loading page

Allergic Contact Dermatitis Detection with Machine Learning
  • +3
  • Kyriakos S Panagiotidis,
  • Anna Tagka,
  • Ioannis A Vezakis,
  • Ioannis Kakkos,
  • Aikaterini Kyritsi,
  • George K Matsopoulos
Kyriakos S Panagiotidis
Biomedical Engineering Laboratory School of Electrical and Computer Engineering, National Technical University of Athens Athens

Corresponding Author:[email protected]

Author Profile
Anna Tagka
First Department of Dermatology and Venereology "Andreas Syggros" Hospital National and Kapodistrian, University of Athens Athens
Ioannis A Vezakis
Biomedical Engineering Laboratory School of Electrical and Computer Engineering, National Technical University of Athens Athens
Ioannis Kakkos
Biomedical Engineering Laboratory School of Electrical and Computer Engineering, National Technical University of Athens Athens
Aikaterini Kyritsi
First Department of Dermatology and Venereology "Andreas Syggros" Hospital National and Kapodistrian, University of Athens Athens
George K Matsopoulos
Biomedical Engineering Laboratory School of Electrical and Computer Engineering, National Technical University of Athens Athens

Abstract

Allergic contact dermatitis (ACD) is a prevalent immune-mediated skin condition, affecting millions of people worldwide. Significant diagnostic challenges occur due to the subjectivity inherent in the current diagnostic approach, which involves skin patches. To address this limitation, the present study explores the potential of machine and deep learning algorithms in automating ACD diagnosis, thereby facilitating more objective and accurate assessments. A dataset comprising 1579 skin patch images from 200 patients was collected, to train and evaluate the proposed diagnostic models. The dataset underwent extensive feature extraction, resulting in 732 distinct features. These features were utilized to train traditional machine learning models, such as Random Forest, Support Vector Machines, and XGBoost, with the objective of identifying correlations related to ACD. In a second approach, Convolutional Neural Network (CNN) architectures such as EfficientNet, ResNet, and MobileNet, were evaluated in recognizing patterns in different image types, such as Texture and Redness, correlated with ACD cases. The results have indicated that machine learning algorithms can achieve a success rate of 83% in ACD detection, with the fusion algorithm of the two approaches boosting the success rate to 85%. The significance of this research lies in the enhancement of overall diagnostic accuracy achieved through the combination of information from various sources. It highlights that machine learning and CNNs can automate ACD diagnosis, making it more objective and efficient. This advancement can greatly assist clinical diagnosis, and benefit regions with limited medical resources and many patients worldwide.