Mayo researchers say patients will benefit from faster, more accurate reads

By Jay Furst

Artificial intelligence may never replace human insight, expertise and judgment. Then again, no one's complaining about how it can help, especially when it comes to analyzing kidney biopsies, a laborious process that's an intricate mix of art and science. Accurate and thoughtful interpretation of biopsies is of life-changing importance for patients.

"The kidney is a pretty complicated structure, and rating the biopsy, taking that data and using it to assess what's already happened and what will happen is complicated," says Mark Stegall, M.D., a Mayo Clinic transplant surgeon. "When a pathologist reads a biopsy, he or she is looking mostly for an overall diagnosis and not detailed quantification of changes in every single part of the biopsy. However, there's a lot of data that is lost that could be helpful in predicting outcomes if we had a way of measuring it."

That's where a deep neural network can help.

The development of deep or convolutional neural networks — a type of machine learning, otherwise known as artificial intelligence (AI) — has made it possible for researchers to conduct more advanced analyses of medical images. Neural networks need human training, however. They need to know what to look for and analyze, and they take vast amounts of computing capacity and storage space.

An international team of researchers, including Dr. Stegall and Mayo Clinic colleagues, recently "trained" a network for analysis of kidney transplant tissue as part of a study that was published in the Journal of the American Society of Nephrology. "This is the first publication of using AI to read kidney transplant biopsies — and certainly with the degree of completeness that we can look at certain features of the biopsy," says Dr. Stegall, a West Texas native who joined the Mayo Clinic staff 21 years ago and has been gathering kidney biopsy samples virtually since he arrived.

"Here at Mayo Clinic we have a long tradition of following patients, so we have something like 17,000 kidney biopsies that we've done with transplant patients who we've followed for up to 10 years," says Dr. Stegall. "We didn't know it at the time, but we were one of the first collectors of 'big data' in our transplant patients."

Those biopsies are a gold mine of data for researchers looking for clues to what's happening in the kidney before, during and after transplant. "Leonardo da Vinci said that in a painting, you don't only see the present moment, you can see the past and also sort of predict what will happen in the future. A biopsy is like that," Dr. Stegall says. "You can see something that's already happened and you can observe features that will help you predict what's going to happen next."

The trick is to read it correctly and completely every time. Dr. Stegall and his Mayo colleagues worked with researchers from Radboud University Medical Center in Nijmegen, Netherlands, to develop and test a convolutional neural network that was capable of advanced analysis of digitized kidney tissue samples.

Training a machine takes time, content and patience

They trained the computer program using 40 images of stained kidney transplant biopsies that had been marked for abnormalities in the biopsies and then used it to analyze 10 tissue classes in 10 transplant biopsies from Radboud and 10 from Mayo Clinic. Walter Park, a senior research technologist in Dr. Stegall's lab who worked on this project, says the Radboud team developed the computer algorithm that conducted the analysis, and the Mayo group is working on complementary algorithms for specific transplant lesions, inflammation and other features.

Training the computer takes content, though, and lots of it, which is where Mayo Clinic's deep archive of kidney transplant biopsies makes all the difference, along with the expertise of Mayo Clinic staff. Dr. Stegall was a study co-author, as were Mariam Priya Alexander, M.D., associate professor of laboratory medicine and pathology at Mayo Clinic College of Medicine and Science, and Byron Smith, Ph.D., a biomedical statistician and computer scientist.

The biopsies have to be prepared, segmented and scanned into the computer for hours on end. "Any training run takes six to eight hours per training set," says Dr. Smith. "Each time a training model is created, it's generally built on a set of user-input 'hyperparameters,' settings that control the behavior of an algorithm and can dramatically impact performance. For this reason, many, many hyperparameters are tested to assure the best possible performance."

In this case, the work was complicated further by differences in protocols and equipment in the Netherlands and the U.S. "Our lab group has different staining protocols and a different slide scanner that creates images in a completely different format," says Dr. Smith. "Despite these differences, the model performed well across major renal compartments."

