Siri. Self-driving cars. Chatbots.
Applications for artificial intelligence (AI) are everywhere these days – and headed toward the stains and slides of a pathology lab.
Computer scientists see the potential of AI in clinical diagnostics, where computers are taught to read slides of biopsy tissue and make inferences. Google’s LYNA research project, for example, aims to speed up slide analysis and diagnostic accuracy, essentially codifying the expertise of a pathologist.
At OHSU, two Ph.D. candidates envision a different shift for pathology: Speedy Histological-to-ImmunoFluorescent Translation or SHIFT.
Their project uses a computing approach called artificial neural networks to detect pancreatic cancer in a new and less expensive way. It leverages powerful, machine-learning principles and sophisticated multiplexed imaging data generated at OHSU.
Award-winning
Earlier this year, Erik Burlingame and Geoff Schau of the Biomedical Engineering Graduate Program and Computational Biology Program, pitched their proof-of-concept to judges at the InventOR Collegiate Challenge, hosted at Portland State University.
Sporting blazers and business cards, they won the competition’s Impact Award, taking home a $5,000 prize and validation that SHIFT had business potential. Early funding for their prototyping came from a $40,000 grant from the Biomedical Innovation Program (BIP) funded by the Oregon Clinical and Translational Research Institute (OCTRI).
“We envision this type of system as a diagnostic aid that pathologists can use to accelerate the rate at which they function, not something that will replace them,” said Schau.
Power of multiplexed imaging
Pathologists evaluate biopsy tissue using standard histology stains, looking for signs of disease. But the complexity of many diseases, including different cancer types, might necessitate a deeper probe that could help clinicians evaluate targeted therapies.
“The field of precision medicine is based on the idea that tumors vary in their responses to modern anticancer drugs because they differ in molecular composition and organization that are not readily apparent in standard pathological preparations,” said Joe Gray, Ph.D., professor of biomedical engineering, OHSU School of Medicine, and director, OHSU Center for Spatial Systems Biomedicine (OCSSB).
These special stain techniques currently in development help pathologists identify protein markers that can reveal the nature of the tumor more precisely.
“By using these protein markers, we can determine if tumor found in the liver, for example, originated from the colon or from the lung,” said Christopher Corless, M.D., Ph.D., professor of pathology, OHSU School of Medicine, and director, OHSU Knight Diagnostic Laboratories.
One such special stain is cyclic multiplexed immunofluorescence (cmIF) – a type of multiplexed imaging developed by Koei Chin, Ph.D., research associate professor of biomedical engineering, OHSU School of Medicine, and Dr. Gray.
CycIF and other multiplexed imaging technologies deliver powerful insights into a tumor microenvironment; they’re currently used in research studies and clinical trials across select biomedical research institutions, including the OHSU Knight Cancer Institute.
“New generation staining procedures can reveal these diagnostic differences but they are expensive and time consuming to generate,” said Dr. Gray.
The process also necessitates additional biopsy tissue, which might mean using up tissue before a pathologist can arrive at a final diagnosis.
Code that spots cancer
Burlingame, who is a third-year ARCS Foundation Oregon scholar, and Schau saw an opportunity. Recognizing that there might be a relationship between tissues stained with either histological or cmIF/IF stains, they devised a way to test a machine-learning algorithm’s ability to learn this relationship, building off work led by their research mentor, Young Hwan Chang, Ph.D., assistant professor of biomedical engineering, OHSU School of Medicine, and a faculty member in the Computational Biology Program.
Using images of the same tissue visualized with both histological or IF stains, the doctoral students trained their algorithm to identify patterns in standard histology images that correspond to pancreatic cancer cells labeled with a pan-cytokeratin (panCK) IF stain. They built up the program’s knowledge base, using tumor data from the Brenden-Colson Center for Pancreatic Care and the Gray lab.
Once the algorithm was trained, the students tested its ability to infer panCK labeling in histological images not used during training. Their testing demonstrated that panCK labeling could be inferred with up to 95 percent accuracy across large areas of histologically-stained tissue in a matter of seconds, effectively enabling virtual IF staining of standard histology slides.
Faster, cheaper
“We’ve demonstrated proof of principle, and we plan to validate our work using additional tumor data,” said Schau. “The takeaway is that SHIFT can make reasonable inferences that detect cancer in this particular setting and may translate to other areas of biology.”
By training an algorithm to see abstract objects in a standard histology slide and make logical inferences from it, Burlingame explained, pathologists get an enhanced interpretation of the tumor microenvironment without needing additional biopsy tissue.
“When validated, the new SHIFT algorithms promise to increase the information that can be obtained from standard pathological preparations in a way that will improve our ability to select the right drugs for the individual patient while keeping costs low and speed high,” added Dr. Gray.
That will make the diagnostic aid particularly useful in underserved medical communities, says Burlingame.
Shift for pathology?
The project is still in its infancy, observed Dr. Corless, and he cautions there are nuances in diagnostics that a computer algorithm can’t detect, but he sees potential.
“Being able to do something instantaneously which, right now, takes half a day would be a boon for pathologists,” he said.
Schau and Burlingame, for their part, credit success of their invention to guidance from mentors, close collaboration with members of the OCSSB, support from the BIP, OCTRI and Jonathan Jubera, M.B.A. through OCTRI, and partnership with OHSU’s Advanced Computing Center.
As multiplex imaging matures and becomes commercially viable five to ten years from now, it has the power to transform pathology’s workflow, say these researchers, and tools such as SHIFT could become an indispensable part of advancing precision medicine.
