Vision transformer assisting rheumatologists in screening for capillaroscopy changes in systemic sclerosis: an...

Area of Science:

• Ophthalmology and AI
• Rheumatology and Medical Imaging
• Digital Health and Diagnostics

Background:

• Systemic sclerosis (SSc) diagnosis relies on identifying microangiopathy via nailfold capillaroscopy (NFC).
• Automated analysis of NFC images using artificial intelligence (AI) can potentially improve diagnostic accuracy and consistency.
• Vision Transformers (ViT) represent a powerful deep learning architecture for image analysis tasks.

Purpose of the Study:

• To implement and evaluate an 'off-the-shelf' Vision Transformer (ViT) deep learning model for detecting microangiopathy in SSc patients' NFC images.
• To compare the diagnostic performance of the ViT model against that of experienced rheumatologists in classifying NFC findings.

Main Methods:

• Analysis of 17,126 NFC images from 289 SSc patients (EUSTAR and VEDOSS registries).
• Cross-fold validation was used to assess the ViT's classification performance for disease-associated capillary changes and the scleroderma pattern.
• A reliability set of 464 images was used to compare ViT performance against a panel of rheumatologists.

Main Results:

• The ViT model demonstrated good performance in identifying microangiopathic changes, with an area under the curve (AUC) ranging from 81.8% to 84.5%.
• Rheumatologists achieved higher average accuracy and a superior balance of sensitivity and specificity compared to the ViT.
• Variability in rheumatologist performance was observed, with one in four showing classification measures equal to or lower than the ViT.

Conclusions:

• The ViT is a capable and accessible AI tool for analyzing microangiopathy patterns in NFC images, potentially aiding rheumatologists in generating consistent reports.
• Despite the ViT's utility, expert clinical judgment remains essential for the definitive diagnosis of the scleroderma pattern in individual cases.