CEO Summary: Pathologists at the Hospital of the University of Pennsylvania are using a new system that combines image analysis software and algorithms to evaluate images containing numerous stains and biomarkers. Pathologists teach the system to identify tumor cells and distinguish them from non-tumor cells. Now used for research purposes, this sophisticated digital pathology system is designed to do much of the manual activity required of a pathologist when assessing an image while producing highly accurate quantitative data.
EVEN AS WHOLE SLIDE SCANNING and digital pathology systems take root in the anatomic pathology profession, more technology surprises lie ahead. In research settings, smart software is already being used to analyze digital pathology images.
These systems combine automated image processing with analytical software that “learns” with each pathology image processed. One such system is “inForm”, developed by Cambridge Research & Instrumentation, Inc. (CRi), of Cambridge, Massachusetts.
An early customer of CRi’s technology now using the inForm system for research is the Pathology Department at the University of Pennsylvania in Philadelphia, Pennsylvania. Pathologists there are developing ways to evaluate pathology specimens using multiplex assays that incorporate numerous stains and biomarkers. Because so many different biomarkers are used, the analysis is quite complex.
This complexity goes beyond the capability of the human eye. What makes it possible to evaluate these pathology images is a combination of computer software and algorithms. Essentially, the computer uses its “smarts” to sort through the combination of stains and biomarkers to present the pathologist with its analysis of the cells, along with a precise mathematical score for selected variables.
“It is important to understand that, although the inForm software and Vectra multispectral imaging hardware ‘automates’ the analysis of these images, it is the pathologist user who directs the advanced features of the system to teach it what to recognize,” stated Michael D. Feldman, M.D., Ph.D., Associate Professor of Pathology and Laboratory Medicine at the Hospital of the University of Pennsylvania (HUP) and a CRi collaborator.
“There is rapid progress in the field of digital pathology and pathologists should understand that our current work with these intelligent algorithms and a system like inForm adds one more dimension,” noted Feldman. “These algorithms allow us to maximize the utility of hardware devices. For that reason, pathologists should expect to see a variety of different imaging algorithms enter the clinical market in the near future.
Seeking an Imaging Platform
“The Pathology Department at UPenn got involved with CRi because we wanted a technology platform capable of handling advanced imaging,” he continued. “Specifically, our goal was to put down multiple protein marker antibodies onto tissues in a way that would allow us to identify their location in the tissue.
“For example, were they in a tumor, or a lymphocyte, or a blood vessel?” asked Feldman. “Once identified, we wanted a way to quantitatively and cytometrically analyze the information. That would allow us to see how these markers associate in such locations as the nucleus of a tumor cell, in the cytoplasm, or in the membrane.
“This technology goes beyond standard whole slide imaging,” he said. “We realized from our clinical trials work that we needed a different analytic platform. We wanted what we call ‘multispectral capability.’
“There was a host of specific biologic questions we couldn’t answer with standard whole slide imaging,” commented Feldman. “That led us to work with CRi. With CRi, we obtained a small business grant from the National Institutes for Health (SBIR) to develop and use this technology.
“We are also developing our own software in collaboration with Dr. Badri Roysam’s lab at Rensaleer Polytechnic Institute to do certain types of cytometric analysis. Independent of our cytometric software, CRi is developing its own cytometric platform,” he noted.
“Not only does this technology go beyond whole slide imaging, but it is downstream from computer-aided diagnosis as well,” Feldman continued. “We work with tissues for which pathologists have already made the diagnosis of cancer.
“When I have a stained area in a field of view, I can tell the system to find the cancer region and the non-cancer region,” he explained. “That allows me to then associate one or more stains with individual cells, along with their subcellular regions.
“If we know there is cancer on the slide or in the field of view, we can ask the inForm software to map every area where cancer exists,” commented Feldman. “Next, the software will give me a quantitative analysis within those areas.
“What makes this a notable development is that the robot and software are doing the time-consuming manual part of the pathologist’s work,” he continued. “By doing this work for the pathologist, it extends his or her ability to diagnose more cases, thus boosting productivity.
“Think of it like this. No longer will a pathologist sit there and circle the areas which show the tumor regions. The system and the algorithms will do that for the pathologist with a very high level of confidence,” said Feldman.
