IT challenges don't drown the drum of digital

November 2007
Feature Story

Anne Paxton

Virtual versus hands-on. Algorithm versus diagnosis. Computer systems versus anatomy. Software versus wetware. It would seem that the worlds of information technology and pathology could hardly be more different. But pathology's embrace of digital technology is moving the profession solidly into digital image capture for archiving, trading, and training—with some institutions already on the leading edge of more complex uses.

Anil Parwani, MD, PhD, director of the Division of Pathology Informatics at the University of Pittsburgh Medical Center, says he and his colleagues are going full steam ahead: "It may take three to four years, but based on the things we're already doing, at our institution there is strong interest in making the journey toward complete digitation." Pathology and the glass slides that pathology laboratories generate have caught the attention of big corporations, and they're putting their resources into addressing the issues, says Victor E. Reuter, MD, vice chair of the Department of Pathology and director of the immunohistochemistry core facility at Memorial Sloan-Kettering Cancer Center, New York City.

"This fact is evidenced by the mergers and acquisitions that have taken place in the digital microscopy field as well as other large corporations that are now working on digital solutions to pathology.

"I've been very impressed at how rapidly there has been progress in the quality of image and ability to scan faster than two or three years ago," he says, "and I expect that to continue over the next couple of years, given the players involved."

Ole Eichhorn, chief executive officer of digital pathology provider Aperio Technologies, says the company has been in business for six years, but a third of the 300 systems it has installed have been added in the past year.

At the beginning, Aperio's chief customers were medical schools and research firms, primarily veterinary pathologists and large drug companies. "The third and largest, but slowest to adopt, is the clinical market." That is changing thanks to increased FDA approvals of digital technology, and general comfort with it and understanding of the "value proposition" that digital technology poses.

"When it first became a possibility to scan a whole slide, people talked about converting their entire workflow to digital," Eichhorn says. "That's in the future but not the way to start, because you have to gradually accommodate to the laboratory workflow. Instead you find useful niches such as secondary consults, preparation for tumor boards, and so on," and begin filling these one by one.

The digital imaging niches in the University of Pittsburgh Division of Pathology Informatics include consultation, quality control, telepathology, and proficiency testing.

At least once a day, for example, a neuropathologist at one of the UP hospitals will render a diagnosis using scanners in live mode from another hospital three miles away. "That's one thing that's been helpful in our practice," Dr. Parwani says.

Digital technology has also been used to distribute CAP proficiency testing for the past year. "We've been digitizing those 10 glass slides because we have seven hospitals under one department, and it saves the logistics of passing one set of slides around, having one pathologist look at them for two or three days, and so on. Now we just send everyone a digital set."

Key slides for each subspecialty have also been digitized. "We have some subspecialists who only look at gyn or GI specimens, so we have them work with their counterparts in the medical field and create an archive of interesting, more difficult cases that will be useful for training and for quality assurance."

Since current technology provides very high resolution for digital slides, Dr. Parwani is hopeful that digital slides will one day serve for primary diagnosis. In a recent study at Pittsburgh, "we created cases for pathologists in the LIS and gave them no glass slides—only the digital images. We found good concordance between diagnoses with glass and diagnoses with digital images."

So far, adoption of digital imaging at Mount Sinai Medical Center in New York City has been cautiously but steadily expanding under the guidance of Alan Schiller, MD, professor of pathology and chair of the Department of Pathology, says Victor Brodsky, MD, neuropathology fellow. Like Pittsburgh,"We use telepathology when a diagnostician is not available in house, for example when a frozen section has to be done after 7 PM and the pathologist has gone home. We put the slide in the machine and the resident gives the pathologist a call. It doesn't scan the complete image; it's more of a live-view microscope." The other applications include making images for courses available to students and taking digital gross photographs of autopsy and surgical specimens.

Among the more technical issues in IT, the main challenge is still the file size of images produced with a scan of a slide. A single properly done scan can potentially result in a 50-gigabyte file. And unlike in radiology, which produces digital data using the machine directly, the scan of pathology slides has to take into account that cells on a slide are located at different levels. "The bandwidth necessary to transmit such images over the hospital network is definitely scary for the hospital IT department," Dr. Brodsky says.

