Artificial intelligence: Potential to improve knee arthritis care

Feb. 16, 2021

Although artificial intelligence is upending medical care, its potential applications to orthopedic surgery haven't been widely studied. Mayo Clinic found that a deep learning algorithm can identify and classify knee osteoarthritis on radiographs as accurately as fellowship-trained knee arthroplasty surgeons.

"This technology has the potential to significantly decrease the likelihood of inaccurate assessment of radiographs in the diagnosis and treatment of knee arthritis," says Adam J. Schwartz, M.D., an orthopedic surgeon at Mayo Clinic in Phoenix/Scottsdale, Arizona. "There is currently no standardized approach, and quite a bit of variability, in reading these radiographs. Reducing that variability can facilitate clinical decision-making and improve outcomes for patients."

The researchers trained a convolutional neural network to identify critical aspects of radiographs indicating knee osteoarthritis, and to rate the severity of the condition using the International Knee Documentation Committee scoring system. Four orthopedic surgeons were then asked to evaluate 576 knee radiographs taken from consecutive patients who made routine visits to Mayo's orthopedic clinic.

The surgeons' ratings were compared to one another and to the neural network's ratings of those 576 knees. Statistical analysis found broad agreement between the assessments from the surgeons and the artificial intelligence tool.

"A convolutional neural network can accurately identify the critical components of a standing posterior-anterior flexion knee radiograph without human control," Dr. Schwartz says. "In many instances in our study, the neural network correlated to a surgeon more accurately than many surgeon-to-surgeon comparisons."

Deep visual learning

A convolutional neural network is a type of deep learning tool that is commonly used to evaluate visual imagery. Deep learning is a method of artificial intelligence that processes raw data — such as images — with learning algorithms to create layers of nodes, each receiving information from, and learning from, the other.

"Before implementing this type of solution in clinical practice, we need more data and additional training, to minimize the potential for errors in classifying the severity of osteoarthritis," Dr. Schwartz says. "But once you have an artificial intelligence model, you can find all sorts of ways to improve efficiency and reduce variability."

Mayo Clinic's ability to perform this type of cutting-edge research rests on the center's multidisciplinary expertise. In addition to Dr. Schwartz and other orthopedic surgeons, Matthew R. Neville, M.S., a Mayo Clinic biostatistician with experience in full-stack web development and artificial intelligence techniques, participated in the project.

"Developing convolutional neural network models takes quite a lot of data," Mr. Neville says. "To assist with data collection, we designed a variety of web-based tools that made it relatively painless for the surgeons to assess radiographs."

This multidisciplinary approach facilitates research that ultimately benefits patients. "Combining expertise allows for the sum to be greater than each part," Dr. Schwartz says. "We see this technology as facilitating shared decision-making by patients and physicians about surgical and nonsurgical interventions."