Roughly 300 patients receive kidney transplants at Mayo Clinic in Rochester annually, part of nearly 1,000 kidney recipients annually among all Mayo Clinic hospitals.
Caring for those patients is a team of physicians, nurses and pathologists all working together before, during and after surgery. Artificial intelligence might join that transplant care team in the near future.
In October 2022, Mayo Clinic researchers published a paper in Clinical Research detailing how an artificial intelligence program can find inflammation in biopsies from transplanted kidneys.
Rochester-based Mayo clinicians already perform routine biopsies on kidney transplant recipients in the months and years after they receive their new organ, even if the patients feel fine and don’t have any symptoms of organ rejection, said senior research technologist Walter Park.
“It really gives us an opportunity in the research realm to really understand the patients a lot better because we’re looking at almost every patient,” Park said. “Hopefully we can understand the mechanisms of what leads to a successful engraftment and failures.”
AI: Another set of eyes
Typically, a team of pathologists manually review the biopsies.
“If you have five pathologists — and that’s the standard of care currently — they would look at this (biopsy) and give it a score,” said Dr. Aleksandar Denic, a Mayo Clinic researcher and assistant professor. Pathologists use a 0-3 scale, with 0 meaning minimal or no damage and 3 indicating severe damage. “Those five pathologists will generally agree with 0 and 3, but 1 and 2, it’s a flip of a coin, and they can disagree.”
The AI under development at Mayo Clinic can act as an extra set of eyes on a biopsy, helping pathologists figure out what’s going on inside that patient’s new kidney. It can also save pathologists some time, said biostatician Byron Smith.
“Counting the amount of information is very tedious, very time-consuming and our pathologists are really stretched,” Smith said. “This is kind of a way to generate more robust, unbiased data.”
Clinicians retrieve kidney biopsies by making a small incision on a patient’s abdomen and using a thin needle to collect a 1 centimeter-long piece of kidney tissue about the size of a piece of mechanical pencil lead. That sample is then submerged in a small block of paraffin wax, and about a dozen extremely thin slices are taken from the block, which are then stained and analyzed by pathologists.
A lot of information
“That’s a 3- or 4-micron slice of that piece of tissue,” Park said. “It’s really tiny.”
But there’s a lot of information to be found in those tiny samples. Each digitized biopsy image can contain hundreds of megabytes or even a gigabyte’s worth of data — for reference, an iPhone 14 with the largest storage capacity available, 512GB, would only be able to store the images for around 50 biopsies. Mayo performs about 7,000 kidney biopsies a year in Rochester.
Because these biopsy images are so large, Smith said that he and his team cut the image in to multiple parts so the files are small enough for the AI to analyze.
“In particular, what we want to do is identify specific objects on the images,” Smith said.
Those specific objects include the glomeruli, structures in the kidney that filter waste out of the bloodstream. They show up as round dots within the biopsy tissue.
“If you have damage within your glomeruli, then that means you have damage generally within your kidney and you’re going to have kidney trouble — kidney failure or some sort of elevated lab values,” Smith said.
The AI program also spots signs of inflammation in the tissue, which could indicate that a patient’s immune system is attacking the donated kidney.
“The presence of inflammation is actually a really big deal in transplant,” Park said.
Extra evaluation
The AI’s findings could help clinicians evaluate whether or not treatment to prevent rejection is working in a particular patient. The AI’s analysis of the biopsy images can be stored and reviewed again to help providers see if a patient’s condition is improving or worsening over a period of weeks or months. Smith adds that the AI could also point to issues in the donated kidney before tissue damage and physical symptoms show up.
“A lot of centers will do for-cause (biopsies), which means that a person will get a blood draw or some sort of urine test,” Smith said. “They’ll have elevated protein levels … but that means the damage has been done. By doing protocol biopsies and looking at the amount of inflammation or things like that, we try to preempt that — you want to treat or intervene in some way before the damage is already done.”
Of course, the algorithm didn’t automatically know how to spot inflamed tissue and glomeruli. Denic and other researchers spent hours upon hours reviewing thousands of biopsy images, tracing the glomeruli to feed that information to the AI and teach it how to identify those structures.
Using this tool
“You can’t just have massive computer power and images to make this work. You need the morphometry to really say, ‘This is the gold standard of what a glomeruli is,'” Park said. “It’s not as simple as, ‘We’ll throw a bunch of computer scientists at this and it’ll work.’ The clinical information is critically important.”
Now that Smith and his team have shown that the AI is an effective tool, the next step is making sure that clinicians will actually be able to use the tool as they care for patients.
“One of the things that we’re discovering as we try to transition from research to the actual clinical practice is that if AI is not developed in a user-friendly way that the pathologists are interested in, they won’t do it,” Smith said. “It’s even more work for them to apply something that we thought was helping them.”
Digitized biopsy images have already become the standard of care for Mayo Clinic’s pathologists, said transplant surgeon and researcher Dr. Mark Stegall, so applying the AI program could be the next step. It’s part of what Stegall sees as the future of medical technology.
“We’re at the point where I think we’re being more quantitative and using numbers more in the management of patients,” Stegall said. “That’s a good thing, but that doesn’t mean the computer’s making the decision about the patient. … The computer just gives you a lot more information.”
Stegall said algorithms like the program Smith and his team have developed could become as commonplace as tools such as CT scanners — technology that was once unavailable but has now become run-of-the-mill.
“You wouldn’t practice medicine without these things, right?” Stegall said. “I think these automated tools are going to be the same thing because you’re going to want that data.”
