Skin Image Analysis for Erythema Migrans Detection and Automated Lyme Disease Referral

This study develops approaches for the automated referral of individuals with Lyme disease using erythema migrans rash (EM) images with clinical-grade or ‘in the wild’ characteristics. We develop a pre-screener using a Deep Convolutional Neural Network (DCNN) that classifies EM vs. other conditions, including either control/unaffected skin, or skin presenting with other confuser lesions. We test and report performance metrics for the proposed approach on this dataset including Cohen’s Kappa coefficient, area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity. The machine classification yields accuracy (and error margin) of 93.04% (1.49), AUC of 0.9504 (0.0156), and Kappa of 0.7549 (0.0586), which is a significant improvement over previously published state-of-the-art methods. Results also suggest substantial agreement between machine and expert clinician annotated gold standard images. The DCNN model developed for this skin classifier is made publicly available and can potentially be used by others for transfer learning to other types of skin lesion classification models including those for skin cancer.