Artificial Intelligence Speeds Identification of Failed Joint Components on X-Rays
Key findings
- When planning revision of total joint arthroplasty, a key step is to identify the failed hip or knee components
- Researchers at Massachusetts General Hospital trained a convolutional neural network (CNN) to analyze radiographs and distinguish among three different implant systems commonly used in total hip arthroplasty
- The CNN was tested on 25 radiographs that had been isolated from the training and validation process
- The CNN achieved 100% accuracy in classifying the radiographs, spending about five seconds per radiograph
When planning a revision of total joint arthroplasty, a key step is to identify the failed hip or knee components. Manufacturers are required to label devices with unique identifier numbers, which must be entered into medical records, but these rules were instituted only in the last few years. For the large number of patients who already had joint implants, surgeons usually use radiographs to identify the device in question.
Subscribe to the latest updates from Orthopaedics Advances in Motion
Alireza Borjali, PhD, research fellow in the Department of Orthopaedics at Massachusetts General Hospital, Kartik M. Varadarajan, PhD, principal investigator in the Technology Implementation Research Center, and colleagues have created a form of artificial intelligence that, in seconds, can identify hip implant designs using radiographic images. They describe their approach in the Journal of Orthopaedic Research.
Study Methods
The goal of the study was to create a deep convolutional neural network (CNN) to analyze the images. CNNs are sets of algorithms that analyze visual images, and deep machine learning indicates a network that can train itself to perform the task.
The researchers retrospectively collected postoperative anteroposterior radiographs of 252 patients who underwent primary total hip arthroplasty at Mass General in 2018 and received one of three different implant systems. They randomly divided the images into sets that were separately used for training, validation and testing of the CNN.
Training the CNN
The CNN refined its accuracy through a process in which it was presented with each of 198 radiographs 350 times, with slight variation each time (e.g., rotation or magnification).
Validating and Testing
In the validation subset of 29 radiographs, the retrained CNN was 100% accurate in identifying the type of hip implant system used. It was also 100% accurate in classifying the test subset of 25 radiographs, spending about five seconds per image.
Machine Learning Becomes Machine Teaching
In making its decisions, the retrained CNN looked at the tip of the stems of all three implant systems, among other features. On close examination, the researchers realized they, too, could differentiate between the implant designs just by looking at the stem tips.
The machine was never directly programmed to look at the stem tips. It learned to do so and to ignore the rest of the image.
Greater Efficiency, Better Patient Health
Manual identification of an implant from radiographs is time-consuming and prone to error. In some cases the device can't be identified preoperatively, which can lead to more complex surgery with increases in blood loss, bone loss, operating room time, recovery time and costs.
Integrating this CNN into orthopedic care could save a substantial amount of time and improve detection accuracy, potentially decreasing the cost of revision hip arthroplasty and even improving patient outcomes.
view original journal article Subscription may be required
Learn more about the Technology Implementation Research Center
Refer a patient to the Department of Orthopaedics