The speed of scanning is not enough to keep up with Mount Sinai's workflow, which can exceed 1,000 slides a day. But new solutions are being developed that would address this shortcoming. Says Dr. Brodsky, "Instead of simply leaving laboratory workflow as is, if slides were produced and scanned in batches over the course of a night with an array microscope like DMetrix's, 1,000 slides could be scanned in approximately 25 hours, and two such microscopes could do it in about 12 hours."

Other technical issues involve tradeoffs between speed and storage requirements. For example, one 50Ą version of a slide could be scanned and stored, but if the pathologist wanted to look at a power of 4Ą, unless the computer has a pre-made image, it needs to take the time to scale it down.

"To avoid delays," Dr. Brodsky says, "once the image is scanned and stored at high resolution, the 4Ą or 10Ą powers of the image need to be pre-made before the pathologist ever gets to them, resulting in a pyramid of images produced for each slide. You can see how storage space needs to be sacrificed."

But it may not matter in five or 10 years, he adds. Only yesterday, in 1988, the price of a gigabyte of storage was $11,540—an inconceivable sum compared with its price tag now: 18 cents.

User interface issues can present other important challenges. For example, says Dr. Brodsky, because there are variations in viewer equipment, including differences in monitor resolution, there are very specific guidelines in radiology for how good the resolution of the monitor needs to be for it to be acceptable for diagnosis. The same thing would be needed in pathology. But the density of pixels in some currently produced monitors, he notes, 204 per inch, is not far from laser printer quality, and may be good enough to be able to diagnose correctly.

Color is an additional consideration. "When one pathologist sends an image to another for a consultation, the colors that are reproduced on the monitors may actually be very different, so one would need to ensure that the equipment is properly calibrated," Dr. Brodsky says.

Perhaps surprisingly, the speed at which the image is scrolled and zoomed is an enormously important factor. "I can pretty much guarantee that pathologists will not give a system the time of day if they sit down in front of it and the scrolling or zooming speed is perceived as slow," he says. He has seen prototypes of systems that other pathologists agree are fast enough to be usable for diagnosis, though the systems are not yet widely available.

But special attention may have to be paid to input devices. "Mouses and touchpads are very much universal and easy to learn, but if we ever want pathologists who have been working with microscopes for many years to adopt digital pathology systems, we may need something better. Someone who has looked at thousands upon thousands of slides using a microscope might need a specialized input device basically emulating those motions."

Once these challenges are overcome, there will be much less reliance on text, Dr. Brodsky predicts. "After a specimen is grossly described, and the dictation is typed, very often pathologists simply put down the gross description text, and, looking at the resident, ask, 'What did you see?' Or 'Why didn't you take more gross pictures?'"

"If digital imaging is properly implemented, you could potentially have four cameras aimed at the surface of a table, and the person who received the specimen would simply lay it out, press a pedal, get four pictures, and automatically add them to the medical record. A second pedal press would result in the same four pictures now taken with a grid projected on the table, thereby creating a record of measurements. If such images are saved into the laboratory information system, I think the necessity for gross description becomes very minimal."

Genzyme Genetics, a national large-volume reference laboratory in Westborough, Mass., is incorporating a digital imaging system now. "We have a current system, but it is centered around what's called the field of view, not whole-slide images. We use it on our breast cancer cases, but this is our evolution or next phase," says Genzyme IT expert Hugh Wax.

With the new technology, which consists of a BioImagene system and Genzyme customizations, pathologists are going to look at the whole slide. "It's the next level of investigation," Wax says, "so the pathologist, either external or internal, instead of going to the microscope and determining the field of view and doing the analysis, will be able to start from a macro view and move down to a micro view through digital pathology."

With the advent of whole-slide imaging, Dr. Parwani says, "we have the capability of hypothetically digitizing everything, but there are many obstacles that do not allow it.

"The scanners in today's market are getting smaller, faster, and cheaper, and storage today is cheaper. But a pathology practice with 10,000 surgical cases every year with an average of five to six slides apiece—even with the best compression scheme out there, you're still talking about 300 to 400 megabytes per case, or terabytes of information." That's one reason Yale University Medical Center's pathology department has not yet moved to adopt whole-slide imaging in the clinical practice. Yale has been doing digital gross photography on autopsies since 1996 and since 2001 on surgical pathology, says John H. Sinard, MD, PhD, professor of pathology and ophthalmology and director of pathology informatics. "We also do microscopic image capture here, although we don't do it routinely. We have used some dynamic nonrobotic telemicroscopy for consulting or for after-hours interpretation."

