Background/Purpose: Rheumatoid arthritis (RA) is the most common form of inflammatory arthritis affecting a large sector of the global population. The disease is associated with a high socio-economic burden owing to disability caused and costs of treatment. Conventional diagnostic methods rely on the skill of trained clinicians in interpreting and prognosing the severity of RA from X-ray images. This is slow, costly, and often error prone. We propose novel AI models that can perform an automatic quantification of X-Ray damage.

Methods: Data is retrieved from the Prince of Wales Hospital in Hong Kong. Patients diagnosed with RA were recruited with physician’s diagnoses are recorded at the first clinic visit according to ICD9-CM. Inclusion criteria are 1) patients who carried the diagnosis code of RA (ICD9-CM: 714), 2) >= 1 follow up visit after disease onset, 3) serial XR hands available ( >= 1 set) after disease onset, 4) age >=18 at disease onset, and 5) DAS28 < 5.1. The final dataset consists of a cohort of ~600 X-ray images of hands and feet. Each X-ray image is labeled with a corresponding value for the van der Heijde Modified Sharp score (vdH score). Our trained deep learning model analyzes a typical X-ray image in two distinct ways:

1. Classification: we partition the dataset according to the vdH value into 11 disjoint sets. Each set corresponds to images with vdH score within a specific range. The classifier is then tasked with deciding the range of vdH values to be assigned to a typical X-ray image;

2. Regression: our trained model directly outputs a numerical value of the vdH score based on its analysis of a typical X-ray image. This task is computationally harder than classification as the model predicts continuous values and therefore can be used in disease progression analysis of X-rays collected over time.

These approaches are prefaced by preprocessing pipelines designed to take the intricacies of images into account. Our deep learning frameworks are based on the Visual Transformer (ViT) architecture. For regression, our framework can take prior information regarding RA severity into account, much in the same way that a physician performs annotation on an image. Prior information is integrated into the regression framework in the form of a trained shape model.

Results: We obtained accuracies between 70-75% on the validation dataset, for the 11-bin classification problem. We also computed the Area-Under-Curve score (AUC), which is a more accurate reflection on the robustness of our classification in the presence of imbalanced data. We obtained an AUC of 0.8. For the regression problem, we obtain relational errors of 0.25-0.3 (computed as fractional deviation of prediction from truths), competitive with a clinician.

Conclusion: We implement deep learning frameworks to automatically assign severity scores to X-rays of bones via classification or regression. With both frameworks we show that it is possible for deep learning prognostic pipelines to be robustly implemented and deployed for use in a conventional clinical workflow, thus paving the way for expedited prognosis and increased quality of care for RA patients.

« Back to ACR Convergence 2023

ACR Meeting Abstracts - https://acrabstracts.org/abstract/deep-learning-approaches-to-rheumatoid-arthritis-severity-prognosis/