Triagem de Covid-19: um Sistema com múltipla Interpretabilidade baseado em deep learning capaz de determinar a severidade da lesão pulmonar

Abstract

COVID-19 is a pandemic-level disease that has taken thousands of lives around the world. Since the beginning of the pandemic, research has been carried out seeking for immunization (vaccine), screening and diagnosis by researchers in theMedical area, and also for automatic diagnosis by the areas of Engineering, Computer Science and Statistics, using computational intelligence methods to identify infected patients. In this work, materials (data) and theoretical information are gathered on the main techniques available for building a complete patient triage system, based on automatic analysis through computational intelligence models. As material, lung tomography images were used, which were segmented looking for the region of interest (lung parenchyma), using semantic segmentation models. Then these images were used to train the Vision Transformer model, which is based on attention mechanisms that allows the explanation of its classifications. Finally, segmentation models were used to identify the two types of lesions that are commonly generated by the action of the virus in the lung, ground glass opacity and consolidation. The experimental results for the lung segmentation models reached a Dice index of approximately 97%. While the lesion segmentation model achieved a Dice index of approximately 85% for consolidation, and a Dice index of approximately 77% for ground glass opacity. The classification model reached a recall score of approximately 91%, precision and specificity of approximately 98%.