Yale does use some whole-slide imaging for medical education and is planning to expand that use. However, adopting whole-slide imaging for routine clinical work would require multiple scanners running in parallel, Dr. Sinard says. "For a high-resolution scan you're still talking about five to 20 minutes per slide, and the only way we could do all our slides is if we bought 10 scanners at $90,000 each. And we don't have a million dollars lying around."

The pathology department at Memorial Sloan-Kettering installed an Aperio XT Scanscope system a few months ago, but Dr. Reuter agrees that memory is not the principal problem. Memorial Sloan-Kettering generates about 500,000 slides a year, plus another 104,000 immunohistochemistry stains, plus other slides for second opinions—in all, more than three-quarters of a million.

Many elements, including the recovery system, firewalls, and security systems, will cost more than memory, Dr. Reuter maintains. "But what cannot be solved right now with money alone is the technology to scan these slides quicker, which will take a couple more years. And everybody's working on it, believe me."

Because pathology is basically where radiology was 10 years ago, Dr. Parwani says, pathologists usually have to use the same monitor to look at the pathology report and the whole-slide image, going back and forth to view them on different software. "So it's not very well integrated into their workflow."

For digital imaging to work, he says, "not only does it have to do the work of the microscope, you also have to have one or two additional monitors as a radiologist would have, and everything should be linked into your LIS, which is not what happens with today's scanners."

What Genzyme is trying to do is deliver the fastest image to the client to make a diagnosis "but still have a system that's manageable from an IT perspective," Wax says.

Under CAP guidelines, laboratories are required to save final reports for at least 25 years. "We plan to continue doing that with the new system in very traditional documents in PDF format," Wax says. "But for the whole-slide files—'typically one gigabyte per slide and 10 slides per case—what we envision is an 'assumed dwell' from the moment we get the slide until the time the pathologist actually accesses the image."

"During that time, the image just sits there taking up server space. So we'll actually start a countdown: Once the pathologist accesses the image, they have so many days to do the analysis and pull the field of view or areas of interest into a report. Actually, we literally do a destructive overwrite," Wax says. "But it's not as destructive as it may sound because the slide is stored, the report is stored, and we can always scan that slide back in, complying with CAP guidelines."

Smooth integration of digital imaging with laboratory information systems, Dr. Sinard points out, is not easy and has drawbacks. The main disadvantage of the integrated solutions currently available from vendors for regular (that is, not whole slide) images is "you need to first of all be running the LIS to get an image." The hardware used as acquisition devices has to be hardware that the vendor has worked out the details of integrating. "That significantly limits your hardware choices and can affect longtime compatibility," he says.

Another compatibility issue arose at Yale because the LIS could run only under a Windows environment. "All our hardware was Macintosh, so we needed to develop a solution that was operating-system-independent to accommodate the hardware."

That stand-alone system worked but required them to figure out how to get images into the LIS, which in their case involved arranging back-door access into the LIS courtesy of the lab's vendor, Cerner.

Many pathologists are pushing to take digital imaging to the next step, Dr. Parwani says. "They want to make archives more digitized, more annotated, more accessible, and they want to keep adding more cases to the archives." To make it happen, there needs to be more pressure on the LIS vendors. Without push from pathologists, he says, "it's not something they are ready to deploy on a wide scale."

"When you have something on glass, I think that is the end point. But once it's digitized, you can access it anytime, you don't have to look in a file, and you can do additional things with it," Dr. Parwani says. For example, his laboratory is installing new software now that will quantitate immunohistochemistry. It is also already using a system at one of the hospitals for breast specimens that allows quantitation of breast markers.

"At our institute, we're incorporating images into pathology reports for a small number of cases," Dr. Parwani says. For example, in a case of skin rash, the pathologist will view it microscopically, "then we put the images directly into the LIS and embed them into the pathology report."

"We're also starting to send gross images directly into the radiology system, so that radiologists and clinicians and surgeons can look directly at them. For example, they might look at the gross image of a liver with the MRI of the liver at the same time."