“When presented with this information by the system, the pathologist then applies his or her own domain knowledge,” he explained. “The pathologist next views the work of the software and acknowledges that it correctly identified the areas of interest.
“This is the beauty of automated image processing and computer-assisted diagnosis,” stated Feldman. “The software does automatically what would be time-consuming manual work for the pathologists, then presents the data for the pathologist to validate, verify, and interpret.
“This is how the system reduces manual steps that are time intensive,” he added. “Once a pathologist adapts to this process, it speeds the analysis time for each image.
“Here’s an example of how it works in our lab,” said Feldman. “We’re doing a clinical trial with the Eastern Cooperative Oncology Group. In this trial, we must assess how patients respond to different therapeutics. We have approximately 150,000 fields of view to evaluate.
“Imagine the time required for pathologists to properly analyze 150,000 views of information in the traditional manner, with a glass slide and a microscope,” he observed. “There are not enough pathologists here at UPenn to handle all those images.
“For this clinical trial, we train the inForm system to evaluate the biomarkers and targets with a very high confidence,” said Feldman. “This project, with its 150,000 images, would be nearly impossible to do manually. However, suddenly we can do this because of the automated functions and the machine learning algorithms within inForm. Moreover, as a pathologist using this system, it becomes a multiplier of my knowledge.
“This next-generation technology positions anatomic pathology for a huge breakthrough,” he predicted. “The diagnostic state of the art today is working with one stain at a time,” Feldman commented. “That’s all we can do. But tomorrow, we will ask more complex questions that can’t be answered with one stain at a time. Tomorrow the gold standard in pathology will involve working with two, three, and four stains at a time.
“Flow cytometry provides a parallel example,” he continued. “Back in the 1970s, when pathologists introduced flow cytometry, we did one stain at a time. Over time, it became possible to do two stains at once—and then three, four, and more. Now it is routine to do flow cytometry with six color stains or more for routine clinical work.
“Biology doesn’t happen one molecule at a time,” observed Feldman. “There is an integration of information on and within cells that includes both positive and nega- tive regulators. To understand what’s going on in cells, it is necessary to look at more than one biologic pathway at a time.
“Hematopathology illustrates these principles and points us to the future of these concepts in surgical pathology,” he added. “Will this also apply to clinical laboratory testing in the future? Yes, I have absolutely no doubt about that.
“However, at the moment, this type of work—translational research—is limited to clinical research,” said Feldman. “During a clinical trial, when a physician gives a drug to a patient, we try to measure the response. That’s one form of translational research.
“In doing translational research, as we develop and validate assays that become meaningful for routine clinical practice, these new assays will migrate from the research space and into wider use by clinical laboratories.
Beta Testing Begins
“Our working collaboration with CRi started about 10 years ago,” he commented. “It was last year when we got this digital pathology robot system (Vectra). That’s when we initiated beta testing.
“The robot has shown us that we need to consider the variation from lab to lab in the handling and processing of tissue and make sure that our systems account for variation in staining and other variables in how we handle specimens,” he added.
THE DARK REPORT believes that the research unfolding at the University of Pennsylvania provides a timely peek into the next generation of digital pathology systems and its potential to change clinical practices in surgical pathology. Feldman and his colleagues are demonstrating how robotics and software can automate many of the manual functions that require time when a pathologist evaluates an image.
Equally intriguing is the ability to assess multiple spectra and biomarkers within the same issue—and accomplish this with mathematical precision. Again, the research activities with the advanced digital pathology system in use at UPenn show that technology makes it feasible to achieve this level of performance today. Collectively, these developments point to an exciting and profitable future for anatomic pathology.
Increased Specificity Demonstrates Need for Pathology Labs to Eliminate Variation in Specimen Handling and Processing
ELIMINATING VARIATION IN LABORATORIES is an important issue for the field of pathology,” observed Michael D. Feldman, M.D., Ph.D., Associate Professor of Pathology and Laboratory Medicine at the Hospital of the University of Pennsylvania in Philadelphia. “It is time for our profession to pay increased attention to our analytes and develop better standards.
“There has been progress in addressing variability across laboratories,” he added. “However, in surgical pathology, we need to start the process of standardizing how tissue is handled, processed, and manipulated.
