surface
artefacts, such as air bubbles.

Trained detectors

During a second scan, at 200X magnification, an algorithm-based image
processor examines every field in search of potentially abnormal objects. The
algorithm defines an abnormal object on the basis of size, shape and opaqueness.
On average, this identifies between 20 000 and 80 000 potentially abnormal
objects, which are then analysed by the two neural networks.

As in other neural computers, each of PAPNET鈥檚 networks is made up of a web
of processors that are roughly analogous to neurons in the brain. The processors
are arranged in several layers, and the output from a processor is connected to
the input of every processor in the succeeding layer. Each connection has a
weighting, which attenuates the signal it carries, and each processor fires off
an output only when the sum of its inputs reaches a predefined threshold
(see Diagram).FIG-21225001.jpg

PAPNET鈥檚 trainers worked with thousands of images of cervical smears,
selected to represent a wide variety of normal and abnormal cells. Through a
long process of trial and error, they tweaked the connection weightings for each
processor until the nets learnt to give a high output when fed any positive
smear and a low output for any negative image. The networks鈥� 鈥渆xpertise鈥� resides
in the final pattern of weightings. It took about two years to train the
networks, according to Jim Herriman of Neuromedical, a process which he
describes as more art than science.

One of PAPNET鈥檚 neural nets screens for abnormal cell clusters while the
other looks for single abnormal cells. The machine assigns each field of view a
score depending on how close the networks鈥� outputs are to their maximum or
minimum. During a third and final scan at 400X magnification, the 128 fields
with the highest scores are captured as high-resolution digital images, along
with their locations on the slide.

Even with its array of fancy technology, PAPNET alone cannot judge whether a
smear is positive or negative. This decision is still made by a human. A
cytotechnologist examines images of the 128 fields of view on a screen. Many of
these will be negative because although PAPNET pulls out the highest-scoring
fields, the actual scores may be close to zero. If the cytotechnologist is
concerned with a field of view, he or she can re-examine the original slide.

In the US, the Food and Drug Administration (FDA) approved PAPNET for
secondary screening in November 1995. As a complement to manual screening, it
has proved a valuable tool. A host of trials around the world show that PAPNET
can reduce the false negative rate by between 5 and 50 per cent. The vast
variation can be attributed to differences in the original manual testing and to
differences in doctors鈥� skills when preparing slides, according to Laurie Mango
of Neuromedical Systems.

In one of the most recent studies, conducted last year, researchers led by
Leopold Koss from the Montefiore Medical Center in New York took 487 negative
slides from an archive and fed them to PAPNET (Human Pathology, vol 28,
p 1196). These slides were chosen because the women they were taken from had
been told they were clear, yet went on to develop high-grade precancerous
lesions or cervical carcinoma. On secondary screening, the researchers found
that 98 of the slides were false negatives鈥攁bout 20 per cent of the
original sample. But the study did not end there. Koss and his colleagues had
used as controls 9666 negative smears from women who were not known to have
cervical cancer. Surprisingly, the secondary screening highlighted precancerous
cells on 1.3 per cent of these slides, too.

Unresolved question

But this is all for secondary screening. Coleman鈥檚 trial in Britain is
studying PAPNET as a primary screener. It is comparing manual screening of 20
000 smears with screening by PAPNET, supervised by a cytologist.

There is in addition the big, unresolved question of economics. For the
cash-strapped NHS, using PAPNET for secondary screening is not an option. 鈥淭he
company tried to introduce PAPNET into the UK for quality control,鈥� says
Coleman, 鈥渂ut the test cost 拢15 and that is in addition to the 拢7 or
拢8 for the initial manual screening.鈥� Though these sums make secondary
screening unattractive, Neuromedical allowed Coleman鈥檚 group to evaluate 2000
slides using PAPNET. The experience encouraged Coleman to persuade the NHS to
fund the present trial of primary screening. In parallel with this study, Nick
Bosanquet, professor of health policy at Imperial College School of Medicine, in
London, is evaluating the economics of PAPNET.

Only when Coleman鈥檚 and Bosanquet鈥檚 results are analysed, will women know
whether the technology of the Strategic Defense Initiative can improve on what
is still鈥攄espite all the legitimate anxieties鈥攐ne of the great
successes of preventive medicine.

* * *

Fighting for a future

REVOLUTIONS may begin for all the right reasons, but history shows that they
tend to be followed by a period of blood-letting as rival camps vie for
superiority. The technological revolution sweeping cervical cancer screening is
no exception.

Even if the present NHS studies shows that PAPNET is valuable for primary
screening, its future is not assured. Another advance in cervical cancer
screening, known as fluid cytology, is forcing Neuromedical to reconsider the
future of its product. In fluid cytology, a cell sample is suspended in a gel,
then centrifuged onto a slide to give a clean, thin layer of cells of a similar
size. This array is easier to examine than the higgledy-piggledy patterns found
on traditional smears.

In May 1996, the US Food and Drug Administration (FDA) approved a version of
the technique developed by Cytyc of Massachusetts. Neuromedical is considering
whether to test PAPNET on Cytyc鈥檚 slides.

Another company, AutoCyte, is developing an automated version of the fluid
cytology technique which includes neural network technology. Observers believe
that Neuromedical might sue AutoCyte. 鈥淏oth are very confident of their
positions, says R. Marshall Austin, a specialist in women鈥檚 pathology from the
Roper Hospital in Charleston, South Carolina.

Neuromedical is already suing another competitor, NeoPath of Redmond,
Washington, for allegedly infringing its patent. NeoPath manufactures AutoPap,
which the FDA approved for primary screening last month
(This Week, 7 February, p 12).
Like PAPNET, AutoPap was designed to be compatible with conventional
smears, not slides prepared using fluid cytology. Unlike PAPNET, which
identifies 128 fields of view on each slide for further analysis, AutoPap
estimates how probable it is that a particular slide contains abnormal
cells.

Austin, who believes that the future of cervical cancer screening will
combine fluid cytology and neural networks, has kept a close eye on the
litigious activities of the competing companies. 鈥淚t鈥檚 been quite unpleasant and
detracted from the new technology,鈥� he says.