by Michael Schneider, Pittsburgh Supercomputing Center

Pittsburgh, PA. — What do umpires and pathologists have in common? An umpire calls balls and strikes, and a pathologist makes the call on biopsies – looking through a microscope at slivers of tissue to decide whether they’re malignant or benign. With a good umpire, a pitch off the outside corner is a ball. But the next at bat, even with the same ump, a pitch at the same spot might be strike three, you’re out.

We like to think pathology is different, more of a science, but there’s considerable research showing what should come as no surprise: Pathologists are human. The differences between malignant and benign tissue can be subtle. Interpretations are often subjective. A recent study at Johns Hopkins, for instance, found that 2 to 3 percent of diagnoses are either wrong – misreading malignant tissue as benign or vice-versa – or could have been more accurate, potentially leading to unnecessary or inappropriate treatment.

Can computers help? It’s a question that occurred several years ago to Dr. Michael Becich, a pathologist at the University of Pittsburgh Medical Center. What if you could build a database of accurately diagnosed tissue and do a computerized search to find close matches for an undiagnosed sample? He took his questions to the Pittsburgh Supercomputing Center and, a few years later, the progeny of this PSC-UPMC partnership is a still-evolving software method called Computer-Based Image Retrieval. CBIR does essentially what pathologists do when they flip through the pages of a reference book, but objectively, more comprehensively and in a matter of seconds.

“We’ve developed a tool,” says Becich, director of genitourinary pathology and informatics at UPMC, “that allows pathologists to classify microscopic images on the basis of image content. Instead of using text, where you’d go to a database and say show me your images of prostate cancer, for instance, CBIR uses computerized image-classification to create image signatures. For an unknown sample, the image signature then acts as the search key.”

The image-search algorithms are computationally intense and require high-performance computing. The aim, emphasizes PSC scientist Art Wetzel, who developed the image-analysis and signature-matching software, is not to replace the training and skill of doctors, but to use technology to provide more information in a more objective manner: “When pathologists get stumped, they go to their shelf and look for examples in a book, but a book has a couple hundred pictures at most. It can’t cover the range of examples we can build into a database. Nor is it easy to find matches in a hurry if it’s a tough case. We’re trying to save search time and broaden the potential choices. There are many situations where a dollar of computing can save much more than a dollar of pathologist’s time.”

The PSC-UPMC team, which includes pathologists John Gilbertson and Rebecca Crowley and grad student Lei Zheng, who created a web-based user interface, has built an archive of about 120,000 images, representing a wide range of tissue categories – muscular, glandular, gastrointestinal, etc. This archive, which is still expanding, is itself an invaluable resource for research and teaching. The images along with associated signatures occupy about a half terabyte (a trillion bytes) in PSC’s archival storage, where it shifts as needed between on-line disk and low-cost tape. Using this data, CBIR has shown impressive ability at matching unknown samples to the correct tissue category.

In going beyond classification to diagnosis, the research team has focused so far on prostate tissue, a specialty of Becich. Although grading prostate tissue is notoriously subject to interpretation, CBIR has matched up well with grading by pathologists and has shown its potential to become a widely used tool.

As the most frequently occurring non-skin cancer in American men, prostate cancer is an apt candidate for improved diagnosis. Recent data indicates more than 200,000 new cases each year and more than 40,000 deaths, second only to lung cancer for cancer-related mortality in U.S. males.

Gleason grading, named for the doctor who developed it in the 1970s, is the best available method for evaluating severity and long-term prognosis of prostate cancer. Pathologists examine the tissue under a microscope and assign a number – from one through five, one being least severe – according to the structural patterns they observe. By adding the two most predominant Gleason grades in a sample, the tissue gets a score between two and 10.

This score is a major factor in treatment planning. The problem, however, explains Becich, is its inherent subjectivity, with research showing a 20 to 40 percent variation when the same samples are scored by different pathologists. Most of this variability comes into play in the range of scores between five and eight, a range often decisive in choosing between surgical removal of the prostate gland or, in more advanced cases, hormonal and chemotherapy.

With CBIR, the major challenge for Wetzel has been to create signatures that define images in a medically useful way. To do this, he implemented a series of image-classification algorithms, some based on coloration, others on mathematical methods – which allow the computer to discern image features that often aren’t visually apparent. For prostate grading, he turned to a method – called “spanning trees” – by which, in effect, the software draws lines connecting cell nuclei. The distribution of lengths and angles of these lines correlates to the Gleason grade.

Using this method on tissue samples where pathologists have identified a region of interest within the sample, CBIR achieves 80 percent agreement with Gleason grading by UPMC pathologists, and most disagreements are confined to one grade level. It’s difficult to gauge what this means with respect to accurate diagnosis, says Becich, due to the inherent variability in human grading, which means there’s no standard: “It’s hard to know if we’re doing better or worse, until we develop a new, objective scale, but it’s clear – since pathologists have this 20 percent variability – that CBIR can do as well as humans.”

“Part of what CBIR offers with prostate grading,” says Wetzel, “is putting a stake in the ground and getting an objective evaluation of where the boundaries are between grades. It’s not so much that there’s always a right or wrong answer; it’s that we need a consistent answer.”

Wetzel emphasizes that CBIR’s ability at Gleason grading at this stage is confined to cases in which a human picks out the region of interest. Though relatively easy for pathologists to do, it’s a challenging software problem, and one of the directions of future work: “How do you know which parts of a slide are important? How do you ignore artifacts like torn edges, poor staining, things that people can look at and know immediately to ignore? The things human vision can do automatically are hard to program. We need to move into some of these areas.”

With prostate grading, Becich is tracking the results of cases in which CBIR picks out matches using signatures unrelated to the structural patterns of Gleason grading. “We’re currently looking at a large number of prostate cases in which we have clinical follow-up to see if CBIR picks out high-grade, poor prognosis tumors better than human grading. One of the possible outcomes is we may be able to provide a grading system that supercedes Gleason at being able to identify more prognostic information.”

The PSC-UPMC team sees CBIR as a resource that will eventually work with high-performance networking to provide “telemicrosopy” for pathologists at remote locations, to save time and improve diagnoses. A “differential diagnosis” provided via CBIR would rank matches from a CBIR search to provide a range of choices. “That’s the way medicine works,” says Becich. “When you get a set of symptoms and signs on a patient, you want to know not only the conditions it most likely represents but also others that may be less likely. Things aren’t always what they seem. When you broaden the range of choice, you improve the ability to arrive at correct diagnosis and appropriate treatment.”

For more information, including graphics, see: http://www.psc.edu/science/wetzel.html .

============================================================