AI Solution Shows Promise in Assisting Pathologists Diagnose Transplant Rejection
An artificial intelligence-based system developed at Penn Medicine shows promise at generating diagnoses about transplant rejection that are just as good as the grading performed by expert pathologists.
In a recent interview, Eliot Peyster, M.D., a heart transplant surgeon at Penn Medicine’s Cardiovascular Institute, discussed his work on this system and the paper his team recently published in the European Heart Journal.
Organ rejection is a serious concern in heart transplant medicine, occurring in up to one-third of recipients. In an effort to provide surveillance, transplant physicians perform a dozen or more scheduled tissue biopsies of the heart during the first year after transplant. Although this type of biopsy with histologic grading is the diagnostic standard for identifying rejection, poor inter-pathologist agreement creates significant clinical uncertainty and confusion, Peyster said.
“The pathologists assign a grade between zero and three — three being serious, two being pretty bad rejection, one being sort of a mild thing, and zero being nothing,” Peyster explained. “And generally speaking, twos and threes are clinically actionable. The problem is that the grading criteria are not very reliable because the pathologists don't agree with each other. We've seen the pathologists agree with each other at best two-thirds of the time,” he said. “That leads to confusion and makes it difficult to do multicenter research. It’s hard to compare grades from different pathologists across different centers, and you just don't really know where the truth lies sometimes.”
Peyster and colleagues digitized the pathology slides and trained an automated tool they call the Computer-Assisted Cardiac Histologic Evaluation (CACHE)-Grader to read them. He said that the tool has proven both highly sensitive for serious rejection and essentially never assigns a falsely reassuring grade. This means when it assigns a “no-rejection” label, this can be highly trusted, and when it assigns a “potentially high-grade rejection” label, workflows could be developed to forward the sample to a pathologist for urgent expert review.
“If the tool suggests that it might be a higher grade, that may be one that gets rapidly forwarded to an expert pathologist for a second look or for an expedited review, because it's been flagged as potentially dangerous,” Peyster explained, “so it could be a workforce multiplier. In that regard, it could speed things up. Pathologists might like it just as a second opinion. There are a lot of borderline cases — that's why there's so much disagreement. And they may find it helpful to have a second reader right there next to them a click away that can help nudge them one way or another, if they're ambivalent.”
Another application that Peyster thinks the CACHE-Grader is already ready for is as a cloud-accessible software platform to help standardize grading across centers to enhance the quality of transplant research.
This work has implications for kidney transplant and lung transplant research as well, Peyster said. They started with the heart because with those transplants they do more biopsies than on other organs and have larger datasets.
More importantly, once you have all these variables describing the tissue, you can move beyond grading, he stressed. “Instead of with a grade, what if I correlate my variables with how sick the patient is, how sick they'll be next month?” he said. “Once you turn the tissue into data, you can correlate it with anything. It gives you an opportunity to mine deeper into these valuable pieces of biologic information. These tissue samples contain so much information that can be looked at and they are underutilized. Doing this type of workflow allows you to access that information. Then it's up to you to figure out all the things you can connect with it. That's really the heart of precision medicine and how computer vision can enable it. That’s what the next-generation projects are focused on — predicting who's going to get sick in the future and who's imminently going to be sick. Now, how do we change their immunosuppression, based on that information for every person? That's the next step.”
In addition to helping with pathology grading, the tool might also help with things that pathologist don't do currently with these tissue samples. “That's really the vacant space where the same technological approach can gain traction as a precision medicine clinical tool — answering questions people didn't even know they should be asking yet and don't have a current framework around,” Peyster said. “That is the very near future. In the present, this is something that I think will make research better and maybe clinicians’ jobs a little easier. But the future is really empowering a more personalized approach by maximally extracting data from these biosamples in such a way that modern computers let us do qualitative descriptions of histology.”