Artificial intelligence, machine learning may identify TJA implants prior to revisions
Click Here to Manage Email Alerts
GRAPEVINE, Texas — Researchers at the Cleveland Clinic Machine Learning Laboratory trained deep learning artificial intelligence algorithms to correctly identify arthroplasty implants from plain radiographs, a presenter said.
At the American Association of Hip and Knee Surgeons Annual Meeting, Viktor E. Krebs, MD, discussed one way that artificial intelligence (AI) can impact orthopedic practice during the computer vision portion of a symposium on the ways AI may advance arthroplasty.
With the burden of hip and knee arthroplasty revisions expected to increase dramatically, orthopedic practices and staff will be spending more time and money identifying in situ total hip arthroplasty (THA) and total knee arthroplasty (TKA) implants in patients prior to revision procedures. “In these situations, we’ve got to identify the implants,” Krebs, of the Cleveland Clinic, said.
“Implant identification, if you don’t identify immediately, can take up to 30 minutes of staff time on average and/or getting implant records. And then there’s 10% of implants that could not be identified preoperatively, even by the experienced eye,” Krebs said.
Enter computer vision, which is “ ... is a function of AI by which machine learning and complex neural networks can be applied to enable computers to capture, analyze and interpret information from clinical images and visual inputs,” he said.
Krebs and colleagues put computer vision to work, developing an implant identifier and training it to identify features of THA and TKA implants from a registry. They then had the AI system identify implants from the plain radiographs of real patients.
The algorithm used focuses on “ ... areas of the implant that can be identified exactly by the computer and, using a heat map, the AI computer vision goes through the process” and then focuses on a particular implant model when the process is done, Krebs said.
During initial work in the lab during the machine learning process using TKA implants, Krebs and colleagues captured nine TKA implant designs after working with 682 anteroposterior (AP) knee radiographs.
For comparable work the group did more recently for THA, they started with more images, 1,972 AP radiographs in all, according to Krebs.
“We captured 18 different femoral stems, and this was our first attempt at internal validation of the machine learning algorithm. A similar process was used, and heat maps of models focused at the end,” he said. “The results: Our accuracy on the hip was 99.6%, specificity 99.8% and total sensitivity 94.3%. So there was a clear positive correlation, in the end, with the number of images used for training in the registry with the sensitivity.”
Researchers have taken these developments a step further through external validation that yielded similar information from 2,954 radiographs on eight distinct hip femoral stem designs, Krebs said.
“Our limitations are due to poor performance with rare implants that the system is not trained on and that we don’t have in our registry,” he said. “Future studies will include focusing on performance with more images through additional multicenter studies and then development of a web-based application that can be used to provide this tool to orthopedic surgeons around the globe.”
References:
Karnuta JM, et al. J Arthroplasty. 2021;doi:10.1016/j.arth.2020.11.015.
Collapse
Krebs VE. Implant identification. Presented at: American Association of Hip and Knee Surgeons Annual Meeting; Nov. 3-6, 2022; Grapevine, Texas.
Click Here to Manage Email Alerts