Related Experiment Video
Updated: Jun 30, 2026
05:10
An Automated Method for Assessing Visual Acuity in Infants and Toddlers Using an Eye-Tracking System
Published on: March 17, 2023
AI-Assisted Automated Two-Stage Patch Test Interpretation System Using Vision Transformer
Jin Ju Lee1, Yon Soo Jeong2, You Won Choi1
1Departments of Dermatology, Ewha Womans University College of Medicine, Seoul, Korea.
Contact Dermatitis
|June 29, 2026
Summary
Automated patch testing using a Vision Transformer system improves allergic contact dermatitis diagnosis by reducing inter-observer variability. The 3-class model offers flexible implementation, while the 6-class model requires two-stage processing for optimal results.
Area of Science:
- Dermatology
- Medical Imaging
- Artificial Intelligence
Background:
- Patch testing is the standard for diagnosing allergic contact dermatitis (ACD).
- Current patch testing methods suffer from significant inter-observer variability.
- Automated interpretation systems are needed to improve diagnostic consistency.
Purpose of the Study:
- To develop and evaluate a Vision Transformer-based automated system for patch test interpretation.
- To compare the performance of different model variants for ACD classification.
- To assess the system's robustness across varied imaging conditions.
Main Methods:
- A retrospective study of 424 patients with 734 patch test images.
- Four Vision Transformer model variants were tested: 3-class and 6-class classification, each with one-stage and two-stage processing.
- Performance was evaluated using accuracy, balanced accuracy, F1-score, PPV, NPV, and inter-rater agreement (Cohen's κ).
Main Results:
- Inter-rater agreement in manual interpretation showed substantial variability (κ: 0.393-0.557).
- The Vision Transformer system achieved high accuracy, outperforming CNNs in binary classification.
- For 6-class classification, a two-stage approach significantly outperformed a one-stage approach (accuracy 91.6% vs. 86.0%).
Conclusions:
- The 3-class automated patch test interpretation system demonstrates robust performance and implementation flexibility.
- A two-stage processing approach is recommended for 6-class classification.
- The system's consistent performance across diverse imaging conditions suggests strong real-world applicability for ACD management.