The weighted mean coefficients, or accuracy, in all classes were 0.80 in the 10 biopsies from Radboud and slightly better, 0.84, in the 10 from Mayo. The neural network excelled at identifying glomeruli, the clusters of small blood vessels at the tip of a nephron in the kidney, achieving nearly 95% accuracy in that area.

In addition to biopsies, the neural network detected 92.7% of all glomeruli in kidneys surgically removed from a patient for other reasons, with 10.4% false positives. Overall, "we found significant correlations between visually scored histologic components and network-based measures," according to the study.

There's more training to be done to take the neural network beyond the basics. You might say it has a good college education at this point, but now it's on to graduate school. "Damaged kidney tissue, such as sclerotic glomeruli and other indicators of chronic kidney disease, were not well-represented in the training set, and not surprisingly, the computer model did not identify these objects well," says Dr. Smith. "We're working on improving those areas."

Dr. Alexander says another step for this project is to delineate the border between cortex and medulla within a biopsy, which will make calculations of glomerular density possible. Glomerular density is an important indicator of the health of the kidney. Decreasing glomerular density has been noted with increasing age, obesity and increasing serum uric acid levels.

AI assists but doesn't replace human decision-making

Going forward, the hope is that pathologists will be able to use the computer analysis to sort out or screen biopsies that need further attention. "The two criteria that we're using are, if it's really hard or tedious for a pathologist to read the biopsy, we're going to see if we can fix that right off the bat," Dr. Stegall says. "And then there are things we know are vitally important in terms of predicting outcome for kidney transplant patients, and we can train the neural network to look for those features."

Two years from now, it may be that the computer can process the scanned slide and do part of the tedious, repetitive and quantitative analysis for the pathologist, depending on the adequacy of the biopsy. "The long-term goal is to have the convolutional neural network read as much of the biopsy as possible," he says. Since a kidney biopsy typically includes about 10 slides, this could save time and improve both accuracy and productivity.

Both the quicker read and the accuracy can directly improve the quality of care for patients, he says. It also allows pathologists to channel their time and effort on the most diagnostically challenging cases.

Deep learning-based analysis also can speed up clinical trials and get new therapies and drugs through the regulatory process to help patients. "The more accurately we're able to score biopsies, the more valuable the biopsies are in predicting outcomes and the more efficiently we can move clinical trials and research forward," Dr. Stegall says. "That's what it's all about. At the end of the day and the end of my career, it would be nice to know that we have improved the reading of kidney biopsies, which allows us to identify patients who are at high risk for graft loss (a transplanted kidney that ultimately fails) and to come up with interventions for those patients."

Though there's some irony involved in teaching a machine to potentially replace the teacher someday, Dr. Smith says he's not overly concerned about that in the near term.

"There's been a lot of discussion about artificial intelligence taking jobs and to what extent staff may be replaced by computer algorithms," says Dr. Smith. "In my opinion, that won't happen in medicine for a very long time. Most deep learning methods are very specific to certain programs and they fail to make logical jumps that humans can reach. These algorithms are excellent when you can apply a mathematical logic, but it would be difficult if not impossible to capture the intuition and experience that a renal pathologist can bring."

Neural networks can rearrange the workload, though, as well as reallocate resources and help pathologists make more objective decisions.

An editorial in the Journal of the American Society of Nephrology regarding the Mayo-Radboud study used the term "augmented intelligence" to describe the combination of machine learning and human expertise. That expertise only comes from deeply human interactions among doctors, nurses, other medical staff and patients, and all the learning those human interactions generate.

"The key always in any of this stuff is to have people do what only people can do," Dr. Stegall says. "I perceive AI as an enhancement of what we're already doing. We know a lot. We can know a lot more, and more accurately, with AI. But we need to understand the technology so we as clinicians can apply it correctly to our patients."

###

Tags: Aleksandar Denic, artificial intelligence, Byron Smith, collaboration, Findings, Innovations, kidney disease, kidney transplant, Mariam Priya Alexander, Mark Stegall, nephrology, News, pathology, team science, transplant, Walter Park