Using Deep Learning AI to ‘Read’ Hip/Knee Implant Scans

Computers can “see” better, faster and with less bias than humans. For medicine, that can be a game changer.

A team from the Bioengineering Department, Department of Orthopaedic Surgery at Massachusetts General Hospital/Harvard Medical School decided test a certain type of deep learning algorithm on CT scans of large joint implants.

The resulting study, “The Ability of Deep Learning Models to Identify Total Hip and Knee Arthroplasty Implant Design From Plain Radiographs,” appears in the May 1, 2022, edition of The Journal of the American Academy of Orthopaedic Surgeons.

Notably, the team did NOT compare their deep learning algorithm results with human CT scan readers.

Study co-author, Young-Min Kwon, M.D., Ph.D. professor of Orthopaedic Surgery at Harvard Medical School and vice chair in the Department of Orthopaedic Surgery at Massachusetts General Hospital, explained the purpose of this study to OTW: “Total hip arthroplasty (THA) and total knee arthroplasty (TKA) are commonly performed procedures, with over 1 million TJAs performed annually in the U.S. alone. However, complications such as infections and aseptic loosening may ensue, requiring revision surgeries.”

“Implant identification is a critical aspect of preoperative workup in this subset of patients and is currently performed manually by reviewing patient records and operative reports. It is also complicated by the fact that most patients undergo revision TJA at a different center than that of their primary surgery. Current procedures of implant identification are cumbersome and labor-intensive, and we aimed to devise a more time-efficient and cost-effective modality for pre-operative implant identification.”

“Artificial Intelligence Deep Learning is an emerging technology that employs advanced computational algorithms to analyze and interpret extensive and complex datasets with minimal human input. Therefore, we developed and validated a convolutional neural network deep learning model to identify commonly used primary and revision hip and knee designs from plain radiographs in our study.”

The team’s own convolutional neural network deep learning model was trained on 11,204 radiographs (8,963 radiographs, 80%, for model training and 2,241 radiographs, 20%, for model validation) to identify 24 THA designs and 14 TKA designs.

Overall, Dr. Kwon told OTW, “The convolutional neural network deep learning model was able to identify primary THA implants with an overall accuracy of 98.2% (sensitivity = 95.8%, specificity = 98.6%, and AUC = 0.98) and primary TKA implants with an accuracy of 97.4% (sensitivity = 94.9%, specificity = 97.8%, and AUC = 0.97).”

For revision THA designs, “The model was able to correctly predict 98% of implants, with a sensitivity of 94.9%, specificity of 98%, and an AUC of 0.98. In addition, 96.3% revision TKA implants were accurately predicted, with a sensitivity of 94.5%, specificity of 98.1%, and AUC of 0.96.”

In conclusion, said Dr. Kwon, “Our findings demonstrated excellent accuracy of the deep learning model for the identification of 24 THA and 14 TKA commonly used implant designs from plain radiographs. This emphasizes the potential of artificial intelligence deep learning models to assist in preoperative surgical planning of failed arthroplasty patients by accurately identifying these implants, thereby avoiding the current cumbersome and labor-intensive manual implant identification process, and improving the patient outcomes.”