鶹ý

Uh-Oh, Is a Robot Coming for Your Radiographic Analysis Job?

— Computer matches live expert in measuring joint space and erosions

MedpageToday

SAN DIEGO -- Joint space narrowing (JSN) and erosions in rheumatoid arthritis (RA) were measured and graded about as well by computerized x-ray analysis as by a human rheumatologist, a researcher reported here.

With the live physician's judgment serving as the reference, the machine-vision system was 92% accurate in determining for hand and wrist joints, according to Carol Hitchon, MD, MSc, of the University of Manitoba in Winnipeg, speaking at the American College of Rheumatology annual meeting.

In practice, she told 鶹ý after her talk, the 92% figure is far greater than the typical level of agreement between two human experts in grading joint x-rays -- or even by a single expert reviewing the image on different days. It was entirely possible that, in some cases, the machine system had provided the more accurate score.

But rheumatologists who specialize in x-ray analysis don't need to worry about their jobs. Not only is her group's system not ready for real-world routine use -- the current version has no learning ability and its interface isn't "clinician-friendly" -- but JSN and erosion detection is only a small part of the RA diagnostic process. The system does nothing with other joint structures that are at least as important in RA pathology.

On the other hand, Hitchon noted that just the business of assigning SVH scores to a hand or wrist joint is difficult and time-consuming. This quantitative scoring for JSN and erosions is critical in objectively tracking disease progression. Rheumatologists should appreciate having it done in a matter of seconds as accurately as they can do it themselves, she suggested. Moreover, availability of machine-based analysis should improve RA patient care in areas where rheumatologists trained in SVH scoring are scarce.

The system was developed in the typical way for machine-vision systems, with initial training on different sets of x-rays for which SVH scores had been assigned, and then validating it on additional image sets. For the current work, some of these were from pediatric patients and some from adults, so as to maximize the system's flexibility and scope. (For what it's worth, Hitchon's group included Désirée van der Heijde, MD, PhD, the Leiden University researcher for whom the scoring is co-named.)

Hitchon and colleagues built the system from the ground up, first training it just to determine what is and what is not a joint in a radiograph, and then to quantify the structures in order to distinguish abnormal from normal. Validation was conducted with 54 images from four adult RA patients who had at least 10 expert-graded radiographs taken over a decade or so.

Accuracy was then checked by running more than 2,200 hand x-rays, from a Canadian cohort of early-RA patients (less than 1 year from symptom onset). It was measured with the F1 statistic, which counts correct predictions across a whole dataset. Hitchon and colleagues also calculated the root mean square error (RMSE; lower values = greater accuracy) as another gauge.

F1 values were 0.92 for SVH scoring of erosions and also for JSN in hand joints. RMSE values were 0.36 and 0.40, respectively. The system was somewhat less accurate for wrist joints, with F1 values around 0.7 and RMSEs above 1.0 for erosions and JSN.

Hitchon said her group expects to be able to add a learning ability to the system without having to start over. They also hope to improve the interface, such that a clinician can simply upload routine patient images for automated scoring.

  • author['full_name']

    John Gever was Managing Editor from 2014 to 2021; he is now a regular contributor.

Disclosures

The study had no commercial funding.

Authors including Hitchon reported extensive relationships with industry.

Primary Source

American College of Rheumatology

Hitchon C, et al "Artificial intelligence models for computer-assisted joint detection and Sharp-van der Heijde score prediction in hand radiographs from patients with rheumatoid arthritis" ACR 2023; Abstract 0745.