Multistage deep learning methods for automating radiographic sharp score prediction in rheumatoid arthritis

Area of Science:

• Radiology
• Artificial Intelligence
• Rheumatology

Background:

• The Sharp-van der Heijde score (SvH) is essential for quantifying joint damage in rheumatoid arthritis (RA) from radiographic images.
• Manual scoring of radiographic images for RA assessment is laborious and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate a multistage deep learning model for automated prediction of the Overall Sharp Score (OSS) from hand X-ray images in rheumatoid arthritis patients.
• To establish an efficient and objective method for assessing joint damage in RA, potentially reducing reliance on manual scoring.

Main Methods:

• A four-stage deep learning framework was employed: image preprocessing, hand segmentation using UNet, joint identification with YOLOv7, and OSS prediction via a custom Vision Transformer (ViT).
• The model was trained and validated using stratified group 3-fold cross-validation on a dataset of 679 patients and externally tested on 291 subjects.
• Performance was evaluated using metrics such as Intersection over Union (IoU), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Huber loss, and Intraclass Correlation Coefficient (ICC).

Main Results:

• The joint identification component achieved 99% accuracy.
• The Vision Transformer (ViT) model demonstrated strong performance in OSS prediction, particularly for patients with lower disease burden (Sharp scores < 50).
• Achieved a Huber loss of 4.9, RMSE of 9.73, and MAE of 5.35, with a significant correlation to expert scores (ICC = 0.702, P < 0.001).

Conclusions:

• This study introduces the first application of a Vision Transformer (ViT) for Overall Sharp Score (OSS) prediction in rheumatoid arthritis.
• The proposed automated deep learning approach provides an efficient and reliable alternative for assessing joint damage in RA.
• This methodology has the potential to streamline the evaluation process and reduce the subjectivity associated with manual scoring of radiographic images.