Down the line, Dr. Parwani hopes, the images might be integrated into the hospital information system. "With the LIS vendor there are more enhancements in the software that allow us to create images in reports. We can actually create image-rich PDF files of pathology reports which have not only pathology images but also gels, electron microscopes, or anything that could potentially be used in a report." These can be viewed now by the Web server, but the HIS is not capable of taking these images in reports and displaying them.

Memorial Sloan-Kettering is also working on incorporating digital images into the electronic health record. "That may come to fruition in a year or two, but to do it we will need to tie into the LIS," Dr. Reuter says. "Aperio is working with our LIS vendor on that." When it happens, and proper firewalls are established, other physicians will theoretically be able to access microscopy slides through the medical record. "You can already acquire a static image, but we don't yet have the interface to be able to look at whole slides."

As for the networks to transmit these records, Aperio's Eichhorn says, "It's very important to be able to use the public Internet. It's ubiquitous and very low cost compared to private networks. Obviously, HIPAA privacy concerns are very important and we have to use encryption of data, but fortunately there is already a lot of good encryption available from other fields such as online banking."

Still at the blue-sky phase, but getting far more attention these days, is the possibility of using digital imaging to actively help with diagnosis.

"A pathologist might look at 1,000 cells and say about 70 percent are stained brown, but a computer can measure that to 15 decimal places," Eichhorn points out. "So quantification is one of the key values of digital imaging. That's simpler than some of the other possibilities, which are really pattern recognition, and people are starting to look at various ways of pattern recognition on digital slides, looking for things that are abnormal or rare events."

"There might be only one or two cervical cancer cells in an entire Pap smear, and if you miss those you're missing the chance to do an early diagnosis that might improve the chances of the patient. So it's a very fruitful field and there are a lot of clinical trials on it."

The point will not be to replace the pathologist but to serve as a tool, Eichhorn says. "Computers can do things that would be very time-consuming for humans—for example, looking at an aerial photograph of a city and finding all the red Volkswagens. But humans are better at pattern recognition."

"One of the things that has really driven digital mammography is the availability of tools that do computer-aided diagnosis on mammograms. The same thing may end up being even more important in pathology because the images are so much larger."

Ulysses J. Balis, MD, associate professor and co-director of the Division of Pathology Informatics at the University of Michigan Health System, Ann Arbor, is working on several prototype systems that seek to provide computer-aided diagnosis based upon real-time searching for images with images themselves.

"With prior image analysis approaches, there was a requirement for some degree of expert knowledge. In this approach, the goal is to implement turnkey search platforms, with this category of algorithms being self-directed; it scans and matches the features of interest autonomously," Dr. Balis says.

"For example, in prostate biopsies, one is interested in separating all benign glands from the occasional, and often rare, glands that have lost additional epithelial lining, which can signal dysplasia or at worst carcinoma," he says.

Conventional image analysis requires significant work to accomplish this, but an algorithm he has had in development for the past 10 years uses a number of representative images that have been input into the system along with suitable classifiers that distinguish benign from malignant tissue, and searches for matches.

Radiology, now almost entirely digital, offers the closest comparison to what Dr. Balis is doing. "Even though their data sets are 1,000 times smaller, they've had only limited success, probably due to the fact that their imaging data sets have significant variation in structure from one region to another, so they're not amenable to highly automated, self-classifying algorithms."

"Pathology, which has organ, with repeating structures, or cells, which are inherently repetitive, as its subject matter, has much more redundancy within the image. This feature made these classes of images amenable to image compression, but also allows for the creation of a vector quantization vocabulary or a statistical map of features. It's a one-to-one structural equivalent of the original image but in highly tokenized fashion."

At this point Dr. Balis has it working on small servers, and has extracted a pilot version that even runs on a laptop machine. The processing has already improved from an overnight computational exercise to something that can be carried out in real time. But to be employed as a clinical decision support tool, where the pathologist is looking at the field of interest and needs a tool to return an answer immediately, "the tool has to be able to provide those answers practically instantaneously."

While different compression algorithms have to be developed for each organ system, nothing at all prevents the profession from using them as decision support tools, Dr. Balis says.

So far he has had good results with developing vectors for liver and for prostate. Liver fibrosis, for example, provides for a reasonably consistent grading of particular features that can be compared to archetypes previously entered into the system as vocabularies.

