Document resource
Introduction Objective: This study aimed to evaluate the diagnostic performance of a pre-trained Vision Transformer (ViT) model on nailfold capillaroscopy images and compare it to expert rheumatologist assessments.Material and Methods We retrospectively analyzed 104 anonymized nailfold capillaroscopy images (23 normal, 81 pathological) sourced from publicly available datasets. A general-purpose ViT model (‘google/vit-base-patch16-224’), trained on ImageNet, was applied without any fine-tuning. Two board-certified rheumatologists independently evaluated the same image set. Diagnostic accuracy and inter-rater agreement (Cohen’s Kappa) were calculated.Results The ViT model generated 0 clinically applicable predictions out of 104 images (0%). Its outputs were unrelated to medical relevance (e.g., ‘accordion,’ ‘nematode’). In contrast, the consensus between rheumatologists achieved 94.2% accuracy with a Cohen’s Kappa of 0.827. Individual assessments showed high sensitivity (93.8% and 97.5%) but variable specificity, with one rater overcalling abnormalities. This inter-rater variability underscores the importance of consensus evaluation in capillaroscopic interpretation.Conclusions Pre-trained general-purpose ViT models, when applied without domain-specific adaptation, show no clinical utility in nailfold capillaroscopy. Expert human evaluation remains the most reliable method. Future work should focus on fine-tuning models using capillaroscopy-specific datasets, incorporating explainable AI techniques, and benchmarking against expert consensus to ensure clinical safety and utility in rheumatology.Accepted for publication in Revista da Associação Médica Brasileira (RAMB), October 22, 2025. In press.Abstract P.223 Figure 1