AI Analyzes MRI Images to Help Predict Rheumatoid Arthritis
Images
Researchers from the Netherlands presented an abstract at EULAR, the annual meeting of the European Alliance of Associations for Rheumatology, demonstrating the use of artificial intelligence (AI) to interpret MRI images and provide more accurate predictions than visual scoring in early rheumatoid arthritis (RA).
Early inflammatory arthritis is often undifferentiated, but it may develop into established RA or another arthropathy. Alternatively, it may resolve spontaneously, or remain undifferentiated for indefinite periods. Erosion is a key prognostic factor which can be detected with MRI. In addition, MRI allows direct visualization and assessment of (teno-) synovitis and bone marrow edema.
Predicting early RA from MR images of the hands and feet can help people access timely treatment, which may possibly change the disease course. Traditionally this is done by radiologists and rheumatologists using a scoring sheet to manually identify key features from the MRI scans.
An abstract presented by Li and colleagues from Leiden University Medical Center details how deep-learning AI can automatically analyze scans in order to predict RA at an early stage in patients with clinically suspect arthralgia. The model was first trained to understand anatomy, then to distinguish patients from healthy controls, and finally to find image features predictive of RA development. The AI analyzed scans of the hands and feet from 1,974 people with either early-onset arthritis or clinically suspect arthralgia, of whom 651 went on to develop RA. Results from a held-out test set showed the model could predict RA with accuracies close to those achieved by human experts. This worked for scans of either hands or feet.
The authors conclude that AI interpretation of MRI scans could provide automatic RA prediction. Further training for the model using MRI data from healthy controls will probably improve the accuracy, and future research will focus on predicting RA in specifically undifferentiated arthritis, as a subgroup of the early onset arthritis group. Additionally, this new method not only confirms the significance of known inflammatory features such as synovial inflammation, but in the long term may also be able to identify new imaging biomarkers, further enhancing our understanding of the underlying disease process in early RA.