Automatic Implant Identification of Orthopedic Implants using Artificial Intelligence

This project aims to train an AI algorithm to recognize different types of joint replacements from radiographs (X-ray) so that they can be quickly and accurately identified before corrective surgery.
Sometimes patients who have had joint replacements will experience problems requiring revision of the joint replacement. Revision total joint arthroplasty (TJA) is a technically challenging procedure associated with increased resource utilization and perioperative risk compared with primary TJA.
Sometimes the original procedure may have been years ago with the patient's notes at another hospital or the procedure was done outside of the hospital, it is then very difficult to know exact make and model simply from its X-ray appearance. Of course, surgeons are skilled and familiar with many popular makes and models of implants and can safely proceed with the surgery by having replacement parts ready before the surgery. If an accurate identification cannot be made prior to surgery, then no identical replacement parts may be available, the surgery may be performed only to reduce the symptoms or require the reuse of old-implant parts.
Implant parts vary and are not generally interchangeable. Numerous different makes and models of implants, often with small but important differences are available. An unidentified model is seen in the X-ray that will delay surgery making it difficult for the correct replacement parts to be used. Delay in the identification and sourcing of replacement parts leads to poor outcomes for the patient necessitating further surgery, with increased costs to hospitals and pain and discomfort to the patient. The AI implant identification algorithm works similarly to a smartphone's face recognition does: the surgeon will upload X-ray images of a person's knee or hip to our identification website which will either definitively identify the implant from its X-ray appearance or a small gallery of similar implants. This identification will be almost instantaneous. However, the AI algorithm will need to be trained on 1000s of examples to recognize the look of a variety of implant makes and models. A prototype is built which has been proven to be more accurate than surgeons on 12 different implant models. The project will build, test and deploy the next versions of the application to recognize over 100 implant models used in several joints: knee, hip, shoulder, and ankle, and it will be developed into a system easily accessible to doctors. This requires many 1000s of images from the UK and other parts of the world, as well as working with implant manufacturers to obtain the model designs. 

This project will be in collaboration with NHS hospitals including Luton & Dunstable University Hospital.

Skill required

Machine Learning, Deep learning, image processing and programming

Further reading

Artificial Intelligence based identification of Total Knee Arthroplasty Implants