Despite John’s wincing every time his leg was stretched, the crepitation that reverberated across the room, and the clear bone-on-bone osteoarthritis displayed by the x-ray, John still did not want the surgery. “My knee is fine! How do you know that I absolutely need surgery? I feel like you just want to cut into me to get more money!” he said, growing frustrated.
Surgeons have often been accused of “incision syndrome,” the practice of pushing through conservative treatments and opting for more invasive procedures (i.e., surgery). The avoidance of medical care due to a lack of understanding and potential mistrust of medical authority inspired us to partner with surgeons from Joint Implant Surgeons (JIS) to develop an AI algorithm that has the potential to help both surgeons and patients identify patients that are candidates for total knee arthroplasty (TKA), unicondylar knee arthroplasty (UKA), or are not arthroplasty candidates based on three x-ray images of the knee.
In this work, recently published in the Journal of Arthroplasty, we first cleaned and curated a dataset of thousands of knee x-rays using our network of labelers on the DiagnosUs app. We then trained an AI model that ingests the images from three different views and outputs a prediction of whether or not the patient had eventually gone on to have a successful UKA, TKA, or no surgical intervention at JIS. Future annotation and model development could continue to improve the accuracy and robustness of the model – especially since it was trained only on data from JIS – but the model proved extraordinarily accurate on a holdout test set: the ROC AUC scores were 0.97, 0.96, and 0.98 for TKA, UKA, and no surgery, respectively.
We believe that this model could add value in a number of ways. For example, it could provide a more objective “second opinion” for patients to consider alongside a surgeon’s subjective recommendation, help surgeons identify patients who may be candidates for less common UKAs, and help non-specialists identify which patients might be candidates for surgical procedures they need. More broadly, this work is a testament to the value contained in datasets like those provided by the surgeons at JIS, and the potential for AI to improve healthcare.
Read the full article here.
Understand why traditional labeling pipelines are so hard to scale and learn how our solution can 10X your labeling pipeline in a shorter time frame and with higher accuracy