Standardization Is Needed
“It will take some time to accomplish this, but standardization has to happen. Our entire industry needs to pay attention to this goal,” said Feldman. “In addition to these standards, the pathology profession needs computer systems to help achieve this standardization. Today no lab information system exists that can capture all the data and all the fields required to provide confidence that the analyte to be evaluated is well-controlled.
“When asked, many pathologists respond by saying it is a daunting challenge to seek such standardization,” recalled Feldman. “They say it’s not likely that the profession will ever achieve standardization.
“On other hand, I have experimental data that suggests that, if you don’t achieve levels of standardization, then the results of the complex analysis that sophisticated software can do today will be meaningless. That is because the variability of the sample is greater than the ability of the systems to measure. It’s an example of the well-known adage: ‘garbage in, garbage out.’
Impact of Ischemic Time
“Here’s an example. We did one research experiment on the impact of ischemic time,” he explained. “We wanted to know how shorter and longer ischemic times would affect our ability to analyze the critical molecules that are the target of the drugs in the clinical trial.”
“We did a simple experiment to compare the results produced by a resection specimen and the biopsy specimen from the same patient,” continued Feldman. “The resection specimen usually has one or two hours of ischemic time. The biopsy has almost no ischemic time. When an organ is removed, its blood vessels are clamped and tied before it is taken from the body. That entire process exposes the tissue to ischemia.
“We wanted to know what happens to the specific molecules that are the targets of the drugs being studied in the clinical trial,” noted Feldman. “These are the molecules we need to measure.
“Is there a difference if I measure these molecules in a biopsy versus a resection sample?” he added. “We know that the very robust stains in the biopsy have almost zero expression in the resection specimen. There’s a complete absence of the target molecules when the organ is taken out and compared with the biopsy sample. The range of intensity changes we saw was significant. Some tissue samples showed no loss of signals. Some have partial loss and some have complete loss.
“This issue is critical for laboratories,” declared Feldman. “As a profession, we need to get our hands around how samples are handled. What are the best types of samples to be tested? What critical features need to be tracked as tissues are processed?
“The answer to each of these questions will have a profound impact on the amount of the protein or phosphoprotein epitopes that are present in the tissue we want to examine.
“The advanced pathology systems under development today have so much analytic sensitivity and precision that this tissue handling variability makes a huge difference,” stated Feldman. “In the future, pathologists will no longer score these images as 1+, 2+, or 3+.
Once this Pathology System Learns, It Can Continually Analyze Thousands of Images
PATTERN RECOGNITION and a machine classification system are at the heart of a sophisticated digital pathology system currently used for research purposes.
“The Vectra robot system from CRi gives us the capability to use numerous stains and biomarkers to perform very complex analyses of tissue,” stated Michael D. Feldman, M.D., Ph.D., Associate Professor of Pathology and Laboratory Medicine at the Hospital of the University of Pennsylvania. “It can automatically map the areas of an image that have tumor and non-tumor cells.
“The pathologist teaches the system from these images,” he continued. “For example, as it learns, I can continually refine its knowledge base. Upon achieving proficiency, I can then run tens of thousands of images through the system and it will work night and day without complaint or pause.
“This system is an imaging platform that is linked to the intelligent software,” stated Feldman. “The robot scans slides at low magnification. The pathologist then uses these low magnification images—where the good fields of view are—to teach the system to find different tissue regions on the slide.
“Next, the system’s software images the high magnification areas based on how it has been trained and how many fields of view need to be acquired to analyze the slide,” he continued. “The system then fetches a pathologist-determined number of high-resolution fields. It uses these fields to collect statistically relevant data which are output as cytometric results.
“It takes more than one field of view to determine if there are enough cellular events to be statistically meaningful,” he added. “The tissue must be sampled at multiple areas to get an adequate number of data points. The flow cytometer is a good metaphor. A flow cytometer doesn’t analyze 100 cells. Rather, it analyzes tens of thousands of cells.
“This sophisticated digital pathology system works in a similar way,” observed Feldman, “Enough fields of view need to be selected from the tissue specimen to represent thousands of events. That way, the statistics will bear out positive and negative results. The automated system allows the pathologist to find those areas of interest and regions at low magnification and then pull them in at high magnification.”