AI Accurately Identifies Both Primary and Revision Arthroplasty Implants From Plain Radiographs
Key findings
- This study aimed to develop a convolutional neural network (CNN) capable of identifying primary and revision arthroplasty implants from both total hip and total knee procedures within a single system
- The CNN was trained on 8,963 radiographs and validated on 2,241 radiographs, representing 24 brands of THA implants (17 primary, seven revision) and 14 brands of TKA implants (eight primary, six revision) from four manufacturers
- In the validation phase, the CNN was at least 96% accurate in predicting the manufacturer and brand of a patient's implant
- The findings suggest the CNN has great potential to assist in preoperative planning for revision or re-revision arthroplasty
Identifying an implant before revision arthroplasty is a key step in optimizing outcomes, especially when there will be an exchange of modular implants. This process currently requires a manual review of surgical reports, which is time-consuming and particularly difficult if the primary procedure occurred at a different institution.
Subscribe to the latest updates from Orthopaedics Advances in Motion
This challenge routinely leads to delays in care, increased perioperative morbidity, and greater costs. Christian Klemt, PhD, a former research fellow in the Bioengineering Laboratory at Massachusetts General Hospital, Akachimere Cosmas Uzosike, MD, MPH, a resident in the Harvard Combined Orthopedic Residency Program, Young-Min Kwon, MD, PhD, director of the laboratory and the vice chair in the Department of Orthopaedic Surgery, and colleagues developed a potential solution: a convolutional neural network (CNN).
As they report in the Journal of the American Academy of Orthopaedic Surgeons, the CNN proved to be highly accurate in identifying implants used in both primary and revision total joint arthroplasty.
Methods
A deep learning CNN is a form of artificial intelligence that's trained on many medical images and "learns" how to interpret additional images. For this project, the researchers collected 11,204 plain anterior–posterior radiographs from 8,763 patients with primary or revision total hip arthroplasty (THA) or total knee arthroplasty (TKA).
24 THA brands (17 primary, seven revision) and 14 TKA brands (eight primary, six revision) from four manufacturers were included. For each design, 80% of the radiographs were randomly allocated to be the training dataset (total n=8,963), and the other 20% served as the validation set (total n=2,241).
Overall Model Performance
In each category, the performance of the CNN in identifying the manufacturer and brand of implants was:
- Primary THA—Accuracy, 98%; sensitivity (true positive rate), 96%; specificity (true negative rate), 99%
- Primary TKA—Accuracy, 97%; sensitivity, 95%; specificity, 98%
- Revision THA—Accuracy, 98%; sensitivity, 95%; specificity, 98%
- Revision TKA—Accuracy, 96%; sensitivity, 95%; specificity, 98%
Model Performance for Specific Implants
- Primary THA—Accuracy of 96%–100%
- Primary TKA—Accuracy of 93%–100%
- Revision THA—Accuracy of 95%–99.6%
- Revision TKA—Accuracy of 93%–100%
Commentary
Although other deep learning systems have been developed for identifying arthroplasty implants, this project CNN was trained on the largest sample of THA and TKA implants to date. It is also the first system automatically identifying revision THA and TKA designs from plain radiographs.
The findings suggest the great potential of the validated deep learning model to facilitate preoperative surgical planning of revision and re-revision arthroplasty.
view original journal article Subscription may be required
Learn more about the Bioengineering Lab
Refer a patient to the Hip & Knee Replacement Service