But calling computer-assisted detection "diagnosis" is too bold, he says. "It's simply a decision support tool. The pathologist is still firmly in control of every aspect—the difference is that now you have two tremendous opportunities, first for a tool to provide differential diagnosis of perhaps unusual cases, and second, the opportunity to apply the algorithm en masse to the entire surface area of a whole slide image to make sure the pathologist didn't miss an important morphological feature. Neither changes the fundamental practice, which is the pathologist reviewing slides and rendering a diagnosis."

The point at which FDA oversight would become mandatory would be if you wanted to take the pathologist out of the loop, as has been done with Pap tests, where laboratories are able to sign out a case as negative with only a percentage of rescreening required. But "that's a question for later generations," Dr. Balis says.

For now, "it's a type of technology which in many sectors would be labeled as disruptive," he concedes. His discussions with other pathologists have confirmed that, to support production workflow, the technology would have to have all organ systems covered, and it currently doesn't. Promising research continues, but other algorithms often take significant time to compute. "It's possible to combine these algorithms in real-time testing," he says, "but there's nothing ready for prime time in clinical decision support."

Dr. Balis became interested in image compression 12 to 15 years ago, before people had high-speed Internet access. "Back then there was no whole-slide data set. It could literally take hours to send around information." But now, since storage is far less expensive, "to me the opportunity to carry out a search in a file thousands of times smaller than before is really much more compelling than simply compressing the image."

To what extent are human factors a driver of or impediment to adoption of more digital imaging applications? Dr. Balis thinks the issue is more "use case justifications."

"Clearly we don't put microscopy images on our surgical pathology reports because they are generally for the consumption of the ordering physician—typically not a pathologist, and the image really has no bearing on what they'll do, whereas with radiology, many types of medical specialists make direct use of the data."

So the use cases may not quite exist yet, because the technology is not up to the task of real-time matching. But Dr. Balis predicts digital imaging will prove useful in the future for decision support and rare event detection.

"Say you want to quickly go through a sentinel lymph node and say there are no epithelial elements or malignant ones in the surface for review. It would go a long way to have a computer-assisted re-evaluation of that whole surface area that confirms that assertion, whereas now there is always concern that you missed a spot."

Dr. Sinard, who participated recently in a "point-counterpoint" panel with Dr. Balis, feels that until usable computer-assisted diagnosis becomes a reality, there is limited usefulness from routinely imaging all glass slides. "If we could scan slides in zero time and had infinite storage space and infinite bandwidth to move images around and zero latency, basically eliminating all the technical barriers and everything was free, what are really the factors that are driving scanning of slides?"

"What does it allow us to do better or easier than what pathologists currently do? When you look at it objectively, there are really very few benefits in routine practice. So my argument is, until the scanning will give us a quantum leap forward in our ability to work on these cases, and it probably will eventually come in the form of computer-assisted diagnosis, I don't think there's enough of a driving force to push people in that direction."

Though computer-assisted diagnosis is starting to make its way into the radiology field, Dr. Sinard says it's still very early. "Many radiologists I've spoken with have suggested the systems are able to diagnose only very simple things, about what a junior radiology resident can do."

However, radiology's data sets are far smaller than pathology's, and Dr. Sinard doubts that computers can be taught to diagnose histological images reliably. "Pattern recognition is not something computers are great at. Density interpretation and quantization, yes—computers are very good at that. But pattern recognition is one of the things the human eye and brain are better at than computers."

Dr. Reuter, too, believes that content-based retrieval is far off. "Basically pattern recognition in computers is a theoretical possibility, but there are so many subtleties to pathology that I think it is very remote."

"I think there's a better chance imaging will help you do a lot of morphometric studies guided by a pathologist that might help us do a little better job in establishing the path or the aggressiveness of a tumor, or read signals like antigen detection or quantification," he suggests. "Digital microscopy is going to help a lot more with those things than with diagnosis."

Dr. Brodsky believes resistance to change continues to present a significant obstacle. As a digital technology advocate, he was caught short when a highly ranked retiring member of his department told him straight on during a conversation, "You're not here to change pathology." "Whether it's active or passive," Dr. Brodsky says, "I think it's one of the top reasons why digital pathology may be a little slow to be adopted."

It will become more popular, though, and pathology practices should not feel confined to exploring only certain solutions, Dr. Sinard says. "They should think outside the box a little bit and find other ways of incorporating this technology to the extent they find it useful."

Anne Paxton is a writer in Seattle.