Detecção da psoríase utilizando visão computacional: Uma abordagem comparativa entre CNNS e Vision Transformers

Abstract

This work performs a comparative analysis between Convolutional Neural Networks (CNNs) and Vision Transformers (ViT) models on the task of multiclassifying skin images, detecting psoriasis among other images of skin diseases visually similar to it, such as dermatitis, lichen planus and pityriasis rosea. The aim of the experiment is to identify the most suitable model for detecting psoriasis, so that it can be used as a diagnostic tool for the disease. During implementation, each model is pre-trained from ImageNet and adapted to the collected image set. During the comparative analysis, it is observed that both CNN and ViT architectures present excellent prediction metrics in the experiment, but ViTs stand out by achieving slightly better performance with significantly smaller models. Finally, DaViT-B (Dual Attention Vision Transformer - Base) stood out from the rest and was therefore selected as the most suitable model for detecting psoriasis, with an f1-score of 96.4%.