RESUMO
INTRODUCTION: Melanocytic nevi present microscopic patterns, which differ in their associated melanoma risk, and can be non-invasively recognized under Reflectance Confocal Microscopy (RCM). AIMS: To train a Generative Adversarial Network (GAN) deep-learning model to produce synthetic images that recapitulate RCM patterns of nevi, enabling reliable classification by human readers and by a Convolutional Neural Network (CNN) computer model. METHODS: A dataset of RCM images of nevi, presenting a uniform pattern, were chosen and classified into one of three patterns - Meshwork, Ring or Clod. Images were used for training a GAN model, which in turn, produced synthetic images recapitulating RCM patterns of nevi. A random sample of synthetic images was classified by two independent human readers and by a CNN model. Human and computer-model classifications were compared. RESULTS: The training set for the GAN model included 1496 RCM images, including 977 images (65.3%) with Meshwork pattern, 261 (17.4%) with Ring and 258 (17.2%) with Clod pattern. The GAN model produced 6000 synthetic RCM-like images. Of these, 302 images were randomly chosen and classified by human readers, including 83 (27.5%) classified as Meshwork, 131 (43.4%) as Ring, and 88 (29.1%) as Clod pattern. Human inter-observer concordance in pattern classification was 91.7%, and human-to-CNN concordance was 87.7%. CONCLUSIONS: We demonstrate feasibility of producing synthetic images, which recapitulate RCM patterns of nevi and can be reproducibly recognized by human readers and by deep-learning models. Synthetic image datasets may allow teaching RCM patterns to novices, training of computer models, and data sharing between research centers.
