AI model predicts early-stage rheumatoid arthritis using MRI scans

It is now possible to predict rheumatoid arthritis (RA), thanks to interpretations of magnetic resonance imaging (MRI) scans made by an artificial intelligence (AI) model.

A recent study presented at EULAR 2023 reveals that the performance of AI is almost on a par with that of human experts. This novel visualization method confirms the significance of known inflammatory features and reveals new imaging biomarkers, which offer a different perspective of understanding RA.

“Predicting early RA from extremity MRI facilitates timely treatment, possibly preventing chronicity,” said the researchers, led by Yanli Li from Leiden University Medical Center, Division of Image Processing, Department of Radiology, Leiden, Netherlands. “Image interpretation by AI may provide more accurate predictions than visual scoring.”

Li and his team developed a deep learning AI method that automatically analyses extremity MRI scans to predict RA at an early stage. They collected scans of the hands and feet from 1,974 patients, of whom 1,247 had early onset arthritis (EAC) and 727 had clinically suspect arthralgia (CSA). In the EAC cohort, 538 developed RA in 2 years, while in the CSA cohort, RA developed in 113.

The researchers preprocessed the MRI scans automatically via background removal, slice-by-slice normalization, and central slice selection. They then pretrained a self-supervised deep learning model to fill in parts of the image that have been blinded by square patches. After transferring the resulting weights, they adjusted the model to predict RA development.

Model evaluation was done through fivefold cross-validation and in a held-out test set of 312 EAC patients and 146 CSA patients. The area under the receiver operator curve (AUC) was used to assess the accuracy of the model. Finally, the researchers developed an improved class activation map and applied this to indicate which areas were most significant to the AI decision.

Expert level

On the test set, the AI model that used MRI scans of the hands (wrist and metacarpophalangeal [MCP] joints) achieved a mean AUC of 0.683 in the EAC arm and 0.727 in the CSA arm. Models trained separately on the wrists, MCPs, and feet obtained a mean AUC of 0.679, 0.647, and 0.664, respectively, for EAC and 0.688, 0.669, and 0.715, respectively, for CSA. [EULAR 2023, abstract OP0002]

These accuracy levels were close to those of experts’ using RAMRIS-based prediction, with reported AUCs of 0.74 and 0.69 in predicting RA in CSA patients. [Arthritis Res Ther 2019;21:249]

Based on the proposed visualization method, the AI models could predict RA using similar patterns of known (teno-)synovial inflammation and bone marrow oedema.

“Automatic RA prediction with AI is feasible. It can perform close to the level of experts. It can be improved with more [clinical] data, but the generalization abilities require more data for evaluation,” said Li. “AI-based prediction could be reliable because AI also looks at known inflammatory signs.”

In addition, “the heatmaps may point to some new imaging biomarkers. That would require the help from clinicians [to be of use in the future],” he noted.