Radiologists and ortho docs make scales to classify how bad structural damage is to a knee. But how does that match the pain felt by a patient? Can CNNs see things humans miss?

We started eager. What if we give a CNN no preconceptions and just ask it to learn how to predict patient pain just from an X-ray. The result was this paper:

An Algorithmic Approach to Understanding Osteoarthritic Knee Pain

Brandon G. Hill, Travis Byrum, Anthony Zou, Peter L. Schilling

JBJS Open Access  ·  01 Oct 2023  ·  doi:10.2106/JBJS.OA.23.00039

How much of a osteoarthritis patient’s pain is structural? This work moves beyond traditional/human derived biomarkers of knee health and sees how much pain can be predicted by a CNN looking at a knee X-ray.

deep-learning pain osteoarthritis medical-imaging oai

From there, we wondered if a CNN could see more pain in an MRI. But before diving into all the work to train on MRIs, why don’t we look at how well structural features determined by humans (i.e., trained radiologists and orthopaedic surgeons) to indicate knee health corresponds to patient pain? That led to our next paper (waiting for publication date).

It seems that neither humans or computers can find structures in the knee that are predictive of pain. If you are a clinician and thinking the human side of this is obvious, it is known. It just hasn’t been confirmed statistically on this scale (almost 5,000 patients).