AI-Assisted Automated Two-Stage Patch Test Interpretation System Using Vision Transformer

Patch testing, the gold standard for allergic contact dermatitis (ACD) diagnosis, suffers from inter-observer variability. We developed a Vision Transformer-based automated patch test interpretation system.

This retrospective two-centre study included 424 patients (734 images) using the Korean Standard Series (25 allergens). Four model variants-3-class and 6-class classification, each with one-stage and two-stage (binary segmentation + classification) processing-were compared. Inter-rater agreement was assessed using Cohen's κ. Performance metrics included accuracy, balanced accuracy, macro F1-score, PPV, and NPV, with 95% CIs by bootstrap resampling.

Inter-rater agreement showed substantial variability (κ: 0.393-0.557). DeiT demonstrated generally favourable performance compared to CNN architectures in binary classification, achieving higher accuracy than Xception (96.1% vs. 92.3%) and substantially higher recall than ResNet-50 (88.0% vs. 66.9%). For 3-class classification, one-stage and two-stage approaches performed comparably (accuracy 92.9% vs. 92.8%, balanced accuracy 88.0% vs. 86.8%, all p > 0.05). For 6-class classification, two-stage processing was superior (accuracy 91.6% vs. 86.0%, balanced accuracy 70.1% vs. 67.5%, p < 0.001). Institution-stratified analysis confirmed robustness across different test marking conventions.

The 3-class approach offers robust performance with implementation flexibility, whilst 6-class classification requires two-stage processing. Consistent performance across a dataset with heterogeneous imaging conditions supports real-world applicability in ACD management.